A Modeling Study of the Sensitivity of Natural Frequency of Vibration to Geometric Variations in a Turbine Blade by Daniel A. Snyder A Project Submitted to the Graduate Faculty of Rensselaer Polytechnic Institute in Partial Fulfillment of the Requirements for the degree of MASTER OF ENGINEERING in MECHANICAL ENGINEERING Approved: _________________________________________ Ernesto Gutierrez-Miravete, Project Adviser Rensselaer Polytechnic Institute Hartford, CT December 2010 (For Graduation August 2011) © Copyright 2010 by Daniel A. Snyder All Rights Reserved ii Contents Contents ............................................................................................................................ iii List of Tables ..................................................................................................................... v List of Figures ................................................................................................................... vi Acknowledgement ........................................................................................................... vii Abstract ........................................................................................................................... viii 1. Introduction and Background ...................................................................................... 1 1.1 1.2 1.3 High Cycle Fatigue ............................................................................................. 1 1.1.1 Zero-Mean Stresses ................................................................................ 1 1.1.2 Non-zero Mean Stresses ......................................................................... 2 1.1.3 Modified Goodman Diagram ................................................................. 2 Vibratory Response ............................................................................................ 2 1.2.1 Design Philosophy ................................................................................. 3 1.2.2 Frequency Prediction ............................................................................. 3 1.2.3 Campbell Diagram ................................................................................. 3 The Problem........................................................................................................ 5 1.3.1 1.4 Defining a Design Space ........................................................................ 5 Previous Work on this Topic .............................................................................. 6 2. Theory and Methodology ............................................................................................ 8 2.1 2.2 Monte Carlo Simulation ..................................................................................... 8 2.1.1 Geometric Parameter Scheme ................................................................ 8 2.1.2 Blade Root Geometry........................................................................... 10 2.1.3 Tip Shrouds .......................................................................................... 10 2.1.4 Generating Random Variable Combinations ....................................... 11 Modal Analysis ................................................................................................. 13 2.2.1 Execution of Computations .................................................................. 14 3. Results and Discussion .............................................................................................. 16 iii 3.1 Mode Shape Identification ................................................................................ 16 3.2 General Trends in Response ............................................................................. 16 3.3 3.2.1 First Mode Shape ................................................................................. 16 3.2.2 Second Mode Shape ............................................................................. 16 3.2.3 Third Mode Shape ................................................................................ 17 3.2.4 Fourth Mode Shape .............................................................................. 17 3.2.5 Fifth Mode Shape ................................................................................. 17 Regression Functions ........................................................................................ 17 3.3.1 Linear Regression................................................................................. 17 3.3.2 Nonlinear Regression ........................................................................... 18 3.3.3 Partial Least Squares Regression ......................................................... 19 4. Conclusion ................................................................................................................. 20 5. References.................................................................................................................. 21 6. Appendices ................................................................................................................ 22 6.1 Appendix 1: Turbine Blade Parametric Model ................................................. 22 iv List of Tables Table 1: Geometric Parameters and their Ranges ............................................................ 