Multivariate statistical modelling based on vibration signal analysis was performed at commercial scale grinding. The source digital signals consist of three channels of mechanical vibrations obtained at the axial, horizontal and vertical directions. The feed rate, power draw, pulp temperature were collected automatically by the control system while samples of the feed material and ground product of the ball mill were manually taken to determine the particle size distributions and pulp densities. Using projection to the latent structure (PLS) and/or principle component regression (PCR), empirical models between grinding parameters of interests and the vibration signals were built based on the training data set collected in two weeks, thus the new grinding parameters could be automatically predicted whenever the vibration signals were known. The modelling results show that both the PCR and PLS model can be used to predict grinding parameters online.