Multi Time Scale Continuous Wavelet Transform Ensemble Learning for Acoustic Emission-Based Tensile–Shear Fracture Mode Identification

Accurate identification of tensile and shear acoustic emission (AE) modes is important for revealing rock damage–fracture mechanisms and early warning of dynamic hazards. The conventional RA–AF empirical criterion relies on manually defined boundaries and cannot stably separate overlapping point clouds. CWT time–frequency images with deep learning offer an end-to-end alternative, but extraction-window duration is usually set empirically and its independent effect remains unclear. This study proposes a multi-time-scale AE framework (MTS-AE) that combines signal-level multi-scale detection, event-level strictly paired extraction and hierarchically locked partitioning, and representation-level unified CWT parameters, forming a model matrix of eight deep-learning architectures and seven window durations (0.5–10 ms). Odds ratio analysis reveals architecture-dependent class-conditional window preferences: ResNet18 prefers 3 ms for tensile-associated records (OR = 3.86) and 10 ms for shear-associated records (OR = 2.73). Same-architecture cross-window decision-level integration yields positive net gains for all eight architectures, improving ResNet18 from 98.86% to 99.75%; a seven-window ResNet18 ensemble (99.75%) slightly exceeds the full 56-model ensemble under the present evaluation protocol at one-eighth the number of ensemble members. These findings demonstrate the value of complementary temporal representations for AE classification and provide guidance for extraction-window design and multi-window ensemble learning.