Context-Aware Self-Supervised Learning Framework for Persistent MultiGait Activity Intelligence in Wearable Systems

Abstract

Human activity recognition using wearable sensor data has gained significant attention due to its wide applications in healthcare, fitness monitoring, and smart environments. With the rapid growth of sensorenabled devices, large volumes of continuous motion data are generated, creating the need for efficient data analysis techniques. Traditionally, activity recognition relied on manual observation or rule-based systems that used predefined thresholds to classify activities. However, these approaches were limited in handling complex patterns, noisy sensor signals, and large-scale datasets, resulting in low accuracy and poor generalization. The primary challenge lies in accurately identifying activities from highdimensional and continuously generated sensor data while ensuring scalability and reliability. Traditional methods lack adaptability and fail to perform well under varying conditions such as changes in user behaviour and sensor placement. This creates a strong need for intelligent, automated systems capable of learning patterns directly from data. To address these challenges, this study presents a machine learning-based activity classification system that utilizes multiple models including Greedy Tree (GT), K-Nearest Neighbors (KNN), Logistic Regression (LR), Naive Bayes (NB), and Adaptive Boosting (AB). The system integrates data preprocessing, feature scaling, model training, evaluation, and prediction within a unified framework. Experimental results demonstrate that the Greedy Tree (GT) model achieves the highest accuracy of 99.00% for the target column “activity,” outperforming KNN (77.85%), NB (57.40%), LR (51.10%), and AB (51.02%). The proposed approach significantly improves classification performance, robustness, and scalability compared to traditional systems. It enables efficient handling of continuous sensor data and supports both real-time and batch prediction.