Thermographic data analytics-based damage characterization in a large-scale composite structure under cyclic loading

Large-scale composite structures such as aircraft wings and wind turbine blades undergo cyclic loading in operation. In-service damage often generates excessive heat due to material frictions and it could be detected by thermography. This study develops a methodology to quantitatively analyze such structural damage based on thermographic data analytics. A full-scale composite wind turbine blade is inspected using passive thermography when it is subject to cyclic loads in laboratory. The damage region is identified from thermographic images and it is tracked automatically using image processing. The damage region is subsequently characterized on both overall and detailed levels. The change of the damage status versus fatigue cycle number is analyzed and the information regarding the growth of the damage area and the damage severity is provided. The initiation and the progress of damage are investigated based on the temperature and the enthalpy change in the damage region. This study provides a viable solution for efficient structural health monitoring and damage prognosis of large-scale composite structures under cyclic loading.


Introduction
Infrared thermography allows for non-contact and efficient damage detection in materials, structural components and full-scale structures [1][2][3]. When loaded dynamically, the damage region within a material can generate more heat than the intact region, e.g., due to material friction [4]. The damage region might be detected from thermographic images. Several studies have demonstrated successful application of thermography to detect cracks and delamination in large-scale composite structures such as wind turbine blades [5][6][7]. Nevertheless, damage detectability is still challenged by many factors such as surface emissivity and reflectivity of the structure [8,9]. Thermal image processing techniques, e.g., matching filters and thermal signal reconstruction [6], are used to improve the damage detectability. In [10], it was found that the norm of the first derivative of temperature and continuous wavelet transform in 2D were the most effective in detecting cracks that are perpendicular to the direction of heat propagation, whereas the Fourier transform and 1D continuous wavelet transform were successful in detecting delaminations. In [11], the authors described temperature evolution in the damage zone based on the thermoelastic equation. Three distinctive regions in the temperature curve were observed with the initial linear decrease the end of which corresponds to so-called damage stress below which the material did not show any detectable damage. Thermograms of a glass fiber composite plate containing hole defects of different sizes were analyzed in [12] using differential absolute contrast between the damaged and the damage-free regions. Progressive damage of a full-scale wind turbine blade subject to fatigue loading is detected using passive thermography while the blade is in motion [13].
The aforementioned studies focus on the detecting damage using thermography. After the damage is detected, it is important to evaluate the severity of damage. The damage evaluation requires quantitative analysis of the thermal images and an in-depth understanding of the underlying thermal-mechanical mechanism. This motivates the current study, which aims to develop a method to analyze the damage based on thermal images and provide insights into damage evolution in a structure under cyclic loading. The deveopled method is demonstrated on a full-scale composite wind turbine blade subject to fatigue loads in laboratory. The damage region of the blade is detected remotely based on passive thermography while the blade is in motion. The detected damage region is automatically tackled in the thermal videowhich is analyzed frame by frame using thermodynamics principles and image processing techniques. The growth of damage size and severity with the increase of fatigue cycle number are calculated and analyzed. Temperature variations inside the damage region and the energy dissipation mechanism are investigated in detail. As such, the novelty and the significance of this work are as follows: − We propose an efficient method to characterize progressive damage in large-scale composite structures under dynamic loading by analyzing thermal images. − The method evaluates the growth and the severity of damage noncontact and in near real-time using passive thermography without applying external heat source to the structure. − The method provides a possible solution for efficient structural health monitoring and damage prognosis of large-scale structures under dynamic loading.
The paper is organized as follows. Section 2 describes the framework of the proposed method. The image processing of the raw thermal video is presented and the image analysis is performed on both overall and detailed levels. Section 3 presents the experiment of the full-scale blade test under dynamic loading and how the raw thermal video is taken. Section 4 discusses the results from thermal image data analysis for quantitative damage characterization. Based on these findings, Section 5 summarizes the major conclusions and outlooks of this study.

