A Self-Supervised Framework for Anomaly Detection and Predictive Maintenance in Steel Rod Mills

High-speed steel rod mills operate under extreme conditions and demand uninterrupted production. Unplanned equipment failures, such as cobble events where tangled hot steel disrupts operation, cause costly downtime and safety hazards. Predictive maintenance (PdM) aims to anticipate such failures and improve uptime. Traditional PdM methods face challenges due to scarce labeled failure data in these environments.This paper presents a self-supervised anomaly detection framework for steel rod mills that learns normal operational patterns from sensor data and identifies deviations, focusing on cobble detection. We implement an unsupervised Gaussian Mixture Model (GMM) to model the distribution of multivariate sensor readings across mill zones and flag outliers as anomalies. Tests on a wire rod mill with strands B, C, and D show anomalies matching two cobble incidents in strands C and D, offering timely alerts.Analysis by strand and zone reveals over 23 000 detected anomaly instances, with high concentrations in the failed strand’s finishing block. A key challenge is distinguishing idle periods during downtime from true anomalies, since the model is trained solely on active operation data. The framework, requiring no labeled failure data, suits scenarios where faults are rare but critical. We discuss how this approach can complement preventive maintenance to reduce cobble occurrences and downtime, and outline strategies for managing alerts in non-operational states.