Motion Control Tunes into AI Methods

Techniques include such names as fuzzy logic, artificial neural networks, genetic algorithms, and expert systems. "Artificial intelligence" (AI) is the family umbrella that loosely links these techniques and separates them from traditional computing methods. Each mimics some aspect of intelligence—human problem solving, learning, knowledge, or evolution.

By Frank J. Bartos, Control Engineering May 1, 1999
KEY WORDS
  • Motors & motion control

  • Artificial intelligence methods

  • Software for controls

  • Advanced controls

Sidebars:
Research leads the way

Techniques include such names as fuzzy logic, artificial neural networks, genetic algorithms, and expert systems. ‘Artificial intelligence’ (AI) is the family umbrella that loosely links these techniques and separates them from traditional computing methods. Each mimics some aspect of intelligence-human problem solving, learning, knowledge, or evolution. For example, fuzzy logic (FL) works with linguistic rules and partial truth; artificial neural networks (ANNs) can learn through training; genetic algorithms (GAs) simulate evolution in biological processes; and an expert system (ES) reasons based on rules and experience derived from experts. For an overview of AI methods, see Control Engineering ‘s July 1997 cover story.

Originally, AI methods were meant to solve problems too difficult for classical approaches. That’s still the case for complex nonlinear, multivariable control problems. In other applications, including motion control, they perform auxiliary and supervisory functions. A newer direction for AI methods promises performance comparable to other solutions, but offering faster, less costly implementation. This can translate to additional motion control applications.

A wider view of motion control reveals AI methods at work in anti-lock brakes and cruise controls of vehicular systems, and in anti-sway control of cranes. More formal industrial motion applications are still ahead. As in the process arena, early-adopters of this technology-vendors and users alike-tend to be tight-lipped about upcoming developments.

Inform Software and Texas Instruments showed fuzzy logic control of ac
induction motors can shorten design time, offer similar performance, and
be more robust than traditional field-oriented control. Fuzzy control 1
uses one fuzzy block; neuro-fuzzy control 2 adds a second fuzzy block.

Enhancing flux vector control

Inform GmbH of Aachen, Germany (Inform Software Corp., Oak Brook, Ill.), a specialist firm in fuzzy logic, cites enhanced control of ac induction motors using FL combined with digital signal processors (DSPs). This refers to a joint application project with Texas Instruments (Houston, Tex.) that employed TI’s TMS320C series DSPs. Performance of traditional field-oriented (flux vector) control was compared to two types of flux controllers enhanced by fuzzy logic and neuro-fuzzy methods (fuzzy control 1 and 2, in diagram).

Given the speed step change shown in the diagram, the motor with flux vector control reached set speed within 0.25 sec with no overshoot. Performance falls, however, with motor heating and slight changes in motor-to-motor characteristics. Implementation effort took 3 person months, according to Inform.

In the first alternative method, time to setup and optimize the fuzzy controller is claimed to be just 4 person days. Here, a fuzzy flux controller handles the nonlinear relation between slip frequency and stator current, holding magnetizing current constant in all operating modes. A NeuroFuzzy module within Inform’s fuzzyTech design software generated the necessary fuzzy rules from existing sample data. Overshoot performance is nearly the same as for flux vector control, but time to reach set speed almost doubles. Computation time for the controller is given as 150 msec using the TMS320C 40-MHz DSP.

For fuzzy control 2, total development effort increases slightly to 7 person days. The enhanced fuzzy flux controller, said to be ‘more robust’ under motor parameter changes than field-oriented control, adds a second fuzzy block-a nonlinear fuzzy PI controller to improve performance. This replaces a standard PI controller in the outer control loop. Fuzzy control 2 reaches set speed almost as fast as the field-oriented controller and without overshoot. Performance is superior to fuzzy control 1. Computation time for the entire controller is 220 msec on the same DSP.

Constantin von Altrock, director of Inform’s Fuzzy Technologies Div., notes motion control applications of its fuzzyTech software range from fine positioning in CNC machines to robotics. Most of these customers use FL to improve motion system robustness against momentum changes. ‘Fuzzy logic has demonstrated it can well cope with nonlinearities in motion control and time-varying conditions such as magnetic saturation and strong dependence of electrical motor parameters on temperature,’ he says.

Inform and TI have an ongoing relationship to develop DSPs as the computational engine for fuzzy logic control.

Traditional control dominates

‘Traditional linear control algorithms are still overwhelmingly popular in industrial motion control because of their predictability and ease of analysis,’ explains Curtis S. Wilson, vice president of engineering and research at Delta Tau Data Systems Inc. (Chatsworth, Calif.). However, for the traditional approach to work well, a motion system must exhibit ‘strongly linear plant dynamics.’

At Delta Tau, AI algorithms find their main usefulness in system setup tasks. In particular, they help identify nonlinear effects in motion systems. One example cited is the setup procedure for flux vector control of induction motors. Determining the magnetic saturation characteristics of the rotor is an important consideration, since it limits the motor’s field strength and low-speed torque constant, according to Mr. Wilson.

Neural network algorithms are integrated into Delta Tau’s standard setup software for flux vector control. ANNs are very adept at identifying the linear portion of the rotor’s magnetization curve (see diagram). For a given quadrature current, the ‘knee’ of the curve corresponds to maximum linear motor speed. Staying within the linear range keeps standard control algorithms appropriate. An interesting twist here is use of neural nets, intended to solve nonlinear problems, helping to keep things linear.

