Given the rise in the stage of composite materials, new joining technologies such as adhesive bonding have gained popularity in the aerospace industry. Adhesively bonded structures have a high strength-to-weight ratio and can join dissimilar materials and complex geometries. Additionally, adhesive joints preserve the structural integrity and smooth surface of composites compared to mechanical fasteners like rivets—i.e., composite structures—might be damaged via fiber breakage during riveting. However, not being able to determine the inclusions in the adhesive–composite interface may lead to significant strength reduction as well as unexpected catastrophic failures. In order to expand the application of adhesive joints, the geometry, size and position of these inclusions have to be identified via nondestructive testing (NDT) techniques [1
]. In our recent studies, interface defects in adhesive-bonded structures are inspected with ultrasonic and thermography nondestructive testing techniques [2
Ultrasonic NDT had been used to investigate the adhesive bond quality with adherend integrity and interface quality evaluation, including disbond detection [4
]. In addition, investigations to detect weak and kissing bonds have been performed in various studies [10
]. Conventional longitudinal pulse-echo ultrasonic inspection as well as advanced measurement techniques, such as acoustic microscopy, air-coupled ultrasound, and guided waves, have been used to evaluate bonding quality [13
]. Moreover, nonlinear behavior has been related to bonding quality with nonlinear ultrasonic NDT [20
]. Ultrasonic NDT has advantages to detect and position defects since it is a directional technique. While the classical pulse-echo technique outperforms the through transmission technique by being a one-sided inspection, it requires structure to be coupled with specific substances such as water. Air-coupled ultrasonics overcome this limitation; however, the high impedance difference between air and structures causes a significant loss in signal amplitude [14
]. On the other hand, guided wave inspections allow large specimens to be inspected in a short period of time; Lamb waves have been reported to be an effective technique to determine bonding quality [23
]. However, the analysis of the results has a high level of complexity, and it is usually specimen-specific.
Additionally, active thermography is a promising NDT technique to investigate bonding quality [25
]. Active thermography has advantages such as being very responsive, sensitive, noncontact, and suitable for automation; therefore, it is used to detect manufacturing defects within adhesive bonding [28
]. Defect detection with traditional light-based active thermography is highly influenced by thermal diffusion and the anisotropy of the structures. In the case of conductive adherends such as CFRP, induction thermography can reduce this limitation [2
]. Although induction thermography has many strengths and increased effectiveness by volumetric heating, the thickness of the structures and the complexity of the results limit its application [30
Moreover, shearography can have a very high resolution and short response time; however, it is only effective in the case of surface and subsurface defects and requires high-stress solicitation [31
]. It is reported that shearography is a suitable nondestructive testing technique to detect disbonding and subsurface defects in aluminum bonds [26
]. Also, where possible, X-ray tomography can be used to investigate inner defects in bonded structures [27
]. However, for composite-adhesive joints, similar diffraction coefficients and structures with a high level of aspect ratios might create limitations in this expensive NDT technique [32
]. Recently, electromechanical impedance mismatching and an adhesion quality test with laser shock had been proposed to evaluate bonding quality [31
]. While extended NDT for adhesive bonding is promising, these systems are expensive and costly to maintain.
The nondestructive evaluation of bonding quality is a challenging task because adhesive bonding is an interfacial phenomenon involving a thin layer of material, usually less than 10 microns [35
]. Although adhesive bonding evaluations with different nondestructive testing techniques have been performed over the past decades, the challenges continue to rise to establish the ultimate reliable NDT technique [3
]. Each NDT technique is limited to deliver a reliable evaluation of bonding quality due to its methodological and physical capabilities. Hence, we propose a combination of ultrasonic and induction thermography with feature-based data fusion.
Data fusion has been introduced to nondestructive testing and evaluation by Gros et al. and the research interest continues to rise [36
]. While the detailed categorization of data fusion reveals the advantages for sensor applications [38
], the survey on data fusion techniques for nondestructive evaluation also highlights numerous studies [39
]. The application on concrete samples mostly used ground-penetrating radar (GPR), impact echo and ultrasonic testing as data fusion resources while deploying several data fusion algorithms such as fuzzy logic [40
], artificial neural networks (ANN) [42
], Hadamard, and the Dempster–Shafer rule of combination [43
]. Considering the variety in nondestructive evaluation of composite structures, data fusion studies focused on several different combinations of NDT techniques. While Gusenbauer et al. [44
] improved porosity determination in composites with X-ray tomography and interferometer; Cuadra et al. [45
] monitored the damage in composites with acoustic emission, digital image correlation (DIC), and thermography. Cao et al. [46
] employed convolutional neural networks (CNN) in order to improve lock-in thermography imaging. Specifically, Daryabor and Safizadeh [1
] worked on the image fusion for ultrasonic and thermography nondestructive evaluation of epoxy patches between composite and aluminum structures. They compared several basic and complex fusion algorithms, namely minimum, maximum, average, principal component analysis, wavelet transformation and pyramid.
This work focuses on the evaluation of bonding quality with the fusion of ultrasonic inspection and induction thermography data. Composite-adhesive single-lap joints containing three different artificial debonding defects were investigated by both ultrasonic NDT and induction thermography. Saved data had been preprocessed for data fusion. The feature matrices emphasizing the defect presence have been extracted from ultrasonic and thermography data. These feature matrices have been used as the source of data fusion algorithms. The data fusion algorithms have been evaluated with quantitative sensitivity analysis. In addition to the previous works that focused on data fusion with ultrasonic nondestructive testing techniques and thermography, this work investigates different types of defects in composite-adhesive bonds and utilizes information theory-based data fusion algorithms. Also, this work contributes to the quantitative bonding quality evaluation efforts with receiver operating characteristic curves and area-under-curve calculations.
