Abstract
Small modular reactor (SMR) is a complex, time-varying and strongly coupled nonlinear system, which adopts integrated design with short pipeline and small coolant inventory. Sensor failure of SMR under frequent load-following conditions threatens the safety of reactor. It is an effective way to realize sensor fault tolerance by abnormal sensing value reconstruction with machine learning method. In this paper, a sensing value reconstruction model combining hybrid dilated convolution neural network (CNN) and long short-term memory (LSTM) neural network is proposed. The hybrid dilated CNN is used to extract the feature information in the data set as the input of LSTM. LSTM is trained through training set to obtain the reconstructed sensor parameter time series. Besides, the traditional LSTM model is trained as a comparison model to verify the advantages of the reconstruction model. The training set is obtained from the SMR simulation model established by RELAP5 code. Multiple sensing values are input to each other’s reconstruction model. In order to avoid the one abnormal sensing value affecting the reconstruction behavior of other sensing value reconstruction models, the abnormal sensed value is replaced by its reconstructed value as the input of other models. The test set completely different from the training set is used to verify the behavior of the reconstruction model. The results show that the reconstruction model proposed in this paper not only fit the real data well, which means good reconstruction behavior on sensing value reconstruction, but also have better reconstruction effect compared with the traditional LSTM model.
