Motion Control Resources
Observer Control Improves Motion
by Kristin Lewotsky, Contributing Editor
Motion Control & Motor Association Posted 02/19/2008
In motion control applications, arriving at the endpoint accurately and rapidly is often essential. Traditional control approaches tend to be linear -- they order the components to operate at a given velocity for a set amount of time, then stop. The problem is that phenomena such as stiction, wind-up and so on are nonlinear, and often result in the load overshooting the mark or oscillating about the end point (ringing).
Applications like electronic assembly pick-and-place or semiconductor metrology can’t afford to have ringing. The load needs to arrive at the targeted end-point as quickly as possible so that the operation can take place and throughput can be maximized. To eliminate ringing and overshoot, designers are turning to a technique known as observer control.
Observer control uses modeling to predict the error in motion, continuously updating that error with feedback from the encoder. “It’s really a best of both worlds,” says George Ellis, chief engineer of servo systems at Danaher Motion (Radford, Virginia) and author of Control System Design Guide, 3rd Edition. “You don’t rely completely on the model because models are always flawed, and you don't rely completely on the sensor because sensors are typically slow. You take this combination of the model and the sensor to give you the best combination available.”
How It Works
Observer control, as you might guess, requires an observer. A state observer typically combines system input/output with a mathematical model to predict the behavior of that system. For motion control applications, the classic starting point is the Luenberger observer. The mathematical model of the Luenberger observer, which is based on a double derivative of position, contains a correction term that is proportional to the expected error of positioning -- in other words, the system compares where it expects to be with where it actually is and uses that data to correct future operations. How it is actually implemented varies from vendor to vendor. In some cases, it’s an algorithm, in others, it takes the form of look-up tables.
The function of the observer is to help the system arrive at its end point with as little settling time as possible. “Let’s say we accelerate at half a G and let’s say we already know that at half a G, we see 50 radians of error, at 1 G we see 72 radians of error, and at 1.7 G we see 103 radians,” says Steven Reese, business development manager at Parker Hannifin (Wadsworth, Ohio). “[The observer] knows that it is accelerating at half a G and it knows that there should be an offset of 50 radians in the command signal and so it puts that in there. It basically offsets the error before it happens. The error’s nonlinear but who cares? At this rate it’s this and at that rate it’s this, so you literally build these massive tables and they get better and better as time goes on.”
In the initialization phase, of course, it’s not that simple, because the observer will have to compare its predictions with the actual numbers and create data to more accurately model the motion and drive the error toward zero. “What it is trying to do is figure out if it moves off the zero point, how much error is it getting? It will try that at it a couple of different rates and then it’ll start filling in the blanks in between. You can imagine in your calculus you might have three points when you start up and then later you'll have 100,000 points. You’ll probably have a lower error at 79,342 bits than you did when you first started up, and you'll get better as the day goes on, as the cycle goes on.”
“Think of the model portion as being very responsive but not that accurate and the sensor as being very accurate but not that responsive,” says Ellis. “The observer is built so that the high-frequency output is based on the model and the low-frequency output is based on the sensor.”
Observer Control in Action
From the standpoint of an end user, OEM or system integrator, observer control can -- and should be -- easy to the point of invisibility. Commissioning an observer can range from seconds to hours, depending on the design, so it’s important to discuss the details with the vendor.
In fact, this specter of complexity is one of the barriers to market acceptance that observer control faces. At a glance, the technique can seem more complex than the time-honored linear proportional-integral-derivative (PID) loop. As a result, the concept of observer control can put potential users off or simply make it difficult for them to truly appreciate the value proposition. That’s an issue that vendors are currently grappling with. “I think for the market to accept it, it's just got to be a simpler story than what we have today,” says Reese.
Part of that story is that although the concept seems complex, when properly executed, observer control is essentially transparent to the user. “It’s integrated into drives. You turn it on and it’s automatic,” Reese says. “[Users] don’t need to understand it. How does it work? It works great,” he quips. Of course, there are some manufacturers who choose that as their solution -- they simply don’t talk about the fact that observer control is integrated into their drives. Machine builders and end-users want precision motion, the thinking goes. They aren’t really so concerned with the details.
Except, of course, details like the 30:1 improvement in error that’s provided by observer control. “Any application in which people have to settle within less than 5 ms, they should start thinking observers,” says Ellis. “It's pretty hard to settle in less than 5 ms without one. You can do it,” he’s quick to clarify. “You can take standard products and settle in under a millisecond. You’ve just got to use everything available to get those kinds of settling times.”
Or you can buy a drive or system with observer control. Yes, the drive will need a processor with some muscle, but the computational burden of observer control is actually modest compared to the demands imposed by the current controller, Ellis says. “When you run a current loop, you have to run at such a high rate that the little bit you add with an observer isn’t much. [Observers] typically run at something between 5 and 10 kHz in a modern servo system, but the current loop is out there running typically at 16 KHz.”
Sweet spots for observer control include applications requiring fast settling times, such as semiconductor metrology; those requiring high throughput, like packaging or stamping; or those requiring steady speed, like grinding.
“[Observer control] can do things that PID loops can’t do, in particular handle huge inertia mismatches, dynamic resonances and nonlinear mathematics,” notes Reese. Linear error is relatively straightforward to compensate for a standard PID loop. Nonlinear error, though, can be nightmarish especially, when it results in ringing and overshoot.
“In the ‘80s, I'd be out there with an RC network trying to figure out where to put the notch filter,” he says. “Now, I don’t have to. [Observer control] can do so many more things. To a guy like me, it’s magic.”
Thanks go to Dan Jones, president of Incremotion Associates Inc. (Thousand Oaks, California), for helpful discussions.