In response to the limitations observed in existing communication modulation recognition technologies, including a restricted number of recognition types and a suboptimal overall recognition rate, an automatic modulation recognition method for communication signals based on feature fusion (ACCLNN) was proposed. Initially, the method leveraged ASPP to capture diverse input information across different scales. Spatial features were then extracted through a parallel convolutional neural network (CNN), incorporating the CBAM attention module to enhance feature representation in critical data areas and extract low-level features. Subsequently, a long short-term memory network (LSTM) was employed for temporal feature extraction, accompanied by the introduction of a self-attention module for profound feature extraction. The final step involved fusing low-level features and deep features to achieve comprehensive feature extraction and mapping. Experimental results demonstrate the efficacy of the proposed method in identifying communication signal categories in low signal-to-noise ratio (SNR) environments, achieving an overall recognition rate exceeding 93%.