Control-aware affine network: a novel framework for reliable fault diagnosis of marine diesel engines in dynamic maneuvers

Maneuvering factors are often insufficiently accounted for in fault diagnosis, which may lead to biased results. Traditional approaches that merely concatenate control signals with condition-monitoring signals as inputs typically fail to deliver satisfactory performance. To overcome this limitation, this paper proposes a novel Control-Aware Affine Network (CAAN), which converts control signals from passive disturbances into active calibration mechanisms. Firstly, a control-aware calibration module employs real-time control signals to dynamically apply affine transformations to fault features, actively mitigating operating condition variations. Secondly, a dual-feature decoupling method decomposes calibrated features into fault severity and maneuver-invariant components. Fault severity features are derived via knowledge distillation to retain category details, while maneuver-invariant features are extracted via adversarial training to enhance cross-domain generalization. Finally, the complementary features are fused to form the final fault representation, and the reliability of the proposed CAAN is evaluated across cross-maneuvering tasks using diesel engine injection system simulations.