Abstract

1. Introduction

1. Power stations - nuclear and conventional power plants.
2. Metal industry - steel producers, steam and pressure vessel construction for the inspection of cracks, defects, and any other flaws and their characterization, fatigue estimation, quality assurance, wall thickness and coating thickness testing, determination of hardness etc.
3. Petro-chemical industry.
4. Transportation - railways.
5. Food industry - Quality assurance of food products.
6. Medical sciences.
7. Civil engineering - inspection of concrete structures, bridges, infrastructure due to aging problem.
8. Aircraft - fatigue estimation in aircraft surface and other parts.
9. Pipe inspection - Inspection of pipes and piping systems in industrial plants. The pipes are used for carrying oils, gases, waters, milks etc.
10. Others.

1. Visual
2. Magnetic
3. Ultrasonic
4. Acoustic
5. Radiography
6. Eddy current
7. X-ray.

II. Configuration of planar electromagnetic sensors

(a) Meander type	(b) Mesh type
Fig. 1 Configuration of planar electromagnetic sensors
Fig. 2 Structure of the sensor	Fabricated meander and mesh type sensors

III. Finite Element Model Formulation

So, eq (1) can be written as

A finite element model formulation has been carried out to solve the above equations with the necessary boundary conditions for the calculation of transfer impedance of the sensor. Fig. 4 shows a part of the model of the mesh-type sensor used for the finite element analysis. FEMLAB, a finite element analysis software package configured around Matlab has been used for the field analysis [14]. A 3-D quasi-statics model of combined conducting, magnetic and dielectric materials has been used for the analysis. The advantage of this model is that the transfer impedance of the sensor can be computed irrespective of the operating region with respect to operating frequency and subdomain characteristics. The same model can take care the both magnetic diffusion equation as well as high frequency microwave operation. The governing PDE (partial differential equation) for this model is given by

(4);

where v is the velocity of the coil. The model gives Electric potential, V, magnetic potentials, x-component, A_x, y-component, A_y and z-component A_z as the output variables. The output parameters are used for the calculation of desired variables.

Fig. 4 Finite element model of the sensor for the calculation of transfer impedance

The post-processing from the finite element analysis are used for the calculation of necessary parameters, the main parameter in this case is the transfer impedance. The transfer impedance is calculated for varying operating parameters. For each parameter the model is run separately. A few results are explained here. Fig. 5 shows the variation of the real part of the transfer impedance with the near-surface conductivity for different values of lift-off. It is seen that with the increase of the conductivity the real part of the transfer impedance is decreased. The operating frequency in this case is kept at 500 kHz.

Fig. 5 Variation of the real part of transfer impedance with conductivity at 500 kHz

Fig. 6 shows the variation of the imaginary part of the transfer impedance as a function of the near-surface conductivity with various values of lift-off at an operating frequency of 500 kHz. The effect of conductivity on the reactive part of transfer impedance is appreciably less compared to the resistive part.

Fig. 6 Variation of the imaginary part of transfer impedance with conductivity at 500 kHz

Figs. 7 and 8 show the variation of real and imaginary part of the transfer impedance with lift-off of the sensor for different values of conductivities of the near-surface. It is seen from Fig. 7 that the real part decreases with the increase of lift-off and also the value of the real part of the transfer impedance is less for higher values of conductivity. It is seen from Fig. 8 that the imaginary part of the impedance increases with the lift-off and the value of the imaginary part of the transfer impedance is actually lower at higher values of conductivities. The operating frequency for this data is the same of 500 kHz as in the earlier case.

Fig. 7 Variation of the real part of transfer impedance with lift-off at 500 kHz

Fig. 8 Variation of the imaginary part of transfer impedance with lift-off at 500 kHz

There can be many other characteristics depending on the parameters of interest. For the inspection of material properties, the estimation of conductivity, permittivity and permeability are the most important parameters. During measurement the distance between the top surface of the metal to the outer surface of the sensor i.e., the lift-off can change a lot, in spite of that the measurement of conductivity and permeability should not be affected.

In order to determine the conductivity from the measured transfer impedance data using the values obtained from the Figs. 5 to 8, the grid system as is shown in Fig. 9 is generated. The grid system, FIRST reported by Goldfine [1, 2, 3] is obtained by plotting the imaginary part of transfer impedance against the resistive part of it for varying values of conductivities and lift-off.

