Abstract: There are problems such as arbitrariness, subjectivity, and discrepancy in the judgment of flame state of the municipal solid waste incineration (MSWI) process by domain experts. The high-quality flame images for building data-driven recognition model are scare. Aiming at these problems, a flame combustion state recognition method is proposed based on image enhancement realized by denoising diffusion probabilistic model (DDPM) and multiple types feature fusion. First, DDPM was used to generate virtual flame images to make up for the scarcity of high-quality images. Second, the LeNet-5 model was used to extract the depth features of the modeling data obtained by mixing the real images and the virtual images, and the physical features such as brightness, region and color of the flame image were extracted. Finally, a recognition model based on deep forest classification (DFC) by using the mixed feature was constructed. Based on the actual flame images, the effectiveness and superiority of the proposed method were verified.