Frequency-domain dielectric analysis and machine learning classification of heat transfer oil degradation states

Heat transfer oils (HTOs) appears as a critical fluid in industrial thermal systems because it offers high-temperature stability and operational safety compared to conventional heat transfer media. However, the efficiency and safety of operation, and equipment longevity is often compromised by the degradation of HTO. HTO suffers from oxidation, cracking, and contamination because of long term operation. Traditional laboratory analyses such as viscosity, total acid number (TAN), and Karl Fischer titration are usually performed to identify the oil quality, it is effective but time-consuming. So, it is unsuitable for in-situ measurement. This study presents a non-destructive dielectric-based approach for assessing HTO quality by using an open-ended coaxial probe coupled with machine learning classifiers. Dielectric constant, ε r ’ were measured over 1–9 GHz for fresh and used samples collected from an incineration plant operating with Shell Heat Transfer Oil S2 (Shell Global). Statistical analysis identified 1.75 GHz and 1.84 GHz as optimal diagnostic frequencies, showing strong separation among good, “Medium”, and degraded oils (p < 0.001). This work has applied three classification models, namely Support Vector Machine (SVM), Random Forest (RF), and k-Nearest Neighbour (kNN) for training of oil classification based on dielectric data and viscosity-based labels. The dataset was partitioned into training and testing sets. The training phase utilized 85 samples (35 “Good”, 20 “Medium”, and 30 “Bad”), while 18 samples (8 “Good”, 7 “Medium”, and 3 “Bad”) were reserved for testing. Based on the analysis, it was found that SVM achieved 100% precision, sensitivity, specificity, and accuracy at 1.84 GHz. The outcome of this work highlights the potential of integrating dielectric sensing with machine learning in predicting the HTO quality. It provides the proof-of-concept and theoretical framework intended for future real-time, in-situ applications.