Abstract: A quantitative evaluation method for flame image stability was proposed. First, convolutional autoencoder was used to extract features from flame images, and then quantitative evaluation index was employed for feature analysis. The quantitative evaluation index, with a numerical interval of 0, 1, was established based on cluster analysis and statistical analysis of the image feature. The convolutional autoencoder adopted a novel loss function based on reconstruction similarity to improve training efficiency. The effectiveness of the quantitative evaluation method for flame image stability was verified through experiments on the ethylene combustion platform. Results show that the convolutional autoencoder can extract image features in an unsupervised manner, and its performance is obviously superior to traditional feature learning methods. In addition, the established quantitative evaluation index can quantitatively characterize the flame image stability, showing strong generalization ability.