A High Efficiency Li-ion Battery Charger with Auto-switching Modulation Modes and Cost-effective Dataset Analysis for Underwater Object Recognition

• Main Authors: Guan-Xian Liu, 劉冠賢
• Other Authors: Chua-Chin Wang
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/z77z65

碩士 === 國立中山大學 === 電機工程學系研究所 === 107 === With the increasing demand of marine technology, many researchers have focused on the research of unmanned underwater vehicles. To improve the efficiency of unmanned underwater vehicles, the management of battery-powered system and the ability to analyze information are extremely important. Therefore, this thesis mainly explores the design of a key component for unmanned underwater vehicles Li-ion battery system and dataset establishment for underwater object detection. The first part of the thesis demonstrates a high-efficiency Li-ion battery charger with auto-switching modulation modes. By switching constant current and constant voltage (CC-CV) charging modes, the battery charging time is reduced and the battery life will be extended. The constant current charging mode is deemed as a heavy load, where pulse width modulation (PWM) is applied to maintain high efficiency. By contrast, the constant voltage charging mode is considered as light load to enable pulse frequency modulation (PFM) and reduce the switching loss. The pulse swallow method and an OR logic gate are used to carry out automatic switching of these 2 modes. Moreover, we improve the efficiency by shutting down those unused the operational amplifiers. The proposed design realized using TSMC 0.50 μm CMOS High Voltage process is featured with the input range of 8-10 V, and the output range of 2.5-4.2 V. The maximum charging current is 1.5 A, and the peak efficiency is 87.2% by simulation results. The second part of the thesis explores the feasibility of underwater object detection dataset based on deep learning. Taking clownfish and diver as examples, the models are established and the background of the picture is changed to analyze the effect of self-built dataset imposed on the recognition rate. YOLOv3 is used to realize the proposed method and verify the performance.