12 v List of Figures Figure 1: Modified Goodman Diagram ............................................................................. 2 Figure 2: Campbell Diagram ............................................................................................. 4 Figure 3: Parametric Model of Fictitious Turbine Blade................................................... 9 Figure 4: Generic Tip Shroud Design .............................................................................. 10 Figure 5: Airfoil Section Definition ................................................................................ 11 Figure 5: Top View of Fictitious Parametric Turbine Blade Model ............................... 22 Figure 7: View of Fictitious Turbine Blade Root Attachment Geometry ....................... 23 Figure 8: Close-up View of Turbine Airfoil Section Definition ..................................... 23 Figure 10: Random Turbine Blade – 1 ............................................................................ 24 Figure 11: Random Turbine Blade – 2 ............................................................................ 25 Figure 12: Random Turbine Blade – 3 ............................................................................ 26 Figure 13: Random Turbine Blade – 4 ............................................................................ 27 Figure 14: Random Turbine Blade – 5 ............................................................................ 28 Figure 15: Mode Shape 1 at 299 Hz ................................................................................ 29 Figure 16: Mode Shape 2 at 712 Hz ................................................................................ 29 Figure 17: Mode Shape 3 at 1267 Hz .............................................................................. 30 Figure 18: Mode Shape 4 at 1694 Hz .............................................................................. 30 Figure 19: Mode Shape 5 at 2905 Hz .............................................................................. 31 vi Acknowledgement Thanks to Jeff Beattie for help with ANSYS and for giving me the concept for this project. Thanks to Grant Reinman for consultation on statistical methods such as partial least squares regression and principal component analysis. Thanks to Pratt & Whitney for providing me with an environment that fosters research and new ideas. vii Abstract A Monte Carlo simulation is used to investigate the effect of geometric variation in a turbine blade on its natural frequencies of vibration. First, a parameter scheme is selected and implemented in a parametric 3D solid model. The parameters chosen are those that affect section properties of the airfoil at three different sections throughout the airfoil. Other parameters control other geometric features in the turbine blade. A total of 37 geometric parameters are independently varied to test the response of the first 5 natural frequencies of the turbine blade model. The material model is isotropic with an elastic modulus of 30E+6 psi and a Poisson Ratio of 0.3 and a density of .289 lb/in 3. A random set of values for the parameters is generated using a Latin Hypercube function implemented in Matlab. A total of 50 combinations of 37 independently varied parameters result in 50 sets of natural frequencies. The impact of each geometric parameter on each natural frequency is quantified by calculating a linear regression. The linear regression function approximates each natural frequency as a linear function of each independent variable (geometric parameter). This result shows that the fundamental mode of vibration is affected mainly by the moment of inertia of the blade about the fixity point. All parameters that tend to increase the moment of inertia of the blade tend to decrease the first natural frequency. The second natural frequency is also bound to a similar relationship with mass moments because its mode shape is also a bending mode. The third frequency is associated with a torsional mode shape. The parameters that increase airfoil root stiffness increase this frequency. The parameters that decrease the polar moment of inertia of the tip tend to increase this frequency also. The importance of this type of study is to explore and understand the implications of varying parameters and combinations of parameters in the design space. Further design studies like this could allow the design engineer to tune certain frequencies out of undesirable ranges while leaving other modes frequencies unchanged. viii 1. Introduction and Background This project deals with prediction of the effects of manufacturing/geometry variation on turbine blade vibration. It is important to prevent resonant vibration of turbine blades during operating conditions. Turbine blades experience unsteady forcing functions at many different frequencies. If the blade experiences an excitation frequency equal to its natural frequency, the blade will usually fail very quickly in a failure mode known as high-cycle fatigue. This fatigue mode is characterized by relatively low fluctuation in stress and very high frequency of fluctuation. This failure mode embrittles the material and causes it to crack in regions of high steady stress. Characterization of materials in this failure mode gives rise to the Goodman Diagram. 