Table 3.
Detailed description of the convolutional neural network used in this study. The first convolutional together with the batch normalization layers are used once for the acceleration data and once for the gyroscope data.
| Layer Type | Hyperparameter | Output Shape | # of Parameters |
|---|---|---|---|
| 1D Convolution | Filter: 128, kernelsize: 8 | (64, 128) | 3200 |
| Batch Normalization | Momentum: 0.9, epsilon: 0.001 | (64, 128) | 256 |
| 1D Convolution | Filter: 256, kernelsize: 5 | (60, 128) | 164,096 |
| Batch Normalization | Momentum: 0.9, epsilon: 0.001 | (60, 128) | 240 |
| 1D Convolution | Filter: 128, kernelsize: 3 | (58, 128) | 98,432 |
| Batch Normalization | Momentum: 0.9, epsilon: 0.001 | (58, 128) | 232 |
| Global max-pooling | (128) | 0 | |
| Fully Connected | (64) | 16,448 | |
| Dropout | dropout: 0.2 | (64) | 0 |
| Dense | (5) | 325 |