Modern intelligent methods for information processing are used to estimate the filling level of the ball mills. This problem has been topical for a long time. Moreover, the increasing number of the used ball mills makes it even more important. The main objective is that the ball mill's functioning mode, which is optimal from the point of view of the energy efficiency, can be reached, if the mill is loaded with the ore as much as possible. In its turn, such a mode may cause an overloading in case of any additional ore supply. This leads to the balls' and ore grain's coarse blowout and, as a result, the mill emergency stopping. As a consequence, the mill downtime results in economic losses. The main aim of this research is to develop a method for processing of the vibration acceleration signal from the ball mill's pin in order to discover the hidden dependencies in it. Such dependencies will allow us to estimate the mills' filling level more effectively, but they cannot be discovered by the classical methods, like spectrum analysis of the signal amplitude. A pilot ball mill is used in the laboratory conditions in order to reach such results. A vibration acceleration sensor is set at its pin. A ball load is changed during experiments. A training set for a neural network is formed as a result of the spectrum analysis of the signal obtained from the sensor. The neural network helped to find a relation between the vibration acceleration spectrum and the ball mill filling level. After the conducted experiments a conclusion can be made that such a network is insensitive to the noise caused by a change of the ball load. This insensitivity is higher in comparison with the methods, which are used as a basis for the conventional vibration-acoustic analyzers.