請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/83160
標題: | 利用自走車輔助學習以改善行動網路室內定位效能 Improving Indoor Localization for Cellular Networks with Robot-Assisted Learning |
其他標題: | Improving Indoor Localization for Cellular Networks with Robot-Assisted Learning |
作者: | 洪健豪 Chien-Hao Hung |
指導教授: | 謝宏昀 Hung-Yun Hsieh |
關鍵字: | 室內定位,行動網路,自走車,通道狀態資訊, Indoor localization,LTE,Channel state information,CSI,Robot, |
出版年 : | 2022 |
學位: | 碩士 |
摘要: | 對室內定位的需求急劇增加,許多應用都需要高精度的室內定位技術,如智能工廠、智能配送、智能旅遊等。和基於測距的室內定位技術相比,基於學習的室內定位技術更能適應環境限制,基於學習的室內定位定位技術對設備的要較低。然而,基於學習的室內定位通常需要耗費人力和耗費時間來建立指紋數據集。此外,隨著時間、季節和溫度的變化,需要對模型進行實時調整。在這兩種情況下,都需要重建指紋數據集來訓練模型,這增加建立系統的成本。我們應用該機器人構建了一個可以邊走邊採集LTE無線電特徵的系統,並利用 SLAM演算法計算出的軌跡數據來輔助無線電特徵的標註。我們提出了MICNN+RNN串接模型,串接模型的性能可以達到0.879 m,實現了亞米級室內定位。對於 MICNN 模型,我們提出應用基於參數的遷移學習方法將從源域系統學到的知識遷移到目標域,該方法可以將 MICNN 的模型性能提高6.7%平均距離誤差 (MDE)。分析不同縮放器對模型性能的影響,我們發現 MinMax 縮放器有助於目標模型性能和微調模型性能。 The demand for indoor positioning has increased dramatically. Many applications require high-precision indoor localization technology, such as smart factories, smart deliveries, smart tours, etc. Compared with ranged-based indoor localization technologies, learning-based indoor localization technologies are more adaptable to environmental constraints, and learning-based indoor positioning technologies have lower requirements for instruments. However, learning-based indoor localization is often labor-intensive and time-consuming to build a fingerprint dataset. Also, with the change in time, season and temperature, the model needs to be adjusted in real-time. In both cases, the fingerprint dataset needs to be rebuilt to train the model, which increases the cost of building the system. We applied the robot to build a system that can collect LTE radio features while walking, and use the trajectory data calculated by the SLAM algorithm to assist in radio feature annotation. We propose the MICNN+RNN cascaded model, and the performance of the cascaded model can reach 0.879 m, which achieves sub-meter indoor localization. For the MICNN model, we propose to apply the parameter-based transfer learning method to transfer the knowledge learned from the source domain system to the target domain, and this method can improve the model performance of MICNN by 6.7% for mean distance error (MDE). Analyzing the effect of different scalers on model performance, we found that the MinMax scaler is helpful for the target model performance and fine-tuned model performance. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/83160 |
DOI: | 10.6342/NTU202203944 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 電信工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-2309202220284100.pdf | 7.15 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。