These results indicate that data fusion algorithms can improve the debonding type defect detection performance for bonding quality evaluation. In this work, three different cases of bonding quality have been investigated by ultrasonic immersion inspection and induction thermography. The obtained data had been preprocessed for data fusion with several steps. The feature matrices that have been extracted from each nondestructive testing method results were used as the source for data fusion algorithms. The data fusion algorithms have been evaluated with quantitative sensitivity analysis.
In the case study one, a composite-adhesive single-lap joint with release film debonding artificial defect was investigated. Ultrasonic inspection with 10 MHz central frequency focused transducer detects the artificial defect fairly well (Figure 7
a) due to high acoustic impedance difference between the air within the double-sided release film and single-lap joint interface. However, the induction thermography feature does not correlate well with the defect position (Figure 7
b); because the electrical conductivity level of release film is quite similar to the epoxy adhesive. In this case, for induction thermography inspection, the thermal wave dominates over Joule’s effect. Therefore, the fusion algorithms that are more focused on ultrasonics, such as weighted average 5UT-1TH (Figure 7
e) has higher performance than thermography dominant fusion algorithms (Figure 8
). As the difference fusion matrix shows defect presence in Figure 7
f and is evaluated with high values in area-under-curve calculations (Figure 8
), it can be said that two NDT techniques contradict each other in case study one.
In case study two, a composite-adhesive single-lap joint containing large brass film artificial debonding was investigated. Ultrasonic immersion investigation performs well with respect to qualitative and quantitative evaluation (Figure 9
a and Figure 10
). However, it does not indicate clear results as good as case study one even though the defect dimensions are the same. On the other hand, compared to case study one, the high electrical conductivity of interfacial inclusion results with high temperature contrasts in thermography investigation (Figure 9
b). Hence. The contradiction between two data fusion sources is much lower, as seen in the difference fusion algorithm (Figure 9
e) and observed low-values for the difference in area-under curve calculations (Figure 10
). While information theory-based fusion algorithms like DS perform quite well with detecting defects (Figure 9
h), the basic algorithm average is evaluated better in receiver operating characteristic curve (Figure 10
In case study three, the composite-adhesive single-lap joint with relatively small brass inclusions was investigated. Even though ultrasonic inspection results are improved by choosing a focused transducer rather than a flat transducer as in the previous work [3
], the defect detection performance with ultrasonic NDT is still limited, as seen in Figure 11
a and as evaluated by ROC curves (Figure 12
). Since the brass inclusion has high electrical conductivity like in case two, induction thermography performed well in defect detection qualitatively, as seen in (Figure 11
b) and quantitatively as calculated in the area-under-curve results (Figure 12
). While the contradiction between ultrasonic NDT and thermography is low according to the difference fusion algorithm (Figure 11
e and Figure 12
), both basic and information theory based fusion algorithms have increased the performance of separate techniques.
When considering ultrasonic inspection of adhesively bonded structures, transducer selection plays a significant role. The small defect detection performance is increased by changing from flat transducers to focused transducers. On the other hand, in order to obtain a clear ultrasonic response in the time domain -clear from the multiple reflections within the composite-adhesive bond, a high central frequency of the transducers is required. However, due to the high frequency, the highly attenuated composite adherend causes a drastic ultrasonic amplitude decrease, which makes the defect detection challenging.
The induction thermography results show that the brass inclusions have been detected with high performance. However, the release film inclusion at the interface is not detected with the same precision as the brass inclusions. This difference in the detection performance is caused by their electrical conductivity levels. While the brass is an electrically conductive material, which allows eddy current to form within, debonding with release film only affects the thermal diffusion. Therefore, induction thermography is a successful technique to detect inclusions that are electrically conductive, even for the small sizes.
It is important to mention that both ultrasonic inspection and induction thermography have advantages and limitations for bonding quality evaluation due to their physical and practical characteristics. Although ultrasonic inspection with the immersion technique is a successful method to detect debonding with release film inclusion, it requires the samples to be underwater, which may not be applicable to every specimen. Induction thermography is, on the other hand, a noncontact NDT technique that does not require any contact medium. However, the nonconductive material inclusions and air-induced delamination may not be determined as successful as ultrasonic inspection. As these inclusions represent possible foreign object introduction to the bonding area during the manufacturing stage, both conductive and nonconductive inclusions are significant. However, at the maintenance scenario where air gap and porosity at the bondline causes debonding, only nonconductive inclusion results should be considered.
Considering three cases, it is observed that the data fusion of ultrasonic NDT with induction thermography can increase the detection performance of defect detection. While information theory-based fusion algorithms like DS perform well, the basic fusion algorithms such as Hadamard and averaging cannot be disregarded. In case study 1, ultrasonic testing performs the best; therefore, each data fusion algorithms that are favoring ultrasonic inspection, such as weighted average 5UT-1TH performs well. Also, it is seen that the area-under-curve values for the difference is close to 1, which indicated that the data fusion sources (induction thermography and ultrasonic inspection feature results) are in contradiction. In case study 2, it is observed that averaging, DS, and Hadamard improves the results from different NDT techniques. On the other hand, case 3 highlights the importance of information theory-based method DS: while averaging evaluated as lower performance than thermography, DS-positive performs very well on defect detection.
Composite-adhesive bonding nondestructive evaluation is considered one of the most challenging NDT applications. This application study only covers the detection of debonding and might not be applicable to weak and kissing bond predictions. Also, the proposed nondestructive evaluations might not suit perfectly for bonding structures with different material properties such as dissimilar joints and aluminum bonded structures. It is important to point out that the contradiction between sources and the preprocessing steps affects the performance of data fusion significantly. The limitations observed in this work might be overcome by deep learning algorithms to emphasize different features from different sources and evaluate the contradiction with statistical-based algorithms.