VAE4RSS: A VAE-based neural network approach for robust soft sensor with application to zinc roasting process. (September 2022)

• Record Type: Journal Article
• Authors: Wang, Chengzhu; Li, Yonggang; Huang, Keke; Yang, Chunhua; Gui, Weihua
• Is Part Of: Engineering applications of artificial intelligence. Volume 114(2022)
• Publication Date: 2022-09
• Subjects: Robust soft sensor -- Variation autoencoder -- Neural network -- Probability density estimation -- Zinc roasting process; Engineering -- Data processing -- Periodicals; Artificial intelligence -- Periodicals; Expert systems (Computer science) -- Periodicals; Ingénierie -- Informatique -- Périodiques; Intelligence artificielle -- Périodiques; Systèmes experts (Informatique) -- Périodiques; Artificial intelligence; Engineering -- Data processing; Expert systems (Computer science); Periodicals; 620.00285
• DOI: 10.1016/j.engappai.2022.105180
• Languages: English
• ISSNs: 0952-1976

Abstract

Abstract: Soft sensor plays a progressively significant role in modern industrial processes. However, process variables usually have complex distribution characteristics, which can adversely affect the performance of soft sensor. On the other hand, due to the inevitable presence of noise in industrial data, the effect of traditional prediction models based on point estimates are greatly reduced. To address these problems, this article proposes a variation autoencoder (VAE) based neural network for robust soft sensor (VAE4RSS) approach. Specifically, on the basis of reconstructing the process variables by the autoencoder (AE), Gaussian distribution constraints are added to the latent features, and the unfavorable effects of complex distribution characteristics on prediction can be overcome by converting the original data into constrained latent features. Then, in order to reduce the negative effect of outliers, the probability density function (PDF) is introduced to describe the training errors instead of the traditional point estimates, an error PDF optimization based neural network prediction model is established to improve the robustness of soft sensor. Finally, in order to evaluate the efficiency and superiority of the proposed method quantitatively, we conduct extensive experiments on a numerical simulation case and an industrial zinc roasting process case in comparison with several state-of-the-art methods. The experimental results demonstrate that the proposed method exhibits satisfactory prediction results and is robust to outliers.