有问必答

             ——“有问必答”栏目组

  数据驱动的控制也许不是一个崭新的控制领域,因为早在20世纪40年代Ziegler-Nichols提出的PID调节方法也是一种基于数据的控制方法。随着科学技术、特别是信息科学技术的快速发展,许多生产工艺、设备和过程越来越复杂,依据物理化学机理建立精确数学模型,并对生产过程和设备进行优化和控制已变得越来越困难。另一方面,随着传感和网络技术的快速发展,人们可以得到大量关于复杂系统的实时数据,为基于数据驱动的控制方法提供了必要的基础,也使该领域的研究得到越来越广泛的重视。本期栏目组有幸采访世界著名控制学者,美国得克萨斯大学Arlington分校的Frank Lewis教授,请他谈谈数据驱动的控制与优化的发展现状和未来。Lewis教授是IEEE Fellow、IFAC Fellow,也是东北大学的千人计划教授。他在最优控制、神经网络、自学习系统、多智能体系统等领域都做出过重要贡献。非常感谢他在百忙中接受我们的采访。在准备本期的采访中,我们有幸得到北京交通大学侯忠生教授的大力支持。他为本期的采访提出了许多宝贵的建议,对此栏目组表示衷心的感谢。以下是Lewis教授的访谈录。

Question: Professor Lewis, thank you very much for taking your time to attend our interview amid your very busy schedule. Could you first introduce data-driven control (DDC) and the driving forces behind its emergence and development?

Answer: In modern industrial processes, aerospace systems, vehicle systems, and elsewhere there is no model available of the process, or the process model is too complex to be tractable for controller design. Modeling and system identification are expensive and time-consuming, and models may be time-varying, or nonlinear, or contain delays. However, huge amounts of measured process data are available, both in the form of stored historical data from prior measurements and on-line data available in real time during process runs.

The term 'data-driven control' (DDC), or data-based control, originated in the 1990's in the Information Technology Computer Science and Computational Intelligence fields. It shares the same context as 'big data', 'data mining', and 'data fusion'. The development of DDC was driven by the huge amounts of data measured in complex process control systems, both stored historical data from prior measurements and on-line data available in real time during process runs. In DDC, the intent is to efficiently use the information in huge amounts of process input/output data to directly design controllers that provide guaranteed performance of the process. Ideally, a model of the process is not identified as a preliminary step. Given the process industry's interest in DDC, its relevance to the objectives of system and control theory, and the fact that many techniques developed over many years for automatic feedback control have similar aims, it is natural that the Control Systems Community should have taken an interest in this terminology.

DDC fits into the broader context of data-driven process monitoring, DD prediction, DDC, and DD fault diagnosis. In retrospect, DDC can be used to categorize some existing techniques in feedback control. Moreover, since the introduction of the idea, DDC has served to inspire the development of many new techniques that directly use data to design process controllers. DDC design methods can be classified in several ways. One can distinguish offline DDC design techniques that use stored historical data versus those that use online data measurements made in real time. Alternatively, one can classify DDC methods into techniques that assume a fixed a priori controller structure, and those whose control structure is directly designed using the available measured data.

Question: In your view, what are the relationships between DDC and existing control techniques such as auto-tuning PID control, adaptive control and iterative learning control?

Answer: First of all, some existing automatic feedback control design methods could be called DDC methods. PID control designed, e.g. using the Ziegler-Nichols method, is an offline DDC method with fixed control structure. PID design is turned into an online DDC technique with fixed control structure using the autotuning methodology. Direct adaptive control is a DDC method that has fixed controller structure and uses online data measurements to tune the controller parameters. In the 1990s direct adaptive control structures were extended to nonlinear-in-the-parameters controllers using neural network (NN) approximation methods. Particularly in the discrete-time case where the need for certainty equivalence was removed, these NN adaptive controllers offer significant extensions to adaptive control for nonlinear processes and provide a powerful class of DDC that have rigorous proofs of performance, convergence, and robustness.

Adaptive inverse control offers a variety of techniques for DD controller design based on solving the Weiner-Hopf equation online using measured data by using recursive least-squares, stochastic approximation, etc. to tune controller parameters. These methods effectively compute autocorrelation functions using real-time measured i/o data. The objective is to decorrelate the output tracking error from the prescribed reference trajectory. The newly developed offline DDC method of correlation-based tuning shares the same objective.

Iterative learning control (ILC) can certainly be considered a DDC technique that does not assume a fixed controller structure. ILC is a run-to-run technique that uses measurements of the control input signal and the tracking error signal over a complete process run to determine a control signal for the next complete run that reduces the tracking error. A process model is not required, and ILC has been shown to converge and yield small tracking errors as long as a spectral radius condition holds on the unknown process model. ILC requires the same initial conditions and the same prescribed reference trajectory at each run.

Some newly developed DDC designs that assume fixed controller structures use least-squares techniques to tune the controller parameters based on measured i/o data. The simultaneous perturbation stochastic approximation (SPSA) DDC uses an approximator such as NN as a controller. The approximator parameters are tuned online using a stochastic approximation method that uses a gradient estimated from the measured i/o data. SPSA is an online DDC technique that uses data measured in real time. Offline DDC techniques that use stored historical data and assume a fixed controller structure include iterative feedback tuning (IFT) and virtual reference feedback tuning (VRFT). Both use estimates of gradients computed using measured i/o data and employ least-squares cost functions to tune the controller parameters.

