Troubleshooting servosystems by trial and error and rules of thumb is being replaced by Internet-based expert systems.
By George Ellis
EDITED BY MILES BUDIMIR
Electrical and mechanical servo-control systems are complex, and those who use and tune them often can dedicate only a small part of their busy schedules to developing sufficient control expertise. Analyzing and correcting servosystem problems is often counter-intuitive, requiring time for training and experimenting beyond what most in industry can afford. Consequently, servosystem users frequently find themselves in need of assistance when trying to solve even the most common problems.
However, an approach using expert systems helps diagnose and correct most servosystem anomalies. An expert system is an automated procedure that lets multiple, interdependent problems be addressed simultaneously. A three-step method of abstraction, matching, and refinement moves from observed behavior to general problems, general solutions, and finally to specific solutions. The end result is a ranked list of actions to improve the machine's performance. The expert system is implemented in software that can be executed from a Web browser, allowing unlimited access via the Internet.
Developing the expertise necessary to coax optimal response from a machine requires knowledge in numerous fields. The foundation is fluency in general control theory which includes achieving an optimal combination of stability and responsiveness, dealing with parameter sensitivity, applying feedforward gains, under-standing tuning procedures for multiple-loop control systems, and knowing how to appropriately apply filters. In addition, detecting and correcting insufficient dynamic and static stiffness is required. There are also numerous motion-specific controls issues such as mechanical resonance and nonlinearities resulting from backlash, friction, and compliance such as stick-slip-based limit cycles.
Beyond controls, expertise in servosystems requires a thorough understanding of feedback devices and related concepts such as resolution, accuracy, and phase lag. Also, familiarity with the more common sources of noise such as contacting surfaces (brushes, gears, pulleys, and so forth) and electrical noise caused by electromagnetic interference (EMI) and quantization is essential. Users typically spend most of their time developing expertise in the application of their machine, rather than in a single facet of its operation. It should come as no surprise, therefore, that most users cannot dedicate the time to develop the expertise in servocontrol that they would like to have when attempting to improve machine performance.
Problems with servosystems in machines occur in different combinations and in varying degrees. For example, suppose a gain is raised to improve response time. The gain is raised a bit, but not enough to completely cure the problem. At this point, the higher gain value causes the machine to produce excessive audible noise, another well known effect of high gain. So the machine has two problems simultaneously: excessive noise and slow response. Each problem has its own degree. The slow response time may be critical because it may cause the machine to be out of spec while the noise problem may be less of a concern.
Binary decision trees such as flowcharts won't work in this case because they are meant to handle problems one at a time. The simple example above had two problems. Even there, the expert system must be able to factor in other constraints such as the time available to implement a solution and the importance of cost. This is not to say that flowcharts are inappropriate for some servosystem problems. Tasks that are well supported by flowcharts are fault location and initial tuning procedures. However, when trying to get the optimal response from a machine, there is almost always a combination of problems present simultaneously. The right solution depends on knowing all the problems and the severity of each one.
When users recognize that a problem exceeds their expertise, they are left with few alternatives. First, they can research the problem. However, the time constraints at this stage of machine development are usually high, leaving little time to research the potential causes of the problem. This is because there is usually little time between when a machine is first functional enough to detect a problem and when the machine must be delivered. Another option is to have an expert visit the factory. However, experts are often not available on short notice, and if they are the visit can get expensive in a hurry. Lastly, an expert can be contacted by phone or e-mail. The problem here is that it is often difficult to explain the condition of the machine accurately enough and in sufficient detail so that the problem can be solved quickly. Expert systems are a timely and cost-effective alternative.
Expert systems are algorithms that en-code human expertise in software to perform a complex task. Their biggest advantage is that they are readily available and can simultaneously represent multiple problems occurring in varying degrees. They also provide multiple solutions ranked in order of their likeliness to fix the problem. Machine condition descriptions can be easily modified so that new conditions can be added. By narrowing the source of the problem, expert systems can provide reference material focused on the current condition, helping the user build expertise in the current areas of trouble.
