One-dimensional decoupled convolutional autoencoder with sparse self-attention mechanism for process monitoring

Industrial processes are constantly disturbed by environmental and human factors during operation. Undifferentiated alarming of these disturbances will bring serious alarm disaster problems.Effectively distinguishing the disturbances that have different effects on the process operation state can hel...

Full description

Saved in:
Bibliographic Details
Published in:Process safety and environmental protection Vol. 199; p. 107156
Main Authors: Yang, Yuguo, Shi, Hongbo, Song, Bing, Tao, Yang, Guo, Lei
Format: Journal Article
Language:English
Published: Elsevier Ltd 01.07.2025
Subjects:
Convolutional autoencoder
Data distribution alignment
Process knowledge embedding
Process monitoring
Variable-level feature extraction and interaction
Process monitoring
Process knowledge embedding
Data distribution alignment
Convolutional autoencoder
Variable-level feature extraction and interaction
ISSN:0957-5820
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Industrial processes are constantly disturbed by environmental and human factors during operation. Undifferentiated alarming of these disturbances will bring serious alarm disaster problems.Effectively distinguishing the disturbances that have different effects on the process operation state can help the field operators to make a reasonable risk assessment.To achieve the above purposes, this paper proposes a one-dimensional decoupled convolutional autoencoder network with sparse self-attention mechanism under process knowledge constraints (PKC-SSAM-DCAE). Firstly, aiming at the change of data distribution caused by feedback control adjustment, the window normalization strategy is adopted for the standardized data. Realize data distribution alignment at the input end of the model. Subsequently, one-dimensional decoupled convolutional encoder (DCAE) is constructed to extract the features of each process variable. The sparse self-attention mechanism network (SSAM) is constructed under the constraint of process knowledge to realize the interaction between process variable features. Then the detection index is established according to the network prediction results. When the fault is detected, the variable oblivion contribution plot is given to locate the key fault variables.Finally, through the experiments on Tennessee Eastman process, it is verified that the proposed model can solve the problem of data distribution change caused by process feedback adjustment, and can accurately distinguish process normal adjustment from faults.
ISSN:0957-5820
DOI:10.1016/j.psep.2025.107156