Adaptive Soft Sensor Modeling Based On Local Learning

In modern process industries,in order to meet the increasing demands of raising the economic benefits and enhancing the process safety,stability and environmental protection,soft sensors have been widely used for online estimations of primary variables.Motivated by improving the soft sensor models' estimation accuracy and prolonging their service life,under the framework of local learning this dissertation focuses on the just-in-time learning(JITL)and the ensemble learning,and aims at developing adaptive soft sensors that can simultaneously deal with process nonlinearities and time-varying issues.In this dissertation,solutions to certain involved crucial problems are proposed and simulations and analyses are carried out.Main contributions of this dissertation are summarized as follows.In order to sufficiently take advantage of labeled samples' supervision information when constructing similar sample set in the JITL,a supervised structure preserving projectionsbased JITL method for adaptive soft sensor modeling is proposed,where the JITL is completed in the feature subspace of secondary variables,such that the sample's structure information can be preserved,meanwhile the supervision information can be utilized.Moreover,a database monitoring strategy is designed to cut down the update frequency of the database.Synthetic data and data from a real-life industrial process verify the effectiveness of the proposed method.To tackle some problems in the traditional localization methods,an adaptive process state partition method is proposed based on the statistical hypothesis testing theory.The proposed method constructs statistical indexes which can test if the first-order and the second order information of the predicted residuals deviate significantly or not.This localization method outputs a series of independent local model domains,and accounts for the mapping relationship between secondary variables and primary variables.It has some additional advantages such as the easy online augmentation of the local model set and the adaptive determination of the local model number.The effectiveness of this localization approach is preliminarily illustrated by the industrial debutanizer column process and the sulfur recovery unit.An adaptive soft sensing method based on online switch of local models is developed,which is the foundation of a subsequently proposed selective ensemble learning framework for adaptive soft sensor modeling.In the online local model switch strategy,a quantitative index of the generalization ability for each local model is constructed by exploiting the newest labeled sample in the historical dataset and the labeled neighbors of the query sample,and the primary variable is estimated through one selected ‘optimal' local model.In the selective ensemble learning framework,the local model switch criterion is used for measure the satisfaction degrees of local models,according to which those local models which can not enhance the estimation performance in the ensemble learning are filtered.Simulations on the simulated dataset on a continuous stirred tank reactor,and datasets from the real-life debutanizer column process and the sulfur recovery unit verify the effectiveness of the proposed method.For incorporating the information of unlabeled samples and providing performance assessment model with high versatility for the ensemble learning,a semi-supervised selective ensemble learning based soft sensor modeling approach and a performance assessment model are proposed.In the semi-supervised ensemble learning,a formula for calculating the Distance to Model(DM)is defined to measure the relevance between the sample and the model which can compute the membership of the query sample with respect to each local model more precisely.Then the selective ensemble learning is performed,where the satisfaction degree is calculated on the basis of the DM.Furthermore,the performance assessment method which measures the prediction accuracy of the ensemble learning is independent of the specific regression algorithm,and is therefore highly versatile.Simulation results on the debutanizer column process confirm the effectiveness of the semi-supervised ensemble learning and the performance assessment method.