# Application of Artificial Neural Network, Fuzzy Inference System and Adaptive Neuro-Fuzzy Inference System to Predict the Removal of Pb(II) Ions from the Aqueous Solution by Using Magnetic Graphene/Nylon 6.

^{*}

**Corresponding Author:**Mahmoud Reza Sohrabi, Department of Chemistry, Islamic Azad University, Tehran, Iran, Tel: 0098-21-7700 98 36-42, Email: [email protected]

*
Received Date: Mar 28, 2018 /
Accepted Date: Apr 02, 2018 /
Published Date: Apr 26, 2018 *

### Abstract

In this study, three modeling techniques based on artificial intelligence were used to predict the removal percent of lead(II) ions from the aqueous solution. These models include Artificial Neural Network (ANN), Fuzzy Inference System (FIS) and Adaptive Neuro-Fuzzy Inference System (ANFIS). Magnetic graphene adsorbent supported on nylon 6 was used for removing lead(II) ions. Optimal conditions for the experimental parameters were performed using the Taguchi methodology. The analysis of variance (ANOVA) test at the 95% confidence level was applied to the results of these models which suggested there were no significant differences among these models.

**Keywords:**
Artificial intelligence models; Magnetic graphene; Nylon6; Lead(II) ion.

#### Introduction

Modeling of chemical processes can lead to lower costs during testing. By using the ability of these models, we can predict the optimal conditions for a process. Intelligent computer systems are the suitable tool that can be employed for this purpose. Artificial Neural Network (ANN), Fuzzy Inference System (FIS) and Adaptive Neuro-Fuzzy Inference System (ANFIS) are part of the intelligent models. ANN is a non-linear statistical data modeling that is inspired by biological neurons. Each neuron is related to a mathematical function with determined inputs, a scientific computation method, and outputs [1,2]. Fuzzy set theory was introduced by Lotfi Zahed in 1965. Fuzzy inference is a process of mapping from a given input to an output data set using the theory of fuzzy sets [3]. On the other hand, ANFIS is a fuzzy inference system combined with the computational power of ANN and acts as an adaptive multilayer feed-forward network. ANFIS is a potent approach to modeling the input and output relationship in non-linear and complex systems [4].

Heavy metals can enter a water supply by industrial and consumer waste by means of various human activity such as mining, fertilizer industries, refining ores, tanneries, battery manufacturing, paper industries, metal-based pesticides and fuels, or even from acidic rain breaking down soils and releasing heavy metals into streams, lakes, rivers, and groundwater [5,6]. Among heavy metals, lead is considered as longstanding environmental contaminant. Lead in large dosage can seriously harm human life and aquatic ecosystems. Where exceeding the permissible concentration limit of lead in the human body can end up with acute or chronic problems such as mental retardation, seizures, anemia, kidney and liver disorder, cancer and hepatitis [7-9]. Therefore, removal of Pb(II) ions from the aqueous system is very important. There are various methods that have been explored to decrease lead(II) ions, namely membrane filtration, chemical precipitation, solvent extraction, ion-exchange, oxidation/reduction, electrode deposition, and bio-adsorption [10-12]. However, these methods have some limitations. Among these methods, adsorption process has been widely used because it is a simple and relatively economical process [13,14]. Graphene (G) as an adsorbent is a twodimensional form and honeycomb with sp2-bound carbon atoms [15]. It has many unique properties such as singular high room temperature, very high thermal conductivity, flexibility and tensile resistance. This feature has made it a potential adsorbent for the removal of heavy metal ions like lead from the aqueous solution [16-18]. Graphene is produced from graphite based on Hummers and offeman’s procedure [19]. In this study, Artificial Neural Network with three algorithms, Fuzzy Inference System and Adaptive Neuro-Fuzzy Inference System have been used for predicting the removal percent of lead ions from the aqueous solution using magnetic graphene oxide supported on nylon 6. The effects of various experimental parameters such as pH of the solution, initial Pb(II) ions concentration, and adsorbent dosage were investigated based on Taguchi experimental design to optimize the absorption performance. To evaluate the significant differences between the reported variances of recovery for removing lead(II) ions the results of ANN, FIS and ANFIS models were compared using ANOVA analysis.

