Multi-scale temporal modeling and adaptive channel fusion framework for fault diagnosis of hydraulic systems

Accurate fault diagnosis of hydraulic systems is essential for ensuring safe and reliable industrial operation. To address challenges such as inconsistent sampling rates of multi-source sensor data, feature redundancy, and inefficient inter-channel fusion, this study proposes a multi-scale temporal modeling and adaptive channel fusion framework. The Spearman correlation coefficient is applied to select sensor signals highly correlated with fault patterns, reducing data redundancy. A multi-branch structure processes signals with different sampling rates, where each branch integrates a temporal convolutional network (TCN) and multi-head self-attention (MHSA) to capture local and global temporal dependencies. An efficient channel attention (ECA) module further adaptively weights and fuses multi-channel features, emphasizing key fault information. Experiments on a public hydraulic dataset show that the proposed method outperforms traditional concatenation and other deep learning models, confirming its effectiveness for multi-rate data fusion and complex fault classification.