Soft Sensing Methods For Key Indexes Of Sintering Process Based On Deep Learning

With the arrival of the fourth industrial revolution,the degree of industrial intelligence and informatization in China is also increasing.The popularity of industrial Internet makes the information flow and communication among people,machines and data more and more important,and the value of industrial big data is increasingly prominent.How to mine the highvalue information contained in it and apply it to the actual industrial process has become the core issue of data-driven industrial process modeling.In this thesis,the iron and steel industry in the typical process industry is selected as the starting point,and the sintering process is taken as the main research object.A soft sensor modeling method based on deep learning is proposed.According to the characteristics of sintering process and data,the thesis focuses on solving the following problems:(1)Aiming at the problems of poor data quality,such as missing data and abnormal data,which are common in the iron and steel industry such as sintering process,a method for recovering the fault data of the temperature sensor of the sintering machine bellows based on the MW-GRU soft sensing model is proposed.This method uses the data-driven soft sensor modeling strategy,and uses the proposed MW-GRU method to predict and model the missing data of the sensor,so as to achieve the effect of fault data recovery.The validity of this method is verified in the experiments of the temperature sensor of the windbox of the sintering machine and the temperature sensor of the hearth of the blast furnace,which provides a guarantee for the data reliability of the subsequent soft sensing.(2)A new neural network framework based on Dynamic time features expanding and extraction(DTFEE)was proposed to solve the problem of dynamic sintering process and time delay of data.In the process industry taking sintering process as an example,the tracking effect of linear model can not meet the needs of actual production.In this thesis,the method of dynamic time feature extension and extraction in sintering process is studied,and a new neural network framework based on time extension feature extraction is proposed to predict the Fe O content in iron ore sintering process.The method considers time delay,time difference,extension and serialization.The network framework can effectively extract dynamic time features,reconstruct long-term industrial data according to time distribution,and make full use of the dynamic features of unlabeled data.The model combines the process characteristics and makes corresponding adjustments.(3)Aiming at the problem that there are too few labeled data in sintering process,the semi supervised soft sensing modeling method of sintering process is studied.In this thesis,an encoder-decoder framework(SS-DTFEE)based on semi-supervised dynamic time feature expanding and extraction was proposed to predict Fe O and other parameters in iron ore sintering process.The framework can effectively use a large number of unlabeled data to obtain industrial process information.The encoder-decoder model is constructed by weighted bilstm,which can better learn the hidden information of bi-directional time series.The unlabeled data obtains the network structure and weight based on the pre training strategy,which contains the process implicit information.Under the fine-tuning strategy,the hidden layer information obtained from the pre training is trained together with the labeled data,and a better prediction effect is obtained than the supervised learning.(4)According to the characteristics of multi-source data in sintering process,the soft sensor modeling method for multi-source data fusion in sintering process is studied.This thesis presents a sintering quality prediction model(MDF-DTFEE)based on multi-source data fusion.In order to solve the problem of too few test variables in sintering process,the video data collected by industrial camera is introduced.This thesis proposes a key frame extraction method based on feature height smoothing,a shallow feature construction method based on sinter layering,and a deep feature extraction method based on Res Net.The above methods constitute a sinter quality prediction model based on multi-source data information fusion,which makes full use of multi-source data information from various sources and effectively improves the accuracy of the sinter quality prediction model.(5)According to the characteristics of industrial Internet platform,a sinter quality prediction system based on industrial Internet platform is developed.This thesis develops a sinter quality prediction system based on industrial Internet platform.Using container technology,the soft sensing model of sinter designed in this thesis is packaged and deployed on the industrial Internet platform.The online model updating strategy based on real-time learning is designed,which has a good effect on the actual sintering production line,provides a good guidance for the actual production process,and improves the intelligence of the sintering process production line.