Mismatched Processing of Composite Waveforms with Small Neural Networks

Linear Frequency-Modulated (LFM) waveforms are commonly utilized in radar systems, however, they are also vulnerable to replication by modern jammers. Composite waveforms, such as a primary (LFM) waveform with a secondary noise component, offer the prospect of designing low-probability-of-intercept systems but can also lead to elevated sidelobe levels and lower Doppler tolerance. In this paper, we explore receiver processing of composite waveforms with a trained neural network (NN) as an alternative to the traditional matched-filter. It is shown that fully connected feedforwarding NNs can be trained to extract the primary waveform's characteristics while minimizing the side effects caused by the secondary waveform. The development of NN structures to create novel types of mismatched filters is thus demonstrated. Through simulations, including constant false alarm rate (CFAR) tests, we substantiate the capabilities attainable through neural network-based mismatched filtering.