Connecting concrete technology and machine learning: proposal for application of ANNs and CNT/concrete composites in structural health monitoring

Carbon nanotube/concrete composite possesses piezoresistivity i.e. self-sensing capability of concrete structures even in large scale. By incorporating smart materials in the structural health monitoring systems the issue of incompatibility between monitored structure and the sensor is surpassed since the concrete element fulfills both functions. Machine learning is an attractive tool to reduce model complexity, so artificial neural networks have been successfully used for a variety of applications including structural analysis and materials science. The idea of using smart materials can become more attractive by building a neural network able to predict properties of the specific nanomodified concrete, making it more cost-friendly and open for unexperienced engineers. This paper reviews previous research work which is exploring the properties of CNTs and their influence on concrete, and the use of artificial neural networks in concrete technology and structural health monitoring. Mix design of CNT/concrete composite materials combined with the application of precisely trained artificial neural networks represents a new direction in the evolution of structural health monitoring of concrete structures.


Introduction
Both nanotechnology and neural networks were in the making during the rst half of the twentieth century. The rst idea of nanotechnology from R. Feynman (Caltech, 1959) and the rst stage of development of articial intelligence by B. Widrow and M. Hoff (Stanford, 1959) marked the beginning of a new era of technology. The conference held in Trieste, Italy in 1984 (ref. 1) opened a discussion on the topic of transforming brain functions onto a mathematical level. Johannesma introduced concepts for the analysis of neural responses to external stimuli, which would be used for an objective denition of receptive elds. However, the development was gradual until Ng and Dean created a network that learned to recognize a high-level concept ("cat or no cat") in 2012. Extensive work followed, making machine learning technology approachable and usable in many different areas. Articial neural networks (ANNs) were primarily developed as a representation of a human brain, however over time it became a tool for solving specic tasks. On the other hand, nanotechnology has remained theoretical and was facing a halt until the discovery of fullerenes in 1985 and carbon nanotubes (CNTs) in 1991 by Sumio Iijima. The initial tests showed that mechanical properties of carbon nanomaterials are superior to the materials which were used at that time. This was the main reason for various investigations of their properties since, and for many elds of engineering and medicine to establish a way to apply carbon nanomaterials.
As for civil engineering, it wasn't until the brink of 21 st century that carbon nanomaterials found their use. Even though concrete is a conventional material, it is the most used material in the world aer water as humanity's rst and basic need. Suffice to say that abundant funds are invested every year in concrete structures and their wellbeing. Every civil structure is designed to meet the criteria of safety, durability, serviceability, and sustainability during its service life so constant supervision over its condition is necessary. For exactly that purpose structural health monitoring (SHM) was developed as a eld which deals exclusively with optimizing the structure's behavior during its exploitation period.
SHM has now grown into a highly developed eld in terms of new and emerging technologies. The most recent developments are connected to different types of nanosensors 2,3 such as transducers, 4 electrochemical and optical 5 sensors as well as smart materials. While the development of nanoscale devices and nanoelectromechanical systems utilizing the unique properties of carbon nanotubes (CNTs) is an evolving area of nanotechnology, there is also considerable interest in making macroscopic engineered, smart, materials that can exploit novel material properties. 6 The existing problem with monitoring devices has been the compatibility between the sensor and the concrete structure. Smart materials with the addition of carbon nanomaterials are being developed in order to surpass this issue of incompatibility by making the whole structure able to give feedback about its condition. Cement-based materials with the addition of CNTs manifest piezoresistive behavior, meaning that it is possible to monitor strain by following the change in the electrical resistance 7 under certain type of physical stimuli affecting the structure. This way, structural elements not only fulll a structural function 8 but also behave as sensors apt for self-monitoring, while the data acquired and collected is regarded as usable.
Machine learning has been used in the eld of concrete technology in the past twenty years. Various neural networks were made 9-26 in order to predict primarily the compressive strength of concrete with different additions. Nonetheless, ANNs may be used also for predicting electrical conductivity of smart concrete, similar to the model used for conductive nanomodied polymer materials. 132,144 The idea of using smart materials can become more attractive by building a neural network able to predict properties of the specic nanomodied concrete, making it more cost-friendly and open for unexperienced engineers.
This paper is organized as follows. Section 2 and 3 give an overview of the properties of carbon nanotubes and further focus on the electrical properties of CNT/concrete composite materials. Section 4 establishes the role of articial neural networks in engineering and their application. Section 5 gives a review on the structural health monitoring and state-of-the-art in this eld and proposes a new method of designing smart materials capable of self-sensing.

