Automatic physical activity recognition using multichannel, fusion CNN-BiGRU-Bahdanauattention networks

Precisely recognizing and classifying physical activity in the everyday routines of patients with chronic illnesses can facilitate the implementation of precision medicine in the treatment of conditions like diabetes. Human activity recognition (HAR) is crucial in ubiquitous computing, specifically in the management of chronic diseases such as diabetes. Deep learning architecture have been increasingly popular for sensor-related HAR in recent years, demonstrating impressive performance. Nevertheless, they encounter obstacles when extracting and characterizing features, as well as segmenting continuous actions, particularly when working with time series data. These issues are particularly relevant in the field of diabetes management, where precise tracking of physical activity is crucial for effective therapy and the control of blood glucose levels. This paper presents a multichannel fusion model which integrates a mutichannel convolutional neural network (CNN) and a bidirectional gated recurrent unit (Bi-GRU) with thebahdanauattention mechanism, terminated with extra trees classifier. This model is designed to leverage the strengths of CNN, BiGRU and integration of attention mechanism for comprehensive feature extraction and temporal relationship learning. The efficiency of different machine learning classifiers evaluated by cross-validation to determine the best effective method for the specific task. The performance of the proposed architecture was evaluated using the UCI-HAR dataset. The model achieved an accuracy of 99.52%, precision of 99.56%, recall of 99.55%, andF1 score of 99.55% when combined with the extra trees classifier in the proposed fusion architecture which is better compared to existing models in recognizing undeclared physical activity types.