1.1 High Cycle Fatigue High cycle fatigue is the failure mode of a material experiencing fluctuating stress. Compared to yielding or rupture failure in which stress is said to be steady or unchanging, fluctuating stresses cause materials to fail by becoming brittle and cracking. The characterization of materials in this failure mode involves testing materials subject to alternating stresses of different means and different amplitudes. In an alternating stress condition, the mean stress is defined as the average stress over a long period of time. For stresses whose variation is sinusoidal over time, the mean stress is simply the average of the maximum and minimum stress. 1.1.1 Zero-Mean Stresses Testing materials with an alternating stress that has a zero mean allows the engineer to measure how many cycles before the material fails. Typically, these data are plotted on a chart with number of cycles on the horizontal axis and stress amplitude on the vertical axis. The number of cycles is typically on a logarithmic scale for plotting purposes since number of cycles to failure can become exponentially larger for low stress levels. Some materials, such as steels, have a stress level below which fatigue failure will never occur – this is called the endurance limit. 1 1.1.2 Non-zero Mean Stresses In cases when the temporal average of an alternating stress is not zero, the fatigue life is reduced compared to the same amplitude of alternating stress with a zero mean. When data are collected from material tests in which the mean stress is not zero, the data are plotted on a two dimensional plane with level curves of fatigue life. The plane of interest shows the average, or steady stress on the horizontal axis and the alternating stress amplitude on the vertical axis. There are several methods of data fitting a curve of constant cyclic life. The most common is called the Modified Goodman Line. Figure 1: Modified Goodman Diagram 1.1.3 Modified Goodman Diagram The Modified Goodman Diagram has two lines plotted on the stress plane. The lines delimit the state of stress below which the material will not fail (assuming a certain probabilistic certainty). The Goodman line plotted on the x-y plane of mean stress and alternating stress is described by the equation: m u a e 1 [1] Understanding material response to alternating stress with nonzero mean is required when designing mechanical components that experience vibration. The next section deals with vibratory response of turbine blades. 1.2 Vibratory Response Turbine blades have curved surfaces used to redirect the flow of fluid in order to extract work from it. Jet engines use turbine blades, positioned behind turbine vanes, to extract work from the flow of very hot combustion products of fuel and air. The turbine vanes are static and act as nozzles to direct the flow toward the blades at the proper direction and flow rate. When the blades see the flow, its intensity fluctuates as is goes between nozzles and wakes. In general, the frequency of excitation is related to the rotational speed of the turbine. The excitation frequency is also related to multiplicative factors related to number of disturbances in the airflow around the turbine. For example 2 such factors can be: number of upstream vanes (nozzles) adjacent to the blade row, number of downstream vanes adjacent to the blade row, difference between number of upstream and downstream vanes, number of fuel nozzles in the combustor, and many other geometric features. 1.2.1 Design Philosophy In the design of a turbine blade, the design engineer tries to minimize the number of times when the blade will experience an excitation frequency equal to one of its natural frequencies at a given running condition. One can never prevent all resonant excitations but can try to place them at engine operating conditions that are not used for long periods of time. The major operating conditions to at which the design engineer must avoid resonance are idle, take-off, climb, and cruise. Those operating conditions may all be associated with different engine operating speeds, each of which creates the potential to excite a resonant mode of the turbine blades. 1.2.2 Frequency Prediction The typical method of predicting natural frequency of a given blade design is to use Finite Element Analysis. The natural frequencies and mode shapes (eigenvalues and eigenvectors) can be numerically approximated and used in the design iterations to prevent resonant excitations. In this type of finite element analysis, the equation to be solved is that of 3-dimensional vibration. It is multidimensional extension of the one dimensional vibration equation. The equation is of the form: Ý kx 0 mxÝ [2] Where in the case of a single degree of freedom system, m is mass, k is stiffness, and x is scalar displacement. In the case of multiple degrees of freedom, x is a vector of displacements while k is a matrix of stiffness (between every mass). 