Proceedings of 2015 International Conference on Quality, Reliability, Risk, Maintenance, and Safety Engineering (QR2MSE 2015) July 21-24, 2015, Beijing, China QR2MSE2015-DRAFT Put Paper Title Here (e.g.: Sequential Particle Swarm Optimization and Reliability Assessment of Planar-Type Voice Coil Motor) Given Name Surname1, Hong-Zhong Huang2,*, … 1. Department of Industrial Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan, 611731, P.R. China 2. Institute of Reliability Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan, 611731, P.R. China ABSTRACT Put abstract text here. Planar-type voice coil motor (VCM) is a key component in ultra-precision motion of fine stage of lithography machine. The reliability-based design optimization method given in this work provides a novel criterion to ensure performance of Lorentz motors by evaluating the reliability of force constant. To solve the reliability based design optimization (RBDO) problem in discrete space with the speed of decoupled loop in sequential optimization and reliability assessment (SORA) for global solution, a Sequential Particle Swarm Optimization and Reliability Assessment method is proposed. The reliability boundary shift is put into penalty function for constrained optimization in fitness evaluation of particle swarm optimization (PSO). The presented optimization design model, the geometric parameters of the studied Planar-type VCM in finite element model are treated as design variables whereas the thrust force constant is an output quantity of interests. By using electromagnetic analysis, the desired requirements of Lorentz motors are verified. here. Put body of the paper here. Put body of the paper here. Put body of the paper here. 1 INTRODUCTION The high precision position of wafer stage should not only be measureable, but also be controllable, which is achieved by motors, bearings and feedback loops [1]. A two stage strategy, composed of dual-servo stage including the coarse stage and the fine stage, was promoted to satisfy the requirement of the wafer stage, the key subsystem of lithography machine [2]. It needs to move a long distance (about hundreds millimeters) with ultra-precision positioning, which is driven by the linear motor for the coarse stage and the planar-type voice coil motor for the fine stage [3]. The voice coil motor, as one type of linear Lorentzforce actuator usually used in loudspeaker [4], DVD drives, is developed to be applied in ultra-precision motion stages as a mature technology, for that the VCM is useful in linear precision positioning control systems with little range by exciting its voice coil with a controlled current [5]. A voice coil motor, which includes rotary-type and planar-type, consists of coil, permanent magnetics and iron yokes. The rotary-type VCM is optimized to increase the force generation by Halbach Magnetic Circuit, which could confine the magnetic flux in order to decrease the thickness of the yoke and to increase the force constant [6, 7]. A type of VCM for digital cameras is designed and optimized for maximum force constant by using electromagnetic simulation software of Ansoft Maxwell [5]. The geometric parameters of VCM have been optimized by modified MEC model to enhance the dynamic response [3]. Halbach Magnetic Circuit design, KEYWORDS: reliability based design optimization; particle swarm optimization; sequential optimization and reliability assessment; Planar-type voice coil motor NOMENCLATURE Put nomenclature here. Put body of the paper here. Put body of the paper here. Put body of the paper here. Put body of the paper here. Put body of the paper here. Put body of the paper * Corresponding Author: hzhuang@uestc.edu.cn Tel: +86-28-6181252; Fax: +86-28-61830227. 1 simulation technology, and geometric parameters are three components in the preparation for the optimal design of VCM. The reliability based design optimization for high accuracy stages may takes novel criteria into integrated circuit manufacture industry, especially for whose failure mode is different with the traditional machinery. It is essential to minimize the height of the planar-type VCM for lightweight design while the force constant needs to be guaranteed by reliability constraints. Reliability analysis is divided into analytical methods [8-14] and sampling-based methods [15-19]. Analytical methods include first order reliability analysis (FORM) [9-13], second order reliability analysis (SORM) [14] and so on. The computational cost of sampling-based method is a limit factor, such as Kriging and importance sampling[15], Monte Carlo [18], and so on. The PSO was extended for reliability based design optimization with a move to the boundary of