基于迁移学习的滚动轴承故障诊断方法研究

Research on Rolling Bearing Fault Diagnosis Method Based on Transfer Learning

摘要: 高速列车轴承智能故障诊断是保障轨道交通运营安全的关键技术。随着我国“八纵八横”高铁网络的快速扩展，列车运行密度持续加大，轴承故障预警面临新的挑战：一方面，台架实验数据与真实运行数据存在显著分布差异；另一方面，故障样本稀缺导致传统深度学习模型泛化能力不足。在此背景下，迁移学习为解决这类问题提供了有效的路径。本文提出一种基于特征选择与机器学习的跨工况故障诊断方法。针对源域中正常样本极少的类别不平衡问题，采用SMOTE过采样增加样本。在此基础上，对比了随机森林重要性排序、递归特征消除(RFE)和自编码器三种特征选择方法，并结合随机森林、支持向量机与多层感知机(MLP)三种分类器进行性能评估。实验结果表明，RFE-20特征子集与MLP的组合在源域验证集上取得最佳性能，准确率达95.16%。使用该最佳模型对16个无标签的目标域样本进行预测，93.8%的样本预测置信度超过0.9，预测类别涵盖内圈、外圈、滚动体故障及正常状态。

Abstract: Intelligent fault diagnosis of high-speed train bearings is a key technology to ensure the operational safety of rail transit. With the rapid expansion of China’s “Eight Vertical and Eight Horizontal” high-speed railway network, the operation density of trains continues to increase, bringing new challenges to bearing fault early warning. On the one hand, there exists a significant distribution discrepancy between bench test data and real operational data; on the other hand, the scarcity of fault samples results in insufficient generalization ability of traditional deep learning models. Against this background, transfer learning provides an effective solution to address such problems. This paper presents a cross-condition fault diagnosis method based on feature selection and machine learning. To address severe class imbalance in the source domain, SMOTE oversampling is applied. Three feature selection methods—Random Forest importance, RFE, and Autoencoder—are compared, combined with Random Forest, SVM, and MLP classifiers. Results show that the RFE-20 feature subset with MLP achieves the highest source-domain accuracy of 95.16%. The optimal model is validated on 16 unlabeled target-domain train bearing samples, with 93.8% prediction confidence above 0.9, covering inner race, outer race, rolling element faults, and normal conditions.