Anomaly detection for the control rod drive mechanism(CRDM) is key to enhancing the security of nuclear power plant equipment. In CRDM real-time condition-based maintenance, most existing methods cannot deal with long...Anomaly detection for the control rod drive mechanism(CRDM) is key to enhancing the security of nuclear power plant equipment. In CRDM real-time condition-based maintenance, most existing methods cannot deal with long sequences and periodic abnormal events and have poor feature extraction from these data. In this paper,a learning-based anomaly detection method employing a long short-term memory-based autoencoder(LSTM-AE)network and an extreme gradient boosting(XGBoost)algorithm is proposed for the CRDM. The nonlinear and sequential features of the CRDM coil currents can be automatically and efficiently extracted by the LSTM neural units and AE network. The normal behavior LSTM-AE model was established to reconstruct the errors when feeding abnormal coil current signals. The XGBoost algorithm was leveraged to monitor the residuals and identify outliers for the coil currents. The results demonstrate that the proposed anomaly detection method can effectively detect different timing sequence anomalies and provide a more accurate forecasting performance for CRDM coil current signals.展开更多
基金supported by the Natural Science Foundation of Fujian Province (No. 2022J01566)。
文摘Anomaly detection for the control rod drive mechanism(CRDM) is key to enhancing the security of nuclear power plant equipment. In CRDM real-time condition-based maintenance, most existing methods cannot deal with long sequences and periodic abnormal events and have poor feature extraction from these data. In this paper,a learning-based anomaly detection method employing a long short-term memory-based autoencoder(LSTM-AE)network and an extreme gradient boosting(XGBoost)algorithm is proposed for the CRDM. The nonlinear and sequential features of the CRDM coil currents can be automatically and efficiently extracted by the LSTM neural units and AE network. The normal behavior LSTM-AE model was established to reconstruct the errors when feeding abnormal coil current signals. The XGBoost algorithm was leveraged to monitor the residuals and identify outliers for the coil currents. The results demonstrate that the proposed anomaly detection method can effectively detect different timing sequence anomalies and provide a more accurate forecasting performance for CRDM coil current signals.
