【作者】 吴昊；

【导师】 丁煦；

【作者基本信息】 合肥工业大学 ， 机械设计及理论， 2025， 硕士

【摘要】 随着工业自动化与智能技术的深度融合,机械设备的运维模式正经历着从传统人工检测向智能感知决策的转型。滚动轴承作为复杂机电设备的核心传动部件,广泛应用于轨道交通和航空航天等高端装备领域,其运转状态直接影响到机械设备的安全服役性能。因此,构建高效轴承故障诊断体系对于保障重大装备全生命周期稳定运行具有重大价值。然而,由于滚动轴承的运行受机械结构和环境工况的影响,当前轴承智能故障诊断方法仍面临着诸多挑战,主要表现在以下两个方面:(1)数据驱动的轴承智能故障诊断方法缺乏对轴承故障数据中潜在因果关系分析,仅聚焦数据间的虚假相关性导致轴承智能故障诊断方法在处理缺陷轴承故障数据时,诊断结果的可靠性与准确性受到较大影响。(2)现有基于因果理论优化故障诊断模型的方法更多的是聚焦于解耦混杂特征,缺少针对模型决策过程中各个特征间因果效应的定量分析,限制了故障诊断方法性能的提升。因此,针对以上问题,本文的研究工作如下:首先,为深入挖掘轴承故障诊断过程中各变量间的潜在因果关系,获取具备因果特性的高质量故障数据,本文搭建一个可实现多源信息采集与监测的滚动轴承故障诊断实验台。在此基础上,设计并开发了配套的运行监控系统。系统通过OPC UA协议与下位机进行通信,实现对下位机中变量的读写和监控,有效支撑了面向智能化运维的研究工作。其次,本文提出一种基于信息流与因果效应修正的因果发现方法。该方法利用缺失因果图表征轴承数据缺失机制,立足于因果效应修正标准与信息流规律,构建轴承数据中潜在的因果结构。整体因果发现过程包括三个阶段:粗略因果骨架阶段、精炼因果骨架阶段以及因果定向阶段。通过合成数据以及真实轴承数据的实验验证表明所提方法具备优异的因果结构模型构建性能。最后,本文提出一种基于因果生成网络和效应估计的轴承故障诊断方法。该方法引入通用因果发现范式生成不同轴承数据集中因果结构模型,精准刻画不同数据域间的因果关系变化。同时,采用因果效应估计定量分析故障诊断过程中特征间因果效应变化规律。通过大量跨域故障诊断任务验证,所提方法显著提高了跨域故障诊断的准确性。更多还原

【Abstract】 With the deep integration of industrial automation and intelligent technologies,the operation and maintenance paradigm of mechanical equipment is undergoing a transformation from traditional manual inspection to intelligent perception and decision-making.As the core transmission component of complex electromechanical systems,rolling bearings are widely used in high-end equipment such as rail transportation and aerospace.Their operational status directly affects the safe and reliable performance of mechanical systems.Therefore,constructing an efficient bearing fault diagnosis framework is of great significance for ensuring the stable operation of major equipment throughout its life cycle.However,due to the influence of mechanical structure and environmental conditions during bearing operation,current data-driven intelligent fault diagnosis methods for bearings still face multiple challenges,mainly reflected in the following two aspects:(1)These methods often lack in-depth analysis of the potential causal relationships within bearing fault data and instead rely on spurious correlations among data,which compromises the reliability and accuracy of diagnosis results,especially when dealing with defective bearing data.(2)Existing approaches that optimize fault diagnosis models based on causal theory primarily focus on disentangling confounding features,while lacking quantitative analysis of causal effects among features during the model’s decision-making process,thereby limiting the potential for further performance improvement in fault diagnosis methods.To address these issues,this thesis conducts the following research:Firstly,to deeply explore the potential causal relationships among variables in the process of bearing fault diagnosis and to obtain high-quality fault data with causal characteristics,this study constructs a rolling bearing fault diagnosis testbed capable of multi-source information acquisition and monitoring.On this basis,a corresponding operational monitoring system is designed and developed.The system communicates with the lower-level controller via the OPC UA protocol,enabling the reading,writing,and monitoring of variables within the lower-level system,thereby effectively supporting research efforts toward intelligent operation and maintenance.Secondly,this study proposes a causal discovery method based on information flow and causal effect correction.This method employs a missingness causal graph to represent the missing data mechanism in bearing datasets and constructs the underlying causal structure by leveraging causal effect correction criteria and information flow principles.The overall causal discovery process consists of three stages:the rough causal skeleton stage,the refined causal skeleton stage,and the causal orientation stage.Experimental validation on both synthetic and real-world bearing datasets demonstrates that the proposed method exhibits excellent performance in constructing accurate causal structure models.Finally,this study proposes a bearing fault diagnosis method based on causal generative network and effect estimation.The method introduces a general causal discovery framework to construct causal structure models across different bearing datasets,accurately capturing causal relationship shifts between data domains.Meanwhile,causal effect estimation is employed to quantitatively analyze the variation patterns of causal effects among features during the fault diagnosis process.Extensive validation on cross-domain fault diagnosis tasks demonstrates that the proposed method significantly improves diagnostic accuracy in cross-domain scenarios.更多还原