Learning control: smart algorithms ensure that machines continue to function optimally
Learning control: smart algorithms ensure that machines continue to function optimally
To meet the demand of end consumers for increasingly better, smarter and personalised products, the industry must use ever better performing and more energy-efficient machines. The typical answer to this are increasingly advanced controllers in systems or components. However, current control algorithms have their limits when it comes to performing complex tasks and over time will show deviations caused by ageing machines. Continuously adjusting these algorithms is very labour-intensive and hardly efficient. By deploying learning control, control parameters are automatically and continuously adjusted so that machines continue to work optimally.
Learning control in practice
Machine controllers make sure that the condition of the system is permanently adjusted so that it continues to perform optimally. A typical example is cruise control in cars. It enables car drivers to drive at a constant speed. When riding uphill, the car speed will drop at the same power level. At this moment, the controller will generate extra power so that the car can again go to the pre-set speed, also on uphill stretches.
Traditionally, there are two ways in which a controller operates: through a feedback or feed-forward control system.
In a feedback control system, the controller will compare the target value of the system with the actual value. The controller gives a signal to an actuator (the combustion engine for instance) to apply a correction. In the above example, the controller will but react as soon as the speed drops and then re-increase the speed. In other words, a feedback system must first measure a deviation or ‘error’ to be able to react.
This is not the case in a feed-forward control system. Here, the controller will anticipate changes in the environment on the basis of simulations or models and initiate a pre-defined action. To get back to our cruise control example, the controller would foresee the speed reduction by measuring the steepness of the slope. In other words, a forward-feed control system does not have to wait for an error to occur before being able to intervene. A disadvantage is that it anticipates on the basis of models and therefore has trouble to handle new types of failures.
In both scenarios, the (re)action will not get better, no matter how many times the machine already performed the same task. Yet, this is exactly what machines in an industrial environment do: continuously performing the same tasks. The answer to this is learning control. Learning controllers register the system performances during previous executions of a specific task and permanently adjust the control of this task so that it is performed ever better under any circumstances.
Detecting deviations
Simply put, learning control algorithms look for abnormalities that occur during the execution of a task and will during the next execution try to eliminate these deviations in the best possible way. In other words: machines constantly adjust themselves so as to come ever closer to the targeted accuracy. The more often a task is executed, the better the controller will perform. And this without any intervention of an operator, who no longer needs to adjust the control parameters time and again. The machine will do this automatically.
This is also very practical when machine parts would start getting hot or when components start to show wear. Normally, this would affect the performance of the controller and thus also the behaviour of the machine. Learning control, however, does take this into account. The controller adjusts itself so that optimal quality can be upheld during the entire life span of the machine.
Control on two levels
The control is actually realised on two levels:
- A learning control algorithm is based on deviations that occurred during a previous execution of the same task and gradually learns to predict the machine behaviour better.
- Based on this improved prediction, the control signals of the machine are adjusted so that it will perform the task better next time.
Universally applicable algorithm toolbox
Flanders Make has experience with learning control in industrial applications. We therefore developed a universally applicable algorithm toolbox.
Five benefits of the algorithm toolbox:
- Compatible with very simple controllers. As such, the toolbox can be implemented without requiring too many adjustments to the existing system.
- Requires less modelling than traditional feed-forward systems because any inconsistencies between model and reality are eliminated in subsequent iterations (or repetitions of the execution).
- Can also be applied for similar tasks and thus not only for perfectly repetitive tasks. This is very useful for the production of product families (products that are similar but are slightly adapted to customer-specific requirements).
- User-friendly.
- Open-source: adjustable to the conditions.