#### Theory

**Artificial neural networks model**

Artificial Neural Networks include a contiguous network of nodes, which are separated into many layers. The input and output layers in the neural networks make the main structure of artificial neural networks. In addition, there are a series of hidden layers between input and output layers. The main duty of hidden layers is to evaluate the relationship between unknown and complex by iterative training from many input-output pairs. The hidden layer has some nodes that they have activation functions and numeric weights that are controlled by appropriate learning algorithms to obtain the best possible correlation between an input and output set [20]. **Figure 1** shows the architecture of the artificial neural network pattern. There is also a weight level between two adjacent levels of input, hidden and output [21,22]. In this research, neural fitting with three algorithms: the Levenberg- Marquardt (LM), Conjugate and Bayesian algorithm were used to evaluate the best algorithm based on prediction accuracy.

**Fuzzy inference system model**

Fuzzy Inference System applies a nonlinear mapping from a given input to the output using the theory of fuzzy sets. Several fuzzy if-then rules used in this mapping. The parameters of the if-then rules specify a fuzzy area of the input space, and the output parameters [23]. FIS consist of three important parts: rule base, which contains a selection of fuzzy rules, database, which specifies the membership functions (MF) were applied in the fuzzy rules and a reasoning mechanism, evaluates which rules are suitable at the current time [24]. This structure is shown in **Figure 2**.

Membership functions use the fuzzification and defuzzification steps of a fuzzy logic system. Gaussian, triangular and trapezoidal are most usual membership functions [25]. The Mamdani and Sugeno methods are two types of fuzzy inference systems that can be performed in the fuzzy logic. In the Mamdani method subsequences are fuzzy sets and the final output is based on defuzzification of the overall fuzzy output. In the Sugeno method, the results are real numbers, which they are linear or constant [26]. In this study the Sugeno fuzzy model was used.

**ANFIS architecture**

For a Sugeno fuzzy model a typical rule set with two fuzzy if-then rules can be demonstrated as:

Rule 1: If X is V_{1} and Y is D_{1}, then F_{1}=p_{1}x q_{1}y+r_{1}.

Rule 2: If X is V_{2} and Y is D_{2}, then F_{2}=p_{2}x + q_{2}y+r_{2}.

ANFIS structure based on Sugeno model is shown in **Figure 3** [27-30].

It has five layers as follows:

Layer 1: this layer is input nodes, which every node is an adaptive node.

Layer 2: rule layer, which multiplies the incoming signals and outputs the product.

Layer 3: average nodes, which they have a normalization rule.

Layer 4: consequent nodes. This layer provides consequent parameters.

Layer 5: output nodes. This layer calculates the overall output of all incoming signals.

In this model, all output membership functions are the same type and they must be either linear or constant. Also, the number of output membership functions and the number of rules is equal. The ANFIS structure is adjusted automatically by the hybrid algorithm that contains a combination of the back propagation and least-squares methods [31,32]. This algorithm optimizes the premise parameters which define the shape of the membership functions [33].

#### Experimental

**Instruments and software**

Scanning electron microscopy (SEM, philips XL 30 S-FG) for examining the morphologies of the synthesized particles and flame atomic absorption spectrophotometer (Varan AA 240 Model) were used for measuring the amount of Pb(II) ions removal from the aqueous solution. The results obtained using Microsoft Excel 2010, ANN, FIS and ANFIS Toolbox of MATLAB 8.6.

**Chemicals and reagents**

Commercial nylon 6 was obtained from the DMS Chemical Company, with averaged number molecular weight 67 Kg/mol. In addition, natural graphite powder was supplied from Sigma Aldrich. Ferric chloride (FeCl_{3}·6H_{2}O), ferrous chloride (FeCl_{2}.7H_{2}O), lead nitrate Pb (NO_{3})_{2}, sodium borohydride (NaBH_{4}), sulfuric acid 98% (H_{2}SO_{4}), hydrogen peroxide 25% (H_{2}O_{2}), potassium permanganate (KMnO_{4}), sodium nitrate (NaNO_{3}), formic acid and other reagents such as HCl, NaOH and Ethanol were obtained from Merck Germany. Stock solutions of Pb(II) ions were prepared from lead nitrate in deionized water.

