Graph-Based SAR Signal Denoising for Platform Position Perturbations

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Synthetic aperture radar systems face performance degradation due to position perturbations and noise interference, leading to poor imaging quality, especially when using distributed sensor units on moving platforms.

Innovation Solution

A graph-based denoising method is applied to synthetic aperture radar systems, combining robust perturbation estimation to process noisy array signals, using a graph model to regularize smoothness in the graph domain and sparse gradients in the time domain, and employing dual regularization to jointly denoise signals and estimate position perturbations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional denoising methods (e.g., matched filtering, coherence gating, CFAR detection) are used, then detection capability is maintained under ideal conditions, but performance degrades significantly in perturbed environments with multipath interference and noise

Engineering Contradiction:
Improvedetection precisionVSAvoiddetection reliability
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent introduces a graph neural network as an intermediary between the raw SAR signal and the detection process. The GNN learns to distinguish direct-path signals from multipath interference by modeling signal relationships in a graph structure, where nodes represent signal components and edges represent their relationships. This intermediary processing layer enables robust detection in perturbed environments by filtering out multipath effects before target detection.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent transforms the SAR signal processing approach by changing the parameter representation from traditional time-domain or frequency-domain signals to graph-structured representations. The signal is modeled as a graph where nodes contain signal features and edges encode spatial or temporal relationships, allowing the GNN to learn optimal parameter transformations for denoising and detection under varying perturbation conditions.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If traditional signal processing methods are applied, then computational simplicity is maintained, but ability to handle multipath interference and noise is insufficient

Engineering Contradiction:
Improveanti-interference capabilityVSAvoidprocessing complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent replaces traditional mechanical signal processing methods (filtering, gating, thresholding) with a data-driven graph neural network approach. Instead of applying fixed mathematical operations, the GNN learns adaptive processing strategies from training data, substituting rigid mechanical processing with flexible neural network-based processing that automatically adapts to different perturbation scenarios.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent adds a graph-structured dimension to the signal processing framework. By representing signals as graphs with nodes and edges capturing complex relationships, the method transforms one-dimensional signal processing into multi-dimensional graph processing, enabling the model to capture intricate interference patterns and signal relationships that traditional methods cannot handle.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Measurement precision

If graph neural network processing is implemented, then detection precision in perturbed environments is improved, but computational load increases

Engineering Contradiction:
Improvedetection precisionVSAvoidcomputational energy
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent performs preliminary action by pre-training the graph neural network on extensive training data that includes various perturbation scenarios. During training, the model learns optimal graph constructions, node feature extractions, and message-passing strategies. Once trained, the model can process test signals efficiently without requiring intensive computational resources during actual detection, as the heavy lifting was done during the preliminary training phase.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4165437B1Graph-based array signal denoising for perturbed synthetic aperture radar
Publication Date: 2026.04.29 MITSUBISHI ELECTRIC CORP
  • EP4165437B1 patent drawingFigure 1A~1C
  • EP4165437B1 patent drawingFigure 2
  • EP4165437B1 patent drawingFigure 3A~3C

AI summary

A radar image processing device is provided for generating a radar image from a region of interest (ROI). The radar image processing device receives transmitted radar pulses and radar echoes reflected from the ROI at different positions along a path of a moving radar platform and stores computer- executable programs including a range compressor, a graph modeling generator, a signal aligner, a radar imaging generator and a focused image generator. The radar image processing device performs range compression on the radar echoes by deconvolving the transmitted radar pulses and a radar measurement to obtain frequency-domain signals, generate a graph model represented by sequential positions of the moving radar platform and a graph shift matrix computed using the frequency-domain signals, iteratively denoise and align the frequency-domain signals to obtained denoised data and time shifts by solving a graph-based optimization problem represented by the graph model, wherein the approximated time shifts compensate phase misalignments caused by perturbed positions of the moving radar platform, and perform radar imaging based on the denoised data and the estimated time shifts to generate focused radar images.