A Semi-Supervised Ladder Network-Based Indoor Localization Using Channel State Information
Yi-Wei He, Tsai-Ting Hsu, Po‐Hsuan Tseng
Abstract
We propose a ladder network-based fingerprinting (LadderNetFi) method, which combines unsupervised learning with supervised learning in the neural network model for channel state information (CSI)-based indoor localization. The unsupervised part of LadderNetFi, which serves as a denoising autoencoder with skip connections from its encoder to the decoder, focuses on detailed features related to supervised learning. By dealing with the measurement uncertainty in the architecture design, a better generalization of fingerprints from the pre-processing of CSI amplitude leads to performance improvement compared to other deep learning-based methods. As semi-supervised learning, we verify that LadderNet trained using only a small portion of labeled measurements with the unlabeled measurements, e.g., 2,000 versus 20,000, and only a part of reference points, e.g., 19 versus 31, provides better performance than state-of-art methods using all labeled measurements. By reducing human labor from labeling, the deployment for LadderNetFi is scalable.