Quantification of process-induced effects on fatigue life of short-glass-fiber-filled adhesive used in wind turbine rotor blades

Cracks in the adhesive joints of wind turbine blades can often cause severe structural blade damage. Effects induced by the manufacturing process can have a significant impact on the cycle number toward crack initiation in the bonding material. This research compared two experimental series of neat adhesive, dog-bone-shaped specimens with different degrees of cure, fiber orientation, and porosity content, which were subjected to static and cyclic tension-tension loading. To enhance the comparability of the two coupon series, the experimental data was normalized. Normalization factors were introduced to compensate for the different properties of the two series. These factors were derived from correction models. After normalization, the S-N values of the two data series were in good agreement in the high-cycle fatigue regime. In the very-low-cycle fatigue regime, however, there was still a lack of agreement in some cases, including the static strength. Moreover, the failure mechanisms were investigated by analyzing images from the fracture surfaces and dedicated samples using X-ray microscopy. The correction models presented can be applied to estimate the design strength of adhesives. This knowledge can then be used to make rotor blade structures more resistant to crack initiation.


Introduction
Cracks in adhesive joints of wind turbine rotor blades can often cause structural damage [1].During operation, tunneling cracks can be initiated in the adhesive layer, and propagate into the surrounding laminate, making structural repairs necessary [2].Such tunneling cracks can be initiated early in the operating life and even earlier still, during manufacture [3].Beside the impact of thermal residual stresses that develop in the adhesive layer during the manufacture [4], other process-induced effects of the manufacture can have a significant impact on the stress-life of a bond line.
The mechanical properties of the adhesive depend on several parameters.Some polymeric materials exhibit anisotropic properties due to the addition of short glass-fibers as reinforcement [5].Holst et al. [6] experimentally investigated the relation between short-glass-fiber orientation and the resulting coefficient of thermal expansion (CTE) and static tensile strength.Moreover, the manufacturing process itself induces effects that affect the performance, i.e., the curing cycle and the degree of cure produced [7]; the mixing of the adhesive resin and hardener and 2. Methods 2.1.Material, specimen manufacture, and testing The fatigue life of Hexion's EPIKOTE TM resin MGS TM BPR 135G3 and EPIKURE TM curing agent MGS TM BPH 137G [15] was investigated.This type of adhesive is based on epoxy resin and filled with short glass-fibers.
Two sets of specimens were considered, which were manufactured by different methods: (i) machine-assisted (MA) specimens and (ii) manually manufactured (MN) specimens.The MA specimen plates were manufactured by means of a mixing and dosing machine.A bead of adhesive was applied and pressed to a thickness of 8 mm between two aluminum plates, which were pretreated with a release agent.For the MN specimens, the resin and hardener were mixed by speed mixer and the adhesive paste was applied manually to a first glass fiber-reinforced polymer (GFRP) plate, which was pretreated with a release agent, in a meander pattern aligned in the transverse, or s-direction, of the specimen (Fig. 1).Then, another pretreated GFRP plate was placed on top to adjust the adhesive layer to a thickness of 8 mm, cf.[16].The MA plates were cured at 75 • C for 4 hours without exothermicity.The MN plates were cured at 80 • C for 4 hours with an exothermic peak of up to 120 • C. All specimen series were then cut with a water-jet.Beside their standard thickness, the specimens had a standard dog-bone geometry according to [17] (Fig. 1), and were used for both the static characterization with a stress ratio of R = 1.0, and the tension-tension fatigue characterization with a stress ratio of R = 0.1.The surface and edges of the MA specimens were smoothed with 180-grit and 1000-grit sandpaper.The specimen dimensions were measured with a dial gauge with a tolerance of ±1.6 µm.Static tensile and cyclic tests were conducted on calibrated 25 kN and 50 kN uni-axial servo-hydraulic tension-compression test machines, respectively, each with a class 0.5 load cell according to [18] with a measurement uncertainty of 0.18% and 0.16%, respectively.The frequency for the cyclic tests was kept between 1 Hz to 2 Hz to prevent a temperature increase caused by hysteretic energy dissipation.The target ambient temperature during the cyclic tests and a series of static tests was the standard class 1 room temperature of 23 • C [19].The surface temperature of each specimen was measured using a 100 Ω platinum Resistance Temperature IOP Publishing doi:10.1088/1757-899X/1293/1/0120423 Detector (RTD) from YAGEO with a tolerance of ±0.1 • C. Type GFLAB strain gauges from Tokyo Sokki Kenkyujo with a relatively low stiffness were installed parallel and transverse to the load axis, one on each side of a set of the static specimens.The glass transition temperature T g of each specimen tested was measured via Differential Scanning Calorimetry (DSC) using a DSC 204 F1 from Netzsch.Statistical evaluation was used to determine a standard deviation of 0.6 • C for the determination procedure.For further details, see Appendix A.
The maximum stress σ max , the stress ratio R, the cycles to failure N , the average surface temperature T , and the glass transition temperature T g of each tested specimen are summarized in Table 1.
Table 1.Summary of static and cyclic test results. No.

