Cross-domain fault diagnosis method for nuclear power plant bearings based on deep transfer learning under small sample conditions

Abstract

In recent years, methods based on deep learning have attracted attention in the fault diagnosis of rotating machinery in nuclear power plants (NPPs). However, these methods are typically developed under the assumption that sufficient fault samples are available. In practice, rotating machinery in NPPs operate in healthy state most of the time and faults occur rarely and last a relatively short period of time. This work proposes a fault diagnosis method based on deep transfer learning to overcome the issue of small sample conditions in the bearing fault diagnosis task of NPPs. The bearing vibration signals collected by the sensor are converted into a time-frequency map by synchrosqueezed wavelet transforms, they are used as input of the deep convolutional neural network. In the learning phase, the deep learning model first learns domain-related knowledge from real devices, then the model parameters are transferred to the target task, and the model is fine-tuned based on the target domain knowledge. The proposed method was applied to two case studies: bearing fault localization and fault severity assessment. Experimental results demonstrated that, for the fault localization case, the method achieved average accuracy, precision, and F1 score of 95.21 %, 95.35 %, and 95.17 %, respectively, under four small sample conditions (with 10, 20, 30, and 40 samples per category in the training dataset). For the fault severity assessment case, the method attained average accuracy, precision, and F1 score of 95.03 %, 95.45 %, and 94.94 %, respectively, under three small sample conditions (with 3, 5, and 8 samples per category in the training dataset), demonstrating its potential value for NPPs bearing fault diagnosis.