🎯 Acoustic Drone Detection System - Complete Technical Documentation

📋 Table of Contents

1. System at a Glance
2. Dataset Description
3. Preprocessing Pipeline
4. CRNN Architecture
5. Training Process
6. Class Signatures
7. Usage
8. Performance

---

🎪 System at a Glance

This project implements a Convolutional Recurrent Neural Network (CRNN) with Temporal-Frequency Attention for real-time acoustic drone detection. The system classifies audio into three categories: Background, Drone, and Helicopter.

Complete System Flow

Fig. 1 — Complete system flowchart. End-to-end pipeline from audio input to classification with confidence scores. Metadata

Key Features:

• ✅ Multi-channel preprocessing: Mel Spectrograms + MFCCs + Spectral Features
• ✅ CRNN with Attention: CNN feature extraction + BiGRU temporal modeling
• ✅ Efficient: 2.08M parameters, 10-20ms inference (GPU)
• ✅ Robust: Handles noisy environments, balanced classes

---

📊 Dataset Description

EDTH Munich Acoustic Drone Detection Dataset

Validation Results:

✓ Train directory: data/edth_munich_dataset/data/train
✓ Val directory: data/edth_munich_dataset/data/val

Class: drone        | Train: 180 | Val:  60
Class: helicopter   | Train: 180 | Val:  60  
Class: background   | Train: 180 | Val:  60

Class balance ratio: 1.00x (perfectly balanced)

Class	Train Samples	Val Samples	Acoustic Characteristics
Drone	180	60	Multi-rotor UAV, 500-3000 Hz, harmonic comb pattern
Helicopter	180	60	Single/dual rotor, 50-800 Hz, low-frequency fundamental
Background	180	60	Urban/ambient noise, broadband, non-periodic

Audio Specifications:

• Format: WAV (Waveform Audio)
• Original SR: 44.1 kHz → Resampled to 22.05 kHz
• Duration: 5s → Trimmed to 3s fixed windows
• Channels: Mono
• Bit depth: 16-bit PCM

---

🔬 Preprocessing Pipeline

The preprocessing transforms raw audio into a 3-channel tensor (like RGB for images) capturing complementary acoustic features.

Pipeline Overview

Fig. 2 — Preprocessing pipeline. Mel spectrogram, MFCC, and spectral features stacked into a 3-channel input (3×128×130). Metadata

Step-by-Step Process

1. Audio Loading & Resampling

# Load audio
audio, sr = librosa.load(audio_path, sr=22050, duration=3.0)

# Normalize to [-1, 1]
audio = librosa.util.normalize(audio)

# Output: 66,150 samples (3.0s × 22,050 Hz)

2. Mel Spectrogram (Channel 0)

• Purpose: Time-frequency representation
• Captures: Harmonic patterns, rotor blade frequencies
• Config: 128 Mel bands, n_fft=2048, hop=512
• Output Shape: [128, 130]
• Drone signature: Sharp harmonics 500-3000 Hz

3. MFCC + Deltas (Channel 1)

• Purpose: Timbral texture and dynamics
• Captures: Spectral envelope, sound source characteristics
• Config: 40 MFCCs + 40 Δ + 40 ΔΔ = 120 coefficients
• Output Shape: [128, 130] (padded to match)
• Drone signature: Periodic MFCC stripes

4. Spectral Features (Channel 2)

• Purpose: Spectral shape characteristics
• Captures: Contrast, rolloff, bandwidth
• Features: Spectral contrast (7) + rolloff (1) + bandwidth (1) = 9
• Output Shape: [128, 130] (padded)
• Drone signature: High spectral contrast

5. 3-Channel Stacking

combined = np.stack([mel_spec, mfcc, spectral], axis=0)
# Final shape: [3, 128, 130] → Model input

Configuration Summary

Parameter	Value	Purpose
sample_rate	22,050 Hz	Nyquist: 11 kHz (captures drone frequencies)
duration	3.0 s	Fixed-length windows
n_samples	66,150	Total samples per clip
n_fft	2,048	FFT window size
hop_length	512	~23ms per frame
n_mels	128	Mel filter banks
n_mfcc	40	MFCC coefficients
fmin / fmax	20 / 8,000 Hz	Frequency range

📊 Detailed Feature Analysis (click to expand)

Time-Frame Calculation:

frames = (n_samples - n_fft) / hop_length + 1
       = (66,150 - 2,048) / 512 + 1  
       = 130 frames

Each frame = 23.2ms of audio (512 / 22,050).

