Research Article Volume 6 Issue 1
^{1}Department of Civil Engineering, Abubakar Tafawa Balewa University, Nigeria
^{2}Federal Roads Maintenance Agency, Nigeria
Correspondence: Elinwa, Augustine Uchechukwu, Department of Civil Engineering, Abubakar Tafawa Balewa University, PMB 0248, Bauchi State, Nigeria
Received: January 28, 2020  Published: February 17, 2020
Citation: Uchechukwu EA, Austin O. Artificial neural network application to the compressive strength of palm kernel shell concrete. MOJ Civil Eng.2020;6(1):110. DOI: 10.15406/mojce.2020.06.00164
This work is on the application of Artificial Neural Network (ANN) to study the effects of using palm kernel shells (PKS) as aggregates on the compressive strength of concrete. ANN with an input neuron of 6 factors, 2 hidden layers, 18 and 12 each and 1 output neuron for the compressive strength were used for this work. A mix ratio of 1: 1.5: 3 with cement content of 382kN/m^{3}, watercement ratio of 0.55 were used for the work, and cured for 90 days. A total of fifty concrete mixes containing PKS in various proportions of 0 % to 40 % by wt. of the coarse aggregate were used for the training. For the validation and testing ten mixes were used.. Therefore, sixty (60) data sets were generated for which approximately eighty (80) percent was used for the training, and twenty (20) percent for the validation and test. The results showed that the distribution characteristic of PKSconcrete using ANN is adequate for the prediction of compressive strength. The predicted and experimental results are strongly correlation, with a model equation with an intercept, 1.5 and a slope of 0.93. The characteristic distribution results of the predicted with the experimental showed that the parameter estimates (ANN and Statistics), are within the 95 % confidence limits (CI), and very significant (P<0.05).
Keywords: ANN, palm kernel shells aggregate, compressive strength, statistical characteristics, PKsconcrete age
ANN, artificial neural network; NWA, normal weight aggregate; MLRA, multiple linear regression analysis; LWAC, lightweight aggregate concrete; NND, neural network design
The development and construction pressures on our conventional materials and the growing needs for sustainability have opened up new areas for research, thus motivating researchers to focus their investigations on the use of waste or recycled materials for potential use as construction materials. The use of palm kernel shells as replacement materials for normal weight aggregate (NWA) have attracted the interest of researchers in the palm oil producing nations for use in the concrete industry. Research works on the use of palm kernel shells as aggregate materials to replace conventional aggregates in concrete have been extensively carried out and some of these works were cited as references.^{1–12 }on the “characteristics of palm kernel shellconcrete.” The testing for strength of concrete specimen is complicated. It is equally a time consuming task. The performance of concrete is affected by many nonlinear factors, more importantly, is the cost of running these experiments, when the expected strengths are not achieved using the conventional materials and methods of mix design. The reasons are there are no fixed formulations for mixing concrete constituents to obtain the compressive strength because concrete mixing is predominantly a qualitative knowledgebased approach subjected to variations.
Faraqui et al.,^{13} and Khashman and Akpinar,^{14} have argued the fact that reliance on such an approach compromised the precision and accuracy of concrete properties and hence, necessitates the development of a reliable mixing formulation. They also stated that statistical modeling techniques like multiple linear regression analysis (MLRA) which have been used in the past have failed to accurately predict the compressive strength, because of the highly nonlinear relationship between the concrete proportions and its properties. Researchers therefore, have come with a new trend in modern concretes involving the applications of new methods like the artificial neural network (ANN) which will enhance creditability and acceptability of our concrete works containing additives and non conventional materials. ANNs are software constructs that can be trained by example to recognize (clarify), or estimate (predict) results from a variety of inputs. The ANN ability to learn so quickly is what makes them so powerful and useful for a variety of tasks, and contains three main sections, classified as, input layer, hidden layer, and output layer.
