Radio Network Position Prediction with Neural Error Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing localization methods based on radio networks, such as fingerprinting, suffer from inaccuracies in determining the position of objects, particularly in indoor environments, due to high positional errors.
Innovation Solution
A method involving a two-tier neural network structure is employed, where a first neural network predicts positions based on channel impulse responses, and a second neural network estimates positional errors, discarding high-error predictions to improve accuracy by averaging only low-error estimates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional geometric approaches or single neural network fingerprinting methods are used for localization, then the system complexity remains relatively low, but the localization accuracy and reliability deteriorate due to high positional errors
Solution Approach 1:
The localization system is segmented into two independent neural networks: a first neural network that generates position estimates from channel impulse responses, and a second neural network that estimates positional errors. This segmentation allows each network to specialize in a specific function, improving overall localization accuracy while maintaining manageable system complexity through modular architecture
Solution Approach 2:
The second neural network acts as an intermediary that evaluates the reliability of position estimates produced by the first neural network. By introducing this intermediate error estimation layer, the system can identify and discard unreliable position estimates, thereby improving final localization accuracy without requiring complete system redesign
2Measurement precision
If multiple neural networks are employed to improve localization accuracy through error mitigation, then the localization precision improves, but the computational complexity and processing time increase
Solution Approach 1:
The system performs partial action by selectively processing only those position estimates that meet reliability criteria. The second neural network evaluates error estimates, and only position estimates with acceptable error margins are retained for final positioning, avoiding unnecessary processing of clearly unreliable estimates and reducing overall computational waste
3Reliability
If a two-tier neural network structure with error estimation is implemented, then the reliability of position estimates improves through mitigation of high positional errors, but the device complexity increases due to additional network components
Solution Approach 1:
The neural network system is divided into two specialized components: a first neural network dedicated to position estimation and a second neural network dedicated to error estimation. This segmentation improves reliability by allowing each network to optimize its specific function while keeping individual network complexities manageable
Solution Approach 2:
The neural network system performs self-service through automated error detection and filtering. The second neural network automatically evaluates the reliability of position estimates generated by the first network, and the system autonomously discards unreliable estimates without requiring external intervention or complex manual validation procedures
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a computer-implemented method for generating a predictor for a position of an object within a predetermined environment (EN) by the use of a radio network (RN), where the radio network comprises a plurality of anchors (ai, ..., ai+3) being receivers for radio signals within the radio network, the method comprising the following steps: i) providing a plurality of data points (du), each data point comprising a different ground truth position of a sender (S, S') and several channel impulse responses, where each channel pulse impulse belongs to a respective anchor; ii) learning a first neural network structure, where a respective first neural network outputs a position; iii) learning a second neural network structure, where a respective second neural network is associated with a respective first neural network and outputs a positional error; iv) configuring the predictor such that the predicted position of the object is calculated.