基于多维变换的模拟电路故障诊断的神经网络方法

Based on a bilinear transformation in multidimensional spaces, a neural-network based nonlinear analog fault diagnostic system for actual circuits is proposed. The proposed method uses multiport functions to diagnose the circuits without sufficient accessible nodes. The approach, which is based on transferring parameter loci to multidimensional spaces, makes the distances between the loci in the multidimensional spaces be greater. This fact leads to better fault resolution and robustness against the influence of component tolerances and measurement errors. Besides, the method based on transformation in multidimensional spaces is adopted here to select fault signatures can decrease the ambiguity groups and improve the performance of fault diagnosis. This approach allows diagnosing single, double, and multi-faults in linear circuits. For nonlinear circuit, after the nonlinear component substituting by their piece-wise linear (PWL) models, the method can be used in nonlinear circuit fault dignosis. Under considering the characteristics of the transformation in multidimensional space, the particle swarm optimization (PSO) algorithm-based BP neural network is applied to diagnose faults. Finally, the realization of the proposed strategy is expounded by using two examples.