reliability constraints [20]. Furthermore, the features of auto-tuning and boundary approaching are added in PSO algorithm to resolve the RBDO problem with the help of subset simulation [21]. And the position updating may be according to its own best experience at early iterations, whereas according to the best experience of whole swarm in late iterations [22]. Also, the evolutionary algorithm with the safety-factor method, which is derived from the Karush–Kuhn–Tucker optimality conditions, eliminates the need for reliability analysis, however, only the design variables can be treated as random variables [23]. The boundary-shifting is used to increase the convergence rate and particle-position-resetting is used to enhance the diversity of particles to keep away from local solution [24]. Here, the boundary shift means to a move towards a boundary of a constraint function [24], whereas the boundary shift in SORA is based on reliability analysis[25], so the boundary shift in SORA is called reliability boundary shift in order to distinguish. This paper is organized as follows: In Section 2, penalty for particle swarm optimization is introduced to the constrained optimization in discrete solution; In Section 3, the proposed method for reliability based design optimization is Sequential Particle Swarm Optimization and Reliability Assessment Method; In Section 4, the Planar-type VCM is designed and optimized with reliability constraints; In Section 5, the results are examined by perturbation plots and finite element method; In Section 6, there is a conclusion for the reliability based design optimization for Planar-type VCM and future work. searching behavior of birds flocking [26]. In standard PSO algorithm, a swarm of particles, on behalf of candidate solutions, are initialized and move through the search space towards the compromise between the best position historically and the best among all particles until converging to the optimum [27]. PSO possesses two apparent advantages over the genetic algorithms: 1) less parameters; 2) fast constringency [28]. The new feature of auto-tuning boundary-approach is applied to PSO in order to take a move toward the boundary of probabilistic constraints with non-smooth performance functions of RBDO [29]. The method is proved to be also useful in cam design problems such as minimizing the magnitude of the negative acceleration and the cycle time respectively [30]. Firstly, the initialization of a population of particles with random positions and velocities is limited to the problem space and the maximum and minimum of the velocities. Secondly, when the fitness of each particle is evaluated, the assessment of the fitness is compared both for the best potion of current than the old history evaluation and the best position for the entire population concluding positions of all the time. Then, the position and the velocity will be updated based on three categories of parameters: two random parameters for random steps; two constant numbers for the balance between the best position of the history and the best position of all particles; one parameter for the influence form last position [31]. (1) Updating particles The position of particles is updated by adding a change velocity: (1) xi t 1 xi t vi t 1 (2) Updating Velocity Each particle updated the velocity based on the last velocity and the best position of history and the whole particle swarm: vi t 1 w vi t c1r1 x pbest xi t (2) c2 r2 x gbest xi t where c1 and c2 are acceleration constants, usually equal to 2; r1 and r2 are random numbers uniformly distributed in the range of (0,1); w is an inertia weight to control the influence of the previous velocity, which could be an constant or updated by w(t 1) w(t ) eta , while eta is a constant; x pbest represents the best position in history of the particle, whereas x gbest is the best position in global candidate solutions. 2 PENALTY OPTIMIZATION FOR PARTICLE SWARM 2.2 PSO with Penalty Function for Constrained Optimization In the process of fitness valuation, the penalty function method could be used to consider the constraints for the particle swarm optimization. A typical penalty function is presented for example, while other penalty 2.1 Particle Swarm Optimization Particle swarm optimization, which was an evolutionary computation algorithm proposed by Kennedy, is, in fact, an optimizer, based on the food2 function could also be implanted in the assessment in the fitness to adapt PSO to the constrained optimization. There are two types of the penalty functions, the interior and exterior penalty functions. Because the exterior function is not required to start in feasible solution, it is more applied in evolutionary