#### Procedure

**Synthesis of nano-adsorbent**

Graphene oxide was synthesized from natural graphite by Hummers and offeman’s method [19]. Then, the magnetic graphene oxide (MGO) was synthesized by co-precipitation of FeCl_{3} and FeCl_{2}. Finally, the magnetic graphene of nylon 6 (MGN_{6}) was prepared by reduction reaction using sodium borohydride. The pH of the solution, initial lead(II) ions concentration and adsorbent dosage were determined by using optimization based on Taguchi method are given in **Table 1**. The amount of lead(II) ions removal from the aqueous solution was evaluated using a flame atomic absorption spectrophotometer (Varan AA 240 Model).

Run | pH | Dose (g) | Concentration (mg/L) | Recovery (%) |
---|---|---|---|---|

1 | 5 | 0.01 | 2 | 20 |

2 | 5 | 0.02 | 4 | 55 |

3 | 5 | 0.05 | 6 | 60 |

4 | 7 | 0.01 | 2 | 88 |

5 | 7 | 0.02 | 4 | 92 |

6 | 7 | 0.05 | 6 | 93 |

7 | 9 | 0.01 | 2 | 96 |

8 | 9 | 0.02 | 4 | 89 |

9 | 9 | 0.05 | 6 | 91 |

**Table 1:** Experimental data obtained using Taguchi method.

#### Results and Discussion

**Scanning electron microscopy (SEM)**

The SEM images of the nylon 6 (N_{6}) and nanocomposite (MGN_{6}) are shown in **Figure 4a** and **4b** respectively. **Figure 4a** represents that nylon-6 exhibits a rough and lumpy surface with some uneven and irregular morphology. The morphology of MGN_{6} in **Figure 4b** shows magnetic nanoparticles with a size ranging from 33 to 74 nm, evenly distributed over the graphene/nylon6 composite.

**Artificial neural network model **

The plots of mean squer error (MSE) versus the number of neurons and layers for Levenberg-Marquardt, Conjugate and Bayesian algorithms are shown in **Figure 5**.

The equation of MSE is defined as follows:

Where* y _{pred}* is the predicted value and Type equation here

*y*is the actual value.In Levenberg-Marquardt algorithm, the least error is related to hidden layer 2. In Conjugate algorithm, hidden layer 10 has less error. In the Bayesian algorithm, the hidden layer number 3 has the least error. The predicted values against the actual values of validation and training set for the best neurons and layers are shown in

_{obs}**Figure 6**. As can be seen in

**Figure 6**, correlation coefficients (R2) clearly show better predictive ability of the algorithms. The statistical results obtained by different training and validation algorithm sets are reported in

**Table 2**.

Sample Number | Actual | Recoveries (%) of Levenberg-Marquardt | Recoveries (%) of Conjugate | Recoveries (%) of Bayesian |
---|---|---|---|---|

Training Data | ||||

1 | 20 | 102.85 | 100 | 100 |

2 | 55 | 100 | 98.18 | 88.34 |

3 | 88 | 99.99 | 101.3 | 100 |

4 | 89 | 101.14 | 100 | 100 |

5 | 91 | 101.45 | 98.9 | 100 |

Mean Recovery | 101.09 | 99.67 | 97.66 | |

RMSE | 0.79 | 0.815 | 2.86 | |

Validation Data | ||||

18 | 92 | 99.99 | 103.26 | 100 |

19 | 93 | 99.87 | 100 | 100 |

20 | 96 | 99.99 | 101.25 | 100.4 |

21 | 60 | 100.01 | 102.73 | 100 |

Mean recovery | 99.96 | 101.81 | 100.1 | |

RMSE | 0.006 | 1.81 | 0.195 |

**Table 2:** Recovery data obtained by application of the Levenberg- Marquardt, Conjugate and Bayesian algorithms.

As it can be observed in **Table 2**, among the three algorithms the Root Mean Square Error (RMSE) of the Levenberg-Marquardt algorithm is lower than the Conjugate and Bayesian algorithms. The RMSE is defined by the following equation [34]:

Where *y _{pred}* is the predicted value in the sample,

*y*is the actual value of the sample, N is the number of samples and e is the mean of the actual value.

