A Study on Establishing a Dynamic Color Schlieren System to Observe Airflow and Predict Temperature Changes

Color schlieren is a technique capable of visualizing fluid waves (e.g., airflows and sound waves) that are beyond the visibility of the naked eye. This study aims to evaluate the airflow temperature in color schlieren through the distribution of different dynamic colors and verify the temperature prediction's feasibility. This approach is extensively helpful for analyzing high-temperature and non-invasive fluids. Compared to infrared thermal images, the advantage of color schlieren images lies in the capability of observing airflow changes in detail; at the same time, the temperature of airflow and measured objects can be obtained from the color distribution. In this Schlieren system, color filters made from low-cost, transparent projector films were adopted, which could change the color of the image captured from the system along with temperature changes, and provide intuitive perceptions. Since scratches from laser printing appeared on the low-cost color filter, the haze would be generated during imaging by the color schlieren system, which reduced the image saturation. To solve the problem of reduction in image saturation, the image dehazing technique was applied in this study to correct the shortcomings. Moreover, the dynamic information of color Schlieren images was also utilized to explore the relationship between Schlieren color distribution and actual temperature, where a temperature prediction model was established with a Feedforward neural network (FNN). Lastly, the Pearson correlation coefficient was applied to evaluate the degree of correlation between the FNN data sets; the Mean square error (MSE) value served as the evaluation index of error between the prediction result and verification data, where the Error histogram reflected the error values for the training, test, and verification datasets. The results indicated that the Pearson correlation coefficient of the correlation analysis was 0.99848, MSE at 0.6663, and the Error histogram reflected 0.4372 as the error between most data. Therefore, the temperature prediction model presented excellent predicting capability.