1 Introduction

Flaws in composites

The complex behaviour of composite materials is related to their anisotropic and non-homogeneous microstructure [1, 2] and a variety of defects introduced during both manufacturing and service [3,4,5,6]. Various types of flaws are possible including, for example broken/misaligned fibres [7], porous matrix [8], voids [9], debondings [10], and delamination [11]. With many ways to increase the quality of a composite such as optimisation of the stacking sequence [12] or addition of external compaction pressure [13], accumulation of defects reduces the performance and life-cycle of composites. The inhomogeneous microstructure and involved processing techniques challenges flaw detection in composites. In this regard, non-invasive procedures can be used to characterise the material properties and its flaws.

Non-destructive testing

Non-destructive testing (NDT) evaluates the material without altering its properties or structural integrity [14]. The application of NDT ranges from process/design optimisation and manufacturing inspection to quality assurance and structural monitoring [15]. In most of the NDT methods, a specific field is analysed for a signal, e.g., eddy-current [16], ultrasonic testing [17], shearography [18], X-ray [19], and thermography [20]. Herein, thermography is used to characterise artificial flaws introduced in fibre-reinforced composites.

Types of thermography

Thermography is a powerful NDT method with many advantages such as rapid scanning of large structures and non-contact inspection [21]. It helps to ensure long operational life-cycles by early detection of defects and prevention of material failure [22]. There are two main types of thermography:

  1. 1.

    ‘Passive thermography’ investigates thermal radiation between a material and its environment due to a temperature gradient which is not caused by any external stimuli. It is used for structural health monitoring [21] or medical inspections [23].

  2. 2.

    ‘Active thermography’ requires external stimuli (heat) prior to inspecting the surface temperature field. Variation of temperature can be monitored in the transient or steady-state regimes [24]. Active thermography is commonly used in the NDT of fibre-reinforced composites [11, 25, 26].

Types of active thermography

Active thermography techniques can be categorised based on the type of heating stimulus, e.g., lock-in (modulated heat wave) [27], pulsed (short heating pulse) [28], pulsed phase (lock-in + pulsed with transformation algorithm to create phasegrams) [24] and step-heating thermography (long heating pulse) [25]. By controlling the stimulation regime, the results of the process can be fine-tuned. For instance, long duration step-heating thermography detects deeper flaws better as it applies a slower heating rate [29]. Therefore, it is often used for quantitative flaw detection of composite materials [23, 26]. For a controlled investigation, artificial flaws are often introduced into the sample.

Artificial flaws

Artificial flaws can be induced in composite parts using several techniques. To emulate debonding/voids, inserts with a thermal conductivity different to the composite specimen are used. For instance, Teflon® was incorporated in between carbon fibre-reinforced polymer (CFRP) laminae during manufacturing [30]; ROHACELL® and Polytetrafluoroethylene (PTFE) plates were used as inserts in cross-ply CFRP samples [31]. As a simple alternative, flat-bottom holes of various diameters are used in glass- and carbon-reinforced composite samples to create a validation matrix and challenge various thermography techniques [26, 31, 32]. The latter method, along with some image enhancement techniques, is adopted in this study.

Image enhancement

The obtained thermal images might not produce adequate contrast or some environmental artefacts might exist [33]; thus, some image enhancing is often carried out to improve the flaw detection process. Fast Fourier transform, principal component thermography (PCT), thermographic signal reconstruction, and partial least-squares thermography were employed in [33,34,35,36] to improve the signal-to-noise ratio (SNR) of the images. Adaptive histogram equalisation was used as the automated part of the pre-processing in [30]. Other spatial/frequency domain methods and their combinations are also present in the literature, see [37, 38] for instance.

Principal component thermography

Principal component thermography improves internal flaw detection by minimising non-uniform heating patterns and enhancing the contrast of thermal images [39]. It is done by applying principal component analysis (PCA) to a sequence of thermal images which intensifies the recorded response [35]. Consequently, the contrast between the defects and intact regions is improved. Although the effect of image enhancement techniques cannot be denied, the method has some inherent limitations; such as regional rendering (instead of point-by-point rendering) and being computationally expensive in time-derivative based computations [35].