algorithm than the interior penalty function [32]. The fitness is the sum of the cost function or the objective function and the penalty item x fit ( x) cos t ( x) x (3) x = i (4) approach (PMA). Whether in RIA or in PMA, reliability analysis with First Order Reliability Method (FORM) is transformed into an optimization problem with equality constraint as limit state function, which is defined in Uspace [33-35]. In RIA, the MPP is u*G U 0 , calculated by: minimize U subject to g U 0 In PMA, the MPP is u* t calculated by: minimize g U subject to U t i The is dynamic dependent on iteration such as the i (5) j Where qi max(0, gi ) , i j indicates the penalty coefficient corresponding to i th constraint and j th violation level. i could be equal to 2. 3 SEQUENTIAL PARTICLE SWARM OPTIMIZATION AND RELIABILITY ASSESSMENT METHOD Reliability-based design optimization is to offer a reliable solution considering uncertainty related to variables, parameters, and models. Design variables may be discrete due to product standards, whereas the distributions of the variables may be continuous or discrete. For example, Diameter of copper coils could be chosen by several discrete choices, but it is actually subjected to a continuous distribution. 3.1 Reliability-Based Design Optimization The general RBDO method can be expressed Design Variable DV d, μX Minimize f d, X, P S.t. Pr gi d, X, P 0 R i (8) PMA depends much less on the nonlinear transformation than RIA, and is adapted to the variety of distributions, whereas the RIA may fails to bound and extreme type distributions [34]. The coupling of optimization and the reliability analysis is classified into three ways: (1)double-loop method [36]; (2) decoupled method [25]; (3) single-loop method [37, 38]. Based on benchmark study of numerical methods for RBDO, with the lower efficiency and more robustness and accuracy than the single-loop method, SORA could be suitable for complex structural systems [39]. iteration or be static value, while could be expressed as: i i qi 1 (7) 3.2 Sequential Particle Swarm Optimization and Reliability Assessment Method The proposed method, decouples optimization process and reliability analysis by SORA strategy, implements performance measure approach for MPPbased reliability analysis, and takes penalty function method to transform constrained optimization into unconstrained optimization problems. In Figure 1, the flowchart of the proposed method is illustrated with the main processes. In the first cycle of SORA, deterministic optimization is in the first place, whereas, the MPP-based reliability analysis is applied for all particles to find the MPPs. Then, in order to distinguish from the boundary shift in the augmented particle swarm optimization, the boundary shift from SORA may be regarded as reliability boundary shift. Based on the position of MPPs, the reliability boundary shift is generated in order to plug the results of reliability analysis into reliability constraint [25], which is formulated as (9) g μx s 0 (6) μ LX μ X μ UX , d R NDV X R NRV i =1,...,NC where f is the objective function. μ X is the mean value of the random design vector X; d is the vector of the deterministic design variables; P is the vector of the random design parameters. The probabilistic constraint is Pr gi d, X, P 0 , i=1,...,NC ; R i is the reliability Where, si μ X ,i x MPP,i for i th constraint. And the penalty functions are used to connect the constraint and the evaluation of the fitness by: fit μ x f μ x gi μ x si (10) target for the i th constraint; NDV is the number of the deterministic design variables; NRV is the number of the random design variables; NC represents the number of the probabilistic constraints [25]. In RBDO, the assessment of probabilistic constraints could be solved in two different methods: the reliability index approach (RIA), or the performance measure In particular, the fitness for sequential particle swarm optimization and reliability assessment method is: fit μ x f μ x iteration g μ i i 3 i x si max 0, gi μ x si 1 i (11) The shift vector contains the reliability information to switch the boundary in the constrained optimization and participating in determine the penalty levels for the distance from the boundary. In this way, it is obvious that the constrained optimization is turn out to be unconstrained optimization with a plug in fitness evaluation, and that the uncertainty is considered by shift vector, a transfer from reliability analysis to the position and velocity updating by fitness evaluation. The evaluation method not only could be used in particle swarm optimization, but also could be used in other evolutionary algorithm. The advantage is to fit the