Enhancing Product Quality in Manufacturing: Detecting Centrical Gear Parts Employing a Modified ResNet50 Architecture ResNet50 Architecture

Companies in the manufacturing sector convert raw materials into semi-finished or finished products with market value, significantly contributing to employment and the national economy. In Indonesia, this sector continues to grow, prompting businesses to develop new products and maintain quality standards. However, substandard product quality remains a common issue. This study aims to detect centrical gear parts by improving product quality outcomes in companies using the ResNet50 architecture. We utilized three distinct stages to recognize part models: dataset collection, part model labeling, and training and testing. We evaluated performance based on high-accuracy detection results. The modified ResNet50 achieved an overall class accuracy rate of $\mathbf{9 1} \%$ using a confusion matrix with 10,000 images. The average precision value reached 0.981 against the recall value. Based on the implementation results or training, with an average recall of 1.00 at full confidence level, an average precision of 0.981, and an overall accuracy of $\mathbf{9 1 \%}$, the CNN-based centrical gear part detection model delivered excellent results. The model’s performance was confirmed with evaluations using the F1 score and confusion matrix analysis, making it a reliable tool for industrial applications involving centrical gear part recognition. The model’s application is expected to enhance the accuracy and efficiency of automatic gear part detection processes.