sensors Article

Artificial Neural Network-Based Early-Age Concrete Strength Monitoring Using Dynamic Response Signals Junkyeong Kim 1 , Chaggil Lee 1 and Seunghee Park 2, * 1 2

*

Department of Civil & Environmental System Engineering, Sungkyunkwan University 2066, Seobu-ro, Jangan-gu, Suwon-si, 16419 Gyonggi-do, Korea; [email protected] (J.K.); [email protected] (C.L.) School of Civil & Architectural Engineering, Sungkyunkwan University 2066, Seobu-ro, Jangan-gu, Suwon-si, 16419 Gyonggi-do, Korea Correspondence: [email protected]; Tel.: +82-31-290-7525

Academic Editors: M. H. Ferri Aliabadi and Zahra Sharif Khodaei Received: 15 March 2017; Accepted: 5 June 2017; Published: 7 June 2017

Abstract: Concrete is one of the most common materials used to construct a variety of civil infrastructures. However, since concrete might be susceptible to brittle fracture, it is essential to confirm the strength of concrete at the early-age stage of the curing process to prevent unexpected collapse. To address this issue, this study proposes a novel method to estimate the early-age strength of concrete, by integrating an artificial neural network algorithm with a dynamic response measurement of the concrete material. The dynamic response signals of the concrete, including both electromechanical impedances and guided ultrasonic waves, are obtained from an embedded piezoelectric sensor module. The cross-correlation coefficient of the electromechanical impedance signals and the amplitude of the guided ultrasonic wave signals are selected to quantify the variation in dynamic responses according to the strength of the concrete. Furthermore, an artificial neural network algorithm is used to verify a relationship between the variation in dynamic response signals and concrete strength. The results of an experimental study confirm that the proposed approach can be effectively applied to estimate the strength of concrete material from the early-age stage of the curing process. Keywords: early-age concrete strength estimation; artificial neural network; electromechanical impedance; harmonic wave; embedded piezoelectric sensor

1. Introduction Concrete is one of the most common materials used to construct civil infrastructures, such as buildings and bridges. However, it is also one of the most difficult materials to manage because concrete is a mixture, consisting of cement, water, sand, gravel, and other aggregates. Due to its heterogeneous property, it is difficult to mathematically predict the compressive strength of concrete during and/or after curing. In particular, predicting the compressive strength during the curing process is important to reduce the construction time and cost by determining the appropriate curing time to achieve sufficient strength. The in situ strength of concrete structures can be determined with high precision by performing a destructive strength test and material analysis on core samples extracted from the structure [1]. However, this method can lead to destruction of the integrity of the host concrete structure. To overcome this problem, a range of nondestructive testing (NDT) methods has been developed to monitor the strength, without damaging the host structure. These NDT methods can be categorized into two classes: thermal property-based NDT, and mechanical property-based NDT.

Sensors 2017, 17, 1319; doi:10.3390/s17061319

www.mdpi.com/journal/sensors

Sensors 2017, 17, 1319

2 of 12

Thermal property-based concrete strength NDT methods focus on the hydration heat occurring during the curing process. Concrete hardens due to the hydration process between water and cementitious materials. Because the hydration process involves an exothermic reaction, the concrete generates a certain amount of heat during the curing process. The strength of concrete can be estimated from the degree of hydration, and it can be calculated by measuring the hydration heat. Therefore, the strength of concrete can be estimated by measuring the thermal history of concrete through thermocouples, fiber optic sensors, or other thermal sensors [2,3]. The physical property-based concrete strength estimation methods are based on the change of mechanical properties. Ultrasonic-based methods are general NDT methods used for the early-age monitoring of concrete. The properties of ultrasonic wave propagation, such as velocity or attenuation, are affected by the change of physical properties [4–8]. Thus, the strength of concrete can be monitored by tracking the changes in ultrasonic wave propagation. Also, an electromechanical impedance method using piezoelectric sensors could use to estimate the strength of concrete. The strength of concrete can be estimated by measuring the resonant frequency of impedance [9,10], calculating the RMSD (root mean square deviation) of impedance signals [11], or impedance spectrum analysis [12]. Furthermore, a range of methods based on the acoustical, electrical, magnetic, optical, radiographic, and other mechanical properties of concrete have been studied [13]. Because the concrete is heterogeneous material, the strength of concrete is hard to estimate using a formal equation. To overcome this limitation, the neural network was used to estimate the strength of concrete. The strength of HPC (High Performance Concrete) can be modeled using a modified neural network architecture [14] and fuzzy-ARTMAP neural network by an analysis of the mix proportion [15]. The main benefits in using a neural network are that all of the behavior of a material can be represented within the unified environment of a neural network. Also, the neural network-based model is built directly from experimental data using the learning capabilities of the neural network [15]. In this context, this study proposes a novel method to estimate the strength of concrete at the early-age stage by integrating an artificial neural network algorithm with dynamic response signals of the concrete material. The dynamic response signals of the concrete, including both electromechanical impedances and guided ultrasonic waves, are obtained from an embedded piezoelectric sensor module. The cross-correlation coefficients of the electromechanical impedance signals and the amplitudes of the guided ultrasonic wave signals are selected to quantify the variations in dynamic response signals according to the strength of the concrete. Furthermore, an artificial neural network algorithm is used to verify a relationship between the variations in dynamic response signals and concrete strength. The results of an experimental study confirm that the proposed approach can be effectively applied to estimate the strength of concrete materials from the early-age stage of the curing process. 2. Early-Age Concrete Strength Estimation Technique 2.1. Embedded Piezoelectric Sensor Piezoelectric sensors can interconvert mechanical energy and electrical energy. Due to this piezoelectric effect, a piezoelectric sensor can be used simultaneously as both an actuator and a sensor. This study employs a lead zirconate titanate (PZT) patch to generate vibration and waves to the concrete structure, and measure the dynamic responses of the concrete [16,17]. Table 1 shows the size and major properties of the PZT used in this study. Table 1. PZT material properties. APC 850 WFB Series Size (mm)