Since the generated grid system is obtained off-line, the correspondence of the calculated data from the finite element model and the experimentally obtained data should be very close. If the calculated values are widely different from than that of measured values a correction factor is to be introduced in the calculation. In order to get more accurate estimation from the grid system a grid can be generated from the experimentally measured data. A lot experimentally obtained data are required for generating the grid.

Fig. 9 Conductivity versus lift-off grid system obtained from finite element model

Once the grid system is available and the real transfer impedance is measured, the next step is to plot the transfer impedance on the grid system as is shown in Fig. 10. The plotted data lies inside a small grid with four nodes. The conductivites and lift-off of the four neighboring node points are known. So the output parameters corresponding to measured data are obtained by an interpolation technique. Table 1 shows some calculated values obtained from Fig. 10.

Fig. 10 The estimation of near-surface properties from the grid and measured data

After obtaining the system properties the next task is to correlate this data to obtain the necessary information for example, the fatigue of metal. There is a degradation of the conductivity with age of the metal. If there is a crack or some other kind of damage taken place just inside the metal, there will be a sharp drop of the conductivity values which can be very easily detected. In order to accurately predict the remaining life of the system the measurement system should have a lot of such data obtained from the mechanical testing of the particular materials and stored in computer.

One important parameter for this measurement system is the selection of operating frequency. Since the skin-depth decreases with the increase in frequency, in order to determine the defect in inner surface of the material, the higher value of the operating frequency is restricted. Fig. 11 shows the variation of skin-depths as a function of frequency for a few metals. It is seen from Fig. 11 that for titanium to inspect a defect at a depth of 0.5 mm the operating frequency should not be more than 500 kHz. If the crack lies at a depth of more than 0.5 mm the operating frequency should be lower so that flux enters more than the required depth and an appreciable change of flux takes place due to the defect.

Fig. 11 Variation of skin-depth with frequency

IV. Applications of planar electromagnetic sensors

(a) Inspection of printed circuit board

The first target application of planar electromagnetic sensors was to inspect the defect of the mother printed circuit board of a pentium processor as shown in Fig. 12 [7, 8, 9]. The printed circuit board has got many long conductors. In order to generate a magnetic field to be linked by the conductors on the PCB, meander type exciting coils have been chosen. The sensing coil may be of the meander type, mesh type and/or a Figure-of-Eight type. The operating principle of the sensing system is based on the flow of eddy current. Due to this it is also known as eddy current testing (ECT) technique. In the absence of a defect there is no voltage available across the terminals of the sensing coil while the presence of any defects will be manifested as output voltage across the terminals of the sensing coil.

Fig. 12 The target PCB (PCB of pentium processor)

The operating frequency can be varied between 1 MHz to 5 MHz for this application. The actual operating frequency is to be chosen based on the depth of the PCB wire. The experimental set-up used for this application is shown in Fig. 13. The voltage across the sensing coil is measured with the help of a Lock-in amplifier and is analyzed for the existence of any defects in the PCB. Only the same frequency component of the sensing voltage is considered. Fig. 14 shows a typical output across the sensing coil when the PCB has some defect. It is seen there is an offset in the sensing voltage which should be taken care of by an additional electronics circuitry.

Fig. 13 The experimental set-up for PCB inspection

Fig. 14 Amplitude of the voltage across sensing coil with a defect in the PCB

With the development of high density PCBs for the modern computer, the testing system should be able to cope up with the latest development. In order to address this problem the sensing coils are now made up of GMR sensors. Multiple sensors are used so that the measurement can be done once.

(b) Estimation of near-surface material properties

The use of the planar type sensing system is extended for the evaluation of near-surface properties such as conductivity, permeability etc and can also be used for the inspection of defects in the near-surface materials. It is impossible to test the mechanical strength, fatigue etc. for a part of a product in working condition. There are different causes such as environmental stress, corrosion, thermal treatments, thermal aging, irradiation embrittlement etc. for which the material gets degraded. Material degradation of nuclear power plant facilities (as well as steam generator tubes in power plants, the aircrapt's outer surface) has been a matter of concern, which includes fatigue, neutron irradiation embrittlement of ferrite steels and thermal embrittlement of duplex stainless steels. Detecting these problems by nondestructive methods as early as possible for prevention of a possible accident will contribute greatly to the improvement of the operation of nuclear power plants as well as securing the reliability.

