A Novel Framework for Enhancing Sensor Data Analysis: Label-Preserving Augmentation and Probabilistic Balancing

The availability of sensors has revolutionized various sectors, including healthcare, industrial automation, environmental monitoring, and more. However, despite their significance, most sensor datasets encounter two significant challenges: insufficient data and class imbalance. Data balancing and augmentation are crucial steps in improving the performance of machine learning models when dealing with such challenges. Data augmentation techniques have been widely successful in domains such as image and audio data. Unlike visual or auditory data, sensor data augmentation can inadvertently alter labels, making it difficult to maintain data integrity. In this study, we address these challenges by introducing a novel data-balancing framework based on probability and randomization for sensor data augmentation. Our approach focuses on label preservation techniques using Pearson correlation and Dynamic Time Warping (DTW) to ensure the label integrity of the augmented data. By applying this framework in conjunction with a lightweight Convolutional Neural Network (CNN) model, we observed a significant improvement in performance—specifically, a 4-5% increase in accuracy compared to existing works. The peak accuracy achieved using our methodology reached an impressive 97.5% on the WISDM dataset, highlighting the effectiveness of our approach.