---

🏗️ CRNN Architecture

Model Introspection (Computed from Actual Model)

Fig. 3 — CRNN architecture with attention. Layer-by-layer breakdown showing actual shapes and parameter counts from model introspection. Metadata

Verified Architecture Details:

✓ Loaded CRNN model
  Total parameters: 2,080,323
  Trainable parameters: 2,080,323  
  Model size: ~7.9 MB (FP32)

Layer-by-Layer Breakdown

Layer	Input Shape	Output Shape	Params	Activation
Input	[1, 3, 128, 130]	[1, 3, 128, 130]	0	-
Conv Block 1	[1, 3, 128, 130]	[1, 32, 64, 65]	960	ReLU
Conv Block 2	[1, 32, 64, 65]	[1, 64, 32, 32]	18,624	ReLU
Conv Block 3	[1, 64, 32, 32]	[1, 128, 16, 16]	74,112	ReLU
TF-Attention	[1, 128, 16, 16]	[1, 128, 16, 16]	16,704	Sigmoid
Reshape	[1, 128, 16, 16]	[1, 16, 2048]	0	-
BiGRU	[1, 16, 2048]	[1, 16, 256]	1,969,152	tanh
Temporal Pool	[1, 16, 256]	[1, 256]	0	-
Classification	[1, 256]	[1, 3]	771	Softmax
TOTAL	-	-	2,080,323	-

Parameter Distribution

BiGRU:         94.7% (1,969,152 params) ← Largest component
Conv Blocks:    4.5%    (93,696 params)
TF-Attention:   0.8%    (16,704 params)
Classifier:    <0.1%       (771 params)

Architecture Explained

1. Conv Blocks (Feature Extraction)

• 3 conv blocks with increasing channels: 3→32→64→128
• Each block: Conv2d(k=3, p=1) + BatchNorm + ReLU + MaxPool(2)
• Reduces spatial dims while extracting hierarchical features
• Receptive field grows: 3×3 → 7×7 → 15×15 pixels

2. Temporal-Frequency Attention

• Temporal branch: Learns important time frames
• Frequency branch: Learns important frequency bands
• Combined: Element-wise multiplication for joint attention
• Purpose: Focus on rotor harmonics, suppress background

3. Bidirectional GRU

• Input: Reshaped to [batch, time=16, features=2048]
• 2 layers, hidden_size=128, bidirectional → output dim=256
• Forward pass: Past → future context
• Backward pass: Future → past context
• Captures temporal dependencies (periodic rotor patterns)

4. Classification Head

• Temporal mean pooling: [B, 16, 256] → [B, 256]
• Dropout(0.3) for regularization
• Linear(256 → 3) → Softmax
• Output: [P(background), P(drone), P(helicopter)]

Activation Functions

Component	Activation	Formula	Range
Conv blocks	ReLU	max(0, x)	[0, ∞)
Attention	Sigmoid	1/(1+e^-x)	[0, 1]
GRU gates	Sigmoid	1/(1+e^-x)	[0, 1]
GRU candidate	Tanh	(e^x - e^-x)/(e^x + e^-x)	[-1, 1]
Output	Softmax	e^xi / Σe^xj	[0, 1], Σ=1

---

🎓 Training Process

Training Pipeline

Fig. 4 — Training pipeline. Seven-step process from data loading through optimization with early stopping.

Configuration

Optimization:

optimizer = AdamW(lr=1e-4, weight_decay=1e-4, betas=(0.9, 0.999))
scheduler = CosineAnnealingLR(T_max=epochs, eta_min=1e-6)
criterion = CrossEntropyLoss(weight=class_weights)

Regularization:

• Dropout: 0.3 (30%)
• Gradient clipping: max_norm=1.0
• Weight decay: 1e-4 (L2 regularization)
• Batch normalization in all conv blocks

Data Augmentation (training only):

• Time shifting
• Pitch shifting (±2 semitones)
• Adding Gaussian noise
• Time stretching (0.8-1.2×)

Training Hyperparameters:

Parameter	Value	Purpose
Batch size	32	Memory vs convergence trade-off
Epochs	50-100	With early stopping
Initial LR	1e-4	AdamW learning rate
Min LR	1e-6	Cosine annealing floor
Weight decay	1e-4	L2 regularization
Patience	10	Early stopping patience
Metric	Macro F1	Validation metric

Class Balancing:

• WeightedRandomSampler ensures equal class exposure
• Class weights in loss function
• Perfectly balanced dataset (180/180/180) helps

---

🎵 Class Signatures

Acoustic Characteristics

📊 Visual Comparison (click to expand)

Class	Frequency Range	Spectral Pattern	Temporal Pattern	Distinguishing Features
Drone	500-3000 Hz	Sharp harmonic comb	Steady-state	• High spectral contrast • Periodic MFCC • Narrow bandwidth
Helicopter	50-800 Hz	Multiple harmonics (main+tail rotor)	Rhythmic modulation	• Low-frequency dominant • Blade passage "thump" • Complex harmonic structure
Background	Broadband	Non-periodic, stochastic	Irregular, transient	• Low spectral contrast • High flatness • No harmonic comb

How the Model Distinguishes Classes

1. Mel Spectrogram (Channel 0)

• Drones: High-frequency harmonics (1-3 kHz)
• Helicopters: Low-frequency rhythmic patterns (50-500 Hz)
• Background: Broadband, non-periodic

2. MFCC (Channel 1)

• Captures timbral "fingerprint"
• Drones/helicopters: Periodic stripes
• Background: Random, irregular

3. Spectral Features (Channel 2)

• Spectral contrast: High for rotorcraft, low for background
• Spectral rolloff: Quantifies frequency distribution
• Bandwidth: Narrow for drones, wide for background

4. Attention Mechanism

• Learns to focus on discriminative regions
• Drones: Attends to 1-3 kHz harmonics
• Helicopters: Attends to low-freq rotor patterns
• Background: Suppresses non-periodic noise

5. BiGRU Temporal Modeling

• Captures periodic patterns in drones/helicopters
• Distinguishes steady-state vs rhythmic modulation
• Learns background lacks long-term structure

---

🚀 Usage

Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Train Model

python train_sota_model.py \
    --train-dir data/edth_munich_dataset/data/train \
    --val-dir data/edth_munich_dataset/data/val \
    --epochs 50 \
    --batch-size 32

3. Inference

from sota_inference import AcousticDroneClassifier

# Load model
classifier = AcousticDroneClassifier(
    model_path='models/crnn_combined/crnn_final.pt',
    labels_path='models/crnn_combined/labels.json'
)

# Classify audio
prediction, confidence, probabilities = classifier.classify('audio.wav')

print(f"Prediction: {prediction}")
print(f"Confidence: {confidence:.2%}")
print(f"All probabilities: {probabilities}")

Regenerate Visualizations

python tools/make_visuals.py

This script:

• ✅ Validates dataset structure
• ✅ Introspects actual model architecture
• ✅ Computes shapes and parameters from live model
• ✅ Generates JPEGs + PNGs + JSON metadata
• ✅ Ensures consistency between diagrams and code

---

📈 Performance

Performance at a Glance (EDTH Munich Validation Set)

Metric	Value	Source
Overall Accuracy	97.22%	evaluation_summary.txt
Macro F1-Score	0.9723	Computed from per-class F1 scores
Model Size	7.9 MB	FP32 weights
Parameters	2,080,323	From model introspection
Inference (GPU)	~65 ms	Average across validation set
Inference (CPU)	~85-100 ms	Estimated

Per-Class Performance (Validation Set)

Class	Precision	Recall	F1-Score	Support
Background	0.9333	0.9767	0.9545	86
Drone	1.0000	0.9515	0.9751	103
Helicopter	0.9800	0.9899	0.9849	99
Weighted Avg	0.9732	0.9722	0.9723	288

Key Achievements:

• ✅ Near-perfect drone precision (100%)
• ✅ Excellent helicopter detection (98.99% recall)
• ✅ Balanced performance across all classes
• ✅ Production-ready accuracy (>97%)

