Human Activity Recognition (HAR) is crucial in healthcare monitoring and smart home systems, tracking patient movements, de- tecting falls, and monitoring daily activities. Despite its importance, HAR faces significant challenges due to the scarcity of large-scale, diverse datasets and the lack of data representing abnormal activities, essential for detecting rare but critical health events. This paper addresses these challenges through advanced synthetic data generation and state-of-the- art classification techniques. We introduce a Generative Adversarial Net- work (GAN) designed for time series data to generate synthetic samples, significantly expanding the WISDM dataset. Our method includes gener- ating an ’abnormal’ activity class, enhancing the dataset’s diversity and real-world applicability. The synthetic data quality is rigorously evalu- ated using Dynamic Time Warping (DTW) to ensure fidelity to original data distributions. For classification, we leverage transformer-based mod- els to interpret and classify human activities from accelerometer data. This approach demonstrates the adaptability of advanced language mod- els to numeric, time-series data, opening new avenues in HAR research. Transformers, known for their success in natural language processing, show promise in capturing complex patterns in human activity data, leading to improved accuracy and robustness. Our methodology aims to expand the dataset, particularly for rare and abnormal activities, and enhance classification accuracy. This work contributes to HAR by pro- viding a framework for dataset enhancement and classification, paving the way for more robust and versatile activity recognition systems, espe- cially where data collection is challenging. The implications for health- care are significant, potentially leading to better patient outcomes and more efficient healthcare delivery.
