NDT Solution
Data Fusion for NDT: What, Where, Why and How
Today's NDT technology has advanced beyond its traditional role as a discontinuity detection technology. It has become essential for NDT professionals to keep abreast of these advances in NDT technology to fully utilize its potential. One of the critical technological areas where great strides have been made is the area of data fusion or data management. This month's article will be useful for the readers who want to know all about data fusion and its potential applications in NDT.
G.P Singh
Associate Technical EditorIntroduction
Nondestructive testing (NDT) has traditionally been involved only with discontinuity detection. In point of fact, NDT is a systematic approach to detect discontinuities in parts. Though discontinuity detection is a necessary part of the process to ensure that a system is capable of performing to necessary standards, the bare facts of a discontinuity and its location are not sufficient to today's decision makers. NDT is the only means to examine a material or component without destroying it and therefore is the primary input for the testing of system performance. For a proper decision to be made regarding a component or system, however, the identification of a discontinuity is needed as well as a determination of the effect of that discontinuity on the performance of the material. The NDT of today, by necessity, must be more than just discontinuity detection. In conjunction with the detection measurement, the raw data measurement must be correlated to the issues of interest in the test. In other words, the information regarding the specific measurement must be obtained. This extrapolation of information from the NDT measurement or raw data is what today's decision maker needs to test the material or system performance as a result of the presence of a discontinuity.
NDT has moved beyond its traditional role as a discontinuity detection methodology.
What is Data Fusion for NDT?
The term data fusion is a true misnomer. The purpose of a nondestructive test is to verify that the material or component is in the expected configuration for its purpose and therefore will perform as expected. The test provides measurements which can be interpreted to attest to the presence or absence of a discontinuity. If the ultimate goal is to determine the performance capability of a component, then simply knowing the presence or absence of a discontinuity is not sufficient to make that determination. More than a measurement is necessary when testing system performance. Testing system performance is accomplished through information - rather than data - with the possibility that additional measurements can provide more reliable or complementary information. Many sources discuss a variety of aspects associated with data fusion (Aldrin, 2001; Aldrin, 2002; Aldrin et al., 2002; Center for Energy Studies, 2002; Golis, 1990; National Oceanic and Atmospheric Administration, 2002).The concept of data fusion is integrally tied to the philosophy of knowledge management. Knowledge management is a process to understand how decision makers make informed decisions. This concept includes the following algorithm for understanding (Drucker et al., 1998):
data are discrete objective facts about events the message present in the raw data is the contextual information knowledge is derived from contextual information at the moment, along with experience, values, insight and many personal intangibles. Therefore the correct terminology which should be used here is not data fusion, but information fusion.
For NDT, this information fusion starts with the test data obtained on a component. From the measurement data, the presence or absence of discontinuities is one piece of information which can be gleaned, but there is other information that can often be taken from the NDT data - other material properties associated with the material under examination. It is this conglomeration of pieces of information from a single NDT data set that can be fused to provide contextual information on the material. In addition, multiple NDT techniques used to test a component would most assuredly increase the amount of information available from the tests.
Where Is This Information?
Every material has physical properties that define its nature. Each NDT technique measures one or more of these properties, since each NDT technology is a measurement of a physical phenomenon's reaction to specific material properties. However, no one NDT technology measures all properties and typically each technology measures only a limited set of these properties.In Figure 1a, we see a space of properties represented by an arbitrary component. Each of the circles in Figures 1b and 1c represents the information about the material property space contained in that particular NDT measurement. The information in any one of the NDT measurements is a subset of the total material property information space.
Why Fuse Information from Multiple NDT Techniques?
Fusing information is only relevant after carefully identifying and determining the question to be answered, that is, identifying a very focused problem space to be analyzed. What exactly is the ultimate decision to be made? From this question will come the context upon which the information must be determined. The specific context provides some additional understanding of the data needed by the test. So the first step in the analysis is to ask two questions: what decision is to be made and what information is needed to make that decision?Examination of the two questions will often show that the amount of information needed cannot be obtained from one data source, meaning that more than one data source will be necessary to gather the needed information to input into the knowledge base. This generates the need for information fusion of multiple NDT data sources.
The oval in Figure 1b and circle in Figure 1c represent the information concerning the material property space contained in particular NDT applications. Overlap exists between techniques (Figure 2). Where multiple NDT technique fusion comes into play is in the ability to take the information that is not in common between techniques and supplement one technique with information not present in the other. Through fusion, information from one technique is able to complement another technique and provide a broader knowledge of the overall materials property space for the material or component under test.
Just as information not in common between NDT techniques provides complementary information available for use, when two techniques provide data for the same material property, this is referred to as competing information (Osegueda and Holguin, 2001). This is represented by the overlap area in Figure 2. Competing information has an additional burden asso-ciated with its analysis. Unless one of the techniques is viewed as a benchmark measurement, then a computational combination of the multiple information sources can become extremely complicated. The potential does exist for combining the competing information to enhance a specific measurement for higher accuracy or reliability. Many technologies can be utilized depending on the problem being addressed. Competing information can be analyzed statistically or can be used to validate each of the techniques for accuracy and reliability. The value of competing information analysis is dependent on the problem being addressed. For instance, verifying that a discontinuity exists at a specific location is trivial if both NDT techniques detect it; however, improving the dimensional accuracy of a discontinuity size by fusing multiple NDT measurement data is very difficult and complicated.
Figure 2 shows how powerful adding NDT techniques and extracting the resultant information can be in the analysis of the entire spectrum of the material properties.
How Does One Fuse Information?
