A Self-Supervised Masked Autoencoder Leveraging Temporal-Frequency Representation for CSI Localization

Channel state information (CSI) inherently contains rich temporal-frequency features, offering significant potential for high-precision indoor localization. However, conventional supervised methods, which primarily focus on learning direct mappings to location labels, often fail to fully exploit these inherent structures and ignore the vast amount of unlabeled CSI data available in real-world deployments. These limitations significantly constrain the practical feasibility and accuracy of localization systems. To address these challenges, we propose a novel self-supervised learning (SSL) framework, referred to as TF-MAE, by leveraging temporal-frequency embedding (TFE) and masked autoencoders (MAE). TF-MAE mainly consists of two key phases: self-supervised data recovery and MAE-enhanced CSI localization. In the self-supervised data recovery phase, we first apply stochastic channel augmentation to unlabeled data to simulate environmental interference, thereby enhancing the framework's robustness against dynamic conditions. Next, a TFE module extracts fine-grained features and reshapes them to match the framework's input format. A masked autoencoder processing (MAEP) method is then employed to learn robust and generalizable representations from incomplete input, enabling accurate reconstruction of CSI data. In the MAE-enhanced CSI localization phase, the pretrained framework rapidly adapts to the target environment using only a few labeled samples, achieving high-accuracy indoor localization. Experimental results based on both a public dataset and real-world measurements show that TF-MAE significantly outperforms state-of-the-art localization methods in terms of both accuracy and robustness.