Fault Diagnosis of Pump Control Valves Based on Multi-Frequency Differential Transformation and a Residual Attention Network
Accurate fault diagnosis of pump control valves is essential for ensuring the safe and reliable operation of water supply systems. However, heterogeneous sensor signals collected from industrial monitoring systems are typically sampled at different frequencies, making effective information fusion and fault feature extraction challenging. To address these issues, this paper proposes a fault diagnosis method based on Multi-Frequency Differential Transformation (MDFT) and a Residual Attention Convolutional Network (RACNet). First, multi-source sensor signals are grouped according to their sampling frequencies and transformed into RGB trajectory images through MDFT, enabling effective fusion of heterogeneous data while preserving frequency-specific characteristics. Then, RACNet integrates dilated convolution, Efficient Channel Attention (ECA), and Global Response Normalization (GRN) to enhance feature extraction and improve the identification of subtle fault patterns. Experiments are conducted on a real-world pump control valve dataset containing four health conditions: healthy, slight lag, severe lag, and approaching failure. The proposed method achieves an average diagnostic accuracy of 97.73%, outperforming several representative fault diagnosis approaches. Ablation studies further demonstrate the effectiveness of the MDFT encoding strategy and the RACNet architecture. The results indicate that the proposed method provides an effective solution for intelligent health monitoring and fault diagnosis of pump control valves.