A Hybrid T-S Model Identification Algorithm Based On Harmony Search Algorithm

This work was supported by the National Natural Science foundation project under grant No.60874084. In this thesis, the T-S model identification methods were discussed.During recent years, The T-S model has attracted wide concern for it's an effective mathematical way to build the model of a nonlinear system from the observed data. Generally, the essential parts of building a T-S model are selecting the input variables, determining the membership function type and parameters optimization. After the input variables and membership function type are chosen, the T-S model identification process can be considered as a multi-dimensional nonlinear parameter optimization problem for most T-S model identification methods to solve.It is well known that the regular T-S fuzzy model identification methods, such as the Fuzzy C-Means (FCM) algorithm and least-squares method, usually fail to find the optimal solutions, since they determine the consequent parameters by the least-squares method just based on only one certain group of premise parameters. That is to say, these techniques are usually trapped into the local optima in the multi-dimensional parameter space. In this paper, a new hybrid identification algorithm (HIA) is proposed to overcome the above drawback. Our method can simultaneously optimize the premise and consequent parameters by merging the Harmony Search (HS) algorithm, FCM algorithm and the least-squares method together.HIA partitions the whole parameter space into two parts: the premise and consequent parameter subspaces. Two different optimization methods are used in these two parts, respectively, i.e., HS algorithm in the premise subspace and the least squares method in the consequent subspace. The HS algorithm, a meta-heuristic stochastic optimization method free from the gradient information, is crucial in our HIA, because it can guarantee that the optimization process always converges to the global optima. Moreover, the error feedback is used in the HIA to build the interaction between the two optimization processes, which is helpful to further improve the search efficiency as well as identification accuracy. In this thesis, we use three numerical examples to examine the performance of the HIA. The HIA is also employed in the T-S modeling of the fiber optic gyro-stabilized platform. The identification results have demonstrated that it can effectively escape from the local optima, and yield a superior performance over the conventional identification methods.