Soft sensor modeling based on cotraining-style kernel extreme learning machine

Most traditional soft sensor modeling requires the labeled training samples that contain both subsidiary and key variables. However, key variables are difficult to be obtained online due to lack of detection information or high measurement cost. In this paper, a novel semi-supervised learning algorithm, called cotraining-style kernel extreme learning machine, is proposed to exploit unlabeled training samples to reduce the labeling cost. This algorithm employs two soft sensor models trained by kernel extreme learning machine, each of which labels the unlabeled samples for the other during the training process. The confidence in labeling an unlabeled sample can be evaluated by training error which reflects the fitting capability of the soft sensor model and the final prediction is made by combining the estimates by both soft sensors. Industrial application case study shows that the proposed semi-supervised learning algorithm exhibits a good capability to exploit unlabeled training samples, which can improve the performance of the soft sensor.