Fracture mode recognition in hard rocks based on acoustic emission monitoring and hybrid unsupervised-supervised learning

Accurate identification of rock failure modes and instability precursors is critical for the safety of underground engineering. Based on acoustic emission (AE) monitoring, we proposed a clustering-based dynamic ensemble learning model (CDEL-RF) for real-time recognition of fracture modes in hard rocks. First, Brazilian splitting and variable-angle shear tests were conducted on granite and sandstone to establish a class-balanced AE signal database. Using a feature distillation module, eight AE parameters were selected for model learning. Subsequently, the Balanced Iterative Reducing and Clustering using Hierarchies Algorithm was employed to cluster the AE signals based on their intrinsic similarities, establishing cluster-specific random forest (RF) learners, with the dream optimization algorithm dynamically fusing their recognition results. On the mixed-lithology test set, the CDEL-RF model achieved a recognition accuracy of 0.9840 for tensile and shear fractures. After extending the CDEL framework to multiple base learners, their average recognition accuracy increased by 4.4%. Based on the model, shear fracture proportion (SFP) was proposed as a complementary acoustic indicator for characterizing rock failure states. It is defined as the average probability of all signals within an event window being identified as shear fracture. In uniaxial compression tests on granite and sandstone specimens with cavities, SFP accurately predicted particle ejection from the hole wall and quantified the transition of rock failure modes.