Research On Echo State Networks Classification Algorithm And Application

Artificial neural networks is a mainstay in classification research. However, training process in classical neural networks is always complex or easy to fall into local minima, which restricts the classification performance and application effect. Based on a new type of neural networks - echo state networks (ESNs) which has the characteristics of simple training process and global-optimization, this dissertation lays stress on the classification method research. Reservoir optimization is a central issue in echo state networks research. Moreover, feature selection with ESNs for complex classification problems needs to be studied. Accordingly, considering the influence of reservoir structure, parameters optimization and feature selection on the classification performance of ESNs, this dissertation focuses on reservoir optimiza-tion and feature selection of ESNs classification method and its applications.The main research work includes the following aspects:1. A classification method using ESNs with corresponding clusters is proposed, which is inspired by complex network topologies imitating cortical networks of the mammalian brain. Multiple-cluster reservoir is constructed by adopting the time windows functions. It is different from the classical method which generates a reser-voir randomly completely. The number of clusters corresponds with the number of classes in specific classification problems to improve the classification accuracy. Ex-perimental results show that the proposed method achieves lower classification error rate compared with the original echo state networks.2. Optimization of reservoir parameters in ESNs classifier has great influence on the classification performance. A reservoir parameters optimization method based on differential evolution algorithm is proposed. Differential evolution is easy to use, robust, and has excellent global optimization ability, which is arguably one of the most powerful stochastic real-parameter optimization algorithms. Experimental re-sults demonstrate that the proposed method is simple and effective, with excellent performance in efficiency and accuracy.3. Feature selection is a key issue in classification method using ESNs, especial-ly for high-dimension and complex data. A wrapper feature selection method based on differential evolution algorithm is proposed. It employs the classification error rate as the evaluation criterion for the classification performance of ESNs which di- rects the search of differential evolution algorithm. Experimental results indicate that the proposed method can find the reduction feature subset for ESNs classifier and reduce the time comsuming. On this basis, a simultaneous ESNs parameters optimi-zation and feature selection method based on differential evolution is proposed, which aims to solve the problem of parameters settings influence on feature selection. Parameters optimization is simultaneously implemented during the process of feature selection. Experimental results indicate that the proposed method can find the opti-mized parameters and the reduction feature subset for ESNs classifier and improve the classification performance.4. Inspired by ensemble learning, a random subspace and multiple reservoirs based classifier is proposed. Multiple reservoirs are constructed which correspond with feature subsets generated by random subspace method. Classification combina-tion is embeded in the learning of multiple reservoir ESNs, which is benefited from the advantage of ESNs - only the weights of reservoir-to-output connections are computed. Experimental results show that the proposed method achieves lower clas-sification error rate compared with the original echo state networks.5. To evaluate the validity and practicability of the proposed methods, the im-plementation structure is designed for the practical application research on fault di-agnosis. Simulation and actual experiment results show that the proposed methods are applicable to the fault diagnosis with good diagnostic performance.