Defect characterisation in laminar composite structures using ultrasonic techniques and artificial neural networks

Abstract

This paper presents an approach to defect detection and characterisation in ultrasonic inspection of laminate composite panels. A set of features coupled with gates was identified along with a method for sub-dividing and thresholding the ultrasonic data, which removes most of the location specific information from the defect data thus increasing the generalisation capabilities of the defect classifier. Validation results obtained from independent defect data indicate that the performance of the presented data description tools coupled with an artificial neural network classifier was able to correctly locate and classify defects at different depths. It was concluded that a structured approach to the pre-processing of ultrasonic testing data, combined with the selective feature extraction for artificial neural network classification, greatly reduces the requirement for artificial neural network training data. Furthermore, it allows for improved performance across a variety of panel geometries.

Keywords

Signal processing feature extraction ultrasonic testing artificial neural network defect classification laminate composites

Introduction

A large number of multi-layered composite structures exist in practice. These include fibre-reinforced composite materials, aircraft structures, biological tissues, thin films, modern electronic systems and semiconductors.¹ Composite materials have found an increasing number of applications² due to a combination of low weight, high strength and stiffness. This is predominantly so in aircraft structures as the weight reductions result in reduced running costs and greater payloads.³ A constraint in large-scale composite use relates to the inspection for component integrity.⁴ Besides, the main problem is high attenuation of the ultrasonic signal due to scattering and multiple reflections inside the material caused by different acoustic impedance of the layers or plies or resin/adhesives placed between the layers.⁵ Hence, several researchers proposed that non-destructive examination is particularly suitable for this purpose, as the composite may vary in composition from point to point and removal of samples for destructive analysis would be neither permissible nor particularly useful.^6–11

Of all the techniques available for non-destructive testing (NDT), ultrasonic testing (UT) has shown itself to be one of the most versatile and informative.^4,12 Composite structures offer several unique challenges for UT over metals, most noticeably, high signal attenuation and strong ply/resin boundary reflections. The unique properties of laminar composite materials include a reasonably well-defined material structure leading to an almost discrete set of defect locations particularly for foreign objects (FOs) which occur in the resin bonding layer between plys. This layered structure can be exploited in the development of a defect classifier so that the type and location of defects can be determined using a minimal set of classifier calculations. Also, the data gates allow a trained artificial neural network (ANN) classifier to inspect any similar composite panel regardless of thickness.¹

Teti and Alberti¹⁰ investigated ultrasonic non-destructive evaluation of defects in carbon-fibre-reinforced plastic laminates, displaying different quality levels according to their fabrication cycle. UT, mechanical testing and structural examinations were carried out to quantitatively characterise the carbon-fibre-reinforced plastic laminates and their quality.¹⁰ D’orazio et al.¹¹ addressed the problem of automatic inspection of composite materials using an ultrasonic technique. The authors developed a normalisation procedure to suppress with a uniform step the samples which are under a selected threshold value.¹¹ Consequently, an evaluation was made of the number of signal points that have to be eliminated to reduce the signal length to that obtained from a minimum ply thickness. The reduced signal maintains the same shape and peak as the original signal.¹¹ Diamanti et al.¹³ presented a method for detecting impact damage and repair-based cracks in composite materials using low-frequency Lamb waves that is only effective in plates for structure breakages. In this research, the authors were able to detect and locate defects of critical size (3–20 mm).¹³

Automated signal classification is becoming increasingly popular in many commercial applications, including non-destructive evaluation. Over the past few decades, a substantial interest in the development of feature extraction techniques has been kindled largely by the requirement for addressing pattern recognition, defect detection and image processing problems, where features constitute inputs to a classifier. Selecting features that can effectively classify patterns is often a non-trivial task in many applications.¹⁴ Once a set of optimal features are chosen, a suitable classifier is used to classify the signals. ANNs have been studied extensively in the literature^15,16 and shown to perform well in many ultrasonic NDT classification problems.^11,17–22 Two significant issues with ANN for defect detection and classification are data preparation and generalisation. Data preparation includes ultrasonic signal processing and feature extraction.²³ Generalisation refers to achieving performance in situations not specifically trained for, such as variation in defect depth within the composite panel being inspected.

