SOFT COMPUTING METHODS FOR DESIGN APPLICATIONS IN MICROWAVE DOMAIN

Abstract

The growing commercial market of wireless communication systems/devices over the past decade has led to the explosion of interests and opportunities for the design and development of microwave components. The wireless industry em phasizes on the development of components in shortest possible time and at low development cost [1]. There is also a class of critical design applications such as design of high power devices and components where the accuracy requirement is the prime goal. The design of the most microwave components require the use of commercially available electromagnetic (EM) simulation tools for their analy sis. In the design process, the simulations are carried out by varying the design parameters until the desired response is obtained. The optimization of design parameters by manual searching is a cumbersome and time consuming process, and the chances to get local minima are very high. Moreover, increasing number of design parameters or widening the search range makes it difficult to converge to the global optima. Soft computing methods play important role in the design and optimization of many engineering disciplines including microwave domain. The aim of these methods is to tolerate imprecision, uncertainty, and approximation to achieve robust and low cost solution in a small time frame [2]. Soft computing methods such as Genetic Algorithm (GA), Artificial Neural Network (ANN) and Fuzzy Logic (FL) have been widely used by EM researchers for microwave design since last decade [1, 2, 3]. However, these methods suffer from certain drawbacks. GA is a powerful optimization tool, but it requires a large number of iterations to achieve convergence and to arrive at an optimum solution. ANN has proved its efficiency in modeling microwave components, but it has also suffered with generalization problems. Moreover, due to very small wavelengths involved in microwave design, which requires high precision, it is not easy to model compo nents using conventional methods. In all, modeling andoptimization are essential parts and powerful tools for the microwave design task but they must be applied judiciously. Our present research work deals with the development and use of soft comput ing based methods for tackling challenging design problems in microwave domain. Our aim in the development and investigation of these methods is to obtain the designs in small time frame while improving the accuracy of the design for a wide range of applications. In order to achieve this goal, a few diverse design problems of microwave field, representing varied challenges in the design, such as microstrip antennas, microwave filters, a microstrip-via, and also some criti cal high power components such as nonlinear tapers and RF-windows have been considered as case-study design problems. Different design methodologies are developed for these applications. The first chapter of the thesis presents intro duction and motivation behind the work. It also presents the scope of the overall work. In chapter two of the thesis, a state-of-the-art review on the use of soft com puting methods for the design of microwave components is presented. In this chapter, we have described the conceptual background offive importantsoftcom puting methods namely GA [4, 5], Particle Swarm Optimization (PSO) [6, 7], Bacterial Foraging Optimization (BFO) [8], ANN [9, 10] and Support Vector Machine (SVM) [11]. Although each description is followed by the review of microwave designs based on these methods, the emphasis is made on covering design works using recently developed techniques such as PSO, SVM and BFO. Microwave designs obtained using hybridization of these soft computing methods are also discussed. Chapter three presents a modified particle swarm optimizer and demonstrates its applicability for the design of a specific microwave filter. Commercial growth in the use of wireless communication products necessitates the efficient design of components in a smaller and realistic time frame. A Computer-Aided Design (CAD) approach is adopted to minimize the time required to obtain best possi ble design parameters and reduce the experimental iterations as far as possible. Though GA and PSO can be directly used for microwave filter design problems, an inherent limitation of these algorithms is that they require a large number of iterations to converge to an optimum solution. Therefore, the researchers on PSO have also been continuously working for improving the convergence speed and accuracy of the standard PSO [12]. In this work, we present a novel modifi cation to PSO algorithm, PSO with Multiple Subswarms (PSO-MS), which aims to offer faster convergence while improving quality of solution. The solution is generic as it can be applied to the cases when the design process is complex, computationally expensive and time consuming. In the proposed modification, we have introduced a new paradigm of multiple sub-swarms for searching pa rameter space with the PSO algorithm. The social component of PSO's velocity update equation is modified to consider the effects of multiple sub-swarms. Five benchmark functions have been considered for testing the proposed algorithm. The approach is implemented and tested for two basic variations of PSO, namely, PSO with inertia weight [13] and PSO with constriction factor method [14]. The experimental results illustrate that the PSO-MS algorithm has the potential to converge faster, thus reducing the computational expenses, while improving the quality/accuracy of the solution. The PSO-MS is also used for the design of cou pled microstrip-line band-pass filter which is a computationally expensive process when EM tool is invoked in iterative loop of PSO. The results of the proposed algorithm show improvement over the design results obtained using standard PSO. Chapter four deals with support vector driven evolutionary algorithms for the design of specific microwave components. Full-wave EM simulation techniques provide accurate solution. However, the accuracy, computational time and con vergence of the solution using EM simulators are dependent on the number of constraints to be handled. Since, these constraints have to be handled man ually in full-wave EM simulators, they do not guarantee convergence and also require long run time especially when optimization-based design automation is considered. In order to overcome these obstacles, either