1. Introduction
To determine the actual compressive strength of cast concrete, destructive tests are the most reliable methods. Typically, test samples are taken from the mixed concrete batch on site and then sent to the laboratory for curing and testing. Then, destructive tests are conducted to obtain the actual concrete compressive strength. Nevertheless, the concrete compressive strength test results in the laboratory might not be representative for the in situ cast concrete because factors such as concrete transportation, placement, tamping, and curing would affect the concrete quality and strength. In light of this, one alternative is to take core samples directly from the finished concrete structure and conduct destructive compressive tests on these drilled cylinder samples to determine the concrete strength. However, it is not always possible to take core samples on site and the structure might be damaged during the drilling process. As a result, non-destructive test (NDT) might be a good alternative to measure concrete strength at times when destructive tests are not preferable.
Two popular non-destructive test methods to estimate concrete compressive strength are the rebound hammer test (RH) and the ultrasonic pulse velocity test (UPV). For the rebound hammer test, a spring-controlled mass strikes the surface of the test object and the rebound is measured. The rebound of the mass depends on the hardness of the surface it strikes and the rebound values can be related to the compressive strength of the concrete [
1]. For the ultrasonic pulse velocity tests, two transducers are set against two surfaces of the test object. Ultrasonic waves are sent from one transducer and received by the other during the test. The travelling time of the ultrasonic pulse to pass from one surface to the other is recorded. With a known distance between the two surfaces of the test object, the velocity of the ultrasonic wave can be obtained. Generally speaking, the speed is higher when the concrete quality is better, in terms of density, uniformity, and homogeneity. As a result, the ultrasonic pulse velocity can be related to concrete compressive strength as well [
2].
Although non-destructive test methods are more economical and time-efficient comparing to destructive tests, one major drawback is that they are not very reliable when the test results are used to estimate the concrete compressive strength. For rebound hammer tests, it is shown that concrete compressive strength estimations have an average of over 20% mean absolute percentage error when comparing to the “actual” compressive strength obtained by destructive tests [
3]. Combined methods (combining rebound hammer tests and ultrasonic pulse velocity tests, also known as SonReb) are proposed by researchers to improve the concrete compressive strength estimations and positive results are obtained [
4,
5,
6,
7,
8,
9].
Most of these previous researches attempted to relate the concrete compressive strength with rebound hammer values and ultrasonic pulse velocity using statistical regression (linear or non-linear) analysis. This research proposes a novel regression approach by incorporating an artificial intelligence method (support vector machines; SVMs) as an alternative regression approach. Research has shown that SVMs have better regression ability when comparing to traditional statistical regression analysis [
10,
11,
12] and, hence, this research intends to incorporate SVMs to improve the regression results. In addition, more sample data are collected to improve the reliability of the research results. For most of the previous researches, only 10 to 20 samples are collected; in contrast, this research collected data from a total of 95 test samples. In order to collect these test samples, the research team cooperated with a certified construction material testing laboratory and obtained both destructive and non-destructive test results from 95 concrete cylinder samples. The data collected are used to develop and test the SVMs concrete compressive strength prediction model. The hammer rebound values and ultrasonic pulse velocities are set as the SVMs model input and the actual compressive strengths from destructive tests are set as the model output. Finally, results from SVMs model are compared with the results from traditional statistical regression analysis.
3. Support Vector Machines and Model Development
The support vector machines, SVMs, are a machine learning technique based on the minimization of structural risk and statistical learning theory and was first purposed by Vapnik [
18]. Through nonlinear mapping into high-dimensional feature space, the SVMs are able to classify sample data using linear models. Similar to neural networks, the SVMs also involve training and testing of data instances. The SVMs try to minimize the upper bound of the generalization error and this enables the SVMs to have better generalizability even when dealing with unseen data. Compared to other artificial intelligence techniques, the SVMs have several advantages, such as efficient use of high-dimensional feature space, uniquely solvable optimization problems, and the ability to be theoretically analyzed using computational learning theory [
19,
20]. The concepts of the SVMs model are briefly described below [
10].
