Anomaly identification with few labeled data in the distillation process based on semisupervised ladder networks. Issue 2 (24th October 2020)

• Record Type: Journal Article
• Authors: Li, Chuankun; Xu, Wei; Zhao, Dongfeng; Yuan, Zhuang; Shi, Jihao; Wang, Chunli
• Is Part Of: Process safety progress. Volume 40:Issue 2(2021)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2020-10-24
• Subjects: anomaly identification -- distillation process -- few labeled data -- semisupervised ladder networks; Chemical plants -- Management -- Periodicals; 660
• DOI: 10.1002/prs.12206
• Languages: English
• ISSNs: 1066-8527

Abstract

Abstract: High repeatability of similar information but a lack of typical fault features, in the monitoring data of distillation processes for continuous production, leads to a small proportion of data with labels. Therefore, the requirement for a large number of labeled samples in conventional deep learning models cannot be met, resulting in significant performance degradation in their anomaly identification. In this paper, an intelligent anomaly identification method for small samples is proposed, based on semisupervised deep learning. Specifically, on the basis of a deep denoising autoencoder (DAE), semisupervised ladder networks (SSLN) is constructed to use a large number of unlabeled, process data to assist the supervised learning process, thus improving the performance of the anomaly identification model. In order to construct the optimum SSLN model, the influences of parameters such as the number of deep network layers, the proportion of labeled samples, and the noise intensity on identification accuracy are analyzed while making the information flow in the network more efficient. Experimental results of anomaly identification in the depropanization distillation process show that compared with the conventional multilayer perception (MLP) and convolutional neural network (CNN)‐DAE models, the proposed method can obtain a higher diagnostic accuracy in the case with limited labeled process data.