Limitations of thermography

Next to the many advantages of thermography, there are some challenges that limit its use [14]. Common issues that interfere with the flaw detection procedure are lateral heat diffusion [40]; reflection and emissivity variation [41]; and non-uniform heating [42]. Flaw size and depth further limit the application of thermography [35, 43]. However, one less-considered effect is the anisotropy of sample during flaw quantification. The present work attempts to address this particular aspect.

Motivation and aim

In the literature, the available image enhancement techniques are often applied separately or in a combination to illustrate their effectiveness. However, the performance of these methods in dealing with fibre alignment has been overlooked. This study presents a systematic way of detecting flaws in CFRPs while considering the effect of anisotropy. It was hypothesised that anisotropy would affect the response of the composite undergoing step-heating thermography. This supposition was challenged by applying various heating regimes to two samples, i.e., a unidirectional (UD) and a cross-ply (CP) laminate (Sect. 2). After discussing the qualitative and quantitative results (Sect. 3), the study concludes the findings (Sect. 4).

2 Experimental Procedure

2.1 Material

Two CFRP composite laminates were manufactured [44] using a unidirectional out-of-autoclave prepreg carbon fibre (VTM264, \(200\,{^{\mathrm{g}}} / {_{\mathrm{m}^2}}\) [45]). Each sample consisted of 20 plies (0.1 m square) that resulted in a sample thickness of 4.5 mm. By varying the fibre orientation, two different plates were manufactured: a unidirectional and a cross-ply laminate.

For the simulation of defects, a \(4\times 4\) matrix of artificial cylindrical flaws was introduced at the back of the samples [26, 46]. The defect matrix consisted of flat-bottomed holes (FBHs) of varying diameter (top to bottom: 2, 4, 6 and 8 mm) with various heights, herein described as h, (left to right: 2.5, 3, 3.5 and 4 mm); they are numbered row-wisely in ascending order, see Fig. 1b. In Fig. 1c the back of the sample is shown where its rightmost column has the flaws of the greatest height, located 0.5 mm beneath the surface. The distance to the surface is therefore defined as d, also referred to as the depth of a flaw throughout this paper.

Fig. 1
figure 1

Manufactured laminate: a front of the unpainted sample; b front of painted sample (FBHs are shown schematically where a larger stroke is used for flaws closer to the surface); and c back of the sample with artificial flaws

Full size image

The infrared thermography technique performs better for surfaces with a high emissivity since the thermal interferences are minimised with lower reflectiveness [41]. The enhanced emissivity allows for higher accuracy of the temperature measurements during experiments, without altering the materials structural properties [47, 48]. Therefore, both samples were coated with a thin layer of flat black paint to reduce reflection, compare Figs. 1a (unpainted) and b (painted).

2.2 Methodology

Apparatus

A fully enclosed step-heating thermography setup was used in reflection mode [49] to obtain the thermal images, see Fig. 2. The system consisted of a Gobi-640 infrared camera (IRC) [50] with an un-cooled microbolometer (a-Si), a noise equivalent temperature difference of 50 mK at 30 \(^\circ\)C, spectral band of 8 – 14 \(\mu\)m and a pixel pitch of 17 \(\mu\)m. Two linear globe halogen lights of 500 watts each with a spectral power distribution between 20 and 100 % of 450–800 nm were used for the thermal excitation. The sample was positioned to receive an even heat distribution during thermal excitation; this was done by adjusting the angle of incidence of the lamps and their distance to the sample. Compared to common flash thermography setups [51], a low-power system was selected to produce low heating rates, that can improve the visibility of deep flaws without overheating and possibly damaging the material [52].

Fig. 2
figure 2

Schematic drawing of the test setup

Full size image
Fig. 3
figure 3

Flowchart of the flaw detection procedure; the boxes below each step list the parameters that should be dealt with accordingly before moving to the next step

Full size image

Method

The process of obtaining the experimental results consists of two main steps (see Fig. 3):

  1. 1.

    Setup Calibration; to obtain better quality raw images at the first stage, minor adjustments are applied through the following steps:

    1. (a)

      Calibrate Excitation; the halogen lamps (with a combined power of 1 kW @ 230 V) are positioned to increase the visibility of the flaws. While keeping the IRC fixed, the distance between the halogen lamps and the sample was varied between 15 to 60 cm; the 20 cm distance resulted in better resolution images. The angle between the axis of lamp and the surface of the sample was varied between 20 to 80\(^\circ\). The best image quality was obtained at a distance of 25 cm from the heat source and at a 45\(^\circ\) angle where a minimum surface reflection was obtained. Several thermal heating periods were applied in 5-second intervals in the 5–25 second range.