Framework of the proposed method
This study develops a methodology for characterizing damage evolution under complex fatigue loading. This methodology is based on thermography image data analytics as shown in Fig. 1. Image preprocessing starts with extracting the image frames from a thermography video. The locations of strucutral damage can be identified due to temperature difference. RGB color images are converted into gray-scale to characterize the intensity of each pixel with just one value instead of three to ease further analysis. The damage region is zoomed in and isolated from the original image and is further analyzed. Due to the motion of the damage region in the structure under cyclic loading, an image correlation technique is used to align the coordinates of the damage region in all video frames. Subsquently, an intensity thresholding is applied to the gray-scale images to identify the damage. The thresholded gray-scale values are converted to temperature using minimum and maximum temperature range of thermography images. Finally, temperature distribution of the damage region is obtained for damage characterization.
Evolution of the damage under fatigue loading is analyzed on overall and detailed levels. Analysis outcomes on overall level provide information on the change of damage status, i.e., the growth of damage area and the increase of damage severity in three distinct stages. Detailed analysis, on the other hand, gives information on thermal and mechanical processes inside the damage region. The underlying thermal--mechanical mechanism behind the damage growth is investigated by dividing the damage region into zones of different temperatures and analyzing the interaction between these zones.

Experimental setup
The experiment has been conducted by a previous study [14], see Fig. 2. It is worthy to briefly describe the experimental setup here. The test object is a 14.3-m full-scale composite wind turbine blade made of unidirectional, biaxial, and triaxial fiberglass/polyester laminates, chopped fiber mats, and PVC foams. The blade was subject to cyclic loading during a fatigue test. A dual-axis electro-actuated exciter is used to induce a bi-axial cyclic load on the blade. Amplitudes of loading in flapwise and edgewise directions are set to ±8400 N and ±1200 N, respectively. Loading frequencies are selected to be equal to the fundamental resonant frequencies -2.25 Hz in flapwise and 4.37 Hz in edgewise directions.
Thermography video of the blade under cyclic loading is recorded using an infrared thermal camera FLIR A655SC containing an uncooled microbolometer detector with a resolution of 640 × 480 pixels and a pitch of 17 μm. The accuracy of measurement is ±2 • C or ±2 % of reading. The field of view (FoV) is 25 • × 19 • . The camera is placed 9.3 m above the ground and measures the temperature in its FoV during the fatigue test. No external heat source is applied to the blade. The surface damage grew under cyclic loading and generate thermal footprints that can be detected using passive thermography.

Extraction of the damage region
A thermal video with a frame rate of 10 fps is recorded which produces a total number of 751 image frames. Image resolution is 640 × 480 pixels. A thermal image is shown in Fig. 3 (a), while Fig. 3 (b) depicts a zoomed-in view of the surface damage, which is referred to as the damage region throughout the rest of the study.
The recorded RGB images were converted to grayscale images through the rgb2gray command in Matlab in order to ease the image analysis procedure. The grayscale image at frame 1 is shown in Fig. 4 (a) with the color bar showing the grayscale value intensities. The damage region is marked with a red circle and the pixel size 26 and 40 along x and y axis, respectively, is marked in the image.
A rectangular zoomed-in image of 40 × 26 pixels is selected as shown in Fig. 4 (b). In order to only retain relevant information related to the damage, a threshold was applied to the zoomed-in image region in Fig. 4 (b) to filter out the lower grayscale values. The selection of the threshold is based on the distribution of grayscale intensities and it is set to 150 in this study. The value is approximately in the upper quartile of the intensity range.
The gray-scale values of the damage region were transformed into temperature values. The maximum and the minimum temperatures of each frame in a thermography video were obtained. Subsequently, the temperature range was normalized by the maximum gray-scale value in an image frame as Conversion from gray-scale to temperature maps was achieved according to where T ij is a temperature map and G ij is a gray-scale map of damage region at pixels i = 1 : I and frames j = 1 : J with J = 751.

Alignment of image coordinates
The blade is moving during the fatigue test with respect to the stationary camera. The temperature profiles of the damage region as seen by the camera are not aligned from frame to frame. In order to track the damage growth by comparing the image information in all frames, the coordinates of frame images should be aligned. In this study, a 2D image correlation is used to assess the similarity of two matrices  coordinate misalignment for frames 1 and 2 is presented in Fig. 5 (a), while the same damage regions after alignment with a 2D image correlation are shown in Fig. 5 (b).

