THE EFFECTS OF NOISE ON MEASUREMENTS MADE IN A DISTRIBUTED PARAMETER SYSTEM by Rodney Robert Hersh , B. S. E. E. Pennsylvania State University (1968) SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE at the MASSACHUSETTS INSTITUTE OF TECHNOLOGY June, 1972 Signature Redacted Signature of Author Department of ~EectifAalEngineering, May 12, 1972 Certified by Redacted Signature-~----------- Thesis Supervisor Accented by Chairn -n Signature Redacted Oep t e-nal - __ ---- - - -- - - ---- mm tee on Graduate Students Archives JUN 27 1972 klRA IE9 THE EFFECTS OF NOISE ON MEASUREMENTS MADE IN A DISTRIBUTED PARAMETER SYSTEM by Rodney Robert Hersh Submitted to the Department of Electrical Engineering on May 12, 1972 in partial fulfillment of the requirements for the Degree of Master of Science. ABSTRACT It is the objective of this thesis to explore some of the problems associated with a distributed system driven by noise. Three main areas are investigated. First, the correlation of two sensors placed along a stochastically driven distributed system is considered and an expression relating correlation coefficient, sensor spacing and sensor position is found. Second, the noise bandwidth of a distributed system is examined. Finally, state estimation using the Kalman-Bucy filter and finite-dimensional approximations to the state-equation of a distributed system is explored. The effects of filter order (modeling accuracy) and multiple sensors are considered, and a lower bound on estimation accuracy is found for the single sensor case. A one-dimensional diffusion equation with two different sets of boundary conditions is used for the distributed parameter system. THESIS SUPERVISOR: Leonard A. Gould TITLE: Professor of Electrical Engineering -3ACKNOWLEDGEMENTS The author wishes to thank his thesis advisor, Professor Leonard A. Gould, for his suggesting and supervising this problem. His comments and criticisms throughout this work were most useful. Thanks are also due to the author's colleagues at the Electronic Systems Laboratory. In particular, discussions with Dr. Gervasio Prado and Dr. John Fay were most fruitful, and the computer program donated by Mr. Michael Warren was invaluable. The author is also grateful to the E. S. L. Drafting Department for their reproduction of the many figures in this work. Finally, the author wishes to thank Mrs. Rita Baughman for her patience in typing this manuscript. This thesis was sponsored in part by the National Science Foundation grant NSF 72917. -'4TABLE OF CONTENTS CHAPTER 1 page INTRODUCTION 1. 1 Motivation for this Study. . . . . . . . . . . . . . . . . . 5 1.2 1. 3 Background........... ......................... Outline of Thesis . . . . . . . . . . . . . . . . . . . . . . 6 6 CHAPTER 2 CORRELATION OF SENSORS 2.1 Introduction. . . . . . . . . . . . . . . . 8 2. 2 2. 3 2. 4 2. 5 2.6 Analysis of Fixed and Free End Systems . . . . . . . . . . Discussion of Function . . . . . . . . . . . . . . . . . . . Computation and Results . . . . . . . . . . . . . . . . . . Interpretation of Results . . . . . . . . . . . . . . . . . . Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . 8 11 12 CHAPTER 3 3. 3. 3. 3. 1 2 3 4 . . . .. 15 16 BANDWIDTH CONSIDERATIONS Introduction . . . . . Analysis of Fixed and Noise Bandwidth . . . Summary. . . . . . . CHAPTER 4 . . .. . . . Free . . . . . . . . . . . . . End Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 . . . . . . . 17 . . .. .. 21 . . . . . . . 22 ESTIMATION OF A DISTRIBUTED PARAMETER SYSTEM 4. 1 Introduction . . . . 4. 2 4. 3 4. 4 Mathematical Preliminaries . . . . . . . . . . . . . . . . .24 Truncation of State . . . . . . . . . . . . . . . . . . . . . 27 Multiple Sensor Estimation . . . . . . . . . . . . . . . . . 29 4. 5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . .38 CHAPTER 5 . . . . . . . . . . . . . . . . . . . . . 24 CONCLUSIONS AND RECOMMENDATIONS FOR FURTHER STUDY . . . . . . . . . . . . . . . . . . 39 APPENDIX A MODAL ANALYSIS OF FIXED END SYSTEM. . . 40 APPENDIX B POTTER'S METHOD FOR SOLVING STEADY STATE RICCATI EQUATION. . . . . . . . . . . . 44 1.0 INTRODUCTION 1. 1 Motivation for this Study: The purpose of this thesis is to investigate the behavior of sensors placed in a distributed parameter system and by using the outputs of these sensors to estimate the state of a distributed system excited by a stochastic source. Problems of this type occur in many physical situations such as the motion of an aircraft wing passing through turbulent air, the temperature distribution in a furnace being heated by an uncertain source, or the voltage and current distributions along a power transmission line terminating in a randomly varying load. Of particular interest in this work is the problem of correlating measurements made in a one-dimensional diffusion system driven by white noise at one end. Physically, this situation would correspond to the taking of temperature measurements along a metal bar heated at one end by a random source. Also of interest is the accurate esti- mation of the state of the one-dimensional diffusion system using the outputs from sensors placed along the system. The concept of white noise is used throughout this work, and although white noise is not physically realizeable it is useful because it represents a "worst case" phenomenon and because it makes the problems considered here analytically tractable. academic, The particular physical example, although somewhat is adequate for demonstrating the key issues which arise. -61. 