Recent industrial systems generate massive and complex multivariate time-series (MVTS) data containing heterogeneous signals such as continuous sensor values, event-driven categorical states, and process control indicators. To prevent potential losses, highly sophisticated assessments of the current state must be derived from these signals. However, categorical signals often dominate clustering, obscuring the contributions of continuous process variables. This study proposes a self-supervised clustering framework for visualized multivariate time-series data to quickly reveal latent structures in industrial processes. The framework extends the standard Invariant Information Clustering model by introducing masking strategies that selectively maximize mutual information, enabling robust handling of noisy, irrelevant, or missing signal regions. In this study, we particularly focus on categorical signals, and present Masked Invariant Information Clustering (Masked IIC) as a means to mitigate their dominance. Experimental evaluations across diverse MVTS applications—including network intrusion detection, water treatment safety monitoring, and furnace process control—demonstrate that the proposed clustering method reveals process transitions. These results highlight the framework’s potential for structure discovery without reliance on labeled data, offering a general approach with broad applicability to industrial automation and reliability engineering.