Summary
Wireless networks routinely estimate how signals propagate between devices and base stations. These channel estimates can also support localization. Self-supervised learning can make use of large collections of channel measurements and reduce the number of reference positions that need to be surveyed.
How channel estimates reveal location
Channel state information (CSI) describes how the wireless channel changes a signal’s amplitude and phase across frequencies and antennas. Propagation delays produce patterns across frequency, while the directions from which signals arrive produce patterns across the antenna array. Together, these patterns can serve as a radio fingerprint of a device’s location.
Consider a mobile robot in a factory. Machinery may block the direct path to a base station, but signals reflected from walls and equipment can still carry information about the robot’s position. Reusing communication channel estimates for this purpose is an example of integrated sensing and communication (ISAC). The network can collect CSI during routine operation, while obtaining accurate position labels typically requires a separate survey or tracking system.
Learning from unlabeled CSI
Self-supervised learning uses the measurements themselves to define a training task. The task determines which properties of the channel the model learns to represent. Several approaches are relevant to localization.
- Predicting missing measurements. The model uses part of a channel estimate to predict information hidden from its input, such as selected subcarriers or antenna measurements. This can help it learn relationships across frequency and antennas.
- Learning from different views of the same channel. The model learns to extract similar features from different views of a measurement. These views can be created by selecting different subcarrier groups or varying the overall signal gain. The aim is to retain spatial cues that remain useful across such changes. Self-Supervised and Invariant Representations for Wireless Localization
- Learning spatial neighborhoods. Channel charting organizes measurements so that devices at nearby locations tend to appear close together in a learned map. The resulting chart can describe relative movement. Reporting positions in building coordinates requires a spatial reference, such as surveyed points or known access-point locations combined with propagation constraints. Channel Charting in Real-World Coordinates
From learned features to reliable positions
For coordinate estimation, a common next step is to train a position estimator using the learned features and a smaller set of measurements taken at known locations. The practical question is how much surveying this saves for a given accuracy. That can be tested by varying the number of labeled positions and evaluating on separate measurement runs, including runs with different device orientations, temporary blockages, or changes in the environment.
The network can also influence which measurements are available. A reconfigurable intelligent surface (RIS), for example, can change the reflected paths and provide additional views of roughly the same device position. In an ISAC system, the improvement in localization accuracy has to be weighed against the time spent collecting those extra measurements.
Simplified summary
Radio signals change as a device moves, even when its direct path to a base station is blocked. Learning from these changes can help a wireless network estimate where a device is with fewer measurements taken at known locations.