This article presents a study on applying artificial intelligence, specifically YOLOv8x and YOLOv11x (You Only Look Once) models, for detecting three types of faults (healthy, broken, and missing teeth) in rotary tedders under data-constrained conditions. We trained both models on three progressively augmented datasets (V1-V3) derived from limited video footage, applying rotations, color alterations and filtering to simulate variance. Our results show a clear correlation between data volume/quality and model per-formance: mean Average Precision (mAP0.5) improved from 62–67% (Dataset V1) to 82–85% (Dataset V3). While YOLOv11x achieved marginally better overall accuracy (85% vs. 82%), both architectures struggled with background confusion, particularly for the "healthy tooth" class. This work demonstrates that with strategic data augmentation, off-the-shelf YOLO models can achieve promising detection accuracy even with limited initial data, providing a practical baseline for real-time agricultural fault monitoring systems.