INTERNATIONAL JOURNAL FOR INNOVATIVE RESEARCH IN MULTIDISCIPLINARY FIELD

Abstract:  This paper provides the comparison between two deep learning methods of automatic modulation recognition of difficult DVB-S2X satellite signals. One way is based on the utilization of a convolutional neural network trained over a broad scope of radio data. The other is based on a hybrid CNN-Transformer model, which was trained on a special simulated dataset created specifically to train on modulations of high throughput of DVB-S2X. We aim at establishing which one is more effective in identifying the advanced modulation schemes at very low signal-to-noise ratio.

 The traditional CNN makes use of one-dimensional model. It is able to take raw in phase (I) and quadrature (Q) samples. Thereafter, it derives local features of the signal. It was trained based on an MATLAB generated dataset which is based on DVB-S2X. It is targeted at the ability to test its potential in handling high order modulations in satellite systems under simulated channel conditions. This CNN-Transformer hybrid model comprises of the combination of CNN layers to obtain the local features and after that, Transformer blocks using multi-head attention to obtain long-range and global correlations. They analyze the overall system using the RadioML 2016 (RML 2016) set which is a benchmark work in this type of work

The standard CNN implementation using the special DVB-S2X dataset demonstrated good prospects of satellite systems. It had an average accuracy of approximately 88.28% at all the SNR levels and its performance did not deteriorate even in presence of weak signals.

Hybrid CNN-Transformer Results (RML 2016 Dataset):

The CNN-Transformer hybrid model achieved good results on the heterogeneous RML 2016 dataset. The model also gave an average performance of 55.28 percent on all 11 modulation types across the SNR range. It is seen that the CNN and the Transformer elements are effectively coupled, particularly when classifying low or medium SNR signals.