Semi-Supervised Deep Conditional Variational Autoencoder for Soft Sensor Modeling

Variational autoencoder (VAE) as an unsupervised deep generated model has been widely applied to process modeling for industrial processes due to its excellent ability in nonlinear and uncertain feature extraction. However, soft sensor based on VAE model faces three challenges. First, the constructe...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:IEEE sensors journal Ročník 24; číslo 5; s. 7153 - 7164
Hlavní autoři: Tang, Xiaochu, Yan, Jiawei, Li, Yuan, Zhang, Xinmin, Song, Zhihuan
Médium: Journal Article
Jazyk:angličtina
Vydáno: New York IEEE 01.03.2024
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Témata:
Data mining
Data models
Decoding
Deep learning
Feature extraction
Labels
Machine learning
Modelling
semi-supervised (SS) model
Sensors
soft sensor
Soft sensors
supervised conditional variational autoencoder (VAE)
ISSN:1530-437X, 1558-1748
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Variational autoencoder (VAE) as an unsupervised deep generated model has been widely applied to process modeling for industrial processes due to its excellent ability in nonlinear and uncertain feature extraction. However, soft sensor based on VAE model faces three challenges. First, the constructed supervised VAE model makes it difficult to describe the correlation between input and output based on self-network. Second, the output of the VAE may suffer from instability and uncontrollability. In addition, the limited labeled data in industries are the third challenge. To solve the above problems, a semi-supervised deep conditional VAE (SS-DCVAE) is constructed for soft sensor based on a supervised DCVAE (S-DCVAE) and an unsupervised DCVAE (U-DCVAE). The S-DCVAE model is constructed by injecting unlabeled data, the actual labels, and estimated labels as constraint conditions from the preneural network. Based on such a conditional supervised structure, the input-output correlation can be strengthened and the generated data can be controlled toward the aim direction. Furthermore, the U-DCVAE model can be built by making the latent distribution as similar as possible to S-DCVAE, as well as only using unlabeled data with corresponding estimated labels. In this way, the unlabeled data can be fully utilized and online prediction can be achieved. Finally, combining the decoder of S-DCVAE model with the encoder of U-DCVAE, the SS-DCVAE model is constructed with both advantages. The effectiveness and superiority of the SS-DCVAE model are demonstrated by comparing the prediction results of the proposed model with other deep learning methods based on industrial cases.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1530-437X
1558-1748
DOI:10.1109/JSEN.2024.3351431