Online Fault Diagnosis Using Bioinspired Spike Neural Network
Lie Xu, Daxiong Ji
Abstract
Data-driven fault diagnosis methods suffer from restrictive assumptions that hinder adaptability to varying work conditions, rendering offline modes insufficient. Addressing this challenge, a bioinspired spike neural network (bio-SNN) is proposed, featuring an innovative online learning mode. The network employs a novel spike encoding method with data compression for efficiently transforming time series data into spike sequences. This encoding method involves converting 1-D series data into a 2-D spectrogram using a filter bank, incorporating a stochastic spike rate for a more flexible representation compared to precise spike rates. The application of a biologically plausible learning rule, specifically spike timing-dependent plasticity (STDP), enhances the adaptability of the network. A horizontal inhibition and homeostasis mechanism are also introduced, facilitating effective online updating of synaptic weights. Experimental results on two well-established fault datasets showcase the advantages of the bio-SNN method over existing approaches, highlighting its potential for robust and adaptive fault diagnosis in practical scenarios.