FPGA-based intelligent bearing fault diagnosis under varying-speed conditions

Abstract Although machine learning has made remarkable advancements in addressing complex classification tasks for non-stationary signals, this artificial intelligence technique still faces deployment challenges in real-time industrial applications due to substantial computational requirements and prolonged processing times. In this study, a fault diagnosis framework based on Kalman filter-based empirical mode decomposition (KF-EMD) and a back propagation neural network (BPNN) is implemented on an field-programmable gate array (FPGA) for real-time bearing fault diagnosis. First, angular domain resampling is employed to compensate for speed variations. Then, KF-EMD is applied to decompose the resampled signal in the angular domain, generating multiple intrinsic mode functions (IMFs) in parallel. Finally, a lightweight BPNN is utilized to perform real-time fault classification based on features extracted from these IMFs. The proposed method was validated using a bearing fault dataset from the University of Ottawa, achieving a diagnostic accuracy of 99.61%. Real-time diagnostic experiments using pre-stored data on the FPGA demonstrated the effectiveness of the approach in balancing computational efficiency and diagnostic accuracy.