    2. (b)

      Calibrate Recordings; the rate of recordings (frames per second, fps) was varied between 1 and 50 fps. A frame-rate of 9 fps was chosen to achieve an adequate number of thermal images for a reasonable post-processing cost. To minimise the skewness of the recordings, the camera was positioned in front of the sample where the lens axis was normal to the sample surface. The cooling period was recorded for 120 seconds of which only 50 seconds were analysed.

  2. 2.

    Post-processing; at the second stage, PCT was conducted on single-phase (only heating or only cooling) and combined-phase data [53, 54] which were followed by image processing:

    1. (a)

      PCT; an in-house program was developed in MATLAB® [55] that used each recorded thermogram (thermal image or frame) as input data, i.e., flaw identification was carried out on the heating phase, cooling phase, and their combination. The PCT approach transforms the frame data into their principal components (dimensionality reduction) while preserving precision. The analysis identifies the correlations within the frames by extracting the eigenvectors of the covariance matrix [35]. In Fig. 4, the PCT image of the five-second heating phase for a plate is obtained from 45 single frames. Individual frames show weak indications of the flaws whereas more intelligible results are evident in the PCT image.

    2. (b)

      Binarisation; the resulting greyscale PCT image is passed through a Gaussian filter. The standard deviation for filtering the image using a 2-D Gaussian smoothing kernel is expressed as \(\sigma\). By increasing its value, the pixel density distribution becomes more dispersed around the mean value; thus, the blurred region grows. This step smooths the image and reduces noise. Then, the image is binarised using adaptive threshold by which the colour of a pixel is compared to local average intensity and converted to black/white. White regions denote flaws and are marked using red ellipses. The role of binarisation is objective detection of flaws; the novelty of this approach is selecting the size of signal area accordingly. This is in contrast with using a constant signal area, see [56] for instance.

    3. (c)

      Quantification of results; SNR is calculated using pixel intensities of the signal region (flaws) and noise region (area around the flaw) inside a bounding box. The bounding box is calculated by the in-house program which marks each flaw with a 9-pixel margin. The SNR metric (in decibels) is calculated by [33]

      $$\begin{aligned} \mathrm {SNR} := 10\cdot \log _{10}\big [ \frac{\left( \mu _\mathrm {S}-\mu _\mathrm {N}\right) }{\sigma _\mathrm {N}} \big ]^2, \end{aligned}$$
      (1)

      where \(\mu _\mathrm {S}\) and \(\mu _\mathrm {N}\) are the arithmetic means of all pixel intensities inside the signal and noise regions, respectively; and \(\sigma _\mathrm {N}\) is the standard deviation of the noise pixels. Note that other definitions of SNR are also available, but Eq. (1) is the most common one used in infrared thermography [33]. For instance, maximum and minimum pixel values could be used in the numerator which leads to similar trends [56]. In order to measure the performance of the current methodology, SNR is calculated for the last frame of the heating phase, PCT image, and PCT image plus Gaussian filter, i.e., \(\mathrm {SNR_{Raw}}\), \(\mathrm {SNR_{PCT}}\) and \(\mathrm {SNR_{PCT+G}}\), respectively. In addition, the aspect ratio of each flaw is calculated using

      $$\begin{aligned} \eta := \frac{l_{\mathrm {maj}}}{l_\mathrm {min}}, \end{aligned}$$
      (2)

      where \(l_\mathrm {maj}\) and \(l_\mathrm {min}\) are the lengths of major and minor axes of the identified flaw, respectively.

Fig. 4
figure 4

Combining 45 frames for a five-second heating (9 fps) into a single PCT image

Full size image

3 Results and Discussion

The manufactured UD and CP samples were exposed to several heating durations (5–25 s); then, the evolution of the temperature field in a single phase (heating or cooling) and combined phase (heating and cooling) was analysed. The following notation denotes the type of analysis:

  • The excitation period and the cooling period are separated by a forward slash, e.g., ‘5/50’ denotes five seconds of heating followed by 50 seconds of cooling.

