Inverse Analysis of Internal Crack Propagation Depth in Cantilever Anti-Sliding Piles Based on Artificial Intelligence and XFEM

Cantilever anti-sliding piles are critical for slope stability, yet their subsurface health assessment remains challenged due to the invisibility of internal damage. This study proposes a non-destructive inversion framework that integrates computer vision with computational mechanics to quantify internal crack propagation depth based on surface morphological features. To address the limitations of the standard Artificial Bee Colony (ABC) algorithm, specifically its slow convergence and susceptibility to local optima—an Improved ABC (I-ABC) algorithm is developed. This enhancement incorporates a neighborhood-based optimal guidance mechanism and an adaptive selection probability strategy to strike an optimal balance between global exploration and local exploitation. The proposed framework first employs an optimized YOLOv8 model, with hyper-parameters tuned via the I-ABC algorithm, to extract surface crack parameters. These parameters are subsequently mapped to internal depths using a normalized objective function derived from an Extended Finite Element Method (XFEM) forward modeling database. Experimental validation, utilizing a hybrid dataset comprising physical pile tests and numerical simulations, demonstrates that the proposed method achieves a crack depth inversion accuracy of 96%. Furthermore, statistical comparisons reveal that the I-ABC algorithm significantly outperforms Particle Swarm Optimization (PSO), Differential Evolution (DE), and Grey Wolf Optimizer (GWO) in terms of robustness and convergence speed. Ultimately, this framework provides a reliable and automated tool for the structural health monitoring of geotechnical infrastructure.