1.2.3 Campbell Diagram Using modal analyses at several operating conditions (with different temperatures and rotation speeds) the engineer can produce a Campbell Diagram, Figure 2. The Campbell diagram simply plots natural frequencies versus engine operating speed. The horizontal axis shows the engine speed (in Rev/min). The vertical axis shows the modal 3 frequencies (in Hertz). The Campbell Diagram shown uses fabricated data for illustrative purposes. The horizontal, drooping lines are the natural frequencies of the blade. The straight lines of various integer slopes are the excitation frequencies caused at certain engine speeds. The slope of the line is called the Engine Order. This is an integer factor equal to a rotational symmetry or repetition found somewhere in the path of the turbine blades such as the number of nozzles before the blade or the number fuel nozzles in the combustion chamber. Any repeated feature that the blade will see in its travel around the engine represents a potential for resonant excitation. If an engine order line crosses a natural frequency line at an operating speed it is said to be a “resonant crossing.” This is a visualization tool that allows the engineer to easily see what modes of vibration will be excited at which engine speeds. The following diagram shows an example of a Campbell Diagram with 7 natural frequencies and 15 engine orders of concern. Figure 2: Campbell Diagram 4 Again, the horizontal axis is the engine rotational speed (in RPM) and the vertical axis is the vibration frequency in Hertz. The conversion from RPM to Hertz is 1/60 (revolutions per minute to cycles per second). The natural frequency lines are not perfectly horizontal. Usually the lines are not perfectly horizontal because natural frequency will vary depending on engine speed and temperature. Although engine speed and temperature are not directly related, they are often highly correlated. For the purposes of the Campbell Diagram, we assume a constant known temperature for each operating condition. For turbine blades, the natural frequency lines usually droop with higher engine speed because thermal softening effects overtake stress-stiffening effects. For fan blades, the lines of natural frequency typically increase slightly with respect to engine operating speeds because temperature increase is small and stress stiffening effects overtake. 1.3 The Problem The difficulty is that when a resonant crossing is predicted, it is up to the intuition of the engineer to know what geometric properties of the blade to change in order to affect the natural frequency desirably. As an added complication, changing one natural frequency desirably may adversely affect another natural frequency. Without a comprehensive analysis of the entire design space, one can never fully understand the practical limitations of tuning turbine blade airfoils. 1.3.1 Defining a Design Space The question posed in this study is: can one accurately and quantitatively characterize the effect each geometric parameter, or combination thereof, has on natural frequencies, or combinations thereof? In order to do this, one first needs to define the design space. The design space is comprised of defined, controllable parameters that affect the shape of the turbine blade. Some simple examples are: height, thickness, aspect ratio, etc. Every design feature in the turbine blade can have a numerical parameter associated with it. In order to fully understand the design space, the engineer must devise a way to test every region of the design space equally. 5 For a simple two-variable design space, assuming there are absolute maxima and minima constraining each variable, the design space is rectangular and has four corners. For higher dimensional design spaces, it is not immediately obvious how to explore the boundaries and interior regions of the space. 1.4 Previous Work on this Topic In a paper published by J. M. Brown and R. Grandhi, a similar study was performed on fan blade airfoils (Brown & Grandhi, 2008). In this study, a population of fan blades was measured using a coordinate measuring machine (CMM). The machine measures the 3D Cartesian position of a point on the surface of the object given an approach orientation. The machine can repeat this measurement for many different points around the airfoil. The data collected was then made to have a zero-mean by subtracting the mean value from each variable. The interpretation of this zero-mean data is the “deviation” from an average airfoil. Zero represents a point being equal to the average position and positive or negative represents deviation from the average. The variations to be measured were caused by random manufacturing variation. To study the effect of this variation on the natural frequency of the airfoils, a large number of realistic sets of deviation variables were to be generated. Many of the deviation measurements of the airfoils would be highly covariant. This is because the airfoil, while deviating from an average population, still remains smooth. Points adjacent to one another on the airfoil surface had high covariance. The authors of this paper projected the measured variable space of high covariance into an orthogonal variable space by means of principal component analysis. This is a statistical technique that determines orthogonal linear combinations of variables that most highly explain the variance in the data not explained by precedent variables combinations. The technique involves simply finding the eigenvectors of the covariance matrix of the dataset. The nth eigenvector projects the old variable space into the nth new variable. A matrix whose rows are the eigenvectors of the covariance matrix forms the transformation matrix that transforms the old, highly covariant variable