Given a dataset
, where
xi is the input vector,
di is the target value, and
n is the size of the dataset, through a non-linear mapping (Φ) of
x into the high-dimensional feature space, the non-linear regression in the low-dimensional space can be represented by the linear regression in the high-dimensional space. It can be expressed as:
where ω is the weight vector; Φ is the high-dimensional feature space, and
b is the bias of the hyper plane.
The principles of the SVMs are to minimize the structural risks. As a result, ω and
b can be obtained by minimizing the risk penalty function below:
where
In Equation (2), is the estimated risk based on Lε (ε—insensitive Loss Function). is the penalty item for estimating the structural risk. C is the penalty constant and can be used to control the level of penalty when error occurs.
In order to estimate ω and b, slack variables ξ and ξ* are introduced and the new objective function becomes:
By introducing the Lagrange multipliers,
ai and
ai*, the SVMs decision function becomes:
Next, the Lagrange multipliers are introduced into the penalty objective function and the dual function can be obtained as:
K(
xi,
xj) is defined as the kernel Function, which is the inner product of
xi and
xj in the corresponding feature spaces ψ(
xi) and ψ(
xj),
. Kernel functions are functions that satisfy Mercer’s conditions and there are four common types of kernel functions used in SVMs: Linear Kernel, Polynomial Kernel, Sigmoid Kernel, and Radial Basis Function (RBF) Kernel. Among them, RBF is favorable for its capability of dealing with nonlinearity and high-dimensional computation, and effectiveness in reducing complexity for inputs by adjusting
C and γ [
21]. For this reason, this research uses the RBF Kernel when building the SVMs classifiers. The RBF Kernel is illustrated in Equation (9):
where γ is a constant.
For this research, the SVMs model development is implemented in the Matlab environment using the LS-SVMlab software developed by Suykens
et al. [
22]. The least square support vector machine, LSSVM, is first introduced by Suykens
et al. [
23]. The LSSVM is chosen because it eases the heavy computation burden for the constrained optimization programming for the SVMs by introducing equality instead of inequality constraints. In addition, LSSVM utilizes a least squares loss function instead of the ε—insensitive loss function. In this way, the problem is simplified in such a way that the solution is characterized by a Karush-Kuhn-Tucker (KKT) linear system instead of a computationally hard quadratic programming (QP) problem [
24,
25].
The hammer rebound (
Q-values) and ultrasonic pulse velocities measured from the test samples are taken as the model inputs and the actual compressive strengths are set as the model outputs. Among the 95 obtained test samples, 85 of them are randomly chosen as the training dataset and the remaining 10 samples are set as testing dataset for model validation. Both the training and testing datasets (EXCEL files) are first loaded into the Matlab environment and then used for the model development and testing. The Matlab environment and command window for the model development and testing process is illustrated in
Figure 5. The 85 sample data in the training dataset are used to train the SVMs model using the “trainlssvm” command in the Matlab. Afterwards, the trained SVMs are called up to generate predictions using the input data from the testing dataset (using “simlssvm” command). Then, the model prediction outputs are compared with the actual compressive strengths in the testing dataset to examine the prediction accuracy. It should be noted that the parameters
C and γ for the RBF kernel need to be determined during the SVMs model development. While a larger
C indicates a stronger smoothing, a higher value of γ forces the model to fit the training data more closely [
21]. There is no particular set of rules to determine these parameters and the optimal set of parameters is obtained by minimizing the training prediction bias through a series of trial and error in this research. For this particular sample, the parameters
C and γ obtained for the best SonReb SVMs model developed in this research are 2 and 10,000 respectively.
Figure 5.
SVMs model development process and commands in Matlab.
Figure 5.
SVMs model development process and commands in Matlab.
4. Results Analysis
To improve the non-destructive test for estimating the concrete compressive strength, SVMs models are developed for the combined NDTs (SonReb) using both the results from rebound hammer and ultrasonic pulse velocity tests as the model inputs. As the SVMs are supervised learning algorithms, the actual compressive strengths from destructive compression tests are set as the target output to train the model. Among the 95 collected test samples, 85 of them are randomly selected as the training dataset and the remaining 10 are set as the testing dataset. The ten samples in the testing dataset are “unseen” or new data for the trained SVMs model and they can be used to test the prediction ability of the model.
