| dc.description.abstract | Abstract. To accurately address fault diagnosis problems in hydraulic systems for aerospace component testing, computational methodologies mainly rely on typical neural network structures to design specific models tailored for different types of vibration signals or multimodal sensor data. These methods, though effective for faults with distinct signatures, often exhibit poor generalization when distinguishing between fault classes that have similar features, and thus fail to capture their subtle differences. This causes the model to be insensitive to the severity of the fault, which is precisely the focus of hydraulic system fault diagnosis. This article, therefore, constructs a novel multimodal fusion state-attention hierarchical framework termed transition matrix hierarchical network. Feature-level splicing fuses raw acceleration-based vibration signals, raw time-series acoustic signals and time–frequency representations of vibration signals, enriching state features and improving diagnostic accuracy. In the hierarchical diagnostic network, the first layer of the network classifies the differently distributed fault types. For uniformly distributed fault types, the classification is done by the pseudo-labeling of the states within the window combined with the state transition matrix in the second layer. This improvement enables the algorithm to focus on time sequence, providing additional basis for distinguishing between faults with varying degrees of uniform distribution. The proposed method achieves an accuracy of 95.03% on our self-built 14-class datasets and 98.25% on the public 10-class Case Western Reserve University datasets, exceeding the best competing model by 1.40 and 0.41 percentage points, respectively. These gains indicate superior diagnostic performance and generalization across diverse fault classes. | |
