基于时序双通道图注意力的工业过程故障时序预测模型  

A time series prediction model for industrial process failures based on temporal dual-channel graph attention network

作　　者：王明胜 WANG Mingsheng(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620

出　　处：《智能计算机与应用》2024年第12期26-35,共10页Intelligent Computer and Applications

摘　　要：工业过程是工业生产的基石,工业过程故障会导致整个工业流程中断,造成巨大经济损失,但由于工业过程多元特征的高耦合和时变等特性,传统机器学习方法难以实现对工业过程故障时序的有效预测。针对此问题,本文利用共享的多头图注意力机制提取特征的高耦合关联关系,并结合双向门控循环单元(Bi-directional Gated Recurrent Unit,BiGRU)和多尺度时间卷积网络(Multi-Scale Temporal Neural Network,MSTCN)构成时序学习层(Temporal Learning Module,TLM)提取特征的多尺度时序信息,提出一种基于时序双通道图注意力(Temporal Dual-Channel Graph Attention,TDCGAT)的工业故障时序预测模型。最后,通过堆叠时序学习层和共享的图注意力实现对工业过程故障时序的有效预测。Industrial process is the cornerstone of industrial manufacturing,whose failure can lead to the interruption of the entire industrial processes,resulting in huge economic losses.However,due to the high-coupling and time-varying characteristic of the multivariate features of industrial processes,traditional machine learning methods have difficulty in achieving precise prediction of industrial process failures time series.To address this problem,a time-series prediction model for industrial faults based on Temporal Dual-Channel Graph Attention is proposed in this paper.In the model,a shared multi-head graph attention is used to extract the highly coupled correlations of industrial features,a BiGRU and MSTCN are combined as a Temporal Learning Module to extract the multi-scale temporal characteristics of features.Finally,the effective prediction of industrial process faults multivariate time series is achieved by stacking TLM and shared graph attention.

关 键 词：多元时序预测 工业过程故障 图注意力 BiGRU TLM 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]