Short-fiber orientation correction
The short-fiber orientation in the adhesive plates manufactured was estimated according to the method described in [6].A cube-shaped thermo-mechanical analysis (TMA) sample was taken at four measuring points of an MN plate and four measuring points of an MA plate.According to [6], the determination of the linear CTE along all spatial axes allows the short-fiber orientation in the different types of specimen plates to be estimated on average in the form of a smeared fiber orientation tensor.The estimation method yielded a fiber orientation tensor entry of α MN zz = 0.08 in the longitudinal, or z-direction, of the specimen for the MN specimens and an entry of α MA zz = 0.75 for the MA specimens.Taking the relationship between tensile strength and fiber orientation from [6], which investigated the same material, the fiber orientation information in the longitudinal, or z-direction, of the specimen was used to adjust the tensile strength of the MN data points by a factor of R t (α MA zz = 0.75)/R t (α MN zz = 0.08) = 70 MPa/58 MPa = 1.2.

Temperature and degree-of-cure correction
The experimental results for the cyclic specimens listed in Table 1 were normalized to remove effects stemming from the average test temperature T and the degree of cure, the latter being affected by process parameters such as mix ratio, exothermicity, or post-curing cycle.On the basis of the hypothesis of a linear relationship between the static strength R t and T /T g as shown in [12,14], the data points were adjusted according to where T n stands for the normalization temperature, σ n max denotes the maximum stress under normalized conditions, and σ max stands for the actual maximum stress determined by the maximum force over the cross-sectional area F max /A.The gradient of the linear approximation is denoted as m R .Temperature values inserted into (1) must be in Kelvin.Therefore, a gradient of m R = 469.2MPa K K −1 (Fig. 2) was used, which was derived from the static MA specimens listed in Table 1.For the normalization of the data points, we chose the average T /T g value of the MA specimens: T n /T n g = 0.859 K K −1 as highlighted by the dotted vertical line in Fig. 2.

Porosity correction
To take account of the stress increase within the critical fracture cross-section, the sizes of the voids within the fracture surface were measured.The machine force measured during the test was ratioed to the cross-sectional area actually remaining.The ratios between the cross-sectional areas of voids on the fracture surface and the total area A v /A tot are given in Table 1.Neither the eccentric positioning nor the 3D shape of the voids was taken into consideration in the normalization of the stress data.

X-ray microscopy
For 3D X-ray microscopy (XRM), the sample MN S3 was scanned in a 360 • rotation conducted with the ZEISS Xradia 520 Versa system operated at the MAPEX Center for Materials and Processes, University of Bremen, Germany.The sample was imaged with 16.2 µm per voxel, using a setting of 80 kV, 7 W, and ZEISS filter LE1.Correction of ring artifacts and reconstruction of the spatial information on the linear attenuation coefficient of the sample was done using the ZEISS software.The images were post-processed through thresholding using OpenCV [20].

Test temperature and glass transition temperature dependence
We observed that the static strength R t of the MA specimens increased with decreasing glass transition temperature T g , see Fig. 2.Although the normalized strength values for porosity (P) and fiber orientation (F) of the MN specimens lay significantly below the trend line of the MA specimens, the gradient observed for the normalized values of the MN specimens is similar although the statistical basis of the data was small, only three data points being available.

