Determining the compressive strength (CS) of concrete in existing buildings is time-consuming, costly, and complicated. The most precise method for doing so is the destructive method of coring, whose disadvantages include damage to the area from which the core is taken, the loss of section capacity, after-coring repair and modification costs, testing costs, and a delay of at least one day between taking the core and obtaining the CS value.

Building codes - such as ACI-318-19 (2018), TS-500-2000 (2000), and TBEC-2018 (2018) - provide some rules for taking concrete core samples. They must be taken from an existing building if the supervising engineer requires tests to determine the CS of the concrete, and the coring must be done from locations that do not reduce the strength of the structure. Moreover, building codes allow the use of non-destructive testing methods (e.g., surface hardness and sound velocity) for correlation with the concrete coring results, but these alternative methods are less effective than taking core samples. For example, the success rate of the ultrasonic sound velocity method, which is frequently cited in the literature, is known to be only 65-75% (Sbartai et al., 2012; Breysse, 2012; Ferreira and Jalali, 2010).

There have been numerous studies involving different methods for determining the CS of concrete, and the success of each method has been assessed in comparison with that reported for the concrete pressure test in the literature (Bogas et al., 2013; Sbartai et al., 2012; Breysse, 2012; Ferreira and Jalali, 2010; Bingöl and Çavdar, 2018; Thapa et al., 2019). It has been established that each method has its own success rate, which is unfortunately inversely related to its ease of application. Therefore, it is highly necessary to devise a method that is both easily applicable and cost-effective while providing results both quickly and as close as possible to those of coring.

In recent years, images of the surfaces of newly produced (Doğan et al., 2015, 2017; Al-Kamaki et al., 2017; Beskopylny et al., 2022; Shiuly et al., 2022) or hardened cylinder and cube samples have been taken to calculate the CS of concrete via image processing technology. As an alternative to mechanical methods, image processing technology (Jang et al., 2019; Naderpour et al., 2018; López et al., 2009; Chang et al., 2009; Gencturk et al., 2014) allows measuring without direct contact with the surface of the material or requiring a complicated and expensive experimental setup.

Unlike previous studies, this study assesses the feasibility of determining the CS of reinforced concrete (RC) in existing buildings not by taking cores but via concrete-surface images analyzed using digital image processing (DIP) and an artificial neural network (ANN). The main objectives of this research are as follows: (i) to develop a new test method that engineers can use with an acceptable accuracy without damaging the structure; (ii) to make it easier to quickly reach results in the most time-consuming part of fieldwork regarding processes such as risk analysis in existing RC buildings and earthquake performance calculations; (iii) to provide significant cost and time savings by ensuring that concrete CS (one of the most critical factors in the behavior of a structure) can be determined both rapidly and economically; (iv) to offer an alternative and innovative method that uses expert systems instead of an uneconomical and time-consuming process that damages the structure (e.g., coring).

Compressive strength (f_c) is one the main parameters of concrete since it directly affects other mechanical parameters. Especially in the last major earthquakes (such as the 2023 Kahramanmaraj-Turkey earthquakes, in which approximately more than 50 000 people died and hundreds of thousands of RC buildings completely collapsed), it has been clearly seen that a low concrete strength may cause structure collapse. The method proposed in this study shows the extent to which concrete strength can be determined in existing RC buildings with an artificial intelligence-based algorithm and without the need for destructive methods. The novelty of this study is the ability to determine the concrete strength detectability ratio from the surface in buildings without taking core samples via a technology used in many engineering problems. This work fills the gap on this subject while showing huge differences between the experimental studies carried out in a laboratory environment and the tests carried out in real buildings.

To this effect, 32 different samples were produced while varying five parameters (water/cement ratio, aggregate type, the amount of binder, the compression to be applied to fresh concrete, and the amount of additive) that affected the CS and surface appearance of the concrete. The locations of 192 cores to be taken from these concrete samples were determined, and these regions were visualized before coring. The CS values of the core samples taken from these regions were matched with the surface images of the concrete, which were digitized via image processing. The DIP images formed the input layer of an ANN, whose output layer was the core CS of the imaged regions. The 192 data were used to train and test the ANN, and the tested model was validated in existing RC buildings. For the verification process, 20 different concrete images taken from randomly selected concrete buildings were used.

