A carbon nanotube-based thermoelectric generator integrated into a smart composite for structural health monitoring

The next generation of advanced composite materials needs to simultaneously address issues such as energy harvesting and structural health monitoring (SHM). The objective of this study is to explore, for the first time, the possibility of utilizing a build-in thermoelectric generator (TEG) to fulfil self-sensing purposes. To this end, carbon nanotube-based (CNT) inks are employed to print TEGs onto a glass fiber substrate, which is then incorporated into a glass fiber reinforced polymer (GFRP) laminate. The output characteristics of the TEG-enabled specimens are measured, displaying an exceptional performance. The specimens are subjected to static, quasi static cyclic and dynamic loading. Adopting a novel idea, the conductive, fully integrated printed path is then exploited to serve as a strain/damage sensor. For this reason, its resistance is monitored online during mechanical loading. To corroborate the findings, acoustic emission (AE) is simultaneously applied. Results reveal that the self-sensing multifunctional composite can successfully monitor its structural integrity. In fact, it demonstrates high sensitivity with a gauge factor approximately equal to 3. Moreover, when the TEG operates as a piezoresistive sensor, it is characterized by reliability. We thus believe that the herein suggested approach unveils new prospects regarding the efficiency and the sustainability of composite structures.


Introduction
In recent decades, advanced composites have seen a rapid rise in high-tech applications, thanks to their outstanding physical and mechanical properties.Unlike metals, composites provide greater specific strength, corrosion resistance, and superior fatigue behavior.Moreover, their mechanical characteristics can be tailored allowing for high structural strength at minimal weight [1], making them the primary load-bearing material in applications where high stiffness to weight ratio is critical.However, composites are intrinsically complex as they consist of two or more distinct phases, nonhomogenous, laminated, and lack plasticity.Thus, they exhibit various damage modes, such as matrix cracking, fiber pull-out, delamination and fiber fracture [2].These complexities necessitate the incorporation of higher safety factors in composite structure design, resulting in increased material usage, and consequently higher costs [3].
To overcome these drawbacks, prevent catastrophic failures and fully exploit the benefits of composites, regular inspections are needed.Periodic inspections are ordinarily carried out with the application of Non-Destructive Evaluation (NDE) techniques, used to assess the structural integrity of materials [4].However, these assessments often require the structure to be out of operation, incurring both time and cost, including standdown expenses.Therefore, there is a strong need for a Structural Health Monitoring (SHM) system capable of continuously monitoring a structure's integrity during its operational lifespan.The implementation of an SHM system directly correlates with reduced maintenance costs, enhanced safety, reliability, and efficiency [5].
Several techniques exist to diagnose the structural integrity of composite materials on-line.An approach for electrically conductive materials involves the monitoring of their electrical resistance, establishing a correlation between a material property and structural health [6].This approach falls into the category of self-sensing since the material itself acts as sensor.After Schulte et al [7] first demonstrated that the conductive fibers of a Carbon Fiber Reinforced Polymer (CFRP) can be used for strain/damage sensing, a considerable amount of research effort has been put into this method [8].The Electrical Resistance Chance Method (ERCM) has been proven effective in detecting flexural strain and damage [9], as well as delamination [10].In this approach the ΔR/R 0 parameter is adopted as an index of the material's structural integrity, which gradually increases with strain and abruptly when damage occurs.Some researchers have combined ERCM with acoustic emission (AE) to corroborate damage detection [11].Vavouliotis et al [12] employed resistance measurements to predict the remaining fatigue life of CFRP laminates.
However, the ERCM cannot be applied for nonconductive materials.To overcome this challenge, the scientific community has explored the possibility of incorporating electrically conductive carbon nanophases in non-conductive materials.CNTs not only reinforce structural materials but also introduce additional functionalities, such as sensing capabilities [13].When CNTs are dispersed in a composite matrix above the percolation threshold, they form a three-dimensional conductive network, rendering the initially insulating material, conductive [14].This enables resistance measurements to monitor strain and damage accumulation within the polymer matrix [15].As demonstrated by Grammatikos et al [16], the inclusion of CNTs into the matrix of CFRPs facilitates the employment of ERCM for detecting structural damage.In recent studies, graphene has been explored as an alternative to CNTs to impart the sensing functionality [17].
Another route for achieving sensing capability in composites involves depositing or printing a conductive path onto them.To this end, an ink or a film comprising conductive nanoparticles, such CNTs or silver [18] may be employed.Loyola et all [19] used spray-deposited thin CNT films to detect spatially distributed damage, while Hehr et al [20] embedded CNT threads into GFRPs for crack detection and monitoring.Other researchers have reported on the employment of piezoresistive CNT films or bucky papers to reliably measure strain [21][22][23].CNT based sensors may also exhibit sensitivity to local defects in the material [24] and can be simultaneously used to achieve strain monitoring and damage detection [25].
Thermal energy harvesting is another functionality that CNTs have demonstrated, owning to their semiconducting characteristics [26,27] Thus, they are suitable for use in thermoelectric (TE) materials, which can directly convert a temperature difference into electrical voltage, making them ideal for harnessing dissipated thermal energy.As reported by C. K. Mytafides et al [28] CNT-based TE devices exhibit outstanding flexibility and can therefore be used as wearables to harvest human body heat.The integration of this capability into composites could contribute to making structures more sustainable.In this direction there are already several studies reported.Karalis et al developed multifunctional laminates capable of generating electrical voltage from an in-plane [29] and a through-thickness applied temperature difference [30].Thermal energy harvesting in honeycomb structures is also recently documented [31].
Despite the plethora of works utilizing CNTs for sensing or thermal energy harvesting purposes independently, the possibility of simultaneously combining these two functionalities has not yet been explored.Therefore, the main goal of this study is to systematically investigate the feasibility of utilizing a TEG as an SHM sensor.In pursuit of this, p-and n-doped CNT-based inks are deposited with a specific architecture onto a glass fiber fabric, to form functional TEGs.The fabric is then incorporated as the top lamina of a cross-ply GFRP laminate, from which TEG-enabled specimens are extracted.The thermoelectric device exhibits a voltage output of 38.49 mV when exposed to a temperature difference (ΔΤ) of 75 K.Subsequently, the printed conductive path is exploited as a strain/damage sensor.Electrical resistance measurements performed during tensile, cyclic, and fatigue loading reveal that variations in the electrical resistance provide a means for assessing the strain and damage status of the material.AE is concurrently used to validate these findings.
Considering the herein successfully demonstrated dual functionalities, combined with the straightforward device fabrication process, suggest that the implementation of a TEG/SHM sensor shows high possibility of enhancing the safety, reliability, and energy efficiency of future lightweight composite structures.

