Explainable assessment of military lubricant quality using learnable edge functions in Kolmogorov–Arnold Networks

Reliable assessment of in-service lubricant condition is essential for predictive maintenance in heavy-duty and military vehicle fleets, where undetected oil degradation can lead to equipment damage and reduced operational readiness. This work investigates Kolmogorov–Arnold Networks (KANs), a class of neural architectures with learnable edge-wise univariate functions, for assessing military engine lubricant quality using real operational data from the Spanish Army. KAN-based models are evaluated under a unified experimental protocol and compared with multilayer perceptron (MLP), support vector machine (SVM), and gradient boosting (XGBoost) baselines, considering realistic constraints of class imbalance and limited data availability. Results indicate that while overall accuracy and F1-score are similar across architectures, several KAN variants achieve higher recall for non-conforming lubricant samples, reaching up to 0.971. This improved sensitivity to degradation events is particularly valuable in safety-critical maintenance scenarios, where missed detections are more costly than false alarms. KAN-based models also achieve competitive balanced accuracy and area under the Receiver Operating Characteristic (ROC) curve, demonstrating robustness under asymmetric error costs. Beyond predictive performance, the learned edge-wise functions provide an intrinsic glass-box representation of the diagnostic process, enabling direct inspection of feature-specific response regimes. These responses qualitatively align with American Society for Testing and Materials (ASTM) based viscosity and contamination thresholds defined in the O-1236 specification, supporting expert validation and auditability. Overall, the results position KANs as a transparent and operationally suitable decision-support tool for lubricant condition monitoring in high-assurance engineering environments.