A physically and biologically interpretable spiking neural network method for rotor system fault diagnosis

Abstract Efficient and reliable diagnosis of rotor system faults is crucial for ensuring the safe operation of industrial equipment. Artificial neural networks (ANNs) face two primary challenges in fault diagnosis: One is that insufficient interpretability can lead to misdiagnosis risks; the other is that high power consumption limits their deployment in resource-constrained scenarios. To address these issues, this paper proposes a physically and biologically interpretable spiking neural network (PBISNN) method for rotor system fault diagnosis. First, Gramian angle difference fields (GADF) are employed as model inputs to represent signal impact characteristics visually. Subsequently, a directional wavelet dictionary convolution (DWDC) with physical significance is designed. This approach improves Morlet wavelet convolution by incorporating angle and scale parameters, utilizing a learnable dictionary to update wavelet kernels for improved matching of image features. Finally, leveraging the biological property of spiking neural networks (SNNs) to simulate brain neuron firing spikes within time steps, a spatio-temporal fusion module (STFM) is designed. This module first performs positional encoding on feature maps to reduce positional information loss, then employs parallel fusion through two branches—contextual spatial attention and step attention—to capture global structure and cross-time step information during training. We validated PBISNN’s effectiveness across three datasets, achieving accuracies of 94.75%, 94.17%, and 99.00% respectively—significantly outperforming other mainstream methods. Post-decision analysis demonstrated biological plausibility and energy efficiency, supporting engineers in deploying miniaturized portable devices for field diagnostics.