SHAP to Shop Floor: Towards Trustworthy Reduced-Order Anomaly Detection for Privacy-Sensitive Manufacturing Environments

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

As artificial intelligence (AI) and machine learning (ML) techniques become increasingly ubiquitous for digital twins in manufacturing, it is essential to create reduced-order models that offer interpretable insights and are compatible with real-time production data. This paper introduces a novel framework for explainable AI (XAI) integration in industrial anomaly detection scenarios, in which we create reduced-order models that simultaneously serve as auditors that effectively identify instances of shortcut learning in existing ML-based anomaly detection pipelines. We address two realistic constraints: privacy-encoded inputs and limited or unlabeled data. We validate our approach on two real-world case studies in manufacturing. The first examines anomaly detection for a high storage system, where we find that utilizing simple dimensionality reduction techniques offers resilience to randomized nonlinear transformations used for privacy encoding. The second case study focuses on deriving reduced-order explanations for multiple anomaly conditions in a high-dimensional pick-and-place process. We demonstrate the effectiveness of an approach that combines XGBoost for anomaly detection with TreeSHAP explanations, which enables the formation of interpretable anomaly clusters characterized by 1-2 term rules based on measurable input features. In a high-dimensional and privacy-encoded case study featuring over 130 time series and overlapping anomaly conditions, our methodology is able to describe 9 out of 14 discovered anomaly clusters with a precision of at least 0.80. Importantly, our findings align with real-world insights, showcasing the utility of the approach for auditing conventional ML models and knowledge discovery in manufacturing and anomaly detection systems.