There can be unlimited applications of this technique but a few are only listed here which includes pre-crack fatigue assessment, cumulative fatigue monitoring prior to crack formation, plastic deformation and residual stress monitoring, crack detection and characterization, new material characterization, age degradation monitoring, coating thickness and property characterization, quality control for processes such as heat treating, shot peening/burnishing and curburization etc.

It has been reported that the physical properties of the material undergo deterioration due to fatigue and aging [1, 2, 3, 4, 5, 6]. In order to avoid the unforeseen accidents it is necessary to do the preventive maintenance to know the changes of the physical properties and to assess the actual state of the material. The problem of bringing the test sample/specimen has forced the testing technique to be compatible with the test object. The testing equipment has to be taken into site and as a result it should be flexible and compact. The measurement of impedance of coil placing on the tested object and studying the change of impedance with the help of an inverse approach for the determination of material is one conventional approach. It needs processing time which makes it difficult to get real on-line measurement. Planar type meander coil has been used for the on-line measurement of near-surface conductivities and coating properties by Goldfine[1, 2, 3]. Goldfine[1, 2, 3] proposed a very simple approach of inversion technique utilizing off-line generated grid model which makes it possible to determine the desired parameters in real time.

The use of meander type sensing coils for the evaluation of near-surface materials has been reported in [1, 2, 3, 10]. Sometimes it is quite difficult to inspect a crack parallel to the meander coil a shown in Fig. 1a, as the eddy current will not be affected due to the alignment of the crack parallel to the exciting coil. Under this situation the inspection using meander type sensor needs to be carried out two times and in orthogonal direction to each other to preserve all information and provide better sensitivity[15]. In some application this problem of taking measurement twice can be overcome by using mesh type sensor [11 - 13, 16 - 19].

Fig. 15 A typical damage due to thermal stress

The operating principle of the estimation of near-surface material properties is based on the measurement of transfer impedance. The transfer impedance of the sensors is defined as the ratio of the voltage across the sensing coil to the current of the exciting coil. The transfer impedance of the sensor is measured by impedance analyzer and both the amplitude and the phase of the transfer impedance are used in the estimation.

The transfer impedance is a complex function of many parameters such as conductivity, permeability, permittivity of the near-surface materials, frequency, lift of the sensors. Since the measured transfer impedance is dependant on many parameters such as conductivity, permeability, permittivity, frequency, coil pitch etc. in a very complex way, it is mathematically very complicated to determine material properties from the measured impedance data. To determine the system properties from the measured impedance is known as inversion problem and there are a lot of research papers reported on this topic. In this paper an off-line generated grid system has been utilized to determine the material properties. Goldfine [1, 2, 3] has proposed the grid based inversion method for the first time which doesn't require complex mathematics once the grid is constructed.

The measurement grids provide a generalized and robust approach to a wide range of applications, and permit rapid adaptation to new applications with varied material constructs and properties of interest. The data of the grid system may be obtained either from analytical solution or by some other numerical method. In this paper the data for the grid are obtained from the finite element calculation of transfer impedance of the sensor.

To reduce the error, the measurement requires a single calibration measurement in air. For the estimation of surface properties the measured transfer impedance is plotted on the grid system as shown in Fig. 16. The actual measured impedance lies inside four neighboring nodes which has got fixed conductivity and lift-offs. By interpolation technique the two parameters, in this case conductivity and lift-off are determined. In order to utilize the grid more effectively the range of the system parameters can be restricted. In Fig. 16, the range of the conductivity is between 1E7 S/m to 6.0E7 S/m. If the material is like aluminium the range of conductivity can be restricted between 3.2E7 S/m to 4.2E7 S/m. So Fig. 16 can be used as a FIRST grid to get the initial values and then a more refined grid can be used for more accuracy.

Fig. 16 Estimation of surface properties

An alternative of the grid system is to adopt a neural network aided estimation [20 - 26]. Fig. 17 shows a neural network aided model developed for the estimation of near-surface material properties. The network is to be trained with the off-line generated data or the actual measured data. The output can be more than two depending on the requirement. If the system is trained with more data it will give much better accurate estimation.

Fig. 17 Neural network aided property estimation

Table 1 shows a comparison of the estimations obtained from the grid system and the neural network method. It is seen that the neural network aided estimation gives more smooth result compared to grid system but the level of maximum error is quire large.

