ERI: AI-Enhanced Dynamic Interference Suppression in Cognitive Sensing with Reconfigurable Sparse Arrays

3D认知传感，由可重构稀疏阵列（RSA）促进，通过有效地减少来自不同来源方向的干扰，同时保持相同数量的昂贵射频（RF）前端组件，优于固定阵列配置。RSA通过快速射频交换在选择的有源天线位置之间共享它们，从而实现了昂贵天线组件的减少。使用RSA的认知感知模拟认知过程的感知-行动周期（PAC）。RSA从不同的空间天线位置收集实时数据，感知周围环境，并动态适应天线位置和随后的阵列处理，称为波束形成。鉴于有源天线位置可以根据所需的源和干扰方向以及其他参数而改变，本工作解决了阵列可重构性的三个主要挑战：频繁和复杂的天线切换，计算最佳有源阵列结构，并在快速PAC内实现波束成形。克服这些挑战需要快速迭代优化算法，人工智能（AI）技术，如基于离线深度学习（DL）的神经网络训练，以及简化天线切换标准的实施。RSA设计探讨了两个关键应用：认知无线电（CR）中的频谱感知（SS）和认知雷达感知中的定位和跟踪。通过SS功能，CR可以检测未充分利用的频段，并自动对RF信号进行分类。在CR中集成RSA设计可以增强干扰缓解，提高业务质量、带宽可用性和频谱利用率。通过增强态势感知、监视能力、频谱管理和共存措施，这些优势有利于监管机构和政府实体。虽然CR通常依赖于被动或仅接收感测，但启用rsa的认知雷达系统通过优化发射器和接收器的天线位置来增强性能，尽管采用不同的方法。这一进步将推动最先进的天气监测、军用雷达、自动驾驶汽车雷达、室内人类活动分类、跌倒检测和远程生命体征估计。提出的研究基于开发快速迭代算法，包括DL技术，通过快速智能地从天线位置的均匀网格中选择天线子集来实现动态干扰抑制。目前的算法仅在非常有限的情况下有效：对操作环境的假设先验知识通常是未知的，并且由于优化阵列拓扑的高运行时间，实时可重构性是一个相当大的挑战。提出的研究旨在通过两个新思路来改变当前的SS和信号调制分类范式：(i)在考虑实际信道损伤的同时，在CR中采用多频段RSA作为一种机制来增强多个频段的源ID分类，以及（ii）通过包括机器学习和凸优化的进展来优化RSA用于多频段传感。DL模型的设计期望在一个统一的框架中处理多个感兴趣的频段。该提案将数据依赖技术和先验环境知识集成到深度学习模型中，从而产生具有更高精度的新架构，通过克服特定于数据依赖实现的瓶颈来推进自适应传感。所提出的研究将通过(i)有效地解决高分辨率RSA多输入/多输出（MIMO）雷达公式，以产生非常精确的接收波束模式，对未知干扰和杂波环境具有鲁棒性，以及（ii）通过解决复杂的优化问题来训练DL模型，实现端到端发射器设计，以预测发射天线的位置，从而提高雷达传感的性能。以及用于向目标位置最大化功率的发射波形。由于各种环境因素（如噪声、杂波和干扰信号）限制了雷达感知能力，因此该提案提供了一个有前途的解决方案，因为它涉及将RSA集成到MIMO雷达跟踪中，以增强抗干扰能力。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

3D cognitive sensing, facilitated by a reconfigurable sparse array (RSA), outperforms fixed array configurations by effectively minimizing interference from various source directions while maintaining the same number of expensive radio frequency (RF) front-end components. The RSA achieves this reduction in costly antenna components by sharing them between active antenna locations selected through fast RF switching. Cognitive sensing using RSA emulates the perception-action cycle (PAC) of cognitive processes. RSA gathers real-time data from different spatial antenna locations, perceives surroundings, and dynamically adapts antenna locations and subsequent array processing, known as beamforming. Given that active antenna locations can change depending on desired source and interference directions and other parameters, this work addresses three main challenges specific to array reconfigurability: frequent and complex antenna switching, computing optimum active array structures, and implementing beamforming within the fast PAC. Overcoming these challenges requires fast iterative optimization algorithms, artificial intelligence (AI) techniques such as offline deep learning (DL)-based neural network training, and enforcement of simplified antenna switching criteria. The RSA design is explored for two key applications: spectral sensing (SS) in cognitive radio (CR) and localization and tracking in cognitive radar sensing. Enabled by SS functionality, a CR detects underutilized frequency bands for opportunistic use and autonomously classifies RF signals. Integrating RSA design within CR enhances interference mitigation, improving service quality, bandwidth availability, and spectrum utilization. These advantages benefit regulatory authorities and government entities by enhancing situational awareness, surveillance capabilities, spectrum management, and coexistence measures. While CR typically relies on passive or receive-only sensing, an RSA-enabled cognitive radar system enhances performance by optimizing antenna locations at both the transmitter and receiver, albeit through distinct approaches. This advancement would propel state-of-the-art weather monitoring, military radar, radar for self-driving cars, indoor human activity classification, fall detection, and remote vital sign estimation.The proposed research is predicated on developing fast iterative algorithms, including DL techniques, to enable dynamic interference suppression by swiftly and intelligently selecting subsets of antennas from a uniform grid of antenna locations. Current algorithms are effective only under severely limited scenarios: the assumed prior knowledge of the operating environment is often unknown, and real-time reconfigurability is a considerable challenge due to the high run times of optimizing the array topology. The proposed research aims to transform the current paradigms for SS and signal modulation classification via two novel ideas: (i) the adoption of multi-band RSA in CR as a mechanism to enhance source ID classification across a range of frequency bands while considering realistic channel impairments, and (ii) the optimization of RSA for multi-band sensing by including advances in machine learning and convex optimization. The design of DL models is expected to process multiple frequency bands of interest in a unifying framework. The proposal integrates data-dependent techniques and prior environment knowledge into DL models, resulting in novel architectures with greater accuracy that will advance adaptive sensing by overcoming bottlenecks specific to data-dependent implementation. The proposed research would advance the performance of radar sensing by (i) efficiently solving high-resolution RSA multi-input/multi-output (MIMO) radar formulation for generating exceptionally accurate receive beampatterns that are robust to unknown jamming and clutter environments, and (ii) training DL models through solving complex optimization problems, to realize an end-to-end transmitter design to predict transmit antenna locations, as well as transmit waveforms for maximizing the power towards target locations. Since radar sensing capability is limited due to various environmental factors such as noise, clutter, and jamming signals, this proposal offers a promising solution as it involves integrating RSA into MIMO radar tracking to bolster interference rejection capabilities.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.