Some Applications Of Machine Learning Algorithms For Chemical Sotf-sensor Modeling

Soft sensor technology is one of the most important research directions in the area of theprocess monitoring. Because most object properties of the industrial processes is difficult todetermine, and soft sensor method based on mechanism modeling is limited, and the softsensor modeling method based on empirical data has been widely applied. The rapiddevelopment of machine learning theory provides a new theoretical foundation for the softsensor modeling technology based on data. In order to improve the estimation accuracy andgeneralization ability of a soft sensor in the chemical process, this paper is improved from thefollowing aspects to establish the effective soft sensor models:1. Probability kernel learning machine is a new direction in the area of machine learning.An ensemble model is proposed based on Boosting and Gaussian process algorithms. UsingGaussian process as a base learner, a leveraging learner is constructed by Boosting algorithm.The ensemble model is obtained by dynamically averaging the regression functions trained byleveraging learners. Finally, the algorithm is applied to a soft sensor model for a productionplant of Bisphenol A(BPA). Simulation results show that the integration algorithm has higheraccuracy and generalization ability comparing to a single Gaussian process model.2. Naive Bayes algorithm is an effective simple classification algorithm, but it can notuse between classes information effectively. In order to solve this problem, an improvedalgorithm of Naive Bayesian Classifier combined with Weight Kernel Fisher DiscriminantAnalysis is proposed. This algorithm is the key to search the optimal projection matrix ofmaximum separation between classes, and then the original samples are projected transformand new samples are obtained. These new samples are classified by Naive Bayesian Classifier.An on-line monitoring data set from an industrial for Bisphenol-A is classified by the method.Simulation results show that the improved Classifier has better performance of classificationcomparing to a Naive Bayesian Classifier.3. In view of multiple models can signficantly improve model’s estimation accuracy andgeneralization performance. A combination model for soft sensor is presented based onGaussian process and Bayesian committee machine. The original data are classified intoseveral subclasses, and then, the sub-models are built by Gaussian Process Regression. Inorder to get a global probabilistic prediction, Bayesian Committee Machine is used tocombine the outputs of the sub-estimators. Finally, the algorithm is applied to a soft sensormodel for a production plant of Bisphenol A. Simulation results show that the integrationalgorithm can make full use of sample information in the actual production, and the estimatedaccuracy of model is improved, and the generalization ability is better, comparing to thetraditional switch or a weighted combination of multiple model.4. Using integrated learning, a model for soft sensor is presented based on SelectedGaussian Process Ensemble Algorithm. The algorithm selects a number of sub-training setsfrom the training sample using Bagging technique and the sub-training sets are built byGaussian Process Regression. Then the ensemble model is obtained by Bayesian CommitteeMachine with the selected regression functions using Particle Swarm Optimization.