室內定位系統跨領域訓練之資料擴增方法研究

Abstract

在基於機器學習的室內定位系統中，當環境改變時，原本為特定環境建立的模型可能不再能有效的運作。因此，我們必須重新收集目標場域中大量的數據，這需要消耗許多時間。因此，這種方法是不實際的。在我們的研究中，我們的目標是設計一個數據增強系統，以減少蒐集資料的時間成本；我們使用軟件定義無線電（SDR）硬體和開源的平台（OAI 5G）建立定位系統。我們可以通過這個系統收集基於LTE的特徵，並將它們放入機器學習模型中以預測用戶的坐標。在此之後，我們使用Cycle GAN進行數據增強，這通常用於改變圖像的風格。我們提出了一種基於Cycle GAN的方法，稱為Semi Cycle GAN，用於使用在新場域中的少量資料和使用在舊場域中的大量資料以產生類似於新場域的充足資料量。這種方法可以省下大量時間，並提高模型的精準度。它比 Cycle GAN 更適用於低維度資料。與Vanilla GAN相比，我們可以提高定位精準度，並從目標場域和源頭場域中使用數據。我們的研究可以通過上述方法提高資料使用率，並減少時間。此外，我們使用帶有不同評分機制的選擇方法以及數據混合技術，以提高模擬數據的質量和多樣性。總而言之，我們可以通過平均距離誤差（MDE）提高36％的定位性能。

In machine learning-based indoor localization systems, the initial model built for a particular environment may no longer be useful when the environment is altered. As a result, we must recollect a big amount of data, which takes time. This approach, however, is ineffective. In our work, we aim to design a data augmenta tion system to reduce the time cost of the site survey; we used the Software-Defned Radio (SDR) hardware platform and open-source software platform (OAI 5G) to build the localization system. We can collect LTE-based features by this system and ft them into a machine-learning model to predict the user’s coordinates. Af ter that, we used the cycle Generative Adversarial Network (cycleGAN) for data augmentation, which is usually used to change the style of images. We propose a method based on Cycle GAN called Semi Cycle GAN to utilize a small amount of data in a new domain and a large amount of data in an old domain to produce a suffcient amount of data similar to the new domain. This method can save lots of time and incline the model’s accuracy. It is more suitable for low-dimension data than Cycle GAN. Compared with Vanilla GAN, we can enhance localization accuracy and utilize data from both the target and source domains. Our research can increase the data usage rate and decrease time by the above methods. In addition, we use the selection method with different scoring mechanisms and data mixing techniques to increase the quality and diversity of simulation data. In conclusion, we can improve the localization performance by 36% in Mean Distance Error (MDE).