Data-Driven Robust Control Systems for Sustainability

点击翻译按钮获取中文摘要

As we are entering the era of ubiquitous artificial intelligence and automation, the sustainability of engineered systems can greatly benefit from the development of machine-learning-based adaptive robust control and optimization. The goal is to increase system automation levels towards autonomous operation while maintaining safe operation at scale and preserving the environment for future generations. While AI, or more precisely machine learning, has recently enabled incredible performance enhancement in many fields such as image recognition and natural language processing, its inherent adaptive black-box architecture has made it difficult to use in safety-critical engineering applications such as aircraft flight control or autonomous vehicles with respect to the explainability of its decision-making and its unpredictable interaction effects with the system dynamics in closed loop. Thus, the certification of such machine-learning-based safety-critical systems is currently very difficult if not outright impossible in, e.g., aerospace applications. Yet, machine learning offers a path to the autonomous future that all are preparing for. The main goal of the proposed engineering research program is thus to develop innovative, data-driven, provably robust and sustainable autonomous control systems. The applications range from electric vehicles (EV), electric autonomous vehicles (EAV), renewable energy management systems (EMS), building heating, ventilation and air conditioning (HVAC) systems, and electrified industrial processes. Classical control and optimization techniques have been very successful at increasing the efficiency and performance of individual systems such as production machines, industrial robots, ground vehicles, aircraft and industrial processes. However, such control systems may not adapt well to wildly variable conditions and rapidly changing dynamics. For instance, in the current context of fast development of automated passenger and transport vehicles, the changing conditions under which such autonomous vehicles will be operating far surpasses the ability of regular control systems based on robust, adaptive and model predictive control methods to keep the vehicle in a stable, safe zone of operation at all times. On the other hand, the ubiquity of connected autonomous cyberphysical systems such as AV offers an enormous amount of fresh data from operating peer systems that can be used to improve the behavior of the local system and allow it to respond to the demand of higher efficiency objectives of the network, such as a fleet of robo-taxis. Machine learning techniques have proven apt at capturing approximate models of fast changing dynamical environments based on real-time data, e.g., in autonomous driving. Thus, we propose to investigate the merging of machine learning and feedback control techniques to get the best of both worlds, that is, provenly-robust high-performance control systems that adapt to changing conditions.