To give the real-time prediction of accuracy of the ball mill soft sensor, and to solve the model mixing problem in data decomposition of soft sensor. this paper proposed a new Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Gaussian Mixture Model (GMM) and Gaussian Process Regression (GPR) based ball mill load parameters soft sensor method. The key features of this method are using CEEMDAN-GPR to decompose and classify vibration and acoustical signals time domain signals to a series of intrinsic mode functions (IMFs). GPR is used to provide the predicted values. Comparing to the other soft sensor method, the CEEMDAN-based method is largely avoiding the mode mixing issue coming with the original EMD method. Anomalous signals can be classified while feature clustering by giving a probability threshold to the GMM. The GPR-based predicting method will not only provide the data-driven predict values, but also provide their confidence intervals, and warn the operator if necessary. The experiment result shows that comparing to the other soft sensor method, the proposed method has improvements on mill load predicting accuracy and ability of abnormal warning.