Tuna Şahin

The project combines classical signal processing and deep learning methods. Raw IQ samples are processed using FFT-based spectral analysis and STFT-based time-frequency representations. The FFT pipeline is used for lightweight signal detection, signal parameter extraction, and fast spectral analysis, while STFT spectrograms are used as image-like inputs for CNN-based drone classification. Different neural network architectures were tested for classification, and the system was integrated into a Python-based real-time GUI that displays the spectrum, spectrogram, detected emitters, signal parameters, and classification outputs.

The final system achieved strong performance in both detection and classification tasks while remaining suitable for real-time operation. The FFT-based model reached an overall accuracy of 99.78%, with class-wise F1-scores of 100.00% for noise, 99.68% for drone, and 99.60% for Wi-Fi, showing that the system could reliably distinguish drone activity from common non-drone RF signals. False alarm behavior was also evaluated in an 8-hour crowded Wi-Fi environment: although the system produced internal false positives, the Temporal Consistency Filtering mechanism suppressed almost all of them, resulting in only 0.1 false positive alerts per hour and 0 critical false negatives. Timing tests showed average end-to-end latencies of 127.43 ms for detection, 1818.8 ms for classification, and 2849 ms for multiple detection, satisfying the real-time requirements for the main detection and classification pipeline. The STFT-based models also performed strongly in GUI-integrated field tests, with ResNet-based models reaching around 98.8–99.0% field-test accuracy, while lightweight or transformer-based alternatives showed that deployment performance depends not only on offline accuracy but also on stability, latency, and robustness under real RF conditions. Overall, the project demonstrated that a hybrid FFT/STFT pipeline can provide accurate drone detection, low false alarm output, reliable classification, and practical real-time visualization in crowded wireless environments.

During the development process, the main components were built and tested separately, including the capacitor, spark plug, radiating coil, and coherer. The final prototype was assembled according to the circuit schematic and tested using both measurement equipment and practical demonstrations. The measured induced voltage had a frequency of 139 kHz, which was close to the theoretical value of 150 kHz, giving an error of about 7.3%. Additional tests with a calculator and a coherer showed that the device was able to create observable electromagnetic interference, confirming that the EMP generator operated as intended.

The antenna was implemented and optimized in CST, where the slot offsets were tuned while preserving the excitation ratios determined by the Taylor distribution. The impedance matching was handled through CST optimization rather than being calculated manually. The final radiation pattern shows a main lobe magnitude of about 13.1 dB, an angular width of approximately 18 degrees, and a sidelobe level around −35.4 dB, meaning the design successfully meets the sidelobe requirement. The final S₁₁ result also indicates that the antenna is matched around the intended operating frequency of 10 GHz. Overall, the project demonstrates that a properly tapered slotted waveguide array can provide high gain, directional radiation, and low sidelobe performance at X-band.