Unsupervised regime discovery from industrial multivariate time series using masked invariant information clustering

In industrial process monitoring, inferring the current operational regime is a prerequisite for reliable downstream tasks such as anomaly detection and maintenance planning, because normal signal patterns can vary substantially with operating context. However, discovering discrete regimes from unlabelled, high-dimensional multivariate time series (MVTS) remains challenging, particularly when dominant categorical variables induce shortcut learning and structural missingness precludes reliable interpolation. To address these issues, this study proposes Masked Invariant Information Clustering (Masked IIC), a self-supervised clustering framework that extends IIC by integrating suppression-based masking into mutual-information estimation. Stochastic masking mitigates shortcut learning caused by dominant categorical variables, while deterministic structural masking handles missing-value patterns without requiring interpolation. Quantitative validation on labelled proxy datasets and application to a real-world glass melting furnace show that the framework can recover regime-relevant structure without label supervision and reveal operationally meaningful regime transitions. A remaining limitation is that the number of clusters and the interpretation of discovered regimes require domain expertise. Nevertheless, Masked IIC provides a practical framework for exploratory regime discovery in industrial MVTS.