- 學位論文
行動通訊室內定位系統下考慮人類活動之強健無線電特徵選擇
Selecting Robust Features for Cellular Indoor Localization in Environments with Human Activities
摘要
室內定位已經提出了多年，但常見的室內定位都是在理想環境下去進行的。在 理想環境下能準確定位的信號通常在有人類活動干擾下的定位精度會有所下 降。為了解決這個問題，我們研究行動通訊網路基地台傳送的信號，期望找出 通用而穩健的特徵，能夠在理想和人類活動的室內環境中進行定位。 首先，我們對在環境內人類活動時有反應的信號收集下來作為我們的訓練特 徵，接著我們設計了一種算法來比較卡方 (CHI2)、信息增益（Information Gain）和可解釋 AI (Shap) 三個特徵選擇的結果，來選出我們收集的特徵的重 要性和關係。在比較了本文提出的三個特徵選擇的重要特徵之後，我們將最後 選擇的特徵作為輸入資料，通過機器學習方法進行訓練：卷積神經網絡 （CNN2D）和全連接神經網絡（FCNN）來定位使用者的位置。為了解釋 CNN2D 模型是如何訓練數據集，我們使用算法 Shap 來觀察特徵在 CNN2D 模型訓練中 具有的重要性和該特徵在模型的訓練起到了正面或負面幫助。這 67 個原始特徵 資料量通過我們設計的算法，最後選擇了 11 個特徵用於室內定位。選擇的 11 個特徵具有通用性，可以在理想與人類活動干擾的環境下定位。最終，訓練維 度總共下降了 84%，訓練時間複雜度也比 67 維減少了 6 倍左右。與預比較的 論文中使用的特徵定位結果相比，定位精度提高了 49%，而且我們只使用了 11 個特徵進行定位， 與我們比較的論文則使用了 23 個特徵。
並列摘要
Indoor localization has been proposed for many years, but common indoor localization is localized in an ideal environment. The signals that can be accurately located in the ideal environment have a decrease in the accuracy of localization under the interference of human activities. To address this problem, we study the signals transmitted by base stations in cellular networks, hoping to find general and robust features that enable localization in ideal and human-active indoor environments. We collect the signal that has responded to human activity to be our training features. We design an algorithm to compare three feature selection result which is Chi-Square (CHI2), Information Gain, and Shap to select the importance and relation features we had collected. After comparing the importance of the three feature selection selected features, we use the selected features and train them by machine learning methods: Convolutional Neural Network and Fully Connected Neural Network to locate the position of UE. To explain how the CNN2D model trains the dataset, we use the algorithm Shap to draw out the positive or negative data important to observing the feature importance in CNN2D training. The algorithm we propose selected 11 features for indoor localization from the original features amount of 67 and these features are generality and can be localized in environments ideally interfering with human activity. Finally, the training dimension will decrease by 84%, and the training time complexity will decrease six times than 67-dimension. The localization accuracy has improved by 49% compared to the localization baseline approach, and we are only using 11 features for localization where baseline approach uses 23 features.