内蒙古电力技术

针对电力电缆故障在线检测中扩展频谱时域反射法（spread spectrum time delay reflectometry，SSTDR）技术难以同步实现故障定位与分类的问题，提出一种基于长短期记忆（long short-term memory，LSTM）网络与残差网络（residual network，ResNet），即LSTM-ResNet的电缆故障定位与类型识别系统。首先，基于电缆行波传播理论，建立SSTDR检测信号的传播与反射模型，分析不同故障类型下阻抗突变对反射信号时频特性的影响规律。其次，通过Matlab/Simulink平台搭建电缆故障仿真模型，对开路故障、低阻故障、高阻故障及局部受潮4种典型工况进行仿真。研究结果表明，LSTM-ResNet模型能够在包含4000个样本的测试集上实现99.38%的故障分类准确率。该方法能够为突破传统方法依赖人工特征设计的局限性，解决深层网络训练中的梯度退化问题及电缆故障数字化在线检测提供新思路。

Aiming at the problem that spread spectrum time delay reflectometry(SSTDR) technology is difficult to realize fault location and type classification synchronously in power cable fault online detection, this paper proposes a cable fault location and type identification system based on long short-term memory(LSTM) network and residual network(ResNet), namely LSTM-ResNet. Firstly, based on the cable traveling wave propagation theory, the propagation and reflection model of SSTDR detection signal is established, and the influence of impedance mutation on the time-frequency characteristics of reflected signal under different fault types is analyzed. Secondly, the cable fault simulation model is built by Matlab/ Simulink platform, and the four typical working conditions of open circuit fault, low resistance fault, high resistance fault and local damp are simulated. The results show that the LSTM-ResNet model can achieve a fault classification accuracy of 99.38% on a test set containing 4000 samples. This method can provide new ideas for breaking through the limitations of traditional methods relying on artificial feature design, solving the problem of gradient degradation in deep network training and digital online detection of cable faults.