Ultralow Latency ANN–SNN Conversion for Bearing Fault Diagnosis

Spiking neural networks (SNNs) achieve an impressive performance due to low power consumption and quick inference on neuromorphic hardware. Among various SNN training methods, the ANN-SNN conversion approach can achieve performance levels comparable to those of artificial neural networks (ANNs). However, the spike firing rate of SNNs needs to be aligned with the activation of ANNs over longer time steps, which leads to a performance decline of SNNs at short time steps, limiting their applicability. In response to the challenge, this research proposes an innovative framework for bearing fault diagnosis. The framework first trains an ANN model with a multiscale convolutional attention mechanism (MCNN-AM) that has the capability of feature extraction and noise resistance. Subsequently, the ANN is trained and converted to an SNN using the quantization clip-floor-shift (QCFS) activation function. During the inference phase, an optimization strategy based on residual membrane potential (SRP) is introduced to effectively reduce the SNN response latency while maintaining diagnostic accuracy. The proposed framework enhances the capability of SNNs for diagnosing faults in bearings and enables the deployment of SNNs on compact and mobile platforms.