space into a new set of independent (orthogonal) variables. In many cases, the majority of the variation in the data is explained using a small number of orthogonal variables. The measure data set may have 6 thousands of dimensions but the majority of the variance can be explained by a much smaller number of dimensions or variables. This is referred to as “reduced order modeling.” Using this technique, Brown and Grandhi were able to randomly create realistic combinations of variables that represented plausible airfoils. In this case, plausible means that the deviations were random but the randomly generated airfoils were still as smooth as the measured ones. These randomly selected deviations representing realistic airfoils were then input into a low fidelity finite element analysis to determine the perturbation of the natural frequency of the airfoil. The result of the study was that the natural frequency of the airfoils was significantly affected by manufacturing variation. Brown and Grandhi’s paper illustrates that it is possible to characterize manufacturing variation and to determine its effect on responses such as natural frequency. 7 2. Theory and Methodology 2.1 Monte Carlo Simulation In order to explore the design space affecting the modal response of a turbine blade, Monte Carlo simulation will be used in this study. In this simulation, many geometric parameters will be varied randomly to see their independent effect on the desired response – frequency in this case. Required for this type of analysis is a 3D solid model of a turbine blade using a certain parameter scheme. A scheme of parameters controlling the shape of a turbine blade model is not unique. The size and shape of its features could be defined in many different ways. 2.1.1 Geometric Parameter Scheme For this analysis, the turbine blade will be constructed between two fixed points in space representing the inner and outer flow path surfaces. An airfoil will be defined between these two points using three cross-section curves as seen in Figure 5. There will be a section at the inner radius, outer radius, and half way in between. Each airfoil cross-section curve will be defined by its leading-edge and trailing-edge points. Other parameters defining the airfoil will be its maximum thickness at the middle, section curvature, leading edge diameter, trailing edge diameter, axial chord length, true chord length, and several other parameters fully defining the airfoil section. Since there will be three airfoil sections, spline surfaces used to connect the sections into a solid airfoil will be second-degree (quadratic) in the vertical direction. Using more sections could give extra flexibility to the airfoil but can also lead to reversals in the airfoil shape. Using three sections allows for a maximum of one reversal over the whole airfoil. A reversal is when one part of the airfoil reverses direction on its way up the airfoil. A five-section airfoil could reverse direction four times. Typically, the interpolation spline degree is limited to 3 making it a natural cubic interpolating spline. The following figures show the turbine blade model that will be used. The appendix also contains additional information about how the turbine blade geometry can be morphed using randomly selected values for parameters. 8 Figure 3: Parametric Model of Fictitious Turbine Blade 9 2.1.2 Blade Root Geometry The turbine blade will have a root at the bottom and a tip-shroud at the top. The root will be defined by several parameters, not all of which will need to be varied in this analysis. The main effect that the root will have on the mode frequencies will be due to its mass. Its stiffness will not cause very much variation in the frequency. The airfoil stiffness will be a significant driver of frequency variation. 2.1.3 Tip Shrouds The tip of the airfoil will be attached to a tip-shroud (see Figure 4 and Figure 6). This is a design feature typically used to reduce endwall losses in a turbine. Airfoils without shrouds (unshrouded airfoils) exhibit differential motion between the outer gaspath surface and the airfoil tip. A shroud is like an outer gas-path that moves with the airfoil because it is attached. There is no differential motion between the airfoil and the endwall so the aerodynamic losses are reduced. The shroud can also be used as a vibratory friction damper. Each shroud can be made to interlock with adjacent shrouds and cause frictional damping. While I will include this design feature in the model, I will not be analyzing the variation in damping effectiveness because it is beyond the scope of this analysis. In the vibratory analysis, I will not model any displacement boundary condition at the shroud interface. This vibratory analysis will be simulating the result of a single blade “ping” test with the blade constrained in a root fixture. Figure 4: Generic Tip Shroud Design 10 2.1.4 Generating Random Variable Combinations In order to generate random sets of parameters for the solid models, matlab will be used to create a latin hypercube design space. Matlab implements this using the function lhsdesign(N,P). N is the number of samples and P is the number of variables. The output of the function is uniformly distributed from 0 to 1. Each variable is uncorrelated to every other variable. While uniformly distributed data may not be as