Research On Soft Sensor Modeling Of Chemical Process Based On Deep Learning

The improvement of chemical production efficiency is conducive to the sustainable development of the national economy.In order to achieve good closed-loop control and real-time optimization of product quality,it is necessary to conduct real-time online monitoring of key variables in the chemical process,such as product quality parameters.However,it is difficult to directly measure the key variables through hardware equipment due to production environment and technical reasons.Soft sensor technology provides strong support for realtime prediction of key variables by establishing a mathematical model between the key variables which are difficult-to-measure and process variables which are easy-to-measure.In addition,massive process data are stored in real time according to time stamps,providing guarantee for soft sensor modeling based on deep learning.However,due to factors such as hardware equipment and signal transmission,it is difficult to completely collect process data.At the same time,due to the complex process mechanism and limited offline testing,process data exhibits characteristics such as strong noise,dynamics,and lack of labels.Therefore,this paper takes the deep neural network as the main research method,facing the core issue of soft sensor modeling of chemical processes,and adopts the "incomplete process data inference→Soft sensor modeling of stationary processes→Soft sensor modeling of non-stationary processes "."Measurement Modeling→Soft sensor migration modeling " as the main line of research,with the research goal of improving the accuracy of soft sensor modeling of chemical processes,and proposes a set of soft sensor modeling methods verified by public data experiments and actual chemical applications.The main work of this article is as follows:(1)The complex production environment,frequently interrupted network signals,aging hardware equipment and improper manual operations have resulted in a serious lack of chemical process data.Aiming at the problem that sparse data distribution in chemical processes is difficult to learn,a missing data inference method based on information-enhanced generative adversarial Imputation networks is proposed.First,this method integrates output features into the input through a dynamic feedback mechanism,which enhances the effective information of missing data;secondly,it focuses attention on important observable data through attention weights,enhancing the interactive information between process variables;thirdly,using sample labels as prior conditions for the generator and discriminator enhances the dependency information between process variables and key variables.Finally,through the confrontation training of the generator and the discriminator,the true distribution of the original data is obtained,and effective inference of missing data in the chemical process is achieved.The proposed method is first effectively verified by public data from the University of California,Irvine,thermal power plant data,and debutanizer process data,and then applied to the real polypropylene production process.The results show that the proposed method still has the best stability and calculation accuracy in the case of sparse data.(2)Complex internal operating systems,uncertain external environments and human operating influences bring a large amount of random noise to the chemical process.Aiming at the problem of weak anti-interference ability and low stability of traditional deep neural network methods under stable operating conditions,a stationary process soft sensor modeling method based on low-rank clustering contrastive learning was proposed.First,the method dynamically constructs positive and negative samples of contrastive learning through adaptive clustering,effectively aggregating the similar samples in the time series,and then extracting the inherent consistent and invariant features of the process variables;Secondly,based on the low-rank prior assumption,this method constrains the samples in the same cluster to a low-rank subspace to further enhance the learned feature representation;Finally,the non-convex optimization process of this method is derived based on the expectation maximization algorithm and combined with Transformer to establish a soft sensor model of the stationary process.The proposed method is first verified with thermal power plant data,and then applied to the real purified terephthalic acid solvent dehydration process and single-condition polypropylene production process.The results proved that the consistency invariant feature can effectively resist the complex noise interference.(3)Frequent switching of working conditions,changes in feed materials and operating conditions have caused non-stationary operating conditions in the chemical process.Aiming at the problems of weak generalization and low robustness of traditional deep neural network methods under non-stationary operating conditions,a non-stationary process soft sensor modeling method based on feature decoupled encoder is proposed.This method first designs a novel trend-periodic long short-term memory to extract the decoupled time domain and frequency domain features of the process data;At the same time,a dynamic self-attention convolutional neural network is designed to extract the spatial features of the process data;Secondly,splicing the extracted time domain,frequency domain features and spatial features to obtain decoupled timefrequency-space multiple spatial features.Finally,the decoupled timefrequency-space multiple spatial features are used as the input of the deep residual network to establish a soft sensor model.Verified by Tennessee Eastman process public data and real polypropylene production data,it is proved that the proposd method can effectively overcome the impact of traditional coupling features on modeling accuracy and improve the prediction accuracy and generalization performance of the model.(4)Offline manual testing and expensive equipment maintenance lead to a lack of labels of chemical process.Aiming at the problem that the transfer learning method based on a single space and a single statistical feature is difficult to effectively align the feature distribution of source domain and target domain,a multi-space soft sensor transfer modeling method based on maximizing the mean-variance difference is proposed.Firstly,this method cleverly uses the trend-period short-term memory neural network and the dynamic self-attention convolutional neural network to map the source domain and target domain data into the three spaces of time,frequency and space respectively,which enhancing the expression of characteristics of source domain and target domain.Secondly,the maximizing mean-variance difference algorithm is proposed for the first time to align the feature distribution of the source domain and target domain in multiple spaces,which can achieve effective transfer of knowledge from the source domain to the target domain.Finally,the feature distribution in the source domain and target domain is constrained by a cycle-consistent adversarial loss,which can ensure the reversibility of feature transfer between domains.Verified by public data on activity state identification and multiple sets of real catalytic cracking unit data,it is proven that the proposed method has good migration performance in the target domain.