Fault Diagnosis of Nuclear Power Plants Based on 1D-CNN with Dual Attention Mechanism

Abstract

A nuclear power plant (NPP) is a large, integrated, complex, safety-critical industrial system. A small failure, such as a small loss of coolant accident (LOCA) caused by a pipe rupture, can evolve into a serious accident that threatens the safety of the plant. Thanks to the development of advanced sensor, communication and database technologies, a large amount of monitoring data is connected to the main control room in real time, providing operators with adequate information about the plant's status. However, when a fault occurs, massive alarm information enters the main control room, which usually exceeds the operator's analysis ability. Under great psychological pressure, it is difficult to make the right response, leading to increased risk of human error. Deep learning methods, such as convolutional neural network (CNN), are widely used in image recognition, computer vision and fault diagnosis fields. CNN achieves pattern recognition by extracting abstract features of the input data layer by layer through a deep network architecture. When facing big data scenarios, CNN may have difficulty in capturing important information, leading to long training time or non-convergence problems and low diagnostic accuracy. In this paper, we propose a fault diagnosis model for nuclear power plants based on one-dimensional CNN and dual-attention mechanism (CNN-BiAM). The parameter attention module and feature attention module are introduced to help the CNN model focus on key information and improve the feature extraction capability. Data samples of 10 types of fault conditions such as LOCA, main steam line break (MSLB) and steam generator tube rupture (SGTR), are obtained using the pressurized water reactor simulator software called PCTRAN, and the case study shows that the diagnostic accuracy of CNN-BiAM model is better than that of traditional methods, such as CNN, support vector machine (SVM) and random forest (RF) models.