ANNs are a type of machine learning algorithm inspired by the structure and functioning of the human brain. They are designed to recognize patterns and learn from data inputs in a way that is similar to how the humans process information. An ANN consists of a large number of interconnected processing nodes, which are organized in layers. Each node receives an input from other nodes or from external data, processes it via a mathematical function, and then passes the output to other nodes in the network. The connections between the nodes are weighted, which means that some inputs are more important than others. During the training process, the network is presented with a large number of input-output pairs, and the weights of the connections between the nodes are adjusted in order to minimize the difference between the actual output and the desired output. This is typically done using an optimization algorithm such as a gradient descent.

ANNs are used in a wide variety of applications, including image and speech recognition, natural language processing, and predictive modeling. They have proven to be particularly effective in applications where the input data is high-dimensional and complex, and where traditional rule-based approaches are difficult to apply.

Digital image processing means processing images via computers or other digital devices. This process uses a range of mathematical, algorithmic, and hardware techniques to capture, process, analyze, refine, and ultimately make images more useful. Digital image processing is used in a variety of industries, such as medicine, security, computer vision, robotics, manufacturing, remote sensing, and entertainment. This technology is used to take image data, make them more meaningful, and facilitate analysis. Examples of this include X-ray images, air traffic control systems, image-based sensors in automated vehicles, etc.

Image processing involves emulating, in a computer environment, the visual mechanism of the human eye and the image-interpretation mechanism of the brain. With this technique, certain information that cannot be accessed directly from an image of either an entire object or a region of it can be accessed indirectly. An image can be defined by the function f (x, y) with values in the xy coordinate plane, each of which represents the color value of the image at those coordinates. The color values of monochrome images are defined either as black and white (0 and 1, respectively) or as grey levels (0-255) (González et al., 2004). Continuous density values in continuous space are expressed as discrete values using a digitization method, thereby allowing images to be displayed, improved, segmented, and transformed by computers (Easton Jr., 2010). This process is known as DIP, which, in this study was performed using the MATLAB software (MATrix LABoratory).

This section explains the production and acquisition stages of the core sample data and the application of DIP and the ANN, which constitutes the analytical part of this study.

In this study, the concrete dataset used for training and the test data of the algorithm comprised the 192 concrete samples obtained by taking six cores from 32 different concrete types. Because the CS of the concrete was a random variable, the frequency factor was taken to be 1,28 according to the probability distribution of the concrete CS (assuming an exceedance probability of 10%). The standard deviation of the distribution was calculated, and elements with either very low or very high strength were removed from the 192 concrete data (Figure 7). This process left 160 concrete samples in the data set.

Figure 7 Source: Authors

A 779 X 1 069 pixel image was taken from each of the 160 concrete samples and then converted to grey scale. Then, the average, standard deviation and median values of all the columns in each of the concrete images were calculated, thus yielding data comprising 3 X 1 069 = 3 207 pixels for each concrete image. Thus, a 160 X 3 207 pixel dataset was created for the 160 concrete samples.

The fact that the number of inputs for the ANN was as large as 3 207 had a negative effect on the training process in terms of time and calculation. To counter that problem, a feature-extraction algorithm based on principal components analysis (PCA) was used to reduce the number of properties for each concrete sample to 50, resulting in a data set in the form of a 160 X 50 matrix. The 160 data were divided into 128 training and 32 testing data to perform five-fold cross-validation (a separate 32-group for each cross-validation process was divided as testing data).

In this study, the Levenberg-Marquardt ANN model was used, with 50 neurons in the input layer, 52 in the interlayer, and one in the output layer (Figure 8). Table 5 presents the values of the other network parameters. In the literature, commonly used activation functions are the linear, step, sigmoid, Gaussian, and hyperbolic tangent. In this study, the selected activation function was the (logarithmic) sigmoid one. The ANN training process was terminated either after 50 iterations or if the training error became less than 10-6. The latter tended to happen after 10 to 12 epochs (Figure 9).

