A CBAM-GRU classification model based on the combination of Convolutional Attention Mechanism Module (CBAM) and Gated Recurrent Unit (GRU) network is investigated for automatic modulation identification in non-cooperative communication systems. The pre-processed time-domain amplitude, phase and I/Q values of the signal are combined and converted into a matrix of input sample values, which are entered into the network for signal classification and identification. Simulations are conducted using the RadioML2016. 10a radio dataset, and the CBAM-GRU model are compared with the Convolutional Neural Network (CNN), Long Short-Term Memory network (LSTM), GRU, and Convolutional Long Deep Neural Network (CLDNN). The results indicates that the classification accuracy of the CBAM- GRU model reaches 92.79%, showing improvements of 8.52%, 1.84%, 1.75%, and 8. 61% over the comparison models respectively. Compared to traditional CNN or LSTM models, the CBAM-GRU model is more effective in capturing spatio-temporal features of sig- nals, thereby enhancing recognition accuracy.