Figure 1

What is Artificial Intelligence?
Artificial intelligence (AI) describes a broad range of computer applications that imitate human intelligence and behavior. Examples of successful programs include understanding speech, making judgments, and learning. Most of you have heard of computers that play chess, perform medical diagnoses, and/or recognize patterns. The programs that drive such computers use some form of AI. The medical diagnostic program analyzes information from the disease symptoms, patient medical history, and laboratory test results to suggest a diagnosis to a physician. This medical diagnostic program is an example of an expert system which depends upon rule based inference in which preestablished rule systems process the data. While such systems are very sophisticated, they do not approach the complexity of true human intelligent thought. AI has not made humans unnecessary but has provided us with another tool to make our work more efficient and error free. The example below is a similar rule based program.

 

An AI Application to NDT
An organization for which I worked ran an information center for nondestructive testing (NDT). Part of the job was to provide answers to requests such as "What would be the best NDT method to use on my problem?" These requests came over the phone from governmental and industrial institutions and the caller needed, and usually expected, an immediate answer. When one of our NDT experts was available, that individual provided the best answer he or she could. Likewise, the manager of the center would often be called upon to provide an answer.

There were three major problems that often left the callers less than happy. First, secretaries and clerks could generally only try to transfer the call to experts that had agreed to perform such a service. Second, the experts were frequently unavailable. Finally, the answers given by the experts seemed most often to involve their area of expertise (favorite technique). Consequently, radiographers recommended radiography, UT experts recommended UT, etc.

 

---

"AI can assist in decision making tasks related to NDT."

---

 

An AI system was tested to see if it could answer telephone requests for recommendation of an NDT method with satisfactory answers. The AI system was named the NDE-Advisor and designed to assist individuals not versed in the application of NDT to select the best NDT technique to solve a stated problem. The system was implemented in the OPS5 programming language on a VAX 11/750 that was always available for telephone contact. The program used approximately 350 rules in three separate components and was fast enough to run interactively.

The three components of the program were gathering general information about the problem and determining what general NDT methods might be appropriate, ordering the selected general methods according to the needs and capabilities of the user, and selecting the specific technique to recommend. Since this program was a first trial for such an application, only three specific techniques were completed to be available to the NDE-Advisor. They were thermal, radiographic, and ultrasonic testing (see Figure 1).

 

Selection of a General Method
The first step in reaching a solution was to guide the inquirer through a process of fact gathering. The facts include information about the type of material involved, accessibility to the specimen, and type, size, and location of the anomaly to be found. Next, questions were asked about circumstances of the testing such as availability of equipment and personnel (just as a human expert might ask). Other questions might reflect the importance of certain advantages and disadvantages of the specific NDT being considered, such as cost, difficulty in application, speed, precision, and safety. Answers to these questions better tailored the selected NDT to the user.

The logic or reasoning required to handle the answers to the above questions is called inexact or fuzzy reasoning. AI is appropriate for making such decisions and for performing reasonable compromises, and a unique way of handling these was developed.

 

Component 1: Selection of the General Method
Information on 13 different general NDT methods was gathered from several experts in different methods. This helped remove the bias expected from a single expert. The methods included visual, penetrant, magnetic particle, electric current perturbation, magnetic flux perturbation, ultrasound, eddy current, radiography, acoustic emission, dynamic/vibration, thermal, leak, and composition testing. To select the appropriate general method, the AI system draws upon information provided from the intended user about the problem as classified into three main categories - matrix, geometry, and discontinuity. Each of these three categories had a set of attributes associated with it (see Table 1).

This develops into a type of large, 2-D array with general methods on one axis and attributes on the other axis. AI can relax certain constraints when no one NDT method satisfies all of the requirements just as a human expert might. These decisions are based on a set of rules specifying when a general method is appropriate to the particular problem at hand.

 

Component 2: Ranking of General Methods
At this point, the intended user is asked various yes/no/maybe types of questions concerning cost, speed, accuracy, availability of equipment and personnel, and other issues pertinent to the methods selected as potentially usable. The AI system provided rankings of the possible methods from this kind of information.