Physical properties of CNTs
Since Iijima et al. 27 described "single-shell tubules" three decades ago, there has been extensive research on this topic and many discoveries have been made regarding the properties and possible applications of single-walled carbon nanotubes. Carbon nanotubes are allotropes of carbon with a cylindrical structure. As a part of the fullerene family, nanotubes are members which have a long hollow structure made of one or more sheets one carbon atom thick. The sheet rolling angles are chiral and the radius of the tube determines the properties of the carbon nanotube. 28 [38][39][40] and thermal conductivity (Table 1). These properties offer CNTs great potential for wide applications in eld emission, conducting plastics, thermal conductors, energy storage, conductive adhesives, thermal interface materials, structural materials, 40 bers, etc. 34 The electrical conductivity of an individual CNT is quite high, however, the electrical properties and performance of CNTs depend on several geometric and chemical factors describing a single tube, as well as subsequent treatment to which it is exposed. Some analytical approaches for determination of the electrical conductivity are based on the continuum percolation theory of 3D cylinders, 42 where CNTs were modeled with nite length and diameter. 43 The nanotube electronic property is a strong function of its atomic structure, mechanical deformation, and chemical doping. 34,44 Nanotubes can be electrically conductive or semiconductive, 4-6,34,45-47 depending on the chirality angle ( Fig. 1) and uniformity aer production. 48 Chirality determines the conguration of stacked atoms in the carbon sheet, meaning that if the CNT is chiral i.e. cannot match its mirror image, conductivity needs to be additionally determined, especially for nanotubes with relatively small radii. 49 As Wang 50 concluded aer examining orthogonal motion transmission between aligned CNTs, chirality directly affects motion-transmission factors denoting the effect on mechanical power transmission in nanomaterials with extreme surface-to-volume ratio. Joshi et al. 51 examined the inuence of chirality on the dynamic behavior of single-walled CNTs. The simulation was made for cantilever nanotube considering it as a space frame structure similar to 3D beams with point masses. The results indicated that the existence of defects, and not chirality, affects the bending rigidity of the SWCNT, which gives valuable information about the nature of the inuence of chirality to the behavior of nanotubes.

Actuation behavior
Actuation behavior of nanotubes was rst reported by Baughman et al., 52 showing that the mechanical deformation of  nanotubes appears as a result of the injection of electrical charge. The demonstration of electrochemical actuation of CNTs was made using a single-walled buckypaper working as an electrode of an electrochemical cell. It was concluded that the classical electrostatics is also applicable at the nanoscale, 52,53 opening a new topic and triggering researchers to further examine this occurrence. Later, Li et al. 49,53,120 wrote on the behavior of CNTs exposed to an electrical charge. The authors found that the electrical failure of nanotube depends on the level of charge as well as its length. Many types of research have been conducted on the topic of nanotube characteristics and behavior, giving guidelines for further examinations. What is undoubted is that CNTs represent building blocks 54 for any kind of materials or devices where conductance is paramount. It is indeed this combination of mechanical and electrical properties of individual nanotubes that makes them the ideal reinforcing agents in a number of applications. 4

Inuence of conductive laments on concrete
Concrete is a nanostructured, multi-phase, composite material that ages over time. It is composed of an amorphous phase, crystals, and bound water. The amorphous phase, calciumsilicate-hydrate (C-S-H) is the 'glue' that holds the concrete together and is itself a nanomaterial. [55][56][57][58] The properties of concrete exist in multiple length scales (from nano to micro to macro) where the properties of each scale derive from those of the next smaller scale i.e. processes occurring at the nanoscale ultimately affect the performance of the bulk material. 55,[59][60][61][62][63][64][65][66][67] Viewed from the bottom up, concrete is a composite of molecular assemblages, surfaces, and chemical bonds that interact through chemical reactions, intermolecular forces and diffusion. The combination of exceptional mechanical properties of CNTs along with their low density and one-dimensional structure makes them an exceptional candidate for reinforcement in composite materials. 35,55,59 Incorporating some functional materials ( Fig. 2) such as carbon nanotubes or carbon bers into conventional concrete gives it the ability to sense strain, stress, cracks or another type of damage while maintaining or improving its mechanical properties and durability. 40,56,57,[68][69][70][71][72] When added to concrete, CNTs work as llers in the mesoporous environment of concrete, improving the hydration of cement and decreasing porosity, thus improving the compressive strength and decreasing autogenous shrinkage of concrete. Nanoscale reinforcement can also inhibit crack growth in the initial stages of damage and so prevent damage propagation, enhance the quality of the paste-aggregate interface and increase the amount of high stiffness C-S-H. 35,36,59,[72][73][74] Li et al. 75 investigated the mechanical properties of CNT/cement composites and found that compressive and exural strengths of cement with 0.5 wt% CNT addition were increased by 19% and 25%, respectively, compared to the plain unreinforced cement. 