_{obs}**Fuzzy inference system model**

There are three inputs containing the concentration of the solution (c), amount of absorbent (mg) and pH of the solution (pH).

In the proposed model, 9 rules are established, and these rules can be expressed in the if-then form. For example:

Rule 1: If input 1=4 and input 2=0.02 and input 3=7 then output=89

Rule 2: If input 1=4 and input 2=0.02 and input 3=7 then output=100

**Table 3** shows fuzzy inference rules for the multivariate systems. **Figure 7** shows the rule viewer of FIS model.

Serial No | C | mg | pH | Recovery (%) |
---|---|---|---|---|

1 | 2 | 0.01 | 5 | 20 |

2 | 4 | 0.02 | 5 | 55 |

3 | 6 | 0.05 | 5 | 60 |

4 | 2 | 0.01 | 7 | 88 |

5 | 4 | 0.02 | 7 | 92 |

6 | 6 | 0.05 | 7 | 93 |

7 | 2 | 0.01 | 9 | 96 |

8 | 4 | 0.02 | 9 | 89 |

9 | 6 | 0.05 | 9 | 91 |

**Table 3:** Fuzzy rules of FIS and ANFIS models.

In this model, the relationship between the concentration (c), adsorption dosage (mg), and percent recovery is presented in **Figure 8**.

Also predicted values against actual values for FIS model is shown in **Figure 9**.

In **Figure 9** the R^{2} value is 0.9975 that indicates better ability of the prediction of this model. Mean recovery (%) and RMSE of the proposed method calculated are summarized in **Table 4**.

FIS | ||
---|---|---|

Actual | Predicted | Recoveries (%) |

20 | 20 | 100 |

55 | 55 | 100 |

60 | 60 | 100 |

88 | 88 | 100 |

92 | 92 | 100 |

93 | 93 | 100 |

96 | 96 | 100 |

89 | 96 | 104.49 |

91 | 91 | 100 |

Mean | 100.49 | |

RMSE | 1.33 |

**Table 4:** Recovery and RMSE obtained by application of the FIS model.

The results in **Table 4** show that this model is fairly good for predicting the mean recovery (%) and RMSE values. As can be observed in **Table 4**, the value of RMSE was found to be 1.33.

**Adaptive neuro-fuzzy inference system model**

This model developed using 9 numbers of membership functions of type 'gaussmf ' with 9 If-then rules. The Rules based on Neuro- fuzzy for multivariable are displayed in **Figures 10** and **11** represents the structure of ANFIS model which has three inputs and one output.

Each input is connected to 9 membership functions. This model has 9 rules. Each rule is connected with one membership function and these membership functions produced the output. Several types of membership functions were used in proposed models and the gaussmf shape selected based on the statistical parameters. Hybrid learning algorithm combining the least-squares method and the gradient descent method are used in ANFIS model. The developed model was applied by using 10 epochs and error of tolerance was found to be zero.

The ANFIS graphs for describing the relationship between inputs and recovery factor are presented in **Figure 12**.

The plots of the predicted versus reference values are shown in **Figure 13**.

The correlation coefficient (R^{2}) was 0.9998 which is close to 1 that shows the suitability of the ANFIS model. The statistical results of Adaptive Neuro-Fuzzy Inference System consist of mean recovery (%) are summarized in **Table 5**. **Table 5** shows that ANFIS model has the higher ability of prediction because of lowest RMSE value.

ANFIS | ||
---|---|---|

Actual | Predicted | Recoveries (%) |

20 | 20 | 100 |

55 | 55 | 100 |

60 | 60 | 100 |

88 | 88 | 100 |

92 | 92 | 100 |

93 | 93 | 100 |

96 | 96 | 100 |

89 | 90 | 101.12 |

91 | 91 | 100 |

Mean | 100.12 | |

RMSE | 0.333 |

**Table 5:** Recovery and RMSE obtained by application of the ANFIS model.

**The analysis of variance**

To investigate the existence of the significant differences between the reported variances of recovery for removing lead(II) ions the results of ANN, FIS and ANFIS models were compared with each other using ANOVA test. The obtained results are presented in **Table 6**. Due to the calculated F value which is less than the F critical, there is no significant differences between the variances at the 95% confidence level.