Materials
The p-and n-type SWCNT-based thermoelectric inks were prepared using TUBALL TM SWCNT powder from OCSiAl (Luxembourg).Poly(ethyleneimine) (PEI), PEDOT:PSS ink, and Sodium dodecylbenzenesulfonate (SDBS) supplied by Agfa (Belgium) were used as dopants and dispersant agent respectively.More information about the materials used can be found in a previous study of ours [29].For manufacturing of the GFRP laminates, a unidirectional (UD) glass fiber fabric with 320 gr/m 2 density and 0.26 mm ply thickness from Fibermax (Greece) was used.The resin system of Araldite LY5052 low viscosity epoxy resin and Aradur 5052 (HY5052) polyamines-based hardener from Huntsman Advanced Materials (Switzerland) was utilized as the composite's matrix.Resin to hardener weight ratio is 100:38% w/w, while the T g is 131 °C.Copper (Cu) sheets were used as electrodes.The AGG3790 conductive silver paint obtained from Agar Scientific (UK) and the RS Pro silver loaded conductive epoxy paste supplied from RS Components (UK) were used to establish ohmic contacts between the electrodes and the TEG.

Ink preparation and fabrication of the TEG
The preparation process of the thermoelectric inks is described in detail in [29].The TEG architecture adopted here consists of alternating p-type and n-type elements, forming the, thermocouple, that are connected electrically in series, and thermally in parallel [28,32].This layout can lead to large electrical output, despite the small output of the individual thermocouples [27,28].A drop casting process was implemented to carry out the printing of the carbon-based TEG module.The inks were deposited by hand using a pipette onto the surface of a glass fiber ply.The ply was positioned atop a hot plate operating at 100 °C.An adhesive vinyl mask was initially applied onto the fabric to facilitate the printing process and prevent the formation of p-n junction at undesirable locations.The heated substrate and the relatively high viscosity of the inks are crucial to prevent ink absorption by the fabric, avoiding potential short-circuits between thermoelectric legs.Although the printing process in this study is manual, it can readily be scaled up or automated by integrating blade-coating, inkjet printing, or a robotic precision fluid dispersion system, to meet future industrial demands.
Figure 1(a) depicts the TEG architecture.Each TEG consists of 13 thermocouples, that is 13 p-and 13 nthermoelements, connected in series that create a continuous electric path.Each individual element is 22 mm long and 5 mm wide, with an interspace of 1 mm.To ease the printing process, the elements align with the fiber's direction.The p-n junctions, with dimensions of 11 × 5 mm 2 , were formed by overlapping the two functional inks at the intersection region between the two legs.The total printed area adds up to 32 × 155 mm 2 .The dimensions of the TEG were purposely chosen for extracting coupons that demonstrate significant power output and are suitable for mechanical testing.After completing the printing process, the fabric was left on the hot plate for about an hour to dry.
In the implemented configuration, a continuous conductive path of alternating p-and n-doped SWCNT networks is formed, eliminating the need for additional metal electrical interconnections, which minimizes the contact resistance between thermoelectric elements [27,28,33].Considering that the electrical conductivity increases with the CNT content [32], 20 passes of ink deposition were performed in total.

