Hybrid LSTM-attention-KF model for dynamic soft measurement of rotary kiln temperature

Abstract Rotary kilns, as critical thermomechanical units in the steel industry, require accurate soft measurement of internal temperature to ensure product quality, enhance production efficiency, and prolong equipment lifespan. Direct measurement inside the kiln is often infeasible due to harsh conditions, while conventional sensing approaches are affected by time delay, nonlinear dynamics, measurement noise, and missing data. To address these challenges, this study proposes a hybrid soft measurement model that integrates Long Short-Term Memory (LSTM) networks, a self-attention mechanism, and the Kalman Filter (KF). LSTM is employed to capture temporal dependencies in temperature variations, attention weights the most informative hidden states, and KF suppresses noise while compensating for data gaps. Experimental results show that, compared with the best-performing baseline model (Informer) among the evaluated conventional baselines, the proposed method reduces Mean Absolute Error (MAE) by at least 28.79% and decreases Root Mean Square Error (RMSE) by at least 26.92%, maintaining estimation errors at consistently low levels. These findings demonstrate the model’s effectiveness as a reliable soft sensor for rotary kiln temperature monitoring, offering robust technical support for real-time process perception, precise control, and intelligent optimization in industrial applications.