Compact Dispersive Near-Infrared Sensors for Jet Fuel Property Prediction

Fuels, specifically conventional and sustainable aviation fuels, exhibit significant variations in their properties. As blends of these fuels are increasingly integrated into industry applications, the demand for engines capable of operating on a wider range of fuels has grown. Incorporating an adaptive combustion control system with on-board fuel property detection can enable multi-fuel capable compression ignition engines. These engines can adjust and compensate for fuel property variations identified by an on-board fuel property sensor. One key fuel property for compression ignition engines is the derived cetane number (DCN), which indicates a fuel’s ignition propensity. However, conventional DCN measurement techniques, such as Ignition Quality Testers, are not suitable for on-board applications as they are intrusive, time-consuming, and quite large. Dispersive near-infrared (NIR) absorption spectroscopy offers a low-cost, robust alternative for fuel property sensing, meeting the size, weight, and power requirements necessary for on-board applications. This approach enables real-time fuel characterization, which is crucial for adaptive engine combustion control. To enable multi-fuel capable engines through fuel sensing and combustion control, three prototype dispersive NIR sensors are developed and tested. The sensors are designed to perform intensity reference calibrations through two distinct methods. One approach employs fluidic components, while the other utilizes multiple optical paths to facilitate reference calibrations. Additionally, each sensor is equipped with an inline wavelength calibration filter and a custom algorithm that provides real-time pixel-to-wavelength calibrations. This custom algorithm autonomously performs polynomial curve fits. The use of a non-Euclidean distance metric-Wasserstein distance-is investigated to further enhance the accuracy of wavelength calibrations. Results indicate the effectiveness of the polynomial curve fitting method in establishing precise pixel-to-wavelength relationships. Further, custom machine learning models are developed and deployed on the sensors to enable the prediction of DCN from NIR spectral data. These models are designed to address common challenges associated with on-board applications, such as variations in spectral range, resolution shifts, and baseline drifts. Two modeling approaches—imputation-based and ensemble machine learning—are formulated and evaluated, with the ensemble model showing strong performance. It achieved a coefficient of determination (R² score) above 0.86 across varying resolutions (12 nm, 10 nm, and 2 cm-1), and an average R² score of 0.857 when tested across different spectral ranges. Ultimately, the combination of reference calibrations, pixel-to-wavelength assignments, and the deployment of flexible machine learning models contribute to the development of a robust dispersive NIR sensor, suitable for accurate on-board fuel property determination.