Multiple Model Ensemble Based Soft Sensor Development

Accurate and real-time quality prediction to realize the optimal process control at a competitive price is an important issue in complex industrial processes.Soft sensing technology is an effective way to realize online estimation of quality variables,so it has been one of the main research hotspots in the field of process control.With the popularization of distributed control system and the further improvement of computer storage capacity,data-driven soft sensors have been widely used in complex industrial process modeling and quality prediction.Ensemble learning based soft sensors can effectively explain the nonlinearity and multi-phase characteristics of complex process,track the dynamic changes of production status and operating conditions in real time,so as to ensure the safe,reliable and steady operations of the processes.In this paper,soft sensor development and application are researched under the framework of multi-model ensemble methods.The main layout and contribution of this paper proceed as follows.Traditional single model based soft sensors may have poor generalization performance on quality prediction for industrial processes because of the strong nonlinearity,multiple-phase,and time-varying characteristics.Therefore,a novel soft sensor modeling method with phase partition strategy based on ensemble least squares support vector regression is proposed.Considering the characteristics of batch processes,such as batch-to-batch variations and finite duration,multiway principal component analysis is firstly employed to handle original highdimensional datasets and extract essential correlation information.Then,different operation phases of the process can be identified by the phase partition strategy based on Gaussian mixture model method.Further,multiple localized least squares support vector regression models are constructed to characterize the various dynamic relationships between quality and process variables for local regions.Finally,the posterior probability for each test sample with respect to different phases can be estimated by Bayesian inference strategy,and local outputs are integrated to produce the final quality prediction results.The simulation results of penicillin fermentation process show that the proposed method can significantly improve the overall prediction performance compared with the traditional global models.In order to improve the robustness and adaptive updating ability of traditional multi-model ensemble method in non-Gaussian industrial process modeling,an adaptive soft sensor development with online adaptive ensemble strategy based on partial least square is proposed.On the one hand,the non-Gaussian process data is preprocessed by local weighted standardization method,which makes it approximate Gaussian distribution and can be further used for soft sensor modeling.On the other hand,the moving window strategy is used to update the model adaptively,which improves the generalization performance of the model and ensures its high prediction accuracy in a long working time.Therefore,the model has stronger ability of tracking and interpreting the dynamic changes and phase transition of the processes.When the test sample arrives,the local partial least square models are online integrated to output the final prediction results by applying the Bayesian fusion strategy.The online prediction results of acid gas concentration in sulfur recovery process verify the feasibility and effectiveness of the proposed method.Aiming at the shortage of labeled samples and the high cost of human annotation,a novel active learning framework with hierarchical sampling strategy of ensemble Gaussian process regression is proposed for smart soft sensor design in order to overcome this drawback.The prediction performance of traditional ensemble modeling methods is highly dependent on labeled data samples,hierarchical clustering algorithm is employed to fully exploit the spatial information between samples in a multi-space way.Then,the most representative and valuable unlabeled samples can be iteratively selected with hierarchical sampling strategy.Under active learning framework,labeled datasets are significantly enlarged for traditional ensemble Gaussian process regression model construction at each iteration step.Comparative studies for the penicillin fermentation process demonstrate the reliability and superiority of the recommended smart soft sensing.The cost of human annotation can be dramatically reduced by at least half while the prediction performance simultaneously keeps high.