A machine learning algorithm for retrieving the geometrical characteristic parameters of soot fractal aggregates from polarized light signal

Abstract

Polarized light measurement technology is provided to study geometric characteristics of soot fractal aggregates. The inverse problem is solved by two mature applications of machine learning, i.e. stacked ensemble model and deep network. A new Fractal Aggregates Generative Adversarial Network (FAGAN) is proposed to generate synthetic data to solve the class imbalance problem. The results show that compared with the stacked ensemble model which is sensitive to noise, the performance of the deep network is still satisfactory even under 10 % Gaussian noise, and the maximum mean absolute percentage error is not more than 3.87 %, which means that the synthetic data generated by FAGAN is very similar to the real data. As a whole, the combination of deep network and FAGAN has important guiding significance for complex and time-consuming experiments and engineering applications where noise is difficult to control or continuous and regular data cannot be obtained.