Rolling Bearing Fault Diagnosis Based on Signal Processing and CNN-BiLSTM

To overcome the challenges of weak fault signals and poor feature extraction, which result in low diagnostic accuracy and insufficient real-time performance in bearing fault diagnosis, this paper presents a diagnosis method that combines optimized Feature Modal Decomposition (FMD) with deep neural networks. First, FMD is employed for noise reduction of the signals, highlighting fault information. To determine the optimal number of modal components (N) and filter length (L) in FMD, minimum envelope entropy is used as the fitness function, and the GOOSE algorithm is applied to optimize these parameters for modal component selection. Next, the optimal components are filtered based on a comprehensive metric for signal reconstruction, and time-domain features of the reconstructed signals are utilized to construct feature vectors, which are then input into a CNN-BiLSTM network for training to complete the fault diagnosis. Finally, experimental validation on the XJTU-SY dataset demonstrates the effectiveness of the proposed method. The experimental results indicate that this method possesses strong generalization ability, achieving a diagnostic accuracy of 99.33%, which significantly surpasses existing methods, with a diagnosis time of only 12.68 seconds, thereby enhancing diagnostic efficiency.