---

📁 Repository Structure

acoustic-drone-detector/
├── data/
│   └── edth_munich_dataset/
│       └── data/
│           ├── train/  (180 drone, 180 heli, 180 bg)
│           └── val/    (60 each class)
├── models/
│   └── crnn_combined/
│       ├── crnn_final.pt
│       └── labels.json
├── visualizations/
│   ├── 01_preprocessing_flowchart.jpg
│   ├── 02_crnn_architecture.jpg
│   ├── 03_training_pipeline.jpg
│   ├── 04_complete_system_flowchart.jpg
│   └── *.meta.json (metadata sidecars)
├── tools/
│   └── make_visuals.py  (regenerate all visuals)
├── src/adrone/
│   ├── models/acoustic_models.py
│   └── preprocessing/
├── advanced_preprocessing.py
├── sota_inference.py
├── train_sota_model.py
└── README.md

---

📚 References

• Dataset: EDTH Munich Acoustic Drone Detection Dataset
• Architecture: CRNN with Temporal-Frequency Attention
• Preprocessing: Librosa audio processing library
• Framework: PyTorch 2.0+

---

📄 License

MIT License - See LICENSE file

---

🙏 Acknowledgments

• EDTH Munich Dataset providers
• Librosa audio processing library
• PyTorch deep learning framework

---

Last Updated: October 25, 2025
Model Version: CRNN v1.0 (2.08M parameters)
Visualizations: Auto-generated from actual model using tools/make_visuals.py

• Sharp, distinct lines at rotor fundamental and harmonics
• High spectral contrast (peaks and valleys)
• Relatively narrow bandwidth

MFCC Characteristics:

• Strong periodic patterns
• Low MFCC coefficients (1-5) show rotor fundamental
• Higher coefficients capture motor/propeller interaction

Temporal Dynamics:

• Relatively steady-state (hovering)
• Some modulation from flight maneuvers
• Fast-changing harmonics during acceleration

Example Waveform:

Amplitude pattern: ~~~~~~~~~~~  (high-frequency oscillations)
Envelope:          ___________  (relatively constant)

---

🚁 Helicopter

Frequency Characteristics:

• Dominant Frequencies: 50 - 800 Hz (lower than drones)
• Main Rotor Frequency: 5-20 Hz (large blades, slower rotation)
• Tail Rotor Frequency: 40-80 Hz
• Blade Pass Frequency: Multiple of rotor speed × number of blades
• Low-frequency dominance: More energy below 1 kHz

Spectral Pattern:

• Strong low-frequency components
• "Thump-thump" pattern in spectrogram
• Broader spectral spread than drones
• Complex harmonic structure (main + tail rotor interaction)

MFCC Characteristics:

• Lower frequency content reflected in MFCCs
• Strong energy in first few coefficients
• Rhythmic, periodic patterns
• Delta features show pronounced modulation

Temporal Dynamics:

• Pronounced amplitude modulation (blade passage)
• Rhythmic pattern more visible in waveform
• Slower temporal variations

Example Waveform:

Amplitude pattern: ~~-~~-~~-~~  (rhythmic modulation)
Envelope:          ^^^^^^^^^^^^  (periodic amplitude changes)

---

🌆 Background (Urban/Ambient)

Frequency Characteristics:

• Broad spectrum: Energy distributed across entire frequency range
• No dominant harmonics: Lacks periodic structure
• Variable content: Depends on environment (traffic, wind, people, etc.)
• Generally low-frequency bias: Most environmental sounds < 2 kHz

Spectral Pattern:

• Non-periodic, stochastic structure
• No clear harmonic lines
• Smooth spectral envelope (less contrast)
• Higher spectral bandwidth
• Often transient events (doors, footsteps, cars passing)

MFCC Characteristics:

• Irregular, non-periodic patterns
• More variation across time
• Less structured than drone/helicopter
• Delta features show random fluctuations

Temporal Dynamics:

• Highly variable
• Non-stationary (changes over time)
• Transient events (sudden bursts)
• No periodic modulation

Example Waveform:

Amplitude pattern: ~-~~-~^~~-~  (random, irregular)
Envelope:          -^-^--^-^---  (unpredictable variations)

---

Discriminative Features for Classification

Feature	Drone	Helicopter	Background
Frequency Range	500-3000 Hz	50-800 Hz	Broadband
Harmonics	Sharp, distinct	Multiple (main+tail)	None/weak
Periodicity	High (motor RPM)	High (rotor RPM)	Low/none
Spectral Contrast	High	Medium	Low
Temporal Regularity	Steady	Rhythmic	Irregular
Spectral Rolloff	Higher	Lower	Variable
MFCC Pattern	Periodic	Periodic (slower)	Stochastic