48 Fu and Chung 76 have researched the strain sensitivity of carbon ber-reinforced cement compared to that of normal cement since 1993. Furthermore, Thostenson and Chou 49,77 proposed that the change in the size of reinforcement, from micron-sized ber reinforcement to carbon nanotubes with nanometer-level diameters, enables the opportunity for multifunctional sensing capability. By in situ combining reinforcements of different scales, carbon nanotubes can penetrate the matrix-rich areas between bers in individual bundles as well as between adjacent plies and can achieve a nerve-like network of sensors throughout the arrays of bers in a composite. More studies evaluated the inuence of hybrid reinforcement, such as a combination of carbon ber (CF) and carbon black with CNTs on the self-sensing properties of cement-based sensors. 25,40,72,[78][79][80][81] Anthony et al. 80 developed computational models of a composite for predicting the benets of graing CNTs at the CF surface in order to create a conductive network that can be used for in situ damage detection. Moreover, Han et al. 81 showed that the exural strength and electrical conductivity in the cement mortar matrix with self-assembled CNT/nano carbon black composite llers are enhanced and stable, and sensitive piezoresistivity is achieved at a low ller content. Since the llers are mainly in order of magnitude of microscale or nanoscale, the potential ller-matrix and ller-ller interface surface areas of the composite are enormous. Since these interfaces have an effect on the quality of the electrical communication between CNTs and concrete matrix and also on communication between adjacent CNTs, they are affecting the conductive network and the overall electrical conductivity. 56

Dispersion of CNTs
The mechanical properties and electrical conductivity of concrete/CNT composite depend on the dispersion state of the nanollers. The nature of the dispersion problem for CNTs is somewhat particular because of their small diameter in the nanometer scale with a high aspect ratio (>1000) 4,48,72,82-84 resulting in an extremely large surface area. The aspect ratio must be sufficiently large to maximize the load transfer between the CNTs and matrix material and thus to achieve enhanced mechanical properties. 4, 35,85 Proper dispersion of the nanoparticles in concrete or cement mortar matrix is vital for improvements in the mechanical performance of the composite but it is difficult to achieve for nanomaterials with a large aspect ratio such as CNTs due to their high surface energy and strong interparticle (van der Waals) forces. If not properly dispersed, the addition of nanomaterials to cementitious materials may nally result in a decline in the mechanical properties. 35  Namely, the one-dimensional llers are more difficult to disperse and easier to damage during the fabrication of concrete. As the surface area of the ller increases, the attraction forces between the llers also increase. 56,87 Especially for brous llers, high aspect ratios combined with high exibilities increase the possibility of ber entanglement and close packing. 81 Studies have emphasized repeatedly 74,88-90 that CNTs tend to agglomerate, which hinders their uniform distribution within a matrix. 91 Poor dispersion of nanoparticles may lead to segregation, weak zones or potential areas for aggregation of concentrated stresses. 36,86 As an alternative, Anthony et al. 80 showed how CNTs can be directly graed on carbon bers in order to form a hierarchical reinforcement, combining nanoscale and microscale reinforcements 92,93 and avoid such dispersing problems. Uniform dispersion can be achieved using various types of mechanical and chemical methods including ultrasonication, shear mixing, calendaring, ball milling, stirring and extrusion. Various methods of chemical modication have been proved quite successful in contributing to better nanotube dispersion. In terms of tensile modulus, it has been established by numerous studies 94,95 that chemically modied nanotubes exhibit a signicant increase as compared to the matrix resin. This is mainly due to the fact that functionalization improves both dispersion and stress transfer. 82 Some studies have proved that the carboxyl CNT 41 is much easier to disperse in water and concrete matrix than the plain CNT because of the carboxyl groups situated on the surface of the carbon sheet. Moreover, the carboxyl CNT achieves a better bond with the concrete matrix than the plain CNT. 41 Surely that the selection of a proper method or a combination of several methods and their processing conditions need to be based on the desired properties of the end product. 35 The sensing ability of CNT/concrete composite Concrete has a volume conductivity lower than 10 À10 S cm À1 (ref. 35) which makes it an electrically nonconductive material. 40 However, the addition of CNTs provides it with new properties 2,56,96 making it a multifunctional structural material, primarily because the CNT/cement composites display piezoelectric behavior. 36,43,[97][98][99] Since rst-ever CNT/polymer composites were reported in 1994 by Ajayan et al., 100 many authors 7,82,101-104 showed how the conductive carbon nanollers such as CNTs can form conductive networks in polymeric materials at extremely low weight fractions while simultaneously improving the fracture toughness. Namely, conductive nanollers were found to impart signicant conductive and piezoresistive behavior to polymers making such material systems feasible for electrical applications, such as strain sensors. 8,79,105 The same principle can be applied to concrete which similarly represents an insulating matrix. 