Professional Documents
Culture Documents
To cite this article: Mohammadjavad Arabpour Roghabadi, Mehdi Momeni & Zahra
Zangenehmadar (2018): Prediction of standard penetration test N-value from dynamic probing
light N-value using ANFIS and multiple regression models, International Journal of Geotechnical
Engineering, DOI: 10.1080/19386362.2018.1498578
Prediction of standard penetration test N-value from dynamic probing light N-value
using ANFIS and multiple regression models
a
Mohammadjavad Arabpour Roghabadi , Mehdi Momenib and Zahra Zangenehmadarc
a
Department of Building, Civil and Environmental Engineering, Concordia University, Montreal, Canada; bCivil Engineering Department, Islamic
Azad University of Kerman Branch, Kerman, Iran; cDepartment of Building, Civil and Environmental Engineering, Iran University of Science and
Technology, Tehran, Iran
CONTACT MohammadJavad Arabpour Roghabadi mohammadjavad.arabpour@concordia.ca Department of Building, Civil & Environmental Engineering
(BCEE), Concordia University, Office: H3G 2W1
© 2018 Informa UK Limited, trading as Taylor & Francis Group
2 M. ARABPOUR ROGHABADI ET AL.
collected from the field in different areas of the city of in the southeast of Iran on a thick fine-grained soil mass. The
Kerman in Iran. The actual and estimated SPT data are upper soil layers in the city are classified as lean clay (CL) and
compared to identify the best method for predicting the silt (ML) in the unified soil classification system. In some
SPT values. Results are discussed and conclusions are made. locations, particularly in the western parts of the city, recent
sand dunes are also found. The existing clay is a moist and
collapsible soil with a thickness of 15–20 m (Momeni et al.
Research methodology 2012). These layers extend from the ground level down to
more than 100 m in depth. As shown in Figure 2, SPT and
Data collection
DPL tests were performed simultaneously in all of the bore-
In this study, six sets of data were collected during the field holes at depth intervals of 1–1.5 m up to 8 m. The ground-
investigations in selected locations at the City of Kerman, water level is at a depth of 25 m. Laboratory tests (e.g. soil
which is illustrated in Figure 1. The city of Kerman is located gradation and Atterberge limits) were carried out to
DPL
SPT
determine the properties of soil, which were used after, in Table 2. Description of the input and output parameters in ANFIS.
order to determine the input data of ANFIS. Sets of SPT and Parameter Description Symbols (unit) Range
DPL tests were carried out in the boreholes based on Input Dynamic probing test DPL (N-value) 7–194
Depth D (m) 0.80–8
American Society for Testing and Materials and Stenzel and Actual density Γ (gr/cm3) 1.41–1.99
Melzer (1978) at the same depth intervals simultaneously. As Output Standard penetration test SPT (N-value) 5.95–64.7
a sample, the site condition and the variations in soil proper-
ties relative to depths for three investigated sites are dipected
in Table 1. As can be seen from Table 1, for some depths the determination (R2) between each input variable and output
Plasticity Index (PI) is equal to zero because for non-plastic variable was investigated. The coefficient of determination for
soil, there is no water content interval in which the soil PI, liquid limit (LL), relative humidity (w%), DPL N-value,
behaves like plastic. total density (ɣ), and depth (D) are calculated as 37.4%,
42.8%, 10.6%, 90%, 60%, and 51%, respectively. Among
these five parameters, DPL N-value, total density, and depth
SPT and DPL data collection were chosen as the inputs of system due to highest R2.
The SPT equipment consisted of a standard sampler and a Therefore, the basic equation for SPT N-value in silty clay
63.5 kg automatic trip hammer with a drop height of 760 mm in the city of Kerman can be expressed as a function of those
capable of averagely delivering 70% of the theoretical energy three parameters. All parameters were considered to have a
as measured during equipment calibration by the manufac- linear relationship with SPT N-value. The specifications of
turer. The DPL equipment consisted of a 10 kg hammer with inputs and output parameters are shown in Table 2.