#### Conclusion

In this study, the prediction ability of artificial intelligence models for the removal of lead(II) ions from the aqueous solution using the magnetic nano absorbent were carried out. The artificial neural network with three different algorithms (Levenberg-Marquardt, conjugate, and Bayesian), fuzzy inference system and adaptive neuro fuzzy inference system, were used. The results obtained from the neural network with the Levenberg-Marquardt algorithm showed better performance than the other two algorithms. Moreover, according to the results obtained using FIS and ANFIS models, it can be suggested that the prediction ability of the ANFIS is better than the other model. The results from proposed models compared with the analysis of variance (ANOVA) test at the 95% confidence level showed that there were no significant differences between these models.

#### References

- Fan M, Li T, Hu J, Cao R, Wei X, et al. (2017) Artificial Neural Network Modeling and Genetic Algorithm Optimization for Cadmium Removal from Aqueous Solutions by Reduced Graphene Oxide-Supported Nanoscale Zero-Valent Iron (nZVI/rGO) Composites, Materials. Materials 10: 544.
- Rijwani P, Jain S (2016) Enhanced Software Effort Estimation using Multi Layered Feed Forward Artificial Neural Network Technique. Procedia Computer Science 89: 3070-312.
- Walia N, Singh H, Sharma A (2015) ANFIS: Adaptive Neuro-Fuzzy Inference System- A Survey. International Journal of Computer Applications 123: 32-38.
- Tanhaei B, Esfandyari M, Ayati A, Sillanpa M (2017) Neuro-fuzzy modeling to adsorptive performance of magnetic chitosan nanocomposite. J Nanostruct Chem 7: 29-36.
- Sekulic Z, Antanasijevic D, Stevanovic S, Trivunac K (2017) Application of artificial neural networks for estimating Cd, Zn, Pb removal efficiency from wastewater using complexationmicrofiltration process. Int J Environ Sci Technol 14: 1383-1396.
- Chang YF, Wenb JF, Cai JF, Xiao-ying W, Yang L (2012) An investigation and pathological analysis of two fatal cases of cadmium poisoning. Forensic Science International 220: e5-e8.
- Singha B, Bar N, Kumar DS (2015) The use of artificial neural network (ANN) for modeling of Pb(II) adsorption in batch process. Journal of Molecular Liquids 211: 228-232.
- Sadrzadeh M, Mohammadi T, Ivakpour J, Kasiri N (2009) Neural network modeling of Pb2+ removal from wastewater using electrodialysis. Chemical Engineering and Processing 48: 1371-1381.
- Mitra T, Singha B, Bar N, Kumar DS (2014) Removal of Pb (II) ions from aqueous solution using water hyacinth root by fixed-bed column and ANN modeling. Journal of Hazardous Materials 273: 94-103.
- Nwabanne JT, Igbokwe PhK (2012) Thermodynamic and Kinetic Behaviors of Lead (II) Adsorption on Activated Carbon Derived from Palmyra Palm Nut. International Journal of Applied Science and Technology 2: 245-254.
- Golkhah S, Mousavi ZH, Shirkhanloo H, Khaligh A (2017) Removal of Pb(II) and Cu(II) Ions from Aqueous Solutions by Cadmium Sulfide Nanoparticles. Int J Nanosci Nanotechnol 13: 105-117.
- Ebadi M, Shagholani H, Jahangiri H (2016) High Efficient Nanocomposite for Removal of Heavy Metals (Hg 2+ and Pb 2+) from Aqueous Solution. J Nanostruct 6: 23-27.
- Venkateswarlu S, Natesh KB, Prathima B, Subbarao Y, Jyothi VVN (2014) A novel green synthesis of Fe3O4 magnetic nanorods using Punica Granatum rind extract and its application for removal of Pb(II) from aqueous environment. Arabian Journal of Chemistry pp: 1-9.
- Guijarro-Aldaco A, Hernandez-Montoya V, Bonilla PA, Montes-Moran MA, Mendoza-Castillo DI (2011) Improving the Adsorption of Heavy Metals from Water Using Commercial Carbons Modified with Egg Shell Wastes. Industrial & Engineering Chemistry Research 50: 9354-9362.