However, expert systems do have their limitations. The current implementation assumes that expert systems will eliminate the need for a human expert only about half the time. However, it should educate the user so that, if a human expert is subsequently needed, exchanges between them will be more efficient. Also, the expert system is not expected to work with novices. Users are expected to be experienced with servosystems and with the operation of their machine.
The expert system operates on the cognitive process model of abstraction-matching refinement. In the abstraction phase, observations of specific problems are transformed to general problems. For example, mechanical resonance causes oscillations that often generate audible noise. In abstraction, observations of audible tones must be transformed to the general problem of resonance. In matching, general solutions are found for the general problems. For instance, one general solution to resonance is to stiffen transmission components. Finally, the process of refinement converts the general solutions to specific ones. For example, using a stiffer shaft coupling or stiffer belts are refinements of the general solution to stiffen the transmission.
During the investigation phase, specific problems are characterized. Most servosystem problems fall into four categories: excessive noise, inaccuracy, slow response, and instability.
Suppose there are two problems: an ever-present, excessive, random noise and a sharp, quick-duration (high-frequency) overshoot that develops in response to aggressive velocity command changes. Investigating the first problem suggests the category of excessive noise. At this point, the type of noise must be categorized. There are four types of noise to choose from:
- A near-pure pitch from high-frequency resonance
- A grinding, irregular tone from low-frequency resonance
- Random noise from electrical (quantization, EMI) or mechanical (gear-teeth-inter-face) sources
- Low-frequency hum indicating vibration
Once the problem has been categorized, the severity is specified. The current algorithm recognizes four levels of severity: primary, serious, moderate, and small. A severity indicator is necessary because multiple problems commonly occur together so that steps that improve one problem often worsen another. For instance, increasing a gain may improve system response but may also increase noise. So if the system has serious noise problems and moderate responsiveness problems, raising gains would be unlikely to help. However, were the positions reversed (noise moderate and responsiveness serious), raising gains would become a more likely solution.
The expert system identifies more than a dozen general problems which can exist in any combination and with any degree of severity. Each specific problem is converted to a general problem. This does two things. One, it characterizes the current state of the machine so that the expert system can suggest appropriate solutions. And two, by guiding users through the process a human expert would go through, they are gaining expertise in servosystems. Placing resources that deal directly with the current problems throughout the investigation, such as application notes and articles, ensures that users will be more likely to study the information than they otherwise might.
After the investigation phase, the data is collected and individual ratings for each of the general problems are combined to form a machine condition profile. This profile is augmented with other constraints such as time and cost issues.
The expert system has about three dozen common solutions to servosystem problems. The solutions span the full range of common problems; adjusting gains, improving mechanical and electrical configurations, and correcting feedback problems. Each of these solutions has a profile that identifies which problems improve and worsen, and to what degree. In the recommendation phase, the solutions whose profiles best match the machine condition profile are placed in an action list. Actions are ranked in order of the likelihood of a solution working.
The recommendations are where much of the judgment of the expert is encoded. Here, trade-offs must be made between how well a solution corrects some problems versus how much it exacerbates others.
ACTION DESCRIPTION (REFINEMENT)
In the final phase, actions are described in explicit detail along with information about which problems will be improved and which will worsen. From this information, a final judgment can be made on which actions to select. The action list also helps explain the reasoning used to select this action. This provides a degree of transparency which helps in accepting a solution offered by an automated tool. It also provides information for assessing when the action has been executed correctly versus both incorrect use and overuse.
The user must understand the reason why a particular solution set was chosen in order to accept it. This is required because the user must often go to great effort to implement a solution. For instance, stiffening a transmission or converting to a direct-drive motor may require significant time and cost to redesign the machine. A user will want to be convinced of the likely improvement offered by an action to be willing to take the risks associated with it.
The expert system is implemented in Web-based software so that all phases can be executed by a standard browser. Operations such as profile matching can be implemented within Web-based applications by scripting languages or by submitting information to the Web-site server. Web-enabling the expert system has the advantage that there is no special interface software needed. Also, putting the system on the Web ensures that it is available all the time.
To take the ServoExpert for a test drive, go to http://kmtg.kollmorgen.com/customer_support/servoexpert