5. Extended Non-destructive Testing Technique Demonstration in a Realistic Environment with Technology Assessment

We designed a panel representing a wing application for the production user case with the intention that it should differ from the part used for the repaired fuselage application to highlight the versatility of the ENDT approaches. The manufacturing process of the panel was performed at Aernnova and comprised two different steps: the manufacture of the T-stringers and the manufacture of the skins. Hexcel M21E was used for the manufacture of all samples. Similar to the repair user case, we chose the dimensions to represent the correct order of magnitude of a tested region while providing enough space to test the application of several ENDT methods.

The stringers were manufactured following current industrial production guidelines. The web was shaped by joining two C-shape preforms that had been produced using automated tape laying (ATL). First, a flat panel was laminated and placed on a hot-forming tool to obtain the C-shape. After that, both C-shaped parts were joined. Finally, the element was trimmed to obtain the six T-stringers with the required dimensions. Four of these were dedicated to mechanical testing and two were kept for the full-scale demonstration. The stringers were inspected by ultrasonic C-scan inspections and no particular defects were revealed (Fig. 5.1).

The skin samples were manufactured with the following carbon fiber distribution Table 5.1, which was similar to the panels used for mechanical testing.

The production began with the manufacture of a 3500 mm by 1000 mm flat panel laminated using ATL. The distribution of the skin samples within this panel is depicted in Fig. 5.2.

After the curing cycle, the flat panel was automatically inspected using pulse-echo equipment with water as the coupling medium to ensure the quality of the part manufactured. No defects were revealed (see Fig. 5.3). After the inspection, a trimming operation was performed to obtain the number of skin samples required for the mechanical testing (see Chap. 2) and full-scale demonstration.

The bonding operation of the stringer onto the composite skin was part of the full-scale demonstration workflow. Thus, intentional contamination needed to be done before bonding. It should be noted here that due to the dimensions of the bond line and the presence of two stringers, all the ENDT investigations could be tested using the same panel. More details are provided in the next section.

To demonstrate the ENDT techniques on full-scale specimens, we performed a bonded repair on a real part as we considered this to best represent a real application scenario, wherein the part being repaired has undergone a variety of loads and environmental conditions. For this purpose, fuselage panels were recovered from an A350 aircraft used by Airbus for full-scale static testing. These panels had been part of such a structural test, but had remained in good condition and could thus be employed in the demonstration (see Fig. 5.4). The size of each panel was approximately 1 m², which was adequate for several reasons:

As explained within the workflow and operations section, we assessed three panels. One panel was used for “intrusive” ENDT, such as monitoring based on magnetostrictive sensors and electromechanical impedance, for which it was necessary to bond sensing devices to the part. As the setups required for surface quality measurements using laser-induced fluorescence spectroscopy (LIF) and vibrometry were not moved to the demonstration event location, one panel was kept at IMP-PAN in Gdansk after scarfing and without bonded repair. Finally, one panel was used for the inspection using non-intrusive bonding assessment techniques, namely nonlinear ultrasound (NLUS) and laser shock adhesion testing (LASAT) (as outlined in the workflow and operations section).

As they were taken from the same section, these three panels were similar, but they were not identical in terms of geometry, number of stiffeners or thickness. A detailed view of two of the three panels is presented in Fig. 5.5. The skin thickness was in the range of 1.5–2.5 mm, which corresponds to the coupon and pilot sample geometries because—once repaired—the assembly thickness increased to about 2.5 mm. In addition, either two or three omega stringers reinforced the skin. The stringer foot was about 1.5 mm, which means that the total thickness was a maximum of 4 mm in this region. The panel material was the composite IMA/M21E (Hexcel), similar to the material used for the coupon and pilot samples.

The three panels were used to simulate a scarfed bonded repair. As the focus is on testing the adhesive bond between the scarfed skin and the repair patch, we decided to use these structures without introducing any damage to save time and costs. While it could have been subjected to, for example, delamination caused by a drop tower, this would have been removed by the subsequent scarfing operations anyway. Instead, we defined the dimensions of the bonded repair patch based on the standard for delamination that was not actually performed.

The complex geometry of the stringer, i.e., a realistic part, as compared to the flat coupon samples required a substantial modification of the setup and an adaptation of the measuring process to facilitate the surface monitoring based on LIBS investigations. We achieved a significant increase in the TRL by developing a robot-aided quality assessment procedure that allowed both the sensitivity and the robustness of the standard laboratory-based approach to be maintained. Essentially, we matched the control of the LIBS sensor to that of a standard industrial robot that was set to handle and position the part to be analyzed, as shown in Fig. 5.14. This arrangement offers some advantages for the LIBS measurement process since the focal point of the LIBS laser can be fixed. This fixed focal point (as shown in Fig. 5.15) was used as the origin of the coordinate system with which the robot handled the parts. Focusing on the analytical process, the LIBS measurements were performed on the CFRP surfaces after the peel ply was removed and the defined combined contamination process was performed. A laser wavelength of 1064 nm was used with a 180 mJ laser pulse energy. In order to determine the surface composition of the stringer, surface areas along three lines—each comprising 130 single measurement spots—were analyzed. An image of such a stringer is depicted in Fig. 5.16, while selected LIBS signals obtained during the investigation are presented and discussed based in Fig. 5.17.

In Fig. 5.17, we present the results achieved when aiming for the detection of a contamination that arose from depositing relevant amounts of a Si-containing RA on the C-rich stringer surface. Essentially, the first 65 measurement points obtained within the intentionally contaminated half of the stringer surface indicated a clearly increased \(\left[ {\text{Si}} \right]\text{ / }\left[ {\text{C}} \right]\) concentration ratio. This significant finding is indicative of the presence of RA contamination and is, as anticipated, in contrast to the LIBS results from the non-contaminated right half of the stringer surface. Thus, using LIBS measurements it is clearly possible to detect and allocate the areas corresponding to either the contaminated or clean CFRP surface state. Additionally, evaluating further signals enables the type of contamination to be identified. Figure 5.17 presents the characteristics of the Na signal along with one of the lines composed of 130 measurement spots separated by 5 ± 1 mm, whereby it is possible to distinguish a locally increased intensity within the first 65 measuring points. Single peaks demonstrating an increased Na concentration indicate the presence of a Na-containing FP deposit in the contaminated area. It is more challenging to detect and locate the FP contaminations as the contaminated area was relatively limited compared to the region covered by the broader RA deposits and also because only up to three single LIBS shots were located in each FP contaminated area (Fig. 5.18).

