GNSS Multipath Detection Aided by Unsupervised Domain Adaptation

    •  Zawislak, R., Greiff, M., Kim, K.J., Berntorp, K., Di Cairano, S., Mao, K., Parsons, K., Orlik, P.V., Sato, Y., "GNSS Multipath Detection Aided by Unsupervised Domain Adaptation", ION-GNSS+ Conference, DOI: 10.33012/​2022.18333, September 2022, pp. 2127-2137.
      BibTeX TR2022-118 PDF
      • @inproceedings{Zawislak2022sep,
      • author = {Zawislak, Remy and Greiff, Marcus and Kim, Kyeong Jin and Berntorp, Karl and Di Cairano, Stefano and Mao, Konishi and Parsons, Kieran and Orlik, Philip V. and Sato, Yuki},
      • title = {GNSS Multipath Detection Aided by Unsupervised Domain Adaptation},
      • booktitle = {ION-GNSS+ Conference},
      • year = 2022,
      • pages = {2127--2137},
      • month = sep,
      • publisher = {Institute of Navigation},
      • doi = {10.33012/2022.18333},
      • url = {}
      • }
  • MERL Contacts:
  • Research Areas:

    Artificial Intelligence, Control, Machine Learning, Signal Processing


This paper concerns the problem of multipath detection in global navigation satellite system (GNSS) positioning. An unsuper- vised machine learning (ML) approach is developed using a convolutional neural network in an autoencoder framework with k-means clustering in the latent space. Such methods often rely on large amounts of annotated data during training, which are difficult to collect in the context of GNSS applications. To this end, a joint approach is proposed to train the ML method on synthetic data generated by hardware in the loop (HIL) simulations, and use unsupervised domain adaptation (UDA) to reduce any discrepancies between real and simulated data. For this purpose, an UDA facilitated by a cycle-consistent generative ad- versarial network is proposed. It is shown that the proposed method can significantly improve the multipath detection accuracy on experimental data compared to baseline approaches (including both heuristic approaches and ML-methods). It is also shown that UDA can enhance GNSS estimation performance in the absence of annotated training data, achieving a prediction accuracy in hand-labeled experimental data of as much as 99 %.