The Effect of Carouseling on MEMS IMU Performance for Gyrocompassing Applications by Benjamin Matthew Renkoski Submitted to the Department of Aeronautics & Astronautics in partial fulfillment of the requirements for the degree of Master of Science in Aeronautics & Astronautics at the MASSACHUSETTS INSTITUTE OF TECHNOLOGY June 2008 @ Massachusetts Institute of Technology 2008 . All rights reserved. Author .................... .. . ......"...... ,, . Department of Aeronautics & Astronautics May 23, 2008 I. // Certified by................ "'~"'MI' ,< i/ I - Matthew Bottkol Charles Stark Draper Laboratory Thesis Supervisor Certified by.................. Emilio Frazzoli Associate Professor of Aeronautics & Astronautics Thesis Supervisor I, 4 , I(\ Accepted by .............. d L. Darmofal Chair, Committee on Graduate Studies SDd MASSACHUSETTS IN OF TECHNOLOGY JUN 2 4 2009 ARCHIVES LIBRARIES The Effect of Carouseling on MEMS IMU Performance for Gyrocompassing Applications by Benjamin Matthew Renkoski Submitted to the Department of Aeronautics & Astronautics on May 23, 2008, in partial fulfillment of the requirements for the degree of Master of Science in Aeronautics & Astronautics Abstract The concept of carouseling an IMU is simulated in order to improve the accuracy of MEMS IMUs. Carouseling consists of slewing the IMU through a pre-determined trajectory that is selected based on inherent properties that lead to improved performance. MEMS devices typically have far more uncertainty than standard inertial measurement devices, yet are considerably less massive and require less power, so implementing this carouseling scheme could make the use of these lightweight systems possible even in high-accuracy situations, such as gyrocompassing. In gyrocompassing, the most significant benefit provided by the carouseling scheme is the reduction in the error contribution of gyroscope bias, as this error is almost completely eliminated. Additionally, it was found that although implementing the carouseling scheme required the addition of error states to account for the size effect, in many cases these error states may not be necessary. Overall, the carouseling of the MEMS IMUs was shown to be effective in reducing azimuth error covariance significantly. Thesis Supervisor: Matthew Bottkol Title: Charles Stark Draper Laboratory Thesis Supervisor: Emilio Frazzoli Title: Associate Professor of Aeronautics & Astronautics Acknowledgments First I would like to thank the Charles Stark Draper Laboratory and the Massachusetts Institute of Technology for giving me such an incredible opportunity to broaden my academic horizons. I am profoundly grateful that I had the privilege to attend the most prestigious engineering school in the world and work for one of the most respected labs in aerospace. I would like to thank my thesis advisor Matthew Bottkol for his dedicated support and invaluable guidance during the creation of this thesis. In addition, I would like to thank Olivier de Weck and Emilio Frazzoli, my advisors at M.I.T. who provided me with wisdom and encouragement along the way. Numerous professors and personell at M.I.T., the University of Missouri, and Draper provided me with innumerable amounts of help that assisted me on my journey, and a special thanks goes to the late Michael Ash, who I never had the pleasure to meet, but who did me the great service of bringing me in to the Draper Laboratory. I thank my parents, Angela and Matthew Renkoski for almost 25 years of unwavering support and love. They taught me the benefits of hard work and instilled in me the drive to succeed and do my best in all my endeavors. There is no doubt I never would have accomplished the completion of this thesis if not for their positive and nurturing influence. A great amount of thanks goes to all my family and friends for the encouragement and kindness they have constantly shown me. I can not imagine a more caring or thoughtful group of people to help me through my thesis or my life. In particular I would like to thank Julie Haesemeier for her unconditional support and dedication. Throughout this thesis writing process she kept my spirits high and was a constant source of inspiration. This thesis was prepared at The Charles Stark Draper Laboratory under IR&D and Contract GC 009256, sponsored by the United States Army Night Vision and Electronic Sensors Directorate at Ft. Belvoir, VA. Publication of this thesis does not constitute approval by Draper or the sponsoring agency of the findings or conclusions contained herein. It is published for the exchange and stimulation of ideas. THIS PAGE INTENTIONALLY LEFT BLANK Contents 1 2 Introduction 1.1 Thesis Objective 1.2 Carouseling Background ................... 1.3 Background on MEMS Instruments . .................. ................... ......... . 9 10 ...... 11 Carouseling Description 15 2.1 Motivation ................... 2.2 The Baseball Stitch Slew ................... 2.3 Analytical basis for Carouseling ................... ........... .. 15 ...... 16 .. 19 2.3.1 Gyroscope Bias Elimination ................... 19 2.3.2 Accelerometer Bias Elimination ................. 23 2.4 Expanding on the Baseball Stitch Slew ................. 25 2.5 Comparing Variations of the Baseball Stitch Slew . .......... 28 2.6 Modeling Carouseling ................... 2.7 Error State Navigation Filter 2.8 3 9 ...... ................... 2.7.1 Covariance and State Propagation . ............... 2.7.2 Kalman Filter Update Size Effect Compensation ... 42 44 48 ................... ................... . ... 49 ..... 51 Gyrocompassing Analysis 57 3.1 Basic description of gyrocompassing ................... 57 3.2 Kalman Filter Update 3.3 Simulation Details ...... ................... .......... ....... ....... 58 . 59 3.4 Azimuth Covariance Determination Results . .......... 3.5 Sensitivity Analysis . 3.6 4 . ..... ..... 3.5.1 Error Budget Formulation . ........ 3.5.2 Sensitivity Analysis Results ..... . . 60 . . . 69 ........... 70 ................ ... Size Effect Sensitivity ................... ...... 71 .. Conclusions 77 85 4.1 Carouseling Effect on Azimuth Accuracy .... 4.2 Size Effect ................... 4.3 Other Considerations ................. 4.4 Future Work ........ ............ 85 .......... ................. .. ......... ... 86 87 . .. . . 87 A MEMS IMU Error Specifications 89 Bibliography 89 List of Figures 93 List of Tables 96 Chapter 1 Introduction 1.1 Thesis Objective The objective for this thesis is to demonstrate the utility of implementing a carouseling technique to improve Inertial Measurement Unit (IMU) navigation performance. The main goal is to see the effect of carouseling in conjunction with Micro-ElectroMechanical System (MEMS) instruments to see if the improvements in accuracy can lead to the utilization of these instruments in higher-accuracy applications such as gyrocompassing. This is highly desirable as MEMS instruments are lighter and cheaper than standard IMUs, yet have significantly less accuracy. First the carouseling concept will be investigated analytically in order to explore the theoretical benefits associated with it. Then, to ascertain if carouseling improves performance, the carouseling maneuver will be simulated and navigated with the results compared to those of a standard navigation system. The remainder of the first chapter of this thesis will provide explanatory background information regarding the carouseling scheme in general and MEMS instruments, particularly those developed by Draper. In the second chapter, a more thorough investigation of the carouseling maneuver will be undertaken including an in depth analysis of its theoretical benefits. The process for simulating the maneuver and the navigation scheme including the error filter shall be described in detail. The third chapter focuses on the results of the carouseling maneuver when applied to gy- CHAPTER 1. INTRODUCTION rocompassing. This includes covariance analysis and sensitivity analysis, including error budget comparisons. Specifically, the error contributions of the size effect involved with the addition of lever arm states is investigated. Conclusions and ideas for future work are included in chapter four. 1.2 Carouseling Background The concept of carouseling, also known as "slewing" or "maytagging", is a concept that has been developed for use in high-precision scenarios using high accuracy IMUs. The basic operation is to put gimballed IMUs through an inertially referenced slew. The main design of the slewing maneuver is to direct the instruments so that they are pointed in different directions so that the total directional effect is averaged out over time to prevent the build up of biases (the analytical basis and equations of motion regarding this will be examined more thoroughly in Chapter 2). Theoretically the bias errors of both the gyroscopes and accelerometers will vanish at the end of each period since they have not been allowed to build up in any one direction. However, since the IMUs now have an added angular velocity and are not collocated, this could add errors that previously were negligible that could cancel out the improvements made in the biases. Therefore it would be advantageous to fully simulate and investigate the effects of the carouseling maneuver. There is growing interest at Draper and in the navigation community for use of this estimation scheme with current and future MEMS applications. If it is effective in reducing error in MEMS IMUs, it could be a substantial enough improvement so that MEMS devices could be used in place of standard IMUs with little or no loss in quality. 1.3. BACKGROUND ON MEMS INSTRUMENTS 1.3 Background on MEMS Instruments Micromachined silicon inertial sensors are much smaller and less expensive to produce than standard IMUs. These characteristics make them extremely attractive for reallife systems; especially certain aerospace and military applications where mass is at a premium. MEMS IMUs can be batch produced using semiconductor manufacturing processes which reduces the cost significantly. Additionally, silicon is inexpensive itself and each silicon wafer can contain a large magnitude of inertial measurement devices. The smaller size of the instruments (only a few square millimeters) not only decreases total mass and volume taken up by the instruments, but also leads to reduced power usage; a highly desirable attribute in many applications. All these advantages makes MEMS IMUs an area of much interest, but their lower accuracy has limited widespread use so far. Draper Laboratory has been developing MEMS sensors for over 15 years, and has become a leader in their development. Draper MEMS instruments have been implemented in various military and space scenarios, including being tested in the use of guiding gun-launched munitions and being designed for use on the Space Shuttle. The MEMS gyroscope Draper has developed a silicon planar tuning fork gyroscope, photographed by a scanning electron microscope (SEM) in Fig. 1-1. This model MEMS gyroscope consists of an etched silicon piece anodically bonded to a glass substrate [9]. Like all tuning fork gyroscopes, it contains two vibrating proof masses suspended by beam springs. In this MEMS case, the two symmetrical proof masses are flexurally suspended over capacitor electrodes in the glass substrate. When the device experiences a rotation about its input axis, parallel to the proof mass plane, the Coriolis effect causes the masses to oscillate up and down out of the plane 16]. The amplitude of the resultant motion, which is proportional to the angular rate, is measured by the capacitive plates underneath the proof masses in the glass substrate. This signal is amplified and demodulated by the integrated electronics to provide an output voltage proportional to the angular rate. CHAPTER 1. INTRODUCTION Figure 1-1: SEM photo of MEMS tuning fork gyroscope Draper has also developed MEMS accelerometers which are pendulous mass displacement accelerometers 16]. A MEMS 100-g accelerometer is shown in Fig. 1-2. Figure 1-2: SEM photo of MEMS accelerometer The accelerometer is also manufactured using a dissolved wafer silicon on insulator process that results in an unbalanced proof mass suspended by torsional spring flexures. The proof mass is suspended in a see-saw type configuration so that when there is an acceleration felt by the accelerometer, the proof mass rotates around the 1.3. BACKGROUND ON AMEMS INSTRUMENTS hinged axis. This motion is measured by capacitive plates in the glass substrate and is proportional to the acceleration, similar to the MEMS gyroscope. This output also provides the input for a closed loop control operation that is used to drive torquer plates to rebalance the proof mass. These MEMS devices are becoming accurate enough for use in many systems, with improvement continuing. They have been used in guiding projectiles, avionics, and numerous automotive applications. New high aspect ratio etch technology and highly integrated fabrication techniques are increasing the overall performance, theoretically for use in more high-performance tactical applications in military and space systems. If found to be effective, the carouseling concept could be used to further this performance improvement. CHAPTER 1. INTRODUCTION THIS PAGE INTENTIONALLY LEFT BLANK Chapter 2 Carouseling Description 2.1 Motivation As previously discussed, MEMS instruments provide many desirable attributes for implementation into high-precision applications. However, the desired performance has not yet been achieved in the instruments themselves, so alternative ways to improve performance have been developed. One of these concepts is the carouseling scheme which would theoretically improve accuracy and retain the high weight and cost saving associated with MEMS devices. The main goal of implementing an estimation scheme using carouseling is to eliminate errors due to accelerometer and gyroscope biases. As described earlier, the basic theory is that if the IMU was spun around in a certain trajectory which will average out the pointing of the IMU so that no direction is favored, the build-up of gyroscope biases would be eliminated. If this process can improve accuracy enough, MEMS devices could then be implemented in higher accuracy applications. There would be additional mass associated with the slewing device, but this could be minimized so that the total system would still be significantly smaller and require less power than a conventional IMU. This chapter will investigate the analytical basis for the performance improvement and describe the kinematics of the carouseling scheme. First the slewing maneuver CHAPTER 2. CAROUSELING DESCRIPTION itself will be introduced and its effect on the contributions from gyroscope and accelerometer error shall be analyzed and simulated. Variations of the original slew will be developed and compared to select the optimum slewing motion to eliminate the most error. Then the equations used to simulate the slewing maneuver with error contributions will be developed as well as navigation equations, including filter updates. Finally, the complications of the addition of lever arms and the size effect will be investigated more fully. 