In summary, when using the advanced robot-aided LIBS characterization process and evaluating both the \(\left[ {\text{Si}} \right]\text{ / }\left[ {\text{C}} \right]\) signal ratio and the Na signal intensity along three parallel lines composed of measurement spots separated by 5 ± 1 mm, we achieved a clear and significant differentiation between the clean surface area and the contaminated area of the stringer, the detection of the elemental constituents of the applied contaminations, and consequently the identification of both contaminants contributing to the combined deposit.

The contamination on the production part was a single RA contamination on half of the part. This contaminated area was successfully detected and differentiated from the clean half. We were thus able to master the challenge of identifying and detecting laterally separated water droplets with the use of image processing based on a neural network. An overview of the measurements is given in Fig. 5.19.

Due to the high potential of e-nose techniques with respect to surface quality assessment, we tested the two distinct setups advanced in the ComBoNDT project during the three-day full-scale demonstration event.

During the demonstration, the realistic and real parts of both main user cases were measured both with and without contamination. For the production user case, three different specimens were tested (a sample with the dimensions 20 cm by 20 cm and two production panels). Panel 1 provided a section with an intentionally contaminated area (Panel 1c) that had been prepared following the combined “RA-FP” scenario. All other samples were not contaminated.

Figure 5.20 shows the desorber device, which was developed in the ComBoNDT project by the partners from Airbus, during its application in taking measurements on the realistic parts in the production user case.

Subsequently, we discuss the evaluation of the acquired e-nose data considering two different approaches to data treatment and representation, namely the partial least squares discriminant analysis (PLS-DA) model and the principal component analysis (PCA).

When we evaluated the measured data of the production panels using the PLS-DA that had been obtained from the pilot sample test, we achieved the differentiation and correct classification of all surface states in both the “clean” and “not clean” (i.e., contaminated) classes. In Fig. 5.21, the horizontal threshold line represents the boundary signal level that can be used to determine whether a data point is indicative of a panel surface state that is attributed to the “clean” class (i.e., ready for bonding) or the opposite class (class threshold).

By performing cross-validation with the obtained, the confusion matrix shown in Table 5.3 is obtained.

Concerning the data evaluation based on a single-class identification method, a different method of signal processing was tested to identify the true class with a clean surface state of a specimen panel. In detail, six samples from the sampling areas on the clean Panels 1 and 2 were taken to determine the region in the PCA that correlates to a clean sample surface. We used these samples to build the detection model. Using a confidence level of 95%, we obtained the green domain shown in Fig. 5.22, which indicates the region in which an unknown sample would be classified as clean. Measurements that lie outside this domain would be classified as false, i.e., contaminated or not clean.

When we subsequently evaluated the remaining seven e-nose measurements of Panel 1 and Panel 2 with this built (single-class identification) PCA model, all measurements made on the uncontaminated areas of the panels were located inside the green domain. Meanwhile, all measurements made on the contaminated areas lay outside this domain, which represents clean surface areas, and were, therefore, correctly identified as being indicative of contaminated samples. Figure 5.22 displays all the results for the single-class identification evaluation. We would like to highlight that all “unknown” (i.e., prepared and analyzed on-site) surfaces of Panel 1 and Panel 2 were classified correctly based on the performed intentional contamination procedure and following the e-nose measurements.

The realistic part production user case was also tackled using the latest version of the ENEA e-nose, which is depicted in Fig. 5.23.

Here, we briefly describe the developed data acquisition and evaluation process. Separate models for anomaly detection were produced in order to differentiate contaminated samples from the clean reference samples. Based on nine measurements taken from a (clean) curved realistic reference specimen, we obtained a low dimensional projection using the first three principal components in the PCA rotated sample space. Hypothesizing a multivariate Gaussian distribution of the measured data’s projection on the low dimensional subspace, we used a two sigma coverage to set an ellipsoidal surface acting as a detection threshold for contaminated samples. Results from the measurements of clean surface regions were uploaded to the joint Graphical User Interface (GUI) software to compute the detector surface and prepare the GUI to accept the test measurements. Then, we used findings from nine sample regions to test the developed detector: three from clean areas of the tested part and six from areas that had been contaminated previously following the “RA-FP” scenario. In summary, we measured 18 samples, of which nine were set apart to build the training set of the PCA-based detector, while the other nine were used for testing purposes. The joint GUI was able to show the results in real-time immediately after the measurement was taken. A screenshot of the active GUI is displayed in Fig. 5.24.

Table 5.4 presents the confusion matrix summarizing the results obtained for the production sample scenario.

As can be seen in Table 5.4, based on the advanced process, we achieved a perfect recognition score. Furthermore, the detection threshold was overcome along a similar direction in the same sub-plane for all except one of the contaminated samples, which adds robustness and significance to the results.

The LIBS measurement technique has successfully shown its potential in monitoring composite adherend surfaces, even in the case of complex shapes. During the three-day full-scale demonstration event in the ComBoNDT project, we used the technique for the automated measuring and detection of multiple contaminations with RA and FP, whereby both contaminants were detected and revealed. If a more clearly localized detection of single contaminant-based FP were aspired in the future, then a higher statistical basis, i.e., more measuring points than the 390 measured here for each stringer, would be required.

Using AWT to inspect a realistic production CFRP part, a single RA contamination was detected by means of advanced data evaluation based on a neural network. Finally, a robust and significant detection of a complex combined contamination was achieved using both distinct enhanced e-nose techniques employed for surface quality assessment during the full-scale demonstration event, applying PCA for data evaluation and benefitting from a user-friendly GUI.

The realistic part prepared for the production user case was a CFRP panel with an adhesively bonded stringer, as depicted in Fig. 5.25. One half of the bonding area for the stiffener was intentionally contaminated before bonding. The rear side of this region is shown in the left part of Fig. 5.25a. The combined contamination scenario comprised the application of RA and FP. The second half of the stringer surface was maintained in a clean state and its rear side is shown in the right part of Fig. 5.25a. Technically, within the EMI-based inspection process, this area is used for reference purposes.

