基于CNN-LSTM-LOF的过程故障预测模型  

A prognosis model for process fault based on CNN-LSTM-LOF

作　　者：程志磊 章国宝[1] 黄永明[1] CHENG ZhiLei;ZHANG GuoBao;HUANG YongMing(School of Automation,Southeast University,Nanjing 210018,China)

机构地区：[1]东南大学自动化学院,南京210018

出　　处：《北京化工大学学报（自然科学版）》2024年第1期121-127,共7页Journal of Beijing University of Chemical Technology(Natural Science Edition)

基　　金：江苏省科技计划(BE2021750);江苏省重点研发计划(BE2022135)。

摘　　要：在现代工业过程中,故障预测可以及时预测设备的潜在故障,减少事故的发生以及降低经济损失,因此故障预测对于过程系统至关重要。由于过程系统的复杂性以及运行数据集分布不均,使用正常数据集离线预测运行状态的方法没有较好的泛用性,且不太准确。针对以上问题,将卷积神经网络(CNN)与长短期记忆网络(LSTM)相结合,用于提取设备运行数据的特征,在线预测之后的运行状态;随后将预测结果送入在离线状态下训练好的局部异常因子(LOF)模型中,计算预测出时间序列的异常值;最后将异常值与离线状态下训练出的故障阈值进行比较,大于阈值则认为存在潜在故障。将模型用于田纳西-伊斯曼(TE)时间序列进行验证,并与传统的故障预测方法进行比较,实验结果表明:本文所提模型对于多故障以及单故障预测相比传统故障预测方法均具有更好的效果,可以提前1个采样窗口检测到数据异常,有应用于工业故障预测的潜力。Fault prognosis is important in process systems since it can predict potential faults in industrial equipment in a timely manner and hence reduce the occurrence of accidents and economic losses.Due to the complexity of process systems and the uneven distribution of data sets,the conventional method of using the normal data set to predict the operating state offline is not versatile and inaccurate.In response to the above problems,this paper combines a convolutional neural network(CNN)with a long-short term memory network(LSTM)to extract the characteristics of boiler operating data and predict the operating state after online prediction.In the local outlier factor(LOF)model,the outliers of the time series are calculated and predicted.The results are compared with the fault threshold trained in the offline state,and if it is greater than the threshold,it is considered that there is a potential risk.The model was used in the Tennessee-Eastman(TE)process,and compared with traditional fault prognosis methods.The results show that the model performs well in multi-fault and single-fault prognosis,and outliers could be detected earlier by one sampling window.The results indicate the model has potential applications in fault prognosis in industrial process systems.

关 键 词：故障预测 田纳西-伊斯曼过程 长短期记忆 局部异常因子算法 卷积神经网络 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]