a falling height of 500 mm. The rods are 22 mm in diameter
and are fitted to a 25.2 mm diameter cone of 90° apex angle. Adaptive network-based fuzzy inference system (ANFIS)
For six test locations, SPT was performed at 1.5 m discrete
intervals to a depth of 8 m below the ground level, and results The ANFIS which was started by Jang, Sun, and Mizutani
were recorded as a standard number of blows per 300 mm (1997) is classified as one of the best modelling systems
penetration denoted by Nm. Simultaneously, undisturbed and which permits the fuzzy systems to acquire the variables by
disturbed samples were collected from the boreholes for adaptive back propagation learning algorithm. ANFIS is a
further laboratory tests. DPL tests were performed, while multilayer feed-forward network where each node performs
recording measurements in the number of blows per a particular function on incoming signals as well as a set of
100 mm penetration denoted by N10. Other input parameters parameters relating to this node (Jang 1993). In fact, ANFIS
such as depth and total density were determined by site is known as a network that is capable of mapping unknown
investigations and laboratory tests, respectively. input parameters to their outputs by obtaining the rules
from previously seen data. A simple structure of ANFIS
with inputs of x and y and output of f is illustrated in
Multiple regression model Figure 3 which shows first the fuzzification layer (the pro-
Regression analysis is one of the popular methods in geotech- cess of changing the ‘crisp’ inputs into fuzzy inputs by
nical engineering. In this study, multi-variable regression assigning membership functions that represent fuzzy sets
method was used in SPSS V.20 to find the relationships of input values), and then the input is processed sequen-
between different parameters including independent and tially to the product layer, normalization layer, defuzzifica-
dependent ones and was applied to analyse data and to gen- tion layer (the process of producing a quantifiable result in
erate a model (Jalalifar, Mojedifar, and Sahebi 2014). In this crisp value), and finally the total output layer.
research, in order to determine the dependency between After both the multiple regression and ANFIS models are
different soil parameters and SPT N-value, the coefficient of implemented, their behaviours are assessed through perfor-
mance indices of variance accounted for (VAF), root mean
square error (RMSE), the coefficient of determination (R2),
Table 1. Sample of variation in soil properties with depth for three sites. and mean absolute percentage error (MAPE). The coefficient
Depth SPT DPL of determination is utilized in statistical analysis repeatedly
Site number (m) N-value N-value %W St LL PI γ
since it is simple to calculate and easy to understand and it
1 1 4 13 8.6 CL 28.34 9.72 1.55
1.65 15 26 12.42 CL 26.84 9.41 1.72 evaluates the percentage of total differences between pre-
2.65 14 31 11.09 CL 35.54 14.89 1.61 dicted and measured values with respect to the average. The
3.65 18 21 15.8 CL 29.61 11.45 1.7 VAF value fluctuates between [0, 1] where VAF of 1 denotes
4.56 13 26 15.81 CL 24.84 7.02 1.75
5.65 26 31 13.33 CL 35.14 14.24 1.76 that the predicted output has been estimated exactly the same
2 1.5 9 26 10.03 CL 18.19 1.57 1.41 as measured one. In recent research works, MAPE has been
2.35 6 18 10.38 CL 25.66 7.79 1.49 used mostly to evaluate the accuracy of a model due to its
3.1 21 69 8.5 ML 19.73 - 1.52
4.75 25 72 16.22 CL 21.72 2.3 1.61 simplicity. It identifies error as the proportion of actual data
6 16 37 16.81 ML 16.94 13.34 1.49 and higher accuracy comes with lower MAPE. This index can
3 0.85 14 53 7.71 CL 27.77 8.83 1.55 be divided into four indicators: high accuracy forecast (MAPE
1.85 17 48 14.1 CL 30.58 11.97 1.74
3.95 24 72 6.55 CL 24.64 5.93 1.55 < 10%), sound forecast (10% < MAPE < 20%), feasible fore-
5.35 43 114 13.79 CL 36.54 14 1.92 cast (20% < MAPE < 50%), and error forecast (MAPE > 50%)
6.55 65 207 15.09 CL 42.29 18.16 1.99 (Jia et al. 2015).
4 M. ARABPOUR ROGHABADI ET AL.
Figure 3. Sample architecture of ANFIS with two inputs and one output.
Results and discussion Table 5. Different parameter types and their values used for training ANFIS.