Humans fuse information pretty automatically. Computationally, a computer can be programmed to extract features, then classify them, using many recognized and studied pattern recognition techniques. Classifications that are well and analytically understood can be analyzed in mathematical closed form. If the actual relationships are not well understood, then other technologies such as statistical pattern recognition or neural networks can be employed.Ultimately, information fusion is obtained by one of two basic approaches: either by determining the discriminating feature or parameter that is the information of interest from each NDT technique or by defining a discriminating classification schema to process all the NDT data to a satisfactory result. The key to understanding the appropriate approach is to continually refine the problem space (that is, refine the questions being asked) and test the NDT technology to work within the problem space of interest. This analysis of the problem space can be thought of as a discovery process: one which seeks to understand what information is needed and how to obtain the data that would provide the most information.
Using this discovery process, refining the problem space and then examining the characteristics of the NDT data in terms of information potential will determine which of the two possible approaches for the effort will be most likely. If the features of the data (that is, the parameters to provide the information) are well understood and accurately discriminate the issue of interest, then a simple classification scheme will suffice to provide the information of interest. If the features or parameter space are not well characterized, then a very accurate classification scheme will be necessary. Often the problem reduces to a combination of both approaches. This means, if the features or param-eter space are not completely and thoroughly understood, then the classification scheme to answer the problem must be very discriminating on the features that are not completely accurate.
The feature extraction aspect is referred to in most pattern recognition problems as a preprocessing transformation of some sort, from which the discriminating, invariant features that represent the parameters of interest in the problem space are extracted from the raw data stream. This preprocessing task is typically selected for a particular application and requires very careful consideration to be an effective feature extraction device. Right now, there are no general rules on what is to be applied where and when for continual effectiveness.
There are two approaches to classification within a problem space. The first would be to determine the invariant, discriminating features applicable to the problem and use one classification scheme to fuse or combine the information. The second approach would be to generalize models for each of the NDT technologies and use a classification scheme to combine only that information that is relevant to the question of the decision (Figure 3). This second approach is ultimately moving the complicated process from defining the discriminating features, to an intricate classifier that must be able to wade through the complete NDT technique model and use only that information that is necessary for the information.
In terms of the complementary and competing information spaces, the identification of features and the classification issues must be based on the problem being addressed. As represented in Figures 1b and 1c, each NDT technology represents a different topology in the materials property space. The overlap area represents features common to each NDT technology. The complementary data are the features specific to each technology. How the features are selected, whether competing or complementary or both, depends on the question being asked and how broad a solution is required.
Conclusions
NDT has moved beyond its traditional role as a discontinuity detection methodology. NDT, in conjunction with detection measurements, must provide a correlation between the measurement data and information necessary to assist in the analysis of the materials and components throughout useful service life. NDT provides a very powerful tool to assist in efficient decision making.
Acknowledgments
I wish to thank John C. Aldrin, of Computational Tools, Inc., for his support.
References
Aldrin, J.C., Overview of Mathematical Modeling in Nondestructive Evaluation (NDE), Report NT-SP-01-03, Gurnee, Illinois, Nondestructive Testing Information Analysis Center (NTIAC), 2001.Aldrin, J.C., "The Role of Optimization in Addressing NDE Challenges," presented at Aeromat 2002, Orlando, Florida, June 2002.
Aldrin, J.C., C.V. Kropas-Hughes and J.R. Mandeville, "Automated Signal Classification Development Environment for Nondestructive Evaluation," presented at ASNT's Spring Conference 2002, Portland, Oregon, 18-22 March 2002.
Center for Energy Studies, "The Data Fusion Server," available at <www.data-fusion.org>.
Drucker, P.F., L. Dorothy, S. Susan, J.S. Brown and D.A. Garvin, Harvard Business Review on Knowledge Management, Boston, Massachusetts, Harvard University, 1998.
Golis, Matthew, An Introduction to Nondestructive Testing, Columbus, Ohio, ASNT, 1990.
National Oceanic and Atmospheric Administration (NOAA), "Data Fusion, What is it?," available at <www.ngdc.noaa.gov/seg/tools/gis/fusion.shtml>.
Osegueda, Roberto A. and Ana C. Holguin, "Fusion of NDE Data for Inspecting Composite Patches on Aging Aircraft," Abstracts of the Model Institutions of Excellence 6th Annual Conference, University of Texas, FAST Center for Structural Integrity of Aerospace Systems, El Paso, Texas, 4-8 April 2001, p. 37.
* Air Force Research Laboratory, Materials and Manufacturing Directorate, AFRL/MLLP, Wright-Patterson Air Force Base, OH 45433; (937) 255-9795; fax (937) 255-9804; e-mail <claudia.hughes@wpafb.af.mil>.
Copyright © 2003 by the American Society for Nondestructive Testing, Inc. All rights reserved.
Copyright © 2010 by the American Society for Nondestructive Testing, Inc. ASNT is not responsible for the authenticity or accuracy of information herein. Published opinions and statements do not necessarily reflect the opinion of ASNT. Products or services that are advertised or mentioned do not carry the endorsement or recommendation of ASNT.IRRSP, NDT Handbook, The NDT Technician and www.asnt.org are trademarks of the American Society for Nondestructive Testing, Inc. ACCP, ASNT, Level III Study Guide, Materials Evaluation, Nondestructive Testing Handbook, Research in Nondestructive Evaluation and RNDE are registered trademarks of the American Society for Nondestructive Testing, Inc. ASNT exists to create a safer world by promoting the profession and technologies of nondestructive testing.
