基于改进卷积神经网络的无绝缘轨道电路调谐区故障诊断

针对现有故障诊断中忽略调谐区故障对列车安全影响的问题,建立分路状态下机车信号电压模型,获取调谐区正常状态及不同故障状态的机车信号电压;并考虑补偿电容故障、不同道砟电阻等对机车信号电压的影响,对现有数据库进行扩充,得到调谐区故障数据集。在此基础上,采用卷积神经网络(CNN,Convolutional Neural Networks)实现调谐区的故障诊断,特征提取通过CNN中卷积层实现,并对比不同卷积层参数下诊断准确率及训练时间,选择当前条件下相对最优的卷积层参数;采用dropout函数避免训练中出现过拟合现象,并通过CNN中第2全连接层实现故障分类。针对人为构建数据集时数据标签错误的问题,通过构建标签错误数据集的方式,减小错误标签数据对训练过程的影响。测试结果表明:当信噪比为40 dB时,测试集的准确率为97.92%,即在噪声环境下,该诊断方式仍然有效。

Fault diagnosis of tuning zone of jointless track circuit based on improved convolutional neural network

To solve the problem that the influence of tuning zone fault on train safety is neglected in existing fault diagnosis, a voltage model of locomotive signal in the shunting state is established to obtain the signal voltage of locomotive in normal state and different fault state of tuning zone. Then, the initial database is expanded to obtain the data set of tuning zone fault by considering the influence of compensation capacitance fault and different ballast resistance on the signal voltage of locomotive. On that basis, Convolutional Neural Network is used to realize fault diagnosis of tuning zone. Hereinto, feature extraction is realized by convolution layer in the CNN and the relatively optimal convolution layer parameters are selected by comparing different convolution layer parameters in term of the accuracy of diagnostic and training time under current condition. Furthermore, dropout function is used to avoid overfitting in training, and fault classification is realized through the second full connection layer in the CNN. To deal with human errors in data labeling when constructing the data sets artificially, wrong labeled data sets is constructed to reduce the influence of wrong labeled data on training process. The test results show that the accuracy of the test set is 97.92% when the SNR is 40 dB and this diagnosis method is still effective in a noisy environment.