ANFIS parameter type ANFIS
In this study, 64 data points were used to develop the ANFIS MF type Gbellmf
and multiple regression models. Datasets were randomly Number of MFs 6
divided into 70% and 30% for training and testing purposes, Output MF Constant
Number of nodes 34
with 45 and 19 samples, respectively. The specifications of Number of linear parameters 8
these samples are shown in Tables 3 and 4. The ANFIS model Number of nonlinear parameters 18
was implemented using the fuzzy logic toolbox available in Total number of parameters 26
Number of training data pairs 45
MATLAB R2014. The model is fully automated in network Number of checking data pairs 0
generation and parameter tuning. The different membership Number of fuzzy rules 8
functions of triangular, trapezoidal, sigmoid, generalized bell,
Gaussian and S-shaped, were used as various trial runs to find
the best membership function. The existing data show the prediction of SPT N-value for silty clay is suggested as the
least error while using generalized bell membership function summation of input variables each of which multiplied by
(Gbellmf). The output membership function can be either a corresponding coefficients added by a constant number. The
constant membership function or a linear membership func- coefficients are determined by multiple regression. In other
tion. For the available data, constant output membership words, the input variables DPL N-value, total density, and
function results in the minimum error. The specifications of depth are multiplied by 0.225, 21.926, and 0.936, respectively,
the ANFIS model are demonstrated in Table 5. and the constant number is defined as −31.423.
The available data were inserted into SPSS V.20 software Figures 4 and 5 show the comparison between predicted and
in order to find the relationship between input variables and measured SPT N-values in multiple regression and ANFIS mod-
output. By performing the multiple regression method, the els, respectively. As can be seen from the graphs, the coefficient
of determination of the ANFIS model is 95.7% while it is 94.5%
in the multiple regression model. Since R2 is more than 90%,
Table 3. Sample dataset used for training ANFIS.
No. DPL γ D SPT
70
1 13 1.55 1 4
2 26 1.72 1.65 15
60
3 21 1.7 3.65 18 y = 0.9835x
4 26 1.75 4.56 13 R² = 0.9434
Predicted SPT N-value
5 18 1.49 2.35 6 50
6 69 1.52 3.1 21
7 72 1.61 4.75 26 40
8 37 1.49 6 16
9 53 1.55 0.85 14 30
10 48 1.74 1.85 17
20
50
are shown as 3D fuzzy surfaces in Figure 6. The DPL N-value
40 escalates with an increase in SPT N-value on a depth less than
30
10 m. In deeper depths, DPL N-value starts to decrease while
SPT N-value increases. This is because the soil type is different at
20 a deeper depth and it shows the effect of density and type of soils
on the SPT and DPL N-values. The SPT N-value declines when γ
10
is increased regardless of the depth.
0
0 10 20 30 40 50 60 70
Measured SPT N-value Conclusions
Figure 5. Predicted SPT N-values with respect to measured ones using ANFIS. There are several empirical equations to predict the SPT
N-value through the literature; however, the majority of
these equations consider solely one variable while several
Table 6. Comparison of errors in the proposed models.
factors are influencing SPT N-value in reality. In this paper,
Method RMSE MAPE VAF R2
ANFIS and multiple regression methods were used to predict
Multiple regression 3.09 16.30 95.3 94.5
ANFIS 2.76 13.13 97.8 95.7 SPT N-value from DPL N-value, depth, and total density. The
proposed models were developed and assessed with 64 data
points in which 70% and 30% were randomly chosen as
both models are accurate enough for practical applications. The training and testing datasets, respectively. Results show that
MAPE, VAF, and RMSE are calculated and results are shown in both models are able to predict SPT N-value close to the
Table 6. As can be seen from Table 6, the RMSE of the ANFIS expected results with the coefficient of determination (R2)
model is 2.76% while it is 3.09% in multiple regression calcula- more than 94%. Since R2 of multiple regression model was
tions. Furthermore, the MAPE of both models is between 10% less than that of the ANFIS model, the ANFIS model is
and 20%, which shows that the prediction models have a sound chosen as the more accurate one. Based on the results, the
forecast. The VAF of 97.8% and 95.3% for ANFIS and multiple ANFIS shows robustness in the prediction of SPT N-value.
regression models, which fall near 100%, indicate that the mod- Forecasting the SPT N-value from DPL N-value and other soil
els are good for the prediction of SPT N-value. Although the properties is beneficial since it eliminates the necessary efforts
comparison of results implies that the ANFIS model predicts the for obtaining SPT N-value and consequently reduces the cost.
SPT N-value more precisely than the multiple regression model, The results will be used for future preliminary evaluations
the black box nature of ANFIS is the main limitation of this and prediction of SPT N-value for geotechnical applications.
Figure 6. 3D fuzzy surfaces of depth, actual density, and SPT and DPL N-value.
