For industrial automation systems, dynamic control algorithms should possess the following characteristics: real-time (computation time can be as short as milliseconds or even lower), reliable (always maintaining consistent and effective control of the system, ensuring the stability of the control system), robust (under disturbances or uncertainties, the output deviation is within a given range, and cannot differ significantly or produce completely opposite results), deterministic (outputting deterministic commands to make the system follow the reference input as accurately as possible, rather than making probabilistic judgments), and interpretable (associating with the real controlled object). Based on whether a dynamic model of the controlled object is introduced during the control implementation process, control methods can be divided into two categories: model-based control and data-driven control, as shown in Figures 2(a) and (b), respectively. Model-based control, under reasonable modeling assumptions, first establishes a model of the controlled object to describe its dynamic characteristics, and uses this model as the center to complete controller design, parameter tuning, performance analysis, and real-time computation. Data-driven control, on the other hand, starts directly from the available data of the system, and through the effective organization and processing of the data (represented by a data model), completes the entire control process, including design, analysis, and real-time computation.
Figure 1 Components of an automated system
Figure 2 Implementation flow of different control methods: Model-based control
Model-based control is rooted in the idea that since it involves dynamic control of the controlled object, if the dynamic behavior of the controlled object can be accurately known, a controller can be designed to provide the correct control commands. If the dynamic behavior of the controlled object can be accurately described by a mathematical model, that is, if the system represented by the mathematical model is equivalent to the actual controlled object, then the theoretical control performance obtained based on the model is consistent with the actual performance of the controller. Therefore, as shown in Figure 2(a), the first step in the design of a model-based controller is to establish a model of the controlled object. The final control performance analysis and actual calculations are also based on the model, which can be obtained through identification or derived from the mechanism of action (physical and chemical theorems and laws).
Ideally, if the modeling is accurate and the parameters are precise, the model can correctly reflect the changes of the controlled system under various stimuli/conditions. This allows the real-time output value of the controlled system to be obtained through the model's calculations, replacing the role of sensors. This is crucial for situations where installing sensors is inconvenient or the cost of sensing measurement systems is high. The mechanistic model automatically ensures a structured and interpretable relationship between the input, output, and internal state variables of the controlled system (this relationship is automatically contained within various universal theorems and laws derived from numerous experimental and theoretical discoveries). This is precisely what deep learning, utilizing various deep neural networks, aims to achieve (compared to "shallow learning" methods such as support vector machines, deep learning, within a multi-layered network architecture, processes raw data layer by layer, gradually combining low-level features to form more abstract high-level categories or features). By fully utilizing the information provided by the model, optimal control of the controlled object can be achieved. Since the 1960s, significant progress has been made in various model-based control methods, including optimal control, model predictive control (MPC), indirect adaptive control, and feedback linearization and back-stepping control for nonlinear systems. These methods have demonstrated tremendous power in practical applications. For example, model predictive control (which predicts the future dynamic changes of the controlled system based on a mathematical model and calculates the optimal control input command based on this prediction and current constraints) has successfully controlled the humanoid robot Atlas (@BostonDynamics) to gracefully perform running, jumping, backflips, and high-difficulty gymnastic movements (in which dozens of joints and corresponding actuators throughout the body need to synchronize and precisely complete the prescribed movements).
Because of its high dependence on the model, the accuracy of the model determines the performance of the model-based control system. To simplify the model and facilitate controller design, as shown in Figure 2(a), reasonable assumptions are usually made during modeling, and unmodeled dynamics inevitably exist. Coupled with unknown external disturbances or noise, the actual dynamic behavior of the controlled object deviates from the dynamic behavior represented by the model, rendering the equivalence principle inapplicable and leading to performance degradation or even system instability in the model-based control system. In this case, under the premise of making assumptions about the unmodeled dynamics or disturbances, appropriate robust design can be used to ensure the control performance of the control system.
When the controlled object cannot be accurately modeled (e.g., stochastic or uncertain systems) or is a highly complex system (the complexity of the model used to equivalently describe its dynamic behavior also increases), model-based control methods face challenges such as complex control algorithm structures and the resulting increased difficulty in design, analysis, and real-time computation, sometimes even making it impossible to design an effective model-based controller. On the other hand, as the digitalization of automation systems continues to increase, the amount of available data also increases. These two aspects pave the way for the development and application of data-driven control.
Data-driven control
Literally, data-driven control encompasses both "data-driven" and "control," meaning it utilizes data-driven methods to accomplish control tasks. From a data-driven perspective, a common misconception is that it directly equates to machine learning and neural networks. However, machine learning or neural networks do not represent the full scope of data-driven control. Data-driven control can be understood as using a specific framework or structure to organize and extract relevant information to complete a specific task. In this sense, neural networks, decision trees, and support vector machines can all be understood as a type of "specific framework/structure." For industrial automation control applications, other methods of organizing and processing data certainly exist. From a control perspective, as a type of dynamic control method, data-driven control should meet the general requirements of industrial automation systems for dynamic control algorithms. After synthesizing different descriptions of data-driven control, the references provide a rigorous definition of data-driven control.