Specifically, we bonded two piezoelectric sensors with cyanoacrylate glue to the adherend surface of the CFRP panel (Fig. 5.25a) in order to assess the bonding quality of its joint with the stringer. We placed the sensors on the rear side of the panel in the area where the stringer was bonded. The DD1 sensor was placed on the area of the clean bond, while the DD2 sensor was bonded to the contaminated (following the “RA-FP” scenario) region. The electrical characteristics of both sensors were measured three times. First, the free sensors were measured for referential purposes. Second, the bonded sensors were measured while the panel was lying on the laboratory table (as depicted in Fig. 5.25b). Third, a measurement was made while the panel was being subjected to the external loading situation used for strain measurements. The loading was applied by attaching the panel to the laboratory table with adhesive tape, as demonstrated in Fig. 5.26.

With respect to the sensor signal output and the subsequent data evaluation, Fig. 5.27a presents the obtained root mean square (RMS) values for the conductance obtained in the 3–5 MHz frequency range. After bonding the sensors, a drop in the RMS value was obtained as compared to the free sensors. We observed higher RMS values for the sensor bonded to the contaminated area, which may indicate a weaker adhesive bond. Additionally, when applying an additional load to the panel, the two sensors showed significantly different signal changes. We observed an RMS increase for the clean region, while a decrease was found for the contaminated bond region. As a second characteristic, the conductance maximum peak shift was also tracked, as displayed in Fig. 5.27b. Taking as a reference the frequency at the conductance maximum of the freely oscillating sensor to be bonded close to the bonding region with the clean adherend, the difference to the frequency at the conductance maximum of the second freely oscillating sensor was below 0.5%. After bonding both sensors, a significant increase of this characteristic frequency was observed, which was more pronounced for the bond comprising the contaminated adherend. After mechanically loading the panel, the peak for the clean bonding region shifted leftward towards a lower frequency, while the one for the contaminated region shifted rightward, reaching a shift of almost 2%.

After inspecting the measured conductance curves in more detail, local resonances in the range of 3.75–3.90 MHz were observed. These affected the RMS values calculated for this frequency region as compared to the RMS values for the free sensors (see Fig. 5.28a). Similar behavior was observed for the resistance curves but in a slightly different frequency range, namely from 3.6 to 3.9 MHz, as shown in Fig. 5.28b. As compared to the regions close to the bonds with clean adherends, we observed lower RMS values around the contaminated area, which may indicate a weaker adhesive bond there.

For the investigation of the bonded CFRP specimen from the production user case, the ultrasound excitation transducer was positioned symmetrically in relation to the longitudinal extension of the stringer to be tested, as shown in Fig. 5.29. Such a position provided a symmetrical insonation field for this stringer, and the findings for the fundamental frequency 20 kHz are displayed in Fig. 5.29.

Additionally, the second-highest harmonics were measured in the production specimen. The findings shown in Fig. 5.30 demonstrate quite different signal distributions, whereby the higher efficiency of the local nonlinear generation in the left-hand-side of the stringer can be clearly perceived and identifies a bonding region based on a contaminated adherend surface area with a lower bonding quality.

Before detailing the results obtained when applying the laser shock adhesion test (LASAT) for investigating the realistic bonded CFRP specimen from the production user case, we provide a short and demonstrative introduction to the full-scale test performed during the demonstration event of the ComBoNDT project.

Introduction to the full - scale test

In order to facilitate the LASAT investigations on large-area realistic or real parts with complex, e.g., curved, shapes, we advanced the Hephaïstos setup by implementing profound modifications. With the aim of handling specimens of the required size, the optics had to be substantially rearranged and a robot had to be installed to manipulate the sample. The resulting setup is displayed in the on-site environment in Fig. 5.31.

The second challenge was the confinement. So far, the plasma generating the shock wave within the sample had been confined using a thin water layer. Even though this ENDT technique offers an easy setup on a smaller scale, the new upscaled configuration would have introduced substantial operating costs, e.g., the installation of special pipes to supply the investigated sample region with water as well as to ensure water drainage. To overcome these challenges, solid confinement was used. Numerous materials were tested, and for some of these, the results for the flux-dependent plasma pressure are displayed in Fig. 5.32. This graph describes the generated pressure as a function of the laser intensity for confinements based on three different materials. The LASAT operating regime is around \(1 {-} 3\,{\text{GW}}/{\text{cm}}^{2}\), and in this flux interval, water, epoxy, and a thick transparent PPE adhesive tape turned out to have the same confinement capability. With respect to the aspired application, the latter was eventually chosen for its ease of use, and the setup is shown in Fig. 5.33. The aluminum sacrificial layer was not modified.

Briefly, the workflow for testing was as follows. For each sample, an initial threshold identification was performed in different areas. On each calibration area, a series of laser shots were performed. If no damage was inflicted, the laser intensity over the area was increased. The opening energy of a sound part was defined as the “reference threshold”. Once this threshold had been identified and confirmed for each calibration area, the rest of the sample was characterized by shooting at 80% of this value, regardless of the contamination state.

To test the bonded CFRP panel of the production user case, the whole stiffener was spot-wise shot with an offset of 20 mm between consecutive shots, as presented in Fig. 5.34. This spacing was selected to avoid compromising an area with a crack that could have appeared in an adjacent tested zone.

The threshold of the bonded regions based on either the contaminated or the clean adherend surface area was isolated and a total of 53 shots were realized on one of the stiffeners. The achieved results are summarized in Table 5.5.

Briefly, two distinctive opening fluxes were found. Furthermore, during the performed 53 laser shots, none of the tested areas on the sound part of the stiffener were opened. Finally, all contaminated areas failed in the test, meaning that the symmetrical LASAT (S-LASAT) results were conclusive for the production panel.

The advanced setups and workflows for the EMI, NLUS, and S-LASAT inspections used for assessing the bonding quality of the realistic bonded CFRP parts investigated in the full-scale production user case within the ComBoNDT project enabled a differentiation of bonded regions based on clean or contaminated adherend surface areas. The outcomes of the three-day demonstration event in the ComBoNDT project demonstrated the enhanced technology readiness levels of these techniques for production user cases involving the adhesive bonding of composite adherends.

For the repair user case, we inspected a circular shaft with a defined contamination consisting of DI and FP and approximately 200 mm in diameter. The measurement and analysis of the repair part were conducted as follows and as depicted in Fig. 5.35.

