【作者】 刘霞；

【导师】 高金吉；

【作者基本信息】 北京化工大学 ， 控制科学与工程， 2023， 硕士

【摘要】 滚动轴承是旋转机械中常见的重要部件,对机械系统的正常运转有着重要作用。在实际工业环境中,滚动轴承零部件的复合故障经常发生,但传统的单故障诊断模型无法对复合故障实现解耦诊断。Transformer是一种新兴的深度学习模型,在多个领域的多个任务上实现了更优的效果,是一种特征提取能力更强的高泛化性模型。因此,本论文基于Transformer模型,以滚动轴承复合故障的解耦诊断为研究内容,主要完成了以下工作:(1)针对现有研究中滚动轴承复合故障数据集的复合故障类别不足的问题,本论文设计了滚柱轴承的局部故障加工形式及实验方案,获得了3个转速工况下的包含多类单故障和复合故障的轴承故障数据集,为充分验证后续研究提出的解耦诊断模型的性能奠定了数据基础。(2)针对滚动轴承的复合故障解耦诊断问题,本论文提出了解耦Transformer编码器模型DTE(Decoupling Transformer Encoder)以及对应的多通道信号融合模型DTECF(Decoupling Transformer Encoder with Convolutional Fusion)。这两个模型在滚动轴承复合故障数据集和齿轮-滚动轴承复合故障数据集上分别实现了96.96%/99.32%和100%/99.89%的高准确率的解耦诊断,有效地解决了传统的单一故障诊断模型无法对复合故障解耦的问题。(3)针对DTE模型和DTECF模型存在的泛化稳定性差和训练时间长的问题,本论文提出了一种名为解耦Transformer的模型(简称DT,Decoupling Transformer),并提出了对应的多通道信号融合模型DTCF(Decoupling Transformer with Convolutional Fusion)。这两个模型在滚动轴承复合故障数据集和齿轮-滚动轴承复合故障数据集上分别实现了99.49%/99.16%和99.89/99.52%的高准确率的解耦诊断。与基于CNN和MLP的复合故障解耦诊断模型相比,DT和DTCF模型的抗噪鲁棒性和少样本下的泛化性能都更加优异。此外,它们属于生成式诊断模型,无需人为设定经验阈值,更智能和自动化。本论文还对模型的自注意力权值和互注意力权值进行了可视化分析,初步探索了模型的可解释性。(4)针对变工况下复合故障解耦诊断的问题,本论文提出了一种带有工况辨别器的解耦迁移Transformer模型,称为DTTWCR(Decoupling Transfer Transformer with Working Condition Recognizer)。我们使用3种不同转速下的轴承复合故障数据,设计了6个迁移任务进行实验验证。实验结果表明,所提出的模型能够有效地分离工况相关特征和故障相关特征,从而实现了较好的迁移诊断效果。更多还原

【Abstract】 Rolling bearings are common and important components in rotating machinery,playing a vital role in the normal operation of mechanical systems.In practical industrial environments,compound faults of rolling bearing components often occur,but traditional single-fault diagnosis models cannot achieve decoupling diagnosis of compound faults.Transformer is an emerging deep learning model that has achieved superior performance in multiple tasks across various fields.It is a highly generalized model with stronger feature extraction ability.Therefore,this paper focuses on the decoupling diagnosis of rolling bearing compound faults based on the Transformer model,and mainly completes the following work:(1)To address the problem of insufficient compound fault categories in existing research datasets for rolling bearing compound faults,this paper designs the local fault processing form and experimental scheme of rolling cylindrical bearings,and obtains a bearing fault dataset that includes multiple single faults and compound faults under three speed conditions.This provides a data foundation for fully verifying the performance of the subsequent decoupling diagnosis model.(2)To address the decoupling diagnosis problem of compound faults in rolling bearings,this paper proposes a decoupling Transformer encoder model(DTE)and its corresponding multi-channel signal fusion model with convolutional fusion(DTECF).These two models achieve high accuracy decoupling diagnosis of 96.96%/99.32% and 100%/99.89% on the rolling bearing compound fault dataset and the gear-rolling bearing compound fault dataset respectively,effectively solving the problem that traditional single-fault diagnosis models cannot decouple compound faults.(3)To address the problem of poor generalization stability and long training time of the DTE and DTECF models,this paper proposes a decoupling Transformer model(DT)and its corresponding multi-channel signal fusion model with convolutional fusion(DTCF).These two models exhibit high accuracy decoupling diagnosis of 99.49%/99.16 and99.89/99.52% on the rolling bearing compound fault dataset and the gear-rolling bearing compound fault dataset respectively.Compared with compound fault decoupling diagnosis models based on CNN and MLP,the DT and DTCF models have better noise robustness and generalization performance under fewer samples.Moreover,they belong to generative diagnosis models and do not require manually setting the empirical threshold,making them more intelligent and automated.This paper also visualizes the self-attention weights and cross-attention weights of the model to explore its interpretability.(4)To address the problem of decoupling diagnosis of compound faults under variable operating conditions,this paper proposes a decoupling transfer Transformer model with a working condition recognizer(DTTWCR).We use bearing compound fault data under three different speeds and design six transfer tasks for experimental verification.The results show that the proposed model can effectively separate the features related to working conditions and fault features,achieving good transfer diagnosis performance.更多还原