2 Background: Previous work on the subject of noise in distributed parameter systems is rather limited. A digital simulation of a diffusive system was mdde by Hisiger (4) and the degree of correlation between two sensor placed along the system was studied. This study provides a qualitative feel for sensor correlation. The analytic description of distributed systems using modal analysis techniques has been presented by Murray-Lasso (5), Gould (2), and Prado (7), while the use of Laplace transformation and fre- quency domain methods is discussed in Could (2) and Hildebrand (3). State estimation methods for continuous time systems were developed by Kalman and Bucy (5) and these methods have been extended to infinite-dimensional systems by Falb (1). Estimation of the modes of a distributed system was considered by Prado (7) but in the context of pole-allocation and three-mode con- trol of a diffusion system where the uncertaintly was in the initial conditions and not the input. 1. 3 Outline of Thesis: This thesis is organized into three main chapters which treat, first, the degree of correlation between sensors placed along a distributed system, second, the noise bandwidth of a distributed system, and, finally state estimation of a distributed system . In each chapter the same distributed system described by a diffusion equation was considered, although two cases with different boundary conditions were studied in the first two chapters. -7Chapter 2 develops an analytic expression for the correlation between two sensors placed along the diffusion system for each set of boundary conditions. This expression is then evaluated numerically and a simple formula relating correlation coefficient, sensor spacing, and a measure of sensor position along the system is developed. In Chapter 3 the bandwidth of the diffusive systems is considered. The transfer function for these systems is developed and using communication theory techniques a measure of the noise bandwidth is found giving an indication of the behavior of a diffusion system when driven by noise. Chapter 4 is concerned with state estimation of a diffusion system being driven by white noise. The trade-offs between modeling accuracy and number of sensors used are considered. For the single sensor case a lower bound on estimation accuracy (for a common comparison among different order models) is developed. Appendices are included to derive fully one of the important results in Chapter 2 and to explain an interesting and useful way to solve the steady state Riccati equations that arise in the design of the Kalman- Bucy filter. -82. 0 CORRELATION OF SENSORS 2. 1 Introduction: In this chapter the question of how the outputs of two sensors in a noisy distributed parameter system are correlated is considered. Perfect point sensors were assumed and the distributed system of interest was described by a diffusion equation with two different sets of boundary conditions. Analytical results are derived for the corre- lation coefficient between two sensors, but numerical techniques were employed to evaluate the expressions. 2. 2 Analysis of Fixed and Free End Systems: The two diffusive systems studied are both described by the normalized diffusion equation 2 _ 3ax(zt) x(zt) 2.1 3Z at The systems are driven at the z = 0 end by a stationary white noise process, w(t), with zero-means and variance T~ S . Figure 2. 1 shows these two systems with their respective boundary conditions w(t) w(t) z. = 0 x(o,t) = w(t) z ; x(1,t) = 0 1 z = 0 x(o,t) = w(t) z= 1 ; iz x(z,t)j = 0 z=1 x(z,0) = 0 x(z,0) = 0 Fixed End Case Free End Case Figure 2.1. Two Diffusion Systems with Boundary Conditions -9Using the modal analysis techniques presented by Prado (7) for the fixed end case, the partial differential equation was represented by an infinite-dimensional set of ordinary differential equations which will be refered to as the state equation. .02O e o o For the fixed end case ,,,'~ Tr XW -7Tr + [ J T3 W~f AX W)+ applying the variation of constants formula for semi-groups from Prado (7), the state equation was solved to give cktc (Ym Hence, the equation for the output of a point sensor, y1 (t), placed at a point zf along the system is -10It is obvious that y1(t) is a zero-mean random process because E = 0 , and by squaring yi (t), taking expectations, W(t)j and evaluating integrals the variance of the steady state output is found to be A complete derivation of Eq. 2. 