closed-form expressions or mathematical curve fitting techniques, which use data obtained from mea surements or EM simulators, can be employed to compute the output response. Various meta-modeling techniques such as ANN, response surface method, kriging, etc., can be used to create approximate model from the empirical data [15, 16]. Most of these models have inherent limitations of accuracy and valid ity over a restricted range of parameter values. ANN has been used by many EM researchers to model microwave components and balance the trade-off be tween computation time and accuracy [1]. However, the generalization accuracy achieved bythe ANN basedmodels ofmicrowave components needs improvement to increase the effectiveness of CAD. In this chapter, we have presented a more accurate model of microwave components using SVM. Similar to ANN, SVM is also a learning technique to learn from empirical data to deal with the accuracy and complexity trade-off, by minimizing upper bound on the generalization error [11, 17]. Adetailed description for SVM based microwave modeling is presented and models for specific microwave components such as a one-port microstrip via and two microstrip antennas are developed. The accuracy of the SVM models is compared with other meta-models developed using ANN. Another contribution of the chapter is in presenting a hybrid approach combining SVM with evoluiv tionary algorithms such as PSO/GA. In this approach, the model of microwave component obtained using SVM is invoked in the optimization loop of evolu tionary algorithms. The significant advantage obtained with SVM model is that it responds quickly (approximately in milliseconds) compared to iterative para metric analysis using EM simulation tools for which the response time is large (approximately in minutes depending upon the complexity of the problem). The hybrid method, support vector driven genetic algorithm, is demonstrated for the design of circular polarized microstrip antenna at 2.6 GHz band, while another similar hybrid method, support vector driven particle swarm optimization, is demonstrated for the design of a simple aperture coupled microstrip antenna. Chapter five of the thesis deals with the design of a nonlinear taper using swarm intelligence based algorithms. The design of high power microwave sources and their components belong to a class of problems in which high precision is required with very less tolerance. Vulnerability of one component may cause the failure of the entire system and spell catastrophic damage. Nonlinear taper is one such component which is used in the output system of high power gyrotrons to connect output section of cavity with the main waveguide system. For high power applications, the design of a nonlinear taper requires very high transmission (above 99%), with minimum spurious mode content. In this work, the design optimization of a nonlinear taper to be used in a specific gyrotron (42 GHz, 200 kW, CW gyrotron operating in the TE0,3 cavity mode with axial output collection) has been taken as a case study. The taper synthesis has been carried out considering a raised cosine type of nonlinear taper and the analysis is done using a dedicated scattering matrix code as it is very fast and accurate for taper analysis [18]. The design of nonlinear taper is carried out using two swarm intelligence based algorithms, namely, PSO and a modified BFO. The classical BFO ignores the effects of swarming, and all the bacteria are assumed to have the same swim length. However, varying the swim length according to the fitness may give better convergence speed. In order to improve the convergence and quality ofsolution, the BFO algorithm is modified such that it includes memory of the bacteria, global swarming effect and variable swim length. The modified BFO (MBFO) is tested on a set of benchmark functions. The optimization of a nonlinear taper is carried out using both the PSO and MBFO algorithms, which show very good agreement with the desired objective. The best optimized taper design shows excellent response with very high transmission (99.86%) indicating the effectiveness of these methods. In chapter six, the design of disc-type RF-windows is presented using Multi- Objective Particle Swarm Optimization (MOPSO) method. High power mi crowave and millimeter-wave sources such as gyrotrons, klystrons, and other gyro-devices produce very high output power at wavelengths in microwave and millimeter-wave ranges [19]. RF-window is a critical component in the output system of these devices. The design requirement ofthe window is that it should withstand high power, mechanical and thermal stresses, be leak tight and loss less. Therefore, the challenge is to select a proper window material, and obtain an optimized design that minimizes power reflections and absorption for a better transmission [19]. Hence, these components have to bedesigned carefully. Many real time design problems require optimization ofmore than one objective. In this case, it is desired to find a solution that optimally balances the trade-off between multiple objectives. Multi-objective optimization is implemented with PSO by several researchers[7]. In this work, the design of two types of RF-windows, namely, double disc window [19, 20] and pillbox-type window [21] for use in high power devices is presented using a specific implementation of MOPSO [22], which uses the mechanism of crowding distance and found to be highly compet itive in converging towards Pareto front. The role ofMOPSO is to find physical dimensions for both the types of windows while optimizing the trade-off between matching of desired resonant frequency and maximizing bandwidth around the resonant frequency. The design of double-disk window is carried out for a specific 42 GHz, 200 kW, CW gyrotron (with two different disc materials) and for a 170 GHz, 1 MW, CW gyrotron, while the design of pillbox-type window is carried out for a 2.856 GHz, 5 MW, pulsed klystron. The results show the best resonance matching for each window design and prove the applicability of MOPSO to wide range of high power microwave devices/components. Finally, in chapter seven, the contributions made in the thesis are summarized and scope of the future work is outlined. In summary, the thesis contributes towards improvement of the efficiency and accuracy of the design problems in microwave domain by proposing and investigating soft computing methods, their modifications and hybridizations.