Research Information System
Engineering Applications of Artificial Intelligence, vol.153, 2025 (SCI-Expanded)
The rapid growth and development of sensor technology in wearable devices has encouraged researchers to develop new Human Activity Recognition (HAR) models. The data collected from various devices is affected by different factors, such as users, environments, and sensor frequencies. These factors cause heterogeneity in the collected data. Traditional machine learning models and single deep learning approaches struggle to effectively manage and handle data heterogeneity. To address this challenge, a novel deep neural network based on combining Long Term Short Memory (LSTM) layers with convolutional layers as a hybrid approach is proposed in this paper. This model streamlines data preparation through efficient segmentation and standardization, avoiding time-consuming techniques and methods used in our previous study. Our compact architecture minimizes computational overhead while enhancing interpretability, making it suitable for resource-constrained environments. The LSTM layer of the model is used for temporal feature extraction, whereas the convolution layer is used for spatial feature extraction from the data. The parameters of the model were reduced by replacing a fully connected layer with a batch normalization (BN) layer, which offers a significant advantage over other methods in the field. The performance of the model was evaluated on four heterogeneous HAR datasets. The overall F1-score accuracy model of the datasets: phone accelerometer, watch accelerometer, phone gyroscope, and watch gyroscope reached 97%, 89%, 89%, and 87%, respectively. The results prove that the proposed hybrid Long Short Term Memory and Convolutional Neural Network (LSTM-CNN)the model has superior performance compared to other reported results. By effectively addressing the challenges of heterogeneous data and obtaining superior scores, this model not only demonstrates robust performance but also contributes valuable insights for future research in human activity recognition.