Wavelet neural networks for interference suppression in chirp radar
Loading...
Date
2026-10-13
Authors
Σουκαράς, Γεώργιος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Pulse-compression radars in the FR3 band (7.125–24.25 GHz) estimate the delay of a received chirp signal with a matched filter, which loses its ability to lock onto the echo when strong narrow-band interference or continuous-wave jammers share the band. This thesis studies a small trainable wavelet neural network (WNN) that denoises the received record before the matched filter. Its first layer is a bank of 16 complex Morlet wavelets of which only the centre frequency and the width are learned; a batch normalisation, a hidden 1×1 layer with ReLU and a 1×1 output layer then reconstruct the clean chirp. The network has 369 trainable parameters and works at full time resolution. Chirplet and band-pass sinc kernels were tested in place of the Morlet wavelets and give similar results.
The denoiser is evaluated in simulation with radar metrics (lock-failure rate, peak-to-sidelobe ratio and its gain) against a convolutional denoiser and a bank of freely trained kernels with exactly the same number of parameters. The WNN keeps the matched filter locked for narrow-band noise up to a jammer-to-signal ratio (JSR) of 30 dB, for a single tone jammer at a random frequency in the band up to 25 dB, and for one to six jammers in a 2 GHz slice up to 40 dB, where the baselines fail in 66 % to 100 % of the records. At moderate interference all denoisers restore lock, and the baselines sometimes reach a higher peak-to-sidelobe gain. Several jammers spread over the whole band remain unsolved.
The trained network folds exactly into eight 64-tap FIR filters. A fixed-point Vitis HLS description of it was synthesised for an FPGA with a clock period of 4.7 ns and without DSP blocks; it is a functional baseline and not yet a real-time design at the radar sample rate.
Description
Keywords
Wavelet neural network, FMCW radar, Matched filter, Interference mitigation, FPGA, High-level synthesis