Electromechanical Coupling Factor

Piezoelectric Charge Constant (10−12 m/V)

Diameter

Thickness

κ33

κ31

d33

d31

30.00

0.508

0.72

0.36

400

−175

Sensors2017, 2017,17, 17,1319 1319 Sensors

of12 12 33of

Sensors 2017, 17, 1319

3 of 12

To obtain dynamic responses from the inside of concrete structures, the sensors should be To obtain obtain dynamic responses from the the inside of concrete concrete structures, the sensors should be To dynamic responses from of sensors should be embedded within the concrete. However, theinside PZT can be easilystructures, broken bythe stresses, such as the embedded within the concrete. However, the PZT can be easily broken by stresses, such as the embedded within concrete. However, the PZT can be the easily broken stresses, as the thermal stress and the shrinkage stress of concrete. To protect PZT in the by concrete, ansuch embedded thermal stress and shrinkage stress of concrete. To protect the PZT in the concrete, an embedded thermal stress and shrinkage stress of concrete. To protect the PZT in the concrete, an embedded piezoelectric sensor has been developed [18]. The novel embedded piezoelectric sensor is fabricated piezoelectric sensor has been developed developed [18]. The The novel novel embedded piezoelectric sensor fabricated piezoelectric sensor been [18]. embedded piezoelectric sensor fabricated to improve the signalhas quality, using a hemi-spherical hollow Styrofoam case, as Figure 1aisisshows. The to improve the signal quality, using a hemi-spherical hollow Styrofoam case, as Figure 1a shows. to improve piezoelectric the signal quality, a hemi-spherical case, as Figurecondition, 1a shows. The embedded sensorusing allows one side of the hollow PZT to Styrofoam maintain free boundary even The embedded piezoelectric sensor allows one side of the PZT to maintain free boundary condition, embedded piezoelectric sensor allows one side of the PZT to maintain free boundary condition, even though the sensor is embedded within the concrete media. Thus, the embedded piezoelectric sensor evenmeasure though the sensor is embedded within theasconcrete media. Thus, piezoelectric though the sensor is embedded within the concrete Thus, the embedded piezoelectric sensor can the electromechanical impedance ifmedia. it were attached tothe theembedded concrete surface. The sensor can measure the electromechanical impedance as if it were attached to the concrete surface. can measure the electromechanical impedance as if it were attached to the concrete surface. The electromechanical impedance is measured using a single sensor with a self-sensing technique, and The electromechanical impedance is measured using a single sensor with a self-sensing technique, electromechanical is ismeasured single sensorone with a self-sensing technique, and the harmonic waveimpedance propagation measuredusing usingatwo sensors; is used to generate the harmonic andharmonic theand harmonic wave propagation is measured using two sensors; onetoisgenerate used sensor tothe generate the the wave propagation measured using two sensors; one is used harmonic waves, the other senses theis propagated waves. The embedded piezoelectric module harmonic and senses the other the propagated waves. The embedded piezoelectric sensor waves, and the other thesenses propagated waves. The embedded piezoelectric sensor module consists ofwaves, two embedded piezoelectric sensors to simultaneously measure the impedance and module consists of two embedded piezoelectric sensors to simultaneously measure the impedance consists ofwave two propagation. embedded piezoelectric sensors to simultaneously harmonic Figure 1b shows the size of the module. measure the impedance and and harmonic propagation. Figure 1b shows the of size the module. harmonic wavewave propagation. Figure 1b shows the size theofmodule.

(a) (a)

(b) (b)

Figure 1. Embedded piezoelectric sensor. (a) Schematic of the embedded piezoelectric sensor; (b) The Figure piezoelectric sensor. (a) Schematic of the embedded piezoelectric sensor; (b) The proposed embedded piezoelectric sensor module. Figure1.1.Embedded Embedded piezoelectric sensor. (a) Schematic of the embedded piezoelectric sensor; proposed embedded piezoelectric sensor module. (b) The proposed embedded piezoelectric sensor module.

2.2. Electromechanical Impedance Measurement for Concrete Strength Estimation 2.2. Impedance Measurement for Estimation 2.2.Electromechanical Electromechanical Impedance Measurement for Concrete Concrete Strength Strength The electromechanical impedance (EMI) method has beenEstimation developed for structural health The electromechanical impedance (EMI) method has been developed for health monitoring, damage detection, and NDT (EMI) [19,20].method If the PZT attached to the host structure and an The electromechanical impedance hasisbeen developed for structural structural health monitoring, damage detection, and NDT [19,20]. If the PZT is attached to the host structure and an alternating electric voltage is applied to the PZT, the elastic waves generated by the PZT are monitoring, damage detection, and NDT [19,20]. If the PZT is attached to the host structure and alternating electric voltage is applied to the PZT, the elastic waves generated by the PZT are transmitted to the host structure. responses the the waves represent mechanical impedance an alternating electric voltage is The applied to theon PZT, elastic wavesthe generated by the PZT are transmitted to the host structure. The responses on the waves represent the mechanical impedance of the host structure as shown in Figure 2. The structural impedance directly reflects the effective transmitted to the host structure. The responses on the waves represent the mechanical impedance of structure as Figure reflects the electrical impedance thein mechanical coupling effectimpedance between thedirectly PZT and host structure. The of the the host host structurethrough as shown shown in Figure 2. 2. The The structural structural impedance directly reflects the effective effective electrical impedance through the mechanical coupling effect between the PZT and host structure. The electromechanical impedance themechanical PZT, as coupled to the hostbetween structure, is given by [16] electrical impedance throughofthe coupling effect the PZT and host structure. electromechanical impedance of theofPZT, as coupled to the structure, is given by [16]  1 The electromechanical impedance the PZT, as coupled tohost the host structure, is given by [16]

k str ( )  1 1  Z ( )  1 1   312  str ((kωstr))( )  −1 i1 C  1  2312 k PZTkkstr Z (  )  Z (ω ) =  kkstr iC 1 − κ31 kkPZT ))  iωC PZT + str ((ω

