Research On Soft Sensor Modeling Based On Active Learning

In the complex industrial process,it is very important to monitor some key variables in real time.However,due to the limited technical conditions,expensive detection devices and bad field environment,these variables are usually difficult to be detected by hardware sensors.In this case,the soft sensing technology is applied,and the mathematical model is constructed through the training set to realize the real-time estimation of the new sample quality variables.Soft sensing technology usually needs a large number of labeled samples to complete highprecision model training.However,in the actual industrial process,the number of unlabeled samples is often large,the number of labeled samples is small,and the acquisition cost is high.Therefore,in this paper,active learning is introduced into soft sensing.Under the active learning framework,according to the characteristics of different methods such as Gaussian process regression,Bayesian extreme learning machine and kernel extreme learning machine,the corresponding sample selection mechanism is designed,which effectively improves the prediction performance of soft sensing model and achieves the purpose of maximizing the performance of soft sensing model with minimum marking cost.The specific research contents are as follows:(1)In order to solve the problems of less labeled samples,more unlabeled samples and higher labeling cost in complex industrial process,an active learning soft sensor modeling method based on subspace integration is proposed.Firstly,principal component analysis is used to integrate the unlabeled samples into subspace.Then,based on the prediction output of all Gaussian process regression sub models,the uncertainty of unlabeled samples is calculated,so as to evaluate the information of unlabeled samples and select the most valuable samples for manual labeling.Finally,by analyzing the function of unlabeled samples and labeled samples,the performance index of training set is introduced to complete the design of termination conditions.Through numerical simulation and application simulation of real industrial process data,it is verified that the proposed method can not only reduce the cost of labeling,but also improve the accuracy of the model.(2)Furthermore,considering the prediction variance of training samples,an active learning regression algorithm of Bayesian extreme learning machine is proposed.This method uses Bayesian extreme learning machine to predict the variance of unlabeled samples,and constructs a sample selection strategy based on the total variance change without considering the change of parameters,so as to improve the generalization performance of soft sensor model.The proposed method is applied to the soft sensing of numerical simulation and hydrogen sulfide concentration prediction.The experimental results show that the proposed method can maximize the performance of the model with the minimum marking cost,which provides a new idea for active learning soft sensing modeling.(3)In order to improve the efficiency of active learning method,a fast active learning method based on kernel extreme learning machine is proposed by combining active learning with kernel extreme learning machine.Firstly,the information of unlabeled samples is evaluated by kernel extreme learning machine,and the confidence of unlabeled samples is taken as the evaluation criterion of sample selection.The most valuable unlabeled samples to improve the performance of the model are selected for labeling.Secondly,considering the operation information of each iteration process,the matrix inversion formula is introduced to optimize the sample selection strategy to improve the efficiency of sample evaluation.Finally,the matrix similarity theory is applied to measure the information of the labeled sample data in the iterative process,and it is used as the basis for the termination of the iterative process to improve the performance of the model with the minimum cost of labeling.The proposed method not only has low marking cost,but also improves the speed of iteration and the performance of active learning algorithm.