| 林俊亭,牛鹏远.基于DCNN的ZPW-2000A无绝缘轨道电路故障诊断研究[J].电子测量与仪器学报,2024,38(6):171-180 |
| 基于DCNN的ZPW-2000A无绝缘轨道电路故障诊断研究 |
| Research on fault diagnosis of ZPW-2000A jointlesstrack circuit based on DCNN |
| |
| DOI： |
| 中文关键词: 无绝缘轨道电路 深度卷积神经网络 BP神经网络 k折交叉验证 故障诊断 |
| 英文关键词:jointless track circuit deep convolutional neural network BP neural network k-fold cross-validation fault diagnosis |
| 基金项目:国家自然科学基金（52162050）项目资助 |
|
|
|
|
| 摘要点击次数: 45 |
| 全文下载次数: 1296 |
| 中文摘要: |
| 针对ZPW-2000A无绝缘轨道电路故障发生的多样性和不确定性导致的故障诊断效率低的问题，从故障特征提取和故障分类的角度出发，提出一种基于深度卷积神经网络（DCNN）的轨道电路故障诊断方法。通过故障分析总结出12种轨道电路故障状态，并将不同故障状态下的轨道电路监测数据进行标准化处理，作为DCNN模型的输入。模型采用卷积 池化结构提取轨道电路的关键特征并滤除冗余特征。BP神经网络作为模型的全连接层，并结合Softmax函数进行故障分类。通过k折交叉验证法优化模型结构，确定最佳模型。实验结果表明，采用4层卷积 池化层结构的轨道电路故障诊断模型在诊断准确率方面达到了98.48%，较同为最优模型的长短期记忆网络（LSTM）模型、深度前馈网络（DFN）模型、双向长短时记忆网络模型（BiLSTM）与CNN-LSTM组合模型分别提升了6.06%,6.06%,3.33% 与 2.27%，训练收敛速度分别快了大约1 250、4 250、1 250与1 450次，且训练时的损失波动更小。本研究提升了轨道电路故障诊断效率，为轨道电路的故障诊断任务提供了一种新的有效方法。 |
| 英文摘要: |
| Aiming at the low efficiency of fault diagnosis caused by the diversity and uncertainty of fault occurrence in the ZPW-2000A jointless track circuit, this paper proposes a deep convolutional neural network (DCNN)-based fault diagnosis method for jointless track circuit from the perspective of fault feature extraction and fault classification. Twelve rail circuit fault states are summarized through fault analysis, and the monitoring data of rail circuits under different fault states are standardized as inputs to the DCNN model. The model adopts the convolution-pooling structure to extract the key features of the rail circuit and filter out the redundant features. The back propagation neural network (BPNN) is used as the fully connected layer of the model and combined with the Softmax function for fault classification. The model structure is optimized by the k-fold cross-validation method to determine the best model. The experimental results show that the track circuit fault diagnosis model with a four-layer convolution-pooling layer structure achieves 98.48% in diagnosis accuracy, which is 6.06%, 6.06%, 3.33%, and 2.27% higher than the optimal models of long short-term memory (LSTM), deep feedforward network (DFN), bidirectional long and short-term memory (BiLSTM), and the combination of CNN-LSTM, respectively. Additionally, the training convergence speed is about 1 250, 4 250, 1 250, and 1 450 times faster, respectively, with less loss fluctuation during training. This study improves the efficiency of fault diagnosis of rail circuits and provides a new, effective method for the task of fault diagnosis of rail circuits. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