Faruqi et al.,^{13} developed a neural network model for predicting the compressive strength of concrete for different mixdesign parameters. This was based on five (5) hidden layers and trained using the results of a series of previously conducted experiments. Each experiment consisted of five (5) parameters and a corresponding compressive strength obtained from 28days cylinders tests. They observed that the neural network model performed with satisfactory results in predicting the 28day compressive strength of concrete.
Yoon et al.,^{14} worked on the mechanical properties of lightweight aggregate concrete using ANN. The study presented the ANNbased prediction for compressive strength and elastic modulus of lightweight aggregate concrete (LWAC), and concluded that the ANN model showed acceptable prediction accuracy with respect to the compressive strength and elastic with respect to compressive strength and elastic modulus of LWAC, and that the highest prediction accuracy was obtained by the ANN model compared to the use of statistical linear and nonlinear regression model.
The reviewed works on this topic were largely to some extent on the strength at 28days, and are equally based on empirical and statistical methods. These methods have limited applications, and are characterized with the issues of accuracies and precisions. Therefore, the present study on the compressive strength of PKSconcrete addressed these issues, firstly, curing beyond the conventional 28 days to 90 days, and secondly, by employing the ANN to address the issues of accuracies and precisions Khashman and Akpinar,^{14} Golizadeh and Namini.,^{15} Faruqi et al.,^{13} Suryadi et al.,^{16}. The ANN architecture used had six (6) inputs with six (6) neurons, two (2) hidden layers with eighteen (18) and twelve (12) neurons respectively, and one (1) output layer with one neuron.
The physical and chemical properties of the cement are shown in Tables 1 & 2 and conformed to BS EN 1963.^{17} The fine aggregate was river sand which is free from deleterious matters with a specific gravity of 2.64, bulk density of 1528 kg/m^{3}, and moisture content of 0.42. The fine aggregate was uniformly graded and falls into zone 2 of the grading curve. Table 3 is the sieve analysis of the fine aggregate. The coarse aggregate was sourced from a quarry site in Bauchi town and has a maximum size of 20mm. The physical characteristics of the coarse aggregate and PKS are shown in Table 4, while Table 5 is the particle size distribution. Both the fine and coarse aggregates conform to BS EN 1097.^{18} The palm kernel shell used was sourced from Ekpoma, Edo State, Nigeria. Ekpoma is a town in the savannah region, in the southern part of the country.
Parameter 
Value 
Specific gravity 
2.98 
Bulk density(kg/m^{3}) 
1475 
Specific surface area(Blaine)m^{2}/kg 
355 
Loss on ignition(%) 
1.51 
Moisture content 
0.39 
pH 
12.4 
Table 1 Physical properties of ashaka pc
Oxide composition 
Percentage by weight (%) 
CaO 
62.12 
SiO_{2} 
20.69 
Al_{2}O_{3} 
6.14 
Fe_{2}O_{3} 
2.32 
SO_{3} 
1.63 
MgO 
1.22 
Na_{2}O 
0.9 
K_{2}O 
1.01 
Table 2 Chemical properties of ashaka pc
Sieve size 
Cumulative % passing 
5.00mm 
_ 
2.00mm 
93.00 
1.18mm 
78.00 
600μm 
43.80 
300μm 
20.40 
150μm 
8.20 
63μm 
8.16 
Pan 
0.00 
Table 3 Sieve analysis of fine aggregate
Physical properties 
Gravel 
PKS 
Specific gravity 
2.75 
1.33 
Bulk density (kg/m3) 
1,714 
694 
Water absorption (%) 
0.04 
17.48 
Void ratio 
0.38 
0.48 
Porosity 
0.27 
0.32 
ACV (%) 
3.69 
0.36 
AIV (%) 
7.28 
1.91 
Table 4 Characteristics of the coarse aggregate and palm kernel shells
Sieve size(mm) 
Cumulative Passing (%) 