(1) (1)(1)

where, Z () is the electromechanical impedance, C is the zero-load capacitance of the PZT, 31 where, ZZ((ω impedance, C is thethe zero-load capacitance of the PZT,PZT, κ isthe where, theelectromechanical electromechanical zero-load capacitance of the )) isisthe 31 is the electromechanical cross coupling impedance, coefficient ofCtheisPZT, kstr ( ) is the dynamic stiffness31of the electromechanical cross coupling coefficient of the PZT, k ( ω ) is the dynamic stiffness of the structure, str is the electromechanical cross coupling coefficient of the PZT, kstr ( ) is the dynamic stiffness of the structure, PZT is the and k PZT and is thekstiffness ofstiffness the PZT.of the PZT. structure, and kPZT is the stiffness of the PZT.

Figure Figure2.2.Electromechanical Electromechanicalcoupling couplingbetween betweenthe thelead leadzirconate zirconatetitanate titanate(PZT) (PZT)and andthe thehost hoststructure. structure. Figure 2. Electromechanical coupling between the lead zirconate titanate (PZT) and the host structure.

Sensors 2017, 17, 1319 Sensors2017, 2017,17, 17,1319 1319 Sensors

4 of 12 of12 12 44of

The dynamicstiffness stiffnessof ofconcrete concretechanges, changes,according accordingto to the development ofof strength during thethe The dynamic stiffness of concrete of strength during the The dynamic changes, according tothe thedevelopment development strength during curing process.Also, Also,the theEMI EMIsignals signalsmeasured measuredin inthe the concrete media should vary during thethe curing curing process. Also, the EMI signals vary during the curing curing process. measured in theconcrete concretemedia mediashould should vary during curing stage. Therefore, the strength of concrete can be estimated by tracking the variation of the EMI signals. stage. Therefore, the strength of concrete can be estimated by tracking the variation of the EMI signals. stage. Therefore, the strength of concrete can be estimated by tracking the variation of the EMI signals. In thisstudy, study,an anEMI EMImeasurement measurement system system based based on on self-sensing technique with single this study, an EMI measurement with aa single InIn this system based onaaaself-sensing self-sensingtechnique technique with a single embeddedpiezoelectric piezoelectricsensor sensorwas wasused. used.AAvoltage voltagedivider-based divider-basedself-sensing self-sensingcircuit circuitas asdescribed describedin in embedded embedded piezoelectric sensor was used. A voltage divider-based self-sensing circuit as described Figure 3 is suitable for use in cast-in-place concrete because it is inexpensive, and has sufficient because it it is is inexpensive, andand hashas sufficient inFigure Figure33isissuitable suitablefor foruse useinincast-in-place cast-in-placeconcrete concrete because inexpensive, sufficient accuracyto tomonitor monitorthe thedevelopment developmentof ofstrength, strength,even eventhough thoughthe theEMI EMIsignals signalsare areless lessaccurate accuratethan than accuracy accuracy to monitor the development of strength, even though the EMI signals are less accurate than whenusing usingother otherimpedance impedancemeasurement measurementmethods methods[21,22]. [21,22]. when when using other impedance measurement methods [21,22].

Figure of self-sensing basedimpedance impedancemeasurement. measurement. Figure3.3.3.Schematic Schematicdiagram diagramof ofself-sensing self-sensingbased based Figure Schematic diagram impedance measurement.

2.3. Harmonic Wave for Concrete StrengthEstimation Estimation 2.3. Harmonic WavePropagation PropagationMeasurement Measurementfor forConcrete ConcreteStrength Strength Estimation 2.3. Harmonic Wave Propagation Measurement The harmonicwave wavemeasurement measurementmethod method also used toestimate estimate thestrength strength ofconcrete. concrete. The The harmonic measurement method is also used to estimate the strength of concrete. The harmonic wave isisalso used to the of The embedded piezoelectric sensor module consists oftwo twoof embedded piezoelectric sensors,one oneof ofwhich which embedded piezoelectric sensor module consists of embedded piezoelectric sensors, The embedded piezoelectric sensor module consists two embedded piezoelectric sensors, one of introduces harmonic waveto tothe theconcrete, concrete, whilethe the other senses the propagated waves. waves. introduces aaharmonic wave while other senses propagated waves. which introduces a harmonic wave to the concrete, while the otherthe senses the propagated The harmonic wave propagation in the concrete media can be idealized as The harmonic wave propagation in the concrete media can be idealized as one-dimensional The harmonic wave propagation in the concrete media can be idealized asone-dimensional one-dimensional longitudinal wavepropagation, propagation,as asFigure Figure444shows. shows. longitudinalwave wave propagation, as Figure longitudinal shows.

