Study Of Soft Sensor Modeling Based On Deep Learning

In modern industrial processes,some important process variables that are key indicators of process performance are difficult or impossible to measure online due to the limitations of process technology or measurement techniques.Therefore,soft sensor is employed to estimate these process variables which are difficult to measure online by modeling secondary variables with a high degree of accuracy and convenient maintenance.Soft sensors are widely used to estimate process variable and can estimate in real-time.Soft sensor have attracted increasing attention and been regarded as a valuable alternative to the traditional means for the acquisition of critical process variables,process monitor and process fault detection.At present,soft sensing modeling methods based on data-driven include regression based on model,artificial neural networks(ANN),Gaussian process,support vector regression(SVR)and hybrid methods and so on.Deep learning is a promising method to extract useful feature representation automatically and improve the prediction performance and generalization performance with a large number of samples.In this paper,deep learning methods are employed to some important issues in soft sensor modeling and applications must be resolved,e.g.,missing data,co-linearity,semi-supervised learning and so on.The main research work is as follows:1.This paper introduces the primary model methods and development of soft sensor technology.Besides,this passage also introduces the theory,main model and cutting-edge technology of deep learning.2.In this paper,a novel soft sensor modeling method based on deep learning is proposed to handle the issues in soft sensor modeling such as missing data,co-linearity.The method integrates denoising auto-encoders(DAE)with support vector regression(SVR).The denoising auto-encoders are designed to capture robust high-level feature representation of import data.And the SVR model precisely estimates output data based on the feature representation generating by DAE.In case study,the method combining denoising auto-encoders with support vector regression(DAE-SVR)is applied to the estimation of the total Kjeldahl nitrogen(TKN)in a wastewater treatment process.The experimental results show that the proposed modeling method provides a new effective method for soft sensor modeling.3.This paper proposes a novel soft sensor modeling method based on a deep learning method that integrates denoising auto-encoder(DAE)with a neural network(NN)method(DAE-NN).The stacked denoising auto-encoders are trained to obtain pre-trained weights for the supervised neural network,with the benefit of the neural network model achieving better generalization and avoiding overfitting to some extent.In a case study,a DAE-NN based soft sensor is applied to the estimate the total Kjeldahl nitrogen(TKN)in a wastewater treatment process.The experimental results show that the proposed modeling method provides a new effective method for soft sensor modeling.4.The proposed DAE-NN modeling method can address unlabeled data and employed on semi-supervised learning.The unlabeled data can add to pre-training stage to capture the essential information and structure of input data.The prediction performance and generalization performance are improved in this way.5.This paper proposed weighted DAE-NN modeling method on semi-supervised learning.The DBSCAN(Density-Based Spatial Clustering of Applications with Noise)method is employed on cluster analysis of unlabeled samples.The samples in low density space is assign to lower weight to improve the prediction performance of model.6.In case study,the deep learning methods which this paper proposed are applied to the estimation of oxygen-content in flue gasses in ultra-supercritical units.The results show significant improvement of the deep learning methods than the conventional method in predicting output measurements.Deep learning is a promising method to design and model soft sensors in process industry.