This research attempts to address the problem of detecting and classifying FO type defects within multiple laminate carbon-fibre-reinforced polymer (CFRP) structures. In validating the proposed approach, the defect depth, number of defects within a single component and panel thickness of the test panels were varied from the situations presented to the classifier at training.

Panel preparation and data acquisition

In this research, two carbon-fibre-reinforced polymer panels were analysed. Both panels were made of CYCOM 970 fabric with a nominal ply thickness of 0.22 mm. Each panel contained three sections of 2 mm (9 plys), 4 mm (19 plys) and 6 mm (28 plys) with a step transition between each section. Both panels contained two types of FOs, a bonding Mylar 4MMFR2-36 film of thickness 0.002′′ and a non-bonding Teflon Airtech A4000R film of thickness 0.001′′. The first panel, which was referred to as the ‘Simple’ panel, was 300 × 175 mm. It contained defects after first ply, in the middle of the panel and before the last ply. Each of the three sections was similar in terms of defect layout and the schematic for the 2-mm-thick section was included in Figure 1(a). The second panel, referred to as the ‘Complex’ panel, was 520 × 350 mm in size. Some FOs were placed at the same depths of those in the simple panel. Other FOs were overlapping or included at different depths within the panel from the three depths previously stated. The schematic of the 2-mm-thick section was included in Figure 1(b).

Figure 1.

Panel schematics showing defect depth, type and location information. P1-4 mm (left) and P1.4-2.2 mm (right), BO: bonding defect type; NB: non-bonding defect type.

The ultrasonic scanning apparatus consisted of a 5-MHz, 0.25′′ diameter focussed transducer operating in pulse-echo mode in a water tank. Scanning was performed in incremental steps of 0.5 mm resulting in defect coverage of a factor of 12 scans within the diameter of the probe. The panel data presented here was scanned merely on the tool side of the panels. Each set of panel data was further sub-divided due to computer memory limitations. The simple panel data were reduced by 75% as only every second point in both the vertical and horizontal directions was kept as illustrated in Figure 2(a) and the complex panel data were physically larger and so every third point was kept, as shown in Figure 2(b).

Figure 2.

Down-sampling grid for (a) simple panel and (b) complex panel. A-scans kept during down-sampling marked with ✓.

The ultrasonic data were recorded using the UT win software package at a rate of 50 MHz, and the system gain was set so that the front wall echo (FWE) displayed no clipping or saturation. Distance Amplitude Correction (DAC) was disabled in the software as was FWE synchronisation.

Methodology

To assist with the analysis, synchronisation of the ultrasonic data at the FWE was performed using peak detection of the smoothed ultrasonic signal. Smoothing of the A-Scan signal assisted in reducing the number of peaks while identifying the key echoes. The optimum smoothing filter was determined from the following set: uniform filter, triangular filter, Gaussian distribution filter, square of the triangular filter, square root of the triangular filter and decaying exponential filter. The width of each filter was applied and tested on the range of width from one to half of the A-Scan width. All the filters were normalised so that the area under the filter summed to one. Two performance metrics were used to measure the quality of the smoothed A-Scans, percentage reduction in the number of peaks and the absolute difference between the smoothed and original signal. The optimal filter was the one which produced the fewest peaks whilst maintaining the greatest similarity to the original signal. Figure 3 shows the performance for each filter on the 2 mm panel data. Based on these investigations, the triangular filter with width nine was used.

Figure 3.

Performance matrices for smoothing of 2-mm-thick sample panel.

All A-Scans were trimmed; two ply thicknesses prior to the FWE and two ply thicknesses after the first back wall echo (BWE). The number of samples per ply was calculated as 7.12 using equation (1). The velocity of ultrasound was calculated to be 3.09 ± 0.08 mm/µs as discussed in

Samples per Ply = \frac{2 \times Sampling rate \times Average ply thickness}{Velocity of ultrasound in material}

(1)

DAC was implemented as a single exponential gain function applied across all A-Scans equally as illustrated in Figure 4. The median FWE and BWE peak amplitudes for all A-Scans in the 2-mm-thick panel were used to determine the exponential coefficients. The DAC curve was clipped at the FWE and BWE locations to limit the effects of post-BWE signals.

