Selective ensemble modeling load parameters of ball mill based on multi-scale frequency spectral features and sphere criterion

• The ball mill shell vibration and acoustical signals are adaptively decomposed.
• A new adaptive multi-scale spectral feature selection approach is proposed.
• The feature selection parameters of candidate sub-models are selected simultaneously.
• The single-scale frequency spectrum model is extended to multi-scale frequency spectra.

It is difficult to model multi-frequency signal, such as mechanical vibration and acoustic signals of wet ball mill in the mineral grinding process. In this paper, these signals are decomposed into multi-scale intrinsic mode functions (IMFs) by the empirical mode decomposition (EMD) technique. A new adaptive multi-scale spectral features selection approach based on sphere criterion (SC) is applied to these IMFs frequency spectra. The candidate sub-models are constructed by the partial least squares (PLS) with the selected features. Finally, the branch and bound based selective ensemble (BBSEN) algorithm is applied to select and combine these ensemble sub-models. This method can be easily extended to regression and classification problems with multi-time scale signal. We successfully apply this approach to a laboratory-scale ball mill. The shell vibration and acoustic signals are used to model mill load parameters. The experimental results demonstrate that this novel approach is more effective than the other modeling methods based on multi-scale frequency spectral features.