Signal-Guided Masked Autoencoder for Wireless Positioning With Limited Labeled Samples

Abstract

In this paper, a signal-guided masked autoencoder (S-MAE) based semi-supervised learning framework is proposed for high-precision positioning with limited labeled channel impulse response (CIR) samples. To release the overfiting effect of the neural network under insufficient labeled samples, we design a two-stage training strategy based on the proposed S-MAE model, which can be divided into pre-training and fine-tuning stage. In the pre-training stage, we design an effective masking pattern in the antenna domain to learn the latent representation of CIR by utilizing a large number of unlabeled CIR samples. Besides, we introduce the channel attention mechanism to enhance the feature extraction ability in the S-MAE. In the fine-tuning stage, we use limited labeled CIR samples to fine-tune the pre-training model in a manner of supervised learning, where the long short term memory (LSTM) network is introduced to realize the mapping from CIR to user coordinates. Experiment results show that: 1) for the case of limited labeled samples, the proposed S-MAE model has superior positioning accuracy compared to conventional positioning models. 2) For the case of non-ideal CIR scenarios, the robustness performance of the S-MAE is better than that of other benchmark models. 3) The performance gain of the proposed S-MAE under different masking patterns/ratios on the CIR sample is presented, which verifies the effectiveness of the proposed masking strategy.