Table 1 Comparison of related work with the proposed model.
Reference | Dataset | Feature extraction | Model | Accuracy (%) | Weaknesses |
|---|---|---|---|---|---|
Zeng et al.13 | RADIOML2016.10A | Handcrafted (HOS) | CNN | 85.12 | Poor generalization at low SNRs |
Ali et al.19 | RADIOML2016.10A | PCA-based normalization | ANN + PCA | 87.45 | Computational overhead |
Beisun et al.21 | Custom SDR data | Adaptive kernels (spectrogram) | Inception-ResNet | 88.90 | Requires SDR hardware setup |
Zhou et al.24 | RADIOML2018.01A | Raw IQ to tensors | CNN with pre-training | 90.32 | Needs more samples to train |
Yin et al.25 | RADIOML2018.01A | Pre-trained weights | CNN + offline pretrain | 89.75 | Underperforms in noisy environments |
Proposed MFOP-ELM | RADIOML2016.10A and 2018.01A | Deep Features (InceptionV3, ResNet50, VGG16) | MFO-Optimized ELM | 94.19 | - |
