Many difficult-to-measure process variables of the industrial process, such as ball mill load, cannot be measured by hardware sensors directly. However, the frequency spectrum of the vibration and acoustical signals produced by the industrial mechanical devices contain information about these variables. An ensemble modeling approach based on the partial least square (PLS), least square support vector machines (LSSVM) and the information entropy is proposed to estimate the load of the wet ball mill. At first, the PLS algorithm is used to extract the latent features of the mill power, vibration and acoustical spectrum respectively. Then, the extracted latent features are used to construct the mill load sub-models. At last, the finals ensemble model is obtained based on the information entropy of the prediction errors of the sub-models. Studies based the laboratory-scale ball mill show that the proposed modeling approach has better predictive performance.