Laser-to-Piezoelectric Signal Mapping for Ultrasonic Guided Waves

A neural network system to map laser ultrasonic signals to piezoelectric ultrasonic signals for enhanced signal quality.

🎯 Project Overview

Component	Description
Input	21×21 grid laser ultrasonic signals (→ interpolated to 41×41)
Target	41×41 grid piezoelectric ultrasonic signals
Physics	Aluminum plate, 200kHz center frequency
Models	1D DenoiseNet + 2D FluxUNet (CWT-based)

---

🚀 Quick Start

# 1. Install dependencies with uv
uv sync

# 2. Run Gravity Loop (FluxUNet 2D U-Net)
uv run python src/scripts/train_flux.py --sanity-check
uv run python src/scripts/train_flux.py --overfit-test --samples 50 --epochs 50

# 3. Full training
uv run python src/scripts/train_flux.py --epochs 100

# 4. Visualize results
uv run python src/core/visualize.py --num-samples 10

# Legacy: 1D DenoiseNet
uv run python src/scripts/train_denoisenet.py --overfit-test --epochs 50 --samples 100

---

📁 Project Structure

DnCNN/
├── src/
│   ├── models/                  # Neural Network Architectures
│   │   ├── denoisenet_1d/       # DenoiseNet (1D CNN)
│   │   ├── dncnn_cwt/           # DnCNN-CWT (2D ResNet)
│   │   └── flux/                # FluxUNet (Signal-to-Image)
│   ├── data/                    # Data Management
│   │   ├── dataset.py           # PyTorch Dataset
│   │   └── loader.py            # MATLAB Data Loading
│   ├── core/                    # Core Utilities
│   │   ├── transmuters/         # Wavelet Processing
│   │   └── visualize.py         # Plotting Tools
│   └── scripts/                 # Training Scripts
│       ├── train_flux.py        # FluxUNet Training
│       └── train_denoisenet.py  # DenoiseNet Training
├── configs/
│   └── default.yaml             # Hyperparameters and data paths
├── checkpoints/
│   └── gravity_loop/            # FluxUNet checkpoints
└── pyproject.toml               # Dependencies

---

🧠 Network Architectures

FluxUNet (2D U-Net) - Recommended

Signal-to-Image-to-Signal Pipeline:

1D Signal → CWT → 2D Spectrogram → FluxUNet → ICWT → 1D Signal
 (1024)         (2×64×1024)                         (1024)

Component	Details
Transform	Complex Morlet CWT (cmor1.5-1.0)
Channels	2 (Real/Imaginary for phase preservation)
Encoder	3 stages: 64→128→256 with ResBlocks
Decoder	Skip connections + ConvTranspose2d
Parameters	~9.3M

Loss Function:

L_total = MSE + 0.5 × SpectralConvergence + 0.1 × PhaseLoss

DenoiseNet (1D CNN) - Legacy

Encoder: Conv1d(1→64, k=7) → MaxPool(2) → Conv1d(64→128, k=5) → MaxPool(2)
Bottleneck: Conv1d(128→128, k=3)
Decoder: ConvTranspose1d(128→64) → ConvTranspose1d(64→1)

---

📊 Verification

FluxUNet Pipeline

1. Sanity Check: CWT→ICWT reconstruction (Correlation > 0.999)
2. Overfit Test: 50 samples, expect loss convergence
3. Visualization: checkpoints/gravity_loop/random_samples_comparison.png

Metrics

• SNR (dB): Signal-to-Noise Ratio improvement
• SSIM: Structural similarity on spectrograms
• Correlation: Time-domain waveform similarity

---

🔧 Configuration

Edit configs/default.yaml:

data:
  noisy_path: "path/to/41_41.mat"
  target_path: "path/to/51_51.mat"
  
signal:
  sampling_rate: 6.25e6    # 6.25 MHz
  target_length: 1024
  bandpass_low: 100e3      # 100 kHz
  bandpass_high: 300e3     # 300 kHz

training:
  batch_size: 32
  learning_rate: 0.001
  epochs: 100
  device: "cpu"

---

📚 Documentation

Document	Location
DenoiseNet Architecture	src/models/denoisenet_1d/README.md
DnCNN-CWT Architecture	src/models/dncnn_cwt/README.md
FluxUNet Architecture	src/models/flux/README.md

---

🛠️ Dependencies

pip install torch numpy scipy matplotlib pyyaml tqdm PyWavelets