decoupled strategy into the evolutionary algorithm for reliability design optimization by adding the shift vector in the penalty function. In this way, other evolutionary algorithm, such as genetic algorithm and Simulated Annealing, with their own advantage, could be taken into SORA strategy. In the proposed method the advantage of PSO and SORA are included, since which it owns the ability to solve engineering problems with the effectiveness from decoupled strategy in discrete solution space for global search. planar voice coil motor. There are three main characters in this planar voice coil motor: 1) double side magnet circuit; 2) Halbach magnet circuit; 3) moving magnet type 4) planar voice coil motor. The double side magnet circuits reduce the force disturbances in z-axis, because the magnetic flux is in a closed loop. Based on Flemming’s rule, on account of decreased variation of the direction between the flux density and current in the moving magnet type, the fluctuation of force constant is less than that in the moving coil type. The conventional magnetic circuit merely consists of vertically-magnetized magnets, whereas, the Halbach magnetic circuit is made up of a horizontally-magnetized magnet and two verticallymagnetized magnets. Halbach magnet circuit design could offer augmentation of magnetic flux, decrease the thickness of the yoke, and then increase the flux density in the air gap. The Figure 2 describes the design prototype for the planar voice coil motor. The arrows on magnet in Figure 2 indicate magnetization directions of each magnet. The arrows on coils indicate the current direction. YOKE . MAGNET . COIL . Initialization CURRENT . Generate Particles . Updating Reliability Boundary Shift vectors Updating Velocity Fitness Value With Penalty Function 4.2 Response Surface Method Response surface approximations of the limit state functions are used to reduce the calculation expense [40]. The thickness of coil, magnet and yoke are design variables, and the force constant is the response. X1, X2, X3 represent the thickness of coil, magnet and yoke respectively, whereas the k f represents the force constant. The procedure is simple and the choose of the model is based on P-value, as expressed in Table 1. If the P-values of "Prob > F" is less than 0.005, it indicate that the model terms are significant and could be kept. However, the P-value of X2^2 is 0.5385, so it is canceled in the model 2, which improve the F-value in the same time. The P-value of the model is less than 0.0001, which means it is significant. As the F-Value of model 2 is more than that of model 1, the model 2 could be used as the response surface of the force constant. The normal probability plot is implied to detect whether the residuals is subjected to the normal distribution by a graphical technique, with the assumption that every variable is independent with each other from the Figure 3, it is obviously that the points make up an approximate straight line, which indicates the date set of the residuals follow a normal distribution. updating pbest and gbest Not Satisfied YOKE Figure 2 Designed Model of Planar-Type Voice Coil Motor MPP-based Reliability Analysis(MPP) Updating Particles MAGNET Stopping Criteria? satisfied Output Figure 1 Flowchart of The Proposed Method 4 DESIGN AND OPTIMIZATION FOR PLANARTYPE VOICE COIL MOTOR 4.1 Halbach Magnetic Circuit for Planar-type VCM The Planar-type VCM is a moving magnet type with double side Halbach magnet circuit, also may be called 4 It is useful to test the actual response values versus the predicted response values by the intuitive graph, in order to easily provide whether a group of values which are not predicted perfect with the model. The Figure 4 shows the points are split by a line of 45 degree, which presents the model is significant. gap is, the larger the force constant is. And on behalf of the request of cooling system in practice, the minimum value of the Gap is 2mm. In this project, K tar f is constant of 9, and F tar is constant of 30, based on the need of force and the thermal control. To simplify the problem and from the experience in manufactory, it is reasonable to set the variables with a standard variance of 0.1, and be subjected to normal distribution. The discreteness of the thickness of the coil is provided by the supplier, but X1 still is a normal distributed factor in fact. In section 3, the model of the force constant decided and calculated. The PMA method is used for MPP search with t =2 . From section 4.2, the model 2 for force constant could be expressed as: K f =2.16912 0.43450* X1 0.44631*X2 3.60926* X3 0.028311* X1* X2 0.083270* X1* X3 0.014550* X2* X3 0.039223* X12 0.54656* X32 Table 