Laminate manufacturing and specimen preparation
Cross-ply GFRP laminates with [90/0] 2s layup were manufactured by hand lay-up using the UD glass fiber fabrics and the resin system.The fabric with the printed TEG was positioned as the top lamina so as not to adversely affect the mechanical properties of the laminate.The apparent interfacial shear strength between glass fibers and epoxy resin is reduced when the fibers are modified with TE inks containing surfactants like SDBS [34].It is noted that the CNT sensor incurs only a minimal mass penalty, given its negligible weight.More notably, the printing of the CNT array onto the material minimizes the possibility of sensor debonding [35].The [90/0] 2s lamination sequence was chosen since printing thermoelectric legs parallel to fibers facilitates the procedure.After layup, the composites were placed in a hot-press for 24h at room temperature to cure, followed by 4 h post-curing at 100 °C.The applied pressure was set at 3 MPa while a metal spacer of 2 mm thickness was used.
Tensile specimens with dimensions of 240 × 35 × 2 mm 3 were cut from the laminates using a waterlubricated diamond saw.GFRP tabs 35 mm long were attached using a two-component adhesive, resulting to a gage length of 170 mm.Specimen edges underwent grinding with incrementally finer sandpaper up to 2000 mesh, followed by polishing with a 5 μm diamond spray.
Copper sheets were mounted on the two edges of the TEG to perform the electrical measurements.To achieve ohmic contact the excess of resin was removed, until the ends of the CNT path were revealed.The exposed surfaces were then cleaned with and acetone, dried and then painted with conductive silver paint.The electrodes were attached onto the painted areas using silver loaded epoxy paste.After 24 h of curing at room temperature contact was established and the areas were sealed with resin coating.Figure 2 presents a schematic diagram of the steps involved in the printing and specimen manufacturing process, as well as the layout of the TEG-enable specimen.

Thermoelectric measurements
The TE output characteristics of the TEG were measured using a custom setup.Specifically, 10 mm of the longitudinal edge of the TEG, encompassing 5 mm of the p-n junctions plus 5 mm of each TE leg, were positioned on a hot plate to be heated (figure 1(b)).The remaining portion of the TEG was placed on a large metal plate and kept at ambient temperature to establish the desired ΔΤ.To determine the ΔT between the hot and cold sections, temperature was continuously monitored with the P 5180 PEAKTECH data logger using K-type thermocouples.The electrical measurements were performed with an Agilent 34401A6½ digital multimeter using the two-probe method.The internal electrical resistance R TEG , the generated voltage V OC , and the generated electrical current I SC , were recorded for multiple values of applied ΔT.Thermal images (figure 1(c)) were captured with the FLIR ONE pro thermal camera.Power output measurements were carried out using the ELC Resistance Decade Box.

Mechanical testing
Mechanical characterization and subsequent on-line monitoring were conducted using an Instron 8800 Universal Testing Machine equipped with a 100 kN loadcell.All tests were carried out under ambient conditions.Static tensile tests were performed according to ASTM-D3039/D3039M using the Bluehill software.
Testing was performed at a constant crosshead speed of 2 mm min −1 .Five specimens were considered in total.To monitor strain the Instron AVE video extensometer was used.Tension-tension fatigue testing was carried out as per ASTM D3479/D3479M utilizing the WaveMatrix software.Testing was implemented with a load ratio R (=σ min /σ max ) = 0.1 and a frequency of 5 Hz.A load control method was employed, where a sinusoidal wave of constant amplitude was applied.Fatigue tests were performed at stress levels equal to 30% − 80%, increasing by 10% of the ultimate tensile strength.If the specimen did not fail at 10 6 load cycles, the test was terminated.

