Enhancing Sensor-based Human Activity Recognition Using Hybrid Deep Learning and Data Augmentation

Human Activity Recognition (HAR) is essential in various applications, including wellness tracking, automated residences, and fitness monitoring. In the past few decades, sensor-based HAR has become increasingly popular due to advancements in technologies for sensors. Nevertheless, HAR networks' effectiveness dramatically depends on the caliber and volume of the training data, which is frequently restricted and unevenly distributed. This research introduces a novel deep learning method called Multihead-CNN-BiGRU, which integrates one-dimensional convolutional neural networks with bidirectional gated recurrent units (BiGRU) to improve the accuracy of sensor-based HAR. To tackle the problem of insufficient a nd uneven training data, we utilize the synthetic minority over-sampling technique (SMOTE) to augment the data. The suggested model is assessed using the publicly accessible WISDM dataset, which comprises sensor data from diverse human actions. The ID-CNN is employed for extracting localized characteristics from the sensor data, while the BiGRU gathers temporal dependencies and contextual information. The hybrid architecture allows the model to acquire spatial and temporal patterns efficiently. In addition, the SMOTE technique is utilized to create artificial samples of the underrepresented classes, thus equalizing the distribution of classes and enhancing the model's capacity to generalize. The experimental findings show that our hybrid strategy provides exceptional outcomes when paired with SMOTE data augmentation. It obtains the most excellent accuracy of 99.51 % and the highest F1-score of 99.49% compared to the most advanced approaches. The suggested framework provides a reliable and precise solution for sensor-based HAR, which opens up opportunities for improved applications in other fields.