Ensemble Deep Learning Network for Enhancing Performances of Sensor-Based Physical Activity Recognition Based on IMU Sensor Data

The application of wearable sensors to identify physical activities has drawn significant attention in the healthcare and fitness industries. This study presents a novel ensemble deep learning network to enhance the accuracy of physical movement recognition utilizing data from inertial measurement unit (IMU) sensors. The proposed Ens-CNN-LSTM method merges a convolutional neural network (CNN) with a long short-term memory (LSTM) network. This integration capitalizes on their strengths in analyzing sequential IMU data. The model further incorporates a random forest classifier to finalize predictions. Our proposed model was evaluated using the PAMAP2 benchmark dataset for human activity recognition. This dataset includes data from multiple individuals performing various activities recorded with IMU sensors. We assessed the ensemble model's performance against individual deep learning models through a 5-fold cross-validation method. Results reveal a significant enhancement in overall accuracy, achieving 99.63%, compared to 96.19% for CNN models and 97.35% for LSTM models. The Ens-CNN-LSTM model demonstrated remarkable performance with an accuracy of 99.61%, a recall of 99.63%, and an F1-score of 99.61%. It surpassed the individual models in all metrics. Furthermore, the ensemble model showed increased robustness to variations among individuals and noise in sensor data. The proposed approach improves the accuracy of physical activity recognition and extends its applicability to other domains and activities.