Figure diagram harmonicwave wavemeasurement. measurement. Figure4. Schematicdiagram diagramof ofharmonic harmonic wave measurement. Figure 4.4.Schematic Schematic

The wave equationof ofaaaone-dimensional one-dimensionallongitudinal longitudinalwave wave is: The wave equation of one-dimensional The wave equation longitudinal waveis: is: 2 2 22 uu 11 22 uu 22 EE ∂ u2  12 ∂ u2 (c (c2 b  )) xx22 = ccc2b2b ∂t tt22 (cbb = E ρρρ) ∂x

(2) (2) (2)

b

where,uuisisthe thedisplacement displacementof ofan anelement, element,and andρρisisthe thedensity densityof ofmaterial. material.The Theaverage averagepower powerppof of where, where, u is the displacement of an element, and ρ is the density of material. The average power p of theharmonic harmonicresponse responseover overaaperiod periodcan canbe beexpressed expressedas: as: the the harmonic response over a period can be expressed as:

Sensors 2017, 17, 1319

5 of 12

EA2 ω2 p= 2cb

√

EρA2 ω2 1 ,A= 2 ω



4p2 Eρ

1



4

(3)

where, A is the harmonic amplitude, and ω is the angular frequency [23]. Concrete is a cementitious material that has the adhesive and cohesive properties to bond inert aggregates into a solid mass of adequate strength and durability. After casting, the concrete gradually stiffens, until the hydration is completed. The physical properties of concrete, especially the strength, are rapidly changed during curing process. Young’s modulus (E) is the factor that affects the most change, not only of the strength of concrete, but also of the amplitude of harmonic wave propagation. Hence, the strength of concrete can be estimated by measuring the amplitude of harmonic wave propagation through concrete [5]. 2.4. Artificial Neural Network for Concrete Strength Estimation Concrete is a composite material that consists of cement, water, and aggregates. Because of this heterogeneity, the strength of concrete cannot be clearly derived algebraically by mapping the signal variation and the actual strength. The neural network algorithm, a pattern recognition method, is utilized to deal with the complex properties of concrete [24,25]. The artificial neural network (ANN) comprises a number of processing elements that are connected to form layers of neurons, although the networks may be complex. The missing links between sets of inputs and outputs are found by determining the optimal synaptic weights, based on the available training data of the inputs and outputs [26]. A supervised multi-layer feed-forward ANN with backpropagation is typically employed. The input to the ANN are the variables, (x1 , x2 , . . . , x N ), which are weighted by whj,i and bias bhj ; and the output results, (y1 ,y1 , . . . ,yk ), feed the hidden layers. The output of the j-th hidden unit can be described as: ! N

yhj = F1 bhj + ∑ whj,i · xi

(4)

i =1

where, h refers to the quantities of the hidden layer, and F1 is a sigmoid nonlinear function. The output of the ANN is biased and weighted by the sum of the hidden layer outputs: yl = F2 bl0 +

H

∑ w0j,i · yhj

! (5)

j =1

where, o refers to the output unit, H refers to the number of units in the hidden layer, and F2 refers to a linear activation transfer function. The ANN is trained starting with a random set of weights and biases, and the output is calculated for every input data. Then, the error is calculated from the output layer and propagated backwards to modify the hidden layers, and has historically been called the backpropagated error. This learning algorithm is called backpropagation [27]. The training of the network involves a set of inputs for which the specified output is known. The training process is finished when the error is smaller than a desired value, or the process reaches a maximum specified number of iterations [28]. To build the ANN for concrete strength estimation, the network is trained using the set of inputs representing the condition of the concrete, and the output that is measured compressive strength for each set of conditions. The inputs are the dynamic response features extracted from impedance/wave measurements. After training, the strength of concrete can be estimated by inputting the measured data set into the trained ANN.

Sensors 2017, Sensors 2017, 17, 17, 1319 1319

of 12 12 66 of

3. Experimental Study 3. Experimental Study 3.1. Procedure 3.1. Experimental Experimental Setup Setup and and Test Test Procedure An strength estimation method. A An experimental experimental study studywas wasperformed performedtotoverify verifythe theproposed proposed strength estimation method. 50 × 50 × 25 cm concrete specimen was cast, and two piezoelectric sensor modules were embedded A 50 × 50 × 25 cm concrete specimen was cast, and two piezoelectric sensor modules were embedded during shows. Module obtain dynamic dynamic response response data data during the the casting casting process, process, as as Figure Figure 55 shows. Module 11 was was used used to to obtain sets sets for for training training the the ANN, ANN, and and module module 22 was was used used to to verify verify the the trained trained ANN. ANN. Table Table 22 shows shows the the mix mix proportion of the concrete. proportion of the concrete.

Figure 5. Geometric information of the test specimen. Figure 5. Geometric information of the test specimen. Table 2. Mix proportion proportion of Table 2. Mix of test test concrete. concrete. 3)3 ) Cement Water (kg/m Cement(kg/m (kg/m33)) Water (kg/m 155155 433 433

3 ) 3) Sand 3) 3) 3) (%) AE Silica Fume (kg/m (kg/m Gavel (kg/m3(kg/m ) Silica Fume (kg/m Sand (kg/m Gavel AE (%) 22.8 22.8 737 737 941 941 0.45 0.45

The The specimen specimen was was cured cured in in air air at at room room temperature. temperature. The The reference reference strength strength of of the the specimen specimen was was measured Machine measured at at the the early-age early-age of of the the curing curing stage stage from from aa destructive destructive test test using using aa Universal Universal Test Test Machine (UTM), with with Φ10 Φ10 ××20 (UTM), 20cm cmstandard standardcylinders cylinderscast castwith withthe thesame sameconcrete concrete as as that that of of the the test test specimen, specimen, at 16, 25, 38, 51, 75, and 99 h after casting. Table 3 shows the strength of concrete at the early-age of at 16, 25, 38, 51, 75, and 99 h after casting. Table 3 shows the strength of concrete at the early-age of the the curing stage from destructive testing. curing stage from destructive testing. Table 3. Reference strength of of the the test test specimen. specimen. Table 3. Reference strength