  • The analysed period is put within brackets, e.g., single phase analysis of the heating phase for the aforementioned regime is denoted by ‘[5]/50’; single phase analysis of the 50-second cooling by ‘5/[50]’; and their combined phase analysis is shown as ‘[5/50]’.

Note that the qualitative analyses were conducted only on the results of Step 2 (a) for post-processing.

3.1 Qualitative Observations

Single-phase analysis of heating

By applying a short 5-second excitation, the surface flaws become visible in both UD and CP samples, see Fig. 5a and b. In these thermograms, flaws become deeper (\(d=0.5, 1, 1.5, \mathrm {and}\, 2\,\mathrm {mm}\)) from left- to right-hand side whereas they become smaller (\(d=8, 6, 4, \mathrm {and}\, 2\,\mathrm {mm}\) in diameter) from bottom to top. Increasing the duration of thermal excitation augments heat penetration and thus deeper flaws start to emerge in the thermograms; meanwhile, the flaws closer to the surface become less pronounced. More specifically, the leftmost column is more visible during short excitation ([5]/50) whereas deeper flaws seem to be less pronounced. This could be attributed to the heat radiation around the deeper flaws where the surrounding material obscures any apparent thermal gradients. The top row for both samples, flaws # 13–16 (2 mm diameter flaws) cannot be distinguished clearly without further image enhancement due to a smeared temperature field. Similarly, the deepest flaws in the rightmost column (flaws # 4, 8, 12 and 16) are not quite evident in the thermograms. In terms of visibility of the flaws, different fibre orientation of the plates does not seem to play a major role.

Due to the anisotropic properties of the UD plate, the temperature is conducted more vertically in the direction of the fibres; thus, a rather elliptical impression remains in Fig. 5a. In contrast, the flaws are captured in a more circular shape in the CP plate where thermal diffusivity is similar in every direction, as shown in Fig. 5b. Therefore, it can be argued that the real shape of flaws could be lost under the influence of the material anisotropy.

The number of detectable flaws varies as a function of heating duration. Under single-phase heating analysis, shorter excitations result in more distinct flaws compared to longer excitation periods where the heat signature of sound and defected regions are less distinguishable. This is in contrast to the findings of single-phase heating or combined analyses.

Fig. 5
figure 5

Analysis of the single heating phase: a UD plate A; and b cross-ply plate B

Full size image

Single-phase analysis of cooling

Longer excitation periods generate higher temperature fields which magnifies radiation and thus reduces the clarity of thermograms [29]. In order to monitor the temperature field for a longer period and avoid this issue, a single-phase analysis of the cooling period was conducted. Following the procedure in Fig. 3, the first 50 seconds of the cooling period was fed into PCA. Similar to the single-phase analysis of heating, surface flaws of above 4 mm in diameter could be easily detected at low excitation periods (5/[50]) while the deepest flaws left almost no impression—except a couple of shadows for 8 and 6 mm flaws # 4 and 8 respectively, see Fig. 6a and b.

Unlike the heating phase thermograms, increasing the excitation period up to 25/[50] shifts the focus towards the deeper flaws and reduces the impression of superficial ones. Cooling thermograms seem to fail to reveal the small 2 mm flaws (flaws # 13–16), too. However compared to the single-phase analysis of heating, heat signatures interfere less and the overall quality of the thermograms improves.

In terms of anisotropy, there seems to be a direction preference in UD samples along the fibres due to their higher thermal conductivity. Thus, the resulting impressions are more elliptical compared to the cross-ply sample. This is a repeating effect similar to the results of single-phase heating analyses.

Fig. 6
figure 6

Analysis of the single cooling phase: a UD plate A; and b cross-ply plate B

Full size image

Combined-phase analysis of heating and cooling

Analysis of the combined phases reveals similar results to those of the single-phase analyses, see Fig. 7a and b. A similar pattern emerges—the largest superficial flaw seems to be highlighted for [5/50]. Nevertheless, most of the flaws seem to be visible under [10/50] and [15/50]. The geometry of the flaws is distorted due to the anisotropy of the UD sample while a more realistic shape is obtained for the CP sample. It seems that analysing both phases slightly reduces the quality of the PCT image due to higher noise levels. This effect becomes more evident after applying image enhancement, see Subsect. 3.2.

