The large number of types of reactor coolant pump condition sensors leads to three difficulties in current main pump condition anomaly detection: difficulty in modeling complex condition data, difficulty in detecting early abnormalities, and difficulty in assessing the degree of abnormality. The study of rapidly developing deep learning techniques such as neural networks can provide new ideas to solve these problems. To this end, a codec network structure anomaly detection method MSFCNAD (multi-scale FCN-based anomaly detection) based on multi-scale fully convolutional neural networks is proposed. Based on the multivariate temporal characteristics of the main pump state, this method uses the full convolutional neural network codec for pixel-level training to precisely locate the abnormal range of reactor coolant pump state data. At the same time, taking into account the main pump state abnormal time characteristics, the multi-scale feature matrix of reactor coolant pump state is extracted, and the abnormal extent is judged by the graded abnormal range detected by the feature matrix of different scales. On this basis, experiments are conducted using real nuclear main pump data to compare the classification effects of several models such as ARMA, BiLSTM, FCN and AEs. The results show that the MSFCNAD model outperforms the models listed in the paper in terms of recall and F1 score, which are 78.44% and 80.30%, respectively, better than the highest 77.53% and 69.74% of the other models. The experimental results show that this method has better performance compared with other anomaly detection methods, and it can also judge the severity of anomalies by the degree of anomalies and prioritize maintenance processing.

anomaly detection  /  multivariate time series  /  fully convolutional neural networks  /  multiscale analysis  /  deep learning