2.2 The Baseball Stitch Slew Carouseling is meant to reduce gyroscope and accelerometer biases by averaging each axis to be pointing in all directions for an equal amount of time. The concept is that since biases are directional errors that build up along an axis, if the overall direction is averaged out by the slewing trajectory, the net effect of the bias errors will go to zero. If the IMU can then be made to follow a trajectory that properly averages out this effect, then the bias contribution to the total error would vanish at the end of each period. Therefore, the first task is to select a trajectory that accomplishes the desired effect. There are a number of slews that achieve this characteristic, the most common being what is called the Baseball Stitch Slew, so named for its resemblance to the path of the stitching along a baseball. The baseball stitch slewing maneuver can be mathematically described as the combination of two orthogonal rotations with the angular velocities of the two rotations related by an integer ratio. For example, in the commonly used 2:1 baseball stitch slew, the second gimbal rotation has double the angular velocity of the first. To create this motion, each IMU will be dual-gimballed so it is allowed to follow the specified path controlled by servo-mechanisms operating each rotation. This rotation is performed in the inertial reference franme, which is an important property of the baseball stitch slew. The slew is essentially a rotation from 2.2. THE BASEBALL STITCH SLEW body coordinates to inertial coordinates. Figure 2-1 shows the baseball stitch slew in a three dimensional representation, demonstrating the trace of the endpoint of one axis for the duration of the slew. Now that the concept of the baseball stitch slew is ....... ... ......]... ..... ... . ......... Si..... .......... .. ........ i Figure 2-1: 3-D Trace of a 2:1 Baseball Stitch Slew Depiction of the three dimensional trace of the endpoint of one of the rotating axes, closely resembling the stitching of a baseball established, its properties of importance can be analyzed. The equations governing its motion will be produced and summarized in this section to provide the basis for this analysis. In matrix form, the 2:1 baseball stitch rotation, Rbs, can be represented as the combination of the two orthogonal rotations comprising the slew: Rbs(t) cos(wt) sin(wt) 0 0 0 - sin(wt) cos(wt) 0 cos(2wt) sin(2wt) 0 cos(wt) - sin(wt) 0 0 1 sin(wt) cos(2wt) ( - sin(2wt) cos(2wt) sin(wt) cos(wt) cos(2wt) cos(wt) sin(2Lot) - sin(2wt) cos(2wt) (2.1) CHAPTER 2. CAROUSELING DESCRIPTION where w is the angular velocity of the first rotation matrix and t is time. In this case, it can be seen the first rotation is about the z-axis and the second rotations is at twice the angular rate and performed about the x-axis. It is important to remember this slew is inertially referenced, not Earth-fixed. The baseball stitch slew Rbs(t), has a valuable property which will be addressed later in this section that demonstrates the benefit of its implementation. To find the angular velocity, use the following equation to define the time rate of change of any three-axis rotation: R = R[ x] (2.2) where [ x] is the skew symmetric form of the angular velocity vector, = (w, W2 , W3). The skew symmetric matrix representation of the angular velocity takes this form: [Ux] = 0 -w3 L2 W)3 0 -L)1 --W)2 L1 (2.3) 0 Again, this rotation R is the transformation from the body axis to the inertial axis and the angular rates are in the inertial frame. This is important as gyroscopes measure inertial rates, since they are inertial instruments. Using the rotation Rbs(t) from Eq. 2.1 as the rotation matrix in Eq. 2.2, the vector 0 can be solved for to yield 2w Qbs = csin(2ct) . (2.4) w cos(2wt) This angular rate vector Obs is the output of an ideal gyroscope which is fixed to a body that executes the baseball stitch slew with respect to inertial space. 2.3. ANALYTICAL BASIS FOR CAROUSELING 2.3 Analytical basis for Carouseling In this section the knowledge gained from Section 2.2 will be used to show how gyroscope and accelerometer biases can be eliminated using the baseball stitch slew. Basic equations of motion will be introduced that will be implemented later into the navigation scheme, but the main focus of this section will be solely on the effect of the baseball stitch on these equations. The overall navigation scheme will be fully described in Section 2.7. 2.3.1 Gyroscope Bias Elimination In order to better understand the effect of slewing on gyroscope bias error, it is important to see what makes up the gyroscope error contribution to angular velocity, defined here as 6Q. To find this error, first linearize Eq. 2.2, where R(t) is the body to inertial coordinate transformation, to result in b = 6R[ x] + R[6 x] (2.5) where the gyroscope error 6Q is the error in sensed angular rate. It can be assumed that the error in the angular rotation can be represented as an error rotation matrix in body coordinates, called T. This can be defined by Eq.2.6: 6R = R[ix] (2.6) This can be substituted into Eq. 2.5 to yield R[,x] + R[Qx][*x] = R[*x][x] + R[6x] (2.7) CHAPTER 2. CAROUSELING DESCRIPTION which can be be converted into a vector equation in body coordinates as I = -Q x q + mO. (2.8) To express the error in inertial coordinates, differentiate the equation I = R(t)F (2.9) where I' is the error rotation in inertial coordinates. After differentiating Eq. 2.9 and substituting the result of Eq. 2.8, this results in d dt = RW+ MI 4R = R(-Q x x, + 6Q) + R[ x]q, = R(-Q x xI + t x * + 6Q) = R(t)6Q2 (2.10) Therefore the total contribution from the gyroscope error to the error rotation in the inertial frame is simply the gyroscope error multiplied by the rotation matrix. Now the gyroscope error 6Q, can be broken down into its components to further examine the contribution to the error. The gyroscope error can be assumed to be consisting of four main error sources: gyroscope bias error bg, gyroscope scale factor sfg, gyroscope misalignment m,, and gyroscope non-orthogonality ng. Each of these quantities are vectors consisting of 3 scalar values (one in each direction of the body frame). The errors are inherent to the gyroscope itself and are a function of the error characteristics of that gyroscope. They can be combined to form the overall gyroscope error in the following equation: S1 = bg + ([sfg.] + [mgx] + [ng(]) Q (2.11) 2.3. ANALYTICAL BASIS FOR CAROUSELING where [sfg,] represents a 3 x 3 diagonal matrix formed from the 3-vector containing the gyroscope scale factor errors, sfg = (sfl, sf 2 , sf3), in the following way: 0 0 0 s f2 0f3 0 0 sf3 sfi [sfg] = (2.12) [m, x] is a 3 x 3 skew-symmetric matrix formed from the three misalignment error terms of the 3-vector, m 9 = (ml, m2, m3): 0 m2 -m3 0 -m2 -m 1 (2.13) 0 M1 and [ng® ] is a symmetric 3x3 matrix created from the non-orthogonality terms of the 3-vector n, = (ni, n2, n 3 ), such that 0 [ng®] = n2 (2.14) n3 n2 1 7i 0J Now that these matrices have been defined, the change in the rotation error, AI', can be calculated by integrating Eq. 2.10 from time zero to some discrete time T, and combining it with Eq. 2.11: A I - = jI/R(t)60dt + TR(t) (b([sfg.] + [mgx] + [ng0]) Q) dt (2.15) Equation 2.15 can be rewritten by distributing the diagonal, skew-symmetric, and symmetric matrix representations to the angular velocity vector Q, instead of the CHAPTER 2. CAROUSELING DESCRIPTION scale factor, misalignment, and non-orthogonality error vectors, respectively. The results of this re-distribution create Eq. 2.16: ,A* ()(bg + [.]sfg - [lx]m 9 + [ = ]ng) (2.16) This puts the equation into a form that can be more easily used to demonstrate the effect the slewing motion has on the rotation error. Notice that the diagonal and symmetric operations can switch to apply to the vector 0 and cause no change, while changing the skew-symmetric operation from the misalignment vector to Q results in a change in sign. Now the gyroscope error source vectors can be completely separated from the integral, since they are assumed to be constant, and Eq. 2.16 can be rewritten as Ak I = SB(T)bg + SSF(T)sfg + SA,(T)mg + SN(T)ng (2.17) where the matrices of the form Sx are known as "sensitivity matrices" and are integrals used to quantify the effect of the four error sources on the rotation error matrix. The sensitives are all integrals over the time interval 0 - T that result in 3 x 3 matrices. The bias, scale factor, misalignment, and non-orthogonality sensitivity matrices can be respectively described as: SB(T) = S(,(T) = SA,(T) = - S (T) J T /T (2.18) [0( )(t)dt, (2.19) R(t) [Q(t) x]dt, (2.20) R(t) [ (t)(] (t)dt. (2.21) The baseball stitch slew, with rotation matrix Rb, attained from Eq. 2.1, is particu- 2.3. ANALYTICAL BASIS FOR CAROUSELING larly interesting and useful because it has the property, SB(T) = Rbs (t)dt = 0. (2.22) This mathematically expresses the fact that the overall accumulation of the rotation Rbs is zero, meaning over time T no one direction is favored. To prove this, it can be seen that integral of each element in the baseball stitch rotation matrix from Eq. 2.1 equals zero when taken over the time of one period. The integral of the trigonometric functions sin and cosine over one period is always equal to zero and using the property of orthogonality of the trigonometric functions, the integral of the other elements of the rotation matrix will also be equal to zero. Since this entire integral is zero, the sensitivity matrix in Eq. 2.18 is equal to zero, and it eliminates the contribution of gyroscope bias to the overall rotation error A*, in Eq. 2.15. Since the bias sensitivity matrix goes to zero, no matter how large the gyroscope bias error, its contribution will always vanish at time T when the baseball stitch slew is performed. In this way the baseball stitch slew is shown to achieve the goal of eliminating the contribution from the gyroscope bias error, but no other error source. The elimination of these other gyroscope errors will be investigated in Section 2.4. 2.3.2 Accelerometer Bias Elimination While the previous subsection detailed how gyroscope bias errors can be eliminated by carouseling using the baseball stitch slew, contributions from accelerometer bias to the error in acceleration can also be reduced. In fact, it was for the express purpose of eliminating accelerometer bias error that the slew was developed, and was later found to also eliminate the gyroscope bias. However, in the gyrocompassing application, the attitude error is more important than the velocity error so the gyroscope errors are more crucial. As was seen in Eq. 2.15, there is no contribution from accelerometer error to the CHAPTER 2. CAROUSELING DESCRIPTION rotation error. However, accelerometer error plays a large part in the error in velocity felt by the IMU. In the following equation, derived from equations of motion, can be seen the components of error in acceleration: = G(t)6x - 2[QE x]5v + R[ ×x]f + 6a (2.23) where 6a is the accelerometer error. The derivation of this equation and the components of the first three terms will be described in detail in Section 2.7, but for now the only term of concern is the accelerometer error. It can be assumed that the accelerometer error is made up of the contributions from four sources, much like the gyroscope error: accelerometer bias, scale factor, misalignment, and nonorthogonality. They combine to contribute to the total accelerometer error as follows: aa = R(6ba + [f-]6sfa - [fx]6ma + [f0fn) (2.24) Where R is the body to inertial coordinate rotation used in the gyroscope equations as well. Ignoring the first three terms and integrating Eq. 2.23 with Eq. 2.24 inserted for Sa, the contribution of accelerometer error to the velocity error can be seen: T 6adt v = v = SB(T)ba + / R([f.]sf /T0 -Lfxlma + [f®]n)dt (2.25) where SB(T) is the bias sensitivity matrix from Eq. 2.18 in the previous section, and has been proven to vanish over the time interval 0 - T when using the baseball stitch slew. Therefore, the contribution from accelerometer bias error to the velocity error vanishes as well, no matter what the accelerometer bias value is. Even though the area of research for this thesis is the gyrocompassing application, which is far 2.4. EXPANDING ON THE BASEBALL STITCH SLEW 25 more concerned with the angle error, this increased performance in the velocity error is still a valuable attribute of the carouseling scheme. And although this does not directly increase accuracy of the azimuth angle, when filtering is used and all errors are intermixed, better accuracy in any measurement can contribute to overall accuracy. 2.4 Expanding on the Baseball Stitch Slew As was shown in Section 2.3, the baseball stitch slew is effective in eliminating the contribution of gyroscope bias to the total error rotation I. However, it does not eliminate contribution from other error sources, namely the integrals SsF(T), SM,(T), and SN(T) do not reduce to zero at the end of each period. So while the bias term returns to zero, these other errors continue to accumulate. This is undesirable, but it has been found that by modifying the baseball stitch slew, these errors can also be eliminated. In this section, modifications to the baseball stitch slew will be made that will make it more effective at improving accuracy. In order to provide a goal for the improved slew, the sensitivity matrix format will again be used to determine another possible way for carouseling to provide benefits. To investigate this problem, first create the matrix expression M = [sfg,] + [m,x] + [n9 ®] (2.26) which can be seen in Eq. 2.11 as the 3 x 3 matrix multiplied by the angular velocity to contribute to the gyroscope error. Also notice that this matrix M is given by nine free parameters, so it can be said to represent an arbitrary 3 x 3 matrix with nine free entries. From this knowledge it can be proven that the three sensitivity matrices it is desired to eliminate (SsF(T),SnA(T), and SN(T)), will be equal to zero if and only if =-R(t)MA(t)dt 0. (2.27) CHAPTER 2. CAROUSELING DESCRIPTION for any 3 x 3 matrix Al. That is, no matter the magnitude of Al (in this case the scale factor, misalignment and nonorthogonality errors), this will hold true. So a rotation R(t) must be found that satisfies Eq. 2.27. This equation holds true no matter the value of M, which means no matter how poor the accuracy, with the right rotation, the error contribution for all three of these sources can be eliminated at time T. Now a slew will be created that satisfies the above equation while still satisfying Eq. 2.22 as well. The equations governing this slew will be developed and then it will be proven that it does satisfy equation 2.27. This new slew can be created from the concept of reversing the baseball stitch slew. The IMU will travel through the slew completely, and then travel over the same trajectory in the opposite direction. This was created to "unwind" the effects of scale factor, misalignment, and nonorthogonality, while still eliminating the bias contribution. To investigate the effects of slew reversal, define slew R+(t) to be any slew, and create a piecewise slew R(t) defined as R(t) R±(t) O<t<T R(t) T<t<2T (2.28) where R_(t) is the reverse in time of R+(t) and can be mathematically defined as: R_(t) = R(2T - t) T < t < 2T (2.29) From Eq. 2.2, the time rate of change for the rotations can be defined as R+ (t) = R+(t)[Q(t)x] R_(t) = R (t)[_(t)x] To find 0 t <T T < t < 2T (2.30) (2.31) _(t) in terms of R+ and Q+, differentiate Eq. 2.29 and substitute Eq. 2.30 2.4. EXPANDING ON THE BASEBALL STITCH SLEW to obtain R_(t) = -R±(2T-t) = -R+(2T- (2.32) t)[,+(2T - t)x] Setting Eq. 2.32 equal to Eq. 2.31 and using Eq. 2.29 produces I_ (t) = -R+(2T-t)[,+(2T-t)x] t)[Q+(2T-t)x] R_(t)[_ (t)x] = -R+(2T[Q_(t)x] = -[a+(2T-t)x] = -(t) (2.33) 2+(2T- t) over the interval T < t < 2T. Using this relation, it can be shown that Eq. 2.27 is true for any trajectory that undergoes a slew reversal as described in Eq. 2.28. Take Eq. 2.27 and use the definitions of the rotation matrix and angular velocity with slew reversal to show R(t)M 2T R(t)M R+(t)M +(t)dt + (t)dt R+(t)MQ+(t)dt - = R+(t)M +(t)dt + = 1 2T (t)dt (2.34) R+(2T - t)MA+(2T - t)dt R+(t)()MOa(t)dt 0 Notice that this is true for any 3x3 matrix M and by Eq. 2.27, SF(2T),SI(2T), and SN(2T) will all be equal to zero when there is a slew reversal. Notice also that this holds for any rotation R(t), not necessarily the baseball stitch slew. However, since it is based on the baseball stitch slew, Eq. 2.22 still holds true and the contribution due to both accelerometer and gyroscope biases will continue to vanish at the end of CHAPTER 2. CAROUSELING DESCRIPTION 28 each period. Combined with the reversal, the contribution of all four gyroscope error sources is now seen to vanish. Comparing Variations of the Baseball Stitch Slew 2.5 To better compare the effects of the