Study Of Semi-supervised Soft Sensor Modeling Based On Deep Network

It is significant to keep a good measurement and control on key process variables in modern industrial processes,among which some variables are hard or impossible to measure online due to restrictions of measure technology.Therefore,soft sensor which modeling on easily acquired secondary variables is developed to estimate these variables.Soft sensors can make accurate and real-time estimation with convenient maintenance,has becoming a vital component of the process analytical technology and is essential for effective quality control.Traditional data-driven soft sensor methods such as PLS,SVR and PCR are mostly based on statistical theory,and able to perform well on short term steady process status with limited samples.When facing to long term complex process with large input fluctuations,intricate ensemble and adaptive techniques often need to be applied,satisfactory online prediction is hard to achieved.And one of the critical issues,huge amount of low cost unlabeled data records which contain helpful information made no contribution.To improve this situation,considering deep network which perform well on high non-linearity and large dataset,this paper constructed several semi-supervised soft sensor models based on deep network able to incorporate the unlabeled data information and achieve high generalization,robustness and accuracy.Semi-supervised deep soft sensor models are achieved mainly by combining deep unsupervised learning and supervised regression task.Research work is as follows:(1)Reconstruction based unsupervised deep autoencoders as the typical deep representation learning methods can learning abstract and robust disentangled representation which are better for supervised regression task.So semi-supervised soft sensor modeling methods based on deep autoencoder are developed.(2)Industrial data are naturally low dimension manifold on high dimension space,integration manifold embedding and deep neural network by adding a manifold regularization on the loss function of the regression network will obtain a novel semi-supervised model which take advantage of neighbor relationship on large input data.The model can make full advantage of label information to get a better accuracy and high generalization.(3)Generative adversarial network is an effective deep generative model and can learn data distribution well,on the point of generative model based semi-supervised learning,combining GAN with regression network by integrating a regression target on discriminator or adversarial learning on P(X,Y)can use unlabeled data to get high prediction performance.(4)In case study,to evaluating the proposed semi-supervised soft sensor models,experiments on waste water treatment plant benchmark simulation data and ultra-supercritical unit are implemented.