| During chemical production processes,key parameters directly related to product quality and process state cannot be measured online,severely affecting the monitoring,control,and optimization of modern industrial processes.As an important indirect measurement method,soft-sensing technology builds a mathematical model based on a certain mathematical regression relationship to realize online estimation.However,traditional offline batch learning soft sensing cannot cope well with the complex features of nonlinearity,time variation,and label scarcity that exist in industrial processes.In addition,data in actual industrial processes often generate rapidly,in large amounts,and as a single flow.To address these issues and establish a soft sensing modeling method that is more closely related to actual industrial processes,this article focuses on industrial data streams with concept drift and researches online adaptive soft sensing modeling methods.The main work is summarized as follows.(1)To solve the problems of nonlinearity,time variation,and label scarcity in industrial process data streams,a semi-supervised soft sensing modeling method based on online clustering analysis of data streams is proposed.This method achieves online dynamic recognition of process states through online dynamic clustering.Then,an adaptive switching prediction strategy is used to handle slow and abrupt changes in process characteristics.In addition,a semi-supervised learning strategy is introduced to augment the labeled training set to fully utilize the information of unlabeled samples.(2)To address the problem that traditional local learning soft sensing modeling methods are prone to ignoring the latest sample information on the time axis and require a large number of historical samples,an online semi-supervised data stream soft sensing modeling method based on spatiotemporal multimodel ensemble is proposed.This method considers both time and space perspectives,and uses instant learning and self-organizing incremental neural networks to identify the most relevant historical local information to the current query,and then uses an ensemble strategy to merge heterogeneous base models to obtain the final output.Finally,a semi-supervised strategy is introduced to expand the training set to improve prediction performance.(3)To address the two issues of wasting resources caused by discarding historical models and possibly losing local information that is more useful for current prediction,an online data stream soft sensing modeling method based on adaptive selective ensemble is proposed,with a focus on the effectiveness of historical base models for current prediction.This method retains historical base models obtained by local learning,introduces a performance ranking strategy,and then combines the adaptive selective ensemble framework to merge the results of base models.The proposed method solves the complex characteristics of nonlinearity,time-varying,label scarcity,historical local information loss,and time correlation loss in data stream soft sensing modeling.The effectiveness and superiority of the proposed method have been verified through the Tennessee Eastman chemical process and the aureomycin fermentation process. |
