Recurrent Neural Network Position Estimation Under Multipath
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional Global Navigation Satellite System (GNSS) receivers face challenges in accurately estimating position in urban environments due to multipath transmission effects, where reflection and refraction of satellite signals lead to inaccurate position estimation, and existing methods for mitigating multipath are computationally complex or ineffective.
Innovation Solution
A positioning system that uses a recurrent neural network (RNN) to directly estimate vehicle position from noisy phase measurements of satellite signals, including both line-of-sight and multipath transmissions, without the need for explicit multipath detection, by applying attention-based multimodal fusion and training with data from log files to minimize the impact of multipath noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional temporal methods are used to minimize multipath effects, then position estimation accuracy is improved, but computational complexity increases and effectiveness decreases when delays are short
Solution Approach 1:
The patent replaces conventional temporal signal processing methods with a deep learning-based neural network approach. The neural network is trained to directly estimate position from GNSS measurements, automatically learning to distinguish LOS and multipath signals without requiring explicit temporal analysis or delay estimation, thereby reducing computational complexity while maintaining accuracy.
Solution Approach 2:
The patent changes the approach from analyzing time-domain parameters (signal delays) to using a neural network that processes measurements in a transformed feature space. The network learns optimal parameter transformations during training, enabling it to handle short delays that are difficult to resolve with conventional temporal methods.
2Measurement precision
If multiple antennas are used for multipath detection and mitigation, then position estimation accuracy is improved, but device complexity and cost increase
Solution Approach 1:
The patent substitutes the physical approach of using multiple antennas with a computational approach using a single antenna and neural network processing. The neural network compensates for the lack of spatial diversity by learning temporal and statistical patterns that distinguish LOS from multipath signals, achieving similar accuracy without additional hardware complexity.
3Measurement precision
If multipath detection is performed before position estimation, then position estimation accuracy is improved, but processing time increases and rapidness decreases
Solution Approach 1:
The patent merges the multipath mitigation function and position estimation function into a single neural network processing step. Instead of performing separate multipath detection and then position estimation, the neural network performs both tasks simultaneously by learning to estimate position directly from measurements that may contain multipath, eliminating the need for separate processing stages and reducing overall processing time.
Solution Approach 2:
The neural network is pre-trained offline with large datasets containing both LOS and multipath signals. This preliminary training enables the network to automatically adapt to multipath conditions during real-time operation without requiring runtime detection or adjustment, allowing rapid position estimation even in challenging multipath environments.
4Adaptability or versatility
If conventional methods are used, then position estimation can be performed, but adaptability to distributed applications and cloud implementation is limited
Solution Approach 1:
The patent replaces conventional real-time signal processing with a pre-trained neural network model that can be deployed in distributed environments. The neural network enables cloud-based or edge-based implementation where measurements can be processed remotely, facilitating distributed applications while maintaining rapid response times through efficient inference.
Data Source
AI summary
A positioning system for tracking a position of a vehicle includes a receiver configured to receive phase measurements of satellite signals received at multiple instances of time from multiple satellites, and a memory configured to store a recurrent neural network trained to determine a position of the vehicle from a set of phase measurements in a presence of noise caused by a multipath transmission of at least some of the satellite signals at some instances of time. A processor of the positioning system is configured to track the position of the vehicle over different instances of time by processing the set of phase measurements received at each instance of time with the recurrent neural network to produce the position of the vehicle at each instance of time.