Driving Model Predictive Control

Last month I had the opportunity to sit in the driver’s seat of an auto-piloted vehicle as it drove down the highway. I had access to a handful of parameters to describe my preferences like “keep at least four car lengths behind the vehicle in front” and “perform lane changes normally” but I left all the moment-to-moment decisions and driver actions to the autopilot. Perhaps with some experience, I would feel comfortable using the available Mad Max lane change mode but I refrained! The autopilot system performed well and the manufacturer’s representative suggested I assume control of the wheel as we returned to the more chaotic neighborhood streets of a beautiful Saturday afternoon.

Since that day, I have been reflecting on how far supervisory control has come since 1996 when my colleagues introduced a model predictive control offering for industrial manufacturing, now known as Pavilion8 Model Predictive Control by Rockwell Automation. Similar to the automated driving systems of today, Pavilion8 MPC uses live measurements of the “location” of your process and models of how that process reacts to control actions over time. Models and math solvers are used to optimize a trajectory of future actions to take the process closer to the preferences set by its operator/driver such as production and quality targets, while keeping away from boundaries like high temperature, vibration and motor overload. By adding economic preferences, control actions may continuously reduce energy or consumable use while meeting other goals. The operator, like a driver, sets their preference based on the needs of the moment and allows the automated system to make regular control actions regardless of the level of attention they can provide. When needed or desired, the autopilot/model predictive controller can be turned off, completely or partially and the process controlled manually. In an autopilot car, a partial autopilot example would be continuing to use adaptive cruise control but without automated steering.

Then, as now, most manufacturing process models and control calculations are more easily defined than automotive autopilots. Process behaviors are driven by physical laws and (usually) don’t have to account for pedestrians and occasional Mad Max behavior from other drivers. For this reason, industrial processes have taken advantage of model predictive control for decades longer than our friends in the consumer-facing automotive industries.

Scott Jost – December 19, 2018