Infrared thermography coupled with deep learning for fast and reliable predictive monitoring of lubricating oils in dual-use heavy-duty vehicles

The reliable evaluation of lubricating oil condition is critical for ensuring the safety and operational efficiency of heavy-duty equipment in both civilian and defense sectors. Conventional laboratory-based physicochemical analyses, although effective, are inherently time-consuming and do not enable real-time diagnostics or on-site decision-making. In this work, we introduce an innovative approach that leverages infrared thermography coupled with deep learning to achieve rapid, non-destructive, and fully automated classification of lubricating oil samples as either “compliant” (fit for use) or “non-compliant” (unfit for use). The study focuses on two reference lubricants (O-1178 (5W30), gearbox oil and O-1236 (15W40), engine oil) widely deployed in military vehicles, with ground-truth class labels established via standardized laboratory protocols. A comprehensive dataset of over 10,000 thermographic images was generated through controlled cooling cycles, providing the foundation for model development. After comparative analysis of several state-of-the-art convolutional neural network architectures, ResNet-34 and ResNet-50 were selected for their superior performance. The models, trained and validated on stratified and balanced datasets, consistently achieved classification accuracies above 99%, with the ResNet-34 model delivering 100% sensitivity and specificity for the detection of non-compliant samples in both oil types. Complementary metrics, including ROC/AUC (≈1.0) and F1-scores near unity, together with stable training–validation loss convergence, confirmed that the classifiers operated in a saturated performance regime with robust generalization. Interpretation with Grad-CAM heatmaps revealed that the model’s decisions are grounded in physically meaningful thermal micropatterns directly linked to lubricant degradation. This strategy not only minimizes unnecessary oil changes and associated environmental impact, but also elevates predictive maintenance capabilities by enabling immediate, reliable diagnostics in dual-use (civil and military) settings. The proposed methodology establishes a robust and versatile framework for advanced lubricant condition monitoring, readily adaptable to other industrial fluids and diverse operational scenarios requiring rapid, on-site assessment. Future work will extend this framework to additional lubricant types and broader real-world conditions to further consolidate these findings.