Research On Soft Sensor Modeling Based On Affinity Propagation Clustering And Gaussian Process Regression

The actual industrial process generally contains multiple modes,multi-model soft sensor modeling methods are effective for describing the characteristics accurately.The modeling of complex system can be simplified to local modeling through partition,and the performance of the model can be improved effectively.During the modeling process,the performance of soft sensor can be influenced by clustering strategies,modeling approaches and ensemble methods.Because of these issues,this article do some research on multi-model soft sensor modeling methods in actual industrial process from the three aspects mentioned above.The main research contents of the article are as follows:For complicated industrial processes with multiple modes,this article do some research on Gaussian process regression(GPR)soft sensor modeling based on affinity propagation(AP)clustering algorithm.The AP clustering algorithm is applied to divide training samples into different phases,and the GPR method is utilized to construct local prediction models.The final prediction model is obtained by the fusion of different local models based on the weight factor calculated by the distances between new sample and clustering centers.In view of the processes with high dimensions,a multi-model modeling method is proposed based on improved AP algorithm.Firstly,principal component analysis(PCA)and differential evolution(DE)method are applied to promote the performance of AP algorithm,and to remove the influence of the redundant information.The most accurate sub-datasets is obtained based on the optimized parameters.Secondly,the GPR method is used to construct local models.Finally,the prediction variance of the new data is utilized to calculate the posterior probability,and the prediction result is then obtained by combination of different local models.The effectiveness of the proposed method is verified through the simulation results of two benchmark datasets and a sewage treatment process,which has practical reference value in handling modeling problem of high dimensions in industrial process.In order to deal with the degradation of soft sensor models,an online modeling method is proposed based on incremental AP algorithm to update the soft sensor model and database.The training samples are divided using AP clustering algorithm,for the new data,a just-in-time learning(JITL)method combined with GPR method is applied to construct local models,and the online prediction result is obtained by the combination of local models;For the new samples added into the database,the AP clustering algorithm is modified by an incremental method to realize the update of its evidence matrixes.The classification of new samples and update of database can be both accomplished rapidly.The simulation results of penicillin fermentation process indicate that the proposed method is an effective online modeling method.