Table 1 Comparison of results

(c) Quality estimation of saxophone reeds

Another novel application of planar electromagnetic sensors is the quality inspection of saxophone reeds. A saxophone reed as is shown in Fig. 18, is a small piece of bamboo that is attached to the mouthpiece of a saxophone as shown in Fig. 19. When the player blows into the saxophone the reed vibrates creating sound. The reed is therefore, in part, responsible for the tone and ease of use of a saxophone. There is nothing more frustrating for a saxophone player than playing on a bad reed. Reeds wear out after a few weeks of playing and must be replaced.

The problem with reeds is that currently the quality is very variable: in a box of ten, three or four reeds are found to be 'bad' when played and are discarded. The reed being quite expensive ($10 per reed) any improvements in quality control would be very worthwhile. The parts of the reed that affect the quality of the reed the most are the 'vamp' and the 'tip'. The planar electromagnetic sensor being flexible can be easily placed on the sloppy surface of the reeds and the experiments were conducted. Conclusions are drawn from the experimental results such as the differences observed between "good" and "bad" reeds.

Fig. 18 Saxophone reed and planar electromagnetic sensor

Fig. 19 A typical saxophone

Fig. 20 shows the experimental set-up configured around a network analyzer, Agilent 8712ET, maximum operating frequency of 1.3 GHz. Table 2 shows the list of the saxophone reeds used for the investigation. Based on the experience the reeds are categorized with a number indicating the overall quality. The tenor reeds#1, 2 and 4 are categorized as bad reeds and all other remaining reeds are considered as good reeds. The transfer impedance of the sensor has been measured placing the sensor on each and every reed. Fig. 21 shows the variation of the phase angle of the transfer impedance as a function of the operating frequency for all the reeds under investigation. It is seen that the nature of the phase waveforms is almost similar except at the frequency range between 500 MHz to 800 MHz.

The so-called bad reeds show the peak phase at around 579 MHz where as the good reeds shows the peak phase at around 510 MHz. The peak phase for the bad reeds is higher in magnitude compared to the good reeds. This phase information along with the magnitude can be effectively utilized to decide the quality of the reeds and the decision of discarding the bad reeds can be taken without any difficulty.

The reason for different phase is due to different effective permittivity of the reeds. The quality of the reed is basically dictated by the permittivity. The different permittivity of the reed will result in different impedance of the sensor. The frequency range between 500 MHz and 800 MHz is the range in which the permittivity of bad reed differs appreciably compared to good reeds. The permittivity being a function of operating frequency may not differ much for the complete frequency range. To distinguish the bad reeds from good reeds an operating frequency between 500 MHz to 800 MHz can be carefully selected. So the initial investigation has given promising results and further work is carried out to predict more about the qualitative aspect of the reeds.

Fig. 20 The experimental set-up for saxophone inspection

Table 2 The saxophone reeds under inspection

Fig. 21 Phase versus frequency characteristics of saxophone reeds

(d) Preliminary investigation with milk

The possibility of employing planar type electromagnetic sensors for the estimation of properties of dielectric materials such as milk, butter, cheese, yogurt etc. has also been investigated for the purpose of composition analysis of dairy products.

A lot of experiments have been carried out and a few results are presented here. Figs. 22 and 23 show the screen shots of magnitude and phase of impedance of the sensor taken from the network analyzer. These two characteristics are very similar to the characteristics of any microwave sensors in an empty container. It is seen there are a few resonances taking place in the frequency range 300 kHz to 1.3 GHz. This is due to the materials characteristic on which the sensor is built and the effect of other parasitics such as the instrumentation involved, cable and noises. Fig. 24 shows the variation of impedance with frequency for different products such as water, juice, milk, oil and yogurt. It is seen that the impedance is different for different products.

Fig. 25 shows the variation of impedance with frequency of milk with different lift-off. In order to examine the usefulness of this sensor to be used in dairy industry, the milk samples with a known percentage of fat content have been prepared in laboratory and the effect of fat content has been studied. Fig. 26 shows the variation of impedance with frequency for different percentage of fat content at a lift-off of 5 mm. The nature of the impedance characteristics of figures 25 and 26 looks very similar even though the magnitude is different at different frequencies. This is due to the fact that figure 25 is for ordinary milk which has got almost 5% fat. The change of lift has some effect on the impedance magnitude as both the inductive part as well as the capacitive part of the sensors is dependent on the lift-off.