First, the calibration of the robot for manipulating and positioning the head of the AWT measuring device was performed. The acquisition of original image data was achieved by conducting measurements of different sections in the area of interest. All the obtained images were then processed using the Convolutional Neural Network (CNN) for droplet detection. Such processing by the neural network results in the generation of heatmaps that allow an assessment of the droplet count (and the surface density of the droplets), the droplet diameter, the certainty, and the wetting ratio. At this point, the evaluated measurement results were available. The next step was classifying the local state of the surface. We achieved this material surface and process-related interpretation with a support vector machine (SVM) based on the local values of the various heatmaps. A class training was performed using a by-hand delimitation of the defined surface states (namely clean and contaminated by DI or FP). The training of the SVM with the hand classification of the assessed surface states was based on 40% of the data for each class. Finally, the classification results for the entire surface were obtained through the trained support vector machine.

As a final result, we achieved a local classification of the contamination state indicated by an off-color map. This is the output of the SVM trained with the standard local output parameters of the CNN, namely droplet diameter, wetting ratio (percentage), certainty of the neural network output, and droplet count for determining the areal density of the droplets. The findings presented in Fig. 5.36 clearly indicate that the surface regions with either type of intentionally deposited contamination can be differentiated from the clean surface regions.

To characterize the surface of the technically realistic CFRP part, FTIR spectroscopy was applied. An impression of the measuring process using the portable spectrometer is presented in Fig. 5.36. Based on the spectral findings and the application of a partial least squares (PLS) algorithm for the surface region that had been intentionally contaminated with DI, a potassium concentration in the range of 9 at.% was indicated, which is considered plausible. However, a fairly high amount of DI was detected unexpectedly on the reference side of the CFRP surface, which is some centimeters away from the half of the scarfed area that had been contaminated with DI during the sample preparation. As the part was subsequently stored in an oven at elevated temperatures, we suggest that this treatment may have caused some of the DI to be transported to originally clean surface regions, e.g., following migration or after partial evaporation and subsequent condensation on the specimen surface.

During the three-day full-scale demonstration event, we tested the two distinct e-nose setups advanced in the ComBoNDT project.

Figure 5.37 illustrates how we applied the Airbus desorber device to conduct measurements on the technologically realistic parts.

Figure 5.38 presents an overview resulting from evaluating the data obtained on technologically realistic or real parts either in the production or the repair user case, benefitting from the PLS-DA model established during the testing of the pilot samples. We infer that the findings of all measurements can be separated into classes corresponding to clean or contaminated CFRP surface states, even for the complex geometries being tested. Once again, the threshold in the graph indicates the applicable parameter level differentiating surface regions that may be assigned to either class.

Performing cross-validation with the obtained, we obtain the confusion matrix of cross-validation, which is shown in Table 5.6.

Using the ENEA e-nose setup, the same investigation process was performed for the repair user case as for the production user case. Specifically, six measurements were taken from areas in the clean CFRP panel region. These were used to build the ellipsoidal 2-sigma coverage detector scheme for samples contaminated following the “DI-FP” scenario. Six measurements in the scarfed region were recorded for testing purposes: three from clean areas and three from contaminated areas.

The elicited response from the contaminated samples was significant and the contaminated samples were revealed without any false assignments. The obtained confusion matrix is presented in Table 5.7.

In the final step of the research, the LIF technique was used to study the surface composition of a real aircraft part, namely a CFRP panel with a stepped area for bonded repair, as displayed in Fig. 5.39.

In Fig. 5.39, we can distinguish eight numbered circular rings that represent separate areas of the inspection. The material layer numbered 1 incorporates the outer layer of the CFRP part without primer and copper mesh. The subsequent lower-lying seven rings (numbers 2 through 8) constitute the steps used (as adherends) for the bonded repair area. The left half of the specimen’s surface area is contaminated. The right side is clean and provides the reference area for the LIF measurements, which were made at three points at each of the eight areas. Separate measurements were made for the clean region and the contaminated part. The obtained fluorescence intensities at each of the three points were averaged and the standard deviation was calculated. The thus obtained results are presented in the plot in Fig. 5.40. In this presentation, the blue bars correspond to clean areas, while the red bars correspond to contaminated areas. It should be noted that clean areas are characterized by higher LIF intensities, except for areas 2 and 3. In the case of areas 1 and 5, the intensity obtained in the clean subarea is almost double that of the contaminated part. For the remaining cases, the differences are less significant. In conclusion, based on these investigations, the LIF intensity did not allow the surface states to be distinguished across all of the two distinctly prepared areas of the realistic aircraft part. This could be associated with the type of contamination used (“DI-FP”). However, based on the comprehensive set of LIF inspection data acquired here, there is evidence of a significantly inhomogeneous distribution of fluorescence intensities across the scarfed surface regions.

In the final step of the surface quality assessment of an Airbus A350 panel prepared for bonded repair, the vibrometry technique was used to study a realistic aircraft CFRP part.

As displayed in Fig. 5.41, the specimen to be inspected had been scarfed for bonded repair, and thus, a stepped region was locally exposed. In order to excite elastic waves in the part, a piezoelectric sensor was mounted close to the surface region of interest. As can be seen in Fig. 5.41, the disc-shaped piezo element was located on the panel surface area covered with primer, meaning it was some centimeters away from the grinded area (where the black CFRP is contrasted in Fig. 5.41) that surrounds the mechanically treated stepped region prepared for bonded repair. Again, we can distinguish eight (height) levels within the inspection area. The first is the CFRP without primer and copper mesh and the next seven levels correspond to the surfaces of the steps introduced by scarfing. The right half of the blackish area was intentionally contaminated. The left side was left clean and the surface areas in this region served as references for the inspection. As shown in Fig. 5.41, the piezoelectric disc was placed alongside the borderline between the two areas in a somewhat symmetrical position so that the elastic wave could propagate in both areas (the clean and contaminated halves of the surface) in the same way. The spatially resolved plot for the root mean square (RMS) of the vibration energy as obtained from the vibrometer software is depicted in Fig. 5.42. This is overlaid on the top view photograph of the inspected region of the CFRP part. As in Fig. 5.41, the contaminated area is on the right side. The total area prepared for bonding is located between two omega stringers, as may be perceived in the foreground of Fig. 5.41. The excitation of an elastic wave in this area results in concentrations of the wave energy between the stringers, which manifest as vertical lines in Fig. 5.42. Moreover, the central image section of the displayed RMS plot shows a circular pattern corresponding to the stepped area on the specimen’s surface. The data processing used for this part followed the whole area scan employed for the moisture-enriched samples and pilot level samples. Specifically, only the data points corresponding to the stepped area were considered here. Next, this set of data points was divided into two halves. The data points obtained in the surface region shown on the left half represented the referential case, and by processing them, the E (proportional to vibration energy) and 3σ values were calculated. The same procedure was conducted for data points from the right stepped half of the panel surface. The clean area is characterized by the values \({\text{\rm E}} = 3.58*10^{ - 7} V^{2}\), \(3\sigma = 9.96*10^{ - 7} V^{2}\), while the contaminated area showed \({\text{\rm E}} = 3.57*10^{ - 7} V^{2}\), \(3\sigma = 9.68*10^{ - 7} V^{2}\). We discern that the E values are practically identical and that there is a slight difference in the 3σ values. We conclude that these parameters do not allow the two areas of the realistic aircraft part to be distinguished, although the state of half the panel’s surface could be associated with the application of the contamination used in the “DI-FP” contamination scenario.