Figure 4.

DAC gain function for the 2-mm-thick sample panel. DAC: distance amplitude correction.

Data gates

Gates perform the role of segmenting ultrasonic data and filtering out insignificant responses. Gates consist of two descriptive parameters: amplitude or threshold; and also a range of times where the gates are active. In this work, gates are represented mathematically as shown in equation (2)

GateOutput (n, t) = {A - Scan (t) ifminGate (n)) \leq t \leq max (Gate (n)), and | A - Scan (t) | \geq Gate Threshold (n) otherwise

(2)

where [tmin, tmax] are the lower and upper limits on the gate position in the time direction. And GateThreshold is the threshold below which the A-Scan signal was ignored.

The time tmin was set to the median location of the FEW plus one ply thickness, and tmax was set to the median location of the BWE minus one and a half ply thickness. The offsets applied to the FEW and BWE were chosen to remove these echoes from future calculations. The BWE required a larger offset due to panel thickness variations and ultrasonic signal spread.

The gate threshold was determined using the average peak height within the range [tmin, tmax] for the 2 mm simple panel. As shown in Figure 5, the peak height is indicative of the presence of a defect, but ultrasonic variability means that more information was required for confident classification. In this work, the threshold (GateThreshold = 0.14) has been chosen as the lower limit at which defects occurred, so that all defect regions were presented to the classifier even though there is greater possibility for misclassification of the defect-free (DF) ultrasonic data.

Figure 5.

Box plots of maximum amplitude between the FWE and BWE for 2 mm thick sample panel.

Feature extraction

Feature extraction for various ultrasonic inspection applications has been well covered in the literature. Ultrasonic data have been described in the time domain;^17,23–27 frequency domain;^17,23–28 correlation, convolution and deconvolution domains;²⁹ and with wavelets.^{23,26,30–32} Making use of all of the features presented in the literature would lead to a large scale of redundant information and very slow computation times, and so a narrow selection was adopted based on the review by Lee.²³

In this work, three sets of features were selected to describe four signal groups. The signal groups were DF, bonding (BO) Mylar inclusion, non-bonding (NB) Teflon inclusion and loss-of-signal (no signal, NS).

The first feature set is referred to as ‘general’, made use of statistical descriptors of the entire A-Scan and of local descriptors of the BWE. The ‘general’ features were mean (rectified), standard deviation, skewness, kurtosis, BWE peak, BWE ToF, BWE area, BWE width, BWE rise time from 25% to peak and BWE fall time from peak to 25%. The second set is referred to as ‘local’ that created localised gates with a width of two ply thicknesses (14 samples wide) in the regions of strong echoes and extracted the features from the smoothed ultrasonic data within these gates. Strong echoes were defined as any peaks between the FWE and BWE that exceeded the gate threshold level of 0.14 as discussed in Methodology section. The local features were peak amplitude, peak ToF, standard deviation, mean (rectified), peak area, peak width, rise time from 25% to peak and fall time from peak to 25%. The third set is referred to as ‘mixed’ which combines the entire ‘local’ feature set with the mean (rectified) and kurtosis of the entire A-Scan. In all these cases, the feature C-Scans were normalised to the range [−1,1].

Classifier training

The training data for all of the classifiers were taken from the 2-mm-thick simple panel as illustrated in Figure 1(a). The location of the training data for the DF, bonding and non-bonding defects was as shown and consisted of 166 samples, 98 samples and 93 samples, respectively. The non-bonding and bonding inclusion data were distributed evenly between defects after the FWE, in the middle of the panel and before the BWE. The NS data were simulated by including 21 empty A-Scans. The 2-mm-thick simple panel contained a total of 12,509 DF signals, 447 non-bonding defect signals, 443 bonding inclusion signals and 99 instances of NS.

