DATA FUSION METHOD FOR THE OPTIMAL MIXING OF MULTI-FREQUENCY EDDY CURRENT SIGNALS

Z. Liu, M.-S. Safizadeh, D. S. Forsyth, and B. A. Lepine

Structure, Material, and Propulsion Laboratory, Institute for Aerospace Research National Research Council Canada, Montreal Rd 1191, Ottawa, ON K1A OR6, Canada

ABSTRACT. Eddy current testing methods are commonly used in the inspection of aircraft fuselage splice joints for corrosion and fatigue damage. However, the inspection of these components suffers from the spurious effects introduced by liftoff, interlayer gaps, rivets, and paint. To overcome these effects, multi-frequency eddy current testing has been proposed, but has not been widely employed due to the difficulty of correctly mixing the different signals. In this study, we investigated the potential of using a data fusion technique to combine the results of different frequencies and quantify hidden corrosion in service-retired aircraft lap splice joints. An updating mechanism based on a contextual Dempster-Shafer approach was used to combine probability mass values from multiple sources. The final classification results were obtained by making decisions based on the maximum belief of the fused results.

INTRODUCTION Manual eddy current (ET) inspections are the most commonly used NDI procedure in aircraft maintenance. Most often they are called for in the inspection of components for fatigue cracking, but they have also shown potential for detection and quantification of corrosion damage in the faying surfaces of fuselage splice joints. In the case of splice joints, the ET signal is affected by changes in thickness of the joint material, which is then assumed to be due to corrosion. The ET signal is also confounded by a number of extraneous factors including variations in probe tilt, probe-specimen liftoff, and interlayer gap to name a few. In order to eliminate some of the extraneous factors, the mixing of multiple frequencies of ET has been proposed by various authors (see for example [1,2]). However, these procedures have not been adopted into common practice, at least in part due to the complexity of the procedures and the sensitivity of analog signal "mixing". This paper presents results from multiple frequency eddy current testing (MFECT) where the mixing is performed by computer post-test using data fusion algorithms. When multiple inspections provide complementary information about the specimen, combining the data may facilitate the analysis or classification process. Data fusion techniques provide a framework to fuse and integrate information from

CP657, Review of Quantitative Nondestructive Evaluation Vol. 22, ed. by D. O. Thompson and D. E. Chimenti 2003 American Institute of Physics 0-7354-0117-9

577

multiple sensors or sources. Dempster-Shafter (DS) theory is one of the data fusion approaches that provide a mechanism to fuse information from multiple sources. Fusing of MFECT data is not a new topic and some researchers presented various fusion algorithms to achieve a better signal-to-noise ratio [3,4]. The use of DS rule to fuse nondestructive testing (NDT) data was described in reference [5] by Gros. The key difference between DS and other common data fusion methods is the assignment of probability mass, which is the first and crucial step in the process. Unfortunately, there is no common answer to the question. This remains an unsolved problem and largely depends on the application itself. To detect and characterize corrosion and fatigue damage in aging aircraft, in the case of multi-layer riveted joints, the following metrics may be required: material thickness loss by layer, corrosion pit size and distribution, pillowing deformation; and crack size, location, and orientation [6,7]. It is not likely that one single NDT method can characterize or quantify all these metrics. Because multiple NDT methods are employed for these inspections, this raises the potential for applying data fusion techniques to interpret multi-sensor data or quantify the results. At NRCC, the following NDT methods are being used for detecting hidden corrosion in aging aircraft: single/multiple frequency eddy current, pulsed eddy current, ultrasonic, and two computer vision based systems: Edge of Light and DSight inspection. In this paper, we present the application of a DS fusion rule to fuse MFECT inspection data from Boeing 727 aircraft lap splice joints. A probability mass was assigned to each pixel, and after applying the DS fusion rule a contextual process was carried out to update fused results. The decision was made based in the maximum belief of the updated results. APPLYING DEMPSTER-SHAFER THEORY The DS Method

