Comparative Analysis of Recurrent Neural Network Architectures and Hyperparameters for Human Activity Recognition Using Wearable Sensors"

Human activity recognition (HAR) is a significant area of research with numerous applications in healthcare, athletics performance monitoring, and elderly care. The potential for HAR with ubiquitous sensors to enhance human performance and quality of life has attracted significant research interest. Recurrent Neural Networks (RNNs) have emerged as a powerful tool for HAR, as they can model sequential data and capture temporal dependencies in time-series data. Using accelerometer data from wearable sensors, this study investigates the efficacy of various recurrent neural network (RNN) architectures and hyperparameters for HAR. Specifically, we compare the performance of three RNN architectures (Simple RNN, LSTM, and GRU) and investigate the impact of hidden units and sequence length on the accuracy of the models. We use the publicly available HARUS dataset, which consists of accelerometer data collected from 30 subjects performing six different activities. Our results show that the LSTM architecture outperforms the other two architectures, achieving an accuracy of 95.0% on the HARUS dataset. We also discover that increasing the number of hidden units generally improves accuracy, with 128 hidden units producing the greatest results. Increasing the sequence length also leads to higher accuracy, but increasing it beyond a certain point can lead to overfitting. In addition, a separate study found that their RNN model obtained an overall accuracy of 99.54 percent on the test set for recognizing various activities using accelerometer data from a wearable sensor. The model performed particularly well for walking and jogging activities, as well as standing and sitting activities, and performed reasonably well for more complex activities such as walking upstairs and downstairs. Our findings indicate that LSTM is a suitable architecture for HAR tasks and that the number of concealed units and sequence length are crucial hyperparameters to consider. Our findings contribute to the existing literature on HAR by revealing the optimal architecture and hyperparameters for the accurate recognition of human activities from accelerometer data collected by wearable sensors.