Gaussian Mixture Model for Wireless Fingerprint Positioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Fingerprint-based positioning techniques, such as Pattern Matching (PTM), face challenges in achieving high positioning accuracy due to None Light of Sight (NLOS) multipath reflections, which cause biases in Time of Arrival (ToA) and Reference Signal Received Power (RSRP) measurements.
Innovation Solution
The method involves training a Gaussian Mixture Model (GMM) with an adaptive number of clusters based on the Non-Line of Sight (NLOS) level of each bin, using historical measurements and regions to determine parameter sets for Gaussian distributions, which indicate probabilities of device positioning within historical regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single Gaussian distribution is used for fingerprint positioning, then the positioning model is simple, but the positioning accuracy deteriorates due to NLOS multipath reflections causing bias in measurements
Solution Approach 1:
The patent segments the single Gaussian distribution into multiple Gaussian distributions (Gaussian mixture model), where each component represents different propagation conditions (LOS and NLOS). This segmentation allows the model to account for measurement biases under different conditions, thereby improving positioning accuracy while maintaining manageable complexity through structured parameterization.
Solution Approach 2:
The patent applies local quality by assigning different parameters (means, variances, weights) to different Gaussian components based on their specific characteristics. Each Gaussian component is optimized for specific propagation conditions, allowing the model to adapt locally to different measurement scenarios rather than using a uniform approach, thus improving overall positioning accuracy.
2Measurement precision
If the number of Gaussian components is increased to improve positioning accuracy, then the positioning accuracy improves, but the model complexity and computational burden increase
Solution Approach 1:
The patent employs dynamic model selection where the number of Gaussian components and their parameters are adaptively determined based on the specific environment and measurement data. The model can dynamically adjust the mixture weights and component parameters to match the actual propagation conditions, achieving high accuracy without requiring a fixed large number of components in all scenarios.
Solution Approach 2:
The patent utilizes parameter changes by optimizing the means, variances, and weights of Gaussian components based on training data. Through parameter estimation and optimization, the model adapts to different environments by adjusting its parameters rather than increasing model structure complexity, thus achieving improved accuracy with controlled complexity.
3Ease of manufacture
If traditional pattern matching is used, then the implementation is straightforward, but the positioning accuracy deteriorates due to bias in ToA and RSRP measurements from NLOS reflections
Solution Approach 1:
The patent introduces Gaussian distributions as intermediary models between raw measurements and final positioning results. These Gaussian components act as mediators that account for measurement biases under different propagation conditions, transforming biased measurements into accurate position estimates through probabilistic modeling and mixture weight computation.
Solution Approach 2:
The patent replaces the simple pattern matching mechanism with a probabilistic Gaussian mixture model. Instead of directly comparing measurement patterns, the system uses statistical modeling with Gaussian distributions to represent measurement uncertainties and biases, substituting a more sophisticated but accurate mathematical framework for the simple matching approach.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
Embodiments of the present disclosure relate to fingerprint positioning in a wireless communication network. A method for establishing a positioning model comprises: obtaining, at a first device, a set of training data comprising historical measurements and historical regions, a historical measurement being associated with a second device, the second device being within one of the historical regions when the historical measurement is obtained; and training the positioning model by using the historical measurements as input of the positioning model and the historical regions as output of the positioning model to obtain parameter sets of Gaussian distributions of the positioning model, the Gaussian distributions indicating probabilities that a third device is positioned in the historical regions, and the parameter sets corresponding to respective non-light of sight levels of the historical regions. In this way, the positioning accuracy of fingerprint positioning method can be greatly improved.