Proposal of a Method for Inferring the Road Surface from Vibration Using Deep Learning on CanSat

This paper describes a method for accurately classifying whether a CanSat is running on a hard surface or a soft, uneven surface based on acceleration data from vibration during running, using FFT and Deep Learning Image Classification. The CanSat cannot run smoothly on a hard surface, as in the case of ARLISS, and on a soft and uneven surface, as in the case of the Noshiro space event, unless the running speed and steering wheel angle settings are changed. Therefore, we considered the possibility of inferring whether the road surface is hard or soft from the vibration during running. This paper describes a method for converting acceleration data from running data into frequency-domain images by FFT analysis, and for classifying these images with high accuracy by using deep learning image classification. We found that this method can achieve a high classification accuracy of 100% for running data used for training, and more than 92% for other running data that was not used for training.