A compact non-contact microstrip sensor integrated with lightweight CNN for accurate water–oil mixture purity estimation

This study presents a novel non-contact microwave sensing system based on a microstrip antenna array, developed for precise estimation of water content in oil–water mixtures. The system employs two identical narrowband bandpass antennas operating at a resonant frequency of 2.925 GHz, carefully positioned to maintain measurement consistency. A series of oil samples with varying water concentrations (0 %–100 %) were analyzed, and their S₁₁ and S₂₁ parameters were recorded. From these, key features including resonance frequency shift (ΔF), return loss (RL), and insertion loss (IL) were extracted and directly fed into a lightweight convolutional neural network (CNN) without additional feature engineering. The model achieved a root mean square error (RMSE) of 1.80, demonstrating strong predictive accuracy. Furthermore, the sensor exhibited a high sensitivity of approximately 13.4 MHz per unit change in relative permittivity (εr), enabling it to detect subtle dielectric differences across the mixture compositions. The proposed method offers a compact, efficient, and scalable solution for real-time, non-invasive purity monitoring, with potential applications in petrochemical processing, oil quality control, and environmental monitoring.