Quality-related nonlinear process monitoring of power plant by a novel hybrid model based on variational autoencoder. (December 2022)

• Record Type: Journal Article
• Authors: Wang, Peng; Ren, Shaojun; Wang, Yan; Zhu, Baoyu; Fan, Wei; Si, Fengqi
• Is Part Of: Control engineering practice. Volume 129(2022)
• Publication Date: 2022-12
• Subjects: Quality-related -- Process monitoring -- Thermal power plants -- Variational autoencoder -- Information bottleneck; Automatic control -- Periodicals; 629.89
• DOI: 10.1016/j.conengprac.2022.105359
• Languages: English
• ISSNs: 0967-0661

Abstract

Abstract: Quality-related process monitoring with data-driven methods has been widely researched and applied for anomaly detection in modern thermal power plants. However, the balance between the feature decomposition capability and reconstruction capability of the model is neglected, which leads to suboptimal modeling and monitoring performance. In this paper, a novel hybrid model called information bottleneck-based variational autoencoder with supervised regression (IBVAE-SR) is proposed for nonlinear system modeling and process monitoring. The β -VAE model is extended into a hybrid network structure with a regression task to establish the relationship between the partial latent variables and quality-related variables. Moreover, some threshold coefficients are introduced in the loss function and the balance coefficients in these loss terms are fixed, which can improve the disentanglement in learning representations while ensuring the model output accuracy. The statistical metrics corresponding to the latent variables and the output of the model are designed for process monitoring. Finally, the effectiveness of the proposed method is validated via its application on a numerical simulation and an industrial medium-speed coal mill. Highlights: A novel hybrid model named IBVAE-SR based on β -VAE and DVIB is proposed. DVIB method is used to extend a regression network on the β -VAE network structure. The disentanglement of latent features can be improved by adjusting the thresholds. A quality-related process monitoring method based on IBVAE-SR is proposed. New statistical indicators are designed to monitor equipment production.