By using the self-developed true-triaxial rockburst testing machine

the rockburst processes were reproduced in laboratory and the sound signals of rockburst process were monitored. The combination index of Meyer cepstral coefficient

spectral centroid and short-time average zero-crossing rate

which can quantitatively describe the sound characteristics

was used as the feature extraction information of typical destructive phenomenon of rockburst process. Then

Gaussian process

a machine learning method for solving small sample

nonlinear classification problems

was used to construct an intelligent identification model. Thus

the intelligent identification of typical failure phenomena in a rockburst process was realized. In addition

in order to overcome the shortage of traditional rock burst prediction methods

which emphasize on trend prediction but can not distinguish the development stage of rock burst process

a multilevel

progressive and dynamic prediction method of laboratory rockburst was developed based on the strategy of intelligent recognition + trend prediction. The variation laws of acoustic characteristic indexes such as quiet period

harmonic mean value and chromatographic vector mean value before rockburst were taken as the precursor information of rock burst. The prediction results of different laboratory rockbursts indicate that the method is feasible and lays the testing foundation of the sound-based method for in situ rockburst prediction in the further.