The challenge of the ENDT-based surface quality assessment consisted of identifying intentionally deposited combined DI and FP contamination on a scarfed real Airbus A350 CFRP panel. The surface area in contact with DI was more than one order of magnitude larger than in the case of FP. Moreover, indications were obtained that the surface area affected by potassium-containing substances from the DI expanded after the panel was oven-heated at elevated temperatures. Depending on the assessable area for each inspection method, the level of detail of the gained insights was different.

The findings achieved with the AWT clearly indicate that surface regions with either type of intentionally deposited contamination can be differentiated from clean surface regions.

The results obtained when we applied the enhanced e-nose techniques make us confident of the accuracy achieved with these devices for detecting “RA-FP” or “DI-FP” mixed contaminations at the presented levels of contamination, both in the production and repair user cases and with realistic and real parts.

Comprehensive LIF inspection data provided evidence of a significantly inhomogeneous distribution of fluorescence intensities across the scarfed surface regions, but the evaluation of the data subsets did not enable a reliable local differentiation between clean and contaminated regions. Based on the FTIR spectroscopic outcomes obtained with a portable spectrometer and the subsequent application of a partial least squares (PLS) algorithm, indications for an intentional deposition of DI were successfully substantiated.

Given the ample specimen geometry and the expectable substantial expenditure of time, the stepped surface area of the scarfed and stepped CFRP specimen was not investigated by laboratory-based reference analysis for reviewing the actually applied contamination degree.

As indications displayed by AWT and other inspection methods like FTIR differ with respect to the extent of DI-related contaminations, we recommend to utilize the presented enhanced ENDT approaches in future projects in order to perform additional tests for validating the data post-processing, e.g. by applying a PLS algorithm, when inspecting realistic parts.

In Fig. 5.43, we present a photograph of a real aircraft CFRP panel with a tailored and adhesively bonded circular repair patch. The two piezoelectric sensors attached to the top of the patched area are also visible. On the left, the sensor labeled DD5 is positioned over the clean bonding area, while the second sensor (DD8) shown on the right is placed over the contaminated region. Both sensors were attached in a gluing process using a cyanoacrylate adhesive.

The selected electromechanical impedance characteristics of the sensors where measured two times. First, the free sensors were characterized before gluing for referential purposes. Second, the signals from the bonded sensors were acquired on-site while the panel was lying freely on the laboratory table and in an arrangement similar to the one displayed in Fig. 5.43.

When we investigated the realistic part after the bonded repair for the repair user case, we did not observe any additional local resonances using the sensors. This is in contrast to the observations made for the production panel that was structured in a significantly different way. The results of the conductance (G) analysis are depicted in Fig. 5.44a. It can be seen that gluing the sensors causes the RMS values to drop. This finding indicates that the sensor vibrations have a lower amplitude, which is attributed to damping effects. The findings for the sensor over the clean region were quite similar to those for the one over the contaminated region. In Fig. 5.44b, we present the measured frequency shift of the conductance maximum. The respective shift for the sensor positioned over the clean bonding region was more than double for the sensor glued over the contaminated region. So, the frequency shift indicates a significant change in comparison to the behavior of the RMS values.

Next, we investigated the resistance curves in the same frequency range (3–5 MHz). The gluing of the sensor resulted in a drop in the respective RMS values (Fig. 5.45a). But the drop was not so significant as for the conductance. Interestingly, we observe that the differences between the RMS values of the two sensors are almost the same for the two investigated sensor states, i.e., the free sensor state and the glued sensor state. The values for the respective differences between the two sensors were found to be 4.25 and 4.53 Ω, respectively. Also, the frequency shift did not differ significantly between these two states (Fig. 5.45b), since almost the same value of shift is observed between the two free sensors and between the free sensors and sensor bonded to the contaminated bond area. The results for the resistance are interpreted as a low sensitivity of the high-frequency region of resistance to the supposed structural differences between the clean and intentionally contaminated bonding regions in the inspected part.

The fundamental frequency field obtained through the NLUS inspection of the real aircraft CFRP panel of the repair user case with the bonded circular patch is visualized in Fig. 5.46, and it shows a rather conventional standing wave pattern over the whole area of the repaired region. In contrast, the higher harmonic fields measured in the patch region (shown in Fig. 5.47) demonstrate quite different nonlinear field distributions. The second and third harmonics presented at the top and bottom of Fig. 5.47 reveal a strong enhancement of the nonlinearity and, therefore, indicate a lower bonding quality in regions of the patch that are discernible at the right-hand side and lower part of the patch.

To reiterate here, we also observed an enhancement of nonlinearity in joint regions with weaker bonds when we were mapping and imaging bonded joints based on locally contaminated adherends in realistic aircraft components. In both user cases, the accuracy of the accordance between the positions of the intentionally prepared areas with weakened bonds and the regions identified and visualized with the NLUS approach was confirmed by the respective manufacturers of the contaminated specimens.

In view of the challenges of a LASAT inspection, the repaired CFRP panel of the repair user case comprised an adhesive bond at different depths, depending on the positioning of the tested subzone with respect to the region of the repair patch. In consideration of this fact, we defined a specific time delay for each inspected subzone of the bond to set the maximum tensile stress induced during a LASAT shot on top of the respective bond location. Moreover, the CFRP panel featured stiffeners on one side. Therefore, we modified and adapted the testing procedure since access to both sides of the panel was not viable for every shot.

