An FPGA-based lightweight intelligent fault diagnosis framework for rotating machinery using model pruning and knowledge fusion distillation

Abstract Deep learning has been widely adopted for intelligent fault diagnosis in rotating machinery. However, deep models with a large number of parameters typically require high-power graphics processing units (GPUs) or workstations for efficient inference, which limits their applicability in industrial environments where low power consumption and high portability are essential. To address this issue, we propose a lightweight intelligent diagnostic framework that can be deployed on field-programmable gate arrays (FPGAs). The framework first applies structural pruning to construct a compact student model. To compensate for performance degradation due to compression, a knowledge fusion distillation strategy is introduced, which integrates feature-level and logit-level knowledge and surpasses conventional single-level distillation methods, thereby enabling superior performance recovery. Parameter quantization is subsequently applied to further reduce the model footprint. To support various network configurations, general-purpose neural network operator IP cores are implemented on the FPGA. Experiments on two bearing datasets show that the compressed student model uses only 4.21 KB of parameters while achieving over 98% accuracy, with the FPGA implementation consuming just 1.74 W, representing 16× and 32.66× reductions compared to CPU and GPU deployments, respectively. These results demonstrate the framework’s effectiveness for low-power diagnostic measurement in industrial applications.