Addressing unknown faults diagnosis of transport ship propellers system based on adaptive evolutionary reconstruction metric network

Accurate fault diagnosis of the propeller supports the normal operation of each ship and aids in the decision-making for maintenance and industrial dispatch. However, the unpredictable status presents challenges for effective fault diagnosis in real applications, particularly involving unknown operating conditions and fault modes. Therefore, this paper proposes a single-source domain generalization diagnostic method based on an adaptive evolutionary reconstruction metric network, achieving high diagnostic precision for propellers. Specifically, an embedded self-evolution regularization strategy is designed to compel the model to learn the residual label-related features, thereby enhancing the model’s generalization capabilities. Moreover, a reinforcement learning-based adaptive threshold mechanism is built to reinforce the model’s adaptability when facing unknown faults. Relying on a real-world data collection platform for transport ship propellers and a public dataset, the superiorities of the proposed AERMN are demonstrated by comparing it with several strong baseline and cutting-edge methods.