A Dynamic Training-Dataset Distribution for AIOT System with Edge Computing

• Main Authors: TSAI CHENG-HAN, 蔡政翰
• Other Authors: YUH-SHYAN CHEN
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/v8tcg8

碩士 === 國立臺北大學 === 資訊工程學系 === 107 === In order to reduces training time and achieve internet of things (IoT) real-time response. The AIoT system presents an architecture and a dynamic distribution, reducing the training time by distributing training data in fog computing platform and cloud computing platform. The trained model deploys on edge computing platform. The AIoT system presents an architecture which processes the advantages of both edge processing and cloud computing for AI application on sensor data and image recognition. End device transfers sensor data and image to fog computing platform by LoRa, XBee, Wi-Fi. In this thesis, end devices configures to monitor environmental data. The proposed advanced dynamic distribution approach is an integrated strategy. Fog computing platform receives the partial training data and testing data, transmitd from cloud drive. After the advanced dynamic distributing, fog computing platform starts training model by Node-RED. Through the advanced dynamic distribution, cloud passes the partial data to fog computing platform then starts tensorflow training model simultaneously. The main advantage of the proposed strategy is aims to reduce training time. When fog computing platform and cloud computing platform completed model training, cloud transfers trained model to fog computing platform. Fog computing platform receives trained model from cloud computing platform. Node-RED combines trained model from each distributed equipment. Node-RED transfers the combined model to NVIDIA JETSON TX2 and provides TX2 to execute data inference. The AIoT system provides sensor data to monitor on thingsboard.