Based on earlier studies,^17,33 feed-forward neural network with a back propagation training algorithm was selected and implemented using the MATLAB program (The MathWorks, Inc.). Nevertheless, the disadvantage of this algorithm is that the search for the best solution is often trapped in a local optimum, making it difficult for the network to work well.³⁴ Several authors^35,36 have proposed training these networks with evolutionary search techniques as a good solution, but they have a drawback that training takes a long time. In order to address this issue, Cuéllar et al.³⁴ have proposed non-linear programming (BFGS³⁷ and Levenberg-Marquardt) algorithms that can be applied to train recurrent neural networks.³⁴

In this research, the network was trained with two hidden layers each with 2 to 20 neurons. Determination of number of neurons for the hidden layer is often achieved through experimentation. Several combinations of input and hidden layer neurons were investigated, and the mean square error value was evaluated. The lowest error value was obtained for neural network configurations that had 10 inputs for the ‘general’ feature set, 8 inputs for the ‘local’ feature set and 10 inputs for the ‘mixed’ feature set. All three configurations had three outputs organised using the well-known one vs. all scheme.^38,39 Four-fold cross-validation was used in training, and the best training results for each training algorithm were recorded in Table 1. It was observed that the ‘general’ feature set scored highest during the training stage. The data selection criteria for four-fold validation were chosen from every fourth point from the training set.

Table 1.

Summary of four-fold cross-validation training results for all artificial neural network training algorithms on the 2 mm simple panel data.

	‘General’ feature set training			‘Local’ feature set training			‘Mixed’ feature set training
	Data	HL1	HL2	Data	HL1	HL2	Data	HL1	HL2
bfg	97.64	6	18	94.69	16	14	96.38	10	16
cgb	97.77	18	12	94.52	20	8	96.81	8	18
cgf	97.14	14	20	94.48	20	16	96.61	6	0
cgp	97.62	10	14	94.74	20	16	96.66	4	20
gda	96.48	8	8	94.68	18	4	96.69	16	20
gdm	91.20	18	12	92.41	12	16	93.43	8	20
gdx	95.34	10	14	94.56	10	20	96.27	16	0
lm	98.43	8	6	94.95	18	20	97.11	18	12
oss	97.22	8	0	94.62	14	8	96.70	18	14
rp	97.64	6	18	95.00	20	20	96.89	16	10
scg	97.77	18	12	94.73	12	14	96.78	20	16

Post-classification image processing was performed to remove the occurrence of isolated classification results. The coarsest step size across either panel was 1.5 mm for the complex panels, and the smallest defects were the 0.5′′ × 0.1′′rectangle. All defects were represented across multiple adjacent pixels in the classification panel view image (i.e. C-Scan). As a result, any single point which does not have a minimum of three occurrences in any 3 × 3 neighbourhood was assumed to be in error and was set to the most frequently occurring value in that neighbourhood.

Results and discussion

Testing of the classifier was first performed on the remaining data within the 2-mm-thick simple panel, and in all cases, performance was a nearly perfect match based on the panel layout as illustrated in Figures 1(a) and 6(a). The next stage of testing followed the example provided in D’orazio et al.¹¹ The trained classifier was tested on thicker panels where the defects were located in the same configuration as used in training, specifically immediately after the FWE and before the BWE. Results were recorded in Table 2. The ‘general’ classifier performing well on the 2-mm-thick panel, but performance was not satisfactory for the thicker panels. The best ‘general’ performance was achieved by the ‘gdx’ training algorithm, which was the second worst performing during the four-fold validation training. In addition, the ‘general’ classifier tended to identify more defects than what was present in the sample and frequently classified the DF region as a bonding defect.

Figure 6.

Classification results using ‘mixed’ feature set and gdx trained classifier. Colours assigned: DF , NS , NB , BO .

Table 2.

Classification results for all classifiers on data from the simple panel.

