基于迁移学习的高速列车轴承故障智能诊断建模研究

Research on Intelligent Fault Diagnosis Modeling of High-Speed Train Bearings Based on Transfer Learning

摘要: 高速列车轴承作为走行系统的核心旋转部件，其运行状态直接关系到列车的安全性与可靠性。长期处于高转速、交变载荷等复杂工况下，轴承表现出故障率高、易损坏等显著特点，已成为高速列车安全运维的关键监控对象。轴承故障若未能被及时准确诊断，可能引发运行延误甚至脱轨等灾难性后果，因此对其开展精准、鲁棒的故障诊断研究具有重大工程安全价值。传统故障诊断方法多依赖于专家先验知识或传统信号处理技术，在诊断精度、泛化能力及实时性方面难以满足复杂运营场景下的高标准需求。近年来，基于深度学习的智能诊断方法凭借其强大的特征学习能力展现了巨大潜力，然而在实际工程应用中，仍面临两大挑战：一是振动信号易受强噪声干扰，导致故障特征被淹没；二是故障样本稀缺导致的类间样本不平衡问题，限制了有监督学习模型的泛化性能。值得注意的是，与稀缺且获取成本高昂的真实运行故障数据形成鲜明对比的是，在受控的台架实验环境中，研究人员能够方便地获取大量标签完备、覆盖各种故障类型与严重程度的轴承数据，且其物理失效机理与真实列车轴承具有高度相似性。如何有效利用这些丰富的实验室数据来弥补实际运维中故障数据的不足，成为破局的关键。在此背景下，迁移学习技术作为一种能够将在一个领域(源域，如实验室台架)中学到的知识迁移应用到另一个相关但数据分布不同的领域(目标域，如在线运营列车)的机器学习范式，为解决上述数据稀缺与分布失衡的困境提供了极具前景的思路。本研究正是在这一现实需求与技术机遇的交汇点上展开，旨在探索基于迁移学习的解决方案，以推动高速列车轴承智能故障诊断技术走向真正的工程实用化。

Abstract: As the core rotating component of the running gear system for high-speed trains, the operating status of train bearings is directly related to the safety and reliability of the train. Operating for long periods under complex conditions such as high rotational speeds and alternating loads, bearings exhibit prominent characteristics, including a high failure rate and vulnerability to damage, making them a key monitoring target for the safe operation and maintenance of high-speed trains. Failure to diagnose bearing faults in a timely and accurate manner may lead to catastrophic consequences such as operational delays or even derailment. Therefore, conducting research on precise and robust fault diagnosis for such bearings is of great significance for engineering safety. Traditional fault diagnosis methods mostly rely on expert prior knowledge or conventional signal processing techniques, which struggle to meet the high-standard requirements of complex operational scenarios in terms of diagnostic accuracy, generalization ability and real-time performance. In recent years, intelligent diagnosis methods based on deep learning have shown tremendous potential owing to their powerful feature learning capabilities. However, in practical engineering applications, they still face two major challenges: first, vibration signals are easily disturbed by strong noise, resulting in the obscuration of fault features; second, the inter-class sample imbalance caused by scarce fault samples limits the generalization performance of supervised learning models. Notably, in sharp contrast to the scarce and costly real-world operational fault data, researchers can easily obtain a large amount of fully labeled bearing data covering various fault types and severity levels in a controlled bench test environment. Moreover, the physical failure mechanism of such test bearings is highly similar to that of actual train bearings. How to effectively utilize this abundant laboratory data to compensate for the shortage of fault data in actual operation and maintenance has become the key to breaking through the dilemma. Against this background, transfer learning—a machine learning paradigm that can transfer knowledge learned from one domain (source domain, e.g., laboratory test bench) to another related domain with different data distribution (target domain, e.g., online operating trains)—provides a promising solution to the above dilemmas of data scarcity and distribution imbalance. This research is carried out at the intersection of such practical demands and technological opportunities, aiming to explore transfer learning-based solutions and promote the engineering practical application of intelligent fault diagnosis technology for high-speed train bearings.