基於遷移學習降低新環境訓練成本的行動通訊網路室內定位系統

調查基於機器學習室內定位系統的相關文獻，當環境發生變化時，針對特定環境所訓練的原始模型可能會變得無效。因此，重新收集足夠數量的無線電特徵資料並重新訓練針對新環境的定位模型將是必要的。然而，對於部署定位系統而言，這種針對於新環境的定位議題，所衍生出的訓練成本，可能會使得大量部署定位系統成為不切實際的目標。為了解決增加訓練成本的問題，我們的研究旨在找到一種可以有效利用原始定位模型和知識的定位解決方案。我們使用軟體定義無線電 (Software-Defined Radio，SDR) 硬體平台和開源軟體環境 (OAI 5G) 來建置具有大頻寬及低延遲特性的行動通訊網路。基於這個行動通訊網路，我們可以取得基於 LTE 的無線電特徵，例如：接收訊號強度（Received Signal Strength，RSS）和通道狀態資訊（Channel State Information，CSI）。將這些無線電特徵，通過機器學習的方法進行訓練，可以以所期望的準確度來定位用戶設備 (User Equipment，UE)。從我們的實驗結果顯示，50% 的測試資料可以達到定位誤差小於 50 公分（cm）。此外，將 XGBoost 模型與 FN (Fusion Network) 模型相結合，我們可以實現 35.77 cm 的更好定位誤差。由於，這種方法需要相對大量的資料來進行機器學習模型訓練，因此，我們應用生成對抗網路（Generative Adversarial Network，GAN）來進行資料擴充。為了在遇到新環境時能夠降低訓練成本，我們嘗試以模型微調來轉移這些領域知識。我們還應用超參數最佳化來尋找模型的最佳超參數，從而提高微調模型的性能。結合超參數最佳化和資料擴充的技術，微調模型的性能可以提高 28%，性能上限和我們微調模型的性能之間的性能差距可以減少大約18%。

關鍵字

遷移學習；室內定位；行動通訊網路；通道狀態資訊；接收訊號強度；機器學習；深度學習

並列摘要

A literature review on indoor localization system based on machine learning indicates that the original model trained for a specific environment could become ineffective when the environment is changed. Re-collecting a sufficient amount of data and re-training a new localization model thus could be required. However, such site survey cost could be undesirable for deploying the localization system. To address this problem of increased site survey cost, our study aims to design a localization system that can effectively use pre-trained models and knowledge. We start from a cellular network by using the Software-Defined Radio (SDR) hardware platform and an open-source software platform (OAI 5G) to build the baseline localization system. Relevant LTE-based radio features such as Received Signal Strength (RSS) and Channel State Information (CSI) are captured to sent to a machine-learning model. Through machine learning-based methods, the User Equipment (UE) can be located with the desired performance. Our experimental results show that 50% of the testing data can achieve a localization accuracy with error less than 50 centimeters (cm). Furthermore, with a XGBoost model combined with a fusion network (FN), we can achieve an even better result with a prediction error of 35.77 cm. As such methods require a relatively large amount of data for training, we apply a Generative Adversarial Network (GAN) for data augmentation. To reduce site survey cost when adapting the model to a new environment, we first apply model fine-tuning for transferring domain knowledge. We then apply hyper-parameter optimization to find the best hyper-parameter set for the model. By combining hyper-parameter optimization and data augmentation, the performance of the fine-tuned model can be improved by 28%, and the performance gap between the performance upper bound and the performance of our fine-tuned model can be reduced by about a percentage of 18%.

並列關鍵字

Transfer Learning ； Indoor Localization ； Cellular Network ； Channel State Information (CSI) ； Received Signal Strength (RSS) ； Machine Learning ； Deep Learning

參考文獻

[1] K. Liu, H. Zhang, J. K.-Y. Ng, Y. Xia, L. Feng, V. C. Lee, and S. H. Son, “Toward low-overhead fingerprint-based indoor localization via transfer learning: Design, implementation, and evaluation,” IEEE Transactions on Industrial Informatics, vol. 14, no. 3, pp. 898–908, 2017.

Google Scholar

[2] R. W. Schafer, “What Is a Savitzky-Golay Filter? [Lecture Notes],” IEEE Signal Processing Magazine, vol. 28, no. 4, pp. 111–117, 2011.

Google Scholar

[3] 3GPP, “Feasibility study on new services and markets technology enablers; stage 1 (version 14.1.0 Release 14),” TR 22.891, June 2016.

Google Scholar

[4] F. Zafari, A. Gkelias, and K. K. Leung, “A Survey of Indoor Localization Systems and Technologies,” IEEE Communications Surveys Tutorials, vol. 21, no. 3, pp. 2568–2599, 2019.

Google Scholar

[5] J. A. del Peral-Rosado, R. Raulefs, J. A. López-Salcedo, and G. Seco-Granados, “Survey of Cellular Mobile Radio Localization Methods: From 1G to 5G,” IEEE Communications Surveys Tutorials, vol. 20, no. 2, pp. 1124–1148, 2018.

Google Scholar

國際替代計量

基於遷移學習降低新環境訓練成本的行動通訊網路室內定位系統

全文下載

主題瀏覽