Intelligent Fault Diagnosis of Centrifugal Pumps via Self-Attention and Convolutional Fusion Method Based on Acoustic-Vibration Signals

The current fault diagnosis methods for centrifugal pumps mainly rely on the analysis and research of vibration signals. However, there are certain difficulties in collecting vibration signals under certain operating conditions, and a single signal performs poorly in the presence of noise labels, while sound signals also have rich characteristic information that can be utilized. This article focuses on the performance of fault diagnosis for centrifugal pumps under acoustic and vibration signals. Firstly, short-time Fourier transform is used to obtain time-frequency domain features from one-dimensional vibration signals and acoustic signals. Then, the two signals are fused and recombined as input signals, and self-attention and convolutional fusion methods are added for fault diagnosis classification. On the centrifugal pump dataset, the proposed method has significantly improved accuracy compared to traditional convolutional neural networks under different noise label ratios.