baseball stitch slew and its derivatives in terms of their contribution to T, this section will focus on exploring the error contributions in Eq. 2.17 graphically for each slew. For the regular baseball stitch slew, first observe the motion of the gimbal angles over a period of 60 seconds (implementing Eqs. 2.1 and 2.4). Carouseling Gimbal Angles ~%U '4 <2 0 10 20 30 40 50 Angular rate vector in Body Frame -50 1 0 10 20 30 Time (s) 40 50 Figure 2-2: Plots of gimbal angles and angular rates of a 60 second Baseball Stitch Slew It can be seen in the first subplot that the gimbal angles continue to increase 2.5. COMPARING VARIATIONS OF THE BASEBALL STITCH SLEW 29 as the slew progresses an the gimbals continue to rotate constantly in one direction, with the outer gimbal increasing at twice the rate as its rotation has twice the angular velocity. The second subplot shows the angular velocity components for the entire slew. Notice two of the angular velocity components follow sinusoidal behavior while the third element remains a constant equal to the angular rate of the second slew, as was shown in Eq. 2.4. Now performing the integrations of the sensitivity matrices in Eq. 2.17, the total contribution to the rotation error can be plotted for the baseball stitch slew. Values for the gyroscope bias, scale factor, misalignment, and nonorthogonality errors are assumed to be those of the current Draper MEMS IMU model, which will be discussed later in Section 3.3 (the specifications can be found in Table A.1 in Appendix A). The exact values of the errors are not important at this point, but rather the pattern of behavior of the error contributions over time for one period. In Fig. 2-3 is plotted the contributions of each error term to the 9 elements of the AI matrix in Eq. 2.17 for the baseball stitch slew. Notice the bias terms vanish at the end of the period as was shown from Eq. 2.22. Also the misalignment and non-orthogonality contributions do vanish, but two elements of the scale factor error continue to build up. Note that the bias, misalignment, and non-orthogonality errors not only vanish at the end of the period, but also every 15 seconds which is equal to one fourth of a period. As can be seen in Fig. 2-2, this corresponds to the times when both gimbals have performed a whole number amount of revolutions. Also observe that the scale for the plot of the error contribution due to gyroscope scale factor is an order of magnitude larger than the scale use for the other three error sources. These errors combine to form the total rotation error shown here in Fig. 2-4. In Fig. 2-4 it can be seen that most of the error elements return to zero at the end of each period, but two elements attributable to the scale factor error still continue to build up. The errors that do return to zero oscillate in the manner observed before CHAPTER 2. CAROUSELING DESCRIPTION T error due to Gyroscope Bias x 10 .4 10 30 .20 40 50 T'error due to Gyroscope Scale Factor x 10.3 0 !I I | I 20 30 40 50 -2 -41 0 5 10 E Y error due to Gyroscope Misalignment " x 10 5 0 -5 -10 10 20 30 40 50 '1 error due to Gyroscope Non-orthoganality x 10.4 -EJ -21 0 10 20 30 Time (s) 40 50 60 Figure 2-3: Contributions of different error sources to the overall rotation error for the Baseball Stitch Slew 2.5. COMPARING VARIATIONS OF THE BASEBALL STITCH SLEW 0.5 x 103 1 1 10 20 Total Error in y , 0 -0.5 -1.5 -2 -2.5 30 Time (s) 40 50 Figure 2-4: The total contribution to rotation error from gyroscope errors for the Baseball Stitch Slew in which they also return to zero ever one fourth of a period. As a point of reference, compare the above plot to the total error for a non-rotating MEMS IMU (i.e. the rotation matrix is equal to identity in Eq. 2.17), which yields the total error plot seen in Fig. 2-5. First, notice that only one element changes from zero, and that is known to be due to the bias. The other errors contribute no error as the body is not moving. So even though the baseball stitch slew does an excellent job of eliminating the gyroscope bias, by introducing rotation, the other errors become a factor. And even though misalignment and non-orthogonality are also cyclically eliminated, the scale factor error builds up which effectively eliminates the improvements made in reducing the bias. In fact, Fig. 2-4 shows that the additional scale factor error grows slightly CHAPTER 2. CAROUSELING DESCRIPTION 1 x 10.3 Total Error in T 0.9 0.8 0.7 0.6 " 0.5 0.4 0.3 0.2 0.1 10 20 30 Time (s) 40 50 Figure 2-5: The total contribution to rotation error from gyroscope errors for a nonrotating body larger than the original bias error seen in Fig. 2-5. So the normal baseball stitch slew performs well at eliminating bias, misalignment and nonorthogonality, but the addition of the scale factor makes it unattractive for implementation. However, as discussed in Section 2.4, if the baseball stitch slew performs a reversal as defined in Eq. 2.28, the scale factor error will also reduce to zero at the end of each period. For the baseball stitch slew with reversal, the gimbal angles and angular velocity now behave as seen in Fig. 2-6. As can be seen in the figure, both gimbals are reversed halfway through the 60 second slew. Instead of the gimbal angles continuing to increase, they now reverse direction and are in fact both equal to zero at the end of the period. Notice the 2.5. COMPARING VARIATIONS OF THE BASEBALL STITCH SLEW Carouseling Gimbal Angles --- Outer Gimbal I-nner Gimbal .... .... ........... .. .. .... ..... ...... .... Angular rate vector in Body Frame 40 ° ° 20 cc 0 -20 -40 -. ... ... ........ .......... ........ ..... ....... .... .... ..... ....... ... .......... ' | 0 10 | 20 30 Time (s) I ! 40 50 E Figure 2-6: Plots of gimbal angles and angular velocities for Baseball Stitch Slew with Reversal elements of angular velocity reverse sign at the time of slew reversal as well. Once again using Eq. 2.17, but this time integrating the baseball stitch slew with reversal rotation, the contributions to the 9 elements of the A'I matrix can be seen for the baseball stitch slew with reversal in Fig. 2-7. Now all nine elements of the four error source sensitivity matrices return to zero at the end of the slew. This is the exact result desired. It would appear that the baseball stitch slew with reversal would be the best slew to use for carouseling, but there is another aspect that has not been taken into account. Although the focus of gyrocompassing is on the azimuth angle so the velocity error is not the main concern, in the real navigator using a Kalman filter, it is still highly desired to have accurate velocity measurements. So to better ascertain if the CHAPTER 2. CAROUSELING DESCRIPTION T error due to Gyroscope Bias x 10-4 5 3 4 0 2 20 25 30 25 30 - x 10, -10 15 T error due to Gyroscope Scale Factor x 10 10 10 y error due to Gyroscope Misalignment x 5 10 15 20 '1 error due to Gyroscope Non-orthoganality x 10,4 0 -2 0 5 10 15 2n Time (s) Figure 2-7: Contributions of different error sources to the overall rotation error for the Baseball Stitch Slew with Reversal 2.5. COMPARING VARIATIONS OF THE BASEBALL STITCH SLEW 35 baseball stitch with reversal is the optimum slew, it would be valuable to investigate the gyroscope error contribution to the velocity error as well. Recall from Eq. 2.23, that the I error actually does contribute to the velocity error in the third term, which is R['IFx]f. If the result for AkF from Eq. 2.17 is inserted for * ,this can be integrated to determine the contribution of the gyroscope errors to the velocity error (represented as Av in this analysis): Av(T) = = Aq x f(t)dt jT(SB(T)b + SsF(T)sf + SM(T)mg + S(T)ng) x f(t)dt(2.35) In this equation it can be seen that instead of the error contributions merely being the sensitivity matrices, it is now the integral of the sensitivity matrices that determines the contribution to the velocity error. Performing the integrations in Eq. 2.35 for the baseball stitch slew with reversal yields the results shown in Fig. 2-8. Notice the contribution due to gyroscope bias continues to vanish, however the other three sources continue to build up. Particularly the error contribution due to scale factor is a problem, as it is an order of magnitude greater than the misalignment and non-orthogonality errors. So the slew reversal fails to eliminate these errors. However, it has previously been found that this reversal concept can be extended to provide better performance. As there are two separate gimbals, both gimbals do not have to be reversed simultaneously. It was discovered that a concept involving only reversing the outer gimbal every other time would improve upon the performance of the baseball stitch slew with reversal. The plots of the gimbal angles and angular velocities for this scheme, called the baseball stitch slew with reversal extended can be seen in Fig. 2-9. This figure demonstrates how the outer gimbal is not reversed at 30 seconds when the inner gimbal is reversed. Also notice that when the next period starts the inner CHAPTER 2. CAROUSELING DESCRIPTION Error in velocity due to Gyroscope Bias 0.02 -" '--" ' -----".. _. . : - _ _--- - 0 -0.02 - I I 0 Error in velocity due to Gyroscope Scale Factor U. -0.5 0 10 5 15 20 25 31 25 30 Error in velocity due to Gyroscope Misalignment 0.02 0 -0.02 -0.04 3 10 5 15 20 Error in velocity due to Gyroscope Non-orthoganality 0.05 I • 0 I I -0.05 Time (s) Figure 2-8: Contributions of different gyroscope error sources to the change in velocity error for the Baseball Stitch Slew with Reversal 2.5. COMPARING VARIATIONS OF THE BASEBALL STITCH SLEW Carouseling Gimbal Angles - 10 D Outer Gimbal Inner Gimbal 40 30 20 50 61 Angular rate vector in Body Frame i t~I I -50 0 t lt iVI' I~ 30 Time (s) Figure 2-9: Plots of gimbal angles and angular velocities for Baseball Stitch Slew with Reversal Extended gimbal will be reversed but the outer gimbal continues to rotate in the same direction. To verify that this slew continues to provide the nulling capabilities of gyroscope error contribution to A* of the baseball stitch with slew reversal, Eq. 2.17 can again be used to plot out these results for the baseball stitch with slew reversal extended, shown here in Fig. 2-10. Notice that all the error contributions continue to vanish at the end of each period, and also the scale factor contribution vanishes at the 30 second mark. The other three sources show the familiar behavior of all elements returning to zero at each one fourth of a period. It is interesting to note that the error contribution plots for each source can be divided at the 30 second point and the two halves are reverse images of each other. CHAPTER 2. CAROUSELING DESCRIPTION Y error due to Gyroscope Bias x 10-4 1,I~ I 0 I I i 10 20 30 x 10 3 T error due to Gyroscope Scale Factor 10 20 30 40 50 T error due to Gyroscope Misalignment x 10-5 10 2 I 40 20 30 40 50 T error due to Gyroscope Non-orthoganality x 10.4 SI I I tr(0 -203 SI 10 20 i I 30 Time (s) 40 50 E Figure 2-10: Contributions of different error sources to overall rotation error for Baseball Stitch with Reversal Extended 2.5. COMPARING VARIATIONS OF THE BASEBALL STITCH SLEW 39 Now to investigate if the baseball stitch with reversal extended provides any benefit in the velocity error, the error contributions to the velocity error are presented in Fig. 2-11. The scale factor error contribution is now reduced back to zero at the end of the period, along with the contribution from the gyroscope bias. Unfortunately, the misalignment and non-orthogonality errors persist, but their error contribution is significantly less than the previous contribution from scale factor for the simple reversal in Fig. 2-8. These misalignment and non-orthogonality errors are therefore not of critical importance to eliminate, and as will be seen in the simulations, their contributions are minimal and do not greatly impact azimuth error. The total error contribution from gyroscope errors to F for the baseball stitch slew with reversal extended can be seen in Fig. 2-12: Now the contributions from all error sources vanish at the end of the period. The scale factor error does not build up like it did for the normal baseball slew as seen in Fig. 2-4, and the bias terms do not continue to grow as seen when no rotation at all is applied as seen in Fig. 2-5. All errors reduce back to zero and do not grow larger than 8 x 10- 4 radians in magnitude at any point, with most elements remaining significantly smaller than that, including all reducing to zero at the mid-point as well. It is important to note that all of these results are for ideal cases where there is no filtering and no updates whatsoever. In real applications there will be filter updates that will affect the contributions of these error sources. Cross-correlations will develop and other error sources such as random walk are incorporated into the navigation as well. Therefore, the exact results found in this section should not be expected to be repeated in the simulations. However, the belief is that the trends seen here will be reflected in the results, as in, it is expected to see that the carouseling will reduce the effects of gyroscope error sources on angle error. It is not expected that these contributions will be exactly zero or that any of the other errors will show identical behavior to what has been demonstrated in these idealized plots. CHAPTER 2. CAROUSELING DESCRIPTION Error in velocity due to Gyroscope Bias 0.02 Hi 0 -0.02 10 30 20 40 50 i Error in velocity due to Gyroscope Scale Factor 0.5 --L l -- ~---L 0 I i - -0.5 Error in velocity due to Gyroscope Misalignment 0.1 0 -0.1 - I I I Error in velocity due to Gyroscope Non-orthoganality 0.1 --~ 0 -0.1 I i i 1 I C Time (s) Figure 2-11: Contributions of different gyroscope error sources to velocity error for Baseball Stitch with Reversal Extended 2.5. COMPARING VARIATIONS OF THE BASEBALL STITCH SLEW Total Error in Y x 10- 4 0 10 20 30 Time (s) 40 50 60 Figure 2-12: Total Contribution to angle error from gyroscope errors for Baseball Stitch Slew with Reversal Extended Now that the baseball stitch slew with reversal extended as been shown to be superior to the regular regular baseball stitch slew, it will be selected for use in the navigation simulations. There would be no practical reason to implement the regular baseball stitch slew instead of the slew with reversal extended as it performs either the same of better in all areas. As the term "baseball stitch slew with reversal extended" is a long and convoluted phrase, for the remainder of this thesis, the slew will be referred to simply as the "baseball stitch slew", unless otherwise specified. So when the baseball stitch slew's performance is referred to from now on, it is understood to be referring to the baseball stitch slew with reversal extended. CHAPTER 2. CAROUSELING DESCRIPTION 2.6 Modeling Carouseling The theoretical benefits of carouseling have been shown, but the real goal of this thesis is to show what these benefits equate to in a numerical simulation which models the true performance of MEMS IMUs in real-life applications. The first step is to simulate the output of the IMU experiencing the slew with ideal accelerometers and gyroscopes at the center of the rotation. In reality, the IMUs will most likely not be collocated at the center of rotation and this has potentially serious implications, called the size effect, which will be investigated in Section 2.8. This section will be concerned with establishing the equations used to model the baseball stitch slew and generate the IMU outputs angular velocity and specific force. First the ideal outputs will be created (i.e. no accelerometer or gyroscope errors added), then the errors will be incorporated to create a simulated IMU output. The ideal gyroscope output is the angular velocity vector Q, which has been defined for the baseball stitch slew in Eq. 2.4 and for the reversal in Eq. 2.33. The output of an ideal accelerometer is the specific force f felt by the accelerometer in the body frame. This includes the acceleration of the