Data-driven Control: The Needed Hero for the Present-day Control Challenges

Oh, how naive I was! Let’s flashback to 2014 when I sat in my first linear systems and controls course during my undergrad. Even though there was too much math to comprehend, I was intrigued by the idea of deriving control formulations from system models. It seemed like the world was at my fingertips. All I needed to control a system was a mathematical model, and I could run off and design compensators for it.

Photo by Rohan Makhecha on Unsplash

Little did I know, those first and second-order models that worked so well for those courses would nowhere come close to modeling even the simplest of systems out in the real world. If only controller performance requirements would be as simple and well-defined in terms of overshoot, rise, and settling times. You could be working with systems with

Little did I know that those first and second-order models, which worked so well for those courses, would not even come close to modeling even the simplest of systems in the real world. To add to the mix, if only controller performance requirements could be as simple and well-defined in terms of overshoot, rise times, and settling times. You could be working with systems that have:

1. Dynamics that are unknown, uncertain, or too complex to model
2. Parameters that vary over time
3. Uncertain operating points
4. Internal and external disturbances that are difficult to characterize
5. Requirements that are high-level and changing based on operating condition/region.

All of these potential roadblocks will make it a challenge for control engineers to tune controllers. In the industry, where 90% of the controllers used are PID controllers, there are cases where such controllers, and even gain-scheduled PID controllers, cannot achieve good performance against these conditions.

The trusty steed of the masses!

So, the million-dollar question is: How do we design controllers that can account for such uncertainties and provide us with the performance that we need?

Get Siddharth Jawahar’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

Enter data-driven control!

We have tons of data that we can collect from sensors on the system and the environment where the system operates and use it to our advantage. With data, you can:

1. Develop system models that then can be used to design model-based controllers such as LQR, MPC etc.
2. Design controllers that can directly learn and adapt from data. These ‘data-driven’ controllers can operate independently or supplement model-based controllers to help meet the control system requirements.

Model-Based Control Vs Data-Driven Control

In the next series of posts, let’s explore some of these data-driven control techniques and discuss how they work, when they can be useful and how you can get started using them for your control applications.

Data-driven control techniques

Specifically, lets look at each of these techniques below:

1. Active Disturbance Rejection Control
2. Model Reference Adaptive Control
3. Extremum Seeking Control
4. Data-driven Model Predictive Control
5. Reinforcement Learning

All on board? First stop: active disturbance rejection control!