Deep Transfer Learning and Data Augmentation for Automatic Flank Wear Detection and Classification

Flank wear in cutting tools remains a critical issue in the metal cutting industry due to its direct impact on dimensional precision, surface integrity, and overall manufacturing efficiency. Failure to detect or accurately classify wear may result in premature tool replacement or extended use of worn tools, leading to increased operational costs, heat accumulation, machining vibrations, and potential damage to workpieces and equipment. This study provides a comparative analysis of four state-of-the-art deep convolutional neural network (CNN) architectures VGG16, GoogleNet, DenseNet121, and MobileNetV2 for the automatic classification of flank wear across four predefined severity levels. To overcome dataset limitations and improve model adaptability, transfer learning was employed by fine-tuning pre-trained models on a domain-specific dataset of tool wear images captured using a DM9 industrial digital microscope under dry turning conditions. A comprehensive data augmentation strategy was also applied to enhance model generalization and mitigate overfitting. Experimental results show that GoogleNet and DenseNet121 achieved the highest classification accuracy of 96.80%, with GoogleNet being favored for its shorter training time and computational efficiency. These findings underscore the effectiveness of deep learning-based approaches for reliable, scalable, and real-time tool condition monitoring in advanced manufacturing environments