Variable ontology and target-leakage screening in industrial PLC-SCADA data using name-based and behavioral evidence

Contemporary cement plants measure thousands of PLC-SCADA variables, and the semantics of those variables is unclear, since the tags and labels are usually incomplete, inconsistent, and/or deceptive. The problem affects not only control-advisory systems but also data-driven models via inclusion of non-manipulated variables into the decision space or introduction of target-related variables resulting in leakage. In this paper, we propose the two-evidence variable ontology and leakage-auditing approach for industrial PLC-SCADA data. We combine name-based evidence, behavior-based signal evidence, expert validation for exceptional cases, and target-leakage auditing within a unified pre-modelling workflow. Within this workflow, the staircase score is used as a lightweight behavioral descriptor for distinguishing piecewise-constant setpoint-like signals from continuously varying measurements. The proposed approach is applied to the real industrial dataset of a cement clinker production line with 2.46 million samples, 375 variables, and 11 logged operational days. Among the 26 documented control parameters, eight exhibited actively changing setpoint behavior during the analyzed 11-day period, while one additional operational control was confirmed through process-engineering validation. Besides the label anomaly where the continuously changing signal is labeled as the setpoint, three leakage types for the industrial multivariate regression are considered in this paper: definitional/mass balance, component, and computed feature leakage. Removal of leakage drops the performance of the firing-calorie model from R² = 0.9999 to R² = 0.8109. Supplementary analyses showed that the behavioral separation of setpoint-like and continuous signals was unchanged under unsupervised (Otsu) threshold selection, was reproduced in two sub-periods with different logging rates, and required about 21 s for all 375 variables, with runtime scaling linearly with data size. A blinded annotation of all 375 variables by a second, independent process expert agreed with the final ontology for 89.3% of the variables (κ = 0.755) and assigned all nine control variables to CONTROL.