Coarse aggregate 
Palm kernel shell 

75 
 
 
63 
 
 
50 
 
 
37 
 
 
28 
 
 
20 
94.4 
 
14 
65.9 
98.5 
10 
24.9 
64.8 
6.3 
3.7 
10.7 
5 
1.5 
3.1 
3.35 
0.6 
1 
Pan 
0 
0 
Table 5 Particle size distribution of the coarse aggregate and palm kernel shell
The experiment to study the characteristics of PKSconcrete was mounted using a concrete mix ratio of 1:1.5:3, with a cement content of 382 kg/m^{3} and watercement ratio of 0.55 (Table 6). Four mix parameters were used and labeled M00, M10, M20, and M30, respectively. The mix labeled M00 was the control, while the rest were having the various replacement levels of PKS by wt % of the crushed aggregate. The experiment for the compressive strength test was carried out using mould cube sizes of 100 mm, and tested in accordance with ASTM C192/C192M.^{19} They were cured for 3 days to 90 days before testing to failure using a motorized ELEmachine. At the end of each curing regime, three samples were tested to failure and the average recorded. The experimental results are shown in Table 7.
Mix type 
Cement (kg/m^{3}) 
PKS (kg/m^{3}) 
Sand(kg/m^{3}) 
Cement (kg/m^{3}) 
Water(kg/m^{3}) 
W/C 
M0 
1265 
 
543 
382 
210 
0.55 
M10 
1138.5 
126.5 
543 
382 
210 
0.55 
M20 
1012 
253 
543 
382 
210 
0.55 
M30 
885.5 
379.5 
543 
382 
210 
0.55 
Table 6 Concrete mix proportions for the experiments
Mix No 
3d 
7d 
28d 
60d 
90d 
M00 
17.9 
21.3 
24 
28.3 
33.8 
M10 
22.2 
23.3 
24 
24.8 
26.5 
M20 
12.9 
14.5 
16.9 
19 
21.2 
M30 
8.5 
10.4 
11.8 
13.3 
15.6 
Table 7 Compressive strength experimental result
The compressive strength of PKSconcrete Table 7 showed that as the replacement levels increased, the compressive strength decreased. The maximum strength was at 10%. At this replacement the strengths at 60days and 90 days above the strengths at 28days are 3% and 10% respectively The reductions in strength have been attributed to many factors such as the low strength of PKS compared to the crushed aggregate, the irregular shape of PKS which could prevent adequate compaction, and the bonding between PKS and cement paste because of the smooth surfaces of the PKS.^{20}
Table 8 shows the distribution characteristics of the PKSconcrete. Measurements were made on the mean, standard error of the mean (SE.Mean), standard deviation (Std.Dev) and coefficient of variation (Coef.Var) for the curing period curing (90 days). The values achieved on the measurements showed good uniform characteristics of PKSconcrete.
The Reliability/Survival studies of the distribution analysis (Arbitrary Censoring), using the Parametric Distribution Analysis (PDA) method in Minitab 17 Software showed the degree to which these tests were consistent and stable in measuring what were intended to measure, and that the tests actually measured what they claimed to measure. These validate the extent to which inferences, conclusions, and decisions made on the basis of these test measurements are appropriate and meaningful. The results are shown in Table 9. The ninetyfive percent confidence interval characteristics of the distributions for three parameters Mean, Std. Dev and Median are given in table
Age(Days) 
Mean 
SE mean 
Std. Dev 
Variance 
CoefVar 
3 
15.4 
3 
6 
35.5 
38.7 
7 
17.4 
3 
6 
35.8 
34.4 
28 
19.2 
3 
6 
35.4 
31 
60 
21.4 
3.3 
6.6 
43.5 
30.9 
90 
24.3 
3.9 
7.8 
60.1 
31.9 
Table 8 Distribution characteristics
Age (days) 
Parameter 
Estimate 
Standard error 
95 % Confidence 
Interval 
Mean (MITF) 
15.44 
2.52 
11.21 
21.26 