Fig. 7
figure 7

Analysis of the combined cooling-heating phases: a UD plate A; and b cross-ply plate B

Full size image

Summary

The PCT images of single and combined phases revealed similar results, qualitatively. High noise levels in the images directly affect the outcome of post-processing by reducing the clarity of the images [57]. Therefore, single-phase cooling was selected and used for Step 2 (b) of post-processing. Note that the literature is equivocal in terms of selecting the best regime; for instance, single-phase cooling is suggested in [56, 58] while [33] suggests using single-phase heating. Therefore, it is advised to consider all combinations independently before making any decisions.

3.2 Binarisation

Binarisation of the single-phase analysis of cooling

Binarisation can be readily performed on the PCT images of individual or combined phases; however, single-phase cooling was used herein. The aim of this step is to reduce complexity of the qualitative observations and to aid their interpretation. Following a trial-and-error process, the following parameters were selected:

  • for the UD plate, \(\sigma\) for the Gaussian filter was set to four with an adaptive threshold at 0.50, see Fig. 8a; and

  • for the CP plate, \(\sigma\) for the Gaussian filter was set to five with an adaptive threshold at 0.51, see Fig. 8b.

In the binarised thermograms, the black region represents the sound material whereas white regions represent the candidate location for flaws. Identified flaw locations are outlined red; those in yellow require further investigation, see Fig. 8 for example.

Fig. 8
figure 8

Applied binarisation to cooling phase analysis: a UD plate A; and b cross-ply plate B

Full size image

Flaw detection

Location, size, and shape are the parameters of interest in flaw detection. Irrespective of anisotropy, the majority of flaws were located for both samples. In the UD sample, all large FBHs (\(d =8\,\mathrm {and}\,6\,\mathrm {mm}\)) were detected unanimously irrespective of their depths at any excitation level. In contrast, only the 5-second excitation revealed all the large flaws in the CP sample. Therefore, one could confidently point out the location of flaws # 1–8 in both samples (especially for the UD case). Other locations could be taken as candidates for further investigation, e.g., using ultrasonic NDTs.

Signal visibility

The flaw signals seem to decrease in size and/or vanish by increasing the excitation period—starting from deeper and smaller flaws, see Fig. 8. Namely, short excitations detect more critical regions. This behaviour could be attributed to the quick thermal reaction of the large surface flaws that appears as thermal gradients. The same behaviour is responsible for generating the yellow-outlined regions which start to untangle at longer excitations asserting on the physical separation of the corresponding flaws. Overall, it seems that the size of a flaw has a greater impact on its visibility rather than depth. Namely, the thermal signature of small flaws, even surface ones, is spread out more due to their size. The important point is that it is necessary to look at the thermograms of various excitation level altogether, which provide a better judgement for the location of the flaw. Also, one should consider that the size of the signal does not generally match the physical size of the flaw. There are a few requirements for extracting the flaw size from images:

  1. 1.

    a calibration of image size to the physical specimen to identify the number of pixels that represent a unit length,

  2. 2.

    to know the depth of the flaw and the material properties to use in heat transfer equations and estimate the real flaw size.

In practice, it is not easy to characterise parts in service and almost impossible to know the real depth of the flaw. Therefore, only estimations could be provided which were excluded from the scope of the current study.

Excluded signals

Using a matrix of flaws allows for determining the limitations of the apparatus and the procedure in terms of flaw size and depth detection [26, 46]. It is assumed that flaws finer than a specific size and deeper than a certain depth cannot be detected. For instance, if flaws of 4mm and 6mm do not leave a thermal impression, no response should be received from 2mm flaws at the same depth. In the UD plate, flaw # 16 is detected even though it is the deepest and smallest. In contrast, some larger flaws (# 14 and 15) cannot be detected in the same sample. This artefact might be due to edge reflections of the sample, presence of inhomogeneities, or some other manufacturing imperfections. Nevertheless, to avoid any interference with the quantified results, flaw # 16 of the UD sample; and flaws # 15, and 16 of the CP plate were conservatively excluded.

Anisotropic effects

The shape of flaws are clearly affected by the thermal anisotropy of the samples at any excitation level, cf. Fig. 8a and b. Higher thermal conductivity of the UD sample (along the fibres) promotes merging the thermal signatures under quick excitations, i.e. flaws # 5 and 9 at 5/[50] and flaws # 1, 5 and 9 at 10/[50]. This effect increases the difficulty of accurate flaw detection. More importantly, anisotropy distorts the physical shape of flaws where a more elliptic shape is observed in the UD sample. In contrast, the response of the CP sample is in a closer agreement with its physical shape. On the other hand, anisotropy seems to improve the visibility of deeper flaws at longer excitations.

