Abstract: Aiming at the problems of insensitive local subtle features and large memory capacity required in the recognition of communication emitters in Transformer network, we propose a method for extracting and identifying fine features of quadrature modulator of communication radiation sources based on local bispectrum and Longformer networks. Firstly, we analyzed the mismatch characteristics of orthogonal modulators and established a mathematical model for mirror frequency interference in the output communication signal of quadrature modulators; Secondly, we derived the local discrete bispectral representation of the communication radiation source signal after orthogonal decomposition and symbol synchronization segmentation. Then, we analyzed the sparse self attention mechanism combining position encoding local self attention and global self attention in Longformer networks, and proposed an orthogonal modulation fine feature recognition and classification framework based on Longformer networks and softmax classifiers. Finally, We simulated the classification and recognition process of fine features of four quadrature modulators of the same type of OFDM radiation source, and analyzed the recognition performance, noise resistance, and network complexity of this method.The simulation results show that Longformer network has better recognition performance than Transformer network and radial basis function network, achieving an accuracy of over 90%.