An Adaptive Quantization Method for Edge Computing-Based Fault Diagnosis

Deep neural networks (DNNs) perform well in the field of industrial equipment fault diagnosis (FD). However, due to the constraints of practical industrial environments, the substantial computational and memory requirements of DNN make them incompatible with resource-constrained edge devices. Current research on lightweight FD models primarily focuses on network architecture simplification and parameter scale compression. However, there is insufficient research addressing the storage and computational resource overhead caused by floating-point representation redundancy in DNN. Therefore, this article proposes an adaptive quantization interval method based on knowledge distillation (AQIKD) for FD. First, a hardware-friendly adaptive quantization interval strategy is proposed, which balances model performance and hardware implementation complexity by differentiating the quantization methods for weights and activation values. Second, the hyperbolic tangent function is employed to optimize the backpropagation process to enhance the model’s convergence capability. Finally, a progressive soft label supervision learning method is introduced, leading to both acceleration in training convergence and improvement in diagnostic performance for low-bit-width models. The stability and efficiency of the proposed quantization method are validated through experiments on both induction motors (IMs) and permanent magnet synchronous motor (PMSM) scenarios. Furthermore, FD models with different bit-widths implemented based on AQIKD are deployed on an FPGA platform. The results demonstrate that the quantized models can achieve up to 74% hardware resource reduction while maintaining satisfactory accuracy.