Evaluation of CNN Models for Multi-Class Gear Fault Detection Using Waveform Images

Abstract

In this study, the gear fault classification problem, which is of critical importance in industrial mechanical systems, was investigated within the scope of five deep learning models including ResNet18, ResNet34, ResNet50, DenseNet121 and EfficientNet-B0 architectures widely used in the literature. Models were trained on the multi-class gear fault image dataset and their accuracy performances were compared with their numerical values. According to the results, ResNet18 achieved the highest accuracy value with 0.9615, while EfficientNet-B0 showed a similarly strong performance with 0.9594. ResNet34 ranked third with an accuracy value of 0.9541, demonstrating that lightweight ResNet architectures offer high generalization ability in gear fault detection. On the other hand, deeper architectures, ResNet50 with 0.7511 accuracy and DenseNet121 with 0.7500 accuracy, did not provide a significant increase in accuracy despite increasing structural complexity and showed limited performance against the characteristics of the data set. These findings reveal that representation efficiency rather than model depth is the determining factor in gear fault classification problems, and that ResNet18 and EfficientNet-B0 architectures are the most suitable options for real-time fault detection systems.