Flame lift-off detector based on deep learning neural networks

Abstract

This study evaluates the application of deep learning techniques in real-time flame lift-off detection and lift-off length prediction from flame images. A multi-feed test facility equipped with an optical diagnostic system was used to investigate the flame stability of both liquid and solid fuels. A high-speed camera was used to capture flame images, and image processing techniques were employed to extract the data. In the case of flame detection from images captured by high-speed cameras, the temporal resolution of the images is very high, which is beneficial for detecting rapid changes in the flame dynamics. In the current study, the analysis of 10,000 images by images processing methods takes approximately 60 min. Therefore, traditional methods may not be suitable for real-time monitoring, as they may not be able to capture the detailed variations in the flame appearance and intensity. The objective of this study was to find a fast and computationally efficient way to detect flame lift-off in high-speed online flame monitoring system. Two models were used in this study: a convolutional autoencoder and a regression neural network. The results demonstrate the potential of deep learning techniques in online flame diagnostics. The proposed approach provides a fast and efficient way to process 10,000 images in only 3.5 s, making it suitable for implementation in real-time flame monitoring systems.Novelty and significancePrevious studies in the field of real-time flame diagnostics have predominantly used low sampling rates to capture flame images. However, when high-speed cameras are used, conventional image processing techniques prove inadequate to provide instant results for real-time monitoring systems. This study presents a novel deep learning approach capable of directly detecting flame lift-off from high-speed images without the need for separate segmentation, thereby significantly enhancing the analysis speed. The results obtained in this paper emphasize the advantage of using deep learning techniques to extract flame characteristics at high speed and with excellent quality. These capabilities are particularly crucial in high-speed systems.