Early fault identification of high-pressure water pumps based on multi-attention mechanism

Abstract High-pressure pumps serve as critical equipment in emergency water supply systems for geohazard-prone and remote mountainous regions. Early fault identification of such pumps is essential to ensure the continuity and stability of system operation. However, due to extreme working conditions such as gravel impact and environmental corrosion, core components like impellers and bearings are prone to failure. Traditional diagnostic methods often suffer from low accuracy and delayed response under variable operating conditions and high-noise environments. To address these limitations, this study proposes a Multi-Head Attention-enhanced Multi-Scale Convolutional Long Short-Term Memory Network (M2-LSCNN) for early fault diagnosis of high-pressure pumps. Specifically, a Multi-Scale Convolutional Neural Network (MSCNN) is employed to adaptively extract key features from vibration signals across different frequency bands, enhancing the model’s sensitivity to various fault types. Subsequently, a Multi-Head Attention (MHA) mechanism is incorporated to perform weighted fusion of multi-scale channel features, thereby improving feature diversity and model generalization. The fused features are then fed into a Long Short-Term Memory (LSTM) network to capture long-term temporal dependencies, further improving temporal awareness and classification accuracy. Finally, a Softmax layer is used to perform multi-class fault identification. Experimental results on a vibration dataset collected under variable operating conditions demonstrate that the proposed M2-LSCNN model achieves an average fault identification accuracy of 97.57%, validating its effectiveness and superior performance.