57,74 By measuring the changes in electrical resistance of the CNT/ concrete composite material under certain strain, its structural health can be monitored in terms of stress, strain, and damage. 41 Saa 97 presented CNT networks that were embedded into a cement matrix with the idea to develop an in situ wireless and an embedded sensor for damage detection in concrete structures. The results showed that the wireless response of the sensors changed due to the formation of cracks during loading and that the CNT/concrete sensors were able to detect the initiation of damage even at an early stage of loading. Sett 106 rst fabricated beams and rings using the polymer concrete with the CF ller. He tested the relationship between the change in electrical resistance and loading as well as the relationship between resistance change and deection of the beams under four-point bending. 41 Percolation theory. Concrete cannot be used for sensor purposes in its basic form, but if a metallic material is dispersed into it a conductive composite can be manufactured. Moreover, a non-conductive or poorly conductive matrix has greater sensitivity to changes in electrical resistance 40 which makes this combination theoretically perfect. At low volume fractions, 49 the conductivity of the composite remains very close to the conductivity of the pure insulating matrix. However, when a certain volume fraction is reached, the overall electrical conductivity experiences a drastic increase by many orders of magnitude 31,72 due to the formation of continuous electron paths or conducting networks. This phenomenon is known as percolation and the critical concentration of the ller as percolation threshold. 41,57,107 Numerous experimental and theoretical studies agree that the electrical conductivity of CNT reinforced cement-based composites is governed by a percolation process. 107 It is well known that the percolation is associated with the ability of electron transfer along CNTs as well as between adjacent nanotubes at their junctions. 43 Hence, the percolation theory is applied to explain the electrically conducting behavior of composites consisting of conducting ller and insulating matrices. 105,[108][109][110][111][112] Below the percolation transition range, there are no electron paths, the electrical properties are dominated by the matrix material 35,56 and there is no conductivity achieved. 8,107 Small size and large aspect ratio of CNTs grants them an exceptionally low percolation threshold meaning that the volume fraction for achieving conductivity is satisfactorily low. 72,113-115 Pan et al. 116 modelled the CNTs within percolation networks as ellipsoidal inclusions uniformly distributed in 3D space 'assuming contacts at the CNT junctions with zero resistance', and other calculations of the electrical conductivity were conducted using a similar micromechanics model. 117 It can be concluded that the level of distribution within the concrete/CNT composite is threefold: (i) the distribution of functional ller in a binder, (ii) distribution of the binder with llers among ne aggregates, and (iii) distribution of the ne aggregates with binder and llers among coarse aggregates. 41,118 As previously stated, the potential llermatrix and ller-ller interfaces affect electrical bonds between llers as well as between the concrete matrix and the ller material, 119 thereby inuencing the comprehensive conductive network and electrical conductivity of the composite. 41,56 Piezoresistive behavior. Carbon nanotubes express interesting electromechanical properties. When subjected to stress, their electrical properties change indicating a linear and reversible piezoresistive response with respect to different stress levels. It was proven that electrical conductivity strongly depends on the morphology of the CNT network 56,92 i.e. the number of contact points and that the electrical resistance of self-sensing CNT concrete can be changed when it is deformed under applied loading. Han et al. 41 showed that the changes in resistance coming from the alteration of nanotubes' geometry under external loading may be neglected since they are small and reversible. However, the contact resistance must be taken into consideration since the distance between adjacent nanotubes may considerably differ when under load, implying the sensing ability is stable and reversible as long as the deformation is elastic. 40,41 Piezoresistivity of composites with conductive llers has been the subject of several numerical and analytical works, 105,125,126 and simulations of the processes taking place in these materials are taking a lot of interest. Most recently, the authors 107 presented a 3D generalization of the micromechanics model of the overall conductivity and uniaxial piezoresistivity of CNT-reinforced composites subjected to arbitrary strain states. Changes in conductance, as the origin of the piezoresistive behavior, expected when the CNT/concrete composite is subjected to external mechanical strains 105,107,[127][128][129] are here prescribed to the changes in the volume fraction, to ller reorientation, and to changes in the inter-particle properties. The authors 107 proposed an approach shown to be capable of determining the piezoresistivity coefficients by two virtual tests, namely laterally constrained dilation and distortion, which is efficient in terms of computational cost and, more importantly, it reliably presents the complex electromechanical behavior of CNT/concrete composites. Piezoresistive behavior of CNT/ concrete composites is continuously investigated because it is the base point on which the majority of sensoring systems used in SHM today were built on.

