Study Of Semi-supervised Soft Sensor Modeling Methods Based On Deep Learning

In actual industrial processes,soft sensor technology plays an important role in realizing timely and effective process control,optimization and monitoring.Industrial data are often characterized of strong nonlinearity and complexity,and deep learning methods can learn the deeper essential features of the data.At the same time,the acquisition of target variables is very difficult due to constraints such as the high cost of labeling,low sampling rate,and industrial field environment.However,a large number of process variables that are easy to collect also contain a large amount of system operation information,which can help in the modeling of the target variables.Therefore,semi-supervised soft sensor modeling methods based on the deep learning model are studied in this paper.The main research contents of the paper are as follows:(1)Aiming at the problem that the Deep Belief Network(DBN)model is susceptible to the randomly initialized input weights and biases,a soft sensor model based on the improved Harris Hawks Optimization(HHO)optimized the DBN is proposed.Firstly,in view of the problem that HHO cannot jump out of the local optimum and the blindness in initializing the population,Piecewise chaotic mapping,inertia weight and adaptive t-distribution are introduced to improve it.Then,aiming at the problem that the DBN model is easily affected by the randomly initialized input weights and biases,the improved HHO(IHHO)is used to optimize it,and the IHHO-DBN model is constructed.Finally,the effectiveness of the proposed method is verified on the industrial process data of sulfur recovery.(2)Aiming at the problem that there are limited labeled samples in the actual industrial process data,active learning is adopted to expand the labeled data set,and a semi-supervised soft sensor modeling method based on active sample selection strategy is proposed.Firstly,the conditional error variance calculated by the error Gaussian mixture model is utilized to evaluate the uncertainty of unlabeled samples.Secondly,cosine similarity is utilized to evaluate the representativeness.Thirdly,a comprehensive evaluation index based on the uncertainty and representativeness is proposed.Though offline analysis or measurement by experts and various instruments,the most valuable unlabeled samples are labeled and added to the labeled data set,and then the IHHO-DBN model is established.Finally,the effectiveness of the method was verified on the industrial process data of the debutanizer.(3)Aiming at the problem that most complex industrial process soft sensor modeling can only use limited labeled samples for modeling,a semi-supervised soft sensor modeling method combining cooperative and generative adversarial networks(GAN)is used.Firstly,semi-supervised learning is combined with GAN to achieve the purpose of modeling with unlabeled samples;Aiming at the problem that there is a certain amount of noise in the signal transmission process and sensors in the industrial process,which leads to unlabeled samples with low confidence misleading the training process,two generators are trained in the form of cooperative training,and unlabeled samples with high confidence are selected to participate in the training.Finally,it is verified on the industrial process data of sulfur recovery,and the experimental results show the effectiveness of the proposed method.