Resistance change measurements
In order to explore the possibility of the TEG to act as a strain/damage sensor two-point electrical measurements were carried out during mechanical loading.The Agilent 34401A6½ multimeter was employed to record in real time the resistance values, while the Agilent DMM software was utilized for data recording.The sampling rate was set at 10 Hz.Specimens were subjected to monotonic tensile, quasi-static cyclic tensile and fatigue loading.Two specimens were tested at each loading scenario to confirm the repeatability of the results.
For tensile testing, a displacement rate of 2 mm min −1 was employed.The loading-unloading sequence involved incremental loading cycles under a displacement control mode of a 2 mm min −1 .The loading step was 3 kN, increasing by the same amount at each consecutive cycle.For SHM during fatigue testing a frequency of 2 Hz was chosen.Five readings of electrical resistance per fatigue cycle were acquired to accurately represent the sinusoidal function.Tests were carried out at 70% of the maximum tensile strength.Specimens were preloaded up to the σ mean for 30 s prior to the dynamic loading application.To investigate if the fatigue loading affects the temperature of the specimens and by extension the recorded electrical resistance, temperature was monitored with a K-type thermocouple.The results indicated that no correction to the resistance values based on temperature was necessary, as there was no induced temperature change due to the low frequency implemented.

Acoustic emission monitoring
AE was employed concurrently with the ERCM to verify the results arising from the electrical measurements.The AE activity was recorded during mechanical testing using two R15-Alpha narrow band, high sensitivity, resonant sensors by Physical Acoustics Corporation.The sensors were attached to the edges of the specimens' span.Acoustic coupling was achieved by applying a water-based ultrasonic gel.Sensor output was amplified by 40 dB at 28 V with an AE signal preamplifier.The cutoff threshold was set at 40 dB to avoid noise deriving from the environment.The signals were recorded in the 40 MHz PCI-2 data acquisition and digital signal processing board by Physical Acoustics Corporation.The AEwin software from Physical Acoustics Corporation was utilized for real-time viewing and analysis of the recorded acoustic profile.After AE recording, a filter was applied to the acquired raw data so that the events with energy below 30 aJ, that are related to noise activity [36], were cut off.A schematic layout of the experimental testing setup is illustrated in figure 3.  (R TEG ), open circuit voltage (V OC ) and generated or short-circuit current (I sc ) as well as the calculated power and power density with respect to the applied ΔT.The maximum power output of the TEG-enabled laminate is calculated from the equation [27,30]:

Results and discussion
where ΔV is the generated voltage.Power is divided by the area of the TEG to derive the corresponding power density.As expected, all the output characteristics of the printed TEG increase proportionally with ΔT, indicating that the higher the thermal difference the more energy is available for harvesting.Within the applied temperature range, R, V and I exhibit a linear dependence on the ΔT, whereas power increases exponentially [27,37,38].The values of the aforementioned TE characteristics, as determined for different temperature gradients (ΔT = 50 K, 75 K and 100 K), are summarized in table 1.These values are very close to the ones reported in literature for a similar approach [33,38,39].In such configurations, enforced cooling is commonly reported [30,40,41] in order to sustain the desired power-generation characteristics.In the hereby reported TEG configuration, sufficient heat dissipation is achieved to maintain thermal gradient equilibrium, eliminating the need of an external cooling mechanism.
Figure 5(a) depicts the TE voltage and power versus the applied different external load resistance (V-R load , P-R load ), while in figure 5(b) the power output characteristics are presented as a function of the generated electrical current (V-I, P-I).In both diagrams, the continuous curves are derived using the following equation [42].
The individual points correspond to the experimentally measured values.The measurements are performed for several values of resistance connected in-series.The maximum power is generated when the externally applied resistance matches the internal resistance of the TEG, meaning that the maximum amount of available power is dissipated as heat in the load resistor.This is also manifested in the P-I curve, where the maximum power occurs at the output current for which the external resistance matches the internal.It can also be observed that the output voltage continuously increases as a function of R load , approaching asymptotically a stable value.This  value is the V OC .Lastly, the output voltage, for different applied R load is inversely proportional to the generated current.
As is presented in the above figures, the manufactured TEG displays a similar performance compared to literature values [29,31,33].This may be attributed to the reduced electrical resistance of the continuously printed CNT path which avoids metal electrodes to connect the individual TE legs [27,28,33,38].The adopted TEG design allows for upscaling by simply increasing the in series connected TE legs.In that case the circuit resistance would also increase linearly [33].

