Semi-supervised Soft Sensor Modeling Based On Help-training Strategy

In the actual industrial production process,in order to ensure product quality and production efficiency,it is necessary to monitor some dominant variables that are difficult to measure directly or in real-time.Soft sensor technology can indirectly achieve effective estimation of dominant variables by mining the relationship between easily measured auxiliary variables and dominant variables to establish mathematical models.With the continuous development of industrial informatization,data driven soft sensing technology has been widely used in complex industrial process modeling.However,in actual industrial production processes,there are often a large number of unlabeled samples with high nonlinearity and multi coupling characteristics,resulting in traditional soft sensor models often difficult to meet production requirements.Therefore,this paper combines semi supervised learning ideas and conducts research on soft sensor modeling methods for complex industrial processes based on help-training strategies.The main research contents of this paper are as follows:(1)This paper introduces the help-training strategy into soft sensor modeling,summarizes its research status,and analyzes in detail the learner model and sample selection strategy in the help-training process based on the help-training strategy.Through a numerical simulation case,the feasibility and effectiveness of the help-training algorithm are verified,laying a foundation for subsequent research.(2)In order to fully utilize the information of unlabeled samples in complex industrial processes and reduce the impact of process uncertainties on the quality of unlabeled samples,a help-training soft sensor modeling method based on dual learner collaboration was proposed.Using a twin support vector regression mechanism to build a master learner,false tags are added to high confidence unlabeled samples;At the same time,a secondary learner is constructed based on the K-nearest neighbor algorithm to maximize the mean square error of the learner on the nearest neighbor sample set.The sample set to be processed after screening by this indicator contains more data information;The main and auxiliary learners complement each other,improving the generalization of the model to a certain extent;Using the constructed helptraining framework to improve sample utilization,a prediction model is obtained to fully mine the information of unlabeled samples.By modeling and simulating the actual industrial process data of a debutanizer,the results show that the model has good predictive performance.(3)Due to the problems of nonlinearity,multistage coupling,and a small number of labeled samples in complex industrial processes,traditional global soft sensor models are often difficult to accurately describe the entire process.Therefore,a multi model soft sensor modeling method based on help-training strategy is proposed.This method uses fuzzy C-means clustering algorithm to mine similar samples in the sample set and establish several sub models;By introducing a help-training strategy,a collaborative training framework based on primary and secondary learners is formed,and a confidence evaluation mechanism is designed to eliminate error samples while expanding the modeling space of the sub-model;Then,the fuzzy membership degree is used as the probability distribution function of D-S evidence theory to calculate the sub model weight,and the sub model output is fused to obtain the final model prediction result.The simulation results of real industrial process data show that the model has good prediction performance.(4)The harsh and complex production environment,variable working conditions,and fluctuations in equipment operation in actual industrial processes can lead to issues such as missing data sample information and inconsistent dimensions.Therefore,after analyzing and summarizing the soft sensor modeling process,an industrial data analysis and modeling software was designed,which includes functions such as data import,missing value supplementation,data standardization,data visualization,and model establishment.By embedding the proposed algorithm into the software,can achieve concise and efficient process based soft sensor modeling.