Explainable CNN-GRU Model With Self-Attention for Human Activity Recognition Using Wearable Sensors Through Data Augmentation
Nadia Sultana, Sidratul Afrida, Bushra Akter, Tasnia Jahan, Afrin Ahmed, Mohammad Abu Yousuf, Md Zia Uddin · IEEE Access · 2026
Human activity recognition (HAR) has become an essential field of study, with applications in smart homes, sports training, human-computer interaction (HCI), healthcare, rehabilitation, and pervasive computing. Time-series data from wearable sensors, mobile devices, and Internet of Things (IoT) systems have been used to significantly increase the accuracy of HAR, owing to recent developments in deep learning. However, traditional models often struggle to balance accuracy and generalizability when applied to datasets with varying feature and timestep sizes, limiting their applicability across diverse real-world scenarios. In this study, we introduce a deep CNN-GRU-Self Attention model for HAR and employ SHapley Additive exPlanations for explainable AI (XAI) to identify influential features in HAR states. The CNN enhances feature extraction across various scales, whereas the GRU captures temporal connections to improve feature representation. The attention mechanism strategically focuses on the most relevant input features, enabling the extraction of time-dependent signals. This combination produces fewer parameters, leading to high-precision results with minimal processing time. Integrating XAI improves transparency, offering clear insights into feature contributions to DL-driven HAR models, which is crucial for validation and interpretability. Using the proposed sliding-window preprocessing strategy the model achieved accuracies of 99.70% and 99.14% on the UCI-HAR and WISDM datasets, respectively, and was further validated on PAMAP2 and WISDM-19 with accuracies of 96.80% and 96.04%, demonstrating its robust performance and strong generalization capability across a variety of activity recognition datasets. Its ability to capture spatial and temporal information simultaneously ensures resilience to variations in activity execution, making the proposed framework suitable for clinical, rehabilitation, and surveillance applications.