A Hybrid Deep Neural Network with Attention Mechanism for Human Activity Recognition Based on Smartphone Sensors

Human activity recognition (HAR) is an area of study that seeks to automatically and precisely detect an individual’s behavior by analyzing bio-signal data. Bio-signal data can be acquired using sensing technology integrated into smartphones and other worn intelligent gadgets. Nevertheless, collecting sensor data tagged with activity information employing individual smartphones can result in data being analyzed in varying contexts, potentially compromising the accuracy of machine learning prediction methodologies. This paper introduces a novel HAR approach via smartphone sensors. The proposed method incorporates a hybrid deep neural network architecture enhanced with an attention mechanism. The deep learning model under consideration is referred to as the Att-CNN-BLSTM network. This particular network can autonomously extract significant features from smartphone sensor data. This feature extraction aims to effectively classify different human actions with a high degree of accuracy. In order to assess the efficacy of the hybrid model, a series of investigations were undertaken utilizing a publically accessible HAR dataset. This dataset was employed for training and testing purposes, employing a 5-fold cross-validation approach. We additionally carried out an analytical comparison utilizing state-of-the-art deep learning models from prior research. The findings from our experimentation demonstrate superior performance of the Att-CNN-BLSTM model over other advanced AI algorithms, with the highest accuracy rate of 90.54% and top F1-score of 86.88% attained.