Mechanical characterization
Figure 6(a) illustrates the tensile stress-strain curves of the five tested specimens.Specimens fail at 462.13 ± 9.62 MPa corresponding to 2.69 ± 0.03% strain and have 25.34 ± 0.22 GPa stiffness.Damage initiation and accumulation in cross-ply laminates follows a specific sequence.Firstly, transverse matrix cracking initiates in the plies perpendicular to the applied load due to local stress concentration and propagates along the fibers.
Transverse cracks terminate at the interfaces between 90°−0°plies.When crack density reaches saturation, the stress concentrated at the tips of the cracks triggers delaminations.Subsequently, the longitudinal plies have to carry a larger portion of the load resulting in fiber breakage [43,44] and finally in specimen failure.The specimens show an almost linear behavior, with the exception of a knee point typical for cross-ply laminates under uniaxial tension.This occurs at low strain of about 0.2%-0.5% and corresponds to transverse cracking in 90°plies [45], which causes stiffness loss.The stiffness of the specimens is then dominated by the 0°-oriented plies [46].
Figure 6(b) shows the S-N curves obtained by fatigue testing.The ordinate depicts the fatigue stress levels normalized with respect to tensile strength (σ/σ UTS ), which is plotted against the number of cycles (Ν) up to ultimate failure.The experimental fatigue data are fitted with a straight dashed line, or the fatigue fracture line.The line has a negative slope and represents the fatigue strength, meaning that for every possible σ/σ UTS -N combinations above it the material is fractured.For 30% stress level no fatigue limit is observed.The fatigue performance of cross-ply laminates is in general determined by the strength of the axial plies, while the  contribution of the transverse plies may be regarded as negligible [47].It is finally noted that since the TEG bearing ply is the outer and the loading is off-axis its does not affect the mechanical behavior of the laminate [34].
3.3.On-line monitoring 3.3.1.Tensile loading Figure 7(a) depicts the stress-strain curve, presented in conjunction with the electrical resistance change of the TEG normalized by the initial resistance (ΔR/R 0 ) and the cumulative AE hits.A simple comparison of the plots reveals that the printed path can serve as a strain/damage sensor and therefore fulfils self-sensing purposes.This is feasible due to many reasons.The TEG possesses high conductivity as it consists of a deposited thin CNT film, the strain-sensing of which is based on its piezoresistivity.As seen in figure 1, the architecture of the TEG resembles that of a conventional strain gauge.More importantly, due to the resin bonding effect the TEG follows exactly the deformation of the specimen.Therefore, the resistance change of the TEG is due to its geometrical alteration.The intrinsic resistance change of the CNTs may also contribute, while the intertube resistance has no effect, since the TEG device is embedded in the resin matrix and is not a freestanding CNT buckypaper [24,25,48].CNT-based sensors are generally characterized by high resilience and flexibility, regardless if they are in the form of a bucky paper [28] or of a coated glass fiber fabric [29].The polymers used here the TE inks preparation may contribute to this behavior.As a result of those attributes, the application of the ERCM provides information about the structural integrity of the tested specimens.
In more detail, as the specimen length increases due to the applied tensile strain, the conductive path stretches, experiencing the same deformation.Therefore, the TEG serving as strain sensor exhibits an almost linear monotonic increase in its resistance.A very small deviation from linearity can be observed at the knee point attributed to transverse cracking initiation and stiffness loss.This is also the onset point of the AE activity [49], as the preceding microscale damage is excluded due to the cutoff threshold and the filter used.Another bend in the relative increase of the resistance, although this time larger, is observed at about 2% strain.This could be the saturation point of the transverse cracking, as it is in good agreement with the AE findings, where the AE activity also shows a rise.The intrinsic resistance of the CNTs, which increases exponentially with respect to the applied strain could also play a role to the non-linear behavior [48], especially in small strain regime [25].
Generally, in resistive strain sensors, the resistance of the material changes with respect to the applied strain due to the piezoresistive effect.The relative resistance change can be directly correlated to the axial strain via the gauge factor (GF) or strain factor, which determines the sensitivity of the piezoresistive sensor.The GF is the slope of the curve depicted in figure 7(b) and is given by the equation: where R is the resistance, L is the length, and ε is the strain [8].Experimental data can be linearly fitted with R 2 = 0.992, so the linear approximation is retained.The experimental GF value is ca. 3.This value is comparable to the GF of conventional strain gauges, being between 2 and 5, CFRPs, being equal to 2 [8], and strain sensors consisting a continuous CNTs film that have a GF of 1-5 [21,24,25,50].It is noted that the TEG shows only positive piezoresistivity and thus no artifacts due to poor contact resistance occur [8].This is observed even though the two-probe method was used, and the electrical circuit was closing within the testing length of the specimen, a fact that could result in strain induced contact deterioration.Moreover, the TEG responds immediately to the applied deformation from the onset of the test, so no lower sensitivity limit is observed.Lastly, the resistance change curve is smooth, with minimal or no noise.Taking of all the above observations into account it can be deduced that the printed TEG is a sensitive and reliable strain sensor.At the moment of final failure, the normalized resistance increases abruptly by several orders of magnitude, indicating the complete disruption of the conductive path.Prior to this point no other abrupt changes or discontinuities are detected either on the stress-strain curve or on the resistance change curve, indicating no other large-scale damage.Therefore, the TEG can not only serve as a strain sensor but as an integrated and consistent self-sensing system, capable of monitoring both strain and damage.It should be however mentioned that any information regarding the structural integrity of the material refer only to the outer ply, since the TEG was printed onto the surface.Internal damage e.g., delamination may only be detected if it leads to events affecting the whole specimen such as stiffness degradation or load drop.

