Learning-Aided Distributed Estimation and Control for Networked Vehicular Systems

现代网络化车辆系统正在利用连通性的进步来推动地球仪主要市场的创新。例如，智能交通将车辆和基础设施连接起来，从而大幅提高燃油效率（高达22%）和乘客安全（冬季事故减少25%），是世界上增长最快的行业之一。然而，在这些复杂的互连系统中，用于数据融合、估计和控制的传统集中式架构（在计算上）效率极低。集中式系统的灵活性和模块化程度有限，使网络容易受到故障和干扰的影响，而这些故障和干扰可能仅从一个故障点就危及整个系统。基于子控制数据系统构建的新型网络控制系统通过通信网络交换信息，并利用分布式机器学习增强算法，为这些挑战提供了有前途的解决方案。 为日益流行的网络物理系统、自动驾驶系统（ADS）和协作车辆启用快速和可重构机制的潜力，强调了开发更可靠的分布式系统设计的迫切需要。由于此类网络系统的复杂性日益增加，当前基于模型的分布式控制方法正在达到其性能极限，从而影响模型的决策预测能力以及系统对意外事件和通信干扰的恢复能力。因此，拟议的研究计划将推进一个新的学习辅助分布式估计和控制平台的网络车辆系统使用实验数据，描述了每个子系统的主要属性，通过宽带通信（高数据速率）提供这些网络节点。首要的长期目标是为网络化车辆系统开发一种新的控制和诊断模式，通过控制和学习算法的协同设计提高可靠性和性能。在接下来的五年里，该团队将通过结合基于模型和学习的结构，利用5G NR等新无线电接入技术提供的更低延迟和更高数据速率，解决分布式系统在计算效率和可靠性方面的核心挑战。两个综合目标将追求：1）学习辅助分布式估计网络系统;和2）分布式学习控制算法的发展。其结果将是一个可扩展的和弹性的分布式框架连接ADS和智能交通的设计，而不需要确切的全球系统模型是已知的子系统节点提供潜在的突破，在分布式系统的学习和控制能力。该团队还将为加拿大智能交通行业培养下一代创新者。

Modern networked vehicular systems are leveraging advances in connectivity to drive innovation in major markets across the globe. Intelligent transportation, for example, which connects vehicles and infrastructure to enable dramatic improvements in fuel efficiency (up to 22%) and passenger safety (25% fewer accidents in winter conditions), is one of the world's fastest growing industries. However, traditional centralized architectures for data fusion, estimation, and control in these complex, interconnected systems are prohibitively inefficient (computationally). Centralized systems have limited flexibility and modularity, rendering the network susceptible to faults and disturbances that could imperil the entire system from just a single point of failure. New networked control systems built upon sub-control data systems that exchange information through a communication network and that leverage distributed, machine learning-enhanced algorithms present a promising solution to these challenges. The potential to enable fast and reconfigurable mechanisms for increasingly prevalent cyber-physical systems, Automated Driving Systems (ADS), and cooperative vehicles, underscores the critical need to develop more reliable distributed-system designs. Current model-based distributed control approaches are reaching their performance limits due to the growing complexity of such networked systems, therefore impacting the model's predictive capacity for decision-making and the resilience of the system to unexpected events and communication disturbances. Therefore, the proposed research program will advance a new learning-aided distributed estimation and control platform for networked vehicular systems using experimental data that describes the main properties of each subsystem, provided through broadband communication (with high data rates) across these networked nodes. The overarching, long-term goal is to develop a new control and diagnosis paradigm for networked vehicular systems, enabling increased reliability and performance through co-design of control and learning algorithms. Through the next five years, the team will address core challenges to rendering distributed systems more computationally efficient and reliable by combining model- and learning-based structures, taking advantage of the lower latency and higher data rates provided by new radio access technologies such as 5G NR. Two integrated objectives will be pursued: 1) Learning-aided distributed estimation in networked systems; and 2) Development of distributed learning control algorithms. The result will be design of a scalable and resilient distributed framework for connected ADS and intelligent transportation without requiring the exact global system model to be known to the subsystem nodes-offering potential breakthroughs in the distributed system's learning and control capacity. The team will also train the next generation of innovators for Canada's intelligent transportation industry.