A two-step method for predicting rockburst using sound signals.

Rockburst (equal to strainburst in this paper), a typical catastrophic geohazard, poses a major threat to the safety of employees and infrastructure in deep rock engineering. Fortunately, accidents can be reduced or eliminated by predicting rockburst accurately, and the sound signals generated during rockburst are useful information for the prediction. In this paper, a two-step method for predicting rockburst using sound signals was proposed, which mainly includes two steps. Step 1: the wavelet scattering network is first used to extract the features of sound signals corresponding to the macrocracking phenomena (e.g. particles ejecting, slabbing, and fragments ejecting) during rockburst, and then these phenomena are classified by the integrated usage of the k-nearest neighbour (KNN) classification algorithm and voting. Finally, according to the nature of these phenomena, which occur sequentially before rockburst, the rockburst development stages can be recognized by the classification results. Step 2: the possibility of rockburst occurrence is determined by identifying the rockburst precursors in the evolutionary characteristics of the sound signals, such as which the Mel-frequency cepstral coefficient rate and energy rate both decrease sharply and the amplitude difference activity exhibits a distinct change from a significant decrease stage to a low-level stage for a duration and then to a dramatic increase stage. Furthermore, case studies show that the two-step method accurately recognizes the rockburst development stages and determines the possibility of rockburst occurrence in a multilevel and progressive way. Therefore, it is feasible to predict rockburst under deep rock engineering. [ABSTRACT FROM AUTHOR]