A Bias-Variance Balanced Solution for Vibration-Based Fault Diagnosis of Tunnel Boring Machine Main Bearing

Vibration-based diagnosis of Tunnel Boring Machine (TBM) main bearings is challenged by long-tailed fault distributions and low signal-to-noise ratio (SNR), which jointly induce a bias-variance dilemma: models tend to be biased toward frequent classes, while noise increases feature uncertainty and weakens fault discrimination. To address this issue, this study proposes a Bias-Variance Coordinated Temporal Network (BVC-TNet). The framework integrates a bilateral-branch architecture with a dynamic accumulation learning strategy. Specifically, a bias-rectifying branch enhances the learning of underrepresented tail classes, whereas a variance-reduction branch preserves robust feature extraction under the original long-tailed distribution. Their loss contributions are dynamically coordinated during training to balance the bias-variance trade-off between rare-fault recognition and diagnostic stability. Experimental results on a TBM main bearing dataset collected from a scaled-down test rig show that BVC-TNet attains not only the highest overall accuracy (90.24%) but also the smallest performance variance among all compared methods. Moreover, it maintains high and stable performance on head and medium classes, including normal and early-stage fault states, while delivering consistent gains on tail classes, demonstrating a more effective balance between bias suppression and variance control.