Defect detection in complex chemical processes is a key challenge for industrial safety and efficiency. Researchers Georgios Glavanis, Dimitrios Kyriakou and Spyros Voutetakis, in collaboration with colleagues from the Department of Information and Electronics Engineering and the Institute of Chemical Processing and Energy Resources at the International Hellenic University in Greece, and Simila Papadopoulou and Konstantinos Diamantaras from the Faculty of Business Engineering and Management, have presented a new approach to increasing the transparency of fault diagnosis systems. Their research focuses on applying and comparing state-of-the-art explainable artificial intelligence (XAI) methods, integrated gradients and SHAP, to interpret decisions made by a high-precision long-short-time memory (LSTM) classifier trained on a benchmark Tennessee Eastman process. This study is important because it demonstrates how XAI can not only identify failures, but also identify the specific process subsystems responsible, providing important insights for improving process control and preventive maintenance.
Improved fault detection in chemical plants promises safer, more efficient production and reduced environmental impact. understanding why Having the system flag an error is just as important as detecting it, allowing engineers to quickly identify and resolve the problem. New technologies offer a way to open the complex process monitoring “black box” and reveal the underlying reasons for failure diagnosis.
Scientists are increasingly deploying deep learning models to optimize operations within industrial facilities, covering areas such as quality assurance, production variable prediction, and fault detection. Challenges remain regarding the transparency of these models and the reliability of the insights they generate for end users.
To address this, a growing field known as explainable artificial intelligence (XAI) has emerged, aimed at understanding the decision-making processes of complex algorithms. Although XAI has been widely studied in image classification and natural language processing, its application to deep learning models that process multivariate time series data, especially in chemical processes, has received relatively little attention.
Research focuses on increasing the reliability of fault detection systems in complex industrial environments. understanding why The ability of the classifier to arrive at a specific decision is central to increasing the reliability of the prediction. Unlike previous studies that focused on easily detectable failures, this study delves into more ambiguous scenarios.
This study aims to establish a link between the machine learning model’s decisions and the physical interpretation of each failure, and to verify the validity of the explanations provided. The central questions guiding the research are whether the XAI method consistently agrees on the most important variables and whether the resulting descriptions are consistent with established knowledge of the chemical processes themselves.
By comparing IG and SHAP, the researchers aimed to determine which method provides more useful insight into the root causes of failure. This classifier served as the basis for applying explainability techniques and allowed insights into the decision-making process. TEP was chosen because of its established complexity and availability, providing a realistic scenario for fault diagnostic studies.
Training involves feeding the LSTM classifier with historical process data, which allows the LSTM classifier to learn patterns that indicate normal behavior and deviations that indicate anomalies. Then, two state-of-the-art eXplainability (XAI) methods, Integrated Gradients (IG) and SHapley Additive exPlanations (SHAP), were implemented to interpret the output of the LSTM classifier.
IG calculates the slopes of the output with respect to the input features and assigns importance based on the integral of these slopes along the path from the baseline input. Conversely, SHAP uses game theory concepts to assign each feature a value representing its contribution to prediction. By applying both methods, we were able to comparatively analyze the advantages and disadvantages of each in this particular application.
These model-independent XAI techniques were intentionally chosen to provide flexibility beyond LSTM classifiers. Unlike techniques tied to specific model architectures, IG and SHAP can be applied to trained machine learning models, broadening the potential impact of this work. Data from TEP containing a set of process variables was fed to both XAI methods along with a trained LSTM classifier.
This configuration allows us to generate feature importance scores for each failure scenario. To test the validity of the explanation, the identified key features were compared with established knowledge of TEP. This process is well documented, allowing researchers to assess whether the XAI technique accurately identified the subsystems and variables most affected by each failure.
The research team focused on ambiguous failure scenarios that are challenging for standard failure detection frameworks. Such cases pose significant challenges for XAI techniques and require high accuracy in determining the root cause of failures. Integrated Gradients (IG) and SHapley Additive exPlanations (SHAP) were compared, and both methods were used to interpret decisions made by a high-accuracy LSTM classifier trained on the TEP benchmark.
Initial evaluation focused on identifying which process variables had the most impact on fault classification. In many cases, XAI methods converge on similar key features, suggesting consistent reasoning behind the model’s predictions. However, when considering more ambiguous failure scenarios, inconsistencies emerged. Here, the SHAP method often provided insights that more closely aligned with the underlying physical cause of the failure, as understood by process experts.
For example, when diagnosing a specific fault, SHAP consistently highlighted variables directly related to the failure mechanism, whereas IG sometimes indicated factors that were less directly relevant. In some cases, the differences in feature attribution between the two methods were large, indicating that SHAP may improve diagnostic accuracy. Assessing the plausibility of explanations for known process behavior was a central aspect of this work.
The researchers verified the reliability of the model’s inferences by comparing the significance of the features obtained from XAI with established chemical engineering principles. The study then focused on ambiguous results and discussed the determination of fault detection and diagnostic systems applied to complex chemical processes. This study was not limited to easily detectable disorders. Instead, we focused on scenarios that are difficult to diagnose accurately.
In addition to identifying important features, this study also investigated the concordance between the IG and SHAP methods. The results showed that although there was often overlap in the identified variables, the magnitude of their importance scores could differ. The researchers found that the two methods often agreed on the most important variables, but the degree of agreement varied depending on the specific disorder. By comparing two XAI methods, this study provides a means to establish trust between end users and machine learning model decisions.
Explaining machine learning decisions improves fault diagnosis in complex industrial processes
Scientists are increasingly relying on complex machine learning models to monitor and control critical infrastructure. why Getting these models to make concrete decisions remains a major hurdle. This research addresses this challenge by applying explainable AI techniques to the notoriously difficult problem of chemical plant fault diagnosis.
For many years, identifying the root cause of process anomalies has required specialized knowledge and painstaking analysis, often relying on human operators to interpret vast streams of sensor data. Automated systems offer speed and scale, but they lack transparency, making you hesitant to fully trust their results. A comparison of two prominent explanatory techniques, integrated gradients and SHAP, reveals subtle differences in their ability to accurately identify the source of problems within a simulated chemical plant.
Although both techniques successfully highlighted relevant process variables, SHAP may provide a more accurate relationship to the underlying failure. This effort is part of a broader movement to build trust in AI systems used in safety-critical applications. Further research should focus on developing objective metrics to assess the quality of explanations, independent of human interpretation. This could enable widespread adoption of these tools, not just in chemical engineering, but in any field where opaque AI systems make high-stakes decisions.