Figure 8 Source: Authors

Table 5

Properties	Variation range
Activation function	Logarithmic sigmoid
Epoch number	50
Stopping criterion	1e-6
Maximum validation error	6
Minimum error tolerance	1e-10
Marquardt setting parameter	0,005
Marquardt setting parameters reduction factor	0,1
Marquardt setting parameters growth factor	10
Marquardt setting parameters max value	1e10
How many epochs will be updated in the image	50

Figure 9 Source: Authors

The 160 data were divided into 128 training and 32 testing data to perform different five-fold cross-validations. For each cross-validation process, a separate group of 32 was determined as the testing data. After completing the training, the estimated values produced by the network for the 32 testing data were compared against the expected values. If the margin of error was less than 5% (between the estimate and the expected value), then the value was considered to be correct. The margin of error was then increased by 5%, and the correct predicted value ratio was determined. In light of the fact that destructive test methods include error levels of 30-40%, Table 6 shows (i) an accuracy of 73,6% with a 30% margin of error and (ii) an accuracy of 86,4% with a 40% margin of error. Table 6 shows the accuracy and error ratios, and Figure 10 shows a graph of the mean accuracy of the five-fold cross-validation results.

Table 6

Source: Authors

Figure 10 Source: Authors

After training and testing the ANN model, it was applied to the validation set. The laboratory results for the cores obtained from the buildings were compared to those obtained from the ANN. Table 7 lists the experimental CS values obtained from the core samples and the estimated CS values obtained via DIP and the ANN. Figure 11 compares the results graphically.

Table 7

Source: Authors

Figure 11 Source: Authors

Table 7 shows that, when the estimated and measured CS values (MPa) of the core samples are compared, some are close, and the results of different experiments were reached in most samples.

The results obtained in this study are somewhat poor when compared to similar studies in the literature (such as Bagyigit et al., 2012, and Waris et al., 2022, who found success rates close to 95% in their studies). However, the main reason for this is that the images were not taken from the surfaces of the concrete samples, which were tested by cutting (slicing) in a laboratory environment. However, there is not much chance of such a production being practically carried out in field studies. Although the results obtained for the laboratory-produced concrete samples are quite satisfactory, the success for the fieldwork samples in the validation set was limited. The proposed method had a success percentage of less than 50% for these samples. In this case, the training, testing, and validation set yield quite different results. The fact is that the laboratory-produced samples do not fully reflect the concrete in the real RC buildings; the images of the cores taken from the existing buildings are different from those of the concrete samples obtained in the laboratory. The results were affected by changes in the image quality of the samples and the difference in the amount of light. The reason for this was fluctuations in the images due to the roughness of the core samples. In this study, because the prediction success for the existing buildings was low, the authors conclude that using only surface images is insufficient for estimating the CS of concrete. It is thought that using images of internal concrete surfaces revealed by cutting the core samples will increase the success of predicting the CS. Similarly, high success was achieved from new concrete cylinder and cube samples (Jang et al., 2019) taken using the appropriate method, while the success in core samples was not satisfactory.

According to the results obtained in previous studies involving artificial intelligence and examples of various engineering problems, this study's results are poor. The main reasons for this are given below.

Although a comprehensive dataset was prepared for the laboratory test and a wide parameter group was used, the surface images obtained in the laboratory did not represent the results obtained from a real structure.

In the literature, for the determination of concrete CS, the success rates of the methods proposed in comparison with the core test results do not exceed this level. When evaluated from this point of view, it can be said that only the analyses made on laboratory samples are satisfactory. The authors think that a combined non-destructive testing method will be an important tool for determining CS in the future, not only with images, but also with other parameters such as sound permeability and surface hardness.

In this study, an experimental and analytical study was conducted using DIP and ANN together to determine the compressive strength of concrete in existing reinforced concrete buildings without cores. According to the study, the findings presented in the following items were reached.

The laboratory samples obtained an accuracy of 73,6% with a 30% margin of error and an accuracy of 86,4% with a 40% margin.