6 M. ARABPOUR ROGHABADI ET AL.
Disclosure statement Jia, C., L. Wei, H. Wang, and J. Yang. 2015. “A Hybrid Model Based on
Wavelet Decomposition-Reconstruction in Track Irregularity State
No potential conflict of interest was reported by the authors. Forecasting.” Mathematical Problems in Engineering. doi:10.1155/
2015/548720.
Lingwanda, M. I., S. Larsson, and D. L. Nyaoro. 2015. “Correlations of
ORCID SPT, CPT and DPL Data for Sandy Soil in Tanzania.” Geotechnical
and Geological Engineering 33: 1221–1233. doi:10.1007/s10706-015-
Mohammadjavad Arabpour Roghabadi http://orcid.org/0000-0002- 9897-1.
3924-8375 Momeni, M., A. Shafiee, M. Heidari, M. K. Jafari, and M. R.
Mahdavifar. 2012. “Evaluation of Soil Collapse Potential in
Regional Scale.” Natural hazards 64: 459–479. doi:10.1007/s11069-
012-0252-z.
References Müller, R., S. Larsson, and J. Spross. 2013. “Extended Multivariate
Cabalar, A. F., A. Cevik, and C. Gokceoglu. 2012. “Some Applications of Approach for Uncertainty Reduction in The Assessment of
Adaptive Neuro-Fuzzy Inference System (ANFIS) in Geotechnical Undrained Shear Strength in Clays.” Canadian Geotechnical Journal
Engineering.” Computers and Geotechnics 40: 14–33. doi:10.1016/j. 51: 231–245. doi:10.1139/cgj-2012-0176.
compgeo.2011.09.008. Robertson, P. (2012). “Interpretation of In-situ Tests-Some Insights, 5th
Ching, J., -K.-K. Phoon, and Y.-C. Chen. 2010. “Reducing Shear Strength JK Mitchell Lecture”, 4th International Conference on Site
Uncertainties in Clays by Multivariate Correlations.” Canadian Characterization ISC-4, Porto de Galinhas, Pernambuco, Brazil.
Geotechnical Journal 47: 16–33. doi:10.1139/T09-074. Rogers, J. D. 2006. “Subsurface Exploration Using the Standard
Ching, J., -K.-K. Phoon, and J.-W. Yu. 2013. “Linking Site Penetration Test and the Cone Penetrometer Test.” Environmental
Investigation Efforts to Final Design Savings with Simplified & Engineering Geoscience 12: 161–179. doi:10.2113/12.2.161.
Reliability-Based Design Methods.” Journal of Geotechnical and Stefanoff, G., G. Sanglerat, U. Bergdahl, and K. J. Melzer (1990).
Geoenvironmental Engineering 140: 04013032. doi:10.1061/(ASCE) Dynamic Probing: International Reference Test Procedure: Proc 1st
GT.1943-5606.0001049. International Symposium on Penetration Testing, ISOPT-1, Orlando,
Jalalifar, H., S. Mojedifar, and A. Sahebi. 2014. “Prediction of Rock Mass USA. DOI: 10.1099/00221287-136-2-327
Rating Using Fuzzy Logic and Multi-Variable RMR Regression Stenzel, G., and K. Melzer. 1978. “Soil Investigations With Penetration
Model.” International Journal of Mining Science and Technology 24: Tests According to Din 4094.” Tiefbau 20: 3.
237–244. doi:10.1016/j.ijmst.2014.01.015. Tarawneh, B. 2017. “Predicting Standard Penetration Test N-Value
Jang, J.-S. 1993. “ANFIS: Adaptive-Network-Based Fuzzy Inference From Cone Penetration Test Data Using Artificial Neural
System.” IEEE transactions on systems, man, and cybernetics 23: Networks.” Geoscience Frontiers 8 (1): 199–204. doi:10.1016/j.
665–685. doi:10.1109/21.256541. gsf.2016.02.003.
Jang, J. S. R., C. T. Sun, and E. Mizutani. 1997. “Neuro-Fuzzy and Soft Yilmaz, I., and O. Kaynar. 2011. “Multiple Regression, ANN (RBF, MLP)
Computing-a Computational Approach to Learning and Machine and ANFIS Models for Prediction of The Swell Potential of Clayey
Intelligence [Book Review].” IEEE Transactions on automatic control Soils.” Expert Systems with Applications 38: 5958–5966. doi:10.1016/j.
42 (10): 1482–1484. doi:10.1109/TAC.1997.633847. eswa.2010.11.027.