Complex industrial processes are controlled and automated with PID controllers and traditional linear MPC/APC solutions. MPC (model predictive control) evolved from complex process industries, such as oil, gas and chemical production, but is now equally at home in everything from optimising wind turbines in wind farms to controlling driverless vehicles and even satellites. These control mechanisms excel at keeping a complex process stable, steadily producing many essentials of the modern World. However, as soon as something changes, be it process inputs, desired outputs, or other disturbances, a human, the current plant operator, needs to step in and adjust controller setpoints. Taking a complex industrial process from a steady production state to another requires deep understanding of the process and on-the-spot vision of desired outcome, hardly features of a traditional, rule-based computer system. To change that, Curious AI has created a system that we call Neural MPC.
The Neural MPC learns a system model from data and adapts to changes as they happen. The system model is then used to predict and plan action sequences optimized to satisfy given targets and desired prediction certainty. Curious AI’s System 2 approach, separating the model and optimization, is a step-change in the flexibility: The targets can be changed on the fly without retraining any models. Also, model-based optimization reduces the need for training data considerably, making initial setup and adaptation simpler and cheaper than with traditional machine learning methods.
Since the system model produces human-readable predictions, the end-user will gain new possibilities in predicting system future state, analyzing what-if scenarios and suggested optimal solutions, and optionally select and approve them.
Deploying the Neural MPC is straightforward. First customer’s R&D team defines the data stream interface, available measurements, controls and applicable targets for integrating the Neural MPC system with the target product. Second, using existing rudimentary controls or safe exploration, the Neural MPC is trained from real process data. Depending on the process, the length of the training history ranges usually between a day and some weeks. Finally, after initial training and performance checks, online self-learning takes over and Neural MPC can be taken into use, either as an advisor for a human or as autonomous closed-loop control.
Industry 4.0 development has advanced to the stage where real-time, connected sensors, required for Neural MPC, are readily available. The other requirement for successful implementation is the definition of target functions and constraints based on the measurements provided by the sensors.
Neural MPC is a very general approach. It can be successfully applied in industrial control, autonomous vehicle operation, mobile robotics, and even more abstract systems like scheduling, planning and logistic optimization. The sensors can be, for example, orientation and pressure sensors on a robot arm and the controls can be, similarly, the actuator motor variables. It works in cloud, in premises or even on mobile units with edge computing.
Contact us for more information and demos.