AI Lightweight and FPGA Parallel Acceleration for Bearing Life Prediction Based on Contrastive Learning and Pruning Optimization

Aiming at the engineering problems of deep learning models for rolling bearing remaining useful life (RUL) prediction, such as large number of parameters, high inference latency, and difficulty in adapting to low-power real-time deployment on embedded edge devices, an edge real-time prediction scheme combining AI lightweight optimization and FPGA parallel acceleration is proposed. Different from general time-series processing methods, this scheme is customized for the nonlinear degradation characteristics of bearings (stable healthy stage and drastic degradation stage). It adopts the positional encoding-driven variable-density time-series sampling to achieve efficient data compression, constructs a lightweight CNN feature extractor based on self-supervised contrastive learning with clear positive/negative sample construction, and optimizes the decoder with a dual regularization strategy to suppress overfitting. On this basis, unstructured pruning is used to reduce the model scale, a modular parallel FPGA hardware accelerator with clear processing elements, pipeline structure and resource mapping is designed for LSTM, CNN and fully connected layers, and mixed fixed-point quantization is adopted to balance accuracy and hardware occupation. Experimental results on the IEEE PHM 2012 bearing dataset show that the proposed scheme reduces the end-to-end inference latency to 0.193 ms with 0.606 W power consumption, achieving 139.9 times speedup over the original software model, and has no obvious accuracy loss in the early and middle degradation stages. The detailed FPGA resource occupation (42% LUT, 20% FF, 14% BRAM, 14% DSP) is fully analyzed. This scheme meets the edge real-time health monitoring requirements of industrial rotating machinery and provides a feasible technical path for the embedded implementation of life prediction algorithms.