| 基于深度主动学习的MVB网络故障诊断方法 |
| |
| 引用本文： | 杨岳毅,王立德,王冲,王慧珍,李烨.基于深度主动学习的MVB网络故障诊断方法[J].西南交通大学学报,2022,57(6):1342-1348, 1385. |
| |
| 作者姓名： | 杨岳毅 王立德 王冲 王慧珍 李烨 |
| |
| 作者单位： | 北京交通大学电气工程学院，北京 100044 |
| |
| 基金项目： | 中国国家铁路集团有限公司科技研究开发计划（N2020J007） |
| |
| 摘 要： | 多功能车辆总线MVB （multiple vehicle bus）用于传输重要的列车运行控制指令和监视信息，准确地诊断MVB网络故障是列车智能运维的基础，为此，提出一种将主动学习和深度神经网络相结合的MVB网络故障诊断方法. 该方法采用堆叠去噪自编码器自动提取MVB信号物理波形特征，并将该特征用于训练深度神经网络来实现MVB网络故障模式分类；基于不确定性和可信度的高效主动学习方法，可解决实际应用中标记样本不足和人工标记成本高昂的问题，使用少量标记训练样本就能得到高性能的深度神经网络模型. 实验结果表明:为达到90%以上分类准确率，所提方法只需要600个标记训练样本，小于随机采样方法所需标记训练样本数的2 800个；在相同标记训练样本数下，所提方法在3种性能指标下均优于传统方法.
|
| 关 键 词： | 多功能车辆总线 故障诊断 主动学习 深度神经网络 堆叠去噪自编码器 |
| 收稿时间： | 2021-03-17 |
Fault Diagnosis Method Based on Deep Active Learning For MVB Network |
| |
| YANG Yueyi,WANG Lide,WANG Chong,WANG Huizhen,LI Ye.Fault Diagnosis Method Based on Deep Active Learning For MVB Network[J].Journal of Southwest Jiaotong University,2022,57(6):1342-1348, 1385. |
| |
| Authors: | YANG Yueyi WANG Lide WANG Chong WANG Huizhen LI Ye |
| |
| Affiliation: | School of Electrical Engineering, Beijing Jiaotong University, Beijing 100044, China |
| |
| Abstract: | Multiple vehicle bus (MVB) is employed to transmit important train operation control instructions and monitoring information, and accurate diagnosis of the fault types of MVB network is the basis of the intelligent operation and maintenance system. To this end, a fault diagnosis method for MVB network is proposed, which combines the active learning and deep neural networks. It adopts the stacked denoising autoencoder to automatically extract physical features from the electrical MVB signals; then the features are used to train a deep neural network classifier for identifying MVB fault classes. An efficient active learning method based on uncertainty and credibility can solve the problems of insufficient labeled samples and high costs of manual labeling in practical application. It can build a competitive classifier with a small number of labeled training samples. Experiment results demonstrate that to achieve a high accuracy above 90%, the proposed method requires 600 labeled training samples, which is less than 2800 labeled training samples required by random sampling method. With the same number of labeled samples, the proposed method can achieve the better performance as to three different metrics than traditional methods. |
| |
| Keywords: | |
|
| 点击此处可从《西南交通大学学报》浏览原始摘要信息 |
|
点击此处可从《西南交通大学学报》下载全文 |
|
