DNEN-ACVIT: A Deep Neural Ensemble Network for Automatic Modulation Classification Based on Attention Convolution and Vision Transformer

Automatic modulation classification is a key technology in cognitive radio, aimed at identifying radio signals in non-communication environments. Recent advances in deep learning have significantly enhanced the accuracy and robustness of radio modulation classification systems by extracting features from raw signals, reducing reliance on manual feature engineering. This paper presents a novel modulation classification model that introduces an attention-based convolutional architecture, integrating Long Short-Term Memory (LSTM) networks and Vision Transformers (ViTs) to facilitate a learning process from local to global features. The model first employs attention convolution to extract local signal features, then uses LSTMs to capture sequential patterns among the local features, and finally leverages ViTs to capture the global semantic representation of the entire signal. After obtaining the global signal representation, a multilayer perceptron maps it to a lower-dimensional vector space for modulation type classification. Experimental results show that the method achieves accuracies of 61.6 % and 64.12% on the RML2016.10a and RML2016.10b datasets, respectively. Compared to current state-of-the-art methods, the proposed approach improves accuracy by 1.4% to 7.8 % on RML2016.10a and 0.8 % to 5.5 % on RML2016.10b. This study demonstrates the effectiveness of advanced attention-based deep learning architectures, particularly convolutional attention and self-attention, in modulation recognition networks. The proposed model exhibits superior accuracy and scalability under various signal-to-noise ratios and modulation schemes, offering valuable support for cognitive radio and other wireless communication systems in practical applications.