 

Component 3: NDT Expert Systems
The third component of the NDE-Advisor consists of an expert system for each general NDT method. Each expert system determines a ranking of specific techniques within the general method should be considered. In radiography, this would be choices between isotope, X-ray, and neutron sources and between film and real time readout.

 

Testing the AI System
The NDE-Advisor tests involved having NDT experts, NDT technicians, and clerical staff operate the system. Small problems were found and fixed in the early stages of testing. After a short time, the NDE-Advisor was found to give no unacceptable selections and provided the nonNDT staff with a more convenient way to answer technical inquiries. The NDE-Advisor provided names and phone numbers of experts likely to be of help with the problem when a specific technique was not selected.

 

Conclusions
AI can assist in decision making tasks related to NDT. The quality of the decision making will depend upon the information provided to the AI system as well as the quality of the AI system used. The AI system allows nontechnical staff to provide good answers to questions involving technique selection and eliminates time necessary to locate human experts and their possible bias in technique selection.

 

References
Fink, P.K., F.A. Iddings, and Mary Overly, "An Analysis of Knowledge Used for a Structured Selection Problem." Proceedings of The 1987 Joint Computer Conference Exploring Technology: Today and Tomorrow, Dallas, Texas, cosponsored by the Association for Computing Machinery and The Computer Society of the IEEE, October 25-29, 1987, pp 672-677.

Iddings, F.A., P.K. Fink, and G.A. Matzkanin, "NDE Method Selection: An A.I. Approach," Proceedings of the 16th Symposium on Nondestructive Evaluation, San Antonio, Texas, sponsored by NTIAC and the South Texas Chapter of ASNT, April 21-23, 1987, pp 362-366.

 

 Table 1 Listing of attributes for matrix, geometry, and discontinuity

I. Matrix
	A. System
		2. Dynamic 3. Transmit sound 4. Does not transmit sound
	B. Individual Part
		1. Metal
			a. Homogenous
				i. Magnetic ii. Nonmagnetic
			b. Nonhomogenous
				i. Magnetic ii. Nonmagnetic
		2. Nonmetal
			a. Homogenous
			b. Homogenous composite/composite
				i. Fiber reinforced ii. Laminate iii. Other
		3. Other attributes
			a. Transmits sound b. Porous/nonporous
II. Geometry
	A. Access
		1. Side with  discontinuity 2. Side opposite  discontinuity 3. Opposite sides
	B. Surface
		1. Flat 2. Curved 3. Complex (valve, pump, engine, etc.)
	C. Form
		1. Thin plate 2. Thick plate 3. Pipe/rod 4. Cable/wire 5. Complex 6. Internal
	D. Finish
		1. Smooth 2. Rough 3. Irregular
	E. Special capabilities
		1. Stress 2. Pressure 3. Vacuum 4. Motion 5. Heat 6. None
III. Discontinuity
	A. Type
		1. Planar
			a. Crack b. Debond c. Lamination d. Lack of fusion e. Seams, laps, tears
		2. Volumetric
			a. Hole b. Pit c. Void d. Inclusion
		3. Dimensional
			a. Thickness b. Size c. Shape
		4. Property variation
			a. Composition b. Density c. Color d. Residual stress e. Surface treatment f. Mechanical strength g. Temperature
		5. System
			a. Missing part b. Worn part c. Misalignment d. Loose part e. Crack growth
	B. Location
		1. Surface 2. Near surface (< 17.4 mm [0.5 in.]) 3. Deep subsurface 4. Internal (> 76.2 mm [3 in.])
	C. Size
		1. Very small (< 0.8 mm [0.03 in.]) 2. Small (0.8-3.2 mm [0.03-0.13 in.]) 3. Medium (3.2-8.5 mm [0.13-0.33 in.])

*   1635 Rob Roy Ln., San Antonio, TX, 78251; (210) 647-7717; e-mail profiddings@juno.com.

Copyright © 2000 by the American Society for Nondestructive Testing, Inc. All rights reserved.

[ Materials Evaluation ]