7 is found in Appendix A. Using exactly the same techniques, the variance of the output of a second sensor, y2, placed at the point z , is 2 4T2v I The cross-correlation between the two sensors, E y, (y) y(03) for the steady state case is found to be Then from the definition of the correlation coefficient [E { y }- E f Ya- }'A It is worth noting that if a non-normalized diffusion equation C E fYi were assumed, the expressions for , and E Y,% ya would each be divided by G< FiIY, and hence the correlation coefficient would remain the same as in the normalized case. -11In the free end case a different set of eigenfuctions must be used in the modal analysis giving slightly different results for the variance The steady state variance of the first sensor yj is ( The corresponding expressions for Y.10 L , T E Ky, yz } , of the sensor outputs. and e follow easily. 2. 3 Discussion of Doubly-Infinite Series At this point some comments on the doubly-infinite series 51mrAv7Tj WMv\ SjvA-n2 are in order. finite for all values of z in the interval 2 + 0 . (o , I This funct ion is , but diver ges as To observe this behavior consider Figure 2. 2 which 3hows the sign behavior of the arguement of the series as a function of the indices m and n for a constant value of z. A - d C 0 Fig. 2. 2 + + M L~1. a b e Sign of the Series Argument -12a The points and b occur when VIT\ Z and respectively, giving 6.. equals . Qr and IT The points c and d are determined similarly. Now, when summing over a finite number of indices, the series should be summed over an integer number of cycles of 'WTT Z SIVN because the doubly-infinite series is oscillatory about its actual value as index of summation increases. This means that the function should be summed to either n = b and m = d, or to the end of the next cycle n=e and m=f. where For example, let z=0. 20. a and b will correspond to n=5 and n=10 Then the points respectively. the series should be summed over an equal multiple of Thus, 10 indices along each axis. As z goes to zero, the first sign change points infinity leaving only positive terms. 5 1n ;Z )I a and c go to The main diagonal terms are then - CoSaln 7r2)which clearly diverges when summed from m=1 to infinity. The divergence of the series as z goes to zero should be no surprise because the system is being driven by a white noise process which has infinite variance. All the preceding analysis in this section also holds for the free end case where the doubly-infinite series is slightly different. 2. 4 Computation and Results: To examine the degree of correlation between two sensors as a function of their location, some measure of their spacing and position 1.0 b 0.9 q 0 * 0.5 0.6 I I.-' I K. 0.7 Ir I 0 .08 I .16 I .24 I .32 I 40 I .48 I .56 AZ about a point Zc Fig. 2.3 I f1m , p End Correlation between Sensors Equi-spaced about Z - Fixed F i .64 I .72 1.12 NV N0 0.9- N0 0.8P 0.6 -0 0.5- 0) .08 .A6 .24 .32 .40 .48 .56 .64 AZ about a point ZC Fig. 2.4 Correlation between Sensors Equally Spaced about Z - Free End .72 -15along the distributed system must be devised. was used. I, ZC , Zj . Let z 1 The following method and z2 denote the position of two observers, Then A 7 , the spacing between them is 2,- ZI , and their midpoint is (Z, -2,+2 . Numerically evaluating the correlation coefficient for increasing values of A Z about a center point zc yields the curves shown in Figure 2. 3 for the fixed end case and those shown in Figure 2. 4 for the free end case. Figure 2. 3 and 2. 4 reinforce the intuitive feeling that for a given spacing between sensors, their outputs will be more correlated as they are placed farther from the driving noise because of the rapid attenuation of inputs to a diffusive system as one moves into the system. The similarity of the two figures also indicates the relatively small effect that different boundary conditions have on the diffusive system. 2. 5 Interpretation of Results: Observation of Figures 2. 3 and 2. 4 reveals that over the interval 0. 5 F < 0. 9 the correlation curves can be approximated relatively well by straight lines. on Figure 2. 4. Three straight line approximations are shown Furthermore, when these straight lines are extended to the vertical axis they intersect in a common point, e =1. 12. In addition, if the slope of these straight line approximations versus the center point zc is plotted on a log-log scale, a straight line relationship is obtained. The exact relationship for the free end case is Combining this expression with the slightly different one for the fixed -16- sensor spacing AZ , and sensor midpoint ~ /~ 52 - , end case, the following relationship among correlation coefficient zc is obtained. ,3 O54O. Q413 While this expression is an approximation for both the fixed and free end cases of the diffusive system, it is accurate to within 5% over This means that for a fixed value of g , the specified range for '. the sensor spacing is directly proportional to the center point zce Furthermore, for ? = 0. 