Our projects are simulation of dynamic processes to understand the feasibility of the process control. When demonstrated, we tested our concept in a pilot then on industrial units.
There's nothing more boring than a level controller! They worked well without intelligence for decades. Can the "Artificial Brain Control & Supervision Unit" (ABCSU) learn enough to control the levels on its own? Our target is to test if a level control can be executed by an AI unit.
The level controller
In our simulations, the inlet flow (Fl) and the set point (yellow line in graph) were changed. While this happened, the controller was able to quickly learn what it had to do for the process to reach the set point target as fast as possible.
As shown in the graph, the controller successfully adapted the outlet flow and the level reached the appropriate requested levels as soon as possible. To better approach the reality, the supported simulation adds a noise on the level signal.
Even a level controller can be improved!
The flow controller
A completely different challenge emerges with the flow controllers as the system effects are immediate.
We simulated an air compressor with a constant demand of airflow. The compressor fills a tank at a pressure of 8 bars. The pressure is going down when the air is consumed and when the pressure reaches 6 bars, the compressor is refilling the tank to 8 bars, immediately.
To make our simulation tougher to learn, we purposefully did not give the pressure information to the controller. This was to test the adaptability of the controller on fast changes (instant changes), observe the controller learning live and observe its adaptation skills : due to the absence of pressure information, the controller is not able to anticipate the coming pressure changes and has to adapt to them in real time.
And it does! Our simulation shows that regardless of surprise pressure changes, the controller is effectively able to adapt and solve the issue.
The concentration controller
The new challenge here is to feed more information to the supervisor.
Liquid is fed to a tank that contains salt at a measured concentration. The flow and concentration of the inlet liquid is changed over time and must be controlled. Concentrated salt is added into the tank where it is mixed with the liquid to reach the required concentration. In this exercise, we consider that the flow at the exit of the tank is equal to the flow at the inlet (liquid to tank + volume of the salt), meaning that the level is kept constant.
The data provided to the controller are flow inlet, salt concentration inlet and salt concentration outlet, which must be adjusted to a set point target. The intelligent controller has to interpret the data so it can output the right control action. The system is simulated but the controller is not given the physics equations needed to solve the problem.
As the flow and concentration of the solution to the tank change, the controller learns and adjusts the flow of concentrated salt to have the salt concentration to the requested level (set point in yellow in the graph) as fast as possible. The blue line shows the simulated concentration in the graph.
The outlet to the control valve is sometimes a little unstable, showing a lack of learning; however that will be optimized for the industrial control.
The tank supervisor
Definition: A supervisor is a system that has control over more than a single variable at the same time.
The tank supervisor has control over two variables: level and concentration at the same time. The purpose of this simulation is to check the adaptability of the control system when interactions are created during the control. Checking the adaptability will let us know if the system is robust, if it would lose control at some point, and of course, how fast the intelligence can learn, adapt to and correct the deviations. Will the interactions between the different variables prevent learning and adaptability?
Information given to the supervisor are the flow at inlet (flow and salt concentration), the level of the tank, the flow of concentrated salt, the concentration of salt in the tank and the flow out of the tank. The level and the salt concentration in the mixing tank are controlled.
The inlet flow and concentration are changed permanently, as well as the set point target (in yellow on the graphs) of both controlled variables (results of simulation in blue on the graph).
We used the same hypothesis than earlier for the simulation: we have a scientifically unaided supervisor and at the starting point the supervisor didn’t use any of the previous simulation as single loops: it is a blank slate.
The two graphs show the results of the variables tuned at the same time and all the changes introduced to the system. Again and as wanted, the supervisor goes online directly and learns and adapts in real time! Some lack of learning can be perceived mainly from the tuning of the concentration and an output that’s slightly unstable. Despite this, however, the tank concentration is well maintained. This is not so surprising considering that the simulation represents less than 15 minutes with several changes and a supervisor that is a complete virgin at the start.
© Promachem 2017