	‘General’ feature set			‘Local’ feature set			‘Mixed’ feature set
	2 mm	4 mm	6 mm	2 mm	4 mm	6 mm	2 mm	4 mm	6 mm
Bfg	99.06	96.69	64.92	98.63	97.14	92.14	99.34	95.53	95.71
Cgb	98.96	21.21	62.28	98.42	96.28	91.69	99.37	93.74	95.52
cgf	98.66	89.78	9.80	98.54	98.22	92.03	99.48	94.89	95.59
cgp	99.04	94.33	41.94	98.64	97.85	91.27	99.38	91.84	95.57
gda	98.67	31.25	8.76	98.42	97.23	90.26	99.46	99.49	95.57
gdm	95.14	93.37	89.02	98.44	97.43	90.80	99.21	98.83	94.92
gdx	98.72	98.46	82.95	98.01	97.24	92.57	99.36	99.52	95.58
lm	98.46	62.01	7.25	98.51	97.70	93.42	99.11	89.51	95.42
oss	98.54	49.04	10.92	98.61	97.43	92.17	99.31	98.78	95.79
rp	98.84	90.75	13.94	98.58	97.50	93.02	99.38	91.51	95.63
scg	98.92	73.57	34.91	98.69	97.43	92.89	99.49	94.62	95.17

On the other hand, the ‘local’ approach performed well in locating defects regardless of the training algorithm used, yet this classifier assigned the incorrect class. In addition, this approach was unable to identify the NS defects due to the use of data gates for all features. This could be easily circumvented with a logical check for the presence of any ultrasonic echoes above the threshold value, FWE and BWE included. The ‘mixed’ approach appeared to contain the strengths of both ‘general’ and ‘local’ classifiers and outperformed the other two.

The performance results in Table 2 provide a measure of the percentage of exact matches between the classified output and the results template. To understand the results further, Table 4 shows the confusion matrix for the classifiers trained with the gdx algorithm alone. The confusion matrix provides a count of the number of signals assigned to a class compared to the class the signal was meant to be assigned to. A perfect classification result is represented by a confusion matrix with only entries on the diagonal.

Table 4.

Confusion matrices for all classifiers trained with gdx algorithm, tested on simple and complex panels.

Assigned class	Target class
	General				Mixed				Local
	DF	NS	NB	BO	DF	NS	NB	BO	DF	NS	NB	BO
2 mm S
DF	12,506	61	15	44	12,530	99	108	27	12,505	2	4	20
NS	0	38	0	0	0	0	0	0	0	95	0	0
NB	27	0	431	0	15	0	331	0	26	2	437	0
BO	25	0	1	399	13	0	8	416	27	0	6	423
4 mm S
DF	12,772	58	34	43	12,847	101	75	28	12,846	4	7	29
NS	0	43	0	0	0	0	0	0	0	94	0	0
NB	1	0	531	0	0	0	313	0	1	3	549	7
BO	75	0	6	498	1	0	183	513	1	0	15	505
6 mm S
DF	9810	359	95	18	11,414	623	69	17	11,410	356	5	20
NS	0	264	0	0	0	0	0	0	0	260	0	0
NB	0	0	278	0	8	0	192	0	10	7	458	126
BO	1654	0	99	475	42	0	211	476	44	0	9	347
2 mm S
DF	94	238	63	0	16,020	355	152	87	16,020	239	230	39
NS	0	115	0	0	0	0	0	0	0	115	0	0
NB	0	1	664	0	0	0	885	0	0	0	732	0
BO	15,927	1	345	796	1	0	35	709	1	1	110	757
4 mm S
DF	17,124	246	35	707	17,205	353	148	116	17,031	238	19	204
NS	0	107	0	0	0	0	0	0	0	115	0	0
NB	88	0	1036	7	0	0	639	0	169	0	951	32
BO	0	0	0	82	7	0	284	680	12	0	101	560
6 mm S
DF	14,667	484	364	460	14,667	1063	171	79	14,666	481	9	286
NS	0	579	0	0	0	0	0	0	0	582	0	0
NB	0	0	524	0	0	0	449	0	1	0	872	2
BO	0	0	0	290	0	0	268	671	0	0	7	462

DF: defect-free; NS: no signal; NB: non-bonding defect; BO: bonding defect; C: complex; S: simple.

