Supervised convolutional autoencoder-based fault-relevant feature learning for fault diagnosis in industrial processes

• Published in: Journal of the Taiwan Institute of Chemical Engineers 2022-03, Vol.132, p.104200, Article 104200
• Main Authors: Yu, Feng, Liu, Jianchang, Liu, Dongming, Wang, Honghai
• Format: Article
• Language: English
• Subjects: Convolutional autoencoder; Deep learning; Fault diagnosis; Fault-relevant feature
• ISSN: 1876-1070; 1876-1089
• DOI: 10.1016/j.jtice.2021.104200

•A supervised convolutional autoencoder is proposed for multivariate fault diagnosis.•Supervised convolutional autoencoder is used to pretrain the deep network and learn the fault-relevant feature.•A minimum difference transformation function is introduced to the network pretraining.•Classification performance is validated on the CSTR process and the TE process. Convolutional autoencoder (CAE) is an unsupervised feature learning method and shows excellent performance in multivariate fault diagnosis. However, CAE cannot guarantee that the extracted feature is always related to the fault type due to its unsupervised self-reconstruction in the pretraining phase. To solve this problem, a new feature learning method, supervised convolutional autoencoder (SCAE) is proposed to pretrain the network and learn representative feature containing internal spatial information and fault information. In the SCAE, process sample and corresponding label are reconstructed by multilayer encoding-decoding the raw sample. Meanwhile, to prevent label information overfitting the network, a minimum difference transformation function is introduced into the loss function. The obtained fault-relevant features can be obviously distinguished between different fault types. The trained pretraining network provides more appropriate predefined parameters for fine-tuning to improve the classification performance. The effectiveness of the proposed method is evaluated by the continuous stirred tank reactor (CSTR) process and the Tennessee Eastman (TE) process. [Display omitted]