Advanced control methods, which significantly enhance the efficiency of production machines, are of vital importance for the Flemish machine construction industry to preserve their leading position within the world economy. New control methods are necessary to satisfy the growing customer demands regarding both flexibility and productivity. Traditional controllers in production machines have important limitations. Firstly, in many cases it is intricate or even impossible for the designers and operators to optimally tune the parameters of a traditional production machine controller due to the complex nature and the vaguely known dynamics of these machines. Furthermore, traditional control algorithms are not able to track changing system parameters and varying environmental conditions, which often appear in practical situations, and will consequently not adapt the control parameters accordingly. These drawbacks of traditional control algorithms, which result in suboptimal efficiency of the controlled machines, can be solved by the introduction of learning behaviour in machine controllers. This will allow machines to automatically learn the optimal control parameters and adapt to variations in both process parameters and environmental conditions.
The realization of such ‘intelligent machines’ was the long-term goal of the LeCoPro project. The project has made a first step in this direction by creating a knowledge platform in Flanders on learning control strategies for production machines. As Flanders has already a lot of expertise on learning techniques, but on different applications, the basic elements for the creation of such knowledge platform were already available at the start. The more concrete objective of the project were therefore to provide the Flemish machine builders with practical methodologies for the design of learning controllers for their production machines. To achieve this objective, two related research tracks were acknowledged: (i) learning control methodologies for complex (sub)systems and (ii) learning control methodologies for decentralized systems.
Three key features of a mechatronic system play an important role in the selection of the learning control strategy:
- the number of subsystems of the global mechatronic system
- the number of inputs and outputs of these subsystems
- the degree of modelability of the controlled system: this modelability can be limited due to uncertainty of system parameters, non-linearity, complex interaction between subsystems, etc.
Fig. 1: 3D-classification of mechatronic systems
Fig. 1 shows a 3D-classification of mechatronic systems, based on these features. The work program of LeCoPro is set up according to this classification. While the research in WP 1 focused on the learning control of an individual mechatronic subsystem, the goal of WP 2 was to develop learning controllers for a global system, consisting of multiple subsystems. Finally, in WP 3, the goal was to practically validate the results of WP 1 and WP 2 and to integrate these results in a combined controller, where on the one hand advanced controllers are applied on the subsystem level and on the other hand a global controller guarantees an optimal global performance.
The goal of WP 1 was the development of model-based as well as non-model-based learning controllers for individual mechatronic subsystems. For these subsystems, nowadays two model-based learning control strategies are mainly applied: Iterative Learning Control (ILC) and Model Predictive Control (MPC). Although the first implementations of learning control strategies appear in the controllers of subsystems in practical machines, there are still many challenges, which require further research, to increase the applicability of these algorithms to a broader range of mechatronic applications. Therefore, the extension of the existing model-based strategies (ILC and MPC) as well as the application of non-model based machine learning techniques (Evolution-based Machine Learning and Reinforcement Learning (RL)) on mechatronic subsystems was the subject of WP 1 (Fig. 2).
Fig. 2: The objective of WP 1
In most of the current mechatronic applications with multiple subsystems, a controller is developed for each subsystem. However, the interactions between these different individual controllers are not taken into account in the controller design. Such a strategy often leads to a suboptimal global performance of the control system, especially for strongly interacting subsystems. Therefore, the goal of this WP was to evaluate the possibilities of global learning controllers for interconnected subsystems, which try to coordinate the local control actions such that a global optimum is obtained. This principle is illustrated in Fig. 3.
Fig. 3: The objective of WP 2
Similar to WP1, model-based as well as non-model-based control techniques were be investigated in this WP. Distributed MPC is a model-based technique, which was explored in this framework, while Evolution-based Machine Learning and RL are the two non-model-based techniques. The objective of this WP was twofold:
- the application of the three above mentioned learning techniques on global mechatronic systems
- the extension of these techniques: the introduction of fuzzy, non-model based information in MPC; the introduction of models of the interactions between subsystems in the Machine Learning algorithms.
In this WP, the results of WP 1 and WP 2 have been combined into an integrated controller and practically validated. For the integration into a combined controller the main focus was on the more complicated interaction between the global learning controller and the decentralised learning controllers for the subsystems. For the practical validation of the different learning control techniques, the objective was to select some representative development cases. On these cases, learning controllers were developed for the subsystems as well as for the global machine. The performance of these learning controllers was compared with the performance of current machine controllers to check the strengths and weaknesses of the different learning controllers.
Fig. 4: The objective of WP 3
The main objective of this WP was the dissemination of the project results to the user group and interest group, which is the first audience, as well as to other interested parties. In this way, the collaboration between companies and research institutions of Flanders is stimulated.
The objective of this WP was to ensure a smooth running of the project.