Marginal region-integrated regressive conditional variational autoencoder-generative adversarial network: A soft sensing enhancement method

In industrial processes, due to limitations of actual industrial production, many industrial data are difficult to obtain directly, which limits sample size and leads to uneven data distribution, ultimately affecting the fitting performance of soft sensing models. To address this challenge, we propo...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Chemometrics and intelligent laboratory systems Jg. 268; S. 105577
Hauptverfasser: Liu, Guo-yu, Zhu, Qun-Xiong, Luo, Yi, Ke, Wei, He, Yan-Lin, Zhang, Yang, Zhang, Ming-Qing, Xu, Yuan
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Elsevier B.V 15.01.2026
Schlagworte:
Generative adversarial network
Mega trend diffusion
Soft sensor
Virtual sample generation
Virtual sample generation
Soft sensor
Mega trend diffusion
Generative adversarial network
ISSN:0169-7439
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:In industrial processes, due to limitations of actual industrial production, many industrial data are difficult to obtain directly, which limits sample size and leads to uneven data distribution, ultimately affecting the fitting performance of soft sensing models. To address this challenge, we propose a marginal Isolation Mega Trend Diffusion with Regressor Conditional Variational Autoencoder-Generative Adversarial Network (IRCVGAN). designed to improve model accuracy by expanding the sample size. Specifically, the proposed method first applies the isolation forest algorithm to detect sparse marginal regions in the dataset, followed by Mega Trend Diffusion (MTD) to broaden the range of input data by generating virtual samples, thus increasing dataset diversity. Next, an improved regressive conditional Variational Autoencoder-Generative Adversarial Network (RCVAEGAN) is developed to perform fine-grained selection on the virtual samples generated by MTD. Furthermore, the mapping between input variables and production quality indicators is embedded in RCVAEGAN, enhancing the representativeness of the samples and improving the model’s fitting accuracy, the effectiveness of our proposed method is validated through function fitting tests and real-world industrial data from a purified terephthalic acid (PTA) solvent system. •modified MTD method augments data in marginal and sparse regions.•Conditional variables are added as supervisory signals to the VAE-GAN framework.•A regressor is incorporated to couple sample generation with soft sensing.•The method improves soft sensor modeling prediction.
ISSN:0169-7439
DOI:10.1016/j.chemolab.2025.105577