Batch process quality prediction based on denoising autoencoder-spatial temporal convolutional attention mechanism fusion network

In batch processes, the accurate prediction of quality variables plays a crucial role in smooth production and quality control. However, various sources of noise in the production environment cause abnormal data fluctuations that deviate from the real value. Coupled with the dynamic nonlinearity of batch processing and the complex spatiotemporal relationship of variables, which greatly increase the difficulty of prediction and pose a severe challenge to prediction performance. Therefore, a denoising autoencoder-Spatial Temporal Convolution Attention Fusion Network (DAE-STCAFN) prediction method is proposed. Firstly, combining DAE and maximum information coefficient (MIC), multi-level data features are extracted to prepare high-quality input data for the quality prediction model. DAE is used to denoise the original data, and relevant variables are selected through MIC. Then, an augmented matrix is constructed to eliminate the autocorrelation of the selected variables in the time series. Secondly, a spatial temporal convolutional attention fusion mechanism is created to extract the spatial temporal fusion features between the input and output variable sequences. Thirdly, to further enhance the learning ability of the model, a batch attention module is constructed to automatically learn the relationship among sample in small batch. Finally, experiments were carried out on the simulation platform of penicillin fermentation and hot tandem rolling process. In the prediction process of penicillin concentration, RMSE and MAE of the proposed method were 0.0099 and 0.0077, respectively. In the prediction of strip thickness, the RMSE and MAE are 0.0008 and 0.0003 respectively. The results show that the proposed method is effective both in simulation experiment and in actual industrial production in terms of prediction accuracy, stability and generalization ability.