Analog circuit fault diagnosis based on density peaks clustering and dynamic weight probabilistic neural network

Abstract

Fault diagnosis methods based on probabilistic neural networks (PNNs) have been widely used in various products, owing to their simplicity and efficiency. However, in some multi-condition circuit fault diagnoses, the presence of numerous faults and heterogeneity of the training data make the computational efficiency, classification accuracy, and selection of the pattern neuron samples of the PNN challenging. To overcome these difficulties, this study proposes a novel analog circuit fault diagnosis method based on density peaks clustering and a dynamic weight PNN. In this method, density peaks clustering is performed to determine the number of pattern neuron classes. Based on the results of clustering, a priority function is defined, and a novel two-step pattern neuron optimization algorithm incorporating local density and gravity is proposed. Accordingly, the representative boundary data are selected as the pattern neuron samples and numerous redundant samples are reduced. Moreover, a dynamic grey correlation weight determination (GCWD) algorithm between the input and pattern layers is proposed, to determine which classes of the pattern neurons need to be activated and involved in the diagnostic calculation. In addition, a dynamic proportion-amplified weight determination (PAWD) algorithm between the pattern and summation layers is suggested to reduce the adverse effect of the heterogeneity. This not only reduces the number of calculations in the diagnostic process, but also improves the accuracy of the diagnostic model. Case study on a filter circuit clearly demonstrates that the proposed method can achieve high classification accuracy with only a few pattern neurons.