Synaptic regularization for continual transfer: A universal fault diagnosis framework for marine diesel engines under incremental operating conditions

With the advancement of ship intelligence, the generalisation capability of diesel engine fault diagnosis models under variable operating conditions becomes paramount. However, existing cross-condition diagnostic models often sacrifice performance in historical conditions when adapting to new scenarios, while replay or extension strategies increase computational and storage demands. Addressing these challenges, this paper proposes a continual learning diagnostic framework tailored for incremental operating conditions in marine diesel engines. This framework innovatively introduces a continuous learning method inspired by synaptic remodeling. It quantifies the importance of historical tasks through path integration, incorporates a structured regularization term into the loss function to flexibly anchor critical parameters, and designs a dynamic binary mask based on importance thresholds to physically block perturbations to old knowledge. Experiments on two critical subsystems of a diesel engine—demonstrate that the proposed framework effectively overcomes catastrophic forgetting without expanding model capacity or retaining historical data. It significantly outperforms traditional fine-tuning and mainstream regularization-based continual learning methods across key metrics including average accuracy, and forgetting rate. This research establishes a diagnostic paradigm with genuine lifelong learning capabilities, providing theoretical and methodological support for constructing robust, sustainably updatable diagnostic models for marine engines operating under diverse long-term conditions.