Development of articial neural networks
Based on Hebbian learning, as it is conceived on principles of the biological systems' dynamics, articial neural networks (ANNs) have been developed as an interconnected group of nodes inspired by a simplied representation of neurons in the brain. According to Braitenberg (1977), changing the stimuli conditions may induce clusters to form and later "evaporate" in order to build new congurations, more or less mimicking the way the corresponding neurons are involved in a dynamic succession of different assemblies, "ignited" by the changes in the external world. 1 Although the original idea of ANNs was to solve problems in the same way as a human brain would, over time it deviated from biology and advanced towards the direction of solving specic tasks (pattern recognition, identication, and classication, 130 vision and control systems, cancer predictions, 131 etc.). Recently, machine learning has emerged as an attractive tool to reduce model complexity. 132 As a subset of machine learning, ANNs have been successfully used for a variety of applications, 133 some including structural analysis 134 and material science, ranging from studies of atomic properties 135 to the mechanical properties of individual CNTs. 136

Architecture of an ANN
The neural network approach is a way of modeling data based on computer learning in order to perform complex functions. The ANNs are composed of simple elements operating in parallel, adaptive and distributed systems. In most ANN implementations, the output of each articial neuron is computed by some non-linear function of the weighted sum of its inputs. Articial neurons, which are processing elements, are divided into multiple layers (Fig. 3), and their connections called edges, typically have assumed weights that are adjusting as the learning proceeds based on a technique called the backpropagation rule. 137 The architecture of the ANN is formed by input and output layer and a series of hidden layers, each of which is formed by a nite number of neurons. 138 This kind of ANN architecture is called a multi-layer network and is the most widely used. It should be noted that articial neurons in some cases may have a threshold such that the signal is only sent if the total signal crosses that threshold. Aer building the database which is divided into subsets, the neural network is trained for a specic task needed.
The training phase makes it possible to calculate the synaptic weights leading to each formal neuron. 139 The training of a network is a process where the set of adjusted weights is optimized to make the best prediction of the target variable on the basis of background variables. Multiple layer networks are trained with the standard back-propagation error algorithm. 137,139,140 This algorithm presents training examples to the network, examining the shi between the network output and the desired output while modifying the weights and biases of connections until the network produces an output very close to the desired one. 139 The back-propagation algorithm consists of a forward step where the signal propagates through the computational units until it gets to the output layer and a backward step where all the synaptic weights are adjusted accordingly to an error correction rule. 136 There are some factors that affect the ability of ANNs to model problems such as network structure, training and testing dates, network algorithms and nally transfer, training, learning, and performance functions. 141 Fig. 3 Schematic of the architecture of an ANN.