Stress-strain curves
The raw stresses (Fig. 3a) of the static MN and MA specimens were adjusted (Fig. 3b).In a first step, the cross-sectional area for the MN specimens was adjusted to take account of porosity.In a second step, the effect of the fiber orientation was normalized such that the orientation of the MN specimens matched that of the MA specimens.And in a third step, the effect of the test temperature and glass transition temperature was normalized for the two series to match T n /T n g given in Fig. 2. In a final step, we adjusted the stress-strain curve of the MN specimens to match the stress-strain gradient (dσ/dε) of the MA specimens between ε = 0 µm m −1 and ε = 5000 µm m −1 .Now, the two curves should be comparable, meaning that all manufacturing effects have been ruled out enhancing the direct comparison between MN and MA specimens.As a result, the normalized MN curves indicate behavior which is more brittle than that indicated by the MA curves.Local stress and strain effects due to stress concentration are not captured in these global curves.Thus we can derive an experimental stress concentration factor of K e = 64.8/54.2= 1.19 from the remaining stress/strain gap between the MN and MA specimens.Note, however, that K e already includes the porosity correction.

Stress-life curves
After we had normalized the values of the cyclic MN specimens, their stress-life was in good agreement with the S-N trend of the normalized values for the MA specimens (Fig. 4a).The normalized maximum stress values σ n max are given in Table 1.Both data sets were approximated by probabilistic S-N models.As the normalized MN data exhibited a rather linear trend in a double-log grid, we first approximated these data according to Basquin's linear model [21], which is still state-of-the-art in the design of rotor blades, cf.[22].As the MA data was rather sigmoidal in shape, we approximated this data with a non-linear model suggested by Stüssi [23,24], which was slightly modified and reformulated by Rosemeier and Antoniou [13], cf.[25].For the sake of comparison, the Stüssi model was also used to approximate the normalized MN data.
In the cycle ranges where the S-N data of the two specimen series overlapped (gray shaded areas in Fig. 4b), the differences between the Stüssi and Basquin models approximating the normalized values for the MN specimens ranged between 5% in the HCF regime and −20% in the static and VLCF regime.Furthermore, for the static case in particular, these two models differed from a Stüssi model which approximates the stress-life of the MA specimens (Fig. 4b).
The largest impact on the stress level came from the degree-of-cure and temperature correction followed by the fiber orientation correction, and the porosity correction.The porosity correction appeared to have greater impact on the static strength than on the cyclic stress level.Next, a trial was carried out to approximate the stress-life probabilistically using both data sets.As the correction models applied in this research did not sufficiently capture the effects in the static and VLCF regimes, the static MN data points were not included in the approximation; this is discussed further in Sec. 4. Taking all adjusted cyclic data points, an approximation of the probabilistic Stüssi model then produced an overall standard deviation of R σ t,MN+MA,S = 2.14 MPa (Fig. 5).As the scatter of the MN data points was greater, this approximation yielded a larger standard deviation than the Stüssi approximation, which considered only the corrected MA data points.The endurance limit for the 50% failure probability of the MA+MN Stüssi model was lower (R e = 26.0MPa) than that of the MA Stüssi model (R e = 33.5 MPa).

Fracture surfaces
Considering the fracture surfaces of the MA (Fig. 6), we observed two types of crack initiation regions: (a) in the internal specimen structure, and (b) at the surface of the specimen.
While the three static specimens with the highest strength, i.e., MA S1, MA S2, and MA S3, exhibited a largely internal crack initiation, the two specimens with lowest static strength, i.e., MA S4 and MA S5, had crack initiation closer to the side surface, and in particular toward a specimen edge.
When considering the cyclic specimens, three specimens, i.e., MA C1, MA C3, and MA C4, had an initiation location at a void at the surface big enough to be visible to the human eye.The cross-sectional area of the void measured between 0.08 mm 2 and 0.37 mm 2 .The other cyclic specimen also exhibited crack initiation at the surface but concentrated more at an edge.The more load cycles a specimen was subjected to, the larger the fatigue zone, also called plasticization zone (Ger.Schwingbruchfläche [26]), around the crack initiation location became.The ratios between the cross-sectional area of the fatigue zone on the fracture surface and the total area A f /A tot are given in Table 1.
Considering the fracture surfaces of the MN specimens (Fig. 7), we observed no such clear areas of crack initiation as for the MA specimens, i.e., no fatigue zone.
A common feature of all fracture surfaces of the MN specimens is that they had relatively large voids.The cross-sectional area of the voids measured between 1.92 mm 2 and 9.28 mm 2 .Many voids were adjacent or close to the specimen surface.Only the fracture surface of specimen MN C7, which was subjected to the highest number of load cycles, had a void which seemed to be located in the internal structure of the specimen.