Many direct DDC design techniques that do not assume fixed controller structures are based on statistical analysis of measured data. The successful use of principal component analysis (PCA) in DD prediction and fault diagnosis has been capitalized upon to design various DDC methods that seek to provide direct controller designs. The model-data integrated methods such as subspace-aided DDC use the SVD to design a diagnostic observer that then leads to a DDC design.

The specially developed DDC technique of model-free adaptive control (MFAC) uses dynamic linearization of the process where a nonlinear correction term to the linear model is estimated using a 'pseudo partial derivative' that is computed using measured i/o data. No controller structure is assumed a priori. Proof of performance is provided.

Several companies by now offer design packages for DDC, including notably MathWorks. 

Question: A related topic is data-driven optimization (DDO). Could you give some account on its recent development? 

Answer: In modern day systems, there are increased demands for fuel efficiency, conservation of resources, cost and energy savings, and other optimal performance requirements. Most DDC methods do not address such requirements. Since the 1960s, optimal control has been responsible for the successful design of aircraft autopilots, ship controllers, industrial control systems, aerospace control systems, and more. Yet, optimal control generally uses offline design techniques based ion solving matrix design equations that require knowledge of a process model. Guaranteed robustness properties of optimal controllers mean that the design model need not be exact, but it must be yield responses close to the actual process model. 

The class of Approximate Dynamic Programming (ADP) controllers provide DDO methods that use available process data measured online to design optimal controllers in real time. ADP is based on reinforcement learning techniques that seek to optimize user-prescribed performance measures by modifying feedback policies online based on system response to current suboptimal policies. A technique known as Q-learning provides optimal controller design in real time for systems with completely unknown dynamics. For many years ADP has been well developed for feedback control of discrete-time linear systems. Since 2007 proofs of stability and convergence have been provided for nonlinear systems ADP. Extensions were made to H-infinity control and dynamic games. In 2008 ADP was extended to partially known continuous-time systems using the technique of Integral Reinforcement Learning (IRL). Recently developed off-policy IRL techniques allow DDO for completely unknown linear and nonlinear processes. Extensions have been made to H-infinity control.

The development of DDO has been particularly strong in China, where there are great natural resources and enormous industrial process factories have been installed. Pressures for efficient use of resources and reduction of pollution, costs, and energy usage are especially strong in China and numerous government-sponsored labs have been established, including notably the State Key Laboratory of Synthetical Automation for Process Industries at Northeastern University in Shenyang.

Question: Can you elaborate further MBC, DDC and DDO?

Answer: The recent interest in DDC for control of complex processes has naturally spurred an equal interest in non-data-driven control approaches, one of which has been recently termed model based control (MBC). The question arises, what is 'model-based control'? An argument has been made that MBC includes all controller design techniques that require a model of the process, albeit approximate, or that identify a process model as an intermediate step. This includes optimal control, robust control, indirect adaptive control, and more. An alternative definition of MBC is a control system that explicitly embeds a model of the process in the control algorithm. This includes, for instance, internal model control and model predictive control. 

MBC is a term especially used since the 1990s that could be used to classify in retrospect existing control methods as opposed to DDC and DDO techniques. However, it is interesting that many early uses of the term MBC appeared also within the Computer Science and Computational Intelligence fields. Traditional methods of 'intelligent control design' used in computer science did not generally consider any model of the controlled process. Looking at early block diagrams of such control schemes, one is struck by the fact that there is no block labeled 'plant', 'controlled system', or 'process dynamics'. There is generally a block labeled 'environment' that actually contains both the process dynamics and the external environment that acts on the process. Hence, MBC arose from methods that sought to include some proven modeling in the controller design.  In some computer science usage of MBC, the process model comprises C-code that describes various specifications of input conditions and environment conditions and how they result in various output instances.

Nevertheless, today in the DDC community, the terminology MBC has come to mean 'control design techniques that require a model of the process, albeit approximate, or that identify a process model as an intermediate step'. 

Several companies offer design packages for MBC, including MathWorks and dSpace. These packages generally comprise tools for dynamics description and modeling, system identification, simulation, CAD tools, and visualization tools and often amount to a hybrid of control systems and computer science tools. 

Question: What is the future of data-driven control?

Answer: Given current ever-increasing global demands for efficient use of resources and reduction of pollution, costs, and energy usage, DDC and DDO will become ever more important for successful control in large-scale complex process industries. It will become more and more important to provide DD design techniques that have proven performance, convergence, stability, robustness, and optimality properties MBC has reliable rigorous techniques for providing such guarantees including nonlinear systems analysis and design tools such as Lyapunov theory, passivity, linearization methods, optimal design based on Hamilton-Jacobi equations, dynamic programming, etc. Most DDC methods do not have such tools for rigorous performance guarantees, with the exception of ADP methods. Developing new tools for DDC and DDO stability analysis based directly on measured process i/o data is a challenge.

Thank you very much Prof. Lewis for sharing with our audience your views on data-driven control which would be very valuable for their research in the field.

  控制理论专业委员会 ©2011-2014 版权所有

  中国自动化学会 控制理论专业委员会
  电话:86-10-62551965;Email:tcct@iss.ac.cn