The core of the DS method contains three aspects: the concept of probability mass, belief function, and the updating mechanism [5,8]. The frame of discernment 6 is a finite set of propositions that are mutually exclusive and exhaustive. The power set of 9 is 2 , and the elements of this set are all subsets of 9 . The mass function assigns degree of belief across the set of all subsets of #, such that,
jn:2*->[o,l], mfa)=0

The quantity m(A) , known as basic probability assignment (BPA), is a real number between zero and one. It represents the exact belief committed to A . A function "Bel" is called a belief function if it satisfies the following conditions:
Jfe/:2'->[o,l], Bel fa) =0 Bel(A) = m(B) , for all A c 9

(2)

578

Bel(A) is a real number between zero and one that represents a degree of support that all the available evidence provides for A. Given a belief function Bel(A), the function Dbt(A) = Bel(-iA) is called a doubt function and represents the total support for the negation of a proposition. The plausibility function of A, denoted by P1(A), is written as Pl(A) = l-Dbt(A). We assume the current state of the system has the value 7^(4) assigned to all the subsets of 0, and represents the total support from all the previous evidence. The observation of a new distinct piece of evidence by a mass function m2, distributes a new set of mass values m2(Aj) over the set of 29 . These new mass value m2 and the old values
ml are combined to produce updated values mu. The updating mechanism is performed by:

(3)

where mu is called the orthogonal sum and can be written as mu = ml m2. Formula (3) is called the DS rule of combination. The crucial problem of using DS is how the mass function distributes the mass values among the subsets of the frame of discernment. Unfortunately, DS theory does not give the answer. The procedure largely depends on the application.
The Procedure of PS-based Contextual Data Fusion

The procedure that was used for fusing MFECT data is given in Figure 1. To relate the measurement value (voltage) to corrosion damage (material loss), one can use a straightforward calibration approach. That is, given a measured voltage, find the material loss corresponding to that measurement using a calibration curve. However, a measurement value does not uniquely correspond to a material loss quantity, due to noise sources affecting the measured signal. In this work, it is assumed that the actual material loss for a measured value is normally distributed. While this has not been demonstrated for the MFECT data considered herein, this assumption has been widely used for estimating the probability of detection [9]. This has also been demonstrated to hold true for diverse NDE methods [10] in a more general derivation.

ECT 5.5 kHz

hi
ECT 17 IcHz

M M*:_
Data Distribution

Fused Result
,.miM .,
W

OS
Rule

2t 22

Contextual Process ^ ,v max ( A )

.1,

mass value

^^^^^^^^M

FIGURE 1. The procedure to fuse MFECT data.

579

One way to obtain this distribution is to use calibration specimens. In order for the results to accurately model the range of results expected from inspection of actual inservice components, all the relevant variables must be included in the calibration specimens. In the case of riveted lap splice joints, these variables include paint, probe tilt and liftoff, and interlayer gap variations. Construction of these calibration specimens quickly becomes a complex and arduous task. Thus, in these experiments, a section of an actual Boeing 727 lap splice was used as prior knowledge to build the distribution maps. The specimen was inspected with the MFECT technique described below, and then destructively examined. After disassembly and cleaning, the individual layers were thickness mapped using a radiographic technique developed at NRCC. The range of measured values of thickness was divided into 100 data bins, with each data bin corresponding to a modeled normal distribution curve that satisfies formula (1). From the distribution map, the probability mass value is assigned. Usually for DS methods the decision is made on the result of formula (3), but the contextual information is not considered herein. A contextual process integrates the spatial correlation between adjacent pixels in order to improve the classification results [11]. A simple implementation of this process computes the average of a 3 by 3 block region:

'M4),

=I >12(4),,, y q=i-\ r=j-l

(4)

The contextual process can be iterated if the procedure generates newly labeled pixels. Finally, each pixel is assigned to the type of corrosion with the maximum belief value.
EXPERIMENTS AND RESULTS

