Remanufacturing requires effective defect detection, but data scarcity and defect variability challenge ML model training. This paper presents a synthetic data generation pipeline focusing on bevel gears, simulating geometric defects like broken teeth, and applying scene randomization to diversify the dataset. The pipeline successfully generates synthetic data for geometric defects, supporting automated segmentation and annotation. However, the current results are limited to geometric defects. Future work will expand to surface and material defects using MDL graphs to improve the pipeline’s applicability. In conclusion, while the pipeline shows potential, real-world validation is necessary to confirm its robustness in remanufacturing.