Anomaly detection in unsupervised industrial settings, developing human-centric diagnostic and prognostic tools to improve decision-making and transparency

The increasing adoption of sensor-rich, automated production systems under Industry 4.0 and Industry 5.0 frameworks has created both the opportunity and the necessity for intelligent, real-time anomaly detection. In complex industrial environments, unexpected deviations in process data can precede equipment failures, quality defects, and unplanned downtime, yet the rarity and absence of labeled anomalies make conventional supervised approaches largely imprac- tical. This thesis addresses the challenge of developing unsupervised anomaly detection methods that are interpretable, computationally efficient, and aligned with human-centric operational requirements. The research was conducted in collaboration with an industrial bakery company, where multivariate sensor data capturing temperature, mixer speeds, conveyor velocities, and other pro- cess variables were collected across multiple production lines. Three comple- mentary approaches were developed: a statistical threshold-based method for abrupt signal deviations, an Isolation Forest for multivariate anomalies in an un- supervised manner, and a CatBoost regression–residual framework for signal- level predictive monitoring under weak supervision. Methods were integrated with Explainable AI (XAI) techniques, including SHAP and AcME, to provide operators with actionable, human-interpretable insights into detected anoma- lies. AcME emerged as the preferred tool due to its computational efficiency and suitability for near-real-time deployment. The CatBoost framework achieved strong predictive performance with 2 values exceeding 0.96, while the Isola- tion Forest demonstrated complementary strengths in detecting anomalies aris- ing from multivariate interactions. The results show that combining lightweight machine learning models with fast, interpretable XAI methods yields a practical and transferable framework for industrial anomaly detection, supporting human-in-the-loop decision-making without requiring labeled fault data. Limitations and directions for future work are discussed.