DeM-FCN: an ultra-lightweight and purely convolutional framework for edge-native human activity recognition in...

Wearable human activity recognition has become an important component of intelligent fitness tracking, but deploying accurate recognition models on resource-constrained edge devices remains challenging. Existing deep learning methods often rely on recurrent structures, attention mechanisms, or complex hybrid architectures, which increase computational cost and limit real-time deployment.

This study proposes DeM-FCN, a lightweight and purely convolutional framework for smart dumbbell-based resistance-training activity recognition. The model integrates a physics-aware input representation, Gaussian noise regularization, stacked one-dimensional convolutional blocks, Global Max Pooling, and a cost-sensitive focal loss to improve subject-independent recognition. The input representation extends raw inertial measurements by introducing trigonometric encoding of Euler angles and acceleration and gyroscope magnitude features, allowing the model to capture both orientation-related motion patterns and orientation-insensitive motion intensity. The proposed model was evaluated using Leave-One-Subject-Out cross-validation on a custom smart dumbbell dataset containing four resistance-training exercises collected from 15 subjects.

DeM-FCN achieved an accuracy of 0.966, macro F1-score of 0.916, and macro AUC of 0.982, while maintaining only 73.7 K parameters, 14.84 M FLOPs, and a model size of 0.29 MB. Additional evaluations on PAMAP2 and MHEALTH suggested that the convolutional backbone retained useful class-ranking ability on public IMU-based HAR datasets, while the reduced macro F1-scores indicated that hard-label daily activity recognition remains more challenging than constrained resistance-training recognition due to broader activity diversity, sensor-domain differences, and missing modality information. A refined ablation study confirmed that trigonometric encoding and magnitude features provide complementary benefits, with magnitude features contributing more strongly to cross-subject robustness.

The results suggest that DeM-FCN provides a favorable accuracy-efficiency trade-off for wearable resistance-training recognition and offers a practical foundation for edge-oriented fitness monitoring.