Machine Learning for Diverse and Adaptive Waveform Generation

Abstract

Multiple-input multiple-output (MIMO) radar systems offer high flexibility in the design of transmit waveforms, something that can be leveraged for several different radar objectives. This thesis focuses on using neural-network based approaches for designing waveforms that suppress unwanted signal returns (clutter), while maintaining strong target returns. The waveform design problem is formulated around shaping the range-angle ambiguity function, where the objective is to maximize signal-to-clutter-noise ratio (SCNR) under a soft constraint on the bandwidth of the phase-coded waveforms. Unlike conventional optimization-based methods of waveform design, the proposed approach aims to train neural networks capable of generalizing to arbitrary, previously unseen clutter environments at test-time. Several neural network architectures are explored throughout this thesis, namely Convolutional Autoencoders, Vision Transformers, Residual Networks, and Diffusionbased models. The models are trained and evaluated on synthetically generated clutter maps, with varying waveform dimensionalities and clutter distributions. The performance is evaluated using held-out test sets of clutter maps, with corresponding optimization-based waveforms as a benchmark. Additionally, methods for generating diverse sets of waveforms for the same clutter scenario are explored, motivated by the non-convexity of the waveform design objective as well as the potential mitigation against adversarial identification and counter-measure techniques. The results demonstrate that neural networks can achieve performance comparable to the provided optimization-based benchmark on previously unseen clutter scenarios, particularly for clutter environments with fewer and larger regions of clutter. Models trained directly using differentiable waveform performance objectives are shown to significantly outperform a supervised training approach. Moreover, the results show that diverse waveform generation is achievable, although there is an indication of a tradeoff between achieved diversity and average waveform performance, and the effective diversity of the generated waveforms is shown to be significantly dependent on the method of generating multiple outputs. Overall, the findings indicate that machine learning and neural networks offers a promising strategy for adaptive MIMO radar waveform synthesis, while also highlighting some important limitations and challenges, with room for future research.