Table 1 EPRN architecture and training parameters.
Component | Specification |
|---|---|
Input format | Time-series window: 64 frames × 51 features (17 joints × 3D coordinates) |
Wavelet features | Statistical descriptors (mean, std, energy, entropy) from approximation/detail coefficients |
Wavelet transform | Discrete Wavelet Transform (DWT), Symlet 4 (sym4), 3-level decomposition |
Recurrent branch 1 | LSTM, 128 units |
CNN layers | Conv2D (32 filters, 3 × 3) → ReLU → BatchNorm → MaxPool (2 × 2) Conv2D (64 filters, 3 × 3) → ReLU → BatchNorm → GlobalAvgPool |
Fusion layer | Attention-based mechanism to combine LSTM and GRU outputs |
Recurrent branch 2 | GRU, 64 units |
Loss function | Mean Squared Error (MSE) |
Output layer | Fully connected layer (linear activation) for motion sequence prediction |
Optimizer | Adam |
Learning rate | 0.001 (adaptive decay) |
Dropout rate | 0.5 |
Training time/epoch | ~ 22.5 s |
Training epochs | 200 |
Batch size | 32 |
Hardware | NVIDIA RTX 3090 (24 GB), Intel Core i9, 64 GB RAM |
Platform | MATLAB 2023a, Deep Learning Toolbox |