Figure 5.48 presents a sketch diagram showing the selected shot pattern as well as the general geometry of the panel.

The panel is partitioned into three distinct main areas, namely the stiffener area (highlighted in dark brown in Fig. 5.48 where the panel can only be accessed from one side, the stiffener flanges (lighter brown), where the stiffener is bonded to the panel, and the actual panel (light brown). As the whole panel was tested, S-LASAT was complemented by standard mono-shot LASAT in the stiffener area. The obtained results are summarized in Table 5.8.

Without applying the enhanced LASAT technique, we were unable to achieve a clear differentiation between a contaminated and a sound region of the patch on top of a stiffener. Both thresholds were alike, and from the ten LASAT shots performed on the contaminated area, only two showed indicators characteristic of an opening. The comparatively low energy required to open the bond might be explained by the actual state of the repaired panel. Numerous porosities were found by the ultrasonic test performed after the repair. Following the line of action for a quality assurance procedure of a repair process, this finding would be considered an indication that the workers did not perform a high-grade repair task. During the full-scale demonstration, we performed further inspection. Although the state of the panel did not meet the demands, we achieved proper discrimination between the repair region on the contaminated adherend surface area and the one on the sound adherend surface by applying S-LASAT. We observed that only the area tested on the contaminated part failed the LASAT test, and thus it was identified.

Advancing the S-LASAT technique enabled us to differentiate, for both inspected panels, the contaminated area from the sound one. We faced some limitations of the basic LASAT technique in assessing inspection regions with complex geometries: If one side of the CFRP panel was not accessible, the technique could not be applied and a non-optimized testing approach, such as the mono-shot setup, performed inadequately.

We consider the LASAT double-shock setup highly promising for overcoming the existing limitations in the case of complex panel shapes and we will be ready to perform further challenging studies in this vein.

Following our layout, a POD procedure is applicable for a given list of quantifiable features related to instantiated (sub)sets based on:

Regarding the user cases and scenarios from the ComBoNDT project, this means that each POD calculation is valid for one specimen type (i.e., contaminated CFRP for surface quality assessment, and assemblies comprising adhesive joints based on contaminated adherends for bonding quality assessment), one contamination scenario, and one ENDT technique and its associated diagnostic procedure. If one of these instantiations or elements changes, then the actual POD approach might change.

We designed the POD approach following the targeted purpose of accounting for the uncertainties in the capability of an ENDT technique to detect a given defect. Regarding the POD procedure for non-destructive examination (NDE), “The POD(a) function is defined as the proportion of all cracks of size that will be detected in a particular application of an NDE system” [2]. The NDT result is made non-deterministic because of the potentially numerous uncertainties. For example, the actual defect (e.g., fatigue cracks), procedure, material, mechanical repeatability, and human factor will influence the signal response of an NDT inspection. Therefore, a POD evaluation is a statistical estimation of the performance of an NDT with a consideration of the uncertainties, accounted for via the dataset. The POD estimation uses a sample set to gather data and is thus rather more “sample-dependent” than inherently representative of the basic population.

In practice, there are two main approaches to estimating a POD, namely the hit/miss model and the signal/response model. Some key elements are provided below, and more details can be found in the literature, such as Berens [2, 3], and other publications related to non-destructive inspection (NDI) procedures, e.g., ultrasonic inspection [4] and scheduling NDI [5].

The hit/miss approach refers to a binary diagnosis providing as its output a logical variable with the feasible values “detected” or “not detected”. For each flaw size a of the dataset, the “observed POD” POD_obs(a) is calculated as shown in Eq. (5.1). The POD function for all a is then obtained through a parametric regression on the POD_obs values. The Berens parametric regression model uses cumulative normal or log-normal functions.

It is generally recommended to have a minimum of 60 flawed inspection sites with at least three operators.

In this approach, the signal response S(a) from the defect with the flaw size a is used to give a quantified diagnostic with regard to a decision threshold value sth (S(a) > sth or S(a) < sth). For a given defect size a, S(a) is considered a normally distributed random variable. The probability density function of the signal value, g_a, has a mean value µ_a and a standard deviation ε independent from a (see Eq. 5.2). Then, the POD for flaw size a is defined as in Eq. (5.3)

Some hypotheses must hereby be verified:

ln(S) and ln(a) (or S and a) are linearly linked: ln(S) = β₀ + β₁ln(a) + δ, with ß₀ and ß₁ representing regression parameters and δ being a normally distributed variable with zero mean and variance σ_δ; σ_δ is independent of a.

Then, the POD function is a cumulative log-normal(µ, σ²) distribution function, with parameters shown in Eq. (5.4). The POD curve is then evaluated to estimate the β₀, β₁, and σ_δ that best fit the signal data.

It is recommended to have a minimum of 40 flawed inspection sites with at least three operators.

Regardless of the approach, the estimation of POD-based on several datasets for the same measuring problem leads to several POD curves and facilitates the evaluation of a confidence interval that is the dispersion in the POD results for a given amount of data obtained for the sample:

To separate and classify the measured analysis data, one important parameter is the detection threshold, and its determination can strongly influence the POD. We recommend defining it in a way that ensures that the defect signature is identified and differentiated from the noise and the saturation signal. In the case of the subsequently evaluated data obtained in the ComBoNDT research project, a saturation of the signal detection systems was not reached, and we consider the noise to be according to the signal in the case of the reference noise level was defined considering the signal on reference specimen. Therefore, we adjust the detection threshold so that it is, on the one hand, superior to the reference values obtained using cleaned adherend surfaces and, on the other hand, inferior to the lowest contamination values that could be differentiated from the reference values.

The model described above uses CIVA software, which is one of the main simulation software packages for NDT modeling.

In the upcoming section, which reports the results for selected example ENDT investigations, we always display the POD findings using the same layout, which is composed of three windows, as shown in Fig. 5.49. First, the top-left window represents the material-related data points as obtained with the respective techniques in the ComBoNDT project. The underlying system of coordinates shows the contamination values along the abscissa, as determined using a laboratory-based reference analysis. Meanwhile, the values for the feature being quantified based on the ENDT inspection are displayed along the ordinate. The dimensions and characteristics are individual for each technique and they might be, for example, amplitude values. We used the graphical user interface of this chart to adjust the detection threshold. Second, the top-right window displays the data scattering and a control parameter for the numerical model hypothesis, which is not detailed. Finally, the bottom window depicts the calculated POD curves, upon which the introduced probability of detection a90/95 is generally displayed.