The magnitude and phase of the transfer impedance is to be used for the determination of the effective permittivity of the product under test. The effective permittivity for a complex combinations of materials and geometries within a particular volume is equivalent to the permittivity for a single homogeneous material that would produce the same electromagnetic characteristics for the same volume. It is seen from the experimental result that a different amount of fat has a considerable effect. The operating frequency is one of the most important parameters and any one operating frequency doesn't give the best effect. So the selection of operating frequency is very important. Fig. 27 shows the results for an operating frequency of 750 MHz at a lift-off of 5 mm. The variation of impedance with different percentage of fat content has an appreciable effect on the impedance of the sensor.

Fig. 22 The impedance magnitude of the sensor as a function of frequency

Fig. 23 The phase of the impedance of the sensor as a function of frequency

Fig. 24 The impedance magnitude as a function of frequency for different dielectric materials	Fig. 25 The impedance magnitude as a function of frequency for milk with different lift-off
Fig. 26 The impedance magnitude as a function of frequency for milk with different fat content at a lift-off of 5 mm	Fig. 27 Variation of impedance with percentage of fat content at 750 MHz with 5 mm lift-off

The experimentally obtained data are used to determine the composition of dairy products by using some computational technique. The technique can be based on single-frequency or multiple frequency excitation. A very simple technique in single frequency excitation system is polynomial curve fitting. Only the magnitude of the impedance has been used and the fat content has been measured which gives quite accurate result. In order to determine the composition of milk product a multi-frequency excitation method is proposed here. Let us assume that a product has got n components of X_n percentage each so that . The experimental results are collected for n times at n different frequencies. From each measurement the effective relative permittivity is computed utilizing the grid approach. The grid which is to be used for this purpose should have variables like permittivity and lift-off. From the measured impedance and using both the magnitude and phase the permittivity for each operating frequency is to be determined. The effective permittivity is then expressed as ; er1,1 , er2,1 … ern,1 are the relative permittivities of each component at frequency #n and are known. So after n reading the following matrix is obtained.

The matrix representation can be represented as A X = b in which the b matrix is obtained from the experimentally measured resultant permittivities. A matrix are known from the dielectric permittivity vs frequency characteristic of each component. Usually the dielectric permittivity is a function of temperature and operating frequency. So the parameters of the A matrix can be obtained and the X matrix can be calculated. This method is under investigation and it is expected to obtain good results.

