TCN-MBMAResNet: A Novel Fault Diagnosis Method for Small Marine Rolling Bearings Based on Time Convolutional Neural Network in Tandem with Multi-Branch Residual Network

Abstract Centrifugal pumps are key components of marine systems, and the health of their internal rolling bearings is critical for the normal operation of ships. In small ship replenishment systems, the rolling bearings of centrifugal pumps often have short failure times, making it difficult to gather balanced failure data. This paper addresses the challenges in extracting fault features from centrifugal pump bearings in ship water replenishment systems and the low accuracy of fault identification caused by unbalanced data samples. We propose a fault diagnosis method that combines a Time Convolutional Neural Network (TCN) with Multi-branch Mixed Attention Residual Networks (MBMAResNet). This paper introduces TCN to capture dependencies in long time series and optimize the feature distribution of raw data. The captured multi-channel 1-D feature sequences are converted into 2-D feature maps to provide spatially correlated features for the designed MBMAResNet. Multiple Single Path Fusion Residual Blocks (SPFBs) with different scales are designed in the MBMAResNet to further extract effective fault features from the unbalanced samples, and a feature fusion structure is utilized for further feature enhancement. Additionally, improved Convolutional Block Attention Modules (CBAMs) are integrated into the network to increase the significance of valid channels and spatial locations. Experiments on an unbalanced public dataset and a private centrifugal pump faulty bearing dataset achieved diagnostic accuracies of 99.71 and 98.50, respectively, demonstrating that our method offers higher accuracy and generalizability compared to other methods.