Cross-Sensor Masked Self-Supervised Transformers for Domain-Adaptive Human Activity Recognition
Year: 2026
Type: conference-paper
Abstract: Human activity recognition (HAR) from smartphone inertial signals often degrades when a model trained on one dataset is deployed on another due to differences in device placement, orientation, user behavior, and data collection protocols. This paper proposes a cross-sensor masked self-supervised Transformer framework for unsupervised domain-adaptive HAR. During pretraining, either accelerometer or gyroscope groups are masked and reconstructed from the remaining modality, encouraging the encoder to learn transferable inter-sensor motion dependencies. The pretrained Transformer is then adapted from labeled UCI-HAR source data to unlabeled HAPT target data using domain-specific batch normalization and adversarial domain alignment. The framework combines masked reconstruction and NT-Xent contrastive learning, while CORAL alignment and confidence-based pseudo-labeling are analyzed as optional diagnostic components. Experiments under the UCI-HAR → HAPT protocol show that the full configuration achieves 57.54% target accuracy and 55.19% target macro-F1 while maintaining strong source-domain performance. Ablation results confirm that self-supervised pretraining and domain-specific normalization are critical, as removing them reduces target macro-F1 to 9.34% and 9.86%, respectively. The findings show that cross-sensor dependency learning with lightweight domain adaptation is effective for robust inertial HAR under unlabeled target-domain conditions.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7214682318
