Quality Prediction for Nonlinear Dynamic Processes Using Semi-Supervised Soft Sensors: An application on Ammonia Decarburization Processes

As a measurement for the production performance, the online quality variables from soft sensors contribute greatly to obtaining immediate information from the process. The complex correlations between large numbers of process variables and disturbances inherited from the dynamic and nonlinear characteristics of chemical processes put more challenges in constructing the soft-sensor models. The soft sensors which are typically developed in steady-state conditions are not suitable for doing predictions in a dynamic operating system. This paper proposes a semi-supervised latent dynamic variational autoencoder (S-2-LDVAE) to learn features between the process and quality data.Furthermore, the issue of the uneven length of the process and quality data is noteworthy. When there are fewer quality data than the process data, severe degradation to the performance of the trained model may occur. The process and quality data are encoded into the latent space in a temporal way for dynamic feature extraction. In the case of missing quality data, the artificial data generated by the trained prediction network are used to provide quality estimates during on-line prediction. The proposed method is compared with the other methods to show its contribution and performance in terms of quality prediction in a numerical case and an industrial case.

Automated summary

This paper proposes a semi-supervised latent dynamic variational autoencoder (S-2-LDVAE) to learn features between the process and quality data. It also addresses the issue of the uneven length of the process and quality data. In the case of missing quality data, artificial data generated by the trained prediction network are used for online quality estimation.