Visible Image-Based Regression: A Novel Approach for Continuous and Real-Time Diagnosis of Plasma Molecular Temperatures

This research explores the feasibility of extracting plasma molecular temperatures from visible images, an approach aimed at circumventing expensive spectrometer equipment and intricate, time-consuming processes involved in spectral acquisition and diagnostics. In our experiment, the emission spectrum data of needle-plate air discharge plasma under various working conditions are collected, and the optical emission spectroscopy (OES) analysis method is employed to diagnose the rotational and vibrational temperatures of nitrogen molecules. Simultaneously, visible images of plasma discharge are collected to establish a comprehensive dataset encompassing a broad temperature range. The extracted chromaticity information or the features automatically extracted by the convolutional neural network (CNN) from plasma visible images is utilized to conduct regression analysis on the plasma rotational and vibrational temperatures. This machine learning (ML) approach enables continuous prediction and diagnosis of plasma molecular temperatures based on visible images. The experiment results demonstrate that the best-performing CNN regression model achieves high prediction ${R}^{{2}}$ of 0.9986 and 0.9948 for the two temperatures, and the prediction time of a single image is only 36.31 ms, which meets the demand for high-precision and real-time plasma diagnosis.