Applications of ANNs
ANNs offer fast solutions for a wide range of problems, especially the ones with real-time constraints. 139 Currently, it's mostly used in the eld of medicine for developing motor neurons 138 or improving the monitoring of neural activity. 142 Some works have been dedicated to predicting the sensitivity of strain sensing of piezoresistive materials, however computational cost can be considerably reduced by using machine learning. 143 In terms of sensor systems, various problems regarding CNTs and their composites have been covered, processed or solved using an ANN approach. Menacer et al. 139 developed a nite element model for an armchair CNTs using ANSYS simulator and examined their ability of detection of the applied force. It was found that the sensitivity is proportional to the nanotube length and that the non-linearity of the sensor depends on the applied force range. Moreover, this smart nanoforce sensor was proposed and investigated using ANNbased computation. 139 Matos et al. 144 made a FEM simulation of a CNT/polymer composite to produce training data for an ANN in the goal of developing a technique of precise tailoring of the electrical properties of composites. The authors conducted simulations of conductivity of different composites by adopting different conductivities of CNTs and the matrix, geometry of CNTs and their volume fractions of the composites. The objective was to reduce the number of problem parameters and it was concluded that their ANN model could facilitate other types of conductive composites. 144 A further study conducted by the same authors included a non-uniform multiaxial strain eld. 132 Fakhrabadi et al. 141 developed an ANN to predict the natural frequencies of CNTs with different types, lengths, and diameters and the effects of the attached masses on the rst ve natural frequencies of a special CNT. The well-known molecular mechanics and FE models are used to describe the vibrational behavior of this discrete structure before and aer attaching a particle on it. The authors 141 used MATLAB ANN toolbox to create a structure that consisted of several layers: one input, few hidden and one output layers. Hidden layers relate the input layer to output one by multiplying them to weight factors and adding some biases. First, the input data were inserted in ANN and the errors between real target values and the proposed target values by ANN were calculated. Next, the sensitivities start to propagate from the output layer toward the rst layer and then the weights and biases are updated. 141 Hayati et al. 137 proposed an ANN for modeling and simulation of the semiconductive SWCNT transistor device. The proposed ANN model was optimized based on data obtained by using the MATLAB script. 137 The application of machine learning implies very low computational cost and can be immediately used by relatively inexperienced users. 132 ANN can be a powerful tool to model different problems, especially in cases that there are not accurate governing mathematical relations for them. 141 ANNs in concrete mix design. The topic of using neural networks for predicting optimal concrete mix proportions has been explored many times in the past few decades. The research work 9-26 included investigating the compressive strength and workability of plain concrete or concrete with different aggregates and additions. Probably the largest contribution in this area was given by Dr I.-C. Yeh which created the rst database containing 1030 sets of data about properties of concrete specimens (mix proportions, age, compressive strength). The concrete specimens contained the additions of blast furnace slag, y ash and superplasticizer in different proportions. The database was given by the UCI repository and has been used in research works since then. 11,14 Some investigations have been directed toward specic types of concrete such as high-performance, 9,10 green, 16,17,22 self-compacting, 18 and lightweight concrete. Khan et al. 15 estimated the compressive strength of plain concrete using MATLAB ANN simulation and aer comparison of the mathematical results and 55 experimental results, concluded that multilayer feedforward neural network gives satisfactory results. Although the multilayer feed-forward approach is mostly used, 9,10,14,17,21,24 researchers oen use other techniques for predicting the compressive strength of concrete, such as fuzzy logic, decision tree or support vector machine.
The study carried out by Ziolkowski et al. 23 showed how ANNs can be applied in civil engineering practice. The authors designed the optimal ANN architecture and fed it with an extensive database including 741 concrete mix recipe record and a corresponding laboratory destructive test. The nal product was a system for obtaining optimal concrete mixtures which can be used by unexperienced users in everyday practice. Some authors showed the efficiency of ANNs by comparing the results from the ANN model with the results from fuzzy logic model, 20 multiple linear regression model, 25 random forest model, 21 decision tree, 21 support vector machine, 24 etc.

Structural health monitoring systems
Civil engineering structures are generally the most expensive investments, with concrete (cement) as the second most widely used material in the world. Civil and industrial structures are constantly exposed to damage caused by the weather conditions in combination with poorly designed material properties, damage due to the over-usage of the structure, accidental occurrence, earthquake, etc. These damages manifest in the form of cracks, corrosion, displacement, delamination, rotation, strain and more, potentially resulting in a catastrophe if not regularly and properly monitored.
Because a civil structure should meet the criteria of safety, durability, serviceability, and sustainability during its service life, constant supervision over the condition of the structure is necessary. Structural assessment is done manually by visual inspection and in the intervals prescribed by the designer and/ or regulations. These examinations are expensive, timeconsuming and labor-intensive, as well as highly subjective and can only consider the damage that is visible on the surface of the structure. 71,145 Having this in mind, the eld of structural health monitoring (SHM) was developed in order to follow structural changes taking place under different environmental inuences and loads.