Most classifiers missed defects rather than incorrectly classifying the panel as containing extra defects. Part of this result can be attributed to the post-classifier image processing tool which performs some minor trimming of objects which would tend to favour DF classification as defects were less common. The uneven quantity of training data from each class also contributed towards a classifier tendency to favour DF classifications; however, this effect was small. In general, non-bonding type defects, with their stronger echoes, were correctly identified more often than bonding defects. The local classifier is unable to differentiate between bonding and non-bonding defects. This is indicated by only focusing on the immediate echo information and ignoring other information available in an A-Scan making the classification process harder to achieve. Additionally, the ‘local’ feature set makes use of smoothed A-Scan data which affects the amplitude of the echoes and can lead to a downgrading of apparent echo strength from non-bonding to bonding and from bonding to DF. This observed effect was small in the bonding defects but was significant with the non-bonding defects due to the smoothing of the high-amplitude pulses. In addition, the ‘local’ and ‘mixed’ classifiers also identified the second echo from the mid-panel non-bonding defect on the 4-mm-thick panel as a possible bonding type defect. This was caused by the second reflected pulse echo off the defect occurring prior to the BWE and having sufficient strength to exceed the gate threshold. Detection of second echoes is always a risk when ‘local’ inspection techniques are used, and this is discussed further with the complex panel results.

The classifiers, and their ability to identify a wider range of defects, were tested on the complex panel data. The key results for this panel are presented in Tables 3 and 4. Key observations from each classifier in terms of correctly classifying the type of defect were similar to those presented for the simple panels with the following exceptions. The ‘general’ classifier returns only the class for the strongest echo observed within a single scan, meaning that non-bonding defects were reported when there was a superposition of both bonding and non-bonding defects. The ‘local’ and ‘mixed’ classifiers displayed an increase in the degree of misclassification of defects, when bonding and non-bonding defects were overlapping and when the bonding defect was closer to the probe. This was due to the increased attenuation of the ultrasonic signal as it passed through the bonding defect. When a non-bonding defect shadowed a bonding defect, often an ultrasonic signal of insufficient intensity was transmitted through the non-bonding defect to observe the later bonding inclusion, and second echoes were observed instead. It might have been possible to correct the attenuation effects of the defects by modifying DAC gain after a defect has occurred. However, the attenuation would be extremely dependent on the defect geometry, i.e. orientation of the ultrasonic wave, bonding with the composite structure, portion of defect in the path of ultrasound and other factors such as selection of the correct attenuation factor, a non-trivial task and beyond the scope of this research.

Table 3.

Classification results for all classifiers on data from the complex panel.

	‘General’ feature set			‘Local’ feature set			‘Mixed’ feature set
	2 mm	4 mm	6 mm	2 mm	4 mm	6 mm	2 mm	4 mm	6 mm
bfg	9.53	96.24	81.45	97.08	95.31	90.74	97.49	73.49	95.13
cgb	7.30	93.77	93.94	97.19	95.38	90.32	48.89	8.69	15.12
cgf	6.79	94.85	41.59	96.89	95.63	90.19	13.24	9.16	31.23
cgp	8.90	95.58	91.66	97.10	95.37	89.24	92.85	59.72	94.35
gda	9.70	95.15	48.00	97.01	94.92	89.34	95.90	77.32	95.12
gdm	88.06	92.17	72.21	97.01	95.14	89.01	97.00	96.37	93.85
gdx	9.15	94.43	92.47	96.55	95.33	90.90	96.60	96.01	95.47
lm	8.63	15.26	73.65	97.07	96.06	91.33	34.13	8.97	57.78
oss	8.35	26.54	10.37	96.99	95.64	90.17	97.47	42.15	93.82
rp	6.12	8.74	9.30	96.94	95.47	90.95	83.00	8.64	34.26
scg	7.58	90.17	93.46	97.18	96.02	91.31	81.51	7.54	14.80

The ‘mixed’ and ‘general’ classifiers were found to either classify extremely well or poorly on the complex data indicating a statistically significant difference between the ultrasonic signals recorded between the two panels. This indicated that degree of separation between DF and defect was very small and that minor changes, such as resin fill levels, in the panel make-up (in terms of quantity of defects or physical characteristics) had a significant effect on statistical descriptors and on data normalisation calculations. This effect was less pronounced in the ‘local’ and ‘mixed’ feature sets which indicated that echo descriptors were more robust to minor changes in the panel make-up. One final observed occurrence was that second echoes received from defects nearer to the front surface were reported as additional defects later in the sample. This could have been prevented through creation of a dead zone in the detection system after a defect was detected, where additional observed defects were suspected to be second echoes and are ignored. The cautious nature of NDT inspection in aircraft structures led to inclusion of all findings, even the ghost defects. Figure 6 showed the ‘mixed’ gdx classifier output for the 2 mm sections of the simple and complex panels which can be compared with the panel schematics in Figure 1.

