A novel quality weighting analysis based regression Variational Autoencoder method for industrial soft sensing

In modern industrial processes, key variables are often difficult to measure and analyze directly using sensors, necessitating the use of soft sensing methods for modeling and analysis. However, traditional soft sensing models often overlook correlation analysis between variables. To address these issues, this paper proposes a mutual information analysis weighting based regression variational autoencoder model (MRVAE), aiming at improving the modeling accuracy and solving the nonlinear correlation problems of the model. Our method incorporates correlation analysis into a variational autoencoder (VAE) framework and integrates a regressor structure for predictive analysis. On the one hand, mutual information is used to identify and emphasize the most relevant variables, ensuring that the feature extraction process captures meaningful correlations. On the other hand, the regressor structure is designed to predict critical variables, by taking advantage of the time delay of the data, a dual encoder structure is constructed to enhance the model's fitness. The proposed method is applied to predict key variables in industrial cases such as SRU system. Results demonstrate that MRVAE significantly outperforms traditional models, achieving higher prediction accuracy.