Soft Sensor Modeling Based on Multiple Extreme Learning Machine

Abstract

Presented is a multiple model soft sensing method based on extreme learning machine (MELM) algorithm, to solve the problem that single ELM model has lower predictive precision and over-fitting problems. The method adopts Gaussian process to choose secondary variable for soft sensor model. Then, samples data are divided into several groups of data by adaptive affinity propagation clustering, and the sub-models are estimated by ELM regression method. Finally, ELM is regarded as output synthesizer of sub-models. The proposed method has been applied to predict the end point of crude gasoline in delayed coking unit. Compared with single ELM modeling, the simulation results show that the algorithm has better predictive precision and good generalization performance.