Conclusion

Three different approaches to ultrasonic data description for neural network classification of defects in CFRP panels have been presented and compared. A feature set based on the description of the ultrasonic pulse echo in the immediate neighbourhood of an abnormality was found to perform well across a wide range of defects and panel configurations regardless of the ANN training approach. Inclusion of statistical descriptors was found to provide an improvement in classification performance at a cost of reduced robustness in terms of classifier training.

Successful detection and classification of defects were achieved for CFRP panels with different thicknesses based on a small amount of training data from three defect locations in one thickness of panel only. The ability of the classifier to correctly identify and classify defects at different depths within the panel was shown along with moderate success at locating multiple defects positioned coaxially to the ultrasonic probe.

This approach would be of use as an extra reporting and display tool alongside traditional C-Scans and D-Scans as classifier training, and testing would only need to be performed once during the development of the inspection process and the numerical loading of these tools when classifying is light.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The authors acknowledge the support of the Defence Materials Technology Centre (DMTC), which was established and is supported by the Australian Government’s Defence Future Capability Technology Centre (DFCTC) initiative, and would like to thank GKN Aerospace, Advanced Composite Structures Australia, Luscher Teknik and the Defence Science and Technology Organisation (DSTO) for their support in this work.

References

Workman GL and Kishoni D. Ultrasonic testing. In: Moore PO (ed) Nondestructive testing handbook. 3rd ed., Vol. 7. Columbus, OH: American Society for Nondestructive Testing, 2007, pp.380–390.

Aleksendrić

Carlone

Composite materials manufacturing. In: Aleksendrić

Carlone

(eds). Soft computing in the design and manufacturing of composite materials, Oxford, UK: Woodhead Publishing, 2015, pp. 15–38.

McIlhagger

Archer

McIlhagger

Manufacturing processes for composite materials and components for aerospace applications. In: Irving

Soutis

(eds). Polymer composites in the aerospace industry, Oxford, UK: Woodhead Publishing, 2015, pp. 53–75.

Bar-Cohen

. Emerging NDE technologies and challenges at the beginning of the 3rd millennium–Part II. Materials Evaluation 2000; 58: 141–150.

Martin

. Ultrasonic attenuation due to voids in fibre-reinforced plastics. NDT Int 1976; 9: 242–246.

Reynolds

. Nondestructive testing (NDT) of fibre-reinforced composite materials. Mater Des 1985; 5: 256–270.

Sachse

Castagnede

Grabec

. Recent developments in quantitative ultrasonic NDE of composites. Ultrasonics 1990; 28: 97–104.

Aymerich

Meili

. Ultrasonic evaluation of matrix damage in impacted composite laminates. Compos Part B Eng 2000; 31: 1–6.

Steiner

Eduljee

Huang

. Ultrasonic NDE techniques for the evaluation of matrix cracking in composite laminates. Compos Sci Technol 1995; 53: 193–198.

10.

Teti

Alberti

. Ultrasonic identification and measurement of defects in composite material laminates. CIRP Ann Manuf Technol 1990; 39: 527–530.

11.

D’orazio

Leo

Distante

. Automatic ultrasonic inspection for internal defect detection in composite materials. NDT E Int 2008; 41: 145–154.

12.

Palanisamy

Nagarajah

Iovenitti

. Effects of grain size and surface roughness on ultrasonic testing of aluminum alloy die castings. Mater Eval 2005; 63: 832–836.

13.

Diamanti

Soutis

Hodgkinson

. Non-destructive inspection of sandwich and repaired composite laminated structures. Compos Sci Technol 2005; 65: 2059–2067.

14.

Kamarthi

. A cluster-based wavelet feature extraction method and its application. Eng Appl Artif Intell 2010; 23: 196–202.

15.

Bhadeshia

. Neural networks and information in materials science. Stat Anal Data Min 2009; 1: 296–305.

16.

