Nonlinear Non-Gaussian and Multimode Process Monitoring-Based Multi-Subspace Vine Copula and Deep Neural Network

Abstract

This paper proposes a robust copula double-subspace (CDS) model based on a sparse robust autoencoder (SRAE) named SRAE-CDS. By reconstructing training data with SRAE, the resulting SRAE-CDS model is more robust to changes in inputs and is more sensitive to process faults. A five-layer SRAE model is proposed to classify the multimode process and extract abstract features, which is first introduced in fault detection area. The SRAE model not only classifies the operating conditions but also extracts high-level features. According to the abstract features, the CDS method, which is good at handling non-Gaussian and nonlinear data, is used to depict the distribution of the advanced features. To perform process monitoring, the highest density region index under a given control limit is calculated in real time. This paper first proposes the mode identification method with the SRAE model and monitors the industrial process through depicting the distribution of the middle layer in deep neural network, achieving good performance. The effectiveness and benefits of the SRAE-CDS method are illustrated in three experiments: the first is a numerical example, the second is a Tennessee Eastman benchmark process for fault detection, and the third is a real ethylene cracking furnace process. The results show that the SRAE-CDS model achieves good performance for industrial processes.

Multi‐layer features ensemble soft sensor regression model based on stacked autoencoder and vine copula

“…Three-layer AEs are unable to extract more intricate data features, and high-performance modelling requires deeper features as well. Thus, the SAE [32,33] was proposed and has been used subsequently in several research fields. It consists of several AEs, where the hidden-layer output of the previous AE represents the input of the next AE, and a deeper AE with more learning capability will be obtained.…”