🛰️ Radar Signal Denoising using 1D-CNN

Efficient Noise Reduction for Radar Chirp Signals using Deep Learning

This repository provides a deep learning-based solution for denoising radar chirp signals contaminated with additive noise. By utilizing a 1D Convolutional Neural Network (1D-CNN), this project demonstrates how to effectively reconstruct high-fidelity signals from noisy measurements, which is crucial for modern electronic warfare and radar systems.

---

🚀 Key Features

Complex Signal Processing	Implemented a custom RadarDataset that splits complex-valued signals into Real and Imaginary (I/Q) channels, enabling the CNN to process phase information effectively.
Encoder-Decoder Architecture	A deep 1D-CNN architecture designed to extract latent features and reconstruct the original signal while suppressing noise components.
Quantitative Evaluation	Performance is measured using PSNR (Peak Signal-to-Noise Ratio) to ensure high-quality signal reconstruction.
MLOps Integration	Fully integrated with Weights & Biases (W&B) for real-time experiment tracking, loss visualization, and hyperparameter tuning.
Multi-domain Visualization	Includes automated plotting for both Time-domain waveforms and Frequency-domain Spectrograms.

---

🏗️ Architecture & Methodology

1. Signal Processing Pipeline

• Preprocessing: Load .mat files and convert raw complex data into a 2-channel 1D tensor $(2 \times 1024)$.
• Feature Extraction: Two-stage Encoder using wide kernel sizes $(9, 5)$ to capture both local and global signal patterns.
• Reconstruction: Symmetric Decoder to map features back to the signal space.

2. Performance Metric

The model optimizes the Mean Squared Error (MSE) and is evaluated via $PSNR$:

$$PSNR = 20 \cdot \log_{10} \left( \frac{MAX_{signal}}{\sqrt{MSE}} \right)$$

---

📊 Results

Denoising Performance

The model effectively removes the noise floor while preserving the linear frequency modulation (LFM) characteristics of the chirp signal.

Category	Description
Noisy Input	Heavy interference obscuring the signal.
AI Denoised	Successfully recovered waveform with high PSNR.
Clean Target	The ground truth for comparison.

Tip

Spectrogram analysis confirms that the AI model significantly reduces the background noise floor across the entire frequency spectrum, maintaining the integrity of the original chirp slope.

---

Copyright © 2026 Jaehyeon Lee (382). All rights reserved.