This study proposes a deep learning-based inspection system using the YOLOv8 architecture for the real-time classification and localization of defects in bayonet-bulb base holders. A custom dataset of 357 images was curated that covered various production anomalies and was annotated. The YOLOv8 model was trained and parameter optimization was performed using the Taguchi Grid Optimization Technique. An L18 orthogonal array design was selected to efficiently explore combinations of five critical parameters namely model variant, image size, learning rate, data augmentation and optimizer type. Signal-to-noise ratio analysis for each parameter defined the significance of the model architecture, image resolution and optimizer choice in optimizing mean AP, precision and recall. Our experimental results demonstrated that YOLOv8-based inspection combined with systematic parameter tuning can significantly enhance quality control processes for bayonet-bulb base holders, offering a trade-off between detection accuracy and operational cost of the system.