navigation point in the body frame and the gravitational acceleration: f = R' - GB (2.36) where RB is the rotation from the Earth Centered Inertial (ECI or simply, inertial) frame to body coordinates. In this case RB is the inverse of the baseball stitch slew, Rb8 = R . The acceleration of the navigation point, x' in inertial space is defined as 3', which is determined by the trajectory of the overall body, i.e. if the navigated point is standing still on the earth, the only movement will be due to the rotation of the earth (and there will be no movement observed when using local level coordinates). If the IMU is being used for missile guidance, the missile's trajectory and flight path will determine the motion felt by the body. The determination of the spe- 2.6. MODELING CAROUSELING cific trajectory must also be modeled and will be discussed when applying to specific situations. Also note that superscripts on variables indicate the coordinate frame of reference, and for rotations, the subscript indicates the initial frame and the superscript indicates the frame into which the rotation is being performed. The gravitational acceleration contains both the effects of gravity and the centripetal acceleration induced by the Earth's rotation: GB = (g(x) - [QearthX]2XI)R B (2.37) where g(x') is the expression of gravitation depending on the inertial location of the navigation point. This can be approximated by linearizing and ignoring perturbation effects as: g(x) =X (I - 3uuT) (2.38) where p is the universal gravitational constant and u is the unit vector in the direction of x I . The quantity [Kearth X ] is the skew symmetric formed from the earth's rotation vector defined as 0 Qearth = 0 (2.39) [Wearth where Wearth is equal to the angular velocity of the Earth's rotation. As was stated earlier, the specific force and angular velocity are the outputs of ideal inertial measurements, meaning there are no measurement errors. In application, there are accelerometer and gyroscope errors in the measurements. Therefore, these errors must be simulated as well and their contributions taken into account. The simulated outputs, designated as a for accelerometer output and g for the gyroscope output vector CHAPTER 2. CAROUSELING DESCRIPTION are simulated as: a = f + R(ba + [f-]sfa - [fx]ma + [f @]na) + Wa (2.40) g = + b + [Qg.]sf - [x]m, + [W ]ng + wg (2.41) where the subscript a signifies an accelerometer error source and the subscript g signifies gyroscope errors. The bias and scale factor terms are simulated as Markov processes as well as constants and the misalignment and non-orthogonality terms as random variables with covariances determined by the specific IMU. In addition to the instrument errors, random walk effects, w, also affect both the accelerometer and gyroscope output. The random walk can also be simulated as a random process with specified covariances. Using Eqs. 2.40 and 2.41, the output of an IMU undergoing carouseling can be simulated. An estimation scheme must be developed to navigate this data so the performance can be evaluated. 2.7 Error State Navigation Filter In order to navigate the slew (or any standard strap down system) an integration scheme must be developed. In this section the basic concept of the navigator will be described and the equations of motion will be presented and linearized. Subsection 2.7.1 presents the covariance propagation equations and Subsection 2.7.2 will describe the Kalman filtering procedure. The navigator will be designed to take the IMU outputs from the simulation and integrate them back to attain the position, velocity and attitude of the body in the Earth coordinate frame. This Earth-based navigator will use a Kalman filter and compute the gains needed for the external updates. This navigator will specifically keep track of the covariance of navigation errors to determine how accurately the solution is being determined. The main errors desired to be known are in position and velocity errors (Sx and 6v respectively) and attitude error represented by mis- 2.7. ERROR STATE NAVIGATION FILTER alignment vector I. The full list of the 45 error states is shown in Table 2.1 below: Error state Position error Velocity error Attitude error Accel. bias(turn-on) Accel. bias(in-run) Accel. scale factor(turn-on) Accel. scale factor(in-run) Accel. misalignment Accel. nonorthogonality Gyro. bias(turn-on) Gyro. bias(in-run) Gyro. scale factor(turn-on) Gyro. scale factor(in-run) Gyro. misalignment Gyro. nonorthogonality index 1-3 4-6 7-9 10-12 13-15 16-18 19-21 22-24 25-27 28-30 31-33 34-36 37-39 40-42 43-45 Table 2.1: Navigation Filter States The linearized navigation states can be determined by first examining the differential navigation equations (in the Earth fixed frame): = v v RE = (2.42) g(x) - 2[L x]v + REf (2.43) RE[qx] (2.44) These can be linearized for navigation as: 6xC = 6v (2.45) 6, G(t)6x - 2[L x]6v + R E[*x]f + 6a (2.46) -[x]F + 5g. (2.47) S= = CHAPTER 2. CAROUSELING DESCRIPTION where 6a is the accelerometer error and 6g is the gyroscope error. Notice Eqs. 2.46 and 2.47 were used earlier in Section 2.3 for determining the contribution from instrument error sources. G(t) contains the effects of gravity and is equal to (I - 3uuT) - [Qg x] 2 , G(t) = - which is similar to Eq. 2.37, as seen before. (2.48) The three linear equations are implemented in the navigator by integrating them at each time step, with initial conditions applied. The derivations of Eqs. 2.45 and 2.46 are fairly straight forward from Eqs. 2.42 and 2.43 respectively. To obtain the result found in Eq. 2.47, which was used earlier in Eq. 2.8 as well, some derivation is required. First define R as the true rotation (this holds true for any rotation from one frame to another, but in this case applies to the body to earth coordinate transformation), and so R is the estimate for this rotation defined as iR = R(I + [T x]) (2.49) which can be rearranged as R-'jR I + ['x]. (2.50) Taking the derivative of Eq. 2.50 yields R)= [lx] d(R To find (2.51) (R-1 R), first examine derivatives of each part, - dt d1 dt 1R = -[Qx]R - 1 = R[(O + g)x] (2.52) (2.53) 2.7. ERROR STATE NAVIGATION FILTER then take the overall derivative of R- 1R: d(R-) dt = -[Qx]R- 1 + R-'R[( + 6g)x]. (2.54) Substituting the value for R- 1R from Eq. 2.50, results in the following equation: d dt(R ) = - [x ](I + [ = x ]) + (I + [ x ])[(Q + 6g) x] [Tx][x] - [Qx][Wx] + [6gx] (2.55) Combined with Eq. 2.51, the expression becomes [ x] = [*×x][x] - [Qx][9x] + [6gx] (2.56) which can be further simplified using the skew-symmetric property that for any two matrices [ax] and [bx], [ax][bx] - [bx][ax] = [(a x b)x]. (2.57) Using this result, Eq. 2.56 can be re-written and then eliminating the skew symmetric [9x] = [( S= v W' = x x -Q )x]+[ gx] + 6g x + 6g (2.58) which is equal to the result obtained in Eq. 2.47. To determine the accelerometer and gyroscope errors used in Eqs. 2.46 and 2.47, recall that the accelerometer and gyroscope outputs are equal to: a = f + R(ba + [f']sfa - [fx]ma + [fl9]na) + Wa (2.59) CHAPTER 2. CAROUSELING DESCRIPTION g = 0 + bg + [Q.]sfg - [Q x]mg + [Q]n, + w, (2.60) The navigator compensates by subtracting the estimated errors: 6a = Wa (2.61) + [ -.]6sfg - [ x]6mg + [O®]6ng ) + wg (2.62) -R(6ba + [f -]sfa - [fx]6ma + [f®]6na) + dg = -(b, which are simulated as they were in the modeling of the trajectory, but with no knowledge of the true values of the errors. This compensation is done prior to the integration of Eqs. 2.45-2.47. 2.7.1 Covariance and State Propagation Equations 2.45-2.47 describe the propagation of the estimate in the navigator, but the estimate is bound by its covariance. The navigator begins to propagate the covariance with initial estimates for the covariances of the error states in Table 2.1. The initial estimates are formed from the specifications of the specific IMU. These comprise the diagonal of the initial covariance matrix, P, which is a square matrix with dimensions equal to the number of error states. The state and the covariance are updated at each time step, At, by the following equations: X+ = 4X- P+ = I)P-I)T +Q (2.63) (2.64) where X- is the state immediately before the time step and X + is state immediately following the time step. P- is the covariance immediately before the time step and P+ is the updated covariance directly following the time step. Q is the covariance of the process noise and ) is the transition matrix which is estimated as 4 = eFAt (2.65) 2.7. ERROR STATE NAVIGATION FILTER where F is the dynamics matrix and relates the error states to their derivatives: 6X = F6X (2.66) The linearized equations 2.45-2.47 are encompassed into F to propagate each error state. In this way the covariance and state are propagated each time step into Earth coordinates. Notice the navigation is being performed in Earth coordinated even though it was stated earlier that the slew will be inertially referenced. Earth-referenced navigation is easier and more common which is why it is being implemented, while the properties of the baseball stitch slew are inherent to being inertially referenced. The consequence of referencing the slew to the Earth will be investigated later, however. The described navigation filter will perform the integration of the estimate and covariance, but provides no updates to maintain accuracy. To improve state and covariance estimates by incorporating external sources of information, a Kalman filtering scheme can be developed as well. 2.7.2 Kalman Filter Update To navigate using external updates, an extended Kalman filter algorithm can be implemented in the integrated navigation system. The main concept is to combine independent external sources of navigation information with the reference navigation solution to obtain the optimal estimate of the navigation state.[131 An example would be the Zero Velocity, or Zupt update Kalman filter, where at each time step the velocity of the IMU is updated to be zero. This would be implemented when it is known that the body is standing still, such as in the gyrocompassing application. The Kalman filter works to better estimate the solution by updating the full state X at regular intervals as: CHAPTER 2. CAROUSELING DESCRIPTION X + = X- + K(z - HX) (2.67) where X + is the state matrix after the update and X- is the state matrix before the update. K is known as the Kalman Gain and is equal to K = P-HT(HP-HT + R)-' (2.68) where H is the observation matrix and R is the covariance matrix of the observation, which is a diagonal matrix where the elements are equal to the measurement noise squared. In Eq. 2.67, z is the measurement, equal to z = HX- + v where v is the measurement noise vector. (2.69) The covariance is updated by the Kalman filter using Eq. 2.70: P+ = (I - KH)P- (2.70) This update works only when the optimum Kalman gain is used. If a suboptimal gain is used, the following update equation is implemented: P+ = (I - KH)P-(I - KH)T + KRK' (2.71) This equation requires more computation, but it is accurate for any value for the gain, not just the Kalman gain, which is beneficial for some of the sensitivity analysis that will be performed later. In either case the gains are computed at each filtering time step and used to update both the state and the covariance. 2.8. SIZE EFFECT COMPENSATION 2.8 Size Effect Compensation As previously stated, the navigation scheme described navigates a point at the center of the rotation. However, in practice, the IMUs will be mounted some distance away from the center. By sheer logistics all 6 instruments (3 gyroscopes and 3 accelerometers) cannot be located at the exact center of the rotation, so some apparatus is created that can provide the necessary amount of gimbal = x0 + Rx (2.72) where R, is the rotation from body coordinates to inertial coordinates. If the accelerometers are assumed to be on fixed lever arms, then xB would be a constant value. Taking the derivative of Eq. 2.72 results in: xI = ix + RBC + R [~ x]xk (2.73) but the lever arm is assumed to be constant so 5xB is equal to zero and the second term can be eliminated. Thus taking the second derivative of Eq. 2.72 to obtain acceleration yields: X =i + R L x2 (2.74) + [ x] 2 xf) +x]x (2.75) ]xf + R which, after combining rotations can be written as I = R(fo + [ since the specific force of the point at the center of the rotation, fo, is the acceleration of that point rotated into the body frame. Equation 2.75 can be changed into the body frame, and fk can be defined as the specific force felt by the kth accelerometer CHAPTER 2. CAROUSELING DESCRIPTION (fk =RI ). This results in Eq. 2.76: fk = fo + [ x]xB + [X]2X (2.76) and rearranged to solve for fo, which is the specific force felt at the center of rotation and the quantity desired for navigation: fo = f - ×]x - [ 2 xk (2.77) The angular rate quantities can be determined from the gyroscope outputs and the lever arms are assumed to have been measured, meaning that only the fk quantity is left to be determined. This value can be determined from the scalar outputs of the accelerometers. Each kth accelerometer outputs a scalar acceleration sensed in the body frame, ak, felt along the accelerometer's sensitive axis, Sk, and can be expressed as ak = sf.fk (2.78) Substituting Eq. 2.78 into Eq. 2.77, the equations determining fo for all three accelerometers can be written as: X]X B + [ x ]2X B ) (2.79) sTfo = a, - sT([x]xB + [ x]2xB) (2.80) sTfo = a, -s T([2x]X (2.81) sTfo = a, - sT([ + [Qx]2xB) For convenience, Eqs. 2.79-2.81 can be written as one equation using matrices. The 3 x 3 matrix S is defined as (2.82) s= S3 2.8. SIZE EFFECT COMPENSATION From the three scalar accelerometer outputs the vector a can be formed, a1 a= a2 (2.83) a3 And an additional vector named M is produced containing the remaining terms: T([x ]Xl + [x ]2X ) M = S([2X]x B + [ x]2x B ) .