TSCResNet: Residual CNN for Activity Recognition
TSCResNet is a one-dimensional residual convolutional network for time-series classification, developed during my doctoral research and applied to human activity recognition on the public PAMAP2 wearable-sensor benchmark.
How it works
- Input. 5.12-second windows of 40 sensor channels (three inertial measurement units plus heart rate) × 172 time steps.
- Architecture. Strided convolutions and bottleneck residual blocks (a width-1 convolution that halves the channels, a width-3 convolution that restores them, plus the skip connection), with batch normalization and leaky ReLU. The layer stack is defined declaratively in a config and built programmatically.
- No pooling or dense head. The final convolutions shrink the time axis to exactly as many positions as there are classes, and those positions serve directly as the class logits.
- Open set. For the open-set study the model is trained once per activity with that activity held out, and a softmax-confidence threshold rejects windows that look like none of the known activities.
Results
On the standard held-out-subject split (closed set), TSCResNet reached 95.6% accuracy, 95.5% weighted F1, and 95.5% mean F1, improving on the best previously published results it was compared against (93.7% accuracy). The open-set follow-up is published in IEEE CSCE 2023. See Research for both papers.
Engineering
Experiments are tracked in MLflow (hyperparameters, training loss, and per-epoch validation and test accuracy and F1), and the training environment is containerized with Docker.
PyTorch · scikit-learn · NumPy · pandas · MLflow · Docker