Fiber orientation
As methodologically shown in [6], an XRM scan of specimen MN S3 confirmed the fiber orientation tensor entries of α zz = 0.08, α ss = 0.87, and α tt = 0.05, which suggests that the short fibers were mainly oriented in the transverse, or s-direction, of the specimen (Fig. 8b,  c).We observed that the fiber orientation varied in the vicinity of larger voids, i.e., the fiber orientation changed from mainly s-direction (Fig. 8b) to an orientation in the specimen's zdirection (Fig. 8d).Moreover, we observed sharp edges, e.g., lobes or notches, at the large void close to the fracture surface (Fig. 8e).Smaller voids with a cylindrical morphology along the main fiber direction (Fig. 8b) and elliptic cross-section (Fig. 8c) were also present in the volume scanned.

Effect of voids
The correction steps presented in this research appear to be applicable in the HCF regime.Despite all the correction factors applied, there still seem to be differences in the VLCF regime, including in the static strengths of the MA and MN specimens (Fig. 4).This discrepancy can be partially explained.
Considering the relative sizes of the cross-sectional areas of the voids at the fracture surface A v /A tot (Table 1) of the static specimens MN S1, MN S2, and MN S3, they tend to be larger than those of all cyclic specimens MN CX with the exception of specimen MN C2.The stress concentration increases non-linearly with void size, cf.[27,28].This tendency is in agreement with our observations.
Considering the eccentric position of the void within the st-plane, i.e., whether the void is close to one specimen surface: The specimens MN C2 and MN C3 were subjected to the same stress level after correction.In the fracture surface of MN C2 one void which extends approx.50% in s-direction and is very close to the surface.Similarly, in the fracture surface MN C3 there is also one void which extends approx.25% in s-direction, also very close to the surface.Assuming voids are equidistant from the surface and circular in shape, a specimen with a void of a larger diameter is generally subjected to a higher stress concentration than a specimen with a void of a smaller diameter, cf.[28,29].This tendency is also in agreement with our observations.This research has not investigated nor taken into account for correction several effects which affect the stress concentration at a void edge: the eccentric position of the void within the stplane and the three-dimensional (3D) shape and size of the void, as well as the interaction of multiple voids, cf.[28].To quantify these effects on the stress concentration by means of models, XRM scans of all specimens would be required, preferably in the pre-mortem state, in order to capture the full 3D shape.
In general, fatigue experiments lead us to assume that a void subjected to a cyclic loading produces an effect which concentrates stress to a lesser extent than is predicted by theoretical elastic analysis [30].
The fiber orientation can change in the vicinity of large voids, and also locally, thereby affecting the stiffness and resistance to crack initiation locally (Fig. 8d).
Considering the effect of void size and shape on crack initiation in MN specimens: A correlation cannot be derived from images of the fracture surface.XRM scan images in the sz-plane or zt-plane, preferably in the pre-mortem state, could reveal more insights into the original void shape and possible sharp edges.

Effect of glass transition temperature
In contrast to what is evident in pure resin materials [7,14], the static strength of short-fiberglass-filled adhesives increased with decreasing degree of cure in terms of the glass transition temperature T g (Fig. 2).Although it is not clear why this happens, we assume that residual stresses develop at the resin-fiber interface.This could result in premature crack initiation at higher resin stiffness, which increases with degree of cure.Also, the increase in cross-linking with increasing degree of cure can have an effect on the load transfer between and around the short fibers.

Effect of fiber orientation
Considering the effect of fiber orientation on the fracture surface: The MN specimens had a transverse fiber orientation in the specimen's s-direction.Here, no fatigue zone was observed.In contrast, the MA specimens had a longitudinal fiber orientation in the specimen's z-direction and although smaller voids were observed in the fracture surface, a fatigue zone was clearly visible.One explanation could be that the crack propagation rate is faster when a crack propagates parallel to the main short-fiber direction, which is the case for the MN specimens, cf.[31].
The XRM images in Fig. 8 show in detail that the orientation of the short fibers is locally affected by the formation of voids as the adhesive is being applied manually during the manufacturing process.The short fibers rearrange in the direction of the void contour.This local fiber orientation differs from the estimated global smeared short-fiber orientation on which the normalization of the experimental data is based.This can be considered as another uncertainty in the analysis.For a more detailed view, the local short-fiber orientation in these areas can be determined using additional XRM scans in combination with software-assisted fiber segmentation.However, this increases the experimental and analytical effort.

