Högskolan i Skövde

This thesis presents a machine learning-based framework for predicting thermocouple temperatures and analysing feature importance in industrial furnaces. Utilizing high-frequency time-series data from two furnaces with six thermal zones, the study explores the effectiveness of both linear (Lasso, Ridge) and nonlinear (Random Forest, XGBoost, LSTM) models in forecasting localized and aggregated temperature behaviours.

Results indicate that nonlinear models, particularly XGBoost, achieve superior accuracy and execution efficiency, making them ideal for real-time deployment in furnace operations. Feature interpretability is addressed using SHAP and LIME, which highlight critical process variables such as fuel flow rates, zone-level average temperatures, and line control speed. These insights provide valuable input for process optimization and predictive maintenance.

An adaptive thresholding mechanism is implemented using residual-based confidence intervals, enabling dynamic and statistically grounded detection of sensor anomalies. The hierarchical approach by structuring models at the zone level, furnace level, and whole unit level, give engineers the flexibility to pinpoint anomalies with precision and explore thermocouple (TC) changes in specific areas. This adds a layer of adaptability that’s crucial for troubleshooting and operational decision-making. Additionally, the study evaluates the utility of transfer learning for model reuse across zones, finding it beneficial for data-limited settings.

To enhance operational usability, an interactive Power BI dashboard is developed for real-time monitoring, visualization, and anomaly exploration. This end-to-end framework demonstrates how predictive modeling, interpretability, and visualization can be integrated to support smarter, data-driven furnace operations in industrial environments.