[Definition of Data-Driven Control]: Data-driven control encompasses all control theories and methods that design controllers by directly utilizing (online/offline input/output data of the controlled system or knowledge obtained from data processing) rather than by explicitly utilizing (information obtained from the mathematical model of the controlled object); the stability, convergence, and robustness of these theories and methods can be guaranteed through rigorous mathematical analysis under reasonable assumptions.
As shown in Figure 2(a, b), compared with model-based control methods, data-driven control methods start directly from available system data (which includes information such as system structural changes, unmodeled dynamics, and unknown disturbances), thus eliminating the dependence on the controlled object model. This improves the robustness of the control system and can effectively handle control problems of controlled objects that are difficult to model. It is worth noting that although data-driven methods such as machine learning and deep neural networks have been successfully applied to different stages such as external environment perception, reasoning and decision-making, fault diagnosis, and prediction (see Figure 1), their application and theoretical analysis in dynamic control still need further development. For reference, the definition of Machine Learning Control is given here.
[Definition of Machine Learning Control]: Machine learning control is a concept that uses machine learning algorithms to learn an effective control law. This concept is used to address situations where modeling is difficult or impossible (for systems handling complex control tasks). In this definition, machine learning control also addresses situations where model-based control methods cannot be applied, using an existing effective control law as the learning object or content. However, the key question is precisely how to design this effective control law to ensure the normal operation of the system and provide sufficient data for machine learning when facing complex control tasks.
The references provide the following example: In a controlled/training environment, a motion capture system accurately determines the drone's position, and based on this position information, a model predictive controller (MMC) accurately controls the drone to avoid obstacles. Reinforcement learning is then used to train the MMC. In a testing environment, the motion capture system is removed (only onboard sensors are used), and a pre-trained deep neural network is employed to control the drone's movement. This effectively utilizes the powerful fitting and generalization capabilities of deep neural networks, enabling the control system to effectively control even unknown new situations. According to the references' definition of data-driven control, although the deep neural network, as a controller, uses offline data for training, this training data originates from MMC and still explicitly uses the controlled object's model information, thus not strictly falling under the category of data-driven control.
In fact, in fields where deep learning has made significant progress (such as computer vision and natural language processing), the targeted systems already operate independently, and there is a large amount of relatively easy-to-obtain "field" data. For industrial automation systems, however, effectively and dynamically controlling the controlled object to ensure its smooth operation is already a challenging task. Furthermore, the data is mostly small-batch, single-point data, and in some scenarios, it's even impossible to install sensors to acquire data. How to provide sufficient data to ensure the training effect of deep learning for dynamic control is another major challenge. Therefore, how to combine various machine/deep learning algorithms with the specific needs of dynamic control is an open question that requires continuous exploration and research. Considering the characteristics of industrial automation systems, data-driven control methods suitable for industrial automation applications should be developed. In addition to the PID, ILC, MFAC, Unfalsified Control, Lazy Learning, and Iterative Feedback Tuning methods already introduced in previous articles, other data-driven control methods have also made significant research progress and have been practically applied.
Figure 3. Summary of Control Methods
Summarize
Returning to the initial question, when selecting a control method, it's crucial to recognize that no single control method is suitable for all applications, and the advantages and disadvantages of various control methods are relative to the specific application scenario. There should be no bias towards any particular control method; the design of the control algorithm should be based on a deep understanding and grasp of the controlled object and the application scenario. As shown in Figure 3(a), the references provide a method for selecting a control method based on the modeling of the controlled object: for controlled objects with accurate models, model-based control methods should be prioritized (considering that this method can achieve optimal control of the controlled object); for controlled object models that are not very accurate and have uncertainties, various robust and adaptive controls, as well as data-driven control methods, can be used; and for situations where model-based control is insufficient, data-driven control methods should be chosen.
From another perspective, to achieve the best control effect, the two control methods can also be combined based on their respective advantages to form various hybrid control methods. For example, in model-based control methods, data-driven methods are used to handle parts that are difficult to model or unknown disturbances (utilizing the fitting ability of neural networks) to improve robustness or adjust the model-based controller online; in data-driven control methods, mechanistic models are used to provide useful information and data to the data-driven controller, or to correct the control commands given by the data-driven controller, and so on.
In summary, model-based and data-driven control methods are not mutually exclusive "or" relationships, but rather complementary "and" relationships. The combination of the two constitutes the entire dynamic control method system, as shown in Figure 3(b).
To address the digitalization and intelligentization requirements of a user's automated cylinder machining production line, this paper designs the master control system for the automated machining production line based on the Siemens S7-1500 series PLC controller and WINCC host software platform. Through production debugging and commissioning of the machining production line, the functional requirements of the automated machining production line are finally realized.
Author Introduction
Li Lei holds a PhD in Mechanical Engineering from Zhejiang University and was a visiting scholar at Georgia Institute of Technology (2016-2017). He currently works on the research and development of automation control algorithms. During his doctoral studies, he published several articles in journals such as IEEE TMech and TIE. He currently serves as a reviewer for international mechanical engineering journals such as TMech and IJIRA (International Journal of Intelligent Robotics and Applications).