6 the sensor spacing A 2 equals the midpoint zc. The linear relationship between L. and Z c indicates that as zc increases, i. e., the sensors are placed farther from the random driving noise, that their outputs become more linearly dependent. 2. 6 Summary of Results: In this section the problem of relating sensor correlation to sensor placement in a distributed parameter system driven by white noise has been solved. It has been demonstrated that sensor correlation is relatively independent of the boundary conditions for a diffusive system. The results support intuitive feelings about distributed systems and provide quantitative relations for the qualitative results shown by Hisiger (4). Since the concept of white noise was used exclusively, the results of this section provide a lower bound on the correlation of two sensors (if a deterministic input were present the outputs of two observers would be perfectly correlated, (p = 1). -17- 3. 0 BANDWIDTH CONSIDERATIONS 3. 1 Introduction: This chapter is concerned with the bandwidth of a diffusive distributed parameter system. The bandwidth of a distributed system is of interest because it will affect the propagation of noise through the system and reveal some interesting facts about sensor placement in a diffusive system. Analytical expressions for the transfer functions for both the fixed end and free end cases described in Chapter 2 are derived and communication theory techniques are employed to obtain a measure of the bandwidth of these systems. Numerical techniques were again necessary to evaluate the analytic results. 3. 2 Analysis of Fixed and Free End Systems: To obtain a measure of the bandwidth of the diffusive systems being considered the transfer functions for these systems were derived using Laplace transformation techniques. The application of Laplace transforms to partial differential equations is discussed by both Could (2) and Hildebrand (3). Starting with the diffusion equation and boundary conditions for the fixed end case the Laplace transform is taken, 55Xz) - s 0) X (SX X(zo (Zi0 =s0 391 3 -18Selecting a general solution for this equation and applying boundary conditions yields, v ) 1 X(Z s)= C' e Z e -'"Z 3.3 -5A( e e Ce -e Hence the transfer function is C X (Zs>- H(z s) - e 3. - e ef Because the diffusion equation is stable, the real part of s may be set equal to zero to obtain the frequency response, - e-VF(I -7) -7) I-i (z, ~ e- Vj Some limiting cases of interest are as I-z W -3.( H -+ H , 0 I - Of0 H For the free end case the frequency response was found to be H (z,j F I--Z) Z) + e - 3-L 6 with the limiting cases -F+ e-v, H-+ I as z -+o H I f e -V a 1 0 . r x=0.1 0.9o o..0 0*1 0.2 0.70- IHI 0.5 I 0.3 o - - 0 m-W...- 0-- *- Ib * 0 0.5 -o. 0 " k a0 0 N0ft 0 I 0 a 0.7 0.3 0.1 0*N ON ~0 I. 0.9 0I 0.1 on-o m0mw o I 1.0 10.0 O0 *%eN 4 h 100.0 FREQUENCY (rad/sec) Fig. 3.1 Magnitude Transfer Function vs. Frequency for D.P.S. Input One End, Other End Fixed 1.0 0.8 x.0.1 - -0A I 0.41- 0.3N 0.4 0 I 010 0.2 - H| 0 -6 0.8 V 1v 0. 0. 1 I 1.0 %1 0* I 10 .0 100.0 FREQUENCY (rad/sec) Fig. 3.2 Magnitude of Transfer Function vs. Frequency for D.P.S. Input One End, Other End Free -21The frequency response of the fixed end case is shown in Figure 3. 1 with that for the free end case shown in Figure 3. 2. Both are shown for several values of z. 3. 3 Noise Bandwidth: In communications problems it is often of interest to know the mean-square value of the output of a network driven by noise. This desire has led to the noise-equivalent bandwidth defined as the bandwidth of an ideal rectangular pass network having the same mean-square output as the actual system. _H Mathematically, (:3wj)j AL.W 3.7 In distributed systems one is also concerned with how uncertainty propagates in a system and hence how uncertainty will affect measurements of the system. However, the noise-equivalent bandwidth concept that is useful for lumped systems must be modified for use with distributed systems. The reason for this necessary modification is that the frequency response is a function of both frequency and distance. diffusion system being considered here when z fixed case Since Hma approaches 1 In the in the H goes to 0, as does the intergral in the formula for Wn also goes to zero, by and in this case Wn L 'Hopital's rule Wn may be defined, indeed does approach a non-zero limit as was shown by numerical evaluation of Wn , a result difficult to interpret -22physically. Therefore, the measure of bandwidth used in this work was the integral of the magnitude squared of the frequency response. 00 IF(z U)i 01 W 3. E3W (Z) H It is easy to show that for high frequencies the magnitude of H decays exponentially with W and hence the above integral exists. A numerical integration scheme using Simpson's rule was employed to evaluate the bandwidth integral and the bandwidth versus distance is shown in Figure 3. 