Prediction of shear process in granular geomaterials using acoustic emission

The progressive failure of granular geotechnical materials acts as a critical precursor to landslide occurrence. To explore the feasibility of acoustic emission (AE) for early warning of such failures, direct shear tests were performed under eight shear rates to reproduce landslides with different deformation rates. AE signal analysis reveals that energy, ring down count, b-value and spectral centroid (Sc) evolve systematically throughout shearing. Specifically, the spectral centroid redshifts from 350 to 250 kHz, marking the development of macroscopic fractures. Integrating these AE signatures with shear stress evolution, the shear process is categorised into four stages: densification, elasticity, strain hardening and residual strength. Using 11 AE features, a convolutional neural network–long short-term memory hybrid deep-learning model embedded with spatial attention is constructed to automatically recognise shear stages, reaching a prediction accuracy of 94.52%. The attention module emphasises pre-failure key features and enhances model performance and interpretability. It is demonstrated that AE monitoring combined with deep learning can effectively identify precursor signals of shear failure in granular media under laboratory direct shear conditions, providing a preliminary methodological reference for the future development of landslide early warning systems.