Development And Applications Of Just-in-time Learning Based Soft Sensors In Fermentation Processes

In many industrial processes,soft sensor technology can be effectively used to replace physical instrument systems and solve the problem that key variables are difficult to measure online because of wicked environment,economic or technical conditions in production process.However,many industrial processes are severely nonlinear and time-varying due to structural characteristics,complex physicochemical reactions,and external factors,which make it difficult to successfully train and apply traditional soft sensors.Just-in-time learning is an effective way of building soft sensors to solve the mentioned problem,which is attracting more and more attention from academic world and industrial world.In this paper,taking the penicillin fermentation process and the industrial erythromycin fermentation process for research objects,several problems of just-in-time learning(JITL)based soft sensors are studied.The specific work is as follows:(1)A novel similarity criterion combining similarity of input and output(SCIO)is proposed.Traditional JITL modeling methods only select relevant samples based on the similarity between input samples,ignoring possible outliers responding to outputs.This paper proposes a novel similarity criterion combining similarity of input and output(SCIO)for selecting relevant samples.In this method,the similarity of input and output variables are firstly calculated by using the traditional similarity criterion based on distance and angle respectively.Secondly,combine the similarity.Finally,the relevant sample set is selected based on the proposed criterion to establish the GPR soft sensor model.For comparisons,the traditional JIT modeling method based on Mahalanobis distance similarity criterion and the distance and angle similarity criterion are also studied.Results show that the proposed similarity criterion can effectively improve the quality of relevant samples and the model has higher prediction accuracy.(2)For easy selection of relevant samples in JITL,Gaussian kernel function is introduced to construct similarity criterion.And a soft sensor modeling method based on Bagging and JITL is proposed(JIT-Bagging).In JIT modeling method,the determination of relevant samples set is critical to the prediction accuracy,but how to determine the scale of relevant samples set remains to be future studied.In addition,traditional JIT soft sensor method usually use all the selected relevant samples to establish a single global model,and for complex nonlinear and multi-mode production processes,the model is limited in prediction performance.This paper proposes a soft sensor modeling method based on Bagging and just-in-time learning.In this method,Gaussian kernel function is firstly used to improve the combined similarity criterion in just-in-time learning.Secondly,a soft sensor model is established by using Bagging technology on relevant samples.Two modeling methods,GPR and PLS,are adopted.A typical nonlinear system and penicillin fermentation simulation process are taken as research objects.Results show that the proposed JIT-Bagging based model has better prediction performance and robustness.(3)A prediction-correction soft sensor modeling method based on moving window and JITL is proposed(JIT-MW).The traditional JIT soft sensor modeling only considers spatial similarity between the query sample and relevant samples but neglects the dynamic characteristics of local production conditions.This paper presents a soft sensor modeling method based on JITL and MW.In this method,a JITL model is firstly established and used to give the prediction of a query sample;Secondly,neighbors of the query sample in a window are collected and used to build a model with the prediction of the query sample;Finally,give estimate of the query sample by using the model.For evaluating the proposed modeling strategy,GPR is used to model penicillin fermentation simulation process and the industrial erythromycin fermentation process.Compared with the traditional JIT-GPR model,the proposed method can achieve higher prediction accuracy.