Adaptive Data-Driven Soft Sensor for Monitoring and Prediction of Temperature Inside Zinc Rotary Kiln

Accurate temperature monitoring and prediction are crucial for maintaining product quality control in zinc rotary kilns. However, achieving precise temperature prediction and real-time monitoring under dynamic working conditions remains a pressing challenge. This article proposes a methodology to integrate the moving window (MW) technique and just-in-time (JIT) learning mechanism into the partial least-squares (PLS) regression algorithm to construct two types of adaptive soft sensors for quality variable predictions. The feasibility of enhancing simple linear models with adaptive mechanisms for soft sensors is demonstrated through validation using benchmark simulation data, eliminating the need for nonlinear and complex models. Moreover, the results also highlight the potential of these models to serve as a backup for hardware sensors in cases of physical sensor failure or calibration issues. Experimental findings reveal that the MW-PLS mechanism reduces the root-mean-square error (RMSE) in temperature prediction by approximately 15.3% and the mean absolute error (MAE) by 12.5%. The proposed model’s predictions fit actual values under highly dynamic temperature variations, in contrast to traditional models that suffer from significant time delay. Furthermore, the MW-PLS mechanism substantially enhances the model’s real-time adaptability, with a 25% reduction in response time, thereby effectively complementing hardware temperature sensors in scenarios with faults or calibration issues.