Cyclic tensile loading-unloading
In figure 8(a) mechanical stress and strain under cyclic loading are shown versus the relative resistance change.From the diagram it becomes obvious that the change in the resistance directly follows the load pattern.Resistance increases during tensile loading and decreases during unloading.This means that the TEG acting as a sensor can measure the strain signal from the composite.As in tensile testing, the TEG is also in this case sensitive to strain from the onset of the loading.The correlation between the response of the sensor and the applied mechanical loading during every cycle supports its durability.
During each loading cycle the material undergoes constantly higher elastic deformation than in the previous cycle, thus the resistance change reaches higher values.Upon unloading the resistance starts to decrease, until it reaches a minimum value.However, the resistance variation does not return to zero since there is permanent resistance change even in the fully unloaded state.This irreversible increase corresponds to the small residual strain caused by matrix cracking and softening of the material [16,51].More importantly it enables early detection of even small-scale damage, which if accumulated can lead to failure of the material.
The permanent resistance change becomes higher in every cycle compared to the previous one.This can be seen in figure 8(a), but is better distinguished in figure 8(b), where the ΔR min /R 0 and ΔR max /R 0 of each cycle are plotted versus time.The line connecting the minima of ΔR/R 0 or in other words the 'zero-load line' (ZLL) is a linear function of time with a constant slope.As discussed above, this trend is indicative of residual strain due to damage accumulation, while its linearity suggests steady accumulation.Consequently, it represents the irreversible phenomena arising from material degradation.The linear relationship also infers that the printed strain sensor experiences no Poisson effect or other artifacts as is often the case with CFRPs [16].Likewise, the 'peak load line' (PLL) results from connecting the maximum resistance change values of each cycle.Therefore, it includes both reversible and irreversible strain mechanisms.PLL also exhibits a monotonic behavior.The slope of the PLL is significantly higher than that of the ZLL.Those two lines have been reported in literature to serve as damage quantification indices [51].
The distinctive curvature during loading differs from the curvature during the unloading stage, a difference that becomes more obvious as the cycles progress (figure 8(a)).This deviation arises from some inherent hysteretic phenomena observed after each additional loading cycle.The hysteretic behavior manifests itself when plotting the resistance change versus the strain for a cycle, as shown in figure 9(b) where the last full cycle prior to failure is depicted.This phenomenon may be attributed to the inherent electrical behavior of the CNTs as the TEG successively stretches and retracts.Moreover, slight permanent damage could occur to the CNTs conductive path after each additional loading cycle.The smoother the hysteresis curve, the more linear the correlation between resistance change and strain [52].Some researchers [16] have suggested the presence of a memory effect since the deflection points corresponded to the maximum load experienced in the previous cycle, and collimated it with the Kaiser effect.Deflection points are defined as the points where the slope in the ΔR/R 0 curve changes.In this study, the deflection points do not coincide with the maximum load of the previous cycle but may as well indicate memory effect.In particular, the deflection points (figure 9(a)) appear at stress level lower than the maximum of the previous cycle, which implies damage accumulation.
The damage history and the stress memory of the material is validated by the AE results, depicted in figure 10.As seen in figure 10(b), upon reloading, AE activity is recorded before the maximum load of the previous cycle is reached.This phenomenon is called Felicity effect and is encountered in viscoelastic materials that show hysteresis on stress and strain, such as composites [53].The load data are also presented with another AE parameter called RA (figure 10(a)).RA is defined as the ratio of the rise time to the peak amplitude, where peak amplitude is the largest amplitude value in the waveform and rise time is the duration time from the first arrival of the waveform to the peak amplitude [49,54].The RA values are moderately varying between different cycles but are doubled at the moment of final failure.A shift to the higher values implies that shear failure mechanisms such as delamination become dominant [49].

