For FSO communication atmospheric turbulence has a large impact on signal modulation, the convolution-profile stellar data image conversion algorithm proposed in this paper performs data conversion on the received constellation maps, so that they retain more original signal feature images. A classification network based on the channel attention mechanism is proposed to classify the modulated signals by extracting the feature information in the image through the residual structure, and the attention mechanism assigns different weights of the channel features. Under the same data conversion algorithm, the proposed classification network achieves the highest recognition accuracy of 96.286%.