This project demonstrates the classification of drone signals using a VGG16-based deep learning model. Signals were captured from drone controllers such as Elsec, Mavic Classic, Mavic Pro, and Phantom Pro. The raw signals were preprocessed using MATLAB, where spectrograms were generated and noise was added using the awgn function at 15dB. These spectrograms were used to train the VGG16 model, achieving 100% accuracy.
This folder contains 4 subfolders for different drone controllers: Elsec, Mavic Classic, Mavic Pro, and Phantom Pro. Each subfolder contains 801 spectrogram images of size 224x224.
- The
awgnfunction in MATLAB was used to add noise to the signals. - Noise was added at 15dB to the signals.
This folder contains the following results from the trained model:
- Confusion Matrix: Displays the confusion matrix for testing results.
- Precision and Recall Scores: Provides precision and recall metrics for model performance.
- Training Progress: Tracks data from the model training process.
- Test Accuracy: Contains accuracy values from testing the model with the test dataset.
- Accuracy Achieved: 100%
-
generateSpectrograms.m
MATLAB script to generate spectrogram images from raw IQ data. The input is a.txtfile containing binary IQ values, where the least significant 16 bits (LSB) represent the Q value and the most significant 16 bits (MSB) represent the I value. -
Drone_Classification_With_Noise_15db.ipynb
This Jupyter Notebook contains the VGG16 model definition, training, and testing procedures.
VGG16_Drone_Model.h5
The trained model file, which can be loaded at any time using theload_modelfunction for further testing or evaluation.
- Sampling Frequency (fs): 30.72 MHz
- Drone Signals Used: Elsec, Mavic Pro, Mavic Classic, Phantom Pro
For each signal file, spectrograms were generated using 1M samples.
The Report folder contains the complete project report, documenting the process, results, and findings.