Optimizing CGAN and DNN Architectures for Wearable Systems to Support Students with Developmental Disabilities

Abstract

In this study, we propose a 6-Degrees of Freedom Inertial Measurement Unit (6-DoF IMU)-based wearable system for recognizing challenging behaviors of students with developmental disabilities. In the proposed system, 6-DoF IMU data is preprocessed and used as input to the Deep Neural Network (DNN) to recognize challenging behaviors of students with developmental disabilities. Building a dataset is one of the biggest challenges in wearable Artificial Intelligence (AI) systems. Since collecting data samples is expensive, there is a limit to the amount that can be collected. In this study, we collect datasets from participants, build a custom dataset, and augment the data using Conditional Generative Adversarial Network (CGAN). We observe the performance change according to the augmentation ratio of the original data, and evaluate the scalability of the developed model by applying data from new participants that have never been shown during the training process. As a result of applying data augmentation techniques to a DNN model that already has high accuracy, a slight decrease in accuracy was observed for the original test set. However, when data from new participants is applied, an accuracy improvement of up to 10% was observed.