Combining SHAP and Causal Analysis for Interpretable Fault Detection in Industrial Processes
Combining SHAP and causal discovery algorithms to enhance fault detection accuracy and interpretability in industrial processes.
Key Findings
Methodology
This study introduces a fault detection framework integrating SHAP-based feature selection with multiple causal discovery algorithms (PC, FCI, RFCI, LINGAM, NOTEARS). Initially, SHAP values identify the most influential features, reducing dimensionality. Subsequently, these features serve as inputs for causal graph construction, revealing underlying fault propagation mechanisms. The approach achieves high detection accuracy (up to 98%) on the Tennessee Eastman Process dataset, while providing causal explanations aligned with key process variables, thus bridging predictive performance and interpretability.
Key Results
- On the TEP dataset, the combined SHAP and causal analysis approach achieved 98% fault detection accuracy, outperforming traditional models at 92%. The causal graphs consistently highlighted cooling and separation systems as critical fault sources, matching SHAP's key features. Feature selection reduced model complexity by 30%, improving computational efficiency.
- Five causal algorithms produced highly consistent causal structures, validating the robustness of identified relationships. Incorporating causal insights reduced fault classification errors by 15%, demonstrating improved reliability across fault types.
- The integrated framework provided actionable insights, enabling operators to understand fault origins clearly, facilitating targeted maintenance and control adjustments.
Significance
This work advances industrial fault diagnosis by combining predictive modeling with causal reasoning, addressing the opacity of black-box models. It enhances trustworthiness and provides actionable root-cause insights, crucial for safety and efficiency in industrial environments. The framework supports Industry 4.0 initiatives by enabling transparent, data-driven decision-making, thus fostering smarter manufacturing systems.
Technical Contribution
The primary innovation lies in the seamless integration of SHAP feature importance with multiple causal discovery algorithms to produce reliable causal graphs. This fusion improves feature selection robustness and causal structure stability. The approach extends existing methods by validating causal relationships across algorithms, providing a systematic way to interpret complex industrial data, and improving fault detection performance with explainability.
Novelty
This is the first work to combine SHAP-based feature importance with multi-algorithm causal discovery for industrial fault detection. Unlike prior studies focusing solely on prediction accuracy or causal explanation separately, this framework unites both aspects, offering a comprehensive, interpretable solution that enhances trust and usability in industrial settings.
Limitations
- The accuracy of causal graphs depends on the quality of data; unmeasured confounders or noisy data can distort causal inference. Future work should incorporate temporal causal models to handle dynamic processes.
- The approach's computational cost may limit real-time deployment in large-scale industrial systems. Optimization and parallelization are needed for practical use.
- The current framework assumes static causal relationships; in real factories, causal links may evolve, requiring adaptive or online causal inference methods.
Future Work
Future research will focus on integrating time-series causal models to capture dynamic fault propagation. Combining deep learning with causal inference may improve performance in highly nonlinear environments. Additionally, deploying the framework in real industrial plants will validate its practical utility and robustness under operational conditions.
AI Executive Summary
Industrial process monitoring faces significant challenges due to high-dimensional, nonlinear, and often opaque data. Traditional fault detection methods like PCA or simple statistical models struggle to provide both accurate detection and clear explanations, limiting their practical utility. Recent advances in machine learning have improved detection performance, but their black-box nature hampers trust and root cause analysis.
This study proposes a novel framework that synergistically combines SHAP (SHapley Additive exPlanations) with multiple causal discovery algorithms (PC, FCI, RFCI, LINGAM, NOTEARS). The core idea is to first use SHAP values to identify the most influential process variables, reducing the feature space and enhancing interpretability. Then, these key features serve as inputs to causal algorithms that construct directed acyclic graphs (DAGs), revealing the underlying causal structure of fault propagation pathways.
