【摘要】 核电厂传感器故障后果严重，而核电厂一回路系统和设备的固有复杂性为基于精确数学模型的传感器故障诊断带来了困难。本文提出了一种将深度学习算法与时频分析相结合的核电厂传感器智能故障诊断方法，将信号识别问题转化为图像识别问题。先利用连续小波变换（CWT）对核电厂典型传感器7种常见健康状态的时序信号进行处理，以生成捕捉故障信号特征的时频图；然后以预处理和标记的数据集训练经通道注意力机制（CA）改进的卷积神经网络（CNN）模型，提取时频图的细微图像特征，基于这些特征识别和隔离传感器故障。该方法不需建模和设计阈值，鲁棒性强，准确率达到97%以上，通过与长短期记忆（LSTM）神经网络、一维卷积神经网络（1D-CNN）等典型深度学习网络的诊断效果对比，验证了改进CWT-CNN的有效性和优越性。更多还原

【Abstract】 The consequences of sensor faults in nuclear power plants are serious, and the inherent complexity of primary circuit system and equipment in nuclear power plants brings difficulties to sensor fault diagnosis based on accurate mathematical models. In this paper, an intelligent sensor fault diagnosis method for nuclear power plant based on deep learning algorithm and time-frequency analysis is proposed, which transforms the signal recognition problem into image recognition problem. Firstly, the continuous wavelet transform(CWT) is used to process the time series signals of seven common health states of typical sensors in nuclear power plants to generate a time-frequency diagram that captures the characteristics of fault signals. Then, the convolutional neural network(CNN) model improved by channel attention mechanism(CA) is trained with pre-processed and labeled data sets, and the subtle image features of the time-frequency diagram are extracted. Based on these features, sensor faults are identified and isolated. This method does not need to model and design thresholds, and it has strong robustness and an accuracy rate of more than 97%. By comparing the diagnostic effects of typical deep learning networks such as long short-term memory network(LSTM) and one-dimensional convolutional neural network(1D-CNN),the effectiveness and superiority of the improved CWT-CNN are verified.更多还原