3 for both the fixed and free end cases. indicates that higher frequencies, This figure and hence higher modes of the sys- tem, are attenuated rapidly along the system. 3. 4 Summary: From the work in this section a qualitative feel for the bandwidth of a diffusive system has been developed. An important inference from this work is the location of sensors used to estimate the state of a system. Because of the rapid attenuation of higher modes with increasing distance from the driven end, one would expect to place sensors nearer the driven than the fixed or free end. Furthermore, one would expect the "optimal" sensor location to move closer to the driven end as an estimation of more modes is attempted. This qualitative feel for sensor placement will be confirmed in the next chapter. -239590- 80- 70- 60- z50I- Z 40- 30- 20- 10 FREE END FIXED END 0. 0 0 0.2 0.4 0 0.6 0.8 z Fig. 3.3 Bandwidth vs. Distance 1.0 - -2h. 4.0 ESTIMATION OF A DISTRIBUTED PARAMETER SYSTEM 4. 1 Introduction: Chapter 4 is concerned with the problems of state estimation in distributed parameter systems. Estimation is of interest because it is a necessary prerequisite for the control of a system. The appli- cation of modern estimation theory to distributed systems creates problems not normally encountered when working with finite dimensional systems, such as truncation of state (accuracy of model) and sensor placement along the system. Both of these problems are considered here along with multiple sensor estimation of a diffusive system. For the single sensor case a lower bound on the estimation error is developed. 4. 2 Mathematical Preliminaries: A necessary requirement for estimation of any system is that it be observable. Strict conditions for observability of distributed systems are given by Prado (7) which for the problems being considered here reduce to the condition that for a system to be observable a sensor must not be placed at the zero-crossing point of any mode being considered. For a three mode estimator of a diffusive system of unit length this means that sensors may not be placed at z = . 25, . 333, . 50, . 667, or . 75. In general, all zero crossings. sensors must be placed at irrational points to avoid This placement of observers is intuitively correct -25because no information about a mode is gained if it is always zero at the sensor point. Because of the finite dimensional limitation of numerical tech- niques, the infinite dimensional state equation of a distributed system must be truncated for any practical considerations. As has been seen in Chapter 3, the higher modes of a diffusive system are likely candidates for truncation since they are attenuated rapidly in the system. However, when studying estimation schemes of various orders, a common basis for comparison is needed. The trace of the error covariance matrix provides a measure of the magnitude of the errors in estimating the state and will be used in this chapter. Furthermore, because different numbers of modes are estimated by different order filters, the sum of the estimation errors of the first three modes will provide a common basis for comparing estimators. For the remainder of this chapter this criterion r will be denoted . This choice is also motivated by the fact that three modes are often observed when trying to estimate the behavior of a distributed system. The K.alman-Bucy filter (5) was used for the estimation because it provides a minimum mean-square estimate of the state. [() interest here is the error covariance Of particular which is propogated by the Riccati equation (t) = A (t) + (k) + BQB' - $_(t) M (z) R A and B from state equation M (z) Q observation matrix from Driving noise covariance 2. 2 2. 6 M (z) (t) 4.1 -26- C(t) Observation Noise SYSTEM w M) Driving- Noise t) ~Vz x A + 7, M'(z)_R A +f M~z) E {w(t)} = 0, E jw(t)w (t +r) E {g(t)} = Q Fig. 4.1 7- Q8(r) E{ f(t)C'(t+r)} R8(r) Structure of System and Kalman-Bucy Filter R Sensor noise covariance N Order of system, i. e., A is N X N matrix To be considered here is the steady state case of the Riccati equation, (t) = 0. The existence of the steady state solution is guaranteed by the controllability and observability of the system. Potter's method (8), an algebraic routine, was employed to solve the steady state Riccati equation. A description and derivation of Potter's method appears in Appendix B. Figure 4. 1 shows the structure of the system and the Kalman-Bucy filter. 4. 3 Truncation of State: To determine the sensor placement that minimizes the mean- square estimation error, a sensor was moved along the fixed end diffusion system, thereby changing the observation matrix M(z) and the Riccati equation for the error covariance was solved to yield the steady state estimation error. v ~ 3 A plot of steady state error for the first three modes, , versus sensor placement for six different order estimators is shown in Figure 4. 