Neural Network Weight Re-Initialization for Indoor Navigation RSS Fingerprinting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Indoor navigation systems face challenges in maintaining accurate positioning due to dynamic changes in indoor infrastructure, which affect wireless signal strength and require frequent recalibration of neural networks trained on outdated physical layouts.
Innovation Solution
A neural network re-training method that combines machine learning with a postulated mathematical model of received signal strength features, allowing for faster convergence and improved accuracy by re-initializing weights based on output errors exceeding a threshold, thus adapting to changes in indoor environments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual calibration is used to generate and maintain the fingerprint map, then positioning accuracy can be achieved, but the process becomes time-consuming and error-prone
Solution Approach 1:
The patent replaces manual mechanical calibration processes with an automated neural network system. The neural network automatically learns and maintains the fingerprint map by processing wireless signal data, eliminating the need for time-consuming manual calibration while maintaining positioning accuracy.
Solution Approach 2:
The neural network system performs self-calibration by continuously learning from incoming wireless signal data. It automatically updates the fingerprint map without human intervention, making the system self-maintaining and eliminating repetitive manual calibration efforts.
2Reliability
If the fingerprint map is maintained based on static physical layouts, then initial training can be completed, but the system becomes inaccurate when indoor infrastructure changes
Solution Approach 1:
The patent transforms the static fingerprint map into a dynamic system through the neural network. The network continuously adapts to changes in indoor infrastructure by learning from new wireless signal data, allowing the system to maintain accuracy despite physical layout modifications, added substructures, or remodels.
Solution Approach 2:
The system implements continuous feedback by monitoring wireless signal data and comparing it against the neural network's predictions. When discrepancies are detected due to infrastructure changes, the system automatically retrain's the neural network to adapt to the new environment, ensuring ongoing accuracy.
3Measurement precision
If full neural network re-training is performed to adapt to changes, then accuracy is maintained, but computational burden increases
Solution Approach 1:
The patent applies partial re-training by updating only specific portions of the neural network weights that are most affected by infrastructure changes, rather than performing complete re-training. This approach maintains positioning accuracy while significantly reducing computational resources and energy consumption compared to full re-training.
Data Source
AI summary
A method and system of maintaining a trained neural network for mobile device indoor navigation and positioning. The method comprises: determining, in the processor, at a first location relative to a wireless signal source at a second location, a set of received signal strength (RSS) input parameters in accordance with a postulated RSS model, the processor implementing an input layer of a neural network, the set of RSS input parameters providing an RSS input feature to the input layer of the neural network; receiving a set of RSS measured parameters acquired at a mobile device positioned at the first location from the wireless signal source at the second location; computing, at an output layer of the trained neural network, an output error based on comparing the RSS input feature to an RSS output feature generated at the output layer, the RSS output feature being generated at least in part based on a matrix of weights associated with at least a first neural network layer; and if the output error exceeds a threshold value, re-training the neural network based at least in part upon re-initializing the matrix of weights associated with the at least a first neural network layer.