Zhang

. Neural networks for classification: a survey. IEEE Trans Hum Mach Syst 2000; 30: 451–462.

17.

Margrave

Rigas

Bradley

. The use of neural networks in ultrasonic flaw detection. Measurement 1999; 25: 143–154.

18.

Selvakumar

Radha

Narayanasamy

. Prediction of deformation characteristics of sintered aluminium preforms using neural networks. Model Simul Mater Sci Eng 2004; 12: 611–611.

19.

Solís M, Benítez H, Medina L, et al. (eds). Location of material flaws using wavelet analysis and neural network. Proc IEEE Ultrason Symp 2002; 1: 841–844.

20.

Farley

Durodola

Fellows

. A neural network approach for locating multiple defects. Appl Mech Mater 2008; 13: 125–131.

21.

Leo

Looney

D’Orazio

. Identification of defective areas in composite materials by bivariate EMD analysis of ultrasound. IEEE Trans Instrum Meas 2012; 61: 221–232.

22.

Palanisamy

Nagarajah

Graves

. A hybrid signal pre-processing approach in processing ultrasonic signals with noise. Int J Adv Manuf Technol 2009; 42: 766–771.

23.

Lee

. Feature extraction schemes for ultrasonic non-destructive testing inspections. Adv Inf Sci Serv Sci 2011; 3: 125–135.

24.

Camplani M, Cannas B, Cau F, et al. Acoustic NDT on building materials using features extraction techniques. In: International conference on computational science and its applications–ICCSA, 2008, pp.582–595. Berlin, Heidelberg: Springer.

25.

Santos

Perdigao

. Automatic defects classification—a contribution. NDT E Int 2001; 34: 313–318.

26.

Staszewski

. Intelligent signal processing for damage detection in composite materials. Compos Sci Technol 2002; 62: 941–950.

27.

Kesharaju

Nagarajah

. Feature extraction for ultrasonic sensor based defect detection in ceramic components. AIP Conf Proc 2014; 1581: 1933–1940.

28.

Benammar

Drai

Guessoum

. Detection of delamination defects in CFRP materials using ultrasonic signal processing. Ultrasonics 2008; 48: 731–738.

29.

Schillemans

Van Hemelrijck

De Roey

. Defect-identification in composite materials using pattern recognition techniques on ultrasonic data. Adv Compos Mater 1993; 3: 57–71.

30.

Bettayeb

Rachedi

Benbartaoui

. An improved automated ultrasonic NDE system by wavelet and neuron networks. Ultrasonics 2004; 42: 853–858.

31.

Sambath

Nagaraj

Selvakumar

. Automatic defect classification in ultrasonic NDT using artificial intelligence. J Nondestruct Eval 2011; 30: 20–28.

32.

Kesharaju

Nagarajah

Zhang

. Ultrasonic sensor based defect detection and characterisation of ceramics. Ultrasonics 2014; 54: 312–317.

33.

Anastassopoulos

Nikolaidis

Philippidis

. A comparative study of pattern recognition algorithms for classification of ultrasonic signals. Neural Comput Appl 1999; 8: 53–66.

34.

Cuéllar

Delgado

Pegalajar

An application of non-linear programming to train recurrent neural networks in time series prediction problems. In: Chen

C-S

Filipe

Seruca

Cordeiro

(eds). Enterprise information systems VII, Springer: Netherlands, 2006, pp. 95–102.

35.

Blanco

Delgado

Pegalajar

. A real-coded genetic algorithm for training recurrent neural networks. Neural Netw 2001; 14: 93–105.

36.

Martı

El-Fallahi

. Multilayer neural networks: an experimental evaluation of on-line training methods. Comput Oper Res 2004; 31: 1491–513.

37.

Byrd

Nocedal

. A limited memory algorithm for bound constrained optimization. SIAM J Sci Comput 1995; 16: 1190–1208.

38.

Kreßel UH-G (ed). Pairwise classification and support vector machines. Advances in kernel methods. Cambridge, MA, USA: MIT Press, 1999.

39.

Vapnik

. Statistical learning theory (adaptive and learning systems for signal processing, communications and control series), 1st ed. New York, NY: John Wiley & Sons, 1998.