Research On Soft Sensor Modeling Of Polyester Fiber Polymerization Process Based On Deep Learning

The polymerization process of the polyester(PET)fiber production is a highly complex industrial process with diverse production facilities,changing working environments,complicated chemical mechanisms,multiple influencing factors and exquisite industrial techniques.It is the key link that determines the final fiber qualities.Thus,it is extremly important to implement real-time monitoring and control for the key quality indicators of the process.However,in the highly complex polymerization production environment,the melt viscosity indicators cannot be monitored in a timely and effective manner due to the poor measurement environment and expensive and inefficient measuring instruments.To this end,soft sensor modeling technologies have been developed,which establishes a mathematical model between key quality indicators and easy-to-measure process variables to eventually achieve predictive estunation of the key quality indicators.Due to three-reactors structural characteristics of polyester fiber polymerization process and the internal complicated physicochemical reactions,the process data are intrinsically characterized with high-dimensional nonlinearities.At the same time,common uneven sampling rate leads to limited labels and sensor failure results in missing data.Except that,dynamics and multimodalities are also common in the process.Thus,in this thesis,we take deep learning algorithms as the main research methods to carry out soft sensor modeling research work on the polymerization process of PET fibers.The main research contents are as follows:(1)A new soft sensor model based on semisupervised improved gated recurrent unit is proposed to solve the problem of limited labels in PET fiber polymerization process caused by uneven sampling rate of process variables and key quality variable.Due to the special network structure of recurrent neural network,it is naturally suitable for dealing with the limited label problem.Firstly,a semisupervised gated recurrent unit regression model-based soft sensor is proposed by combing gated recurrent unit and kernel ELM neural network.Moreover,with the increase of the length of the sequence,the original gated recurrent unit is difficult to effectively obtain the correlation between samples in the sequence.Therefore,multi-heads self-attention mechanism is introduced to make it easier to obtain the correlation between samples at different times in the sequence,which largely improves the prediction accuracy of the model.(2)A new soft sensor model based on two-stream ?-gate recurrent unit(TS-?GRU)neural network is proposed for nonlinear dynamic data of PET polymerization process.Gate recurrent unit is successfully applied to deal with complex dynamic time series data.However,there is a linear coupling constraint in its state function,which largely impedes the flow of the information adequately passing through.Thus,two ? factors are introduced for altering the linear constraint to enrich the information passing through,forming ? GRU.Then,according to the internal characteristics of the key quality indicators,a two-stream soft sensor framework based on the above ?GRUs is designed,with a temporal correlation stream and a dynamic causal correlation stream.Finally,the learned features from two streams are fused and a supervised learning regression layer is employed to train the model.This soft sensor model can effectively learn the temporal features of key quality indicators and the causal features associated with process variables.(3)A new soft sensor model based on supervised deep variational autoencoders(SDVAEs)network is proposed to solve the problem of missing data in input process variables.Variational autoencoder is an unsupervised generative neural network used for feature extraction,and its superior ability of reconstruction is good for solving the problem of missing data imputation.Based on this,two novel sub-models are firstly constructed,namely supervised DVAE(SDVAE)and modified unsupervised DVAE(MUDVAE)which are used for predicting key quality indicators and extracting features strongly related to labels respectively.Then a new soft sensor framework can be constructed by combing the encoder of SDVAE with the decoder of MUDVAE.At the same time,we extend the proposed soft sensor framework to handle the missing data situation since our designed VAE has superior ability in data reconstruction which works well under the missing data situation.(4)A new soft sensor model based on quality-driven Gaussian mixture variational probabilistic network(QGMVPN)is proposed for multimodal process data in PET fiber polymerization process.This model takes variational autoencoder as the base glgorithm whose latent distribution is assumed to be diagonal Gaussian distribution.It is too simple to describe the real latent features.So,introducing the Gaussian mixture distribution and invertible Householder transformations to expand the original latent distribution to a non-diagonal Gaussian mixture distribution,which makes it more suitable for complex multimodal process data.Then a new quality-driven soft sensor framework is designed by adding the key quality indicators into the model input and utilizing the gate mechanism connection to improve the structures of encoder and decoder.The proposed soft sensor is not only good at predicting the key quality indicator for complex motimodal data but also greatly good for improving the robustness of the model.At the end,a conclusion is made for this dissertation.Some of the inadequacies of current work is stated and some perspective problems in related field which worth further researching is discussed briefly.