(2.84) sT([2 x]xB + [ x]2XB) In this manner, the Eqs. 2.79-2.81 can be expressed simply as fo = S - a - S -1 M (2.85) where all quantities used to determine fo are either known beforehand or are outputs of the accelerometers or gyroscopes. Notice also that if the IMU is designed so the three sensitive axes are aligned with the lever arms, then the following equation is true by definition: s k (2.86) If Eq. 2.86 holds, then the first term in each entry of the vector M will be equal to zero and therefore fo is independent of 2. This can be advantageous as the angular acceleration would not be required to be computed. If the lever arms were known perfectly, Eq. 2.85 could be applied to the accelerometer outputs to yield fo, which would be identical to the specific force used in modeling the slew with no size effect (Eq. 2.36). However, sometimes the lever arms are not known, or even if they are, there is some limit to how well they can be measured or how precisely they can be constructed. So there will be some error in the lever arm values. As the lever arm is incorporated into the equations as xB and has CHAPTER 2. CAROUSELING DESCRIPTION three dimensions, there are three possible errors for each lever arm. In the same way there are three possible sensitive axis errors for each lever arm. These errors must be incorporated into the navigation as additional error states. To incorporate the size effect into the navigator as described in Section 2.7, first the result for fo found in Eq. 2.85 will be used to transform the outputs from the three separate accelerometers into the theoretical output of one accelerometer that was at the center of rotation. This specific force value can then be used in the estimator. However, the possible errors in the lever arms must be accounted for as well. The equations of motion in Eqs. 2.42-2.44 must be re-linearized, because no the specific force error term is associated with it (due to the size effect). The linearized equations now take the following form: 5x = 6v 6 (2.87) = G(t)6x - 2[ = -x l x]6v + Rfo + R[Lx]fo (2.88) + 6g (2.89) where the only difference from the original linearized equations is the added Rf 0o term in the velocity equation. The position and attitude equations remain completely unchanged from the earlier navigation equations. It is necessary then to determine an expression for 6fo so that it can be incorporated into the dynamics matrix for propagation. The goal is to find a linear equation for 6fo in terms of the different error states, that can be inserted into Eq. 2.88 and therefore the dynamics matrix. To accomplish this, linearize Eqs. 2.79-2.81 to yield the resulting equations: 6s fo + s6fo = 6a - s ([(x] + [Lx] 2)X - S([X + s Tfo + sTfo = a2 - 6S([2x] + [{x] 2)x2 - s s fo = a - sfo X 2 )6xB (2.90) X([!x] + [x] 2 )6x (2.91) s ([{2x] + [ x] 2 )xB - s([X + LX]2)6x (2.92) 2.8. SIZE EFFECTCOMPENSATION where the 6 ak terms are the accelerometer errors, the 6sT terms are the sensitive axes errors, and the 6x3 terms are the lever arm errors. Rearranging these equations to isolate the 6fo terms yields: sl6fo = 6a, - sT6fo = 6a2 - sT6fo = 6a3 - 6sT (([ x] + T(([2x] + [ x] 2 )X + fo) - S ([!x] + [QX]2)6x sT(([x ] + [Q x] 2)XB + f0) - ST([! B + fo) - ST( [ x] 2 )X3 (2.93) x ] + [QX]2)6x B (2.94) X + [x2)6xB (2.95) In order to combine Eqs. 2.93-2.95 to obtain a matrix expression for 5fo, each side of the equations can be multiplied by S - 1 and a complicated looking expression can be obtained in order to facilitate the direst correlation between the lever arm errors and the specific force error for the implementation in the dynamics matrix. The equation for 6fo then takes the form 3 fo = S-1'a - E 3 S- 1 AkSk - E S-Lkx~ (2.96) k=1 k=1 where Ak is a 3x3 matrix of zeros except for the kth row defined for each sensitive axis k as Ak(k,:) = ([!ax] + [Qx] 2 )x B + fo (2.97) and similarly, the 3x3 matrix Lk has all elements equal to zero except for the kth row which is defined by the following equation: Lk(k,:) = s([2 x] + [2 x]2). (2.98) These equations can be used to express the specific force error precisely in terms of the sensitive axis and lever arm error states for incorporation into Eq. 2.88 and the dynamics matrix F of the navigation filter. In a real system, it is highly desired to simplify these equations in order to reduce CHAPTER 2. CAROUSELING DESCRIPTION possible errors. If general procedures were to be followed, some assumptions can be made, such as designing the IMU such that the sensitive axes are aligned with the their respective lever arms so that Eq. 2.86 holds true and the effects of angular acceleration can be greatly minimized, and possibly ignored completely. Also it is standard practice to align the lever arms properly so that the three sensitive axes are aligned with the body coordinate axes, meaning that S is equal to the identity. This greatly simplifies calculations involving S or S - 1 as they will both be equal to the identity. In order to modify the original navigation filter, the extra error states are first added and then Eqs. 2.87-2.89 can be implemented into the estimation scheme in place of Eqs. 2.45-2.47 and integrated to find the navigated solution as earlier described. The propagation and Kalman updates are performed in the same manner as before, except now there is the presence of extra lever arm states which will increase computation time. Also recall that now the accelerometer input to the navigator is the vector a which is really a collection of the three scalar non-collocated accelerometer outputs. Therefore a must be converted into fo using Eq. 2.85 before integration. Accounting for the lever arm errors by incorporating extra error states allows the filter to most accurately model the dynamics of the problem and correct for the size effect. In the next chapter, the size effect encountered when not using these extra error states will be analyzed in detail. Chapter 3 Gyrocompassing Analysis 3.1 Basic description of gyrocompassing Gyrocompassing is the process of finding North by measuring the direction of the Earth's rotation vector and the rotation of the gravity plane. The rotation is sensed by the gyroscopes and accelerometers in the IMU and this knowledge is used for azimuth determination. Since gyrocompassing relies on the Earth's axis of rotation and not it's magnetic field, it points to true north, as opposed to magnetic north like standard compasses. Gyrocompassing has another advantage in that it will not be interfered with by ferrous metals in the vicinity and is not subject to variations in the Earth's magnetic field. The method was first used for marine navigation in the late 19th/early 20th centuries and now its use has been expanded to UAVs, personal navigation and military scenarios. For the purposes of this thesis the focus will be on implementing carouseling into a gyrocompassing application for personal navigation use that needs to be lightweight as it will be carried by a person, so the mass benefits of MEMS instruments are highly desired. The concept is a person can carry the gyrocompass and set it down at intervals to take a bearing on the direction of north, much like the utilization of a magnetic compass. As gyrocompassing is typically a higher accuracy application, MEMS devices have CHAPTER 3. GYROCOMPASSING ANALYSIS generally not been selected for use in these scenarios. However, for certain gyrocompassing applications, there are severe mass and power constraints where it would be highly beneficial to use the lighter, more energy efficient MEMS devices. Therefore, this chapter will investigate the implementation of carouseling with MEMS IMUs for use in gyrocompassing. Kalman filtering techniques and MEMS specifications used will be described and then the results of the simulation will be presented. The carouseling scheme will be compared to the standard gyrocompassing method to see if any benefits are produced. Numerous variables concerning the navigation will be tested for their effect on performance. The influence on azimuth error of the size effect in particular will be investigated, including the ramifications of not properly compensating for the addition of lever arms. 3.2 Kalman Filter Update In most gyrocompassing applications, the body is standing still in relation to the Earth when a reading is taken. This extra amount of knowledge can be incorporated into the Kalman filter to improve navigation. Since it is known for a fact that the IMU is not moving in relation to the Earth, this knowledge can be implemented as truth to the filter. One way to tell the filter the body is not moving is to implement a zero-velocity update, known as a Zupt. This is a Kalman update where the earthrelative velocity is updated as zero with each update. In this case the observation matrix H in Eq. 2.68 is a 3 x n matrix (where n is the size of the covariance matrix) and all the elements are equal to zero except for a 3 x 3 identity matrix in the rows corresponding to the velocity error. This zupt update procedure indicates to the filter that the body has no earthrelative velocity at that point and implies there was no motion over the time step. However, this does not provide the explicit information that the IMU has not moved from its original spot. So another option is to use a delay state update instead, in 3.3. SIMULATION DETAILS which the position is updated at each interval as being the same as the position at the previous time step. This implies, but also does not explicitly state, that the IMU did not move between time steps. In preliminary simulations, the delay state method provided similar results to the zupting method in this case, but zupting is more common in this application and will be used for all the simulations. For this simulation, a measurement noise of 0.001 m/s was selected for the zupt updates, which will occur with a frequency of 1 Hz. This seems to be a conservative estimation for the noise that would be felt in this personal navigation setting. The effect of changing the value of the measurement noise and the frequency of the zupts on the results will be investigated in later sections. 3.3 Simulation Details The Charles Stark Draper Laboratory has designed numerous MEMS instruments ready for use in a variety of applications. The error specifications for the most accurate MEMS accelerometer, were used for these simulations. For most of this analysis, the simulations will be performed using the specifications of the current MEMS gyroscope Draper has designed, but the improved model is in development and Draper has estimated specifications for the improved accuracy it will be able to provide. Therefore the current model IMU will be the main focus of this research, but the estimated specifications of the improved IMU will also be studied as it will provide further insight. Both IMUs implement the use of the same accelerometers, it is only in the gyroscope performance that they differ. The complete error specifications of the two IMUs can be seen in Table A.1 in the Appendix. For this simulation, the baseball stitch slew will have a period of 60 seconds and follow the angular rate profile shown in Fig. 2-9, which equates to a maximum 48 0 /s angular rate for the outer gimbal (the outer gimbal performs 8 total revolutions of CHAPTER 3. GYROCOMPASSING ANALYSIS 3600 for a total of 28800 in 60 seconds). Once again, the slew is inertially referenced, as was shown in Section 2.2, although the effects of referencing with respect to the Earth will be investigated as well. The size of the apparatus performing the slewing is assumed to have a radius of approximately 3 cm, so this will be the assumed values for the lever arms, with an error covariance of 1 mm 2 (therefore the standard deviation of the error in all the lever arm measurements will be equal to 1 mm). The variations of the lever arm parameters will have a great influence on the size effect experienced and in turn the accuracy of the IMU, so they shall be investigated later in this chapter. 3.4 Azimuth Covariance Determination Results The goal of this section is to show the effect of the carouseling scheme on the gyrocompassing accuracy. Simulations were run to compare the carouseling method to the standard gyrocompassing method so the differences can be observed. The set-up of the simulation is that of a typical gyrocompassing problem in which the body is at rest on the earth's surface in order to take a bearing. Recall, that although the body is at rest, the IMU will be going through the inertially-referenced carouseling scheme in the body. The IMU navigates and essentially measures the Earth's rotation vector in order to determine North, or the azimuth angle. The navigation scheme will solve for the entire state, but the main focus of gyrocompassing is to determine the azimuth angle. In any navigation problem there are initial error covariances in the states, which have been set to be typical values shown in Table 3.1: Error State Position Velocity Attitude Initial Covariance 10 mn 0.1 m/s 1o Table 3.1: Initial Error States Implementing these initial covariances along with the specifications from the IMU, 3.4. AZIMUTH COVARIANCE DETERMINATION RESULTS the gyrocompassing problem can be navigated and the results will be analyzed. The state estimate is not of much interest as it depends on the initial value and varies according to random variables correlated by the covariance. Therefore, the focus will be on the covariance analysis, as the covariance determines the "envelope" for the error estimate. The analysis begins by simulating the slewing trajectory using the equations described in Section 2.6 to obtain an IMU output of 6v and 60 along with the truth values of position, velocity, and attitude. Then, using the navigator developed in Sections 2.7 and 2.8, the IMU outputs can be integrated to obtain estimates for the errors along with the covariances for all states. The estimates can be compared against the true values for all states, however it is the azimuth error that is of most importance in the gyrocompassing application. In Fig. 3-1 can be seen the azimuth error when using the baseball stitch slew and the error specifications of the current Draper MEMS IMU. In this plot, the covariance is depicted as the dotted red lines and the error estimate is the solid blue line. The navigation azimuth error estimate initially fluctuates but approximately approaches zero after starting a from a randomly distributed point with the covariance range. The covariance decreases to slightly less than 0.3 degrees within 10 minutes from an initial error of one degree. This drastic covariance decrease confirms the results of Section 2.4, and reveals that the baseball stitch slew can provide accurate results. The estimate is shown in this plot to demonstrate its behavior within the covariance bounds, but as there are random value functions involved in the estimate calculation, the focus will be on the covariance analysis. As was seen in Fig. 3-1, the carouseling method does seem to perform well, but in order to accurately gauge the performance of the baseball stitch slew, it is necessary to have a standard for comparison. One standard would be to compare the results to those of the scenario in which the body were stationary and made absolutely no movement over the time interval. However this is unrealistic, as gyrocompassing systems CHAPTER 3. GYROCOMPASSING ANALYSIS Azimuth Error and Covariance ii 1 0.8 .....'C -! i.. . I........ ................. ....... ........ l . . . 0.6 0.4 . . .. .... ... .... . .. . .. .. .. .. ... ... .. • .... . . . . ... . . .. . . . . . . .. . . . . . . . . .... ... . .. . .. .. . . ........... 0.2 0 -0.2 ' .. ............ ......... ......... ..... ... . ..... ....... .. - . . . . . . . =. -0.4 (~ - . . ~~~~~~ -0.6 .. • i ~.-. ~ L I .. -0.8 -1 0 1 2 3 4 6 5 Time (min) 7 8 9 10 Figure 3-1: Current MEMS IMU Azimuth Covariance for Baseball Stitch Slew The estimation error remains inside the covariance bounds which indicate an accuracy of 0.2910 after 10 minutes. implementing IMUs of lower quality take the extra measure of reversing the orientation of the body periodically in what is called 2-Position Gyrocompassing. Some IMUs are accurate enough that this reorientaiton is unnecessary, but for the MEMS devices selected it would certainly be employed. Therefore, this method shall be taken as the standard to which the baseball stitch slew share be compared. 2-Position Gyrocompassing is essentially a less-intensive way to eliminate biases by simply reversing the direction of the object and preventing the build up in one direction of biases. The action can be physically picking up the gyrocompass and turning it around or could be performed by some servo-mechanism. This rotation would be earth referenced in the simulation, not inertially since it is just the physical reversing of direction on the Earth's surface. The time between reversals can vary but for this simulation, the 3.4. AZIMUTH COVARIANCE DETERMINATION RESULTS direction is reversed once every minute. In the following plot shows the covariance of the baseball stitch compared with the 2-position method over a period of 10 minutes: Azimuth Error Covariance I I I I I 0.8 .... ........... . ..... ''I-,,-............ : ......... -"' '' ... 0.6 . . .. ....... . !......... !.........i........ .. . ~........ !........ -... -.. ..... .. .. . . . ... .. . ... .. "-_"-- .. . .. . .. . . . .. . . . :" -. . ... .•. . .. 0.4 ... . ;.........;............ • 0.2 : ........ ..- -.. ... • ......... - - ..,-. -.. ....... : • ' -,- . .......... ._... _ ... _ ... . ... . . . . . 0 -0.2 -0.4 -- --........ 4 ...... ;........ . . .. . ..;................. • - .................... i . .... .... . .... .. . .. .... ... ... ..... .... ... ....... ........... .. ..... .. ;.......................... . .. :. . ..- ... ... .. ... .. .. .. ................ ... .. . .. .. i.... ... . ......... -0.6 -0.8 ... ..... -1 .... ...... i......... ........ ......... ::..... I I -- 1 2 Baseball Stitch Slew -2 p s to I , , , 4 5 E Time (minutes) 7 8 9 2-o it o =Ib • 0 I - 3 10 Figure 3-2: Current MEMS IMU Azimuth Covariance Comparison for Baseball Stitch Slew and 2-Position Gyrocompassing The carouseling method provides a substantial decrease in error of 50% over the conventional 2-postion method. This shows that the baseball stitch slew does indeed provide a significant benefit. After 10 minutes of operation the carouseling scheme has improved accuracy by over 50%, from 0.649 degrees to 0.291 degrees. Using carouseling, the current MEMS IMU can achieve accuracy of less than half a degree in 4 minutes, where the 2-position gyrocompassing method will never reach this level of accuracy. This certainly shows 64 CHAPTER 3. GYROCOMPASSING ANALYSIS there is a substantial increase in accuracy when implementing carouseling. There are some variable factors that affect these results which will need to be analyzed. To further investigate these results, another comparison would be to determine if the benefit gained from carouseling is due to the unique properties of the baseball stitch slew as discussed in Section 2.3, or if it is due simply to the fact that rotating the IMU in any slew will provide similar results. To test this, the baseball stitch slew covariance can be compared to that when the body is simulated to simply rotate 3600 back and forth every minute. These results are shown in Fig. 3-3. Azimuth Error Covariance 0.8 ...... - .- .... .......... ... ... ... ..... ... ..... ..... .. . 0.4 o) =0.2 E -0.8 ...... .......... ........................ - -1 0 1 2 3 ........ ...... • . •Baseball Stitch Slew - - 360* turns 5 6 4 Time (minutes) 7 8 9 10 Figure 3-3: Current MEMS IMU Azimuth Covariance Comparison for Baseball Stitch Slew and ±3600 turns The carouseling technique provides the same benefit over the ±3600 turns case as over the 2-position method. Once again, the baseball stitch slew provides significantly more accuracy. In fact, it can be seen that the ±360' turns provide no discernible benefit to the 2-position 3.4. AZIMUTH COVARIANCE DETERMINATION RESULTS method (the final covariance for the ±3600 method is 0.648 degrees as compared to 0.649 degrees for the 2-position method). So it can be concluded that the special properties of the baseball stitch slew as discussed in Chapter 2 are responsible for this increased accuracy, not simply the rotating motion itself. One facet of the baseball stitch slew that has been mentioned before is that the slewing must be inertially referenced for the full benefits to be felt. However, in most cases it is much easier to reference gimballing in the Earth coordinate instead. Since both coordinate frames are centered at the center of the Earth, the only difference is the small rotational motion of the Earth itself. So it would be valuable to know how the IMU will perform if the slew actually is Earth referenced and not inertially referenced. Figure 3-4 shows the azimuth error covariance results comparing the inertially referenced slew to the earth referenced slew. This plot reveals that although the inertially referenced slew does perform better, as expected, the Earth referenced slew does not increase the error covariance drastically. There is less than a .030 increase with the Earth referenced slew. This would indicate it may be possible to Earth reference the slew to reduce complexity while still attaining an almost 50% improvement in covariance performance over the 2-position method, depending on the required accuracy of the application. This is a 10% increase in error over the inertially slewed case, so if the extra precision is required for performance the slew will need to be intertially referenced to take full advantage of the properties of the baseball stitch slew. Another important aspect that affects both the 2-position and carouseling methods is the zupt update. It was stated earlier that the zupting was performed once a second for these simulations. Although this frequency is not uncommon, it may be desired or necessary to update less frequently, which could potentially have an effect on the results. In Fig. 3-5, the azimuth covariance for both the 2-position and carouseling schemes can be seen when the zupts are performed every 5 seconds instead of every one second. CHAPTER 3. GYROCOMPASSING ANALYSIS 66 Azimuth Error Covariance 0.2 0.4 -- ED -0.2 .Referenced ..... ............. i... ........ ... ... ..... -1..... -- 0 1 2 3 6 5 4 Time (minutes) - Earth Referenced 7 8 9 10 Figure 3-4: Current MEMS IMU Azimuth Covariance for Baseball Stitch Slew referenced in two different coordinate framesThe inertially referenced slew performs better than the Earth referenced slew, but the difference is only marginal. There is a slight increase in the covariance for both methods, but the increase is consistent for both so the overall benefit of slewing is undiminished. The baseball stitch slew still provides a 50% lower azimuth error covariance. The zupting interval is therefore not a large factor in the comparison between the two methods. The marked increase in azimuth performance when implementing carouseling for use with the current MEMS IMU is now apparent, but if in the future the current model can be improved upon, it will be helpful to see what kind of benefit the baseball stitch slew will provide. The predicted specification of the improved IMU can be integrated into the simulation to see how this change in IMU accuracy affects the results. Using the same initial conditions, the carouseling and 2-position methods 3.4. AZIMUTH COVARIANCE DETERMINATION RESULTS 67 Azimuth Error Covariance 0 0 .8 0.2..... 0. .... -0.. ........................................................ . ......... Time (minutes) Figure 3-5: Current MEMS IMU Azimuth Covariance Comparison with zupt update performed every 5 seconds Even with zupting performed over longer intervals, the carouseling method provides the approximate same amount of benefit. can be simulated and their respective azimuth error covariance shown here in Fig. 3-6. Although the slewing continues to provide increased performance over the 2position method, the final reduction in error is significantly lower than for the current MEMS IMU. The carouseling method provides an azimuth accuracy of 0.1160 while the 2-position method provides 0.1450. Both methods in fact perform better than when the carouseling is used for the current MEMS IMU, and now the benefit of slewing is reduced to about 0.040, which is still 20% improvement, but not a substantial magnitude. So if the specifications for the proposed improved design actually can be met, if concerned with longer time applications it makes little sense to add can be met, if concerned with longer time applications it makes little sense to add 68 CHAPTER 3. GYROCOMPASSING ANALYSIS Azimuth Error Covariance I I I 1* 00.2.2 . . ............................. ........ ........ ....... ........ ............. ........... ... ......... ........ ....... o .......... .. ............................ ........ ................ • -0.2 .... ........... ........ 0 ......... -1 0 ............... ....... ....... .......... ........ ................. ........ ............................................... 1 2 3 Baseball Stitch Slew -- 4 5 6 Time (minutes) -2-position 7 8 9 10 Figure 3-6: Improved MEMS IMU Azimuth Covariance Comparison for Baseball Stitch Slew and 2-Position Gyrocompassing With the improved gyroscopes, the carouseling method still shows slightly better performance but it is not as significant as with the current model. complexity by using the carouseling scheme when the accuracy is not improved drastically. However, notice that in the first 4 minutes of the simulation there is still a sizable benefit of using the carouseling method. For example, the carouseling method can reach an accuracy of one half of a degree in about 45 seconds while it takes 3 minutes for the 2-position method to reach that mark. At right before 2 minutes, the accuracy of the carouseling method is about 0.25 degrees as compared to almost 0.8 degrees for the 2-position method. Therefore, if the given application is extremely time-sensitive, there is still a large advantage to using the carouseling technique with the improved MEMS IMU. This would be a likely scenario since it would be undesirable for a person using the gyrocompass to be required to wait over 5 minutes to .0.4._ 3.5. SENSITIVITY ANALYSIS find which direction is North. Carouseling with the improved MEMS model would provide that extra accuracy in a far more timely manner than the 2-position method. These covariance plots have shown the overall benefit in using carouseling for this particular application, but the reasons and sources of this improved accuracy can be investigated further. By performing different types of sensitivity analysis, the contributions to the error can be broken down and compared to see in what way is the carouseling providing benefit, and how changes to the problem will affect it. This is the focus of the remaining of this section. 3.5 Sensitivity Analysis The overall effect on azimuth covariance due to carouseling has been seen, but the components of this error can be even further investigated. This section is devoted to parsing out the total error covariance in order to see the exact benefits that carouseling provides. The goal is to discover in what ways the baseball stitch slew provides improvement and how sensitive the results are to the changing of certain aspects of the problem. It would be desirable to see if the carouseling maneuver is eliminating the errors predicted in the analysis of Chapter 2. Also, insight can gained form investigating if there are any other errors unexpectedly building up in the process. It would be beneficial to study the components of the total covariance error to see what contributions are the most important. A special type of sensitivity analysis known as an error budget can be implemented to reveal this information and show the breakdown of the entire covariance. An error budget determines the separate effects of the error sources on the overall covariance errors 171. The process of creating an error budget will be described in the next subsection. CHAPTER 3. GYROCOMPASSING ANALYSIS 3.5.1 Error Budget Formulation The main principle of the error budget is that when a filter is applied to the navigation solution, all of the error sources (even if they are not modeled in the filter) contribute to the total error in the filter estimate. The error budget delineates these separate contributions to the total covariance error [7]. To create this budget, first, the time history of the filter gain matrix K must be acquired. This can be done by simulating the system in the exact same way as the earlier covariance analysis with initial covariance matrix P*(O), and recording the filter gain matrix at each time step. This time history of the gains can be saved as K*. To determine the contributions of each error source, the simulation is run again with zero measurement noise and initial covariance P(O) where all error elements not corresponding to the selected error source are set to zero. For example, if the contribution due to accelerometer bias is to be measured, all entries into P(O) will be equal to zero except for the turn-on and in-run accelerometer bias covariance terms. This can be parsed out further if for example the contribution due solely to the turn-on bias is desired to be tested. During these simulations, instead of the filter generating its own gains, the gains from the time history K* are employed for the duration of the simulation. The final covariance values are recorded for each simulation, which captures the effect due solely to that specific error source. The final covariance for each simulation is the total contribution to the error solely due to that specific error source. The contribution for each error source can then be recorded, including measurement error. To find the effect of measurement error (in this case the zupt noise), all initial elements of the P matrix are set to zero and so the measurement noise is the only error contribution. After generating the final covariance for each separate error source, it can be seen that squaring all the values and adding them together is equal to the covariance error for the full state errors squared. In this way, the entirety of the error can be accounted for, and the contribution from each error source can be compared. The larger the 3.5. SENSITIVITY ANALYSIS contribution from one error source, the more influence this error source has on the total error covariance and the more sensitive the covariance is to that error source. By this logic, error sources with the largest contributions to the error budget are the most desired to be eliminated or reduced as this will provide the most benefit to the total results. In the next sections, the results from creating error budgets and other sensitivity analyses will be used to determine the effects of changes in critical error sources. 3.5.2 Sensitivity Analysis Results Using the method described in Subsection 3.5.1, error budgets of azimuth error for both the carouseling and 2-position method using the current MEMS IMU can be generated. Table 3.2 compares the azimuth error results of both for 10 minute simulations with initial conditions as specified in Table 3.1. Now it is possible to dissect the covariance errors for both methods and ascertain what gains the slewing procedure is providing. First, notice that the largest reduction in azimuth error comes in the gyroscope bias contribution. The baseball stitch slew practically eliminates this entire error, leading to an RSS contribution of only 1% of the gyroscope error in the 2-position method. This is an extremely attractive feature of the carouseling scheme, and confirms the analytical results obtained in Chapter 2. Using the 2-position method, the gyroscope bias is the largest error contributor, and by using slewing this substantial error source is almost completely brought to zero. In the 2-position method, the gyroscope bias contributes to 26% of the total covariance. This is brought down to less than 1% by using carouseling. This also explains why the slewing provides only marginal benefit over 2-position gyrocompassing when the specificaitons for the improved MEMS model are used. The gyroscope bias of the improved MEMS IMU is less than 10% of the bias for the current MEMS model. Because the bias is already so low, there is less error for the slewing to eliminate. It is also of interest that for the lever arm errors present only in the carouseling CHAPTER 3. GYROCOMPASSING ANALYSIS Error Source Position Velocity Baseball Stitch 2-position 0.000 0.001 0.000 0.001 Angle 0.113 0.240 Gyro. Bias Gyro Scale Factor Gyro. Misalignment Gyro. Non-orthogonality Angle Random Walk Accel. Bias Accel. Scale Factor Accel. Misalignment Accel. Non-orthogonality Velocity Random Walk Lever Arm 1 Lever Arm 2 0.033 0.013 0.002 0.001 0.219 0.107 0.011 0.001 0.001 0.016 0.001 0.001 0.327 0.013 0.002 0.001 0.305 0.076 0.000 0.001 0.001 0.014 Lever Arm 3 0.001 Measurement Noise Total Error 0.158 0.291 [ 0.056 0.647 Table 3.2: Error Budget Comparison of Azimuth Error contributions in degrees for Baseball Stitch Slew and 2-position Gyrocompassing case, the total error contribution due to these errors is practically insignificant. The sensitivity to these errors will be investigated in detail later in Section 3.6, but the fact that these errors do not substantially decrease performance for the assumed scenario is a positive result. This indicates that for the lever arm specifications employed (lover arm length of 3 cm and uncertainty of 1 mm), there is not a large penalty in terms of overall accuracy. Many of the other error sources contribute roughly equally in both the carouseling and 2-position case, for example, the accelerometer scale factor, misalignment, and non-orthogonality errors contribute almost no error to the azimuth covariance. This makes sense as the accelerometers measure change in velocity and do not factor into the attitude error directly. However there are a few anomalies noticeable in Table 3.2. The carouseling scheme also provides significant improvement over the 3.5. SENSITIVITY ANALYSIS 2-position method in the contribution due to initial angle error and the angle random walk. These gains are not as substantial