Розробка soft-сенсору з інтегрованим виявленням аномалій для оцінювання октанового числа риформату

The object of research is the catalytic reforming process, for which the reformate octane number is used as an efficiency indicator. The research addressed the scientific and practical problem of rapidly determining the octane number, which is an important parameter for control tasks. The proposed solution is based on the soft sensor, which is developed based on the Random Forest ensemble model. This approach allows quantifying the uncertainty of the forecast through the variance of the model’s ensemble, which is used to build an integrated anomaly detection system. The input vector consists of ten parameters, critical to the technology process dynamic and are available for direct measurement: reactor temperatures, pressures, flow rates, raw material and reformate densities, and the catalyst activity index. The training data for the soft sensor is generated on a simulation model with technological parameters varying within the nominal operating range. The data volume corresponds to 30 days of reforming process operation with a 5-minute interval. The soft sensor parameters are optimized by random search method with five-fold cross-validation. The resulting soft sensor provides high prediction accuracy (MAE = 0.41, RMSE = 0.54, R² = 0.974) on the validation sample. The research used three criteria for detecting anomalies: monitoring the forecast interval width, checking the forecast output over the historical range, and the Mahalanobis distance in the input feature space. Five abnormal process scenarios are developed and tested, including a sudden change in feedstock composition and gradual catalyst deactivation, etc. An anomaly detection rate of 92–96% is achieved, with a false alarm rate not exceeding 2.1%. The results confirm the feasibility of using ensemble methods for constructing soft sensors and for the early detection of anomalies in oil refining processes.