Semi-supervised Learning Based Smart Soft Sensor Modeling

In the actual industrial production process,product quality,production efficiency and reliable operation of operating equipment are greatly dependent on the real-time measurement of key quality variables(dominant variables).However,it is usually difficult to measure quality variables directly by using hardware sensors due to the harsh environment,monitoring technology and economic costs.In this case,the quality variables can be predicted and estimated via constructing soft sensor models.To handle the non-linearity and multi-stage characteristics of industrial processes,an affine information reconstruction based semi-supervised extreme learning machine soft sensor method is proposed.Firstly,the training dataset is divided into several subsets by nearest correlation spectrum clustering.Then,under the semi-supervised learning framework,the affine information is used to reconstruct the Laplacian matrix of the extreme learning machine,and the predicted results of each sub-model are fused by Bayesian posterior probability to obtain the final output value.The numerical simulation and actual steelmaking process simulation show the effectiveness of the proposed method in dealing with the multi-stage modeling problem with the missing dominant variables.Combining ensemble learning with semi-supervised learning,an adaptive ensemble semi-supervised soft sensor modeling method is proposed.Firstly,the k-nearest neighbor method is used to pseudo-label the unlabeled samples,and the quality of the pseudo-labeled samples is evaluated and selected to obtain the optimal selected dataset.Then,the labeled dataset and the selected sample set are simultaneously generated by the Bagging algorithm,and the two kinds of sample subsets are adaptively matched by the dissimilarity algorithm.The process of sample selection and adaptive matching in the proposed method not only guarantees the diversity of ensemble learning,but also improves the accuracy of ensemble learning,thus the prediction accuracy and generalization ability of soft sensor model are effectively improved.In order to cope with the instability of model performance improvement in active learning,an active learning soft sensing method based on approximate linear dependence is proposed.The proposed method corresponds the additional information contained in unlabeled samples to the minimum representation error between vector groups.The quality of unlabeled samples is evaluated by approximate linear dependency algorithm,which provides the most informative samples for model training,and achieves more efficient performance improvement with less labeling cost.In the experiment of sulfur recovery process,compared with the existing random selection strategy,the active learning strategy based on approximate linear dependence shows higher prediction accuracy and faster convergence speed.