[4Rin1] Interactive Session 2

[4Rin1-44] Adaptive Compressed Sensing Using Deep Learning Suitable for Edge Computing

〇Masatoshi Sekine¹, Satoshi Ikada¹ (1. Oki Electric Industry Co., Ltd.)

Keywords:compressed sensing, edge computing, deep learning, supervised learning, reinforcement learning

In this paper, we propose a method to dynamically and directly estimate the optimal compression ratio based on sensor data in compressed sensing. It is able to be executed at edge devices for efficient data transmission in wireless sensor networks. Generally, the optimization in data compression requires a large amount of calculation. Therefore, in our proposed method, the optimal compression ratio can be obtained with lightweight processing by using the pre-trained model of supervised learning or reinforcement learning in deep learning. As a result of the performance evaluation using acceleration data generated from the sensor placed on a bridge, our proposed method can reduce the estimation delay of the optimal compression ratio. In addition, both the compression ratio and reconstruction error can be reduced by changing compression ratio dynamically based on the sensor data in time domain.