2. These results reinforce the intuitive feel that a higher order Kalman-Bucy filter (better model of the system) will yield a better estimate than a lower order filter. However, with a higher order filter the best estimate is much more sensitive to small deviations in position about the optimal point. Of considerable interest is the change in optimal sensor location as the order of the filter varies. Plotting the optimal sensor location versus the order of the Kalman-Bucy filter on log-log paper yields a straight line -28- 3000 - 25 3rdORDER 20t 4h. 69 o 0 M,15 -N% 10 - 4: R ='.0 12th 5- 01 0 .05 .10 .15 .20 .25 .30 .35 Z DISTANCE ALONE SYSTEM Fig. 4.2 Estimation Error vs. Sensor Placement for Various Order Filters -29relationship as is shown in Figure 4. 3. zopt = 5. N-0.855 Analytically, 36 N6 12. , 4. 2 Strictly this equation is valid only over the interval stated but the author sees no reason why it shouldn't hold for all N ', 3. Vr Figure 4. 4 shows 23 of the best estimate for each order fil Ler versus the sensor location zopt where that estimate was made. A definite relationship between the error of the best estimate and sensor p osition appears; giving +f Substituting Eqn. = 3 0 zolo 3 4. 3 4. 2 into 4. 3 yields IrI Og_. + 3_: = /400 N - 0,QQN 4. 4 This expression gives a lower bound on the estimation error of the first three modes as a fuction of the order of the filter 0.l +V 2 4 00 N 3 ' 4.5 the equality holding only when the sensor is placed at the optimum point given by Eqn. 4. 2. Again one should note that this lower bound has been developed only for 34 N 4. 4 12. Multiple Sensor Estimation: From the previous section it can be seen that the system of a distribu- ted system driven by a stochastic source can be estimated to a very high degree of accuracy provided a large enough model of the system is used in the -30-0 1.0; 0 4- 0. 0 0\\\ z w 'M.10 w Zp>.0 0 z w (I) .5 zop =5.0 N CL 0 I .01 I 2 1 I I I 4 I 8 (ORDER OF FILTER) N Fig. 4.3 Optimum Sensor Placement vs. Order of Filter 12 D 00 - 15 - - 20 trls =8 0 pt * 5 - 0 .05 .10 .20 .15 .25 z Fig. 4.4 Optimal Estimate Error vs. Sensor Position .30 4-32- estimator. However, it is usually not practical to work with systems of large dimension giving rise to the question: Is there a point of Figure 4. 5 shows diminishing returns in the modeling of the system? the error of the best estimate of the first three modes versus the dimension of the filter used to make the estimate. around N = 8 A "knee" in the curve appears indicating that the per unit gain in the estimation accuracy for increases in filter order is steadily diminishing. further work using more than one sensor the value the dimension of the filter. Therefore, for N=8 was chosen for It should also be noted that the computation time needed to solve the algebraic Riccati equation is proportional to N 3 , thus there results a savings of one-half in the time needed to solve an 8th order Riccati equation as compared to a 10th order one. When using more than one sensor to estimate the state, the obvious problem of sensor location arrises. Since this problem for the single sensor case has been solved, a reasonable extension is to place the first sensor at the position deemed optimal in the previous section and vary the second along the system until a minimum of fY ]I 3 is obtained. For the eighth order Kalman-Bucy filter being considered, point for the first sensor is zj = 0. 08485. the optimal Figure 4. 6 shows the effect of using two sensors to estimate the state of the diffusive system. A definite improvement in the estimation error of the first three modes occurs when two sensor are used as can be seen by comparing the two curves in Figure 4. 6. -33- 20- 16- 12- K 0 8 4- 01 0 4 12 8 N Fig. 4.5 Optimum Estimate vs. Filter Dimension 16 a a asa 0 a 25- 20Q = 10.0 R =1.0 15ro NI 100 a -ONE 5 - 0 0 SENSOR a -TWO SENSORS,ONE FIXED AT Z =.08485 .05 Fig. 4.6 .10 .15 Z .20 .25 .30 Comparison of Estimate Error Using One and Two Sensors .35 However, the most interesting facet of the two sensor case is that both sensors should be placed at the same point, z1 = z 2 = 0. 08485. This phenomenon is explained by noting that the best sensor location maximizes the signal-to-noise ratio for all the modes being considered. Because the two sensors were assumed independent of each other, their individual errors can be "averaged out" by the Kalman-Bucy filter. Hence, intuitively one might suspect that each sensor should be placed at the point of maximum signal-to-noise ratio. Furthermore, when more than two sensors are used, the same result still appears, namely that the optimal location for all observers in a diffusive system remains at the maximum signal-to-noise ratio point. When using M sensors (M> 2) in the estimation process the following scheme was used for sensor placement. The first M-1 sensors were placed at the optimal point and the Mth one moved along the system to find a minimum for 'r 23 . In each case the optimal point for all M sensors was the same, 0.