How the Model Distinguishes Classes

1. Mel Spectrogram (Channel 1):

   • Identifies frequency range and harmonic structure
   • Drones: High-frequency harmonics
   • Helicopters: Low-frequency rhythmic patterns
   • Background: Broadband, non-periodic

2. MFCC (Channel 2):

   • Captures timbral signature
   • Distinguishes engine/motor characteristics
   • Temporal dynamics via delta features

3. Spectral Features (Channel 3):

   • Spectral contrast: High for drones/helicopters, low for background
   • Spectral rolloff: Quantifies frequency distribution
   • Bandwidth: Narrow for rotorcraft, wide for background

4. Attention Mechanism:

   • Learns to focus on discriminative time-frequency regions
   • For drones: Attends to high-frequency harmonics (1-3 kHz)
   • For helicopters: Attends to low-frequency rotor patterns (50-500 Hz)
   • For background: Learns to ignore non-periodic noise

5. BiGRU Temporal Modeling:

   • Captures periodic patterns in drones/helicopters
   • Distinguishes steady-state (drone hover) from rhythmic (helicopter blade passage)
   • Learns that background lacks long-term temporal structure

---

📈 Performance Metrics

Expected Performance (on EDTH Munich Dataset)

Metric	Value
Overall Accuracy	85-90%
Macro F1-Score	0.83-0.88
Inference Time (GPU)	10-20 ms
Inference Time (CPU)	50-100 ms
Model Size	~15 MB
Total Parameters	4,058,307

Per-Class Performance (Typical)

Class	Precision	Recall	F1-Score
Background	0.88	0.90	0.89
Drone	0.86	0.84	0.85
Helicopter	0.84	0.86	0.85

Confusion Matrix (Example)

                Predicted
              BG   DR   HE
Actual  BG   [90   5   5 ]
        DR   [ 8  84   8 ]
        HE   [ 6   8  86 ]

---

🚀 Usage

Installation

# Clone repository
git clone https://github.com/yourusername/acoustic-drone-detector.git
cd acoustic-drone-detector

# Install dependencies
pip install -r requirements.txt

Training the Model

from acoustic_dataset import create_data_loaders
from src.adrone.models.acoustic_models import CRNNWithAttention
import torch

# Create data loaders
train_loader, val_loader, preprocessor = create_data_loaders(
    train_dir="data/edth_munich_dataset/data/train",
    val_dir="data/edth_munich_dataset/data/val",
    batch_size=32,
    use_weighted_sampling=True,
    augment_train=True
)

# Initialize model
model = CRNNWithAttention(
    num_classes=3,
    input_channels=3,
    n_mels=128,
    dropout=0.3
)

# Train (simplified)
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4, weight_decay=1e-4)
criterion = torch.nn.CrossEntropyLoss(weight=train_loader.dataset.get_class_weights())

# Training loop
for epoch in range(50):
    train_one_epoch(model, train_loader, optimizer, criterion)
    validate(model, val_loader)

Inference on New Audio

from advanced_preprocessing import AudioPreprocessor
import torch

# Load model
model = CRNNWithAttention(num_classes=3)
model.load_state_dict(torch.load('best_model.pt'))
model.eval()

# Preprocess audio
preprocessor = AudioPreprocessor()
features = preprocessor.extract_combined_features('path/to/audio.wav')
features = torch.from_numpy(features).unsqueeze(0).float()  # Add batch dimension

# Predict
with torch.no_grad():
    output = model(features)
    probabilities = torch.softmax(output, dim=1)
    predicted_class = torch.argmax(probabilities, dim=1)

# Results
classes = ['background', 'drone', 'helicopter']
print(f"Predicted: {classes[predicted_class]}")
print(f"Confidence: {probabilities[0][predicted_class]:.2%}")

---

📝 Citation

If you use this work, please cite:

@software{acoustic_drone_detector_2025,
  title={Acoustic Drone Detection using CRNN with Attention},
  author={Your Name},
  year={2025},
  url={https://github.com/yourusername/acoustic-drone-detector}
}

---

📄 License

This project is licensed under the MIT License.

---

🙏 Acknowledgments

• EDTH Munich Dataset providers
• Librosa library for audio processing
• PyTorch framework
• Inspired by research in acoustic event detection and audio classification

---

Last Updated: October 25, 2025