Summary

A signal dictates the possible location of a flaw but does not necessarily reflect its real size, depth and physical shape. The detection of the realistic shape of a flaw becomes more difficult with increasing distance of the defect to the surface and can result in misguided defect evaluation [46, 59, 60]. When dealing with qualitative observations, the size and shape of the flaws are affected by the excitation period and the anisotropy of the material. In the next subsection, more quantified analyses are carried out to complement the initial findings.

3.3 Quantification of Results

Herein, two metrics are used to further investigate the 50-second single-phase cooling results: image quality is gauged by \(\mathrm {SNR}\), and the shape of flaws are presented by their aspect ratio (\(\eta\)). Note that the metrics are calculated only for the confidently-detected flaws.

Signal-to-noise ratio

Table 1 lists the SNR values of the raw image (a single frame), PCT image, and Gaussian-filtered PCT image of UD and CP samples under excitations of 5–25 seconds. Generally, PCT increases the SNR values, i.e., the number of clearly-detected flaw increases. Some negative SNR values are obtained for raw images that indicate either a very large standard deviation of noise and/or very small distinction between noise and signal; qualitatively, negative values indicate that there is no clear distinction between noise and signal. Therefore, the average SNR of positive values (\(\mu +\)) seems to be a better indicator. It shows an overall increasing trend across the post-processing steps, which peaks at 20/[50].

Overall, increasing the thermal excitation improves SNR; however, its value starts to decrease at very long excitation levels. The 20/[50] regime produces the best contrast which is diminished by further increasing the excitation period. This behaviour could be attributed to the depth of the flaws since the detection of surface flaws are not improved by increasing the heating period. Therefore by increasing the excitation time, a peak for the SNR value would be expected if there are deeper flaws in the specimen. A similar behaviour was reported in [33] for thermal excitations more than 30 seconds.

Almost in every case, increasing the depth of any specific flaw size reduces the SNR value—often to the negative domain which indicates a decrease in their visibility. Similarly, smaller flaws tend to produce smaller SNR values. These trends confirm the qualitative assessments of the previous section, i.e., signals become weaker for deeper and smaller flaws. However, there exceptions exists, e.g., flaw # 5 of the CP sample is smaller than flaw # 1 but produces a better SNR value. This response can be attributed to the inhomogeneity of the sample. That is, the aforementioned statements are correct provided that the sample is completely homogeneous. Namely as an ideal case, it is assumed that the SNR value does not depend on the location of a specific flaw. However in practice, a specimen might contain resin-rich regions or clustered fibres around the flaws that interfere with the consistency of response. Nevertheless, SNR values can be generally used as a comparative metric for quality of thermographic response. Note that a simple post-processing step such as Gaussian filter can tremendously improve the SNR value, e.g., for flaw # 11, SNR changes from \(-0.850\) to 6.277 in 5-seconds excitation.

One side effect of anisotropy is the possibility of merged flaws along the direction of higher thermal conductivity, which occurs under short excitation periods. For instance, flaws # 1 and 5 under 5-15s excitations and flaw # 9 under 5–10s excitations merge to the flaws in the upper row. Surface flaws in the UD case (flaws # 1, 5 and 9) merge into each other due to a higher conductivity along the fibre direction whereas the response of the CP sample is milder along both principal directions. Therefore in practice, there might be some closely-spaced flaws that appear as a single flaw. It is recommended to increase the excitation time to see if any separation of response happens.

Looking at the average SNR values, both UD and CP samples achieve their average peak under 20s excitation and decrease afterwards. The CP sample shows higher average positive SNR values compared to the UD one except for the 5s excitation. Thus, in average the flaws in the CP sample are more clearly detected and the best results are obtained under 20s excitation, which seems to be independent of anisotropy.

Although no conclusive differences can be detected for the SNR values due to anisotropy, there are differences. For instance, if the available SNR values for both UD and CP samples are considered, the response of similar size flaws are more pronounced in the CP specimen. Also, deeper flaws in the CP samples consistently decrease in SNR values whereas for UD samples, there are some discrepancies. Additional tests would be required by varying the degree of anisotropy to confirm other possible trends.