A MIZ 40A eddy current instrument was used in the experiment to drive a sliding probe housing with adjacent coils in transmit/receive configuration. The probe was excited at four discrete frequencies, -5.5 kHz, 8 kHz, 17kHz, and 30kHz, simultaneously during one scan. The Winspect data acquisition software was used to capture the inspection data. The subject specimen was a section of a service-retired Boeing 727 lap joint. It consisted of two layers of nominal 0.045" thickness Al 2024T3, and a stringer of Al 7075-T6. The joint was fastened with three rivet rows of equal 1" spacing between rivets. A photograph of the specimen is shown in Figure 2 and the inspection results are shown in Figure 3. After applying MFECT and other NDT inspections, the specimen was disassembled, cleaned, and the individual layers inspected with a radiographic technique to determine remaining thickness.

FIGURE 2. Photograph of a section of the Boeing 727 lap splice joint used for this study.

580

FIGURE 3. MFECT inspection data at four frequencies. (The scale is in voltage.)

Note that only the imaginary parts of the signals were used in the DS analysis. Since a conventional eddy current inspection procedure was used, the real component of the signal is affected by varying probe liftoff, interlayer gap, and some other factors. For this reason, phase information was not analyzed. Table 1 gives the relation between the frequencies and penetration depths. In general, the 30 kHz and 17 kHz inspections reveal the first layer's characteristics while the 8 kHz and 5.5 kHz frequencies penetrate to the second layer.
TABLE 1. Inspection frequencies and penetration depths.

Frequency (kHz) 30 17 8 5.5

(mm) 0.698 0.925 1.349 1.627

Penetration depth (xlO'3 inch) 27.5 36.5 53 64

X-ray thickness map: top (left) and bottom (right)

m^m
II II

Fusing 5,5 and 8kHz for top

layer (left) and

bottom layer

Fusing 17 and 30kHz data for top layer (left) bottom layer

(right)
0035
OJ1 tJtl

FIGURE 4. Image fusion results for section D. (Grayscale shows the thickness and the unit is inch.)

581

For the procedure of mass function assignment, 20 classes of corrosion types were defined according to the percentage of material loss from the nominal thickness. The MFECT inspection of section D was used to generate the probability mass functions, and the results of the DS fusion on section D are shown in Figure 4. It was found that the areas near the rivets did not follow the same distribution as the other measurements, and these pixels were excluded by applying a preprocessing step to the MFECT images. The square blocks in the images show the positions of rivets and all values in these areas are set to zero. It is important to note that section D, which was used to generate the distributions for the fusion rule, contained very little second layer corrosion. This is one drawback of using service-retired specimens for "training"; it is difficult to know if the desired range of corrosion on each layer is available in the specimen until teardown. Furthermore, it is for this reason that the fusion method was not tested for its ability to discriminate between multiple layer damage. Figure 5 shows the results of section C. There is little corrosion on the bottom layer. To evaluate the classification results, rate of classification is often adopted. However, this criterion is not suitable for this application. For example, misclassifying type 5 to type 6 will be treated the same as misclassifying type 5 to type 16. The difference between these corrosion classes is the percentage of material loss. Obviously, the former "misclassification" is better than the latter one. Figure 6 gives the difference images between X-ray thickness map and fused results (thickness maps) of section D and C. The mean errors for these two images are 0.0018 and 0.0019 respectively. This value is approximately 4% of the single layer thickness.

X-ray thickness map: top layer (left) and bottom Layer (right)
1:

Fusion results of 5.5 and 8kHz for top layer (left) and bottom layer (right)

JLJL
Fusion results of 17 and 30kHz for top layer (left) and bottom layer (right)

0.005

0.01

0.015-'

0.02

0,025

0.03 0.035

0.04 0.045

FIGURE 5. Image fusion results for section C. (Grayscale shows the thickness and the unit is inch.)

582

FIGURE 6. The difference between X-ray thickness map and fused results of section D 1st layer (left) and section C 1st layer (right).