Curing Age (h) 16 25 16 25 Compressive Strength (MPa) 10.24 14.06 Curing Age (h)

Compressive Strength (MPa)

10.24

14.06

38 18.95

51 22.65

75 26.02

38

51

75

18.95

22.65

26.02

99 28.63 99

28.63

The EMI and harmonic wave signals were measured using the NI-PXI DAQ system (1042Q), and The EMI circuit and harmonic wave signals were measured using NI-PXI DAQ system (1042Q), and a self-sensing board was used for EMI measurement. Thethe EMI was measured in the frequency a self-sensing circuitkHz, board was EMI measurement. The EMI waswas measured the frequency range of 5 kHz~200 with a 2used MHzfor sampling rate. The harmonic wave actuatedinwith a 100 kHz range of 5 kHz~200 kHz, with a 2 MHz sampling rate. The harmonic wave was actuated frequency signal, and measured with a 5 MHz sampling rate. The dynamic response signalswith werea 100 kHz frequency signal, measured measured every hour afterand casting, up to with 100 h.a 5 MHz sampling rate. The dynamic response signals were measured every hour after casting, up to 100 h. 3.2. Result of EMI Measurement The measurement result of EMI is shown below. Figure 6 shows the results measured by module 1, and Figure 7 shows the result measured by module 2 at curing ages of 12, 48, 84, and 100 h. It is observed

Sensors 2017, 17, 1319

7 of 12

3.2. Result of EMI Measurement

Sensors 2017, 17, 1319

7 of 12

The2017, measurement result of EMI is shown below. Sensors 17, 1319

Figure 6 shows the results measured by module 7 of 12 1, that the7 EMI signals change during the curing stage. In particular, resonant and both Figure shows the result measured byearly-age module 2ofatthe curing ages of 12, 48, 84, andthe 100 h. It is that both the EMI signals change during the early-age of the curing stage. In particular, the resonant frequency (theboth frequency at the peakchange of the signal) increases curing age. Because observed that the EMI signals duringgradually the early-age of theduring curing the stage. In particular, the frequency (the frequency at the peak ofto the signal) gradually increases during thetime. curing age. Becauseage. the resonant frequency proportional the stiffness, it should increase with resonant frequency (the is frequency at the peak of the signal) gradually increases during the curing the resonant frequency is proportional to the stiffness, it should itincrease time.with time. Because the resonant frequency is proportional to the stiffness, shouldwith increase

Figure6. Electromechanical impedance impedance Figure (EMI)variation variationofofmodule module1.1. Figure 6.6.Electromechanical Electromechanical impedance(EMI)

Figure 7. EMI variation of module 2. Figure 7. EMI EMI variation variation of module 2.

The cross-correlation coefficient index (1-CC) was calculated to provide quantitative information The cross-correlation coefficient index (1-CC) was calculated to quantitative about variation. The 1-CC values were derived the following equation: Thesignal cross-correlation coefficient index (1-CC) wasusing calculated to provide provide quantitative information information

about derived equation: N about signal signal variation. variation. The The 1-CC 1-CC values values were were derived using using the the following following equation:

1  CC  1 

N (Re( Z 0 )  Re( Z 0 ))( Re( Z i )  Re( Z i )) 1  i 1N (Re( Z 0 )  Re( Z 0 ))( Re( Z i )  Re( Z i ))  (Re( Z0 ) − Re( Z 0 ))(Re( Zi ) − Re( Zi )) 1 N 1 1 i i1∑ Z 0 Zi =1

(6)

(6) (6) σZZ0 0σZZii at the baseline (the EMI data before where, Re(Z0) is the real part of the impedance function i) is the real part of the impedance of the i-th hour at each measured data, Re( Z ) embedment), Re(Z 0 real where, Re(Z Re(Z00)) is is the the real part part of of the impedance impedance function function at at the the baseline baseline (the (the EMI data before

11 −CC CC  =11− NN−1 1

and Re(Zi ) Re(Z are the each set,  Z and arehour standard deviation of each dataset the embedment), Re(Z isaverage thereal realof part ofdata theimpedance impedance ofthe i-th houratateach each measured data, Re(ZZ00)) ii))is the part of the of measured data, embedment), Re( Z 0 i-th 0 and Re ( Z ) are the average of each data set, σ and σ are standard deviation of each dataset and N Z0  andRe( N Zisi )the and aretotal thenumber averageofofdatasets. each data set,Z0 Z and are standard deviation of each dataset i Z0 0 is the total number of datasets. Figure 8 reveals that the 1-CC data of modules 1 and 2 show almost the same pattern, which has and N is the total number of datasets. a similar trend to the typical strength curve at the early-age of the curing stage. Figure 8 reveals that the 1-CC data of modules 1 and 2 show almost the same pattern, which has a similar trend to the typical strength curve at the early-age of the curing stage.

Sensors 2017, 17, 1319

8 of 12

Figure 8 reveals that the 1-CC data of modules 1 and 2 show almost the same pattern, which has a Sensors 2017, 17, to 1319 8 of 12 similar trend the typical strength curve at the early-age of the curing stage. Sensors 2017, 17, 1319 8 of 12

Figure 8. Cross correlation variations during the early-age of the curing stage. stage. Figure 8. Cross correlation variations during the early-age of the curing stage.