The model prediction accuracy is measured by the mean absolute percentage error (MAPE) illustrated in the equation below:
where
Ai is the actual concrete sample compressive strength and
Pi is the model predicted value.
For comparison purpose, a total of three SVMs models are developed with different input variable settings: (a) Hammer rebound value only; (b) Ultrasonic pulse velocity only; and (c) Hammer rebound value + ultrasonic pulse velocity. The output variable is set as the actual compressive strength for all three models. After training with the 85 test sample data, the 10 samples in the testing dataset are used to examine the model prediction accuracy. The model prediction outputs (estimated concrete compressive strength) are compared with the actual compressive strengths obtained by destructive compression tests. The model prediction results are summarized in
Table 1.
Table 1.
SVMs model prediction accuracy.
Table 1.
SVMs model prediction accuracy.
SVMs Models With Different Input Variable | Mean Absolute Percentage Error (MAPE) (%) |
---|
(a) Rebound Hammer (RH) Test | 8.92% |
(b) Ultrasonic Pulse Velocity (UPV) Test | 9.25% |
(c) SonReb (RH+UPV) Test | 6.77% |
As shown in
Table 1, the SVMs model with two input variables (hammer rebound value + ultrasonic pulse velocity) yields the best prediction results (6.77% MAPE). The MAPEs are obtained by comparing the model predicted concrete compressive strength with actual concrete compressive strength. As for the remaining two models with only one input variable, rebound hammer test model yields better prediction results (8.92% MAPE) when comparing to ultrasonic pulse velocity test model results (9.25% MAPE).
For comparison, linear regression with two input variables (hammer rebound value + ultrasonic pulse velocity) is also conducted. To be consistent, the same 85 training data are used to develop the multi-variable linear regression model. The linear equation obtained is illustrated in Equation (11). With the obtained linear equation, data from 10 testing dataset are used to calculate the concrete compressive strength estimation. The prediction results are summarized in
Table 2 and the MAPE for the multi-variable linear regression is 8.73%.
where
Y is model output (compressive strength, N/mm
2);
R is hammer rebound value and
V is ultrasonic pulse velocity.
When comparing the overall prediction accuracy, the MAPE for the SVMs model is 6.77% and the MAPE for the linear regression model is 8.73%. From
Table 2, it can be observed that most of the individual estimations from the SVMs are closer to the actual compressive strength. In addition, the variations are smaller for the SVMs model when the estimation errors are examined. Nevertheless, it should be noted that results obtained from this analysis are specific for the sample data collected in this research. It is a broad indication that the SonReb method produces better estimation than the single test method. Also, SVMs can be a good alternative to linear regression when developing the prediction model.
Table 2.
SVMs and linear regression model prediction.
Table 2.
SVMs and linear regression model prediction.
Concrete Compressive Strength (N/mm2) and Mean Absolute Percentage Error (MAPE,%) |
---|
SVMs Model | Linear Regression Model | Actual Strength |
---|
Strength | MAPE | Strength | MAPE | Strength |
---|
28.44 | 3.94% | 29.00 | 6.01% | 27.36 |
28.99 | 12.54% | 28.14 | 15.11% | 33.15 |
29.52 | 12.23% | 25.33 | 24.68% | 33.64 |
26.63 | 9.93% | 25.47 | 5.16% | 24.22 |
28.03 | 13.86% | 26.27 | 6.72% | 24.61 |
26.64 | 3.68% | 25.22 | 8.80% | 27.65 |
27.83 | 3.58% | 29.78 | 10.82% | 26.87 |
26.71 | 5.17% | 27.14 | 6.85% | 25.40 |
26.85 | 0.80% | 26.68 | 1.42% | 27.07 |
27.49 | 1.94% | 27.44 | 1.75% | 26.97 |