Human Activity Recognition Using Multiple Learning & XAI Techniques with Wearable Sensor Data

Abstract

Human Activity Recognition (HAR) is an important area of research in artificial intelligence, machine learning, and ubiquitous computing. It involves identifying or predicting human actions based on sensor data. This paper investigates Human Activity Recognition (HAR), a crucial area in artificial intelligence and multiple learning. It focuses on identifying human actions using sensor data. We analyze various machine learning techniques, including Random Forest, Gradient Boosting, AdaBoost, K-Nearest Neighbors (KNN), and the Voting Classifier, as well as deep learning models like CNN, ANN, CNN ANN Hybrid, and LSTM with our collected 72,095 collected sensor data. We also employ XAI (SHAP) techniques to understand feature importance. Results indicate that, Random Forest leads with an 85% accuracy rate. Among deep learning models, ANN achieves the highest accuracy at 82%. LGB, CatBoost, and XGBoost perform well, each reaching an 84% accuracy rate. In Federated Learning, 72% accuracy is achieved by global model with ANN. We propose an app for activity detection and data collection. These findings emphasize the potential of machine learning in enhancing HAR systems, with implications for applications from healthcare to wearable technology.