Research And Application Of Data-driven Soft Sensor Method For Complex Industrial Processes

Complex industry process is an important component of manufacturing industry and a pillar industry for China's national economic and social development.In traditional industries such as metal smelting and mineral processing,a large number of key process parameters cannot be detected online,resulting in poor comprehensive control level and facing huge industrial upgrading pressure.In order to improve the competitiveness of modern industry,online monitoring and stable control are crucial to ensure production safety and improve product quality.In this dissertation,data-driven soft sensor technology is applied to solve the issues existing in industrial process monitoring and control,which will lay a technical foundation for the realization of complex industrial informatization and intelligent development.The following aspects are mainly studied:1)Batch process is often characterized by strong non-linearity,and multi-stage,the convention single model cannot effectively capture the multi-stage characteristics of the batch process and the transition characteristics between two stages.Therefore,a novel multi-model soft sensor method based on Gath-Geva clustering and kernel extreme learning machine(KELM)is proposed.Firstly,principal component analysis(PCA)is used to extract features from the input variable,and then the Gath-Geva algorithm is used to identify the different operating phases of the batch process.Subsequently,local KELM models are built for each identified phase.Furthermore,for each query sample,the final model is developed by integrating fuzzy membership as sample weight and local KELM prediction.The proposed method is applied to a multiple penicillin fermentation process,and the experiment results show that the proposed multi-model approach has higher prediction accuracy than the single model.2)In order to address the outlier issue of batch process,a novel prediction model based on ensemble random vector functional link network(AB-RRVFLN)is proposed.First,the mutual information method is proposed to select the optimal variables for prediction model.Then,considering the outlier influence of industrial data,a robust random vector functional link network(RRVFLN)based on the iteratively reweighted least squares is developed to build the prediction model.Moreover,AdaBoost strategy is integrated with RRVFLN model to enhance prediction performance.Finally,the model output is further corrected by the offset compensation technique.Based on industrial data from shuttle kiln,the application results on sintering temperature prediction demonstrate that the proposed ensemble method achieves better accuracy and shows stronger robustness than other methods.3)Conventional multi-model approach cannot capture the time-varying behavior and batch-to-batch variation of batch process.In this study,a novel online ensemble extreme learning machine method(OEELM)is proposed to predict quality variable for time varying batch process.Firstly,Gaussian mixture model(GMM)is utilized to identify different operating phases.Further,extreme learning machine(ELM)models are built to characterize the different dynamic relationships within various operating modes.Meanwhile,the posterior probabilities of new test sample with respect to different modes can be estimated by Bayesian inference strategy and used t o incorporate multiple localized ELM models into a global model for quality variable prediction.In addition,the proposed OEELM method provides four types of adaptation strategies to update model parameter including adaptation of GMM parameter,local ELM model updating with incremental learning,mean and variance updating,online offset compensation.The effectiveness of OEELM approach and its superiority over conventional adaptive soft sensor methods is demonstrated through a simulated fed-batch penicillin fermentation process.The experiment results confirm that OEELM method has better prediction performance than traditional adaptive methods,as well as fast learning speed.4)In order to handle strong nonlinearity and time-varying characteristic of continuous process,this dissertation proposes a novel JRELM adaptive soft sensor approach based on just in time learning(JITL).Firstly,considering redundancy between input variables,PCA is used for feature selection.Then,in the JITL framework,a local regularized ELM(RELM)model is built for online prediction.Meanwhile,an efficient adaptive optimization algorithm is proposed to determine RELM model parameters and enhance model prediction ability.Furthermore,moving window strategy is adopted to update historical databases.Finally,offset compensation technology is utilized to correct model output and improve prediction reliability.Based on the industrial data from an alumina enterprise in China,the experiment result comparison show that the prediction accuracy of the proposed JRELM is much better than other adaptive approaches,as well as the higher online efficiency.5)In alumina rotary kiln production,adjusting coal-feeding quantity is the main way to maintain the sintering temperature stability during sintering process,which plays a critical role in improving production quality and saving energy consumption.In this dissertation,a novel integrated method(termed PSR-PCA-HMM)is proposed to predict coal feeding sate for optimal control by integrating PCA and hidden Markov model(HMM)based on phase space reconstruction(PSR).First,PSR is utilized to dynamically choose input vectors of prediction model and C-C(Cheng Church)is selected to determine optimal delay time and embedding dimension.Second,PCA is proposed to efficiently reduce redundancy of the high-dimensional feature space reconstructed by PSR.Then,considering nonlinear dynamic characteristic of sin tering process,three HMM models are respectively built to capture the nonlinear dynamic relationship between thermal variables and corresponding coal feeding state.For any test sample,the posterior probabilities with respect to three HMM models can be estimated by using the forward algorithm.The final prediction of coal feeding is determined by the maximized likelihood estimation.Based on the field industry data,the application results indicate that the proposed method can significantly improve prediction performance and outperform in comparison with PCA-BP,PCA-LSSVR and PCA-HMM.