Multivariate Time Series Prediction Modeling Of Polyester Fiber Based On Graph Neural Network

As a kind of synthetic fiber,polyester fiber has the advantages of excellent spinnability and resource exploitation,so it is widely used in various application scenarios,including civil fabric and industrial fabric.With the continuous development of polyester fiber production technology in recent years,many links from raw material production,fiber molding to processing have realized continuity,automation and intelligence.Therefore,polyester fiber has become a kind of synthetic fiber with high production efficiency and large yield.The polymerization process mainly consists of three stages,namely esterification,precondensation and final condensation.Its main purpose is to esterify terephthalic acid(PTA)and ethylene glycol(EG)as raw materials,the intermediate product after polymerization to get polymer.In these three stages,the sensor monitors the key indicators such as temperature and pressure,and their changes have a great impact on the reaction process.Therefore,accurate prediction and appropriate control of the change trend of these key indicators can ensure the stability of the polymerization process to a certain extent.However,compared with the general multivariate time series,the multiple variables of the whole process monitored by the sensor has higher complexity due to the coupling between multiple variables.Therefore,the impact of other relevant variables on a target variable should also be considered when predicting the target variable.This puts forward higher requirements for establishing multivariate time series prediction model based on polyester fiber polymerization process.In order to make the multivariate time series prediction model established more effective and accurate use of the potential correlation between the multivariate time series variables,this paper regards the sensors as nodes,and describes the topological structure constructed by each sensor node in the sensor network in the form of a graph.So that the model can carry out the information transmission between the target node and its neighbors through the connection relationship between the sensor nodes.On this basis,the multivariate time series prediction task in the polyester fiber polymerization process was studied in depth.The main research content of this paper can be divided into three parts: the first part is the multivariate time series prediction model of esterification process based on causal spatiotemporal graph neural network;The second part is the multivariate time series prediction model of the whole process based on local and global hierarchical graph neural network.The third part is the multivariate time series prediction model of the whole process based on dynamic graph neural network.The main innovation points of this paper include the following three points:1)A multivariate time series prediction model for the esterification process based on a causal temporal graph neural network is proposed.The model constructs a multivariate causal graph by learning the potential causal relationships between multivariate temporal variables at lagged moments and at the same moment,and expands the spatial association between variables in the temporal dimension,making the graph a spatio-temporal graph instead of a mere spatial graph.In addition,the validity of the learned causal relationships is demonstrated by comparing them with the control logic relationships among sensors in the flow chart of the actual aggregation process,which is closer to the industrial production reality than the previous random graph structure that relies entirely on the training process.In addition,most of the previous mechanistic models of industrial processes can only consider the relationship between variables from a holistic perspective,while the causal metric approach allows to understand the correlation between each variable pair,which provides a certain degree of reference for exploring the interactions between variables.2)A whole process multivariate time series prediction model based on local-global hierarchical graph neural network is proposed,which achieves the requirement of graph adjacency matrix sparsity and enables the graph convolution to extract spatial features from both local and global levels by partitioning a whole graph into several subgraphs.In addition,considering the heterogeneity of temporal features at different time scales,a multiscale fusion temporal convolution module is proposed,which enables multiple temporal features at different time resolutions to be interfused in the process of temporal convolution to obtain enhanced features with more representational power.Most of the previous time series modeling methods consider all the time series variables from the macro level,but for a process industrial system like industrial process,the correlation of variables between different links is mostly weak,so the model can focus a lot of attention on only a few most relevant neighborhood variables,which provides some reference for exploring the local model of industrial process,feature selection,etc.to some extent.In addition,the local and global aspects can also help to better understand the hierarchical structure of industrial processes from both perspectives.Finally,it can be seen from the experimental results that the model has better prediction performance on prediction tasks with a large number of nodes compared to other methods.3)A whole process multivariate time series prediction model based on dynamic graph neural network is proposed,which mainly implements the updating process of dynamic graph adjacency matrix and parameter matrix advancing with time,and its basic idea is based on the property that the hidden state of the gated recurrent unit can evolve with time,in which the historical information of graph adjacency matrix and parameter matrix is retained by updating gates as well as resetting gates and update the current moment state.The graph convolution based on this idea can directly extract the spatio-temporal feature information and obtain better prediction results with less network complexity.The experimental results prove that the model has good prediction accuracy in shorttime prediction scenarios,while the rationality of the dynamic graph is verified by comparing the evolution of variable correlations with their actual change curves.In addition,it is closer to the industrial production reality compared with the previous time series modeling methods based on the static relationship between variables,and the model explores the dynamic relationship between nodes over time can help to better understand how different variables affect each other at different stages.