Extracting acoustic emission features that precede hard rock instability with unsupervised learning

The static and dynamic failure processes of hard rocks were simulated by compression tests to investigate the precursors of hard rock instability. In the tests, acoustic emission (AE) signals were acquired, and their clustering features were automatically extracted by the clustering unsupervised learning method. The results showed that a class of AE signals featured ‘high rise time, high ring count, high energy and low peak frequency and low quantity’ precede hard rock instability and can be used as an indicative precursor feature of hard rock instability. In addition, this class of signals can be used to predict the instability mode of hard rocks after occurrence of the slabbing failure phenomenon: the hard rocks eventually will experience spalling instability if the AE signals occur intermittently and tensile failure signals dominate; or rockburst instability if the AE signals occur continuously and shear failure signals dominate. • Unsupervised learning method was used to extract the features of AE signals. • The cluster analysis of AE signals reveals the precursor signal's characteristics. • The precursor signal can be used as precursory indication of rock instability. • Two rock instability modes of rockburst & spalling can be predicted by clustering.