【摘要】 核电厂单一故障识别的方法有很多，但是由于核电厂的复杂性，复合故障识别的难度较大，且传统故障诊断方法存在难以利用核电厂运行数据中时序信息的问题。针对上述问题，提出一种循环神经网络和多标签分类方法相结合的核电厂复合故障诊断方法。该方法首先将故障数据切分为携带时序信息的输入样本；然后，通过循环神经网络提取故障样本中的时序特征；最后，通过多标签分类器完成多个故障标签的解耦输出，实现了复合故障的诊断。仿真实验验证了所提方法无论是对单一故障还是复合故障都具有良好的故障诊断效果。同时，还探究了不同循环神经单元和不同长度的输入样本对模型诊断效果的影响，结果表明：LSTM模型和GRU模型的效果优于常规RNN模型，且增加输入样本的长度并不一定能够提升模型诊断准确率。更多还原

【Abstract】 Most researchers focus on the identification of single faults in nuclear power plants. Howe-ver, because of the uncertainty in the operation of nuclear power plants, compound faults are still possible. Moreover, traditional fault diagnosis methods struggle to utilize the temporal information in the operational data of nuclear power plants. To address those issues, a novel diagnostic method for compound faults in nuclear power plants, combining recurrent neural networks and multi-label classification, was proposed. That method first segmented the fault data into input samples carrying temporal information, then extracted the temporal features within the fault samples using a recurrent neural network, and finally decoupled and outputed multiple fault labels using a multi-label classifier, achieving the diagnosis of compound faults. Through data simulation experiments, the proposed method was validated to effectively diagnose both single and compound faults. Additionally, the impact of different recurrent neural units and varying lengths of input samples on the diagnostic performance of the model was explored. The performance of LSTM and GRU models surpasses that of conventional RNN models, and increasing the length of input samples does not necessarily enhance the diagnostic accuracy of the models.更多还原