Quality prediction of multi-stage batch process based on integrated ConvBiGRU with attention mechanism

It is important for quality prediction and monitoring to ensure the safe operation of the process. When constructing a prediction model, it is crucial to choose appropriate input variables to influence the online prediction performance and quality monitoring. Data-driven techniques have been widely used for prediction and monitoring of quality variables, but there are some difficulties in the application of batch processes, three-dimensional characteristics of data, different initial conditions, and multi-stage characteristics within batches. Therefore, we propose a quality prediction model of multi-stage batch process based on integrated ConvBiGRU with attention mechanism (MI-ConvBiGRU-AM). Firstly, Firstly, the original 3D data are expanded into 2D time slices by the batch-variable expansion method. Secondly, the 2D time slices are clustered to complete stage identification using the improved affine propagation clustering method based on the design of the Markov chain similarity matrix. At each stage, we select product quality-related modeling variables using the Maximum Relevance Minimum Redundancy (mRMR). Then, the selected variables are used to train a convolutional bi-directional gated recurrent unit with an attention mechanism (ConvBiGRU-AM). Finally, ConvBiGRU-AM model for each stage is integrated together a whole prediction model for the entire process to accomplish quality prediction, and the prediction residuals are utilized for quality monitoring. The validity of the proposed method was verified by Industrial-scale fed-batch fermentation (IFBF) process and the Hot strip mill (HSM) process. For the IFBF process, the model achieved an FDR of 99.73%, FAR of 0.54%, MAE of 0.0043, RMSE of 0.0396, MAPE of 0.0121, and R² of 0.9971. For the HSM process, the results were an FDR of 99.95%, FAR of 0.25%, MAE of 0.0053, RMSE of 0.0111, MAPE of 0.1539, and R² of 0.9990. These results demonstrate that the proposed method significantly improves prediction accuracy and achieves better quality monitoring compared to existing methods, highlighting its effectiveness for industrial applications.