realistic as normally distributed data, uniform sampling tests all regions of the design space equally. Such is the goal of Latin Hypercube sampling. I will define a maximum and minimum value for each parameter I want to vary. Then I will transform the random numbers to be between the minimum and maximum for each geometric variable that I have defined using a linear equation. The combinations of geometric parameters will be used to create unique 3-D models of turbine blades representing that point in the design space. The parametric model must be rigorously tested for robustness given highly variable input parameters. In the case that a certain variable is causing a high instance of model failure, the parameter range is reduced. The process of updating the parametric computer model can be done automatically with macros in Unigraphics NX6. Figure 5: Airfoil Section Definition Many of the parameters in the model control airfoil section properties. Some also control non airfoil features such as tip shroud shape and size, root thickness, airfoil 11 length, and so on. The parameters used in the solid model are shown in Table 1 along with their mean, minimum, and maximum values. Mean 0 0 0 0 0 0 1.4 1.4 1.15 0.5 0.5 0.5 10 16 0.25 0.25 0.25 0.07 0.07 0.07 0.02 0.75 0.38 0.25 0.25 0.25 17 1.6 0.08 0.2 0.05 0.025 0.025 0.025 1.85 1.85 1.85 airfoil_x_position_mid airfoil_x_position_root airfoil_x_position_tip airfoil_y_position_mid airfoil_y_position_root airfoil_y_position_tip axial_chord_mid axial_chord_root axial_chord_tip cmax_mid cmax_root cmax_tip flowpath_id_rad flowpath_od_rad hmax_mid hmax_root hmax_tip led_mid led_root led_tip mateface_gap neck_length neck_width peak_chord_percentage_mid peak_chord_percentage_root peak_chord_percentage_tip platform_angle rim_width shroud_bump_height shroud_bump_length shroud_thickness ted_mid ted_root ted_tip true_chord_mid true_chord_root true_chord_tip Min -0.03 -0.03 -0.03 -0.03 -0.03 -0.03 1.3 1.3 1.1 0.4 0.4 0.4 9.5 15.5 0.2 0.2 0.2 0.04 0.04 0.04 0.01 0.65 0.36 0.2 0.2 0.2 14 1.5 0.05 0.15 0.04 0.015 0.015 0.015 1.8 1.8 1.8 Max 0.03 0.03 0.03 0.03 0.03 0.03 1.5 1.5 1.2 0.6 0.6 0.6 10.5 16.5 0.3 0.3 0.3 0.1 0.1 0.1 0.03 0.85 0.4 0.4 0.4 0.4 20 1.7 0.11 0.25 0.06 0.035 0.035 0.035 1.9 1.9 1.9 Table 1: Geometric Parameters and their Ranges 2.1.5 Explanation of Parameter Scheme They are listed here in alphabetical order. The first six are positioning parameters for the airfoil sections. They move the airfoil sections in the X and Y directions independently. There are six of them because there are three sections in the airfoil. 12 The next parameter is airfoil axial chord. This is the distance front-to-back of the airfoil section when viewed from the side. Axial refers to the direction of the engine axis. There are three of these parameters; there is one for each airfoil section. The next set of three parameters is called “cmax” here. It stands for “camber.” What is meant by camber is the curvature of the airfoil section. In the defining section (See Figure 5), the parameter controls the distance from the center of the airfoil max thickness circle to the line tangent to the leading edge and trailing edge arcs. A larger camber distance causes the minimum moment of inertia to increase. The maximum moment of inertia, which is perpendicular to the minimum, is not changed very much. The next two parameters are flowpath inner radius and flowpath outer radius. Airfoil length is the difference between these two parameters. Airfoil radial position is related to the sum of these parameters. The frequencies should depend on airfoil length but should be independent of airfoil radial position. A more intelligent parameter scheme would define airfoil length separately from airfoil radial position. The next three parameters are “hmax.” This refers to airfoil max thickness for each airfoil section. It is defined in the airfoil section as the diameter of the arc tangent to the concave and convex side of the airfoil. Increasing this parameter stiffens the airfoil and adds mass to the airfoil. The next three parameters are “led”. This stands for Leading Edge Diameter. It is the diameter of the leading edge arc for each section. It adds stiffness and mass to the leading edge of the airfoil. The next parameter is “mateface_gap” which refers to the clearance between the platforms of two adjacent blades. Its range is set to be very small (between 0.010” and 0.030”). The next parameter is 2.2 Modal Analysis After that, every blade will undergo modal analysis implemented using the Finite Element approach. I will implement the analysis using ANSYS, a finite element analysis software package widely used in the industry. Modal analysis in its most basic sense solves the homogeneous differential equation: 13 M XÝÝ K X 0 [3] In this equation, X is a vector of the displacements of all the degrees of freedom of the mass system. K is the stiffness matrix that relates the displacements to one another. M is the mass matrix and it is a diagonal matrix that contains all the masses associated with each displacement. In three dimensions, each mass will have three degrees of freedom. Each entry in the displacement vector, X, is one of those displacements. The continuous material of the turbine blade will be discretized into a finite number of elements that have mass and stiffness. Inside