Damage characterization
Evolution of damage is analyzed on both overall and detailed levels. In the overall level analysis, the size and the severity of the damage at different stages of the fatigue loading are investigated. The detailed level analysis elaborates on the relationship between damage development, heat generation and enthalpy change.

Overall damage analysis
The increase of number of pixels in a damage region as seen in a thermal image is proportional to increase of damage area physically. In this study, a relative area increase A/A 0 is used to assess the growth of the damage size with respect to the original damage size in frame 1. The relative increase of the damage area versus the number of fatigue cycles is shown in Fig. 6 (a). It is known that internal energy of an object is proportional to its temperature. Enthalpy of an object is the sum of its internal energy and the product of its pressure and volume. Since in a solid the change in pressure and volume can be assumed negligible, the enthalpy change is proportional to the internal energy change with a good approximation. During fatigue damage of a material, the hysteresis energy is dissipated and will mainly be converted to heat. The dissipated heat will increase the internal energy of the material and, accordingly, its enthalpy. As shown in [15], the enthalpy change is mimicking a trend of hysteretic   energy dissipation. The behavior of hysteretic energy dissipation with increasing cycles of fatigue loading has three distinct stagesinitial growth, steady state growth and final growth. These three stages are attributed to damage progression from initiation to fast growth. Thus, damage severity at an arbitrary instance of fatigue loading can be qualitatively determined from the three stages of the enthalpy change curve. The enthalpy change is calculated as The values for the terms in Eq. (3) for the glass fiber composite material are as followsdensity ρ = 1835kg/m 3 and specific heat capacity at constant pressure c p = 903J/(kg × K) [15]. Ambient temperature was T a = 296.5K(i.e. 23.35 • C). Pixel volume V pix was calculated by calibrating the image resolution (number of pixels) to known dimensions of damage area through (4) where N d l = 40 and N d w = 26, l = 0.4m, w = 0.38m and t s = 0.0036m. After the calculation, V pix = 5.26 × 10 − 7 m 3 and enthalpy changes are illustrated in Fig. 6 (b).
A clear non-linear relationship of the enthalpy change curve with fatigue cycle number is obtained. The boundaries of damage severity stages are marked with vertical dashed lines drawn through the convex and concave points of the cubic polynomial curve. A cubic polynomial is used to fit the enthalpy change based on [15].
It can be seen that the enthalpy change curve qualitatively reflects three different stages of damage severity. It has a non-linear relationship with fatigue cycle number including an initial ramp-up at stage I, a nearly steady-state at stage II and a final ramp-up at stage III. Finding the concave and convex points of a cubic polynomial fit curve allows the partition the enthalpy change curve into progressive damage stagesthe cycle ranges of the three stages I, II and III are n f ∈ (0, 72), n f ∈ (72, 257) and n f ∈ (257, 328), respectively.