Location of crack initiation under static and cyclic loading
Cyclic MA specimens without voids exhibited crack initiation at one specimen edge, while the best-performing static MA specimens exhibited initiation in the inside.This observation indicates that the mechanisms that lead to crack initiation after one cycle and n cycles must be different.Several effects can play a role, e.g., stiffness degradation, or relaxation during cyclic loading.These effects can "equalize" the stress level within the structure and prevent a crack being initiated at a local micro-defect, as is the case in the static MA specimens.Consequently, IOP Publishing doi:10.1088/1757-899X/1293/1/01204212 the specimen edge, being the most vulnerable part, becomes the location most susceptible to crack initiation.
The stress-strain curves (Fig. 3b) represent a global behavior as they are derived from machine force and strain gauge measurements.That is, they do not capture the local strain behavior at the strain concentration.If such strain concentrations are located at the specimen surface, they could be analyzed using Digital Image Correlation (DIC) techniques during cyclic loading, c.f. [32].

Cycle number to crack initiation
With the MA specimens, we observed that the fatigue zone grew larger, the lower the stress level and the higher the cycle number toward crack initiation.To obtain the actual cycle number toward crack initiation, the crack propagation cycle number from initiation until final rupture would be of interest.This correlation leads to an overestimation of the stress-life, which might be even greater at a lower stress level or in the VHCF regime.A method to quantify the crack propagation cycle number is required.This would allow the actual cycle number toward crack initiation to be determined by backward extrapolation.

Conclusions
This research is the first to be published in openly accessible English-language literature about how manufacturing process parameters such as curing cycle and temperature, processinduced effects such as fiber orientation, and effects of defects such as voids, can be quantified simultaneously and taken into account by models for a short-glass-fiber-filled adhesive material widely used in the wind industry.These effects were taken into account by correction models which describe their contribution to the mechanical performance of an adhesive material, under both static and cyclic loading.
The different S-N behavior in the HCF regime of two series of specimens with different degrees of cure, fiber orientation, and porosity was explained by correction models.A method whereby the manufacturing effects can be removed from the data on material strength, both static and fatigue, through the normalization of two different sets of experimental data was presented by way of example.The normalization of the degree of cure and fiber orientation had the largest impact on the stress-life.
All correction models applied were derived on the basis of simple standardized tests which can be performed quickly and with a limited budget.The correction models presented can be applied to estimate the design strength of adhesives.This knowledge can then be used to make rotor blade structures more resistant to crack initiation.
Future work should enhance the porosity correction method to model the stress concentration in more detail on the basis of the void eccentricity and shape, in order to explain the remaining post-normalization differences, including the static strength R t , between the two series in respect of the VLCF regime.

Figure 3 .
Figure3.Static stress-strain curves of neat adhesive specimens manufactured by means of machine-assisted application (MA) and manual application (MN): (a) raw data and (b) data normalized for porosity (P), fiber orientation (F), test temperature and glass transition temperature (T /T g ), and stress-strain gradient (dσ/dε).

Figure 4 .
Figure 4. Effect of manufacturing process on stress-life of neat adhesive tested at R = 1 and R = 0.1; specimens manufactured by means of machine-assisted application (MA) and manual application (MN): (a) adjusted for porosity (P), fiber orientation (F), test temperature and glass transition temperature (T /T g ), and respective S-N model approximations, and (b) differences in the model approximations between the MN and MA data sets.

Figure 5 .
Figure 5. Probabilistic Stüssi prediction considering all normalized data points of MA specimens and normalized data points of cyclic MN specimens.

Figure 6 .Figure 7 .
Figure 6.Fracture surfaces of static (S) and cyclic (C) machine-assisted (MA) manufactured specimens post-mortem; red arrows indicate point crack initiation, red-dashed lines indicate border of fatigue zone.

Figure 8 .
Figure 8. Views of specimen MN S3: (a) fracture surface, (b) thresholded XRM image through B-B plane, (c) thresholded XRM image through A-A plane, and (d,e) zoom into void edges.