Expert system programs provide an overall ease of setup for a motion system. Based on customer demands, Delta Tau has developed ES programs to guide users through the entire setup process. Ability to test and verify decisions at each step, as well as suggestions for appropriate changes, is included. Besides tackling typical tuning problems, the expert system can find such conditions as polarity mismatches, errors in setting resolutions, etc. Delta Tau considers the value of these setup programs to users so high as to devote ‘far more resources’ to them than to the actual control algorithms.

Neural network algorithms easily identify the ‘knee’ of the rotor magnetization
curve to keep control in the linear range, in Delta Tau Data Systems’ setup
software for vector control of induction motors. Quadrature voltage, V q , supplies
the torque-producing current under an acceleration command to the motor. Several
tests of this kind, with different rotor magnetization currents, characterize motor performance.

Supervisory role

Galil Motion Control Inc. (Mountain View, Calif.) looks at AI methods as tools for the future, and so far has not found a direct need for them. ‘We are continually evaluating AI technologies, keeping an open mind for future applications that warrant their use-for example, in more complex motion systems or where larger numbers of variables must be controlled,’ says Jacob Tal, Galil’s president.

Expert systems, incorporating previous experience and existing knowledge, are in use in Galil’s motion products. However, the ES has a supervisory rather than direct control role. These systems are applied strictly as a deterministic tool, currently serving two functions, according to Dr. Tal. In a pre-operation mode they run on a PC, performing design, tuning, and performance optimizing functions. For post-operation, they run on the controller, doing various system supervisory tasks. ‘Expert systems are no better than an engineer, but can work faster,’ he adds.

Siemens Automation & Drives Group (Erlangen, Germany) likewise sees existing methods sufficient to handle motion control needs. The affected control elements ‘can be fully described by mathematical formulas’ says a Siemens spokesperson. However, the company’s Simovert MV (medium-voltage) and ML (multi-level) ac drives use neural network techniques to compensate for harmonics effects. ML drives are very large cycloconverter-fed units with nominal rating in the 3 to 15 MVA power range.

Robotics looks ahead to AI

Fanuc Robotics North America (Rochester Hills. Mich.)-a manufacturer of robots and robotic systems-notes a place for AI methods in future products. As more complex dynamic models evolve to meet optimal process quality, AI technologies will be required. ‘But this is not yet happening. Advanced methods like AI are slow to progress from the R&D stage to practical industrial implementation,’ says Gary Zywiol, Fanuc Robotics NA’s vice president of Product Development.

He sees three prime areas where AI techniques can help enhance robotic process quality: painting, dispensing (e.g., seam sealing in automotive panels), and arc welding. Today’s process models simply can’t accommodate all the adjustments needed. In the case of arc welding, this means precise robot positioning, electrode orientation relative to the part, and welding parameter changes to compensate for variables in manufacturing conditions or inaccuracy in part specifications. Multiple variables to be controlled include welding current and voltage, wire feed speed, and robot motion adjustments to weave amplitude and frequency of the weld.

In particular, ‘Neural networks and fuzzy systems will provide adaptability to process behavior, which is unpredictable using explicit models,’ adds Mr. Zywiol.

Further developments and research will likely expand user awareness and the role of AI methods in motion control (see sidebar).

While motion systems consist of essentially linear elements, linearity itself has a price. Delta Tau Data Systems mentions some ongoing interest in applying ‘strongly nonlinear’ system components to reduce costs. Artificial intelligence methods could make a contribution here. ‘AI algorithms will be key to identify and control these systems, although there is a long way to go in providing general solutions and establishing user trust that these systems will work over the entire range of possible actions,’ concludes Mr. Wilson.

Research leads the way

Research work on AI methods for motion control is active in many universities, globally. Such projects often provide the basis of progress on the industrial scene. The concentration here on U.K. sites is coincidental.

Fuzzy logic, neural networks, and genetic algorithms are among AI methods currently receiving research attention at the Intelligent Motion Control Research Group (IMCRG)-part of the Department of Engineering at the University of Aberdeen, Scotland (U.K.).

IMCRG’s work in fuzzy logic controllers includes ac and dc drive applications, with the goal to simplify methods for effective real-time control. Neural network techniques are used to automate generation of membership functions and rules that are basic to the fuzzy logic approach. This has extended the project to fuzzy-neural controllers, as well. Developments in genetic algorithms (GAs) include a GA-based observer for an induction motor drive under rotor-flux-oriented control. This is a shared project with the University of Bologna (Italy).

Genetic algorithms are also on the agenda of the Robotics Research Group (RRG) at the University of Sheffield (U.K). Motion planning and collision avoidance are prime concerns for robotic systems, involving optimization among infinite possible trajectories of the robot arm’s ‘joint space.’ This is where GAs can be useful. These search/optimization methods are based on biological processes where survival of the fittest evolves into an optimal solution-satisfying one or more objective functions. As in nature, ‘selection pressure’ is used to favor individuals with the highest fitness in the population.

Main benefits of GAs are global search capability that avoids entrapment in local minima, and real-time operation via parallel processing (of population subgroups) in a distributed system of controllers. Typical motion optimization goals include travel time, energy usage, and minimizing torque magnitude applied to the robot’s motors.

Work of the Robotics Research Group is part of a recent book, ‘Genetic Algorithms in Engineering Systems,’ published by The Institution of Electrical Engineers (London), edited by Ali Zalzala and Peter Fleming. It also points out practical complexities and difficulties found in working with genetic algorithms.