The old gold standard for motion control is ladder logic control, and IEC 61131-3 language capabilities on newer programmable controllers are proof of its enduring practicality. But where rigidly executed control isn't efficient (such as for motions that must be simultaneously coordinated or responsive) ever-improving processor and computing power have revolutionized automation. They've spread the use of PID controls, and now sophisticated adaptive control, PID-based and otherwise.
Originally for simpler applications, adaptive controllers are moving into high-bandwidth control functions on machines and robots that need piles of sophisticated data to react predictably on the fly. Nonlinear mechatronic systems such as fluid drives and ship navigation also benefit from augmented or non-PID adaptive control.
On the horizon are new inferential estimations, neural-network-based technologies, modeling methodologies, data analysis, and embedded adaptive control, according to a recent study by Frost & Sullivan, Palo Alto, Calif.
Not just under research anymore, adaptive systems go beyond simple compensation within a set system bandwidth (what traditional closed-loop control can handle) to boost reliability in real working systems, even when loads, speeds, and environment change dramatically.
Model reference adaptive control (MRAC) includes a closed-loop controller with updatable parameters. Output is compared to that of a set target model, and then control parameters are adjusted to match.
“Our Sigma-5 amplifier employs adaptive autotuning by using model-following control to optimize feed-forward gains while the servo is in motion,” says Todd Rhode of Yaskawa Electric America Inc., Waukegan, Ill. The amplifier constantly adjusts speed loop and position gains, speed loop integral time constants, and torque filter time constants to reduce motion settling time — and optimizes reference filter gain, friction compensation variables, and notch filter settings based on load conditions and changing load inertias.
Generally speaking, these systems track error, and then prompt changes in one or more preselected parameters to correct situations. A preset sensitivity determines just how much these parameters change.
Stability is a major challenge here, especially when signal magnitudes are large. So, some systems use reliable gain scheduling, which adjusts gains in an open-loop fashion — so actions are actually more predictable. Gain scheduling algorithms can't be fooled by unmodeled dynamics — by mistaking inertia for friction, for example.
Other model tracking uses fuzzy logic to prevent drives from tripping unnecessarily and keep values within limits. For ac motor control, some effective fuzzy systems integrate the output of differential and proportional components to reach set speeds as fast as field-oriented controls (FOCs) without any of their overshoot.
Direct torque control (DTC) eliminates the need for encoders and lends itself well to adaptive control. “People conceived encoder-based FOC first, and then tried to modify it for sensorless torque control algorithms,” says Mark Kenyon of ABB Inc., Automation Technologies, New Berlin, Wis. So most sensorless vector controls use FOC architecture as the starting point, and then estimate speed using motor current and voltage information.
“But DTC estimates machine flux and torque, so two (hysteretic) loops are closed without intermediate torque and flux-producing loops,” Kenyon explains. The output of these hysteretic control loops is an inverter state optimized to output flux and torque quickly and accurately. “This hysteretic loop runs every 25 µs,” Kenyon adds.
No explicit current regulators means there is no explicit inverter voltage command — and no need for a traditional pulse width modulator. So, hysteretic control loops based on torque and flux errors select the inverter output state — then applied for the next 25 µs.
For an accurate adaptive motor model, a user either supplies motor nameplate data (leaving the drive to identify necessary machine parameters) or spins the motor at no load in a (rarer) procedure for identification.
Another way: All autotuning
Other adaptive controls do away with models; these systems are called self-tuning regulators or model-free adaptive (MFA) controls by one developer, CyboSoft, Genreal Cybernation Group Inc., Rancho Cordova, Calif. This approach doesn't reference any set parameters, or estimate unknown parameters in the controller, but collects them separately, after the actual system situation is appraised.
It's practical, because the algorithm doesn't need to match a model — so upon startup, designers don't need to spend time and money building models for it. An inner regulator loop includes a definition of system moves, and a linear feedback regulator. An outer estimator loop includes a recursive parameter estimator and design calculation to adjust the inner loop's regulator. Applications include industrial conveyor belts and automobile cruise controls.
Genetic algorithms (search techniques designed to emulate the genetic evolution found in nature) may soon augment more autotuning systems. Here, algorithms identify parameters in an estimator loop, with single-input-single-output models representing dynamics for the regulator. Research shows that genetic autotuning controls, though computationally expensive, are robust and can control nonlinear systems.
Traditional controllers sometimes compromise complex systems. Balancing error occurrences and correction identification with steady performance is challenging.
Some next-generation adaptive controllers use model changes and process output to compute an integrated square error (ISE) for each of three process parameters. So after analyzing the low, middle, and higher combinations of all three, it can devise 27 models.
“Through continued iterations, each model is normalized to a total ISE and the best value computed for each of the parameters is used in the next iteration as the middle value,” says research analyst S. Menaka, of Frost & Sullivan. “Thus, the model undergoes interpolation with recentering of the parametric values, to ultimately reach an optimum corrective model.”
Autotuning smoothes motion and eliminates vibration. In some systems, functions monitor load oscillation frequencies and suppress load vibrations — so machine builders can use structures and motion transmissions that are less stiff, to reduce machine cost. “Transitional vibration can be suppressed at frequencies down to 1 to 100 Hz, generated mainly if the machine stand vibrates during positioning moves,” Bill Leang, engineer at Yaskawa, explains. “Detection is based on vibration that may appear as position error. An amplifier algorithm then applies filters to suppress this vibration.”
Application example: Dancer arms
A dancer is a free-moving idling roller used in paper and fabric rolling that can change its position in relation to web tension. When tension is okay, its arm rests in a neutral default position. But if tension is too low or high, the arm moves off that. “A dancer-arm controller is considered a tension controller, but is actually a position controller; it controls the position of the dancer arm,” explains Bradley Briggs, electrical engineering manager at Nexen Group, Inc., Vadnais Heights, Minn.
The controller's job is to return the arm to its running position. So if torque is increased, the roll resists paying out web, and increased tension causes the dancer arm to reduce storage in the web loop. Likewise, if torque is decreased, the roll pays out web more easily, and reduced tension causes the dancer arm to increase the amount of web stored, to take up slack. Adjusting torque to maintain a constant arm position in the middle of this tug-of-war between roll and web machine is not easy. It requires a well-designed dancer arm, high quality torque actuator, and stable dancer arm controller.
“Our RSD200 controller uses a nonlinear second-order PID algorithm with gain adaptation plus other algorithms to compensate automatically for changes in roll diameter and inertia,” explains Briggs. It does this by utilizing a mathematical model of a typical torque-controlled winder. The adaptation process modifies this model in real time to account for current machine and roll conditions.
The models take into account linear effects caused by roll diameter changes, and nonlinear effects caused by roll inertia changes. During initial tuning, the mathematical model is altered to fit the web machine. “While running, another controller of ours, the RSTC1000, adapts its internal gains to the modified model and its estimate of the roll diameter,” adds Briggs. Resulting gain changes follow a realistic nonlinear relationship.