Modeling parameters of mill load based on dual layer selective ensemble learning strategy

Abstract

Different frequency spectral feature sub-sets of mill shell vibration and acoustical signals contain different information for modeling parameters of mill load. Selective ensemble modeling based on manipulate training samples can improve generalization performance of soft sensor model. Based on the former studies, we proposed a new dual layer selective ensemble learning strategy. At first, vibration and acoustical frequency spectral feature sub-sets are extracted and selected by the methods in literature [15]. Then, selective ensemble modeling method based on genetic algorithm and kernel partial least squares (GASEN-KPLS) is used to construct the first layer selective ensemble model for every feature sub-set. Finally, brand and band (BB) and adaptive weighting fusion (AWF) algorithm is use to select and combine the outputs of the first layer models to construct the second layer selective ensemble model. Results indicate that the proposed approach can perform reasonably well on estimate mill load parameters of a laboratory ball mill grinding process.