Chemical Process Monitoring Method Based On Spatiotemporal Sequence Predictive Neural Networks

In order to monitor the operation status of chemical process in real-time,ensure the stable and safe operating of production process,process monitoring technology has become a research hotspot.With the development of computer technology and information technology,and the widely application of distributed control system(DCS),a large amount of data is collected by sensors distributed in production equipment,which provided the motivation for evolution of data-driven process monitoring methods.In recent years,neural networks have been used in chemical process monitoring field and achieved excellent results.However,limited by input form and internal structure of convolutional neural networks(CNN)and recurrent neural networks(RNN),the dynamic characteristic of data and the spatial correlation between measurements cannot be considered comprehensively.To deal with this issue,spatiotemporal sequence data of chemical process was first defined as the data which satisfy the conditions that there is autocorrelation of each individual variable,and there is cross correlation between variables which is stronger between adjacent measurements.Furthermore,a spatiotemporal sequence predictive neural network based process monitoring method is proposed.In offline modelling stage,an improved predict recurrent neural network(Pred RNN++)that contain cascaded dual memories to deal with temporal and spatiotemporal correlation is applied as a autoencoder to extract the spatiotemporal features contained in spatiotemporal sequence data,the reconstruction errors are used as the monitoring statistics and its control limit is established by kernel density estimation method.Data from a simulation spatiotemporal sequence data and a pre-reforming reactor are applied to validate the proposed methods.Compared with the principal component analysis method,autoencoders,convolutional based and long short-term memory based process monitoring method,the proposed method achieved the best performance in two cases.It is proved that comprehensively extracting spatiotemporal features of spatiotemporal sequence data can effectively improve the monitoring results.