Deep Learning TEC Map Reconstruction for Missing GNSS Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning techniques for predicting total electron content (TEC) maps face challenges due to missing data caused by geographical limitations of GNSS receivers, device failures, and data transmission errors, leading to overfitting or underfitting and reduced prediction accuracy.
Innovation Solution
A TEC map prediction system using deep learning that employs a convolutional generative adversarial network (DCGAN-PB) for reconstructing missing regions and a convolutional long short-term memory (ConvLSTM) model for predicting future TEC maps, integrating image processing techniques to generate optimized TEC map data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If deep learning models are trained on incomplete TEC map data with missing regions, then the model training process can proceed, but the prediction accuracy deteriorates due to overfitting or underfitting
Solution Approach 1:
The patent applies preliminary action by using a DCGAN-based generative model to reconstruct missing TEC map data before the main prediction model (ConvLSTM) is trained. This pre-processing step fills in the incomplete regions with synthesized data, ensuring that the prediction model receives complete training data and can achieve high accuracy without suffering from overfitting or underfitting due to missing values
2Use of energy by moving object
If traditional empirical models are used for TEC prediction, then computational efficiency is improved, but the ability to capture high-level spatial and temporal variability deteriorates
Solution Approach 1:
The patent replaces traditional empirical mathematical models with deep learning-based neural network models (DCGAN for reconstruction and ConvLSTM for prediction). This substitution enables the system to automatically learn complex spatial and temporal patterns from data, capturing high-level variability that empirical models cannot represent, while still maintaining computational efficiency through the power of modern GPU-based deep learning frameworks
3Loss of information
If physics-based models are used for TEC prediction, then the ability to trace ionospheric processes is improved, but computational complexity and time requirements increase
Solution Approach 1:
The patent uses copying by training deep learning models to replicate the behavior of complex physics-based models like WAM-IPE. The ConvLSTM model learns to predict TEC maps by capturing the essential dynamics of ionospheric processes from training data, providing a simplified computational alternative that maintains predictive accuracy while reducing computational complexity and execution time
4Measurement precision
If complete TEC maps are generated to train deep learning models, then prediction accuracy is improved, but data processing time increases
Solution Approach 1:
The patent applies preliminary action by pre-training the DCGAN model on available TEC map data to learn the statistical characteristics and spatial patterns of complete TEC maps. Once trained, the DCGAN can rapidly generate missing data during inference, allowing the system to produce complete TEC maps for prediction without the time-consuming process of collecting and processing extensive complete training datasets
Data Source
AI summary
The present disclosure relates to a TEC map prediction system and method using deep learning. The present disclosure relates to a technique for predicting two-dimensional TEC maps using a deep learning model, and more particularly, to a technology capable of more accurately restoring/predicting regional TEC maps with a small-scale structure to provide a precise TEC map.