Sensor fault diagnosis based on a passive observer and hybrid CNN-GRU model for intelligent ships

Abstract Aiming at the problem of unknown model accuracy, high sample representativeness requirements, and the loss of time series features in traditional model-based and data-driven fault diagnosis for intelligent ship sensors, this paper proposes a hybrid fault diagnosis method that combines both approaches. A nonlinear passive observer is constructed to generate system residual signals at first. Then, a convolutional neural network and a gated recurrent unit neural network are used to extract local and time series features, respectively. Moreover, the self-attention mechanism is introduced to further distinguish the important relationship between different time points of the signal. Finally, the fault diagnosis is realized through the classifier. Experimental results based on an intelligent ship model show that the diagnosis rate increased by 7.4% compared to models without an observer. Compared to traditional machine learning and deep learning methods, the proposed model achieves a diagnostic accuracy of over 99%, demonstrating superior performance.