Deep CNN hyperparameter optimization algorithms for sensor-based human activity recognition

Human activity recognition (HAR) is an active field of research for the classification of human movements and applications in a wide variety of areas such as medical diagnosis, health care systems, elderly care, rehabilitation, surveillance in a smart home, and so on. HAR data are collected from wearable devices which include different types of sensors and/or with the smartphone sensor's aid. In recent years, deep learning algorithms have been showed a significant robustness for classifying human activities on HAR data. In the architecture of such deep learning networks, there are several hyperparameters to control the model efficiency which are mainly set by experiment. In this paper, firstly, we introduced one dimensional Convolutional neural network (CNN) as a model among supervised deep learning for an online HAR data classification. In order to automatically choose the optimum hyperparameters of the CNN model, seven approaches based on metaheuristic algorithms were investigated. The optimization algorithms were evaluated on the HAR dataset from the UCI Machine Learning repository. Furthermore, the performance of the proposed method was compared with several state-of-the-art evolutionary algorithms and other deep learning models. The experimental results showed the robustness of using metaheuristic algorithms to optimize the hyperparameters in CNN.