each element, the continuous displacement gradient is approximated as a simple, low order function continuous at the element boundaries. Continuity in spatial derivatives of displacement is not necessarily enforced at element boundaries. The finite element software requires certain inputs from the engineer. The first is the geometry of the object to be analyzed. The engineer also inputs the material properties, namely mass density, elastic modulus and Poisson ratio. For this analysis, all the properties will be assumed as homogeneous and isotropic. Homogeneous means that the properties are constant throughout the material. Isotropic means that the non-scalar properties do not depend on orientation. Pertinent to this analysis, the elastic modulus and Poisson ratio will not vary with respect to orientation. Some turbine blades are made of anisotropic materials such as single crystal nickel superalloys. Anisotropic material properties will significantly affect the vibratory response of the turbine blade. In the scope of this analysis, we will assume the blades are made of an “equiaxed,” isotropic material. This type of material has a random crystal structure throughout. The anisotropy of any individual crystals is averaged out over the entire sample of material. The finite element software program automatically generates the stiffness and mass matrices (and damping matrices if required). Although, this analysis will not include any internal or external damping effects. 2.2.1 Execution of Computations The modal analysis was performed using Unigraphics NX6 Advanced Simulation with ANSYS 12.0 as a solver. The model is prepared in NX Simulation. This includes meshing the model into a finite element model, applying displacement fixity conditions 14 at the root, and defining material properties. The material used for this analysis was a simple structural steel. The material does not actually matter since the desired outcome is just the geometric sensitivity. Material variation affects all results uniformly and therefore cancels out when comparing frequencies relative to one another. The finite element modes were meshed using NX Simulation. The meshes contained between 25,000 and 35,000 nodes. When the mesh was submitted to ANSYS for solution, the solution took 30 to 60 seconds to solve the first 5 mode shapes. The 50 parametric model updates and finite element solutions were performed manually due to the quickness of the solution. A more investigative approach could explore more combinations of parameters by using an automation macro. 15 3. Results and Discussion The parameter scheme chosen for the turbine blade proved to be robust. A wide range of parameters produced stable, realistic turbine blade geometries. Out of 50 parameter combinations of 37 parameters, 100% resulted in valid part geometries. This was due, in part, to the limited range of the parameters. The parametric blade model was tested extensively for robustness. Parameter ranges were determined based on the tendency of the model to produce valid geometries. Examples of failure would be: airfoil falling off platform; shroud falling off airfoil tip; airfoil to platform blend failure. 3.1 Mode Shape Identification Under special conditions of symmetry, mode shape identification can be difficult and natural frequencies can be very close to one another. All of the natural frequencies of this turbine blade were spaced out enough that there was no confusion about which mode was associated with which frequency. For the purposes of this analysis it is assumed that the mode shapes, when sorted by frequency, are the same for every blade. 3.2 General Trends in Response 3.2.1 First Mode Shape The first mode of vibration (see Figure 14) is a bending mode. The airfoil is tending to bend about the root section’s minimum moment of inertia. The analysis shows that the parameters that affect root stiffness have a large impact on the first mode frequency. The first mode frequency is also affected by parameters that affect tip mass. Parameters that affect tip section stiffness have very little impact on this mode frequency. 3.2.2 Second Mode Shape The second mode of vibration was also bending about the root (see Figure 15), but in contrast to the first mode, the deflection was in the perpendicular direction. The blade is tending to bend about the root section’s maximum moment of inertia. The frequency is correspondingly higher due to the increased stiffness in that direction. The tip mass affects this frequency similarly to the first mode frequency. 16 The chord length of the airfoil root section also causes increased stiffness in this vibration mode thereby increasing the frequency. 3.2.3 Third Mode Shape The third mode of vibration is a torsional mode (see Figure 16). This means that the vibration is characterized by twisting of the airfoil tip. This frequency is affected by the torsional stiffness of the root and midsection of the airfoil. The frequency is also affected by tip rotational moment of inertia. 3.2.4 Fourth Mode Shape The fourth mode is not easy to describe (see Figure 17). It exhibits motion in the tip shroud and the airfoil trailing edge. The two motions of the shroud and the airfoil trailing edge are opposing one another as if to counterbalance each other’s effect on the structure. This could be considered a second-order bending mode. The frequency is affected by many geometric factors. The linear regression in the following section will reveal the trends more clearly than one can discern from simply looking at the response matrix. 