SensorGAN: A Novel Data Recovery Approach for Wearable Human Activity Recognition

Human activity recognition (HAR) and more broadly, activities of daily life using wearable devices, have the potential to transform a number applications including mobile healthcare, smart homes, fitness monitoring. Recent approaches for HAR use multiple sensors on various locations body achieve higher accuracy complex activities. While increase accuracy, they are also susceptible reliability issues when one or unable provide data application due sensor malfunction, user error, energy limitations. Training classifiers that subset is not desirable since it may lead reduced applications. To handle these limitations, we propose novel generative approach recovers missing available from other sensors. The recovered then used seamlessly classify Experiments three publicly datasets show with sensor, proposed achieves within 10% no data. Moreover, implementation device prototype takes about 1.5 ms recovering in w-HAR dataset, which results an consumption 606 μ J. low ensures SensorGAN suitable effectively tinyML energy-constrained devices.