A Flexible Winding Identification Method Utilizing Self-Attention and DNN for the Double-Suction Centrifugal Pump

Double-suction centrifugal pumps (DCPs) are extensively utilized in water diversion projects, agricultural irrigation, and urban water supply. Due to the increasing river pollution, they are facing a crisis of entanglement by flexible materials such as plastics and aquatic plants. However, there is limited research on the identification of the flexible winding fault in these pumps, and variations in working flow rates present challenges for accurate winding identification. This article proposes an integrated self-attention mechanism and dense neural network (DNN) model, namely, self-attention mechanism with DNN (SDNN), to effectively identify the flexible winding fault using a small training sample size of pulsation signals and head signals across various flow rates. The proposed model leverages the idea of weight allocation in the self-attention mechanism to achieve key information extraction from multidimensional signal inputs. Experimental studies were conducted under different flow rates for both normal and winding scenarios. The results demonstrate that the SDNN achieves the best identification capability on winding faults than other deep learning models, regardless of whether conventional data inputs, imbalanced sample sizes, or noisy environments are considered. Consequently, the SDNN shows the potential for effective and timely monitoring of the health status of DCPs.