The widespread use of drones has introduced significant challenges in terms of privacy, airspace management and national security. While radar systems represent a robust so- lution for detecting such threats, several factors related to both the structural and opera- tional characteristics of drones and the physical principles underlying radar measurements make it difficult to distinguish them from birds and other similar targets. A promising approach to overcome these limitations involves analyzing the spectral signatures embed- ded in radar returns, which encode features caused by the motion of specific components of these devices, like the rotation of propellers. This enables more effective discrimination between drones and other targets that do not exhibit such behavior. This thesis investigates the use of Deep Learning approaches to develop a model for automatic classification of spectrograms derived from radar measurements. This work benefited from access to a dataset of real radar measurements, which enabled rigorous val- idation of our models. However, given the limitations of the available real-world dataset, a comprehensive simulation framework was developed to generate realistic synthetic radar data by modeling micro-Doppler effects produced by drone rotor blades, as well as thermal noise and environmental clutter. The performance levels of several architectures, includ- ing convolutional and recurrent neural networks, combined with different preprocessing techniques, were compared on the drone identification task. The analysis then proceeded to evaluate our pipeline on the more challenging task of identifying the specific model. Our results demonstrate that Deep Learning models trained on spectrogram data can effectively learn relevant features for both tasks, with convolutional architectures achiev- ing the highest performance in our experiments. Furthermore, the adoption of carefully designed training strategies that combine real and simulated data has proved to enhance the generalization capabilities of the models. In addition, interpretability was examined by applying a visualization technique to the internal activations of the trained models, offering insights into their decision-making process and highlighting the positive impact of the preprocessing routines we implemented. The final classification pipeline, validated on real measurements, achieved a strong performance with 94.8% accuracy on both drones and drone models identification tasks.
L’uso sempre più diffuso dei droni ha introdotto nuove sfide per quanto riguarda la privacy, la gestione degli spazi aerei e la sicurezza nazionale. Sebbene i sistemi radar rappresentino una soluzione efficace per rilevare tali minacce, diversi fattori, legati sia alle caratteristiche strutturali dei droni, sia alle loro dinamiche di volo tipiche, rendono difficile distinguere questi target da altri bersagli simili, come gli uccelli. Un approccio promettente consiste nell’analisi delle firme spettrali incorporate nei segnali captati dai radar, le quali codificano informazioni aggiuntive indotte dal movimento delle componenti mobili di questi apparati come, ad esempio, la rotazione delle eliche per i droni. Ciò consentirebbe di discriminare più efficacemente i droni da altri bersagli che non presentano tale caratteristica. Questa tesi indaga l’utilizzo di tecniche di Deep Learning al fine di sviluppare modelli per la classificazione automatica di spettrogrammi derivati da rilevazioni radar. Durante lo sviluppo abbiamo potuto accedere a un dataset di misurazioni reali che ci ha permesso di valutare i nostri modelli in condizioni realistiche. Tuttavia, considerate le limitazioni dei dati reali a nostra disposizione, abbiamo deciso di generare dati sintetici sviluppando un simulatore basato sulla modellazione degli effetti micro-Doppler tipici dei droni e di altri tipi di bersagli. Abbiamo quindi proceduto a confrontare le prestazioni ottenute sulla task di rilevamento dei droni per diverse architetture, tra cui reti neurali convolutive e ricorrenti, combinate con alcune tecniche di elaborazione dei dati da noi ideate. L’analisi è poi proseguita testando la nostra pipeline su un’estensione della task precedente, la quale richiedeva l’identificazione del modello specifico di drone. I nostri risultati dimostrano che i modelli addestrati su spettrogrammi sono in grado di imparare a riconoscere caratteristiche rilevanti per entrambe le task. In particolare, l’architettura convolutiva si è rivelata la migliore secondo i nostri esperimenti. Inoltre, l’adozione di strategie di addestramento che combinano dati reali e simulati ha dimostrato di migliorare le capacità di generalizzazione dei modelli. Infine, abbiamo esaminato l’interpretabilità delle pipeline finali attraverso una tecnica di visualizzazione, offrendo un’ulteriore punto di vista sul loro processo decisionale interno ed evidenziando l’impatto positivo delle routine di pre-elaborazione impiegate. La pipeline finale, validata su dati reali, ha dimostrato un’accuratezza del 94.8% sia nella task di identificazione dei droni, sia in quella di discriminazione tra i loro modelli.
Learning from micro-doppler signatures: a deep learning approach for UAV identification
CAPACCI, TOMMASO
2024/2025
Abstract
The widespread use of drones has introduced significant challenges in terms of privacy, airspace management and national security. While radar systems represent a robust so- lution for detecting such threats, several factors related to both the structural and opera- tional characteristics of drones and the physical principles underlying radar measurements make it difficult to distinguish them from birds and other similar targets. A promising approach to overcome these limitations involves analyzing the spectral signatures embed- ded in radar returns, which encode features caused by the motion of specific components of these devices, like the rotation of propellers. This enables more effective discrimination between drones and other targets that do not exhibit such behavior. This thesis investigates the use of Deep Learning approaches to develop a model for automatic classification of spectrograms derived from radar measurements. This work benefited from access to a dataset of real radar measurements, which enabled rigorous val- idation of our models. However, given the limitations of the available real-world dataset, a comprehensive simulation framework was developed to generate realistic synthetic radar data by modeling micro-Doppler effects produced by drone rotor blades, as well as thermal noise and environmental clutter. The performance levels of several architectures, includ- ing convolutional and recurrent neural networks, combined with different preprocessing techniques, were compared on the drone identification task. The analysis then proceeded to evaluate our pipeline on the more challenging task of identifying the specific model. Our results demonstrate that Deep Learning models trained on spectrogram data can effectively learn relevant features for both tasks, with convolutional architectures achiev- ing the highest performance in our experiments. Furthermore, the adoption of carefully designed training strategies that combine real and simulated data has proved to enhance the generalization capabilities of the models. In addition, interpretability was examined by applying a visualization technique to the internal activations of the trained models, offering insights into their decision-making process and highlighting the positive impact of the preprocessing routines we implemented. The final classification pipeline, validated on real measurements, achieved a strong performance with 94.8% accuracy on both drones and drone models identification tasks.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_07_Executive_Summary_Capacci_Tommaso.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
538.97 kB
Formato
Adobe PDF
|
538.97 kB | Adobe PDF | Visualizza/Apri |
|
2025_07_Thesis_Capacci_Tommaso.pdf
non accessibile
Dimensione
5.72 MB
Formato
Adobe PDF
|
5.72 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/240107