as the gyroscope bias reduction, but they still do provide a non-trivial amount of the total covariance improvement. Conversely, the 2-position method, besides from not having any error due to the lever arms, has a substantially lower amount of error due to measurement noise. This is interesting, as the zupting measurement noise was equal to 0.001 m/s for both. The error budget indicates that the 2-position method is in fact less sensitive to the zupt noise than the baseball stitch slew. To further investigate this, another type of sensitivity analysis can be done in order to determine the effect of measurement noise on the azimuth performance. Since the error budgets indicated that the measurement noise provides a substantial unforeseen disparity between the two methods, a sensitivity analysis of the measurement noise shall be performed. Now instead of generating an entire error budget, only the measurement noise used for zupting will be changed and the total azimuth covariance error will be recorded for both the 2-position and carouseling schemes. The results can be seen in Fig. 3-7: Notice the points which correspond to a measurement noise of 0.001 m/s reflect the results found in the earlier simulations. However, the 2-position method is relatively unaffected by any change in measurement noise, as the value of azimuth error scarcely shows any change even with orders of magnitude changes in the measurement noise (note the log scale of the measurement noise axis). This is expected from the error budget results, as the contribution from the measurement noise was low for the 2position method. It was seen that the measurement noise contributed to less than 1% of the RSS azimuth error, so any increase or decrease in the zupt noise will not have much effect on the total error. However, when the baseball stitch slew is applied, the azimuth covariance is much more sensitive to the measurement error. A full 26% of the total error in azimuth is due to the measurement error in the error budget, and as seen in Fig. 3-7, this contribution only increases with increasing noise. This is CHAPTER 3. GYROCOMPASSING ANALYSIS 0.7 0.6 1 0.5 0 " 0.4 .E 0.2 2-position I 0.1 --- 0 0.0001 0.001 Baseball Stitch Slew 0.01 Measurement Noise (mis) Figure 3-7: Azimuth Covariance Error after 10 minutes for Varying Measurement Noise As the measurement noise increases, the contribution to azimuth error increases for the carouseling case while remainly practically constant for the 2-position method. valuable information, as it indicates that if the measurement noise can be reduced in some way, the carouseling will provide even greater accuracy. For example, if the measurement noise is reduced by a factor of 10, the overall accuracy improves by 40%. This would not be unreasonable as the measurement noise for zupt updates can vary significantly with a given application. Conversely, if the environment of the application is considerably noisy, much of the benefit of the carouseling sheme could be negated. For further analysis, the error budgets for both the 2-position and baseball stitch trajectories were compiled for simulations lasting 5 minutes and 15 minutes in addition to the error budget for 10 minutes in Table 3.2. In this way, any patterns that develop over time can be investigated. Figure 3-8 shows the covariance contributions of four 3.5. SENSITIVITY ANALYSIS 75 selected error sources: gyroscope bias, zupt noise, accelerometer bias, and initial attitude error. These four were selected as the other error sources either have very minuscule contributions or did not exhibit particular interesting patterns over time. So the columns in Fig. 3-8 are only partial error budgets and their total height is not equal to the total error covariance in this case. 0.35 0.3 0.25 0.2 0 Gyro Bias MZupt Noise . 015 Accel Bias . Initial Attitude 0.1 0.05 2-position Carouseling 5 minutes 2-position Carouseling 10 minutes 2-position Carouseling 15 minutes Figure 3-8: Selected Azimuth Covariance Error Contributions over time As time in- creases, the gyroscope bias contribution increases for the 2-position case while remaining negligible when carouseling. Additionally, the carouseling maneuver eliminates the initial attitude error significantly more quickly. Notice first that the contribution from gyroscope bias for the carouseling maneuver for all three times is so small it is practically not visible. However, the same contribution for the 2-position case is not only substantial, but actually grows larger over time. In fact, the gyroscope bias error completely dominates all other errors by the time the simulation has been running for 15 minutes. So over time the benefit of CHAPTER 3. GYROCOMPASSING ANALYSIS carouseling will only become increasingly more substantial. Not only does carouseling eliminate the gyroscope bias error almost completely, it keeps it from growing and the covariance getting worse As previously mentioned, the measurement noise contribution is seen to be larger for the carouseling case than the 2-position case. This contribution does decrease as time increases for both methods, but still remains a large contributor to the error for carouseling. In fact, as time increases and the carouseling method continues to vastly reduce all other errors, the portion of the total error due to the zupt noise increases. At the same time, the measurement noise error for the 2-position case is hardly a factor at 15 minutes. This again points to the possible improvements that could be made for th carouseling case if the measurement noise could be reduced. Also seen in Fig. 3-8, although the contribution from accelerometer bias is larger for the 2-position case at 5 minutes, for the simulations of 10 and 15 minutes the carouseling actually has a slightly larger accelerometer bias contribution. This is most likely due to the fact that because the carouseling maneuver involves constantly slewing the IMU, whereas for the 2-position case the IMU was mainly stationary. This added velocity creates a larger opportunity for velocity error in the carouseling method. As for the 2-position case, velocity errors will not be as large as there is less total velocity. The last contribution seen in Fig. 3-8 is the that due to the initial attitude error of one degree. It can be seen that the proportional amount of this error decreases drastically for both methods over time, however the carouseling maneuver does a substantially better job of eliminating this error quickly. For example, at 5 minutes, the carouseling scheme has reduced this error to less than one eight of the same contribution for the 2-position method. The relative benefit in this error contribution provided by the carouseling decreases over time, as can be seen for the 10 and 15 minute simulations, as both methods effectively eliminate this error source. But the quicker response of the carouseling method gives it a distinct advantage for more 3.6. SIZE EFFECTSENSITIVITY time-critical operations, especially if the initial error is ever larger than one degree, in which case the benefit will be even more pronounced. The analysis so far has been focused on general comparisons of the two methods in terms of error contributions. This has provided a picture of the benefits of the carouseling scheme when implemented for the selected application. However, of possibly greater concern is the impact of the size effect on any gyrocompassing application possibly employed. For this simulation, this size effect was seen to be negligible, but it is desired to see how this can change in different situations. Investigating the sensitivity to size effect, an area of great concern regarding carouseling, is the aim of the next section. 3.6 Size Effect Sensitivity One of the biggest issues with implementing carouseling is the compensation of the size effect. As seen from the error budget for this gyrocompassing application, the increase in azimuth covariance due to the lever arm uncertainties is minimal (less than 1% of the total error). This was based on assumed parameters for the lever arms and their respective uncertainties as well as the slew rate, which may be true for this scenario but could vary greatly from application to application. Recall from Eq. 2.84 that the lever arm length and angular rate 0 are the main contributors to the difference in the outputs of the uncollocated accelerometers from the specific force at the center of rotation, fo. Since the lever arms are actual error states in the filter, the navigator also attempts to estimate the lengths of these lever arms. As previously stated, in the gyrocompassing simulation the lever arms are assumed to be 3 cm long with an initial uncertainty of 1 mm. Also, the period of the baseball stitch slew was 60 seconds, which equates to a maximum gimbal speed of 480/s. To examine how well the filter is accounting for the lever arm, the time history of the estimate for one lever arm will be examined. In this situation, the lever arm was given a real error CHAPTER 3. GYROCOMPASSING ANALYSIS of 0.2 mm, and it is desired to see how well the filter does at estimating this error. Figure 3-9 shows the time history of the estimate for this lever arm error. The filter does an excellent job of estimating the error as the estimate approaches Arm Error Estimate -Lever - - Actual Lever Arm Error 1 - 0.8 Wm 0.6 S0.4 0.2 I ------------------------------- SI 0 0 1 2 3 4 6 5 Time (min) 7 I 8 9 Figure 3-9: Lever Arm Error The estimate can be seen to approach the true error of 0.2 mm the true value and is within 1% within 8 minutes. Note that although this is the estimate for only one of the three lever arms, the other two show identical the identical behavior exhibited in Fig. 3-9. This shows the filter can accurately estimate lever arm errors, which means this filter could also be implemented to determine the true length of lever arms, possibly during calibration. In this scenario, since even the true lever arm error is already on the scale of less than a milimeter, the total error contribution to the covariance due to the lever arm uncertainty is not substantial. However, in other applications, there could be longer lever arms and greater uncer- 3.6. SIZE EFFECT SENSITIVITY tainty in the lever arm values as well as different slew rates. Since the implementation of these extra lever arm states significantly increases the computation necessary and the complexity of the filter, there is interest in studying if these error states are necessary or if they can be eliminated. If the additional errors for no compensating for the additonal states are not significant, it could be desirable to not implement the additional states to save on computaiton. For example, the output of the three accelerometers with the 3cm lever arms can be simulated as in previous analysis. But instead of compensating for the lever arms, the navigation will be performed on the accelerometer outputs with a navigator that does not account for the lever arms and treats this output as the specific force felt by the body at the center of rotation. The azimuth error estimate when this procedure is followed is compared the the normal covariance in Fig. 3-10. It can be seen that the estimate does fluctuate more so than in the compensated case (as seen previously in Fig. 3-1), but remains within the normal covariance bounds. This would lead one to believe that it would be possible to neglect the size effect in this case so the lever arm states would not need to be included in the filter. As the lever arms are only 3 cm in length, the extra velocity added is not that significant. The impact of the size effect is then not so great that it would require the extra computation and complexity of incorporating the extra lever arm states. If the lever arm length is longer than the 3 cm assumed for this application, the size effect felt due to the added change in velocity will increase and it may no longer be possible to ignore. For example, if the same simulation is run with 10 cm lever arms where there is no compensation, the estimate will deviate significantly, as shown in Fig. 3-11. Clearly this is well out of the covariance bounds and attempting to implement this navigation scheme would yield disastrous results in a real application. In fact, the estimate in this case would be significantly worse than when using the 2-position gyrocompassing scheme. In order to better quantify the value of the added lever arm CHAPTER 3. GYROCOMPASSING ANALYSIS Azimuth Error and Covariance 0.5 0 -0.5 -1.5 0 1 2 3 4 6 5 Time (min) 7 8 9 10 Figure 3-10: Navigated solution compared to Azimuth Covariance when 3 cm lever arms are uncompensated The error estimate deviates but remains within covariance bounds with a maximum outer gimbal rate of 480/s states, a sensitivity analysis can be performed that accounts for filters with a reduced number of states. [71 In this case the navigation filter using the lever arm states is the "optimal" or "truth" filter as it accounts for all states of the true dynamics model, and the filter not implementing the lever arms states is the "sub-optimal" filter. So the optimal filter has more total error states and thus requires more computation than the sub-optimal filter, while more accurately accounting for the dynamics of the system. Using the sub-optimal filter to navigate will produce a time history of the sub-optimal gains used for navigation. Much like with the error budget formulation, these sub-optimal gains can be stored and then implemented into the optimal filter instead of the optimal 3.6. SIZE EFFECTSENSITIVITY Azimuth Error I I I ~ I I I I I i ....................................... • I •; ... .. .. ..: • --.. ... .. .. . . ... .. . . ... . ... ..... .. ... . ... .. ... ... .. .. _... . ,. I. ... . . .. . ... . ... . . .. . . -2 . . . . . . . . . : . . . .. . .. ... .. . . .... . . . . .. . ...... • . . ... . . . ... .. .. ,.. . . .. .. . . ... ... .. ... ... . ... .. ... . . .. ... .. . . .. . . ... ..