1 Response Surface of Force Constant Kf = Model1 P-value Model2 P-value C * X1 * X2 * X3 * X1 * X2 * X1 * X3 * X2 * X3 * X1^2 * X2^2 * X3^2 F-Value 1.486511 -0.4345 0.632471 3.609263 < 0.0001 < 0.0001 < 0.0001 2.169116 -0.4345 0.446306 3.609263 < 0.0001 < 0.0001 < 0.0001 -0.02831 0.0077 -0.02831 0.0077 -0.08327 < 0.0001 -0.08327 < 0.0001 0.01455 0.039223 -0.01241 -0.54656 576.2221 0.3676 < 0.0001 0.5385 < 0.0001 0.01455 0.039223 None -0.54656 649.65 0.3670 < 0.0001 None < 0.0001 (13) The result of the proposed method for the problem is in Table 2. Table 2 The Result of The Proposed Method Method μ X1 μX2 μ X3 12.00 11.3632 11.00 10.00 objection R1 R2 Predicted Response 5.3069 9.00 8.00 PSO_DO 3.40000 7.27345 2.76037 27.46764 0.479134 0.789569 PSO_SORA 3.40000 8.00979 2.74596 28.91150 0.99902 0.97953 7.00 R1 and R2 represent the reliability for the constatint1 and constraint 2 respectively. The probability of each constraint shows the differences between the RBDO and the deterministic optimization by Monte Carlo Simulation, which is a classic evaluation method for reliability analysis with accurate probability for sufficient numbers of simulation. The proposed method could guarantee the reliability constraint with relative small change of the objective evaluation, which is acceptable for the practical solution. Different from the traditional reliability design by the experience of safety factors, the reliability based design optimization could definitively find the quantitative solution except for uncertainties of mathematic models. The PSO not only deal with the discrete solution space, but could offer global search with an adjustable speed between the global position and the history position towards the point of convergence. Since the solution is based on response surface method, which may result in uncontrollable errors occasionally, it is necessary to perturbation analysis for the force constant and the finite element analysis for the design regarded as actual response in practice. 6.00 5.00 5.00 6.00 7.00 8.00 9.00 10.00 11.00 12.00 Actual Response Figure 4 Predicted Verse The Actual Response 4.3 Reliability Based Design Optimization of Planar-Type VCM Reliability based Design Optimization of Planar-type VCM is defined as: Design Variable DV μ X1 , μ X 2 , μ X 3 Minimize μ X1 +2*μ X 2 +2*μ X 3 +2*Gap S.t. Pr g1 0 g2 0 R i g1 K f K tar f ; (12) g 2 μ X1 * K f F tar 1 μ X1 6, and μ X1 is discrete with interval of 0.1 6 μ X 2 9, 1 μ X 3 4, There are three design variables μ X1 , μ X 2 , μ X3 , represent the mean value of thickness of coil, magnet and yoke respectively. The Gap means the thickness of the gap in the design, which is important for the force constant. However, it is a deterministic relevance that the less of the 5 5 PERTURBATION ELEMENT ANALYSIS ANALYSIS AND FINITE Force Constant 5.1 Perturbation Analysis for the Force Constant The perturbation plot for the force constant is applied to illustrate the effect of the three factors at the design point in solution space to find the sensitive factor to the force constant based on response surface. The response of factors in Figure 5 is plotted by varying just single factor over all the range while keep the other factors as the constant at the design point in Design Expert. In this way, the steeper the slope of the line is, the more sensitive the factor is to the force constant such as X1, whereas the flatter the slope of the line is, the less sensitive the factor is just as the X3. The standard error of the response in Figure 6 is also plotted to show the uncertainty caused by response error, which is relative low in design point. The influential factors X2 and X3 in the perturbation plot are compared with each other in the contour plot Figure 7 and Figure 8 respectively, with a constant factor X3, which is the same in the deterministic optimization and RBDO. The plots of the perturbation analysis of the force constant show the reliability design, which keep the distance from the deterministic constraints. Figure 7 Response Surface of The Model with One Constant Factor Force Constant Standard Error of Force COnstant 5.2 Finite Element Analysis Finite element analysis is applied to examine whether the results are reasonable in practice. In reality, the design is calculated by Ansoft Maxwell software. The mean values of the random design variables are put into the final design in finite element analysis. The plots of magnetic flux density in coil and in yoke are plotted in the Figure 9 and Figure 10 respectively. Deviation from Reference Point Figure 8 Standard Error of Response with One Constant Factor Standard Error of Force Constant Figure 5 The Perturbation of The Model Figure 9 Magnetic Flux Density In Coil Deviation from Reference Point Figure 6 Perturbation of The Standard Error of Response 6 proposed method is well applied in the process to produce the optimal design subjected to reliability constraints. 