Intelligent Fault Classification Exploration Inspired by Suprathreshold Stochastic Resonance

The aim of this article is to present a novel intelligent fault diagnosis method based on the principle of stochastic resonance (SR) in array of noisy ReLU activators. To this end, the phenomenon of suprathreshold stochastic resonance (SSR), which is a resonance like collective effect of the array model for the input above threshold, is demonstrated, and the limit of the input-output relation of the average response is explicitly acquired. With the explicit noise intensity dependent relation as an activation function, the new fault diagnosis algorithm is proposed by training the intensity of Gaussian white noise as an independent parameter, and the corresponding initial value is set referring to the minimum point of the loss function defined by cross-entropy. A penalty term is further introduced to the loss function to ensure the nonnegativity of noise intensity. The performance and noise robustness of the proposed algorithms is validated by typical state-of-the-art networks across three open datasets, namely, CWRU bearing fault dataset, SEU gearbox dataset, and HUST varying speed gearbox dataset. It is found that the proposed algorithm outperforms its counterparts both in the absence of noise and in the presence of fixed noise on test accuracy. Particularly, the proposed algorithm can markedly improve the test accuracy when the training samples are limited and noise contaminated.