Std. Dev 
5.02 
1.66 
2.63 
9.6 

3 
Median 
15.43 
2.69 
10.96 
21.71 
Mean (MITF) 
17.47 
2.49 
13.21 
23.1 

Std. Dev 
4.96 
1.67 
2.56 
9.61 

7 
Median 
17.58 
2.64 
13.1 
23.59 
Mean (MITF) 
19.29 
2.42 
15.08 
24.67 

Std. Dev 
4.83 
1.67 
2.45 
9.51 

28 
Median 
19.49 
2.54 
15.09 
25.17 
Mean (MITF) 
21.45 
2.75 
16.67 
27.58 

Std. Dev 
5.49 
1.84 
2.85 
10.58 

60 
Median 
21.65 
2.89 
16.67 
28.13 
Mean (MITF) 
24.34 
3.39 
18.52 
31.98 

Std. Dev 
6.74 
2.15 
3.61 
12.58 

90 
Median 
24.51 
3.58 
18.41 
32.62 
Table 9 Confidence interval (95%) characteristics of distribution of PKSconcrete
Mix proportions
$MSE=\frac{1}{mn}{\sum}_{i=1}^{n}{\sum}_{j=1}^{m}{({t}_{ij}{y}_{ij})}^{2}$(1)
Mix for the training input 

Runs 
Mix no 
Cement(kg/m^{3}) 
Fine agg.(kg/m^{3}} 
Coarse agg.(kg/m^{3}) 
PKS(kg/m^{3}) 
Water(kg/m^{3}) 
1 
M00 
382 
543 
1265 
0 
210 
2 
M00 
382 
543 
1265 
0 
210 
3 
M10 
382 
543 
1138.5 
126.5 
210 
4 
M10 
382 
543 
1138.5 
126.5 
210 
5 
M10 
382 
543 
1138.5 
126.5 
210 
6 
M20 
382 
543 
1012 
253 
210 
7 
M20 
382 
543 
1012 
253 
210 
8 
M20 
382 
543 
1012 
253 
210 
9 
M30 
382 
543 
885.5 
379.5 
210 
10 
M30 
382 
543 
885.5 
379.5 
210 
11 
M00 
382 
543 
1265 
0 
210 
12 
M00 
382 
543 
1265 
0 
210 
13 
M00 
382 
543 
1265 
0 
210 
14 
M10 
382 
543 
1138.5 
126.5 
210 
15 
M10 
382 
543 
1138.5 
126.5 
210 
16 
M20 
382 
543 
1012 
253 
210 
17 
M20 
382 
543 
1012 
253 
210 
18 
M30 
382 
543 
885.5 
379.5 
210 
19 
M30 
382 
543 
885.5 
379.5 
210 
20 
M30 
382 
543 
885.5 
379.5 
210 
21 
M00 
382 
543 
1265 
0 
210 
22 
M00 
382 
543 
1265 
0 
210 
23 
M10 
382 
543 
1138.5 
126.5 
210 
24 
M10 
382 
543 
1138.5 
126.5 
210 
25 
M10 
382 
543 
1138.5 
126.5 
210 
26 
M20 
382 
543 
1012 
253 
210 
27 
M20 
382 
543 
1012 
253 
210 
28 
M20 
382 
543 
1012 
253 
210 
29 
M30 
382 
543 
885.5 
379.5 
210 
30 
M30 
382 
543 
885.5 
379.5 
210 
31 
M00 
382 
543 
1265 
0 
210 
32 
M00 
382 
543 
1265 
0 
210 
33 
M00 
382 
543 
1265 
0 
210 
34 
M10 
382 
543 
1138.5 
126.5 
210 
35 
M10 
382 
543 
1138.5 
126.5 
210 
36 
M20 
382 
543 
1012 
253 
210 
37 
M20 
382 
543 
1012 
253 
210 
38 
M30 
382 
543 
885.5 
379.5 
210 
39 
M30 
382 
543 
885.5 
379.5 
210 
40 
M30 
382 
543 
885.5 
379.5 
210 
41 
M00 
382 
543 
1265 
0 
210 
42 
M00 
382 
543 
1265 
0 
210 
43 
M10 
382 
543 
1138.5 
126.5 
210 
44 
M10 
382 
543 
1138.5 
126.5 
210 
45 
M10 
382 
543 
1138.5 
126.5 
210 
46 
M20 
382 
543 
1012 
253 
210 
47 
M20 
382 
543 
1012 
253 
210 
48 
M20 
382 
543 
1012 
253 
210 
49 
M30 
382 
543 
885.5 
379.5 
210 
50 
M30 
382 
543 
885.5 
379.5 
210 
Table 10 Mix proportions for network training
Mix Proportion for the Validation 

