Control and diagnosis based on learning from data

数据在决策过程中变得越来越重要有几个原因。首先，大多数系统和过程最初都经过了良好的校准，并在良好控制的约束条件下运行。然后可以建立数学模型，并用于正常运行条件下的控制和性能分析。然而，随着自然退化或意外条件，如异常，系统偏离其最初的设计条件，并且这些模型的保真度恶化。开发完整的新模型通常是很费时的。在这种情况下，数据对于性能监控、故障诊断和控制律更新变得非常重要。其次，当系统高度复杂时，数学建模甚至可能无效。在这种情况下，数据驱动的建模几乎成为唯一的手段。总的来说，如何从多种来源的数据中学习已经引起了工业界和研究界的极大关注。基于最近的研究进展，我们将重点关注系统控制和监测中围绕数据利用和开发的几个重要问题：1)在基于模型的控制和故障检测方法中，设计方法对模型不确定性的鲁棒性是最重要的考虑因素之一。关于这个话题已经有了广泛的研究成果。然而，数据驱动方法的鲁棒性问题和不确定性分析并没有像基于模型的方法那样得到解决。数据中的不确定性如何通过所谓的“灰盒或黑盒”模型（通常由机器学习方法呈现）传播值得更好的研究，并且对于测量这些方法的准确性或置信度很重要。2)健康监测系统设计的另一个挑战是知识和数据的融合。最近，随着机器学习和基于神经网络的建模的广泛采用，数据在描述系统行为和估计/预测关键输出方面起着核心作用。然而，过程知识和基于物理的模型不应该被抛弃。如何在数据驱动的方法中利用这些知识是我们感兴趣的。3)为了增强现有控制系统的数据连通性，数据驱动控制的研究近年来日益活跃。在所谓的直接数据驱动控制中，可以直接利用系统的测量数据来计算控制信号，而不需要显式的模型参数，从而省去了模型识别的步骤。虽然线性系统的直接数据驱动控制有了一些新的发展，但对更复杂的系统，如高维和非线性系统，直接从数据设计控制是一个开放的问题。此外，对于自主和安全关键系统，控制系统的容错能力至关重要。因此，数据驱动的容错控制设计是本研究的重点。

There are several reasons why data becomes more essential in decision making. Firstly most systems and processes are initially well calibrated and are operated under well controlled constraints. Mathematical models can then be established and used for control and performance analysis under normal operation conditions. However, with the natural degradation or unexpected conditions, e.g. anomalies, systems deviate from their initially designed conditions, and fidelity of those models deteriorates. It is normally time consuming to develop complete new models. In this case, data becomes important for performance monitoring, fault diagnosis, and updating of the control law. Secondly, when the system is highly complex, mathematical modelling may not even be valid. In this case, data driven modeling becomes almost the only means. Generally speaking, how to learn from multiple sources of data has attracted great attentions from industries and research communities. Based on recent research progress, we will be focused on several important problems centred around data utilization and exploitation in system control and monitoring: 1) In model based control and fault detection methods, robustness of the design approach against model uncertainties is one of the most important considerations. There have been extensive research results on this topic. However, the robustness issue and uncertainties analysis for data driven approaches are not addressed as much as the model based counterpart. How uncertainties in data propagate through the so called "grey- or black-box" models (usually rendered by machine learning methods) deserves better investigation, and is important in measuring the accuracy or confidence of these methods.  2) Another challenge is the fusion of knowledge and data in designing the health monitoring system. Recently, with the wide adoption of machine learning and NN based modelling, data plays a central role in describing the system behavior and estimating/predicting key outputs. However, process knowledge and physics-based models should not be abandoned. How to exploit such knowledge in the data-driven approaches is of interests. 3) To enhance the data connectivity of the existing control systems, research on data driven control has become more active lately. In the so-called direct data driven control, measured data from the system can be directly utilized to compute the control signal without explicit model parameters, which eliminates the model identification step. While there are some new developments on direct data driven control for linear systems, design of control directly from data for more complex systems, such as high dimensional and nonlinear systems, is an open problem. Furthermore, for autonomous and safety critical systems, fault tolerance of the control system is vital. Hence design of data driven fault tolerant control is of particular interests in this proposal.