The equation equation for for estimating estimating the the strength strength was was derived derived using using aa linear method using using the the The linear regression regression method The equation for estimating the strength was derived using a linear regression method using the 1-CC of of the the impedance impedance signal. signal. The 1-CC The estimation estimation equation equation from from linear linear regression regression is is shown shown in in Equation Equation (7). (7). 1-CC of the impedance signal. The estimation equation from linear regression is shown in Equation (7). S ( MPa )  210.5  CC  30.24 (7) MPa)) = 210.5 × CC (7) SS((MPa  210.5 CC−30.24 30.24 (7) Figure 9 shows the result of the strength estimation using that equation during 100 h. The Figure 99 shows the result of the estimation using thatthat equation during 100 h. Figure shows thestrength strength estimation using 100The h. estimation result after the 20 hresult showsofalmost the same value as the referenceequation strength.during However the estimation result after 20 h shows almost the same value as the reference strength. However the The estimation result after 20 h shows almost the same value as the reference strength. However estimation results before 20 h are decreased very steeply, and the estimation strength before 5 h estimation results before 20 20 h are decreased very h the estimation results before h are decreased verysteeply, steeply,and andthe theestimation estimationstrength strength before before 55 h represents a minus value. represents a minus value. represents a minus value.

Figure 9. Strength estimation result using the regression model. Figure 9. Strength estimation result using the regression regression model. model.

3.3. Result of Harmonic Wave Measurement Measurement 3.3. Result of Harmonic Wave Measurement Figures 10 and 11 show the variations of harmonic wave signals measured by module 1 and 2 and 11 11 show show the the variations variations of of harmonic harmonic wave wave signals signals measured by module 1 and 2 Figures 10 and due to the curing process. As the curing age increases, the amplitude of the internal harmonic wave thethe curing ageage increases, the amplitude of theofinternal harmonic wave due to the the curing curingprocess. process.AsAs curing increases, the amplitude the internal harmonic decreases. decreases. wave decreases.

Sensors 2017, 17, 1319 Sensors 2017, 17, 1319 Sensors 2017, 17, 1319

9 of 12 9 of 12

9 of 12

Sensors 2017, 17, 1319

9 of 12

Figure 10. Harmonic wave signal variation of module 1. Figure 10. Harmonic wave signal variation of module 1.

Figure wave signal signalvariation variationofofmodule module Figure 10. 10. Harmonic Harmonic wave 1. 1.

Figure 11. wave signal variationof ofmodule module2.2. 2. Figure variation of module Figure11. 11.Harmonic Harmonicwave wave signal signal variation

ToTo quantify the signal variation, the amplitude of the the harmonic waveisis isextracted extracted as feature of quantify the signal variation, theamplitude amplitude of the harmonic wave a feature of of To quantify the signal variation, the of harmonic wave aa feature Figure 11. Harmonic wave signal variation of module 2.extractedasas wave propagation. Figure 1212shows shows the amplitude changesof ofthe thetwo twomodules. modules. wave propagation. Figure12 showsthe theamplitude amplitude changes changes of wave propagation. Figure the two modules.

To quantify the signal variation, the amplitude of the harmonic wave is extracted as a feature of wave propagation. Figure 12 shows the amplitude changes of the two modules.

Figure 12. Wave propagation amplitude change during the early-age of the curing stage. Figure Wave propagation Figure 12. 12. Wave propagation amplitude amplitude change change during during the the early-age early-age of of the the curing curingstage. stage.

Although the slope of decreasing amplitude is not clear and both data are not the same, the two Although the slope slope of decreasing decreasing amplitudedecreases is not not clear clear and both both data are not notof the same, the the two Although the of amplitude is and are the same, two sets of amplitude data show that the amplitude according to data the increase strength. This sets of amplitude data show that the amplitude decreases according to the increase of strength. This sets of amplitude data show that the amplitude decreases according to the increase of strength. This is caused by the relationshipamplitude between the waveduring amplitude, and theof Young’s modulus Figure 12. inverse Wave propagation change the early-age the curing stage. (E) of is is caused inverse relationship between thewave wavechange amplitude, and the Young’s modulus (E) of the propagation media. However the amplitudes only by 1%, and theYoung’s variation patterns (E) are of caused byby thethe inverse relationship between the amplitude, and the modulus

theAlthough propagation However the amplitudes only by 1%, the variation are themedia. slope of decreasing amplitude is notchange clear and bothand data are not the patterns same, the two

sets of amplitude data show that the amplitude decreases according to the increase of strength. This is caused by the inverse relationship between the wave amplitude, and the Young’s modulus (E) of the propagation media. However the amplitudes only change by 1%, and the variation patterns are

Sensors 2017, 17, 1319

10 of 12

Sensors 2017, 17, 1319

10 of 12

the propagation media. However the amplitudes only change by 1%, and the variation patterns are dissimilar between module 1 and 2. This is due to the fact that the path of wave was different because dissimilar between module 1 and 2. This is due to the fact that the path of wave was different because the placement of aggregates. Thus the amplitude of harmonic wave is not recommended to estimate the placement of aggregates. Thus the amplitude of harmonic wave is not recommended to estimate concrete strength estimation in this specimen. concrete strength estimation in this specimen.

