Bearing Fault Diagnosis Based on Residual Networks and Grouped Two-Level Attention Mechanism for Multisource Signal Fusion

Bearing fault diagnosis is critical in the operation and maintenance of industrial equipment. By fusing signals from multiple sensors, complementary information that may be missed by a single sensor can be compensated. However, with the increase in the number of sensors, current multi-source signal fusion methods encounter challenges like feature dilution caused by signal mixing, increased model complexity, and information redundancy. To tackle these challenges, this paper proposes a novel fault diagnosis method for bearings based on multi-source signal fusion—the residual grouped two-level attention network (Res-GTLA-Net). First, multiple residual blocks with a kernel height of 1 are employed to extract features from the two-dimensional (2D) matrices converted from multi-source signals, which avoids the increase in model complexity caused by multiple network branches and prevents the dilution of each signal’s unique features during feature extraction. Secondly, the proposed grouped two-level attention mechanism performs feature enhancement and fusion at both intra-group and inter-group levels, focusing on the most informative signals. Experimental results indicate that Res-GTLA-Net performs better with more sensors.