Key Process Quality Index Prediction Based on Variable Selection

To improve the prediction accuracy of key process quality index in continuous industrial processes, we propose a novel industrial quality prediction method combining variable selection, deep feature extraction, and temporal analysis. Firstly, we employ a K-nearest neighbor conditional mutual information-based variable selection method to screen key process variables highly correlated with quality variables from high-dimensional redundant data. Secondly, we adopt an improved deep autoencoder to perform feature extraction on the screened variables, where original features and extracted deep features are fused to construct enhanced feature sets, while the introduction of regularization terms enhances model robustness and generalization capability. Finally, we construct quality prediction models using long short-term memory networks to fully capture temporal dependencies in data. In verification tests across three industrial cases, the proposed method achieved high prediction accuracy while demonstrating lower prediction errors and smaller output fluctuation amplitudes under different operating conditions, indicating its certain applicability to continuous industrial processes.