We demonstrate a novel application of supervised machine learning (ML) models to quantify the size, shape, number density, and distribution parameters of a water spray introduced at a gas turbine inlet. Only a limited set of laser scattering and extinction observations, acquired by pairs of photodetectors and cameras, are required for an accurate output. A phase Doppler particle analyzer as well as a conventional extinction inversion method are used to validate the particle size estimation, with the ML method converging closely to both. By measuring a water spray, where a spherical particle shape can be assumed, these size estimate validations could be made, which would have been difficult for a nonspherical particle measurement. By combining all the estimated parameters, the liquid volume fraction as well as the liquid flow rate is estimated and compared to a traceable ultrasonic flowmeter. To our knowledge, this is the first in situ condensation load measurement made at a gas turbine inlet without prior calibration. The ML approach is able to accurately estimate the liquid flow rate, with the majority of the estimates lying within the uncertainty bounds of the flowmeter and a root-mean-square difference of 0.8 L h−1 or 7.4%. Estimating the liquid flow rate using all the particle parameters demonstrates the method’s robustness and readiness for accurately measuring even nonspherical particles. The low number of required optical observations also makes this technique attractive for more generalized inlet particle measurements including sand, dust, and volcanic ash, in addition to condensation.
The inverse scattering problem of non-spherical particle size estimation is solved using a series of supervised machine learning models trained on a library of light scattering data. By establishing a large library with spheres and spheroids as fundamental shapes and through optimization of model hyperparameters, the trained models are able to accurately estimate a precise equivalent volume sphere radius of particles from an external database and simulations, with root mean square errors of 2.6% and 1.9% for the external and simulated particles, respectively. It was found that classification via a
k
-nearest neighbor model and refinement via a trained ensemble regression model performed best for equivalent volume measurements.
We present a novel optical particle sensor technique using artificial neural networks. This method relies on observations of light scattering and extinction by particles as input features to a trained neural network, which provides relevant particle distribution and representative shape for an integrated particle mass flow estimation. The models are trained on artificial data, generated for particles that the sensor is likely to encounter. The feasibility of our method is demonstrated through an experimental measurement of solid sand particles injected into a high-speed wind tunnel. The results show accurate estimations of the injected sand mass flow and particle size statistics, with a sand mass flow root-mean-square error of 0.28 g/min or 4.1% from the monitored rate using a precision scale. This measurement framework paves the way for sensor applications in harsh operating environments with limited
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.