Short-Term Multistep Prediction of Process Industry Product Quality With CNN-BiLSTM Network Based on Parallel Attention Mechanisms

In the process industry, online product quality prediction is crucial for effective quality control. However, the highly nonlinear and dynamic nature of monitoring data, characterized by temporal-spatial relationships, brings great challenges to product quality prediction. To address this issue, a novel model is proposed for short-term prediction of product quality in process industries. This model employs a parallel attention mechanism to extract weight coefficients from raw data's temporal and feature dimensions. Subsequently, a 1-D CNN with SeLU activation performs spatial correlation extraction and noise filtering. The processed time series data is then effectively modeled by a BiLSTM network. Finally, a dense network realizes short-term prediction of product quality. The proposed model was evaluated using data from coal processing scenarios and water quality predictions. The results demonstrated superior multistep prediction accuracy and stability compared to existing models, including BiLSTM, CNN_BiLSTM, and Attention_CNN_BiLSTM. This study proposes a deep learning approach that effectively addresses product quality nowcasting in the process industry, achieving satisfactory prediction performance.