3.2.5 Fifth Mode Shape See Figure 18 for a displacement plot of the fifth mode. The fifth mode shape shows much of its deflection in the corner of one side of the tip shroud. That motion is counterbalanced by motion in the trailing edge and leading edge of the airfoil. This could be considered a third-order bending mode combined with some twisting of the tip shroud. 3.3 Regression Functions 3.3.1 Linear Regression The first step in interpreting the frequency results is to see which parameters most highly affect the frequencies. One approach to this is to create a linear regression. The geometric parameters will be the predictors while the calculated frequencies are the responses. Matlab can generate coefficients for a linear regression for each response and 17 calculate the residual for each data point. The residual is the difference between the actual response and the predicted response using the linear regression at an observation (data point). If the regression is performed using the predictor variables as dimensional quantities, then the coefficients will have units of response/predictor. In this case it would be Hertz/Inch. This means that the response should change by a certain number of Hertz for every inch of variation in the predictor variable. For some of the geometric parameters, one inch is a fairly acceptable variation. For other parameters, one inch of variation is inconceivable and would cause the model to fail. To account for this, the predictors can be transformed to a new space depending on their acceptable ranges of variation. In this new space, zero would represent the minimum value for the parameter and one would represent the maximum value for the parameter. This is referred to as non-dimensionalizing. The linear influence coefficient resulting from the dimensional and non-dimensional predictor variables is quite different. 3.3.1.1 Pareto Analysis The regression coefficients represent the strength of influence each variable has on the response. To create the Pareto Chart, sort the magnitude of all the influence coefficients and plot the values on a bar chart. The typical result is that a small number of the parameters will have a large percentage of the influence on the response. 3.3.2 Nonlinear Regression The responses may be linear over a small range of predictor variables, but the approximation may break down over a larger domain of the variables. The next step after linear regression is nonlinear regression. Instead of just regressing the responses with the individual variables, we also use combinations of two or more variables. FIPER can do this automatically. One can choose to use combinations of different variables and combinations of the same variable (the variable squared). The result of this regression can also be plotted on a Pareto Chart to visually show the influence of variables and combinations thereof. 18 3.3.3 Partial Least Squares Regression Another regression technique particularly well suited to this situation is called partial least squares regression. Herman Wold and his son Svante Wold developed PLS regression in the early 1980’s. It has also been called “Projection to Latent Structures” but the name Partial Least Squares has been more often used. It finds linear combinations of X that most highly explain variability in linear combinations of Y. See the following equation …. 19 4. Conclusion I conclude that this work has proven that… It can be used in the future to predict the frequency response of a design space. 20 5. References Abdi, H. (2003). Partial Least Squares (PLS) Regression. In M. B. Lewis-Beck, Encyclopedia of Social Sciences Research Methods. Thousand Oaks, CA: Sage Publishing. Brown, J. M., & Grandhi, R. V. (2008). Reduced-Order Model Development for Airfoil Forced Response. International Journal of Rotating Machinery , 2008, 1-12. Hastie, T., Tibshirani, R., & Friedman, J. (2008). The Elements of Statistical Learning (2nd Edition ed.). Stanford, CA, USA: Springer. Norton, R. L. (2006). 6.11 - Designing for Fluctuating Uniaxial Stresses. In R. L. Norton, Machine Design, an Integrated Approach 3rd Edition (pp. 356-360). Upper Saddle River, NJ, USA: Prentice Hall. Petrov, E. P. (2008). A Sensitivity-Based Method for Direct Stochastic Analysis of Nonlinear Forced Response for Bladed Disks with Friction Interfaces. Journal of Engineering for Gas Turbines and Power , 130 / 022503-1. Richardson, M. H. (1997). Is It a Mode Shape or an Operating Deflection Shape? Sound & Vibration Magazine (30th Anniversery). 21 6. Appendices 6.1 Appendix 1: Turbine Blade Parametric Model Figure 6: Top View of Fictitious Parametric Turbine Blade Model 22 Figure 7: View of Fictitious Turbine Blade Root Attachment Geometry Figure 8: Close-up View of Turbine Airfoil Section Definition 23 Figure 9: Random Turbine Blade – 1 24 Figure 10: Random Turbine Blade – 2 25 Figure 11: Random Turbine Blade – 3 26 Figure 12: Random Turbine Blade – 4 27 Figure 13: Random Turbine Blade – 5 28 Figure 14: Mode Shape 1 at 299 Hz Figure 15: Mode Shape 2 at 712 Hz 29 Figure 16: Mode Shape 3 at 1267 Hz Figure 17: Mode Shape 4 at 1694 Hz 30 Figure 18: Mode Shape 5 at 2905 Hz 31