75 / 2023-09-02 15:12:16

Multimodal fusion-based fault diagnosis of motor in CRF pump-unit under data imbalance

multimodal fusion, motor fault diagnosis, attention mechanisms, imbalanced data, convolution neural network

Special Sessions > SS16 - Multi-Physical Sensors-Based Detection & Diagnosis Solution Empowered by AI Technology

Fault diagnosis technology has developed from the initial traditional expert experience and signal processing methods to the stage of intelligent fault diagnosis, and the combination of intelligent sensors and deep learning technology plays an increasingly important role. Nowadays, the unimodal fault diagnosis technology can no longer satisfy the actual working conditions. Currently, there are several problems in the motor fault diagnosis of circulating pump sets in nuclear power plants (CRF): hardware failure of sensors, data imbalance, and multi-speed situations all lead to the degradation of the diagnostic performance. In this paper, we propose a diagnostic framework based on multimodal fusion, which first extracts the feature layers of vibration and acoustic data based on a onedimensional convolutional neural network, and then feeds the feature layers into the Multi-head self-attention feature fusion (MAFF) module proposed in this paper, and finally performs the prediction. The results show that the proposed method can improve the fault diagnosis accuracy compared with unimodal sensors, and at the same time, it is also superior to the fault diagnosis accuracy of other methods in the case of extremely imbalanced data.