Multiple Regression Machine System Based on Ensemble Extreme Learning Machine for Soft Sensor

During the practical production process, the soft sensors based on traditional single learning machine can not satisfied the needs of production. In this study, a multiple regression machine system (MRMS) is proposed to establish the soft sensor for overcoming the above shortage. The regression machines will be aggregated by ensemble algorithm to improve the performance of soft sensor, especially to enhance the accuracy of prediction. In the MRMS, a novel three-layer feed-forward neural network Extreme Learning Machine (ELM) with random weights between the inputs and the hidden units is selected as the BaseLearn for its better accuracy of predictability and fast learning speed. An improved ensemble algorithm for regression problem is presented to combine the ELMs. The real production data from 300 t LF in Bao-steel Co. Ltd. are used to train and test the new soft sensor based on MRMS. The results of experiments demonstrated that the new MRMS method can improved generalization performance and boost the accuracy, and the accuracy of the soft sensor temperatures is satisfied for the needs of practical production.