6 CONCLUSIONS This paper design and optimize a planar-type voice coil motor considering uncertainty. The reliability boundary shift, added into the penalty function in fitness evaluation, is the key process to take the reliability constraints into the penalty for particle swarm optimization by SORA strategy. The proposed method improves the reliability constrained optimization in discrete space with the speed from SORA strategy and the global solution form particle swarm optimization. The optimized Planar-type voice coil motor is useful in the fine stage and other applications where a large acceleration speed and high accuracy are needed. In the future work, the reliability based topology for the motor will be applied for lightweight while the force constant is guaranteed. Figure 10 Magnetic Flux Density In Yoke The Halbach Magnetic Circuit design restricts more energy in the middle of the Coil instead of out in the space, which is evident in Figure 9 and Figure 10. There are tremendous variances of magnetic flux density in small distance, whether in the yoke or in the coil, expressed as the rapid changes of colors, which present the concentration of magnetic flux density distribution. Figure 9 illustrates the distribution of magnetic flux density in coil, in which there is a potential optimization in future to put more current in central place to generate more power and reduce the weight. Figure 10 demonstrates the function of the yoke to keep more energy in the motor, in which the yoke could be reshaped by topology technique in the future design to put more thickness in the central place. In Table 3, the simulation result of finite element analysis is presented. The force constant satisfies the initial deterministic design with a margin for reliability constraints. The results of RBDO is not much too conservative to afford, instead, the cost is worthy to nearly improve fifty percentages and thirty percentages in reliability constraints in Table 2, respectively. It not only proves the availability of the proposed method, but also the useful tool of the response surface method, which is simple but robust in many practical engineering problems. Even though other directions of the force are not considered in the initial work, which need to be improved in the future work to reduce the forces in the other directions to enhance the performance of motion with high precision, they are relative low and keep little disturbance to the moving direction, enough to be endured or even eliminated by the control system. Also, there may be some errors from the simulation, which may be different with the actual measurements. ACKNOWLEDGMENTS Put acknowledgments here. REFERENCES [1] H. Wang and H. Pham, Reliability and Optimal Maintenance. Berlin: Springer, 2006. [2] C. R. Cassady, W. P. Murdock Jr, and E. A. Pohl, “Selective maintenance for support equipment involving multiple maintenance actions,” European Journal of Operational Research, vol. 129, no. 2, pp. 252-258, 2001. [3] Y. Liu and H. Z. Huang, “Optimal selective maintenance strategy for MSS under imperfect maintenance,” IEEE Transactions on Reliability, vol. 59, no. 2, pp. 356-367, 2010. [4] A. Lisnianski and G. Levitin, Multi-State System Reliability Assessment, Optimization, Application, Singapore: World Scientific, 2003. [5] S. W. Prmon, C. R. Cassady, and A. G. Greenwood, “A simulation based reliability prection model for conceptual design,” In Proceedings of Annual Reliability and Maintainability Symposium, pp. 433436, 2001. [6] D. L. Fugate, “A reliability allocation method for combination serialparallel system,” In Proceedings of Annual Reliability and Maintainability Symposium , pp. 432-435, 1992. [7] H. Z. Huang, J. Qu, and M. J. Zuo, “Geneticalgorithm-based optimal apportionment of reliability and redundancy under multiple objectives,” IIE Transactions, vol. 41, no. 4, pp. 287-298, 2009. [8] R. Braun, Collaborative Optimization: An Architecture for Large-scale Distributed Design, Ph.D. Dissertation, Stanford University, 1996. [9] UKHSE, United Kingdom Health and Safety Executive, available at: http://www.hse.gov.uk/ index.htm. Table 3 Results of Finite Element Analysis Force_axis F_x F_y F_z F_mag Force value 0.082503 9.1172 0.11423 9.1183 The Perturbation Analysis and Finite Element Analysis testify that the optimized design is reasonable and simulation result is acceptable for the reliability based design optimization solution of VCM. In this way, the 7