Multi-scale Frequency Spectra Kernel Latent Feature Extraction based on Mutual Information for Modeling Mill load Parameters

Ball mills are heavy key mechanical devices of the mineral grinding process. Load parameters within them relate to production quality and process safety directly. Mill shell vibration and acoustic signals with characteristics of multi-components and non-stationary have been used to measure them. In this paper, a new multi-scale frequency spectra kernel latent feature extraction method based on the shell vibration signal for modeling mill load parameters is proposed. At first, the mill shell vibration signals are decomposed into multi-scale sub-signals with different physical interpretations using empirical ensemble empirical mode decomposition (EEMD) technology. Then, these sub-signals' frequency spectral kernel latent features are adaptively extracted using kernel partial least squares (KPLS) and mutual information (MI) methods with some criterions. At last, these selected features are used to construct soft sensor model with the popular used least squares support vector machines (LS-SVM) algorithm. Simulations based on an experimental laboratory ball mill are used to validate the proposed method.