Dynamic loading
In order to assess the sensing capabilities of the TEG and its ability to follow the gradual degradation of the material during fatigue loading, the relative resistance change is plotted as a function of the Young's modulus reduction.To determine the stiffness reduction due to the applied dynamic load the secant slope of the stressstrain hysteresis loop of each individual cycle is calculated and then normalized to the initial values (E/E 0 ) [44].As illustrated in figure 11(a) stiffness decreases rapidly in the initial stages of the fatigue life, as a consequence of the material degradation due to the formation and the subsequent propagation of transverse cracks.A sharp rise in the material's stiffness is noticed during the transition from static loading to dynamic loading [55].This shift  in the applied load type was purposely chosen in order to evaluate the ability of the sensor to track it.After that, stiffness loss is manifested in a slow and steady manner as the delamination propagates.This region constitutes the larger portion of the fatigue life span.Finally, the stiffness is reduced considerably and abruptly with fiber breakage at the final failure of the specimen [12,44,56].
Likewise static and quasi-static cyclic loading discussed above, the piezoresistive sensor is also able to monitor the applied stress and induced strain during the course of the fatigue life.As seen in figure 11(a), the resistance change mirrors the modulus reduction.Due to the multitude of data points, the original graph is approximated here with the application of a percentile filter.This smoothing approach best represents the original electrical response, which is shown as inset in the lower right corner of the figure.At the initial part of the curve, the resistance increases linearly, following the elastic deformation due to the applied quasi static loading during the preloading phase.Then, when the loading switches to dynamic, a sharp increase can be distinguished, as the resistance change curve exhibits a logarithmic growth up to 10% of the fatigue life.After that, the resistance enters a phase of a linear steady increase, represented in the diagram by a mean line.The slope in this stage is much smaller than the slope in the preloading area and completely conforms to the gradual stiffness loss due to damage accumulation discussed above.At the end of the curve an abrupt increase in the resistance occurs, denoting the failure of the specimen and the disruption of the conductive path.
Apart from the mean lines and the general trend of the electrical response, the sensor is also capable of precisely following the oscillation between the maximum and the minimum applied load at the level of single fatigue cycles.This is depicted in the magnified areas of the stress and ΔR/R 0 versus time curves.More precisely, in figure 11(b), where the electromechanical behavior during a part of the preloading and the first ten cycles is shown, the transition point from static to dynamic loading is clearly detected.The fitted mean line here is logarithmic, mirroring to some extend the stiffness loss pattern in this loading stage.In figures 11(c) and (d), illustrating respectively the middle and the last cycles, the resistance change matches exactly the alternating loading.The specimen breakage and the subsequent conductive path interruption is also clearly shown.Overall, the resistance change of the TEG can accurately track not only the deformation and the failure of the specimen, but also the damage accumulation.Moreover, the fact that the ΔR/R 0 completely correlates with the applied stress/strain during each cycle confirm its durability and long-term stability.
Finally, the recorded AE activity versus the modulus reduction is plotted in figure 12.The AE hits follow closely both the stiffness loss and the electrical response of the TEG.Hits accumulate linearly and rapidly during the quasi-static preloading stage.Subsequently, when dynamic loading is applied the AE hits are characterized by an even faster increase.The cumulative hits curve shows a logarithmic relation with time for the initial 10% of the fatigue life, having a similar pattern with its resistance change counterpart (figure 11(a)).Then, it follows a steadily increasing behavior for most of the fatigue life span, extending between 10% and 90%.However, two local increases at about 70% and 75% of the normalized fatigue life are also present.Those increases are related to damage incidents and are absent from the respective resistance change curve.This suggests that they concern internal damage that has no effect on either the TEG or the modulus of the material.Lastly, another significant rise is observed at about 90% of the normalized fatigue life, as the AE hits increase exponentially prior to ultimate failure.Similar results emerge from the RA value.The transition between loading types coincides with a strong increase in the RA and after this point its value is rather stable for the majority of the normalized fatigue life.RA rises again to higher values when the material approaches its final failure, and the RA exhibits a sharp increase by about an order of magnitude.This generally indicates a transition from tensile cracking mode to shear [49].The damage incident at about 75% of the normalized fatigue life is not only detected but also characterized by a high RA value.This implies a shear mode damage such as delamination.Delamination cannot be detected by the resistance change of the TEG, since the sensor tracks only strain and damage of the conductive path.Moreover, it may not lead to significant stiffness reduction.Thereby, the employment of the AE together with the ERCM is deemed essential to compensate for the limitations of the printed sensor.

