Hybrid approaches to AI for manufacturing

Modern manufacturing systems generate vast amounts of multivariate time-series data through distributed sensors and automated control systems. Monitoring such environments remains challenging due to evolving operating conditions, complex inter-sensor dependencies, and the scarcity of labeled anomaly data. Consequently, industries continue to rely heavily on Statistical Process Control (SPC) and manual inspection, both of which become increasingly difficult to scale in complex production settings. This study, conducted in collaboration with a leading industrial machinery manufacturer, investigates anomaly detection in manufacturing processes under realistic unlabeled conditions. The goal is to combine interpretable statistical monitoring with data-driven artificial intelligence while maintaining industrial usability and expert trust. A hybrid anomaly detection framework is proposed that integrates residual-based SPC, self-supervised representation learning, and human-in-the-loop validation. SPC serves as an interpretable benchmark, identifying deviations from steady-state operation through residual analysis. To address the limitations of fixed-threshold monitoring, a contextual contrastive learning approach is introduced to learn normal operational behaviour directly from unlabeled multivariate time-series data. Using cycle-aware window extraction and context-target learning, the framework captures process dynamics without requiring fault annotations. Anomalies are detected through embedding-similarity discrepancy scoring and statistically calibrated intervals. Since ground-truth labels are unavailable, detected events are evaluated through industrial visualization dashboards, enabling domain experts to review anomaly candidates and provide operational feedback. Comparative analysis shows that SPC primarily identifies localized signal deviations, whereas the representation-learning approach detects temporally coherent anomaly intervals that better capture sustained behavioural changes. Additional detections during transitional operating regimes reflect the inherent uncertainty of monitoring complex industrial systems rather than model failure. Industrial validation demonstrates that the proposed framework reduces manual inspection effort while preserving human oversight. The results highlight the complementary strengths of statistical monitoring and self-supervised learning, contributing a reliable and scalable methodology for AI-assisted anomaly detection in manufacturing environments where labeled anomaly data are unavailable.