-08485, for the eighth order filter. Estimation of the fixed end diffusion system with multiple sensors yields the plot of estimation error versus the number of sensors shown in Figure 4. 7. Both the total error in estimating all eight modes, +r 2. and in estimating only the first three modes, Y 3 , are shown. While increases in accuracy for all eight modes occur, a much larger increase in estimation accuracy is evident for the first three modes. Also, a definite "knee" appears in both curves indicating that a point of -36- 40 F 35 h 10 0 HI 30 F a) 8 FO 0 C.) 0 0 251- 6 NA 20 - 4 0 15 2 - 0 ) 0 iI I Fig. 4.7 Ii 2 Ii Ii 4 3 NUMBER OF SENSORS Estimation Error vs. Number of Sensors I 5 0 -37diminishing returns for the addition of more sensors. For an accurate comparison of the multisensor cases Table 4. 1 gives the optimum error covariance of the first three modes as a function of the number of sensors. Number of Sensors 2 1.429 0.752 0.547 4 0.423 V- 222 2. 441 1. 607 1. 270 1. 046 0. 932 332 3.590 2.637 2.177 1.855 1.677 11 Estimate Error 3 Table 4. 1. 5 0.368 Estimate Error of First Three Modes By observing Fig. 4. 7 and Table 4. 1, one can conclude that by using three independent sensors the "knee" of the curve is reached. Hence, any further addition of sensors buys little more as far as increases in estimation accuracy are concerned. To determine if the optimal sensor location were affected by the strength of the driving noise, the estimation process was rerun using larger driving noise covariances, previous work where Q equaled 10. remained fixed at 1. 0. Q, of 20 and 50 as compared to the The observation error covariance No change in optimal sensor placement occured for either the one or two sensor cases indicating that sensor placement is independent of the driving noise strength. -384. 5 Summary: In this chapter a study of estimation of a diffusion system has been presented. The consideration of model accuracy reaffirmed the fact that better models make better estimators but also revealed the fact that there is a point of diminishing returns beyond which large increases in the state of the model used in the estimator are necessary to produce modest increases in estimation accuracy. If the performance criterion is an accurate estimate of the first three modes of the system, a reasonable choice since the first three modes cover a 10 to 1 frequency range, this point of diminishing returns occurs when a finite dimensional system of order eight is used to approximate the diffusion system. Further- more, a lower bound on estimation error with the order of the filter as a parameter was developed. This bound is attainable only if the single sensor is placed correctly. For estimation with multiple sensors again one is faced with a diminishing returns situation. While the accuracy of estimation can be improved with the addition of more sensors at the optimal point, for the performance criterion considered here the use of more than three sensors gains little. -.395. 0 CONCLUSIONS AND RECOMMENDATION FOR FURTHER STUDY It has been the goal of this thesis to explore the problem of measuring the behavior of a distributed parameter system driven by a stochastic input. While the limited case of a one-dimensional diffusion equation was considered, the results shed light on the problems peculiar to distributed systems. Chapters 2 and 3 provide insight into the propogation of uncertainty in a diffusion system and develop a relationship among the variables of sensor correlation, sensor placement and sensor spacing. The work in these chapters also shows the relative independence of the diffusion equation to different boundary conditions. Chapter 4 considered the problems of state estimation of a diffusive system using a finite-dimensional model. sensor location, The questions of model size, and multiple sensor estimation were explored. Much work remains in the area of estimation of distributed systems. Obvious extensions of this work would be the estimation of systems described by hyperbolic or elliptic partial differential equations. Another area open for study is that of the trade-offs between the dimension of the filter used and the number of sensors used. A better knowledge of esti- mating distributed systems will make the job of control much simpler. -440APPENDIX A Consider the following diffusion equation with its related boundary conditions, S(Z,, ~) ETt~L A.1 YL(z , o) IL(I, L) ) t( = w ). where w(t) is a white stochastic process with mean zero and variance U -'-) ) . Because both ends of the system are zero, an orthonormal set of eigenfunctions for this system is Then i xh Ui) i0 2(Z) f) J and I ( (Z) = F Y(Z) YD SA7nT A9 I Y - A .3 integrating by parts I . T 262. 