Self-Supervised learning for industrial time series in process control systems: deployment profiles and acceptance criteria based on a structured review and a bench experiment

Abstract

Purpose. To justify the selection and deployment of self-supervised learning methods for industrial time series within process control systems and to test the proposed deployment profiles on a bench ICS/HIL dataset. The relevance is driven by scarce reliable labels, reliability requirements of monitoring services, and international research on contrastive learning, universal time-series representations, and method taxonomies. Model. The study uses a structured review of openly available full-text publications from 2019–2026; the final corpus includes 20 works. Each publication was coded by method family, runtime mode, resource profile, robustness to missing data and distribution shifts, diagnostic signal, thresholding scheme, and update rules. The experimental part was performed on the open HAIEnd 23.05 dataset: training and threshold calibration used only normal-operation archives, while test labels were used solely for the final assessment of profiles P1–P3. Findings. Four deployment profiles are identified: a lightweight streaming detector, a fixed encoder with lightweight decision blocks, a two-stage alarm-confirmation scheme, and cross-asset transfer. The bench run showed that profile P2 with a masked autoencoder and threshold q = 0.995 achieved the highest F1 score in the considered experiment, F1 = 0.537, and covered 29 of 52 attack intervals, whereas more conservative P2 (q = 0.999) and P3 configurations are preferable when false alarm episodes must be limited. Research limitations. The review corpus is limited to openly available full texts, and the bench experiment is based on the open HAIEnd 23.05 dataset rather than on an archive from a specific industrial enterprise. Therefore, the results should be interpreted as a reproducible bench validation of the acceptance protocol and deployment profiles, not as proof of industrial effectiveness for a particular process control object. Practical implications. A minimum set of acceptance criteria is proposed and tested: detection quality, attack-interval coverage, detection delay, alarm load, robustness to regime shifts, regression checks across modes, version traceability, rollback capability, and fail-safe behavior under input-data degradation. Social implications. The indirect effect concerns reduced operator workload, higher trust in the alarm subsystem, and lower risk of misinterpreting technological deviations. Industrial deployment should preserve human responsibility for operational decisions and provide transparent alarm messages. Originality/value. The contribution is the translation of self-supervised time-series research into engineering language for design, acceptance, and maintenance of computational subsystems in process control, supplemented by an experimental demonstration through shadow-version comparison on HAIEnd 23.05. The results are intended for researchers, developers of industrial monitoring systems, automation engineers, diagnostic specialists, and maintainers of software-hardware complexes.