A final note on the discrepancy of results is the proposed flowchart. Some publications, such as [56], use a constant area to monitor various size of flaws. Therefore, larger flaws would definitely result in a higher SNR value. In contrast, the proposed flowchart updates the monitoring area according to the binarisation step. This is required as the shape of the response changes due to anisotropy and thus, a specific boundary (e.g., a 7 by 7 pixel box) cannot contain the signal; larger boxes would be required which also would add some noise area to the signal. This aspect of anisotropy has not been addressed in other publications.

Table 1 SNR results of the UD and CP samples
Full size table

Aspect ratio

The aspect ratios of the detected flaws under 5- and 20-second excitations are listed for both UD and CP samples in Table 2. Ideally, an aspect ratio of 1.00 is expected for the circular FBHs. Hence, a value close to this allows for a realistic representation of the flaw. With the mean aspect ratios of 1.23 (5/[50]) and 1.08 (20/[50]), the CP sample clearly exhibits a more isotropic nature compared to its UD counterpart where aspect ratios of about 2 are obtained. No valid relationship between the excitation time and aspect ratio could be established. However under 5-seconds of excitation, deeper flaws seem to reflect higher aspect ratio in the UD sample whereas the CP sample shows a lower sensitivity. This effect is less pronounced under 20-seconds of excitation. This behaviour can be justified by the heat conduction of fibres which is further stretched along the fibres orientation before reaching the surface—hence making a larger impression for deeper flaws compared to the superficial flaws of the same size. Higher excitation periods seem to increase the aspect ratio but the increasing trend is not always followed. Interestingly, the fibres of the CP sample work equally in two directions, which makes the aspect ratio almost independent of the flaw depth.

Summary

By increasing the excitation period, the thermal response of surface flaws quickly peaks and drops; in contrast, deeper flaws take longer to react to this increase [61]. Detection of deeper flaws becomes progressively difficult due to the increasing amount of energy needed to obtain a thermographic response. This tends to result in a decrease of their SNR values—and thus visibility [47]. Hence, plotting SNR against excitation time for various flaws could help with comparing their depths.

Anisotropy does not seem to affect SNR values a lot but obviously alters the aspect ratio of the response. Analogous to the concept of thermal ellipsometry [62], the contrast in the thermograms is observed to reflect the direction of thermal conductivity in fibre direction. Higher contrasts are achieved quicker along the fibre orientation of the material. Thus, the shape of thermal impressions does not correspond to that of the physical flaws. Nevertheless, the side effect of merging flaws in the UD case can be eliminated by increasing the excitation period; optimum SNR values are obtained under 20s excitations irrespective of the degree of anisotropy.

Table 2 Aspect ratios (\(\eta\)) for the 5/[50] and 20/[50] analyses of the UD and CP samples
Full size table

4 Concluding Remarks

Conclusion

The aim of the current study was to illustrate the effect of anisotropy in the thermographic response. To this end, artificial circular flaws were introduced into UD and CP carbon fibre laminates at various depths. The thermal response of samples was investigated under various thermal excitations using step heating thermography. In order to provide more objective results, a flaw detection procedure was introduced and followed for qualitative and quantitative analysis of the results. The proposed methodology uses PCT, Gaussian filters, and binarisation for a systematic flaw detection. It starts with the calibration of the setup followed by post-processing of the thermal images using PCT and some image enhancement tools. Single-phase (excitation/cooling only) and combined-phase analyses were compared to each other where single-phase cooling was selected for further study. It was shown that short thermal excitations work better for superficial flaws and by increasing the duration of excitation, deeper flaws start to reveal. The SNR value experienced a peak at 20-seconds of excitation due to the presence of both deep and superficial flaws. Quantitatively, the SNR metric seemed to be insensitive to the anisotropy of the sample. However, the thermal signature of the flaws was scaled to aspect ratios of about 2 in the UD sample, which resulted in distortion of shape and merging of closely-spaced flaws along the fibres. In contrast, the CP sample exhibited a more isotropic response.

Future work

To further challenge the suggested procedure, further experimentation on the degree of anisotropy (e.g., adjusting it by using angle ply laminates), material contrast (e.g., by using glass fibres), failure mechanisms (e.g., delamination), or inhomogeneities (e.g., fibre clustering) could be conducted in the future.