51 ',Y aZ ALZ I But upon application of the boundary conditions the first term on the AAl -):1right-hand side is zero. Integrating by parts again I I' Fa s; vt h -rz dz Y (Z =- t) a 1T C 0 S - -Z J I , '3 Jr P 1 2 -rid, X(Zj) Observing that the last integral is A5 S;V% YOTZ A 7. - V2TT n , and applying ) f boundary conditions gives ~V(Z)v f-I = A2 7T)- flt)L TZC cZ aY M\ + V7hi-T wU-m. Combining all the xn's into an infinite dimensional vector yields the equation 0 X( ) d 0 -14 -1 0 0 9.. J F It .4 2 3 j A.7 From the variation of constants formula for semi-groups presented by Prado (7), (i),Z S (VO(t Xo wcwv +f 0 -42where F-rr e S (vi') 5 U~z e a' Tir ,J--4)'9 0 -4-PaL (n') IQ~Jrj -\ X Since 1 VT (0) V-ITT 0 4 na -v) C 2N Tr Wv&) 0 017- I e rd W (T) cOr e-s) 26 With a point sensor located at o, as;n ITZ, I M z /~(z.) y 1 (Q)= W r-T A. IO (q)d C the output 1 XW OF, s ;V% y 1 (t) is given by ) FTf0 - --1 z, A.11 Combining M (zl) with the above expression for the state gives **O S; VIA -Mz I E I y, (,)} = T2 -)V Y) n W(7-) O(q-. 0 , Clearly Te Y1 (~)z. '~J cc: tooI - w () O V I T:) 0 Sf4Vi rI (t~) r I t - Y1 2 Tr - (f -7) A, 13 -43Then the variance of y 1 (t), value of yi 2 t>O, may be found by taking the expected (t) El: Y, (0)1 n Si M~ z, S;VV1 r j! 7~r~ Wi Y)$~,v11rZ f0 7r - -0 d 7- Cti).. T2 X (V- A. Iq Evaluating the integrals and cancelling out terms leaves 00 i . = YA nSn 2 MTr, s -(d+ Vna T M C Y% TrI7 Ii [ I A.15~ Ignoring the transient term reduces the expected value of yj 2 to y121 sYAiw OI LH, a 7rZ, S'V n TrZ, %2 - A.16 By the same method the cross-correlation between two sensors located at z1 and z2 is found to be M Trz Z hI P1 S;In V" 1 + =kY .51i, qVT Z A, 1 The correlation coefficient is then found using the definition (E -Y I , E:= yV yY1jF A4Ia -144APPENDIX B In finding the steady state estimation error using the Kalman filter it is necessary to solve the following algebraic Riccati equation = ('M R M Z)2 8' - I 1Q A 2 + 2 A' f- 0 13. The algebraic method of solution outlined here is due to Potter (8). For this solution to be valid the following conditions must be met, [A,8]]) Cofri'oila6e t) (z M)] 06seyvcxle Q>0 3) t) R >o. Assuming that the state is an n-dimensional vector, the dv'%x2n matrix v =~. V is formed ~Z) R- M(-Z) -A' First it will be shown that if X -A is an eigenvalue This condition is known as quadrilateral symmetry. of V so is - A Consider the %X 2n vimatrix ~.3 Q T Observe that . TV J 11 Then by direct multiplication 3 -M45Multiplying each side by J again gives e = a o V 13.5- 'Te If XA is an eigenvalue of V with an associated eigenvector e V showing that - = - is an eigenvalue of V Since V and V e Te with eigenvector Je have the same eigenvalues, -A is therefore also an eigenvalue of V . At this point the conditions of controllability and obserability guarantee that no eigenvalues will have zero real parts. Now consider the n eigenvalues of V with positive real parts Arranging the eigenvectors in column form yields a which can be partitioned into two Vxvn matrices Q, ,a , with their respective eigenvectors X VA X andY matrix by X Notice that the following equation holds where A- is the wVxn Jordan canonical matrix of the eigenvalues with positive real parts. It is now the claim that the solution to the Algebraic Riccati equation can be written = xY-1 B. -W46The proof follows easily. From Equation B. 8 AX + C Y DX - XJ\ A-"Y /0. Yr _ Postmultiplying both of these equations by Y (Y can be shown to be invertible) and premultiplying Equation B. 11 by XY 1 ANY-' +C = A _ X Y'DX)Y' -XY'-1A I Subtracting Equation B. 13 from B. 12 XY_' AXWFurthermore, least C or D - 4 for C and D non-singular, gives 91 Xi2Y A)/1 leaves XY-1D XyYc 0. semidefinite and symmetric with at it can be shown that , the solu- tion to B. 1 , is positive definite. In the actual algorithm, first the Qyt X% matrix and then its eigenvectors and eigenvalues are found. V is formed Next the eigenvectors corresponding to eigenvalues with positive real parts are placed in a SA X Y X V V1 matrix. matricies. This matrix is then partitioned horizontally into two The lower matrix tiplied by the upper matrix to the X to give algebraic Riccati equation. Y is then inverted and premul, the steady state solution -147REFERENCES 1. Falb, P. L., "Infinite-Dimensional Filtering: The Kalman-Bucy Filter in Hilbert Space", Information and Control, 11, 1967. 2. Gould, L. A., Chemical Process Control Theory and Applications, Addison-Wesley, Reading, Mass., 1969. 3. Hildebrand, F. B., Advanced Calculus for Applications, PrenticeHall, Engelwood Cliffs, N. J., 1962. 4. Hisiger, R. S., "An Investigation into a Measurement tJncertainty Principle for a Distributed Parameter System Driven by Noise", S. M. Thesis, M. I. T., May 1971. 5. Kalman, R. and Bucy, R., "New Results in Linear Filtering and Prediction Theory", Trans. ASME, ser. D, Vol. 83, 1961. 6. Murray-Lasso, "The Modal Analysis and Control of Distributed Parameter Systems", Ph. D. Thesis, M. I. T., 1965. 7. Prado, G., "Observability, Estimation, and Control of Distributed Parameter Systems", Ph. D. Thesis, M. I. T., Aug. 1971. 8. Potter, J. E., "Matrix Quadratic Solutions", SIAM Journal of Applied Mathematics, Vol. 14, No. 3, May 1966.