Experimental validation on the Tennessee Eastman Process dataset demonstrates the effectiveness of this approach. The combined method achieves a fault detection accuracy of 98%, surpassing traditional models at 92%. The causal graphs consistently highlight critical process components like cooling and separation systems, aligning with domain knowledge and providing actionable insights. The feature selection reduces model complexity by approximately 30%, leading to faster computation and easier deployment.
This integrated approach bridges the gap between predictive power and causal understanding, offering industrial engineers a powerful tool for both accurate fault detection and root cause analysis. It enhances operational safety, reduces downtime, and supports smarter maintenance strategies. The framework’s robustness across multiple causal algorithms further ensures reliability in complex industrial environments.
Looking ahead, future work aims to incorporate temporal causal models to handle dynamic process changes, optimize computational efficiency for real-time applications, and validate the framework in real-world industrial settings. Overall, this work marks a significant step toward transparent, reliable, and actionable fault diagnosis in Industry 4.0 manufacturing systems.
Deep Analysis
Background
随着工业自动化和智能制造的推进，故障检测成为保障生产安全与效率的核心环节。早期方法多依赖统计分析技术如PCA和PLS，但难以捕获复杂非线性关系。近年来，机器学习模型如深度神经网络和XGBoost显著提升了检测性能，但缺乏可解释性，限制了实际应用。SHAP作为一种模型解释工具，逐渐被引入工业场景，改善了模型透明度。同时，因果推断算法如PC、FCI等被用以揭示变量间的因果关系，为故障根源诊断提供理论基础。尽管如此，缺乏将特征重要性与因果关系结合的系统框架，限制了其在复杂工业环境中的应用潜力。本文试图弥补这一空白，通过结合SHAP与多算法因果推断，推动工业故障诊断的智能化与可解释性发展。
Core Problem
工业过程数据高维且复杂，故障类型多样，传统模型难以兼顾检测准确性与解释性。黑箱模型虽能实现高性能，但无法提供故障根源信息，限制了操作人员的信任与干预能力。因果关系的不透明也阻碍了对故障机制的深入理解。如何在保证检测性能的同时，提供清晰的因果解释，成为当前工业故障诊断的核心难题。特别是在实际工业场景中，变量间的复杂交互和潜在干扰使得因果推断具有较大挑战。解决这一问题，需结合模型的预测能力与因果分析的解释性，构建既精准又透明的故障检测体系。
Innovation
本研究的创新点包括：1）将SHAP的特征贡献分析与多算法因果结构发现结合，形成端到端的可解释故障检测流程；2）利用SHAP筛选出影响最大的变量，降低模型复杂度，提升效率；3）采用多算法验证因果关系的稳定性，增强结构可信度；4）在TEP数据集上实现98%的检测准确率，优于传统方法。这一方法突破了以往单一模型或统计分析的局限，为工业故障诊断提供了全新思路。
Methodology
- �� 数据预处理：采用SMOTE进行少数类过采样，随机欠采样平衡数据。• 特征筛选：利用SHAP值评估变量贡献，筛选出前10个关键变量。• 因果推断：应用PC、FCI、RFCI、LINGAM、NOTEARS算法，从筛选变量中构建因果图。• 模型训练：在筛选变量基础上训练MLP和XGBoost模型进行故障分类。• 解释分析：结合SHAP值和因果图，识别故障传播路径和关键因素。• 性能评估：采用准确率、误差和结构一致性指标验证模型效果。
Experiments
在TEP数据集上，设计多场景故障测试，比较不同模型和特征筛选策略。通过交叉验证优化超参数如学习率、树深等。引入多算法验证因果结构的稳定性。结果显示，结合SHAP和因果分析的模型在准确率和可解释性方面优于单一方法。特征子集消融实验验证筛选变量的有效性。整体性能表现出极强的鲁棒性和泛化能力，验证了方法的实用性。
Results
结合SHAP筛选的特征使模型准确率达98%，优于未筛选模型的92%。因果图揭示冷却系统和分离器为关键故障源，与SHAP重点变量一致。多算法因果推断结果高度一致，验证了因果关系的稳定性。模型在多类故障检测中的误差降低了15%，显示出优越的泛化能力。整体上，该方法实现了性能和解释的双重提升。
Applications
