Explainable AI to Improve Machine Learning Reliability for Industrial Cyber-Physical Systems

TL;DR

Industrial CPS reliability is critical and standard ML metrics alone are insufficient to guarantee safe deployment. This paper uses Explainable AI, specifically C-SHAP with a custom time-series decomposition that yields Levels, Peaks, Scale, LF, HF concepts, to explain and improve a CNN-based fault-detection model trained on phase-based time-series windows. The authors show that enlarging the data window from a small to a larger size increases accuracy, with levels driving predictions and more stable SHAP values as context grows. Public data and code are released to demonstrate a practical, scalable workflow for making CPS ML more reliable and generalizable.

Abstract

Industrial Cyber-Physical Systems (CPS) are sensitive infrastructure from both safety and economics perspectives, making their reliability critically important. Machine Learning (ML), specifically deep learning, is increasingly integrated in industrial CPS, but the inherent complexity of ML models results in non-transparent operation. Rigorous evaluation is needed to prevent models from exhibiting unexpected behaviour on future, unseen data. Explainable AI (XAI) can be used to uncover model reasoning, allowing a more extensive analysis of behaviour. We apply XAI to to improve predictive performance of ML models intended for industrial CPS. We analyse the effects of components from time-series data decomposition on model predictions using SHAP values. Through this method, we observe evidence on the lack of sufficient contextual information during model training. By increasing the window size of data instances, informed by the XAI findings, we are able to improve model performance.

Figures (6)

• Figure 1: The experimental setup for machine trace collection and the data processing steps leading to a labelled dataset, suitable for training a CNN model.
• Figure 2: C-SHAP workflow, including concept construction from signal data, masking and calculation of C-SHAP, i.e., attribution scores.
• Figure 3: Global explanation in the form of the mean absolute SHAP values over the test data with respect to the concepts. The SHAP values for the ground truth class of each window are used.
• Figure 4: Change in SHAP values for the test data when increasing the window size from 100 to 400. The change is shown as the mean and standard deviation of the difference across windows.
• Figure 5: The predictions and SHAP values for two selected example segments from the inference data. a), c) and e) correspond to segments of the same NoFan sample, while b), d) and f) correspond to segments of the same Normal sample. a) and b) correspond to the results for window size 100, while c) and d) correspond to size 200, and e) and f) correspond to 400. In the top rows of the images, the signal is shown along with the predicted classes as coloured overlays, marking the associated windows. For the NoFan sample, the purple windows are correct, and for the Normal sample, the white windows are correct. For each component, the component itself is plotted, combined with the SHAP values for each window. For c)-f) only the ""*Levels component is shown for clarity.