Abstract: The components of shell vibration and acoustical signals of ball mill are complexity and difficult to interpret. Moreover, the useful information contained in these signals is redundancy and complementary, and the mapping relationships between these signals and mill load parameters are difficult to describe. Aiming at these problems, a new soft sensor approach is proposed, which analyzes shell vibration and acoustical signals for modeling mill load parameters based on empirical mode decomposition (EMD) technology and selective ensemble learning algorithm. At first, the complexity of the shell vibration and acoustical signals are analyzed based on the production mechanism. Then, these signals are adaptive decomposed into a number of intrinsic mode functions (IMFs) with different time-scales using EMD technology, and the spectral features of IMFs are analyzed and selected based on the mutual information (MI) method. At last, the selective ensemble learning algorithm based on kernel partial least square modeling approach and the brand and bound optimal algorithm are used to construct soft sensor models of mill load parameters. Thus, the selective information fusion based on multi-source frequency spectrum features is realized. The simulation results based on operating data from the laboratory ball mill validate the proposed approach.