Name
Abstract | ||
|---|---|---|
As the use of wireless communication devices increases, so does the need for effective signal generation techniques. We show how different methods of waveform generation with the same signal-to-noise ratios result in varied accuracy performance for modulation recognition when using neural networks (NNs). Our research indicates that two aspects of waveform generation significantly change NN behavior with equivalent SNRs. First, generating purely random IQ constellation points adversely impacts classification accuracy, as real radios do not have completely random constellation points. We illustrate that NNs can use the intermediate IQ samples to improve classification accuracy. Second, we exhibit that adding pure additive white Gaussian noise (AWGN) results in different accuracies compared to using channel distortion models, which include fading, multipath, and phase shifts. We show that relying on SNR alone to characterize signals can result in misleading and incorrect performance evaluations of NNs when applied to radio performance. Our work also demonstrates that care must be taken when applying channels models to simulate noise, as NNs can use specific channel distortion features to identify a modulation. |
Year | DOI | Venue |
|---|---|---|
2021 | 10.1109/CSCN53733.2021.9686084 | 2021 IEEE Conference on Standards for Communications and Networking (CSCN) |
Keywords | DocType | ISBN |
Neural Networks,Deep Learning,CNN,LSTM,Fully Connected Neural Network Matlah,GNU Radio,Synthetic Data Generation,AWGN,Modulation Recognition,Wireless Radio Signals | Conference | 978-1-6654-2350-2 |
Citations | PageRank | References |
0 | 0.34 | 0 |
Authors | ||
7 | ||
Name | Order | Citations | PageRank |
|---|---|---|---|
| Tina L. Burns | 1 | 0 | 0.34 |
| Richard P. Martin | 2 | 0 | 0.34 |
| Jorge Ortiz | 3 | 0 | 0.34 |
| Ivan Seskar | 4 | 0 | 0.34 |
| Dragoslav Stojadinovic | 5 | 0 | 1.35 |
| Ryan Davis | 6 | 0 | 0.34 |
| Miguel Camelo | 7 | 0 | 0.68 |