Here, we explain our cutting-edge approach and highlight the caution that guided how we proceeded during our proof of concept.

As, to our knowledge, the POD approach has never been used for ENDT techniques and the associated material-features, and as we aimed to introduce a quantitative probability of detection for contaminants, we adapted the contamination values included here in a way that the POD-based model converges. Indeed, the model was established for defects that are conventionally monitored in aeronautical applications, such as cracks or delaminations, which are generally quantified referring to a length in millimeters, with relatively small values. Pragmatically and to avoid any numerical issues, we thus decided to adapt and scale the contamination-related values from the laboratory analysis and ENDT to remain in an expected range for the used numerical tool. We list the thus obtained values in Table 5.9, and we briefly outline our value implementation for each contamination scenario. Essentially, we used a consistent procedure for all the contamination scenarios which had been realized centrally and characterized by surface-sensitive XPS laboratory investigations in a standardized way. For the local implementations of contamination scenarios, we relied on the feature values that were individually measured in laboratories. Furthermore, in the case of double-digit feature values, we scaled them as described subsequently.

Finally, we would like to note that for all the contamination scenarios, the respective reference values were deliberately set as hypothetical values, and in short, are supposed to represent features of cleanly prepared CFRP specimens. We chose each value to be at least ten times inferior to the value of the lowest applied and investigated contamination level. To be specific, it cannot be equal to zero, neither in practice nor in the model that is supposed to correctly converge. We estimated this specimen state and the respectively measured abscissa values to be a good material-related approach to provide the noise level for each ENDT technique.

As a first example, we present our findings of the POD calculation for a feature obtained from the aerosol wetting test (AWT) data obtained by using the bonNDTinspect device for the RA contamination scenario within the production user case.

The first feature that we investigated was the average droplet diameter observed after temporarily depositing primary droplets from a water spray onto intentionally contaminated CFRP surfaces. The respective AWT results are presented in Table 5.9. In this case, the POD model hypotheses are respected because the data are linearly correlated with the various contamination levels and the scattering reduces with the increasing contamination. In the next step, we positioned the detection threshold to differentiate the findings of the minor contamination level “1” from those for the cleaned reference surface. We determined that the set of measured feature values obtained for level 1 partially overlap with the reference dataset. Specifically, one point exceeds this tentatively set threshold. This finding affects the POD for the AWT technique for this scenario, although the detection probability remains good and achieves a90/95 = 3.0 at.% for Si atoms (Fig. 5.51).

From the technological point of view, we would like to highlight that we will always be able to prepare CFRP surfaces with RA deposits that are so low that their AWT droplet pattern can barely be differentiated from that of a reference sample (or a set of reference samples). Thus, the technologically relevant aspect here is to highlight the importance that those contamination levels that significantly affect the fracture toughness, as an example of design-relevant criterion, can be reliably detected. As shown in Chap. 2 for the RA-1 contamination scenario, depositing an amount of Si-containing RA that corresponds to 3.0 at.% Si (as measured by XPS) is related to a loss of the average fracture toughness. In detail, G_IC and G_IIC values are reduced by almost 18% and 37% of the value respectively observed for joints prepared with CFRP adherends with surfaces treated according to a qualification process. The second AWT-based feature that we investigated was the wettability instead of the droplet diameter. Based on essentially the same raw dataset, the thus obtained results were slightly better, as can be inferred from Fig. 5.52. Indeed, the evaluated data points for the lowest investigated contamination level (level 1) were not in the range of the data for the reference surface state. When we used this feature, the detection threshold thus completely separated the reference data from the level 1 data. As a consequence, the POD curve was more significant. The calculated a90/95 value corresponded to 2.9 at.% of Si, which is very similar to the outcome of applying the droplet diameter feature.

As a conclusion of these data post-processing options for the AWT findings, our POD investigations have shown that in the case of the RA contamination scenario, the performance of the testing procedure depends on the selection of the feature to be evaluated. We found a better detection performance when using the wettability feature compared to the droplet diameter feature. As a next step, from the data post-processing point of view, we suggest performing a POD calculation by coupling the two approaches to determine whether the performance can be extended even further. From the test setup point of view, the area density of the applied aerosol droplets could be further adapted to be optimal for this scenario. We would like to highlight that at the end of the day, the multi-dimensional optimization of several parameters would be desirable, which would require extended modeling resources.

We performed POD calculations for several infrared spectroscopic (FTIR technique) datasets that had been performed for various contamination scenarios in both the production and repair user cases, namely moisture uptake in CFRP adherends, the deposition of de-icing fluid, and the thermal degradation of specimens. The respectively obtained results are presented as plots shown in Figs. 5.53, 5.54, and 5.55. For all the scenarios, the data are well suited for POD calculation; only the scattering in the moisture scenario is slightly diverging. Otherwise, the detection threshold can be easily placed to separate the reference values from the contamination values. With no overlap between the reference and the contamination data, the POD performances are good. The values are presented in the following:

For the LIBS technique, we performed in-depth POD investigations and here present the findings related to the aspect of assessing and increasing the robustness and reliability of the technological procedure. We highlight the LIBS findings here because numerous datasets, each with a considerable number of data points, were available or were made available upon request. This was supported firstly by the fact that LIBS is a spectroscopic, and thus multimodal, technique that provides several quantifiable features and secondly by the sub-millimeter width of the spot-like CFRP surface area assessed with a single LIBS shot. We performed these investigations for the RA contamination scenario of the production user case for CFRP surfaces. For LIBS, the actual testing procedure during data acquisition includes a high number of measurements across surface regions that are small compared to, e.g., the lateral dimension of the intentionally applied contaminations. This means that during the data evaluation, an average calculation within a (grouped) set of data points may be performed to compensate for the variability among the locally acquired data.

Based on these considerations, we evaluated in more detail how different approaches to data processing affect the quality of the procedure in assessing quality-relevant material surface features. In other words, our QA approach can be outlined as follows: First, the user operates the LIBS device to perform several measurements at various spots in a given area. During the processing of the thus achieved spectroscopic datasets, a spectroscopic feature is chosen—most often an intensity ratio of two spectral signals with one representing the RA deposit and one comprising the CFRP substrate. Then, an average is calculated for a group of feature values and is considered as a quality-relevant data processing result for the respective area. In the LIBS investigations described in Chap. 3, we experimentally established the number of involved LIBS shot measurements based on a confidence interval. Subsequently, we demonstrate how the procedure for refining this crucial criterion in the testing procedure may be enhanced by applying the POD tool. Specifically, for calculating the average values, we vary the size of the groups formed by the data points obtained from the raw data of the LIBS measurement that were acquired at neighboring surface positions.