The eld of SHM has, since its inception, been developed in terms of sensor systems going from cabled to wireless, to MEMS (microelectromechanical systems) and ultimately developing NEMS (nanoelectromechanical systems). 124,146 NEMS is currently the object of intense research activity in the eld of detection, where the main objective is to minimize the MEMS surface and power consumption using carbon nanotubes. 139,147,148 Nevertheless, more advanced and accurate methods for structural health monitoring are available and continue to be developed. 36,149 More advanced methods typically involve the analysis of measurements from sensors such as displacement transducers, 40 accelerometers and strain gauges, 150 which provide real-time information that may be continuously monitored from a central location and give insight into the condition of the structure. 36,145 One of the main limitations of sensors is that they are discrete xed-point directional sensors, not being a part of the material or structure that is being monitored. This on occasion leads to discrepancies in readings, showing the incompatibility between the sensor and the monitored element. The need to develop sensors that can be a part of the material and used for multidirectional and multiple location sensing is, therefore, imposed. 79 Self-sensing CNT/concrete composite is considered a good candidate for solving the durability and incompatibility issue between conventional sensors and concrete structures because it itself functions as both the structure and the sensor. 41,92,151 With the development of smart, self-sensing materials including CNT/concrete composite which provide an electrical response to stress or strain, 36,56,152 selfmonitoring structures can become a civil engineering practice.

Smart materials
Smart materials are engineered intelligent materials with properties adaptable to external stimuli (stress, moisture, temperature, light, acidity, radiation, etc.) in a controlled fashion. These composite materials are usually designed as a compound of the main insulating material and a ller material which gives it multifunctionality. Within the general eld of civil engineering, structural design and SHM are the areas which mostly emplore these materials. Multifunctionality is introduced to structures exposed to severe loading, such as skyscrapers, bridges, structures in seismically active areas and aesthetically important structures. Saa 97 presented how the use of CNTs has opened the door for new smart and advanced sensing materials that could effectively be used in structural health monitoring where wireless and real-time sensing 72,145 could provide information on damage occurrence and progression. CNTs have exhibited extraordinary mechanical properties, providing structural and functional capabilities simultaneously, including actuation, 52,153 sensing, and generating power. These abilities represent the possibility for developing actuators capable of high stress and strain operating at low voltage, and multi-functional electrochemical and mechanical sensors. 34 Piezoresistivity is dened as the change of electrical resistivity under the applied strain, and in order to obtain a piezoresistive cement-based material, an electrically conductive material such as carbon nanotubes has to be incorporated into a cementitious matrix. 124,154 The piezoresistive behavior of CNT networks was rst presented with CNT/polymer composites, whereupon application of mechanical load, the conguration of the networks was affected resulting in a change of the electrical resistance. 36,38,40,98,104 Zhu 43 concluded that piezoresistivity can be signicantly inuenced by the CNT types, aspect ratio, intrinsic electrical conductivity and the dispersion condition that occurs in the processing and curing stages. Kang et al. 37 presented an examination of a polymer lm strain sensor that was cast and bonded onto a glass ber beam in order to test the macroscale strain sensing characteristics. The beam displacement and change of resistance of the sensor were simultaneously measured to build a strain response model and to nd the sensitivity of the sensor. Because of their piezoresistivity, carbon-based structures have attracted special attention in the eld of high-performance nanomechanical sensors 43,56,101,155 and numerous studies in the range of experimental and computational methods have been carried out. 72,[156][157][158][159][160][161][162] State-of-the-art in monitoring systems The authors 107 conducted a detailed parametric study and proposed an approximate simplied model for smart concrete strain sensors that are easy to handle in practical full-scale applications. The conguration of CNTs to aligned bers is recognized as a natural embodiment for a one-dimensional material and the most effective to transfer axial properties to the macroscale. 54,78 Carbon nanotube networks are highly valuable in SHM for strain mapping, damage detection and identication of crack initiation and propagation. 54 Abot et al. 92 produced a thread spun from carbon nanotube forests by growing and drawning CNT forests that were twisted into a ne thread (or yarn) in order to be used as the sensor element. The authors proved 92 that this CNT thread is suitable as a distributed sensor material, allowing the development of a robust and efficient SHM system.
