Research And Application On Soft Sensor Based On Deep Transferable Dynamic Feature Extraction

With the development of modern industry,the control of productive process and product quality are more and more important.It is particularly important to measure and monitor process variables accurately and online.However,in many cases,some variables are in bad conditions,and the direct measurement with actual sensors will lead to measurement lag and high cost caused by sensor loss.Therefore,soft sensor technology emerges as the times require.Its core idea is to build a mathematical model between the process variables measured easily and the quality variable need to be measured,and the model is utilized to estimate quality variable online.With the increase of data accumulated in the industrial process and the development of statistical learning theory,the soft sensor model based on data-driven has become a research hotspot,which owns the advantages of simple model training and fast iterative speed.Nevertheless,it still faces many problems and challenges in actual industrial process applications,such as the process dynamics,the process nonlinearity,the existence of noise in process data,and the scarcity of labeled samplesAt the same time,with the development of artificial intelligence technology,the feature representation ability of deep learning network and the generalization ability of some complex machine learning models are enhanced increasingly.Therefore,this paper has taken the deep transferable dynamic feature extractor built by deep networks and the transferable strategy from feature to the model with strong generalization as the research means,aiming at the process dynamics,nonlinearity,the noise in process data,the lack of labeled samples and application in actual industry of complex models are studied.The main contents of this paper are as follows(1)Aiming at the process dynamics and nonlinearity,the unsupervised deep transferable dynamic feature extractor constructed by gated recurrent neural and encoder-decoder network and the soft sensor model of the ensemble trees based on dynamic feature transfer have been proposed.The dynamic feature extracted is the encoding vector of the model,and it can be explained by theoretical deduction that the dynamic feature is the sequences crossing and nonlinear transformation under the supervision of the decoding network in this model essentially,so as to ensure that the dynamic feature has a positive effect on the regressor.Meanwhile,the unsupervised dynamic features are transferred to the ensemble trees with strong generalization,and the feature extraction ability of deep network and the generalization ability of strong regressor are taken into account,which can improve the prediction accuracy of the soft sensor model effectively.(2)Aiming at the noise in process data,the robust dynamic feature extractor and the soft sensor model of the ensemble trees based on robust dynamic feature transfer have been proposed,that is,a supervised attention network which can indicate noise is added to the original dynamic feature extractor.The attenuation factor is constructed by the weight distribution of attention which can reflect the noise intensity.For the ensemble trees,the attenuation factor is used to smooth the dynamic feature so as to weaken the effect of dynamic feature with noise.The experimental results show that the proposed model can improve the prediction accuracy on the normal data,and reduce the effect of noise on the data with noise.(3)Aiming at the lack of labeled sample,the semi-supervised dynamic feature extractor and the soft sensor model of the ensemble trees based on semi-supervised dynamic feature transfer have been proposed.Owing to the supervised and unsupervised parts in the robust dynamic feature extractor,the proposed model is improved on this basis.The model determines the division rules of labeled and unlabeled sequences,and controls the flow direction of two kinds of sequences in the network to conduct collaborative training for supervised and unsupervised parts,so that the unsupervised dynamic feature extractor is supervised by labels,while the supervised part can use unlabeled samples fully.The experiments show that the model can use the unlabeled sample information effectively when the labelled samples are missing,and the prediction accuracy is improved compared with the supervised model significantly.(4)In order to solve the problem that some complex data-driven models are difficult to be applied in the actual industrial process,a distributed analytical platform for industrial big data,MCBDA for short,is proposed.The platform consists of four parts:cloud sever,industrial field terminal,client terminal and monitor terminal.The cloud server includes load balancer used to configured sub-clusters,distributed database and distributed parallel model library.The model library is the core of MCBDA,which uses three-tier architecture including distributed parallel operator layer,basis model layer and customized model layer.The high-level model can be built by the combination of low-level models and form algorithm interfaces automatically,which make it easy that building complex distributed and parallel models.Meanwhile,the prototyping platform of MCBDA has been designed,and the soft sensor model of the ensemble trees based on dynamic feature transfer which is a kind of customized model has been implemented by the prototyping MCBDA,and a model interface is formed for actual industrial field.