DISCUSSION When pixel level data fusion is considered, an affine transformation that registers multi-sensor images should be performed in advance (see, for example [12]). Fortunately, for MFECT techniques, the discrete frequencies can be captured in one scan so the registration process can be omitted. However, to find the relationship between the material loss and the voltage values of MFECT measurement, there is still the need to match the X-ray thickness maps and MFECT images. The problem is that X-ray thickness map is of much higher resolution. The resizing of X-ray map to fit MFECT images may introduce error and it would be better to perform a low-pass filtering before using it. Observing the material loss and MFECT measurement, we cannot find any linear or near linear relation between them as might be derived from using simple calibration specimens. This demonstrates that without dealing with the effects of liftoff and interlayer gap, use of calibration data may not lead to a correct evaluation result. In our experiments, we used the distribution map of section D as a priori knowledge, and tested the resulting algorithm using section C. This works well for both section C and D; however, if there are too few pixels in certain data bins, the derived distribution curves may be inaccurate. This will affect the mass function assignment especially when the data to be evaluated has more pixels in this range. To solve this problem, more data should be collected to build the distribution map. The results for the first and second layers are obtained by fusing high and low frequency pairs respectively. It is obvious that fusion of the higher frequency images achieved a better result for the first layer. Any combination of low and high frequency images does not improve the results for evaluating first layer corrosion. This implies that data fusion does not assure good results. When complementary information is available, an effective fusion operation will be helpful. The crucial step is the mass function assignment. There are various approaches for this and it largely depends on the application itself. The mass function will be more reliable if a data-driven and objective process is employed.

583

CONCLUSION

In this paper, a data fusion scheme based on Dempster-Shafter theory is presented for fusing multi-frequency eddy current data. The results of multiple frequency eddy current inspections of hidden corrosion in a multilayer lap joint are quantified through this proposed approach. The results can then be used as the input to structural analysis models. Further improvement to the results will be attempted by obtaining more "training" data from service retired specimens in order to better model the relationship between the eddy current signals and the thickness of the layers of the joint. Other NDI techniques could also be used, and results fused with the MFECT results for improved quantification and reliability.
ACKNOWLEDGEMENTS

Funding of this work is provided by National Research Council Canada and Department of National Defence Canada, AVRS.
REFERENCES

1.

2. 3.
4.

5. 6. 7.
8. 9.
10.

11.

12.

Thompson, J.G., "Second Layer Corrosion Detection Using Dual Frequency Eddy Current Techniques", Presented at 1992 ATA Nondestructive Testing Forum, Cincinnati, Ohio 25-27 August 1992. Hagemaier, D.J., Nguyen, K., Materials Evaluation, 52, No.l, 91 - 95 (1994). Liu, Z, Tsukada, K., Hanasaki, K. and Kurisu, M., Research in NDE, 11,165177(1999). Yim, J., "Image Fusion Using Multiresolution Decomposition and LMMSE Filter", Ph.D. Thesis, Iowa State University, 1995. Gros, X.E., NDT Data Fusion, Arnold, Great Britain, 1997. Forsyth, D.S. and Komorowski, J.P., in SPIE Proceedings, Vol.3994, pp47-58 (2000). Fahr, A., Forsyth, D.S. and Chapman, C.E., "Survey of Nondestructive Evaluation (NDE) Techniques for Corrosion in Aging Aircraft", LTR-ST-2238, NRCC (26 October 1999). Hamid, R., "An Experimental Data Fusion Model for Multisensor Systems", Ph.D. Thesis, New Mexico State University, 1989. Alan P. B., NDE Reliability Data Analysis, Metals Handbook Volume 17: Nondestructive Evaluation and Quality Control (9th ed.), ASM International, 659701 (1988). Forsyth, D.S., Fahr, A., in Review of Progress in QNDE, Vol. 19B, eds. D. O. Thompson and D. E. Chimenti, Plenum, New York, pp 2159-2166 (2000). Solaiman, B., Pierce, L.E. and Ulaby, F.T., IEEE Tran. Geos. Remo. Sens. 37, No.3,ppl316-1326(1999). Liu, Z., Forsyth, D.S., Registration of Multi-modal NDI Images for Aging Aircraft, To be published in Research in NDE.

584