Deep learning-based tool wear classification under MQL in Ti-6Al-4V turning using full-field thermal imaging

Abstract

This study presents a deep learning-based framework for classifying tool wear during the turning of Ti-6Al-4V titanium alloy under Minimum Quantity Lubrication (MQL) conditions using full-field infrared thermal imaging. Unlike conventional methods relying on scalar temperature signals, the proposed approach leverages spatial heat distribution patterns to capture wear progression more effectively. Flank wear (Vb) measurements were used to define three physically grounded wear stages: low, medium, and high. A thermal image dataset was constructed from a controlled continuous wear-progression trial, preprocessed through region of interest (ROI) extraction and contrast enhancement, and evaluated using temporal train/test splitting and 5-fold temporal cross-validation to reduce the risk of frame-level data leakage. The study systematically compares end-to-end convolutional neural networks, including ResNet-50, VGG16, and EfficientNet-B0, with hybrid models based on frozen CNN features combined with support vector machine, random forest, and k-nearest neighbor classifiers, as well as a manual thermal feature baseline. Comparative analysis shows that end-to-end optimization significantly outperforms hybrid and hand-crafted approaches. ResNet-50 achieved the highest performance on the natural-distribution test set, with 99.55 +/- 0.60% accuracy and a weighted F1-score of 0.995 +/- 0.006, while EfficientNet-B0 achieved highly competitive accuracy with substantially lower inference latency, making it a promising option for edge deployment. Furthermore, model interpretability analyses using Grad-CAM and t-SNE confirmed that ResNet-50 extracted highly discriminative features that were physically consistent and spatially aligned with the tool-chip interface. The findings indicate that direct deep learning processing of full-field thermal imagery provides an interpretable and highly accurate approach for non-contact tool condition monitoring under controlled MQL-assisted turning conditions; however, broader independent cross-trial validation remains essential before deployment in smart manufacturing environments.