The vibration signals of ball mill bearing are found to be highly divergent and strongly stochastic when being used as a parameter of fill level. Therefore, based on the cloud model, a mathematical tool which has the property of stable tendency and the ability to organically combine the fuzziness and the randomness of the data, a method is proposed to represent the concepts of fill level and efficiently measure the fill level in ball mill. Firstly, the antecedent cloud models are obtained by using normal backward cloud generator to extract the linguistic concept from characteristic sequence generated from the power spectral density(PSD) of the vibration signals. Then the consequent clouds corresponding to the antecedent clouds are figured out by employing the fill level information of the training samples. Finally, the soft sensor of the fill level is realized by uncertainty reasoning based on the cloud model. The comparison experiments show the effectiveness and feasibility of the proposed method.