CSI-Based Indoor Localization Using Deep Metric Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing wireless localization methods in indoor environments, such as those based on triangulation or trilateration, are not robust due to complex multipath issues, and there is a need for more reliable techniques to determine the position of wireless devices using channel state information (CSI) for applications like augmented reality and unmanned vehicles.
Innovation Solution
Utilizing channel state information (CSI) for localization through time reversal resonance index (TRRI) or deep metric learning, which involves training an embedding neural model to compute a distance metric proportional to the true distance between spatial points, leveraging CSI data to determine positions and similarities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If triangulation or trilateration methods are used for wireless localization, then positioning can be achieved, but the accuracy and robustness deteriorate due to complex multipath issues in indoor environments
Solution Approach 1:
The patent replaces traditional geometric localization methods (triangulation/trilateration) with a machine learning-based approach using neural networks. The system learns optimal positioning from training data consisting of wireless signal measurements and known device locations, substituting mathematical geometry with data-driven pattern recognition that is more robust to multipath effects.
Solution Approach 2:
The patent transforms the localization problem by changing from direct geometric calculation to learning a mapping function between wireless signal parameters (RSSI, channel state information) and spatial coordinates. The neural network learns optimal parameter transformations during training, enabling accurate positioning despite signal variations caused by multipath propagation.
2Measurement precision
If deep metric learning with embedding neural model is used to compute distance metrics, then localization accuracy improves, but computational complexity and training requirements increase
Solution Approach 1:
The patent performs the computationally intensive neural network training in advance during an offline phase, before actual localization operations. The trained model is then deployed for efficient online inference, separating the complex learning process from the operational phase and reducing real-time computational requirements.
Solution Approach 2:
The patent uses the trained neural network model as a reusable component that can be deployed across multiple devices or instances. Once trained on comprehensive data, the same model can serve multiple localization tasks, amortizing the training cost across numerous inferences without repeating the complex training process.
Data Source
AI summary
Embodiments in accordance with this disclosure provide methods to compute a distance metric between two wireless data samples, including using a time reversal resonance index (TRRI) or deep metric learning, including different input shaping methods for channel state information or channel impulse response data to be used as inputs to a deep learning model.