A Hybrid Transfomer-BiLSTM-Att Framework for Dynamic Prediction of Thickness Deviation in Cold-Rolled Aluminum Plates

In the cold rolling process of aluminum plate, the exit thickness deviation is one of the important indicators to evaluate the plate quality. To achieve dynamic monitoring of this metric and proactively detect anomalies, this paper proposes a hybrid deep learning model named Transformer-BiLSTM-Att. The model consists of a Transformer encoder, a bidirectional long short-term memory (BiLSTM) decoder, and a self-attention (Att) mechanism. To validate its effectiveness, Transformer-BiLSTM-Att is analyzed for ablation and compared to four other state-of-the-art prediction methods. Experimental results demonstrate that each component contributes meaningfully to the model's prediction capability, enabling it to achieve the best multi-step performance, with MSE = 0.0217, R2 = 0.9071, RMSE = 0.1471, and MAE = 0.1174 for 3-step prediction. Meanwhile, Transformer-BiLSTM-Att also exhibits good generalization performance in real-world production environments, effectively identifying potential thickness deviation abnormalities.