06 / 2023 – 12 / 2023
Accurate localisation in indoor environments plays a crucial role in numerous applications such as asset tracking, indoor navigation, and context-aware computing. Radio frequency (RF) based localisation systems offer a promising solution, but their performance is limited by obstacles and signal propagation anomalies, particularly in none-line-of-sight (NLOS) situations [SH16]. Indoor environments present various challenges for RF localisation systems. Obstacles like walls, furniture, and human bodies can cause signal blockages and multi-path interference, leading to errors in the localization of a target device [CGH+21]. The successful detection of LOS and NLOS conditions in indoor RF localisation can significantly enhance the performance and reliability of indoor positioning systems [FNL22]. Many of the current state of the art machine learning models need labelled data to be trained, e.g. models based on transformer [TBJR22], transformation-CNN [CGH+21], LSTM [CLL+18], LSTM-CNN [NNCK18], ResNet, FCN, encoder [SKMM20]. Supervised models generalisation capabilities depend on sufficient training data, hence an extensive data recording is needed [PNC+20]. However, labelling NLOS is a challenging task in indoor environments due to transmission properties of radio signals, as not all blockages of visual light will also block radio signals due to lower frequency ranges. Therefore, often optical reference systems can not be used. However, it is possible to use unsupervised models instead [SKO+22, KMBG21, FLN21, FA19], which do not require labeled data. We want to improve on the performance of unsupervised models that may only have access to line of sight data during training. Those models are based on either detecting out of distribution (OOD) (e.g. VAE, EM-GMM) data points or they compute a threshold for which to classify a sample as anomalous (e.g. ARIMA). We want to further investigate unsupervised models for anomaly detection, extending existing models to include a dynamical model to exploit the temporal dependencies in the data, e.g. dynamical VAE (DVAE). These models are not only able to learn the distribution of the hidden latent states but may also learn a state transition probability. Therefore we can not only detect states that are OOD but also consider unlikely state transitions in time [KKS+21]. Different DVAE models are presented in the literature [GLB+21], such as Kalman VAE [FKPW17], recurrent VAE (RVAE)[LAPGH20], extended Kalman VAE (EKVAE) [KKS+21]. In this work, different DVAE models are evaluated for unsupervised NLOS identification and compared to deterministic supervised and unsupervised NLOS identification algorithms. The evaluations of the algorithms will be performed on a prerecorded industrial data set proposed in [SKO+22] and in a office setup, where the data collection and labelling has to be done during the thesis
[CGH+21] Zhichao Cui, Yufang Gao, Jing Hu, Shiwei Tian, and Jian Cheng. Los/nlos identification for indoor uwb positioning based on morlet wavelet transform and convolutional neural networks. IEEE Communications Letters, 25(3):879–882, 2021.
[CLL+18] Jeong-Sik Choi, Woong-Hee Lee, Jae-Hyun Lee, Jong-Ho Lee, and Seong-Cheol Kim. Deep learning based nlos identification with commodity wlan devices. IEEE Transactions on Vehicular Technology, 67(4):3295–3303, 2018.
[FA19] Jiancun Fan and Ahsan Saleem Awan. Non-line-of-sight identification based on unsupervised machine learning in ultra wideband systems. IEEE Access, 7:32464–32471, 2019.
[FKPW17] Marco Fraccaro, Simon Kamronn, Ulrich Paquet, and Ole Winther. A disentangled recognition and nonlinear dynamics model for unsupervised learning, 2017.
[FLN21] Laura Flueratoru, Elena Simona Lohan, and DragoÅŸ Niculescu. Self-learning detection and mitigation of non-line-of-sight measurements in ultra-wideband localization. In 2021 International Conference on Indoor Positioning and Indoor Navigation (IPIN), pages 1–8, 2021.
[FNL22] Xu Feng, Khuong An Nguyen, and Zhiyuan Luo. Wifi access points line-of-sight detection for indoor positioning using the signal round trip time. Remote Sensing, 14(23), 2022.
[GLB+21] Laurent Girin, Simon Leglaive, Xiaoyu Bie, Julien Diard, Thomas Hueber, and Xavier Alameda-Pineda. Dynamical variational autoencoders: A comprehensive review. Foundations and Trends in Machine Learning, 15(1-2):1–175, 2021.
[KKS+21] Alexej Klushyn, Richard Kurle, Maximilian Soelch, Botond Cseke, and Patrick van der Smagt. Latent matters: Learning deep state-space models. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 10234–10245. Curran Associates, Inc., 2021.
[KMBG21] Anil Kirmaz, Diomidis S. Michalopoulos, Irina Balan, and Wolfgang Gerstacker. Los/nlos classification using scenario-dependent unsupervised machine learning. In 2021 IEEE 32nd Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), pages 1134–1140, 2021.
[LAPGH20] Simon Leglaive, Xavier Alameda-Pineda, Laurent Girin, and Radu Horaud. A recurrent variational autoencoder for speech enhancement, 2020.
[NNCK18] Viet-Hung Nguyen, Minh Tuan Nguyen, Jeongsik Choi, and Yong-Hwa Kim. Nlos identification in wlans using deep lstm with cnn features. Sensors, 18:4057, 11 2018.
[PNC+20] JiWoong Park, SungChan Nam, HongBeom Choi, YoungEun Ko, and Young-Bae Ko. Improving deep learning-based uwb los/nlos identification with transfer learning: An empirical approach. Electronics, 9(10), 2020.
[SH16] Bruno Silva and Gerhard Hancke. Ir-uwb-based non-line-of-sight identification in harsh environments: Principles and challenges. IEEE Transactions on Industrial Informatics, 12:1–1, 06 2016.
[SKMM20] Maximilian Stahlke, Sebastian Kram, Christopher Mutschler, and Thomas Mahr. Nlos detection using uwb channel impulse responses and convolutional neural networks. In 2020 International Conference on Localization and GNSS (ICL-GNSS), pages 1–6, 2020.
[SKO+22] Maximilian Stahlke, Sebastian Kram, Felix Ott, Tobias Feigl, and Christopher Mutschler. Estimating toa reliability with variational autoencoders. IEEE Sensors Journal, 22(6):5133–5140, 2022.
[TBJR22] Slavica Tomović, Klemen Bregar, Tomaž Javornik, and Igor Radusinović. Transformer-based nlos detection in uwb localization systems. In 2022 30th Telecommunications Forum (TELFOR), pages 1–4, 2022.