Evolutionary design automation for control systems with practical constraints
The aim of this work is to explore the potential and to enhance the capability of evolutionary computation in the development of novel and advanced methodologies that enable control system structural optimisation and design automation for practical applications. Current design and optimisation methods adopted in control systems engineering are in essence based upon conventional numerical techniques that require derivative information of performance indices. These techniques lack robustness in solving practical engineering problems, which are often of a multi-dimensional, multi-modal nature. Using those techniques can often achieve neither global nor structural optimisation. In contrast, evolutionary mechanism learning tools have the ability to search in a multi-dimensional, multi-modal space, but they can not approach a local optimum as a conventional calculus-based method. The first objective of this research is to develop a reliable and effective evolutionary algorithm for engineering applications. In this thesis, a globally optimal evolutionary methodology and environment for control system structuring and design automation is developed, which requires no design indices to be differentiable. This is based on the development of a hybridised GA search engine, whose local tuning is tremendously enhanced by the incorporation of Hill-Climbing (HC), Simulated Annealing (SA) and Simplex techniques to improve the performance in search and design. A Lamarckian inheritance technique is also developed to improve crossover and mutation operations in GAs. Benchmark tests have shown that the enhanced hybrid GA is accurate, and reliable. Based on this search engine and optimisation core, a linear and nonlinear control system design automation suite is developed in a Java based platform-independent format, which can be readily available for design and design collaboration over corporate Intranets and the Internet. Since it has also made cost function unnecessary to be differentiable, hybridised indices combining time and frequency domain measurement and accommodating practical constraints can now be incorporated in the design. Such type of novel indices are proposed in the thesis and incorporated in the design suite. The Proportional plus Integral plus Derivative (PID) controller is very popular in real world control applications. The development of new PID tuning rules remains an area of active research. Many researchers, such as Åström and Hägglund, Ho, Zhuang and Atherton, have suggested many methods. However, their methods still suffer from poor load disturbance rejection, poor stability or shutting of the derivative control etc. In this thesis, Systematic and batch optimisation of PID controllers to meet practical requirements is achieved using the developed design automation suite. A novel cost function is designed to take disturbance rejection, stability in terms of gain and phase margins and other specifications into account in-the same time. Comparisons made with Ho's method confirm that the derivative action can play an important role to improve load disturbance rejection yet maintaining the same stability margins. Comparisons made with Åström’s method confirm that the results from this thesis are superior not only in load disturbance rejection but also in terms of stability margins. Further robustness issues are addressed by extending the PID structure to a free form transfer function. This is realised by achieving design automation. Quantitative Feedback Theory (QFT), method offers a direct frequency-domain design technique for uncertain plants, which can deal non-conservatively with different types of uncertainty models and specifications. QFT design problems are often multi-modal and multi-dimensional, where loop shaping is .the most challenging part. Global solutions can hardly be obtained using analytical and convex or linear programming techniques. In addition, these types of conventional methods often impose unrealistic or unpractical assumptions and often lead to very conservative designs. In this thesis, GA-based automatic loop shaping for QFT controllers suggested by the Research Group is being furthered. A new index is developed for the design which can describe stability, load rejection and reduction of high frequency gains, which has not been achieved with existing methods. The corresponding prefilter can also be systematically designed if tracking is one of the specifications. The results from the evolutionary computing based design automation suite show that the evolutionary technique is much better than numerical methods and manual designs, i.e., 'high frequency gain' and controller order have been significantly reduced. Time domain simulations show that the designed QFT controller combined with the corresponding prefilter performs more satisfactorily.