【作者】 陈俊；

【导师】 丁煦； 王松；

【作者基本信息】 合肥工业大学 ， 计算机技术（专业学位）， 2025， 硕士

【摘要】 近年来深度学习在智能故障诊断领域广泛应用。然而,基于数据驱动的深度学习模型通常是黑盒结构,可解释性不足。并在样本分布漂移的跨样本/跨部件故障诊断中性能不能保持。因果推理作为可解释性建模工具可以通过建立因果模型发现数据中的因果不变性,提高可解释性。并且可以通过干预调整的方式估计真实因果效应,提升泛化能力。尽管如此,因果发现会由于条件独立测试(CI Test)所用的条件集过大而遭受维数灾影响,且由于部分因果结构具有相同的条件独立性导致其获取的马尔可夫等价类(MEC)含有较多未定向边。此外,故障数据中的蕴含的混淆变量可能导致混淆关系,从而让注意力模块关注到虚假的相关性,破坏模型的泛化能力。针对以上问题,本课题进行如下研究:(1)针对复杂故障系统因果发现困难的问题,本研究首先引入迭代因果发现,通过当前迭代的轮次来限制条件集的大小与其离测试节点的距离,以减少所需的CI Test数量。然后,介绍因果方向准则,其利用MEC中蕴含的因果不对称性推断成对节点之间因果边方向。最后,将因果方向准则引入迭代因果发现构建扩展迭代因果发现(EICD)算法,加速迭代因果发现并且相比MEC额外定向因果边。在实验部分和多个对比算法比较所需CI Test的数量以及定向精度的提升。并使用EICD算法从故障数据中恢复故障系统的底层因果图。实验验证,EICD算法比其它算法能够减少所需CI Test数量,并额外定向10%～20%的因果边。(2)针对故障数据中混淆变量导致的虚假相关性问题,本研究提出时序因果注意力深度神经网络(TS-CATT-N),该模型首先引入时间序列令牌器对振动数据进行等效一维卷积操作,从而让一维时序数据直接作为模型输入,并更好提取其故障特征。然后,使用前门调整来消除混淆关系,从而估计真实的因果效应。具体来说,引入因果注意力机制,通过样本内注意力和跨样本注意力的组合来模拟因果干预,提升诊断模型的泛化能力。在实验部分首先利用分布内故障数据集训练TS-CATT-N和其它对比模型,并使用分布外样本对各类模型进行测试,实验证明TS-CATT-N拥有良好的分布内诊断性能,且保持着较好的分布外泛化性能。最后使用消融实验来验证时间序列令牌器和因果注意力模块的有效性。更多还原

【Abstract】 In recent years,deep learning has been widely applied in the field of intelligent fault diagnosis.However,data-driven deep learning models are typically black-box structures with insufficient interpretability,and their performance often degrades in cross-sample or cross-component fault diagnosis under sample distribution shifts.Causal reasoning,as an interpretable modeling tool,can uncover causal invariance in data by establishing causal models,thereby improving interpretability.Additionally,it can estimate true causal effects through intervention adjustments,enhancing generalization capabilities.Nevertheless,causal discovery may suffer from the curse of dimensionality due to excessively large conditional sets used in conditional independence(CI)tests.Moreover,some causal structures share the same conditional independence properties,leading to Markov equivalence class(MEC)with numerous unoriented edges.Furthermore,unobserved confounding variables in fault data may introduce spurious correlations,causing attention mechanisms to focus on misleading dependencies and undermining the model’s generalization ability.To address these challenges,this study conducts the following research:(1)To tackle the challenge of causal discovery in complex fault systems,this study first introduces iterative causal discovery,which limits the size of the conditional set and its distance from the test node based on the current iteration round,thereby reducing the number of required CI tests.Next,a causal direction criterion is introduced,leveraging the causal asymmetry within the MEC to infer the direction of causal edges between node pairs.Finally,this study integrates the causal direction criterion into iterative causal discovery to construct the Extended Iterative Causal Discovery(EICD)algorithm,accelerating the process and orienting additional causal edges beyond the MEC.In the experimental section,the proposed method is compared with multiple baseline algorithms in terms of the number of required CI tests and orientation accuracy.The EICD algorithm is then used to recover the underlying causal graph of the fault system from fault data.Experiments demonstrate that EICD reduces the number of required CI tests while orienting 10%～20%more causal edges than other methods.(2)To address the issue of spurious correlations caused by confounders in fault data,this study proposes the Temporal Causal Attention Deep Neural Network(TS-CATT-N).This model first introduces a time-series tokenizer to perform equivalent 1D convolution operations on vibration data,enabling raw 1D time-series data to be directly fed into the model while better extracting fault features.Then,front-door adjustment is employed to eliminate confounding effects and estimate the true causal impact.Specifically,a causal attention mechanism is introduced,combining within-sample attention and cross-sample attention to simulate causal interventions,thereby improving the diagnostic model’s generalization ability.In the experimental section,TS-CATT-N and baseline models are first trained on in-distribution fault datasets and then tested on out-of-distribution samples.The results show that TS-CATT-N achieves strong in-distribution diagnostic performance while maintaining superior out-of-distribution generalization.Finally,ablation studies are conducted to validate the effectiveness of the time-series tokenizer and causal attention module.更多还原