该框架适用于化工、制造等行业的复杂工业流程监控。通过筛选关键传感器变量，减少硬件成本，提高监控效率。结合因果分析，帮助操作员理解故障机制，制定精准维护策略。未来可集成到工业控制系统，实现自主故障诊断与预警，推动智能制造升级。
Limitations & Outlook
模型对因果推断的依赖较大，若数据中存在未测变量或干扰，可能影响结构的正确性。复杂环境下，因果关系可能随时间变化，需引入动态因果模型。高维数据和噪声仍是挑战，未来需优化算法以增强鲁棒性。实际部署时，计算成本和实时性也是限制因素，需进一步提升效率。
Plain Language Accessible to non-experts
想象你在管理一个大型工厂，里面有许多机器和传感器。每个传感器都在监测不同的参数，比如温度、压力、流量等。有时候，工厂会出现故障，比如设备过热或泄漏。传统的方法就像用放大镜盯着每个参数，试图找出异常，但很难理解为什么会出问题，也难以找到根源。
这项研究就像给你配备了一套聪明的工具箱。第一个工具是“重要性分析”，它告诉你哪些参数最可能导致故障，就像医生用血压和心跳告诉你身体哪里出了问题。第二个工具是“因果图”，它像一张地图，显示这些参数之间的关系，帮你追踪故障的“源头”。
通过结合这两种工具，你不仅能准确发现故障，还能理解为什么会出问题。比如，发现冷却系统和分离器是关键环节，一旦出问题，整个工厂就会出事。这就像你用地图找到问题的“源头”，然后及时修理，避免更大的损失。这种方法让工厂变得更聪明、更安全，也让操作员更有信心应对突发状况。
ELI14 Explained like you're 14
想象你在学校里，有很多不同的班级和学生。老师想知道哪个学生出了问题，比如学习成绩突然变差。以前，老师会盯着所有学生，试图找到原因，但这样很费时间，也不一定能找到真正的问题所在。
现在，假如老师有两个神奇的工具：一个叫“重要性分析”，它能告诉老师哪些学生的表现变化最可能导致班级成绩下降；另一个叫“因果图”，它像一张关系图，显示学生之间的关系，帮助老师追踪问题的“源头”。
用这两个工具，老师不仅能快速发现哪个学生出了问题，还能理解为什么会出问题，比如是因为某个学生受到朋友的影响，或者因为家庭作业没做。这就像用地图找到问题的“根源”，然后帮学生解决问题。这样，老师可以更有针对性地帮助学生，整个班级的学习也会变得更好。这种方法让我们更聪明、更懂得解决问题，就像超级侦探一样！
Abstract
Industrial processes generate complex data that challenge fault detection systems, often yielding opaque or underwhelming results despite advanced machine learning techniques. This study tackles such difficulties using the Tennessee Eastman Process, a well-established benchmark known for its intricate dynamics, to develop an innovative fault detection framework. Initial attempts with standard models revealed limitations in both performance and interpretability, prompting a shift toward a more tractable approach. By employing SHAP (SHapley Additive exPlanations), we transform the problem into a more manageable and transparent form, pinpointing the most critical process features driving fault predictions. This reduction in complexity unlocks the ability to apply causal analysis through Directed Acyclic Graphs, generated by multiple algorithms, to uncover the underlying mechanisms of fault propagation. The resulting causal structures align strikingly with SHAP findings, consistently highlighting key process elements-like cooling and separation systems-as pivotal to fault development. Together, these methods not only enhance detection accuracy but also provide operators with clear, actionable insights into fault origins, a synergy that, to our knowledge, has not been previously explored in this context. This dual approach bridges predictive power with causal understanding, offering a robust tool for monitoring complex manufacturing environments and paving the way for smarter, more interpretable fault detection in industrial systems.