3.4. ANN-Based Concrete Strength Strength Estimation Estimation 3.4. ANN-Based Concrete The features features of of the the dynamic dynamic response response signals signals have have aa certain certain pattern pattern related related to to the the increase increase in in The concrete strength. The ANN was utilized to recognize the pattern and estimate the strength of concrete strength. The ANN was utilized to recognize the pattern and estimate the strength of concrete, concrete, using thefrom features from theresponse dynamicsignals. response transfer function ANN was using the features the dynamic Thesignals. transferThe function of ANN wasof tan-sigmoid tan-sigmoid and the rule was Levenberg–Marquadt backpropagation. Also,had the function andfunction the learning rulelearning was Levenberg–Marquadt backpropagation. Also, the ANN ANN had 10 hidden layers and one output layer. 10 hidden layers and one output layer. The input input data data consisted consisted 100 100 sets sets of of the the 1-CC 1-CC of of the the impedance impedance signals signals and and the the output output data data was was The the reference reference strength. strength. The The unmeasured unmeasured reference reference strength strength was was calculated calculated by by the the interpolation interpolation using using the the measured reference strength and the curing time. The ANN was trained by data from module the measured reference strength and the curing time. The ANN was trained by data from module 11 and the the trained trained ANN ANN was was verified verified by by inputting inputting the the data data from from module module 2. 2. and Figure 13 shows the result of concrete strength estimation using the ANN model. model. The Figure 13 shows the result of concrete strength estimation using the ANN The maximum maximum error between the reference strength and estimation strength was 1.33 MPa. Also, the strength before error between the reference strength and estimation strength was 1.33 MPa. Also, the strength before 16 hhwas wasreliably reliablyestimated. estimated. Because neural network could connect the relationship between 16 Because thethe neural network could connect the relationship between signal signal variation and strength with higher order than linear regression (and also the network trained variation and strength with higher order than linear regression (and also the network trained using all using all of data from module 1), the estimation result from ANN is robust against error than linear of data from module 1), the estimation result from ANN is robust against error than linear regression regression results. According to the the proposed ANN model could estimate the strength of results. According to the result, theresult, proposed ANN model could estimate the strength of the test the test specimen with negligible error. specimen with negligible error.

Figure 13. Strength estimation result using the artificial neural network (ANN). (ANN).

4. Conclusions

trained by by dynamic response signals is proposed to estimate earlyIn this this paper, paper,an anANN ANNmodel model trained dynamic response signals is proposed to estimate age strength monitoring of concrete. The dynamic response signals, thesignals, electromechanical impedance early-age strength monitoring of concrete. The dynamic response the electromechanical and ultrasonic are changed the strength the hostofconcrete. EMI impedance andharmonic ultrasonicwaves harmonic waves arebychanged by thevariation strengthof variation the host The concrete. and harmonic waves were measured the during curing process using two embedded The EMI and harmonic waves were during measured the curing process using twopiezoelectric embedded sensor modules, andmodules, the reference strength was measured a destructive test on standard piezoelectric sensor and the reference strength was using measured using a destructive test on cylinders.cylinders. From a series studies, it was confirmed that the that dynamic response signals standard Fromofa experimental series of experimental studies, it was confirmed the dynamic response obtained from the concrete during during the early-age of the curing processprocess has a pattern according to the signals obtained from the concrete the early-age of the curing has a pattern according strength. According to theto strength gain, gain, the resonant frequency of EMI increased and to the strength. According the strength the resonant frequency of was EMI gradually was gradually increased the amplitude of harmonic wavewave was was decreased. The variation in resonant frequency of the and the amplitude of harmonic decreased. The variation in resonant frequency of two the sensor modules hadhad a similar tendency while thethe variation two sensor modules a similar tendency while variationininamplitude amplitudeof of the the harmonic harmonic wave path difference. difference. Thus, the response of the two sensor modules had a dissimilar result caused by the path EMI variation was selected to estimate the strength of concrete. The cross-correlation coefficient was used to quantify the variation in EMI according to the strength gain. Furthermore, an artificial neural

Sensors 2017, 17, 1319

11 of 12

EMI variation was selected to estimate the strength of concrete. The cross-correlation coefficient was used to quantify the variation in EMI according to the strength gain. Furthermore, an artificial neural network algorithm was used to define the relationship between the variation in EMI and the strength of concrete. The ANN model was trained by the variation in 1-CC of EMI measured by sensor module 1 and verified by the data of sensor module 2. The results conclusively confirmed that the proposed approach could be effectively applied to estimate the strength of concrete material from the early-age stage of the curing process. Acknowledgments: This paper was supported by 63Research Fund, Sungkyunkwan University, 2016. Author Contributions: Junkyeong Kim and Changgil Lee conceived and designed the experiments; Junkyeong Kim performed the experiments; Junkyeong Kim and Changgil Lee analyzed the data; Junkyeong Kim and Seunghee Park wrote the paper. In addition, Junkyeong Kim and Seunghee Park are responsible for the implementation the proposed scheme. Conflicts of Interest: The authors declare no conflict of interest.

References 1. 2. 3. 4.

5. 6. 7. 8.

9.

10.

11.

12. 13. 14. 15.

