Cross-sensor collaborative distillation with dynamic competitiveness balancing for fault diagnosis under imbalanced missing data

Multi-sensor fault diagnosis often suffers from imbalanced missing data rates, which can bias learning toward sensors with higher data availability and degrade diagnostic performance. This paper proposes a cross-sensor collaborative distillation framework with dynamic competitiveness balancing to address this problem. The framework transfers shared diagnostic knowledge across sensors through signal-level logit alignment and semantic-level prototype and distribution matching, while a competitiveness balancing mechanism adaptively regulates sensor contributions during training based on their relative learning progress. By suppressing the dominance of low-missing-rate sensors and preventing the marginalization of high-missing-rate sensors, the proposed method enables balanced multi-sensor representation learning. Experiments on drive motor fault diagnosis datasets demonstrate improved accuracy and robustness under uniform and non-uniform missing data conditions compared with existing methods.