Scientific research: radar, MATLAB and AI

The starting point

A thesis carried out between ESME Sudria and the Beijing Institute of Technology (supervisor: Prof. Xiuzhu Ye). Beyond the topic itself, the project shows a way of conducting research: building a physical measurement chain, connecting it to machine learning, then auditing it methodically with the help of AI agents.

The problem

Small drones are hard to detect: slow and low-flying, they are easily mistaken for birds on a conventional radar. Their micro-Doppler signature, the modulation created by the rotating blades, should in principle make them recognisable. Exploiting it in the W band (77 GHz) requires a dedicated processing chain and rigorous validation, because a model can look excellent while generalising poorly.

The answer

A complete chain links radar acquisition to the decision: MATLAB for signal processing (range profiles, range-Doppler maps, spectrograms), Python for learning (an SVM on physical features against a ResNet-18 network), and a paired statistical test to settle the comparison. AI agents run MATLAB in batch mode to replay the scripts, confront results with the physics and document every correction reproducibly.

• FMCW MATLAB chain: range profiles, range-Doppler maps, STFT spectrograms, cadence-velocity diagrams
• 21 real W-band (77 GHz) recordings processed with quality control
• MATLAB-to-Python bridge: the spectrograms feed the learning models
• SVM versus ResNet-18 comparison settled by a paired McNemar test
• Audit of the chain by AI agents driving MATLAB in batch: 8 documented, reproducible findings
• Leak-free evaluation protocols that keep recordings separate
• Automated generation of the report and the defence slides

Under the hood

Radar acquisitions are processed in MATLAB: beamforming, a Fourier transform for range, then an STFT along slow time to extract the micro-Doppler modulation. The resulting spectrograms feed the Python side, organised into data, training and evaluation modules, where an SVM on physical features competes with a ResNet-18. For the audit, AI agents launch MATLAB from the command line on the 21 recordings, produce diagnostic scripts and replayable CSV outputs, and confront each result with the radar's physical parameters.

The notable challenge

The most spectacular result, a very high accuracy over six classes, turned out to be biased: random frame splitting placed near-identical excerpts in both training and test sets. Rather than keep this flattering figure, the protocol was redone by keeping recordings separate to obtain an honest measure of generalisation; this rigour, documented in an audit, becomes a contribution in itself.

My role

Odilon designed the whole project: the radar processing chain in MATLAB, the learning side in Python, the evaluation protocols, the audit carried out with AI agents, the thesis report and the defence.

Where the project stands

The processing chain works on real data and the comparison bench runs on the public dataset. The audit showed that the initial score overestimated generalisation: the data splits were corrected before any figure was retained, and retraining is the next step, followed by adaptation to the W-band data.

• Re-run training on the corrected splits
• Adapt the models to real W-band data
• Study real-time processing

A few figures

The project’s tools

Signal processing

MATLAB R2025b (STFT, FMCW, Capon beamforming)

Learning

Python, scikit-learn (SVM), PyTorch / torchvision (ResNet-18)

AI agents

Claude Code: batch MATLAB execution, reproducible audit

Data

DIAT-µSAT public dataset, 77 GHz W-band acquisitions

Reporting

NumPy, SciPy, Matplotlib, PDF / PPTX generation