Irie, H.; Yoshida, Y.; Sakurada, Y.; Ito, T. Non-destructive-testing methods for concrete structures. NTT Tech. Rev. 2008, 6, 1–8. Jonsson, J.A.; Olek, J. Effect of temperature-match-curing on freeze-thaw and scaling resistance of high-strength concrete. Cem. Concr. Aggreg. 2004, 26, 21–25. [CrossRef] Roller, J.J.; Russell, H.G.; Bruce, R.N.; Hassett, B. Effect of curing temperatures on high strength concrete bridge girders. PCI J. 2003, 48, 72–79. [CrossRef] Voigt, T.; Grosse, C.U.; Sun, Z.; Shah, S.P.; Reinhardt, H.W. Comparison of ultrasonic wave transmission and reflection measurements with P- and S-waves on early age mortar and concrete. Mater. Struct. 2005, 38, 729–738. [CrossRef] Gu, H.; Song, G.; Dhonde, H.; Mo, Y.L.; Yan, S. Concrete early-age strength monitoring using embedded piezoelectric transducers. Smart Mater. Struct. 2006, 15, 1837–1845. [CrossRef] Sun, Z.; Voigt, T.; Shah, S.P. Temperature effects on strength evaluation of cement-based materials with ultrasonic wave reflection technique. ACI Mater. J. 2005, 102, 272–278. Oh, T.; Kim, J.; Zhang, A.; Lee, C.; Park, S. Concrete strength evaluation in an early-age curing process using SVM with ultrasonic harmonic waves. Insight 2016, 58, 609–616. [CrossRef] Kim, J.W.; Kim, J.; Park, S.; Oh, T.K. Integrating embedded piezoelectric sensors with continuous wavelet transforms for real-time concrete curing strength monitoring. Struct. Infrastruct. Eng. 2015, 11, 897–903. [CrossRef] Shin, S.W.; Oh, T.K. Application of electro-mechanical impedance sensing technique for online monitoring of strength development in concrete using smart PZT patches. Constr. Build. Mater. 2009, 23, 1185–1188. [CrossRef] Kim, J.W.; Lee, C.; Park, S.; Koh, K.T. Real-time Strength Development Monitoring for Concrete Structures using Wired and Wireless Electro-mechanical Impedance Techniques. KSCE J. Civ. Eng. 2013, 17, 1432–1436. [CrossRef] Providakis, C.P.; Liarakos, E.V.; Kampianakis, E. Nondestructive wireless monitoring of early-age concrete strength gain using an innovative electromechanical impedance sensing system. Smart Mater. Res. 2013, 2013. [CrossRef] Qin, L.; Li, Z. Monitoring of cement hydration using embedded piezoelectric transducers. Smart Mater. Struct. 2008, 17, 055005. [CrossRef] ACI Committee 228. In-Place Methods to Estimate Concrete Strength Report. ACI 228.1R-03; American Concrete institute: Farmington Hills, CT, USA, 2003; pp. 1–44. Kasperkiewicz, J.; Racz, J.; Dubrawsk, A. HPC strength prediction using artificial neural network. J. Comput. Civil. Eng. 1995, 9, 279–284. [CrossRef] Yeh, I.-C. Modeling of strength of high-performance concrete using artificial neural networks. Cem. Concr. Res. 1998, 28, 1797–1808. [CrossRef]

Sensors 2017, 17, 1319

16. 17. 18. 19.

20. 21. 22. 23. 24. 25. 26.

27. 28.

12 of 12

Park, S.; Inman, D.J.; Lee, J.J.; Yun, C.B. Piezoelectric sensor-based health monitoring of railroad track using a two-step support vector machine classifier. J. Infrastruct. Syst. 2008, 14, 80–88. [CrossRef] Bhalla, S.; Soh, C.K. Structural health monitoring by piezo-impedance transducers, I: Modeling. J. Aerosp. Eng. 2004, 17, 154–165. [CrossRef] Song, G.; Gu, H.; Mo, Y.L. Smart aggregates: Multi-functional sensors for concrete structures-a tutorial and a review. Smart Mater. Struct. 2008, 17, 033001. [CrossRef] Giurgiutiu, V.; Rogers, C.A. Modeling of the electro-mechanical (E/M) impedance response of a damaged composite beam. In Proceedings of the ASME Winter Annual Meeting, ASME Aerospace and Materials Divisions, Adaptive Structures and Material Systems, Nashville, TN, USA, 14–19 November 1999; ASME Aerospace Division: New York, NY, USA; Volume 59, pp. 39–46. Liang, C.; Sun, F.P.; Rogers, C.A. Electro-mechanical impedance modeling of active material systems. Smart Mater. Struct. 1996, 5, 171–186. [CrossRef] Lee, S.J.; Sohn, H. Active self-sensing scheme development for structural health monitoring. Smart Mater. Struct. 2006, 15, 1734–1746. [CrossRef] Lee, S.J.; Sohn, H.; Hong, J.W. Time reversal based piezoelectric transducer self-diagnosis under varying temperature. J. Nondestruc. Eval. 2010, 29, 75–91. [CrossRef] Achenbach, J.D. Wave Propagation in Elastic Solids, 1st ed.; Elsvier: Amsterdam, the Netherland, 1973; pp. 103–152. Akkurt, S.; Ozdemir, S.; Tayfur, G.; Akyol, B. The use of GA-ANNs in the modelling of compressive strength of cement mortar. Cem. Concr. Res. 2003, 33, 973–979. [CrossRef] Kim, D.K.; Jong, J.L.; Lee, J.H.; Seong, K.C. Application of probabilistic neural networks for prediction of concrete strength. J. Mater. Civil. Eng. 2005, 17, 353–362. [CrossRef] Worden, K.; Tomlinson, G.R. Classifying Linear and Non-linear Systems using Neural Networks. In Proceedings of the 17th International Seminar on Modal Analysis, Leuven, Belgium, 23–25 September 1992; Katholieke Universiteit Leuven: Leuven, Belgium; pp. 903–922. Lippmann, R.P. An Introduction to Computing with Neural Nets. IEEE Acoust. 1987, 4, 4–22. [CrossRef] Lopes, V.J.; Park, G.; Cudney, H.H.; Inman, D.J. Impedance-based structural health monitoring with artificial neural networks. J. Intel. Mater. Syst. Struct. 2000, 11, 206–214. © 2017 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).