- Liu G, Teweldebrhan D , Balandin AA (2011) Tuning of Graphene Properties via Controlled Exposure to Electron Beams. IEEE Transactions on Nanotechnology 10: 865-870.
- Geim AK, Novoselov KS (2007) The rise of graphene Nature Mater 6: 183-191.
- Balandin AA, Ghosh S, Bao W, Calizo I, Teweldebrhan D, et al. (2008) Superior thermal conductivity of single-layer graphene. Nano Lett 8: 902-907.
- Ghosh S, Calizo I, Teweldebrhan D, Pokatilov EP, Nika DL, et al. (2008) Extremely high thermal conductivity of graphene: Prospects for thermal management applications in nanoelectronic circuits. Appl Phys Lett 92: 151-911.
- Hummers Jr WS, Offeman RE (1958) Preparation of Graphitic Oxide. J Am Chem Soc 80: 1339-1339
- Chakraborty R, Maitra A (2016) Retrieval of atmospheric properties with radiometric measurements using neural network. Atmospheric Research 181: 124-132.
- Zupan J, Gasteiger J (1993) Neural Networks for Chemists: An Introduction. John Wiley & Sons Inc, New York.
- Bose NK, Liang P (1996) Neural networks, fundamentals. McGraw-Hill, New York.
- Panchal R, Suryanarayana TMV, Parekh FP (2014) Adaptive Neuro-Fuzzy Inference System for rainfall-runoff modeling. Journal of Engineering Research and Applications. 4: 202-206.
- Nayaka PC, Sudheerb KP, Ranganc DM, Ramasastri KS (2004) A neuro-fuzzy computing technique for modeling hydrological time series. Journal of Hydrology 291: 52-66.
- Mendel J (1995) Fuzzy logic systems for engineering: a tutorial. Proceedings of the IEEE 83: 345-377.
- Kokkinos K, Loukas A, Samaras N (2015) development and evaluation of an adaptive neuro fuzzy inference system for the calculation of soil water recharge in a watershed. European Scientific Journal pp: 105-116.
- Roger JJSH (1995) Neuro-Fuzzy Modeling and Control. Proceedings of The IEEE 83: 378-406.
- Tahmasebi P, Hezarkhani A (2010) Application of Adaptive Neuro-Fuzzy Inference System for Grade Estimation; Case Study, Sarcheshmeh Porphyry Copper Deposit, Kerman, Iran. Australian Journal of Basic and Applied Sciences 4: 408-420.
- Singh HR, Biswas SK, Purkayastha B (2017) A neuro-fuzzy classification technique using dynamic clustering and GSS rule generation. Journal of Computational and Applied Mathematics 309: 683-694.
- Akbari HHR, Sohrabi MR, Abdolmaleki P (2004) Adaptive Neuro-Fuzzy Inference System (ANFIS) Applied for Spectrophotometric Determination of Fluoxetine and Sertraline in Pharmaceutical Formulations and Biological Fluid. Iranian Journal of Pharmaceutical Sciences 10: 19-34.
- Zahedi AS, Sadi M (2013) Application of adaptive neuro-fuzzy inference system for the prediction of the yield distribution of the main products in the steam cracking of atmospheric gasoil. Journal of the Taiwan Institute of Chemical Engineers 44: 365-376.
- Premkumar K, Manikandan BV (2014) Adaptive Neuro-Fuzzy Inference System based speed controller for brushless DC motor. Neurocomputing 138: 260-270.
- Walia N, Singh H, Sharma A (2015) ANFIS: Adaptive Neuro-Fuzzy Inference System- A Survey. International Journal of Computer Applications 123: 32-38.
- Guan J, Zurada J, Levitan AS (2008) An adaptive Neuro-Fuzzy Inference System based approach to real estate property assessment. JRER 30: 395-421.

Citation: Afroozeh M, Sohrabi MR, Davallo M, Mirnezami SY, Motiee F, et al. (2018) Application of Artificial Neural Network, Fuzzy Inference System and Adaptive Neuro-Fuzzy Inference System to Predict the Removal of Pb(II) Ions from the Aqueous Solution by Using Magnetic Graphene/Nylon 6. Chem Sci J 9: 185. DOI: 10.4172/2150-3494.1000185

Copyright: © 2018 Afroozeh M, et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.