Table 1.
Comparison table for highlighted DCGAN architectures. The lower the DTW and FFT MSE metric values, the better is the generated signal. The lower the discriminator loss, the better it is at distinguishing fake from real samples. The lower the generator loss, the better it is at generating fake samples closer to real samples. Model number 6 (4CNN-MBD-MA) shows the best overall results.
| # | Model | Latent Space (z) |
Disc. Loss |
Gen. Loss |
DTW | FFT MSE | EMG Envelope Cross-Correlation |
|---|---|---|---|---|---|---|---|
| 1. | 3CNN-NOISE | Rand (100) | 0.000002 | 16.118095 | 405.038863 | 130.188808 | 0.373913 |
| 2. | 3CNN | Sample (400) | 0.005163 | 3.084711 | 132.185279 | 13.093062 | 0.205428 |
| 3. | WAVELET | Sample (400) | 0.066975 | 2.680009 | 100.145536 | 16.564078 | 0.223018 |
| 4. | 4CNN | Sample (400) | 0.035544 | 7.006461 | 93.439412 | 9.675622 | 0.624258 |
| 5. | 4CNN-MBD | Sample (400) | 0.000203 | 10.275796 | 100.786512 | 18.916364 | 0.739453 |
| 6. | 4CNN-MBD-MA | Sample (400) | 0.004439 | 10.636311 | 98.532786 | 13.531477 | 0.791920 |