In order to address the difficult problem of ball mill load identification during milling operation, the multi-scale fuzzy entropy algorithm is introduced into ball mill load identification and an innovative ball mill load identification method is proposed- the complete integrated empirical decomposition based on adaptive noise (CEEMDAN)-joint denoising with wavelet thresholding-multi-scale fuzzy entropy biased mean value (PMMFE) ball mill load identification method. Firstly, the vibration signals of ball mill bearings are denoised by the CEEMDAN-wavelet threshold joint denoising method and the analysis reveals that this method has obvious advantages over other denoising methods; secondly, the fuzzy entropy, multi-scale fuzzy entropy, and multi-scale fuzzy entropy deviation of denoised vibration signals are computed, the relationship between each entropy feature and the mill load is analysed in-depth and in an information-rich manner. Finally, the least squares support vector algorithm is used to identify the load of the feature vector. The analysis of the measured vibration signals reveals that the overall recognition rate of this method is 84.4%, which is significantly higher than that of other denoising methods and the combination of feature parameters, and the experiments show that the mill load recognition method based on CEEMDAN-wavelet thresholding-PMMFE is able to effectively identify the different loading states of ball mills.