Enhancing Performances of Deep Neural Networks with Ensemble Learning Methods for Complex Human Activity Recognition Using Wearable Sensors
Sakorn Mekruksavanich, Anuchit Jitpattanakul · 2025
Sensor-based human activity recognition (HAR) is a widely explored research area focused on identifying human behaviors by analyzing sensor-generated data, particularly one-dimensional time series signals. Such HAR systems have many real-world applications, including elderly health monitoring and athlete performance evaluation. This paper presents Ens-DeepNet, a novel ensemble learning approach designed for recognizing complex human activities using data from wearable sensors. While deep neural networks have proven effective for HAR tasks, their performance can be further improved through ensemble strategies. To this end, we propose an ensemble architecture that combines five distinct deep learning models—convolutional neural network (CNN), long short-term memory (LSTM), bidirectional LSTM (BiLSTM), gated recurrent unit (GRU), and bidirectional GRU (BiGRU). These models are integrated using a majority voting scheme to boost classification accuracy. The proposed method is evaluated using the UTwente dataset, which includes sensor data collected from ten individuals via tri-axial accelerometers and gyroscopes. Experimental findings demonstrate that the ensemble model achieves an impressive accuracy of 99.28%, surpassing the performance of each standalone model, whose accuracies range from 92.77%to 97.34%. This performance gain is particularly notable in recognizing complex activities such as eating, drinking coffee, smoking, and conversing.