At present, research of cable fault diagnosis based on artificial intelligence mainly takes statistical characteristics as inputs, which means the appropriateness of statistical characteristic selection is directly related to the diagnosis accuracy and the identification results may have certain contingency. Further, most of these methods do not consider the correlation of signals in time. Therefore, this paper proposes a novel diagnosis method for 10 kV single-core cable based on Double-Layer Long Short Term Memory (D-LSTM) network considering timing relationship of multiple observable electrical quantities. Firstly, analysis object is expanded from single electrical quantity to multiple observable electrical quantities, and the relationships among these quantities are analyzed. Secondly, characteristic matrix of combined time series is constructed by time series pairs extracted from multiple observable electrical quantities. Thirdly, the D-LSTM network for processing sequenced input is established according to the features of characteristic matrix. Then, adaptive moment estimation (Adam) method is applied to model training under supervised learning and the model of fault diagnosis is obtained. Finally, recognition experiments are carried out by the proposed method with sample data obtained by simulation of three cable faults and load disturbance. Results show the diagnosis accuracy of proposed method can achieve 99.06%.

目前基于人工智能的电缆故障诊断研究主要以统计特征作为输入，这意味着统计特征选择的恰当性直接关系到诊断的准确性，识别结果可能具有一定的偶然性。此外，这些方法中的大多数没有及时考虑信号的相关性。因此，本文提出了一种基于双层长短期记忆（D-LSTM）网络的10 kV单芯电缆诊断新方法，该方法考虑了多个可观测电量的时序关系。首先将分析对象从单一的电量扩展到多个可观测的电量，并分析这些量之间的关系。第二，组合时间序列的特征矩阵由从多个可观测电量中提取的时间序列对构成。第三，根据特征矩阵的特征建立用于处理序列输入的D-LSTM网络。然后，将自适应矩估计(Adam)方法应用于监督学习下的模型训练，得到故障诊断模型。最后，通过模拟三种电缆故障和负载扰动获得的样本数据，通过所提出的方法进行识别实验。结果表明，该方法的诊断准确率可以达到99.06%。将自适应矩估计(Adam)方法应用于监督学习下的模型训练，得到故障诊断模型。最后，通过模拟三种电缆故障和负载扰动获得的样本数据，通过所提出的方法进行识别实验。结果表明，该方法的诊断准确率可以达到99.06%。将自适应矩估计(Adam)方法应用于监督学习下的模型训练，得到故障诊断模型。最后，通过模拟三种电缆故障和负载扰动获得的样本数据，通过所提出的方法进行识别实验。结果表明，该方法的诊断准确率可以达到99.06%。