Our starting point was a group size of one data point, meaning that we did not actually calculate the average and instead considered each single measurement point to be the result for the respective area that serves as the input data for the POD tool. The result of the POD calculation obtained for this boundary condition is presented in Fig. 5.55. In this case, 60 data points were used per contamination level (also including the reference surface state, which was also attributed to a contamination degree of 0.1). Clearly, these are raw data and no averages were calculated. The displayed data show that along with the increase of the contamination degree, there exists a considerable overlap between the reference data range and the data range for the contaminated samples. This finding hindered us from correctly positioning a detection threshold since either many points from contaminated samples would be below this threshold or some data points from non-contaminated surface positions would be above it. This uncertainty surrounding the POD result means that it is not satisfactory. The a90/95 does not even appear on the graph (as presented in Fig. 5.55). These outcomes indicate, as already outlined in Chap. 3, that it is mandatory to perform average calculations during the data processing to increase the reliability of QA approaches based on the LIBS technique.

Therefore, we moved on to develop testing procedures based on calculating the average values from the raw data obtained from several LIBS shots. First, we calculated the averages over groups of ten points, thereby obtaining the outcomes of the POD calculation, as presented in Fig. 5.56. We worked with 120 points per contamination degree, including the reference. To some extent, we used some of the same data points as in the abovementioned first calculation. This means that we used 12 (averaged) values per contamination degree as input data for the POD calculation. Initially, this dataset appears somewhat inadequate, but it was worth further inspection, and indeed, the POD calculation offered a better curve than previously. It was still not possible to perfectly differentiate data points for the reference specimen from the points for the first two contamination levels, but the third level was correctly determined based on the detection threshold. This enhanced approach thus led to a better probability of detection, even if the a90/95 value was not yet obtained.

Hence, we had gained sufficient evidence to suggest another series of measurements that were performed specifically for our POD investigations. Within this advanced procedure, 36 measurement points were included from each surface region of the CFRP specimen, and a total of 20 regions were inspected for each contamination degree (including the reference). The POD calculation results presented in Fig. 5.57 are thus based on an input dataset comprising 20 data points per degree, with each point being an average of 36 measurements. The achieved results show a much better probability of detection. We did not find any overlap between the data ranges of the different contamination degrees, and we could identify and even reduce the detection threshold. As all data points from the contaminated surface regions were above this detection threshold, the final POD result was better. The a90/95 was evaluated to be about 3.2 at.% of Si. As shown in Chap. 2 for the RA-1 contamination scenario, depositing an amount of Si-containing RA that corresponds to 3.2 at.% Si (as measured by XPS) is related to a loss of the G_IC and G_IIC values by almost 18% and 37%, respectively.

Briefly, in addition to this POD quantification, this in-depth study of surface QA processes based on the LIBS technique for the RA scenario has provided some interesting implications for establishing this testing procedure. For this contamination scenario, it emerged that doing averages over ten measurements for a total of 120 LIBS shots is not sufficient to obtain a good POD. By averaging 36 measurements and performing 720 analytical shots, the POD improves to the point that it can even be quantified. This overall data acquisition and evaluation process can be maintained if the technique meets the detection requirements set for testing the design-relevant mechanical performance of adhesive joints manufactured based on the assessed CFRP adherends. We expect that the detection performance could be further improved by evaluating more and/or larger average groups, as might be revealed by another POD investigation. Clearly, establishing the optimum lateral distance between neighboring LIBS spots should also consider the total measurement time required per CFRP surface area. In this respect, many of the presently applied aeronautical user cases that require a significant number of manually performed operations offer a sufficient time “buffer” for LIBS-based surface quality monitoring (Fig. 5.58).

As an example of ENDT techniques for bonding quality assessment, we performed POD investigations based on datasets obtained using nonlinear ultrasound (NLUS) for five different contamination scenarios from the repair and production user cases. The respectively obtained results for the P-FP, P-MO, P-RA, R-DI, and R-FC scenarios are presented in this section in, Fig. 5.59, Fig. 5.60, Fig. 5.61, Fig. 5.62, and Fig. 5.63, respectively. For this ENDT technique, the hit/miss approach was used. After examining the NLUS datasets for each scenario, this approach was motivated on the one hand by the finding that the correlation between the feature signal amplitude and the applied contamination level was never monotonous or linear. However, the values obtained for the adhesive joints prepared from intentionally contaminated adherends were well separated from those found for the reference specimens, and this finding enables an NLUS expert to make a diagnosis since, based on their experience, such a signal pattern is typical for ultrasound diagnostics. Effectively, when the signal amplitude exceeds a given threshold (called a “gate” in the case of conventional ultrasound), then a defect is indicated. This behavior is thus well suited to a procedure based on the hit/miss model approach.

The results are detailed in the captions of the respective figures and are summarized comprehensively below. We would like to remind the reader here that applying the hit/miss model requires 60 data points per contamination degree. As during our orienting approach we used only ten data points, we would like to highlight that some care shall be taken before drawing in-depth conclusions.

The NLUS data evaluation for the FC scenario required slightly more attention than the other scenarios. Indeed, in the case of the R-FC scenario, only one point from contamination level 3 was below the detection threshold that was placed above the reference data range. Since very few points were available, this point alone had a strong influence on the POD calculation as it had more weight. The a90/95 is equal to 2.5, meaning it lies between bond strength levels 2 and 3. If we consider this point an outlier and exclude it from the data range, with plausible justification, we may estimate what the POD would be without it or with more data (assuming additional points above the threshold). The result of such a consideration is given in Fig. 5.63. This leads to a90/95 = 0.677, i.e., below level 1, which is much better than when considering this assumed outlier to be significant. From a data-driven perspective, we thus conclude that it appears to be implausible to weigh and use this value unless an appropriate justification is given, e.g., through experimental replication or a physics-based explication. In essence, we recommend providing a sufficiently substantial and consolidated set of measured data points for any technically relevant POD approach.