A spiking neural network algorithm with residual membrane encoding for bearing fault diagnosis

Accurate and timely fault diagnosis of rotating machinery is crucial for maintaining operational reliability and safety in modern industrial systems. With the rapid development of neural computing, spiking neural network (SNN) has shown great potential in temporal information processing with their biologically inspired characteristics. However, existing SNN-based fault diagnosis methods show significant limitations in practical applications. The ineffective preservation of temporal characteristics in current spike encoding schemes and the limited capability of traditional neuron models in maintaining historical information severely affect the diagnostic reliability, leading to unsatisfactory fault detection performance. In view of these problems, this paper proposes a Residual Membrane Encoding SNN (RMESNN) for bearing fault diagnosis. The bitwise shift and superposition encoding scheme efficiently processes temporal information by transforming vibration signals into discrete spike sequences, preserving critical fault characteristics in the encoding process. The residual membrane soft-spiking neuron model enhances feature capture by maintaining historical membrane potential information and improving the network’s ability to recognize complex fault patterns. The effectiveness of the proposed method is validated through comprehensive experimental evaluations on multiple benchmark datasets. Experimental results demonstrate that RMESNN achieves superior diagnostic accuracy, reaching 99.4% on the CWRU dataset, improving upon the current state-of-the-art accuracy of 97.92% by approximately 1.5%.