Artificial Intelligence-Enhanced Flexible Sensors for Human Motion and Posture Sensing

Addressing Data Scarcity & Generalization

A central challenge in flexible wearable sensor-based motion and posture recognition is the limited scale of available datasets. When dealing with small-scale wearable datasets, the optimal methodology depends critically on the data characteristics and the target application. If the dataset is well-defined and the deployment distribution is expected to remain consistent with the training distribution, the primary objective is effective model fitting. In such cases, traditional machine learning methods (e.g., Support Vector Machine, SVM [91], eXtreme Gradient Boosting, XGBoost [92]) are typically sufficient for simple data structures, while lightweight deep learning models are preferred for capturing more complex patterns [93]. When related datasets or pretrained models are available, transfer learning can further improve data efficiency by leveraging previously learned representations, often enhancing generalization under limited labeled data. However, “limited data” more commonly refers to scenarios where the training data is scarce, but the model must generalize to diverse and unseen test environments, leading to a distribution shift. In these situations, standard overfitting mitigation techniques like dropout or regularization may be inadequate. Under such conditions, data augmentation and synthetic data generation can serve as complementary strategies, introducing additional variability to better approximate the potential test distribution [94,95]. Furthermore, while self-supervised learning (SSL) has gained prominence for addressing label scarcity, it fundamentally requires a substantial volume of unlabeled data for pre-text tasks to learn robust representations before fine-tuning on small labeled subsets [96,97]. Consequently, in extreme scenarios where even raw data is scarce, carefully designed task-specific augmentation often remains practical baseline approach.