V. Conclusions

V. References

1. GOLDFINE, N.J., "Magnetometers for improved material characterization in aerospace application", Material Evaluation, March 1993, pp. 396-405.
2. GOLDFINE, N.J., CLARK, D., and LOVETT, T., 'Material characterization using model based meandering winding eddy current testing (MW-ET),' EPRI Topical workshop: Electromagnetic NDE applications in the Electric Power Industry, Aug 21-23, 1995, Charlotte, NC.
3. GOLDFINE, N.J., 'Conformable, meandering winding magnetometer (MWM) for flaw and material characterization in ferrous and nonferrous metals,' ASME Pressure Vessels and Piping Conference, Proceeding on International Advancement in PVP Technology, July 27-31, 1997, Orlando, FL, USA.
4. BI, Y., GOBINDARAJU, M.R., and JILES, D.C., 'The dependence of magnetic properties on fatigue in A533B nuclear pressure vessel steels,' IEEE transactions on Magnetics, 1997, 33, (5), pp. 3928-3930.
5. SHI, Y., and JILES, D.C., 'Finite element analysis of the influence of a fatigue crack on magnetic properties of steel,' Journal of Applied Physics, 1998, 83, (11), pp. 6353-6355.
6. BI,Y., and JILES, D.C., 'Dependence of magnetic properties on crack size in steels,' IEEE transactions on Magnetics, 1998, 34, (4), pp. 2021 - 2023.
7. YAMADA,S., KATOU,M., IWAHARA, M., and DAWSON,F.P., "Eddy current testing probe composed on planar coils", IEEE transactions on Magnetics, vol. 31, no. 6, pp. 3185-3187, 1995.
8. YAMADA,S., KATOU,M., IWAHARA, M., and DAWSON,F.P., "Defect images by planar ECT probe of Meander-Mesh coils", IEEE transactions on Magnetics, vol. 32, no. 5, pp. 4956-4958, 1996.
9. YAMADA,S., FUJIKI,H., IWAHARA,M., MUKHOPADHYAY, S.C., and FDAWSON,F.P., "Investigation of printed wiring board testing by using planar coil type ECT probe", IEEE Transactions on Magnetics, vol. 33, no. 5, Sep. 1997, pp. 3376-3378.
10. MUKHOPADHYAY,S.C., YAMADA, S., and IWAHARA,M., "Investigation of near-surface material properties using planar type meander coil", JSAEM Studies on Applied Electromagnetics and Mechanics, Vol. 11, pp 61-69, 2001.
11. MUKHOPADHYAY,S.C., YAMADA,S., and IWAHARA, M., "Evaluation of near- surface material properties using planar mesh type coils with post-processing from neural network model", International journal on Electromagnetic Nondestuctive Evaluation (VI), 2002, IOS Press, Vol. 23, pp 181-188.
12. MUKHOPADHYAY,S.C., YAMADA, S., and IWAHARA,M., "Inspection of electroplated materials - performance comparison with planar meander and mesh type magnetic sensor", International journal of Applied Electromagnetics and Mechanics, vol. 15, pp 323-329, 2002.
13. MUKHOPADHYAY,S.C., "Quality inspection of electroplated materials using planar type micro-magnetic sensors with post processing from neural network model", IEE Proceedings - Science, Measurement and Technology, vol. 149, no. 4, pp. 165-171, July 2002.
14. FEMLAB, Finite Element Software, COMSOL, 1994-2003.
15. Taniguchi, T., Nakamura, K., YAMADA, S., and IWAHARA,M., "An Image Synthesis Method for Eddy Current Testing Based on Extraction of Defect Orientation", Digest of IEEE Intermag conference, paper no. GC-03, Amsterdam, May 2002.
16. MUKHOPADHYAY,S.C., YAMADA,S., and IWAHARA,M., "Experimental Determination of Optimum Coil Pitch for a Planar Mesh Type Micro-magnetic Sensor", IEEE Transactions on Magnetics, Vol. 38. N0. 5, pp 3380-3382, Sep. 2002.
17. MUKHOPADHYAY,S.C., YAMADA, S., and IWAHARA,M., "Optimum coil pitch selection for planar mesh type micro-magnetic sensor for the estimation of near-surface material properties" JSAEM series on Applied Electromagnetics and Mechanics, Vol. 14, pp. 1-9, January 2003.
18. MUKHOPADHYAY, S.C., "A Novel Planar Mesh Type Micro-electromagnetic Sensor: Part I - Model Formulation", IEEE Sensors journal, Vol. 4, No. 3, June 2004, pp 301-307.
19. MUKHOPADHYAY, S.C., "A Novel Planar Mesh Type Micro-electromagnetic Sensor: Part II - Estimation of System Properties", IEEE Sensors journal, Vol. 4, No. 3, June 2004, pp 308-312.
20. RATNAJEEVAN, S. and HOOLE, H.: 'Artificial neural network in the solution of inverse electromagnetic field problems,' IEEE transactions on Magnetics, 1993, 29, (2), pp. 1931-1934.
21. ENOKIZONO, M., TODAKA, T., AKITA, M. and NAGATA, S.: 'Rotational magnetic flux sensor with neural network for non-destructive testing,' IEEE transactions on Magnetics, 1993, 29, (6), pp. 3195-3197.
22. CARLO, F.M. and CAMPOLO, M.: 'Location of plural defects in conductive plates via neural networks,' IEEE transactions on Magnetics, 1995, 31, (3), pp. 1765-1768.
23. ENOKIZONO, M., TSUCHIDA, Y. and CHADY, T.: 'Crack size and shape determination by moving magnetic field type sensor,' IEEE transactions on Magnetics, 1998, 34, (4), pp. 1252-1254.
24. GLORIEUX, C., MOULDER, J.C., BASSART, J. and THOEN, J.: 'The determination of electrical conductivity profiles using neural network inversion of multi-frequency eddy-current data,' Journal of Physics, D, Applied Physics, IOP Publishing Ltd., UK, 1999, 32, pp 616-622.
25. CHADY, T., ENOKIZONO, M. and SIKORA, R.: 'Neural network models of eddy current multi-frequency systems for nondestructive testing,' IEEE transactions on Magnetics, 2000, 36, (4), pp. 1724-1727.
26. CHADY, T., ENOKIZONO, M., SIKORA, R., TODAKA, T. and TSUCHIDA, Y.: 'Natural crack recognition using inverse neural model and multi-frequency eddy current method,' IEEE transactions on Magnetics, 2001, 37, (4), pp. 2797-2799.