针对核电二次回路中给水回热等复杂热工系统故障诊断存在样本稀缺、数据高维耦合以及数据预处理丢失特征信息导致传统诊断模型准确率受限的问题，提出一种基于相关分析进行深度学习的故障诊断方法。首先，构建给水回热数字孪生系统，建立故障诊断数据仓库；然后，利用综合相关分析(CCA)方法建立系统状态矩阵，搭建深度卷积网络，对系统状态特征进行深入挖掘，并通过可视化（VS）方法将网络的内部过程可视化；最后，建立基于深度卷积网络的故障诊断模型，诊断给水回热系统的典型故障，包括管道泄漏、水室短路、传热恶化、阀门卡涩等。结果表明:该数字孪生系统能够实现对给水再热系统正常和故障工况下的精确仿真，并满足后续深度学习模型的数据要求；基于综合相关分析和深度卷积网络算法的故障诊断模型能够实现时变、多维工业数据的故障诊断；采用T分布随机领域嵌入（TSNE）方法对模型可视化发现，模型的不同故障类型有明显的区别，相似的故障数据有明显的聚合性。

Aiming at the problems of sample scarcity, highdimensional coupling of data and loss of feature information in data preprocessing in the fault diagnosis of complex thermal systems such as feedwater regeneration in the secondary circuit of nuclear power, which leads to the limited accuracy of traditional diagnostic models, a fault diagnosis method based on correlation analysis for deep learning is proposed. Firstly, the digital twin system of water supply regeneration is constructed, and the fault diagnosis database is established; then, the comprehensive correlation analysis (CCA) method is used to establish the system state matrix, build a deep convolution network, deeply mine the system state characteristics, and visualize the internal process of the network through the visualization (VS) method; finally, a fault diagnosis model based on deep convolutional network is established to diagnose typical faults of water supply regenerative system, including pipeline leakage, water chamber short circuit, heat transfer deterioration and valve jam. The results show that the digital twin system can realize the accurate simulation of the water supply reheat system under normal and fault conditions, and meet the data requirements of the subsequent deep learning model; the fault diagnosis model based on comprehensive correlation analysis and deep convolution network algorithm can realize the fault diagnosis of timevarying and multidimensional industrial data; the Tdistribution stochastic neighbor embedding (TSNE) method is used to visualize the model. The different fault types of the model are obviously different, and the similar fault data have obvious aggregation.

TL48