- -. ... - . ... . . .. . .... . .. . . .. .. . . .. . .. .. . .. . . . . . . ....... .. .. .. .. .. . . . . . . . . -3 -4t o 1 2 3 4 5 6 Time (min) 7 8 9 Figure 3-11: Navigated solution compared to Azimuth Covariance when 10 cm lever arms are uncompensated In this case with a maximum outer gimbal rate of 480/s, the estimate deviates significantly from the covariance bounds gains, to obtain an azimuth covariance. This sub-optimal covariance result can be compared to the result when the optimal filter performs updates implementing the normal optimal gains. In this manner the penalty for not compensating for the size effect is the difference between the navigation using the sub-optimal gains and the optimal gains. This penalty can be observed for varying lever arm values with the same 60 second baseball stitch slew with the maximum outer gimbal rate of 48 0 /s. These results are compiled in Fig. 3-12. The azimuth error can be seen to increase in an approximately linear relationship with increases in lever arm length when angular rate is held constant. For lever arm values less than 1 cm there is practically no appreciable difference in azimuth covari- 82 CHAPTER 3. GYROCOMPASSING ANALYSIS 140 I I --------- 0---- 0 +I I, I 0 0.05 0.1 0.15 02 0.25 Lever Arm Length (m) Figure 3-12: Increase in Azimuth Covariance when implementing Sub-optimal gains for varying lever arm lengths In this simulation the sub-optimal gains do not model the lever arm states while the truth filter does include these additional states. The size effect contribution appears to increase linearly with the increase in lever arm length while the maximum gimbal rate is held constant at 480/s. ance. And even up to a 4 cm lever arm equates to a 10% increase in error. Therefore for smaller devices where the accelerometers will not be offset by more than 4 cm, it may be advisable to use a filter that does not account for the lever arm states. However, for lever arms longer than this, it becomes apparent that by not accounting for the size effect, the overall accuracy of the IMU will suffer noticeably. From the equations developed in Subsection 2.8, it can be seen that the size effect is not only dependent on the length of the lever arm, but also the angular rate of the rotating point. In this simulation, the baseball stitch slew was always assumed to have a period of 1 minute, which meant the outer gimbal had a maximum angular 3.6. SIZE EFFECTSENSITIVITY rate of 480/s. If the angular velocity were to be increased, the amount of change in the velocity felt by the accelerometers due to the size effect would be increased. In fact, as was seen in the equations developed in Subsection 2.8, the additional velocity is directly proportional to the total angular rate felt by the accelerometers squared, q2. To see the effect this has on the navigation scheme implementing updates, a sensitivity analysis can be performed similar to that seen in Fig. 3-12, where the sub-optimal performance without accounting for the dynamics of the lever arms is compared to that of the optimal filter that includes the additional lever arm states. In this case the gimbal rate of the outer gimbal shall be varied instead of the length of the lever arms (the lever arms are held constant at 3 cm as in the previous simulations). The results can be seen in Fig. 3-13. This figure demonstrates that the increase in error does roughly increase proportionally to the gimbal rate squared. For the 3 cm lever arm, with angular rates less than 800/s, the increase in error is less than 10% and it would be probably be acceptable to not implement the lever arm error states. With increasing gimbal rate, the increase in errors due to the sub-optimal gains increases drastically and it becomes infeasible to implement a filter that does not accurately account for the size effect. As was shown in Figs. 3-12 and 3-13, the size effect can have a substantial impact on the accuracy of the IMU and in many cases will need to be accounted for with extra error states. But it is also very possible to eliminate the need for these extra states by designing the IMU to be compact enough so the lever arms do not exceed about 5 cm, and keeping the slew rate low. The rate of the slew in no way effects the accuracy of the accelerometers, so keeping this low will add a great benefit. This is because as was shown in the development of the sensitivity matrices in Section 2.5, the sensitivity integrals are all of the special orthogonal group in three dimensions SO(3), and therefore the speed of traversing the rotation does not effect the value of the integrals. As long as the slew itself is fixed, the overall accuracy will remain constant as well. So any scheme for carouseling using a low slew rate with small lever CHAPTER 3. GYROCOMPASSING ANALYSIS 160% -----------.-------------------- ---- ------------ ---- 140% ---------- ---- --- - - ------ ------------I------ ------- -- --- -- - - - -- -- --- ---- 120% I I 100% ---------------------- 80% -------------. . .. . n. . . . . . . . . . . . . . . . . . . .- . -. . . . . ------ ----------- ---------. ..- . . ,--- . . . . ---- - - - - - - - - - -- - - - - - - - - - - -. --------- % 60% ---- ---- 40% 20% 4--------- ------------ -- -,- --- - - . . 0 20 40 60 , . -.. -.. . . . - -. . . 80 . . - . . . 100 --- . . - . - . . - . - . 120 . , 140 . . ..-- - . . ..-- 160 Outer Gimbal Angular Rate (degls) Figure 3-13: Increase in Azimuth Covariance when implementing Sub-optimal gains for varying outer gimbal angular rates The sub-optimal gains do not include the lever arm states and do not as accurately model the system. The size effect increases roughly proportionally to the gimbal rate squared arms should safely be able to navigate without implementation of extra error states. It should also be noted that although adding error states does increase computation, depending on the application, there may be enough computing power where this is not an issue and the savings in accuracy is of more critical importance. As with all navigation problems, the trade offs between aspects such as mass and accuracy requirements are specific to different situations and can vary greatly. Chapter 4 Conclusions 4.1 Carouseling Effect on Azimuth Accuracy As was seen in Fig. 3-2, implementing carouseling provides significantly better performance in azimuth error than the standard 2-position method for the gyrocompassing problem when implementing the current Draper MEMS IMU. There is a 50% drop in the covariance, a large portion of that reduction due to the almost total elimination of gyroscope bias. With such marked improvements, this makes implementation of carouseling with use of the current MEMS IMU a highly recommended option for gyrcompassing. However, there is an increased sensitivity to measurement noise when implementing carouseling that could be significant if implemented in a noisy environment. By the same token, if the measurement noise can be reduced, the benefits of carouseling become even more pronounced. Because of the decreased sensitivity to gyroscope bias, if the gyroscope bias of the IMU being used is lowered, as with the improved MEMS model, the comparative benefit of using the carouseling scheme over time is reduced. However, for applications where performance within the first 5 minutes is critical, the carouseling method provides considerable benefits when used in conjunction with the improved MEMS IMU. CHAPTER 4. CONCLUSIONS For the gyrocompassing application using current MEMS technology, implementation of the carouseling scheme would be highly beneficial to overall azimuth accuracy. Using carouseling allows MEMS devices with significantly lower mass, cost and power requirements, to be implemented in high-accuracy applications. If the measurenent noise is not large and especially if the MEMS instrument selected has a large gyroscope bias or the initial attitude uncertainty is large, the carouseling will provide even more benefit. Additionally, in situations where the accelerometers are offset by less than 5 cm, it may not be necessary to add states to the filter as the size effect will be negligible. This could be an added benefit, but even when adding the error states to the filter, the benefits that carouseling provides in lower mass and power requirements could greatly outweigh the burden of extra computation. 4.2 Size Effect In the particular gyrocompassing problem laid out in this thesis, the size effect is not a significant complication and the lever arm error states could possibly be eliminated. However, this simulated scenario has small lever arms and small uncertainties which might not be realistic to assume for other applications. The size effect was proved to be related to the offset of the accelerometers (the lever arm) and the angular rate of the slew squared. It was seen in this simulation, that for lever arms of 5 cm or more, the size effect becomes substantial and for a navigator to be accurate, the lever arm states must be implemented. Additionally, if the angular speed of the outer gimbal is larger than 80 degrees per second, the added velocity from the uncompensated lever arms will also cause a large error and it is necessary for the states to be incorporated into the filter. If these lever arm states are accounted for, the same accuracy can be attained, but more computation will be required as more states must be integrated in the filter. 4.3. OTHER CONSIDERATIONS 4.3 Other Considerations Other factors besides accuracy will determine the feasibility of implementing a carouscling scheme in the gyrocompassing application. It is important to remember that the IMUs must be physically slewed and this rotation must be performed and controlled by some mechanism. This mechanism will add some volume and mass (although still have lower total mass than a typical high-accuracy IMU) but also must be accurate enough to perform the slewing effectively. It was shown however, that the slew can most likely be Earth referenced as this does not drastically change the results of the inertially referenced slew, which can make the controller scheme less difficult to implement. Overall, the added complications of the carouseling scheme are not so severely intensive that they would preclude the implementation of the scheme for use in real-life applications. With the expected benefit of a 50% increase in accuracy, the added mass and designing of the gimballed control scheme can easily be justified. 4.4 Future Work A major area of future work would be to simulate the carouseling maneuver for trajectories and applications other than simply gyrocompassing. There are many other high-accuracy applications that could benefit greatly from the reduced mass and cost of MEMS IMUs, both in space and for terrestrial navigation. Other applications could implement other types of updates including GPS and might be concerned more with velocity error more than azimuth error as was the case for the gyrocompassing application. The specific challenges of different applications could affect the benefit of carouseling and must be investigated. Additionally, the effect of having more uncertainty in the slewing trajectory itself could also indicate how sensitve the results are to the preciseness of the slewing motion. A next logical step in this research that would also be of importance is to design actual hardware for use in real-world experiments to validate the results of the size 88 CHAPTER 4. CONCLUSIONS effect simulations. With real-world data the true value of incorporating the lever arm filter states can be ascertained and the ideal filter model for the gyrocompassing application can possibly be determined. This thesis has been focused on simulations for its results which must be confirmed by experimentation. Appendix A MEMS IMU Error Specifications 0.3 0.06 33 scale factor(in-run) misalignment nonorthogonality random walk 3.3 0.3 170 170 6.25 6.25 0.05 units dph dph ppm 8.25 6.25 6.25 0.01 ppm arcsec arcsec deg/ hr bias(turn-on) bias(in-run) scale factor(turn-on) scale factor(in-run) misalignment nonorthogonality random walk 300 240 50 50 6.25 6.25 0.009 300 240 50 50 6.25 6.25 0.009 p-g P-g ppm ppm arcsec arcsec mps/ /hr 900 900 sec 11 Current Gyro. Gyro. Gyro. Gyro. Gyro. Gyro. Gyro. Accel. Accel. Accel. Accel. Accel. Accel. Accel. bias(turn-on) bias(in-run) scale factor(turn-on) Time constant Improved Table A.I: Design Specifications for the Current and Improved Model MEMS IMUs APPENDIX A. MEMS IMU ERROR SPECIFICATIONS THIS PAGE INTENTIONALLY LEFT BLANK Bibliography [1] Theresia C. Becker. Approaches to optimal inertial instrument calibration using slewing. Master's thesis, Massachusetts Institute of Technology, June 2005. 121 Esmat Bekir, editor. Introduction to Modern Navigation Systems. World Scientific Publishing, Singapore, 2007. [31 Brown and Hwang. Introduction to Random Signals and Applied Kalman Filtering. John Wiley & Sons, New York, 3rd edition, 1997. [4] Charles Broxmeyer. Inertial Navigation Systems. McGraw-Hill, New York, 1964. [51 J. Connelly and A. Kourepenis. A micromechanical inertial measurement unit for tactical applications. Technical Report AIAA-2002-5050, Charles Stark Draper Laboratory, 2002. [6] Charles S. Draper, Walter Wrigley, and John Hovorka, editors. Inertial Guidance. International Series on Aeronautical Sciences and Space Flight. The Macmillan Company, New York, 1960. [71 Arthur Gelb, editor. Applied Optimal Estimation. The M.I.T. Press, Cambridge, MA, 1974. 18] Donald E. Kirk. Optimal Control Theory. Dover Publications, Mineola, NY, 2004. BIBLIOGRAPHY 191 Anthony Lawrence. Modern Inertial Technology: Navigation, Guidance and Control. Mechanical Engineering Series. Springer, New York, second edition, 1998. [10] C.F. O'Donnell, editor. Inertial Navigation: Analysis and Design. McGraw-Hill, New York, 1964. 111 George R. Pitman Jr., editor. Inertial Guidance. University of California Engineering and Physical Sciences Extension Series. John Wiley & Sons, New York, 1962. 1121 A.L. Rawlings. The Theory of the Gyroscopic Compass. Macmillan, New York, 1944. 113] Robert M. Rogers, editor. Applied Mathematics in Integrated Navigation Systems. AIAA Educational Series. AIAA, Inc., Reston, VA, second edition, 2003. [141 Laurie Tetley and David Calcutt. Electronic Navigation Systems. ButterworthHeinemann, Boston, 2001. I151 Walter Wrigley, Walter M. Hollister, and William G. Denhard. Gyroscopic Theory, Design, and Instrumentation. The M.I.T. Press, Cambridge, MA, 1969. 1161 Paul Zarchan and Howard Musoff, editors. Fundamentals of Kalman Filtering. AIAA Progress in Astronautics and Aeronautics. AIAA, Inc., Reston, VA, 2000. List of Figures 12 1-1 SEM photo of MEMS Tuning Fork Gyroscope . ............ 1-2 SEM photo of MEMS accelerometer . ................. 2-1 3-D Trace of a 2:1 Baseball Stitch Slew . ................ 2-2 Plots of gimbal angles and angular rates of a 60 second Baseball Stitch 28 Contributions of different error sources to the overall rotation error for . the Baseball Stitch Slew ........................... 2-4 . .. . . .................. Plots of gimbal angles and angular velocities for Baseball Stitch Slew . 34 . ................ Contributions of different gyroscope error sources to the change in ve36 locity error for the Baseball Stitch Slew with Reversal ......... 2-9 33 Contributions of different error sources to the overall rotation error for the Baseball Stitch Slew with Reversal 2-8 31 32 . ....... with Reversal .................................. 2-7 . The total contribution to rotation error from gyroscope errors for a non-rotating body . . .................. 2-6 30 The total contribution to rotation error from gyroscope errors for the Baseball Stitch Slew 2-5 12 17 ..................................... Slew ...... 2-3 . Plots of gimbal angles and angular velocities for Baseball Stitch Slew with Reversal Extended ................... ...... 37 LIST OF FIGURES 2-10 Contributions of different error sources to overall rotation error for Baseball Stitch with Reversal Extended . ................ 38 2-11 Contributions of different gyroscope error sources to velocity error for Baseball Stitch with Reversal Extended . .......... ... . 40 2-12 Total Contribution to angle error from gyroscope errors for Baseball Stitch Slew with Reversal Extended . .................. 41 3-1 Current MEMS IMU Azimuth Covariance for Baseball Stitch Slew . . 62 3-2 Current MEMS IMU Azimuth Covariance Comparison for Baseball Stitch Slew and 2-Position Gyrocompassing 3-3 . ............. 63 Current MEMS IMU Azimuth Covariance Comparison for Baseball Stitch Slew and ±3600 turns ............... 3-4 ...... Current MEMIS IMU Azimuth Covariance for Baseball Stitch Slew referenced in two different coordinate frames 3-5 . .............. ................... .. 67 Improved MEMS IMU Azimuth Covariance Comparison for Baseball Stitch Slew and 2-Position Gyrocompassing 3-7 66 Current MEMS IMU Azimuth Covariance Comparison with zupt update performed every 5 seconds 3-6 64 . ............. 68 Azimuth Covariance Error after 10 minutes for Varying Measurement Noise ...... ............. ................ 74 3-8 Selected Azimuth Covariance Error Contributions over time 3-9 Lever Arm Error ................... .... . ......... 75 78 3-10 Navigated solution compared to Azimuth Covariance when 3 cm lever arms are uncompensated .... ..... ........... ..... . . 80 3-11 Navigated solution compared to Azimuth Covariance when 10 cm lever arms are uncompensated ................... ...... 81 3-12 Increase in Azimuth Covariance when implementing Sub-optimal gains for varying lever arm lengths ............... ....... 82 LIST OF FIGURES 95 3-13 Increase in Azimuth Covariance when implementing Sub-optimal gains for varying outer gimbal angular rates . ................. 84 96 LIST OF FIGURES THIS PAGE INTENTIONALLY LEFT BLANK List of Tables ........................... . 45 . 60 2.1 Navigation Filter States 3.1 Initial Error States ............................... 3.2 Error Budget Comparison of Azimuth Error contributions in degrees for Baseball Stitch Slew and 2-position Gyrocompassing ........ 72 A.1 Design Specifications for the Current and Improved Model MEMS IMUs 89