Conclusions
The use of CNTs in the design of advanced structural composites is motivated by the multitude of functionalities that they can provide.Among others they are capable of thermal energy harvesting as well as detecting strain and damage evolution.This paper demonstrates the viability of exploiting a TEG integrated into a composite to achieve SHM, a perspective that has up to this point not been considered.
The TEG is fabricated via a versatile and scalable printing method, hence showing high potential for low-cost industrial production.Printing on the outer lamina of the composite is not affecting its mechanical performance.TEG demonstrates a significant power output of 3.31 μW upon being exposed to a temperature difference of 75 K.Subsequently, adopting an innovative strategy, based on the inherent electrical conductivity of the generator, the TEG is utilized as a strain/damage sensor.TEG-enabled specimens are subjected to a series of loading, namely static tensile, cyclic loading-unloading, and fatigue.Concurrently their electrical resistance is measured, employing the well-established ERC method.As no other external or embedded sensors are used, the procedure falls within the scope of self-sensing, with all the advantages that this entails.Nonetheless, their electromechanical response is also coupled with acoustic emission to strengthen the findings.
Results reveal that the TEG is a well-suited and noninvasive sensor capable of monitoring the structural integrity of the composite, as it allows for efficient, reliable and noise free sensing.In all loading cases, the electrical resistance follows directly the applied displacement from the onset of testing, enabling early damage detection and validating its durability.The TEG is characterized by a fairly linear strain response and displays high sensitivity.The gauge factor is approximately equal to 3, being in the same range as the conventional metallic strain gauges.More importantly it exhibits no Poisson effect or other artifacts introduced by spurious mechanisms, consequently increasing its dependability.Analysis of the results deriving from quasi-static and dynamic loading indicates that the irreversible variation in the electrical resistance can track the residual strain and by extension the accumulated damage.This is also suggested by the clear correlation that exists between the electrical resistance change and the stiffness loss which is often used as a damage accumulation parameter.Finally, the ultimate failure is clearly identified in every loading type.
Despite the overall described success there are also some limitations.The herein suggested multifunctional composite may only detect strain or damage that refers to outer ply where the TEG is integrated.It lacks however the ability of identifying internal damage of the host structure, such as delamination, unless they lead to reduction of the overall material's modulus.In such cases the parallel use of other NDT methods, i.e., the AE could prove useful.
In conclusion, this study reports on a robust technology that could hold high potential for real-life applications by combining energy harvesting with structural integrity diagnosing.This combination of functionalities can improve the safety and increase the efficiency of structures, thereby promoting sustainability.
In future studies we intend to take further advantage of these two joint functionalities to achieve self-powered sensing through the utilization of the generated output characteristics.

Figure 1 .
Figure 1.TEG architecture.(a) Top view, overall dimensions in mm, and dimensions of thermoelements.(b) Heated area and thermal diffusion.(c) Captured thermal image.

Figure 2 .
Figure 2. Schematic diagram of the TEG printing process, laminate manufacturing, specimen's preparation, and layout.

3. 1 .
Thermoelectric response of the TEG-enabled GFRP laminate In figure 4 the performance of the TEG is demonstrated in terms of output characteristics and energy harvesting potential as a function of ΔΤ.The plot presents the experimentally determined internal electrical resistance

Figure 3 .
Figure 3. General overview of the on-line monitoring including all parameters (load, displacement, electrical resistance, acoustic emission, temperature).

Figure 4 .
Figure 4. Output characteristics of the TEG (R TEG , V OC , I SC , power, power density) versus the applied ΔΤ.

Figure 5 .
Figure 5. Power output of TEG-enabled composite at ΔΤ 50, 75 and 100 K. (a) V-R load and P-R load curves.(b) V-I and P-I curves.

Figure 11 .
Figure 11.Electrical online monitoring during fatigue loading.(a) Normalized stiffness and ΔR/R 0 versus normalized fatigue life.(b) Stress and ΔR/R 0 versus time during preloading and the first cycles.(c) Stress and ΔR/R 0 versus time during the middle of the fatigue life (d) Stress and ΔR/R 0 versus time during the last cycles.

Figure 12 .
Figure 12.Acoustic emission during fatigue loading.Normalized stiffness, cumulative hits RA versus normalized fatigue life.

Table 1 .
TE performance of the TEG-enabled GFRP laminate at various temperature differences.