This paper proposes a method for estimating the pose of assembled bolts using Convolutional Neural Networks (CNNs) based on synthetic data. Pose estimation of assembled bolts plays a crucial role in quality control and automation systems in industrial settings, requiring accurate and efficient models. However, labeling real data is time-consuming and costly, particularly for non-intuitive information such as angles. To overcome these challenges, this research generates synthetic bolt images using Blender and applies preprocessing by Laplacian filters to emphasize the outline of the target and reduce the domain gap with real data. Comparative analysis of CNN architectures (3-7 layers) with three pooling strategies revealed that shallower models achieve superior accuracy, with striding and max pooling outperforming average pooling. The 3-layer model achieved 0.70% elevation error and 3.54% azimuth error, while the 5-layer model achieved 0.84% and 3.26% respectively. This research demonstrates that synthetic data-driven CNNs achieve industrial-grade accuracy while reducing data acquisition costs by approximately 90% compared to manual labeling, establishing a systematic framework for selecting optimal architectures based on application-specific accuracy-efficiency tradeoffs.