Abstract: To solve the problem of dynamic pattern classification of complex time signals, this paper proposes a new radial basis function process neural network (RBFPNN) model, which combines local kernel functions and global ones. Considering the diversity and complexity of time signals, the kernel functions have the local and representation ability for signal distribution morphology. By linear superposition of polynomial kernel functions for global properties with Gaussian kernel functions for local pro- perties, the new combined kernel function is composed and a new RBFPNN model is constructed, which improves the information model's extraction and memorization properties for dynamic sample complex process, and enhances the recognition capability of RBFPNN for time signal feature.  This paper analyzes the nature of the new RBFPNN model, and constructs the model parameter optimization algorithm based on CGA. By taking reciprocation machinery working status diagnosis based on indicator diagram as an example, the actual data processing results show the effectiveness of the model and the algorithm.

Key words: dynamic pattern recognition, radial basis process neural network, combined kernel function, chaos genetic algorithm