Detailed damage analysis
In order to investigate thermal features within the damage region, the temperature range was divided into several sub-ranges: T 1 ( • C) ∈ [25,27]; T 2 ( • C) ∈ (27,29] and T 3 ( • C) ∈ [29,30]. These three temperature sub-ranges denote three zones within the damage regionthe outer zone at T 1 , the middle zone at T 2 and the inner zone at T 3 . The growth of each individual zone can be explained by accounting for the energy balance of the zone. The energy is in a form of heat that flows from hotter areas to colder areas. The rate and the magnitude of the heat flow is characterized by the heat flux q → . The heat generated in the hottest inner zone under the fatigue loading due to material friction between surfaces of a crack is flowing outwards into the middle zone and further into the outer zone. Heat conduction in solids in two dimensions is expressed by Fourier's law of heat conduction where q(W/m 2 ) ̅̅̅̅̅ ̅→ , κ(W/(m × K)) and ∇T(K/m) are the heat flux vector, thermal conductivity and temperature gradient, respectively. For small temperature variation, κ is assumed constant. The temperature gradient is a vector pointing in the direction of increasing temperatures and its direction is opposite to the heat flux. Hence, heat flows from hotter to cooler zones.
The extracted temperature profiles of the damage region corresponding to the three stages of the enthalpy change curve are illustrated in Fig. 7  The non-linearity of the enthalpy change can be explained by considering Fig. 7. Fatigue loading results in the breakage of the chemical bonds of material. As a result, the rough fracture surfaces start sliding against each other which results in heat generation due to interface friction. Frictional heat generation rises the temperature in the damage region in thermography images, it can be seen as a spot of high temperature. The heat generated by frictional sliding propagates outward from the inner zone of the damage region according to the heat conduction law as shown in Eq. (5).
In stage I, damage initiation results in rapid sliding of the interfaces against each other which have relatively rough surfaces. Rapid sliding and relatively large friction coefficient at the interface lead to large heat generation and thus faster increase of temperature and enthalpy change, see Fig. 6 (b). At the beginning of fatigue loading shown in Fig. 7 (a)), there is only one site of damage initiation in the inner zone with a red color. In stage II, the site of damage initiation from stage I has not grown considerably whereas relatively steady sliding of the fracture surfaces of damages initiated in stage I result in relatively constant heat generation and accordingly a steady state increase in enthalpy change in stage II. There is a good correlation between the enthalpy change in stage II and the steady stage damage growth. Gradually, more damage initiation sites appear in stage II as seen in Fig. 7 (b), however, the damage growth phenomenon dominates the damage initiation. Eventually, the growth of the damages changes from a steady state in stage II to a rapid increase state in Stage III which in turn leads to rapid sliding of the fracture surfaces against each other and thus larger heat generation and enthalpy change. Moreover, the damages initiated in stage II, grow larger in stage III, as seen from the temperature contour of the damage region in Fig. 7 (b). To summarize, the enthalpy change is always increasing with fatigue cycle number due to new damage initiation sites, each of which acts as a heat generation source. The heat flux from the damage initiation area in the inner zone to the surrounding zones is schematically illustrated in Fig. 8.

Concluding remarks
This study presents a methodology to characterize structural damages in large-scale composite structures subjected to cyclic dynamic loading using thermal image analysis and enthalpy change phenomenon. The hysteresis energy dissipated during cyclic loading is converted to heat and accordingly results in enthalpy change of material. Damage analysis results revealed that the damage region grows continuously in size with the fatigue cycle number. The severity of the damage region is increasing nonlinearly with the fatigue cycle number and it can be classified into three progressive stages using enthalpy change.
The damage initiation in stage I results in large heat generation and accordingly a large increase of the enthalpy change. The steady-state damage growth in stage II leads to relatively constant heat generation and thus a steady-state increase of enthalpy change. The change of damage growth from a steady-state to a rapid increase state in stage III indicates severe fatigue damage reflected by considerable heat generation and accordingly enthalpy change.
The method presented in this study uses only thermographic data and thermodynamic principles for damage characterization remotely and quantitatively. The method provides a basis for further development of thermography-based structural health monitoring systems for largescale composite structures such as aircraft wings, wind turbine blades, and other engineering structures under cyclic loading. Further study may include tracking multiple damage regions and field applications considering environmental effects and complex dynamic loading.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments
This study is partly supported by the RELIABLADE project (Improving Blade Reliability through Application of Digital Twins over Entire Life Cycle, 64018-0068) through the Energy Technology Development and Demonstration Program (EUDP) of Denmark and by the AINDT project (AI-based NDT for reduced service costs for wind turbine blades) through the Innovation Fund Denmark.

Appendix A. Supplementary material
Supplementary data to this article can be found online at https://doi. org/10.1016/j.compstruct.2022.115525.  The inner zone at temperature t i is acting as a heat source out of which the heat is flowing to the middle zone, which is at temperature t m . The heat from the middle zone is flowing to the outer zone and, subsequently, to adjacent regions outside the damage.