Kang et al. 37 proposed a new approach for developing smart composite material able to monitor its own health while the actuation capability allows it to actively improve the performance of the structure and extend its life. The high strength and elastic modulus, as well as piezoresistivity, indicate the possibility to produce a long continuous sensor that would be able to measure strain over a large surface area for SHM of concrete structures. The authors 37 developed a long continuous strain sensor (structural neuron) as an element in a biomimetic articial neural system used for SHM, which is capable of measuring large strain and forming a grid over a large area of a structure, and on the other hand, simultaneously has a loadcarrying capability as a structural component.
Fabrication of CNT neurons as long lms has been proposed by some authors. 163,164 The articial neural system is primarily developed within the medical eld where the application of CNTs includes electrical interfaces for neuronal stimulation and recording, 29 neural prosthetic devices for articially restoring impaired neural function, 165 encouraging neuronal growth, 28 retinal implants, 165 etc. In a functioning nervous tissue, neurons carry electrical signals via switching of the membrane potential caused by electrically and chemically gated ion channels in the cell membrane. Although the signal propagating along the length of the neuron is primarily achieved electrically, neurons communicate with each other at chemical junctions, synapses, 142 much so like the communication between CNTs by electron hopping at their junctions. The same principle constitutes the articial neural system which uses multiple neurons for real-time monitoring of strain of the structure. Since the production of neurons involves a mask to dene the pattern, it can take any shape including a grid that resembles that of the peripheral neural system of the human body. Sensors act as nerves and actuators act as muscles where the control center is the brain.
The idea of a self-sensing structure is to mimic a biological system through sensing, actuation, adaptability, self-repair, etc. and the ultimate form of intelligent structures are the ones that have the additional ability to learn in contrast to the preprogrammed response. This learning feature is realized by employing articial intelligence, namely machine learning and vis-à-vis articial neural networks. 145,166 The choice of ANNs to model so many different systems is, in part due to their exibility, adaptability and generalization capabilities and their easy application in soware and hardware devices 138 and materials. 144 Proposal for application of ANNs and CNT/concrete composites in structural health monitoring The authors of this paper propose a novel SHM system that would imply concepts of self-sensing concrete containing an articial neuron system in combination with articial neural networks. Real-time effective and long-distance monitoring of concrete structures is possible to obtain by developing a cementitious material representing a good host for the intricate branches of long continuous CNT neurons on one side and on the othertrained neural network capable of disregarding negligible data and processing relevant information about the structure's condition, so it may generate a 3D model showing possible damage and current state of structural elements. In the case of the maintenance planning of a new structure, ANN could also be used to obtain the optimal mixture of CNT/concrete composite in terms of desired specic properties of the end product such as e.g. high permeability, assigned porosity, high exure, self-cleaning ability, etc.

Conclusions
This paper reviews previous research work which is exploring the properties of CNTs and their inuence on concrete, and the use of articial neural networks in material science and concrete technology. The paper shows how incorporating carbon nanotubes in concrete mixtures or alternatively graing short CNTs to produce excessively long carbon bers, combined with the application of precisely trained ANNs represents a new direction in the evolution of structural health monitoring of concrete structures. The capability to tailor materials at the nanoscale has opened-up a new eld of research aimed at both tailoring their properties as well as developing intelligent systems that are capable of sensing and actuation. Several studies were conducted on the topic of different types of strain sensors 79,150,167 made with nanomaterials as llers. The goal is not only higher strength of the composite material, but also obtaining multifunctional smart and cost-effective structures. The behavior of CNT-based sensors and actuators involves multiple phenomena at length scale hierarchies ranging from quantum to macroscopic scale. Although there are still some challenges regarding the proper dispersion 168 of CNTs in the cement matrix and the exact volume fraction at the percolation threshold, 43 the results of the work presented in this paper are very promising and indicative of the beginning of a new direction in the development of structural health monitoring. Fundamental knowledge of the structure/property relations of CNT composites is still somewhat lacking which highlights the need to develop both experimental and analytical techniques to bridge these scales toward the optimization of SHM systems. 49 Nevertheless, smart nanomodied cementitious materials provide promising and innovative applications. 38 Moreover, despite much work devoted to making a perfect ANN structure, there are no conclusive rules on how to achieve it. The only way to produce a suitable ANN structure includes the 'trial and error' process, so the number of hidden layers and the number of neurons at each layer is adjusted for each specic problem. 140 Future investigation of the authors will focus on developing the methodology of production of self-sensing CNT/ concrete composites and specifying, modeling and training the ANN by using the obtained results of sensitivity testing.

Conflicts of interest
There are no conicts to declare.