Leveraging Causal Inference for System Fault Diagnosis in Incremental Operating Conditions

While data-driven fault diagnosis is essential for industrial operations, maintaining performance under non-stationary conditions remains a challenge due to catastrophic forgetting. This study introduces a replay-free, causality-based framework designed to mitigate forgetting while ensuring privacy. We theoretically characterize the forgetting phenomenon within a causal graph structure, demonstrating that feature representations from the previous task function as a collider. This collider blocks the information flow between historical and current data distributions. To rectify this, we employ a de-confounding strategy by conditioning on the colliding factor. The framework further incorporates enhanced contrastive learning to enforce local topological consistency and a centroid alignment mechanism to ensure global class separability. Empirical results demonstrate this approach effectively balances the stability-plasticity dilemma, delivering superior diagnostic performance in incrementally evolving industrial environments.

Keywords: Fault diagnosis, Causal inference, Structural causal model, Incremental learning, Data-driven, Bearing.