Adaptive Quantization for Lightweight Fault Diagnosis With FPGA-Accelerated Edge Inference

Lightweight fault diagnosis (FD) models are essential for efficient inference on resource-constrained IoT edge devices and practical industrial deployment. While existing lightweight FD research has achieved notable progress in network architecture compression, the substantial benefits brought by quantization techniques remain largely overlooked. However, directly applying mainstream quantization methods to FD tasks is problematic, as they suffer from two main drawbacks: 1) Uniform quantization struggles to simultaneously maintain local resolution in dense regions and dynamic range coverage in sparse regions. 2) Mainstream quantization-aware training (QAT) methods rely on the straight-through estimator (STE) to approximate discrete gradients, introducing gradient bias that further limits model accuracy improvement. To address these issues, this paper proposes an Adaptive Quantization Error Minimization Method (AQEMM) for fault diagnosis, with the following main contributions: 1) An adaptive quantizer is introduced to replace the traditional fixed quantization step size, expanding the search space of quantization points from discrete integer grids to a continuous subspace, enabling quantization levels to adaptively concentrate in data-dense regions. 2) A block coordinate descent strategy is incorporated to alternately optimize quantization parameters, with quantization reconstruction error minimization as the direct optimization objective, fundamentally circumventing the gradient bias introduced by STE. Validation on FD tasks for two typical industrial devices, induction motors (IM) and permanent magnet synchronous motors (PMSM), demonstrates that the proposed 2-bit quantization model saves up to approximately 82% of hardware resources compared to full-precision (FP) deployment on an FPGA platform while maintaining competitive diagnostic accuracy. Furthermore, by optimizing the hardware design to leverage the released hardware resource margin, a significant inference speedup of approximately 20.6× is achieved over pure software inference.