Runs 
Mix No 
Cement(kg/m^{3}) 
Fine agg (kg/m^{3}) 
Coarse agg(kg/m^{3}) 
PKS(kg/m^{3}) 
Water(kg/m^{3}) 
1 
M00 
382 
543 
1265 
0 
210 
2 
M30 
382 
543 
88.5 
379.5 
210 
3 
M10 
382 
543 
1138.5 
126.5 
210 
4 
M20 
382 
543 
1012 
253 
210 
5 
M00 
382 
543 
1265 
0 
210 
6 
M30 
382 
543 
885.5 
379.5 
210 
7 
M10 
382 
543 
1138.5 
126.5 
210 
8 
M20 
382 
543 
1012 
253 
210 
9 
M00 
382 
543 
1265 
0 
210 
10 
M30 
382 
543 
885.5 
379.5 
210 
Table 11 Mix proportions for the validation
Where t is the largest target value and y is the output value.
The activation function was the sigmodal function with the epoch number set to 10000 to avoid over fitting and training. The process is defined as a nonlinear inputoutput relation between the influencing factors (Cement content, FA content, CA content, PKS content, Water content and Age of concrete [390days] and the compressive strength cured [390days].
The Levenberg Marquardt algorithm was chosen as the most efficient one for the training of the ANN. Approximately; eighty(80) percent of the data in Table 12 was used for the training, and was stopped when the network prediction closely matched the experimental results to avoid over fitting of the network. Figure 2 is the MSE/Epoch results for the training output with a minimum final mean square error of 0.0176 (1.76%). This stabilized at an epoch value of 412. Twenty(20) percent of the total data as shown in Table 13 were used for validation and testing. Figure 3 showed the test and validation of the MSE/Epoch results. The minimum final mean square error for the validation and test was 0.0267 or 2 67%, and stabilizes at 229. After the testing and validation the predicted results were compared with the experimental data. Table 14 shows the predicted output with respect to the experimental results and the error is approximately ±5. This shows a very strong correlation between the two results. The output against target model generated for the predicted and experimental results of the compressive strength is shown in Figure 4, and the model equation is given as:
${f}_{predicted}=1.5+0.93{f}_{experimental}$(2)
Training 
Training 

Runs 
Mix No 
Age (Days) 
Comp. str(kN/m^{3}) 
Runs 
Mix No 
Age(Days) 
Comp. str. (kN/m^{3}) 
1 
M00 
3 
18 
26 
M20 
28 
17.2 
2 
M00 
3 
17.7 
27 
M20 
28 
17.2 
3 
M10 
3 
22.7 
28 
M20 
28 
16.5 
4 
M10 
3 
24 
29 
M30 
28 
11.8 
5 
M10 
3 
20.1 
30 
M30 
28 
12 
6 
M20 
3 
14.4 
31 
M00 
60 
28 
7 
M20 
3 
10.8 
32 
M00 
60 
28 
8 
M20 
3 
13.5 
33 
M00 
60 
29 
9 
M30 
3 
8.4 
34 
M10 
60 
25 
10 
M30 
3 
8.6 
35 
M10 
60 
24.8 
11 
M00 
7 
23.2 
36 
M20 
60 
24.6 
12 
M00 
7 
19.2 
37 
M20 
60 
18.8 
13 
M00 
7 
21.6 
38 
M30 
60 
13.5 
14 
M10 
7 
24 
39 
M30 
60 
13.5 
15 
M10 
7 
22 
40 
M30 
60 
13 
16 
M20 
7 
14 
41 
M00 
90 
34.8 
17 
M20 
7 
15.2 
42 
M00 
90 
31 
18 
M30 
7 
10.2 
43 
M10 
90 
17 
19 
M30 
7 
10.4 
44 
M10 
90 
26.4 
20 
M30 
7 
10.6 
45 
M10 
90 
26 
21 
M00 
28 
25,6 
46 
M20 
90 
21 
22 
M00 
28 
23.5 
47 
M20 
90 
21.2 
23 
M10 
28 
24.5 
48 
M20 
90 
21.4 
24 
M10 
28 
23.5 
49 
M30 
90 
16 
25 
M10 
28 
24 
50 
M30 
90 
15.8 
Table 12 Output results (training)
Validation output results 

Runs 
Mix No 
Age (days) 
Comp. Str 
1 
M00 
3 
17.9 
2 
M30 
3 
8.6 
3 
M10 
7 
23.8 
4 
M20 
7 
15.2 
5 
M00 
28 
23 
6 
M30 
28 
11.5 
7 
M10 
60 
24.6 
8 
M20 
60 
19.2 
9 
M00 
90 
35.6 
10 
M30 
90 
15 
Table 13 Validation results
Property 
Experimental versus predicted 

Mix no 
Age(days) 
Experiment 
Predicted 
Error 

Compressive Strength 
M00 
3 
17.9 
22.9 
5 
M30 
3 
8.6 
10.1 
5.2 

M10 
7 
23.8 
19.1 
4.7 

M20 
7 
15.2 
14.7 
0.5 

M00 
28 
23 
25.6 
2.6 

M30 
28 
11.5 
12.1 
0.6 

M10 
60 
24.6 
24.2 
0.4 

M20 
60 
19.2 
19.4 
0.2 

M00 
90 
35.6 
32 
3.6 

M30 
90 
15 
16.5 
1.5 
Table 14 Experimental versus predicted results
with a correlation coefficient (r^{2}) of 97.0 %. This shows a very high correlation between the experiment and the predicted. Sensitivity analysis on the experimental and predicted results using the Minitab 17 Statistical Software is given as:
${f}_{predicted}=3.99+0.806{f}_{experimental}$ (3)
The regression model is significant with a pvalue of 0.000, a standard deviation (s) of 2.532, and a correlation coefficient (r^{2}) of 87.24%, The constant and experiment are significant with pvalues of 0.116 and 0.000, respectively. Figures 5,6 are the normality and residual plots.
Figures 7, 8 are the 3D surface plots of the experimental, predicted and age of PKSconcrete on one hand and the experimental, predicted and the error. The errors are within ±5.
The distribution characteristics of the experiment and the predicted results Table 15 are within the 95% CI, and very significant (p<0.05). The narrower the CI the better it is Mannan MA et al.^{21} If the CI is narrow, we can be quite confident that any effects far from this range had been ruled out by the study.^{22,23}
Goodness of fit. [AndersonDarling Adj] 
Basic statistics 
Estimates 
Std. error 
95 % Normal CI 

Lower 
Upper 

Parameter 
Mean (MTTF) 
19.47 
2.37 
15.33 
24.71 

Experiment Result 
1.443 
Standard Deviation 
7.46 
1.48 
5.05 
11.02 
Median 
19.2 
2.53 
14.83 
24.85 

14.06 
2.56 
9.84 
20.09 

Mean (MTTF) 
19.69 
2.03 
16.09 
24.1 

Predicted Result 
Standard Deviation 
6.4 
1.27 
4.34 
9.44 

1.365 
Median 
19.68 
2.16 
15.87 
24.4 

15.1 
2.32 
11.25 
20.5 
Table 15 Characteristics of distribution for the experimental and predicted results
The application of ANN to study the compressive strength of palm kernel shell (PKS) concrete has been evaluated, and the following are the conclusions.
None.
The author declares that there are no conflicts of interest.
None.
©2020 Uchechukwu, et al. This is an open access article distributed under the terms of the, which permits unrestricted use, distribution, and build upon your work noncommercially.