CHEM. RES. CHINESE UNIVERSITIES 2010, 26(6), 899—904

Application of Back-propagation Artificial Neural Network in Speciation of Cadmium WANG Lin-lin1, ZHANG Jie2, LIU Hai-yan1, ZHANG Hai-tao1, WANG Hong-yan1, YANG Xiu-rong2 and WANG Ying-hua1* 1. College of Chemistry, Jilin University, Changchun 130012, P. R. China; 2. State Key Laboratory of Electroanalytical Chemistry, Changchun Institute of Applied Chemistry, Chinese Academy of Sciences, Changchun 130022, P. R. China Abstract A method for predicting the five species contents of cadmium was developed by combining the back-propagation artificial neural network with graphite furnace atomic absorption spectrometry(BP-ANN-GF-AAS). Based on the strong learning function and the features of the information distributed storage of artificial neural network(ANN), a single ANN was constituted in which only one determination point of every sample was required. The exchangeable, carbonated, Fe-Mn oxidable, organic and residual species of cadmium for 20 kinds of soil samples from the two sections of Changchun(China) were determined by BP-ANN-GF-AAS. The detection limit of the method is 0.024 µg/L and the limit of quantification is 0.080 µg/L. t-Test indicates that there is not any systemic error of the results obtained by the Tessier sequential extraction graphite furnace atomic absorption spectrometry method(Tessier -GF-AAS) and BP-ANN-GF-AAS. Compared with those of the Tessier-GF-AAS, the prediction errors of BP-ANN-GF-AAS are less than 10%. The proposed method is fast, convenient, sensitive, and can eliminate the interference among various species. Keywords Artificial neural network(ANN); Speciation; Graphite furnace atomic absorption spectrometry(GF-AAS); Cadmium Article ID 1005-9040(2010)-06-899-06

1

Introduction

The heavy metals in the soil have the peculiarities of long accumulate time, complexity of existent state and toxicity[1]. Their bioavailability is harmful for the plants, human beings and environment[1] . Cadmium is one of harmful elements in soil[2,3]. It can harm human health via food chain. The superfluous cadmium in body can induce the function change of nephridium, liver, lung and bones. The harm of heavy metals is related to the total concentration of heavy metals and especially to the chemical species of heavy metals[4―6]. So, the researches of heavy metals species and their distribution in soil sample have important guiding significance to the researches on the effects of heavy metals to human health and environment conservation. The sequential extraction method is an effective chemical operational procedure for species analysis[7—9]. So far, Tessier sequential extraction has been used in the different species separation of heavy

metals as the pretreatment process of traditional speciation analytical method[9,10]. Usually, five chemical species are separated by Tessier sequential extraction, namely exchangeable, carbonated, Fe-Mn oxidable, organic and residual species. The exchangeable species is the primary one which can harm plants and a human being via food chain. The carbonate species and Fe-Mn oxide species are potentially dangerous species for plants and a human being. If the pH value of soil environmental is changed, the two species will be released. The residual species is the most stable species and not easily be released to the environment. The separation of each species needs differently selective reagent in the sequential extraction method. However, it is difficult to choose the highly selective extractants and the interference among various species is unavoidable in the sequential extraction procedure. Another problem is that the sequential extraction procedure will finish in a week approximately. So, it is necessary to develop a fast and sensitive method for

——————————— *Corresponding author. E-mail: [email protected] Received February 1, 2010; accepted March 5, 2010. Supported by the Fund of the State Key Laboratory of Electroanalytical Chemistry, Changchun Institute of Applied Chemistry, Chinese Academy of Sciences.

900

CHEM. RES. CHINESE UNIVERSITIES

determining various species of heavy metals in soil simultaneously. Back-propagation artificial neural network (BP-ANN) is a kind of chemometrics methods, it has multiple functions of non-linear mapping, high degree fault-tolerance, self-organization and self-learning. It has been applied in many fields[11—14]. Lee et al.[15] predicted the short-term storm surge in the Taichung harbor of Taiwan by BP-ANN. Sang et al.[16] detected element content in coal using BP-ANN method. Xu et al.[17] optimized the parameters of heat treatment technique of high-vanadium high-speed steel using BP-ANN. Fu et al.[18] modeled the adsorption course of bovine serum albumin on porous polyethylene membrane by means of BP-ANN. Graphite furnace atomic absorption spectrometry(GF-AAS) is one of the most effective techniques for the determination of trace cadmium due to its high selectivity and sensitivity[19—21]. Herein lies a method of speciation analysis for cadmium, i.e., BP-ANNGF-AAS. The method can be used to predict synchronously the five species contents of cadmium in soil samples.

2 2.1

Experimental Instruments and Reagents

A model AA-6800 atomic absorption spectrometer from Shimadzu(Japan) was used, which was equipped with a graphite furnace, deuterium background corrector and an ASC-6100 Autosampler. A hollow cathode lamp(Yongning, China) of cadmium (λ=228.8 nm) was employed as radiation source and it was operated at 8.0 mA. The instrumental parameters were adjusted according to the manufacturer’s recommendations. Argon 99.995%(volume fraction) was used as the purging gas. An electric heater was used to digest the samples. The reagents used in this study were of extra-pure grade. Distilled water was used for all the experiments. All the Teflon containers and glassware were cleaned by soaking in HNO3(10%, volume fraction) for at least 12 h and were rinsed with secondary distilled water prior to use. The standard stock solution of cadmium was a laboratory-made stock solution(1000 mg/L cadmium in 2.0% HNO3). 2.2

Sampling and Sample Treatment Soil samples of 30 kinds were collected from the

Vol.26

two locations of Changchun, the square of Nanhu and the park of Nanhu. The sampling locations were chosen by systemic distribution sampling method(i.e., a point every ten meters at the sampling locations). The range of the depth of sampling was 0—30 cm and the weight of every sample was about 1 kg. All the soil samples collected were dried and milled at room temperature, then passed through a 100-mesh nylon fiber sieve. The soil samples obtained were stored in polypropylene bags for analysis. 2.3

Preparation of Work Solution

A series of cadmium standard solutions was prepared from the stock solution of cadmium(1000 mg/L) and the concentrations were 0, 0.05, 0.1, 0.2, 0.4, 0.8, 1.0, 2.0, 3.0, 4.0 and 5.0 µg/L, to which 10 μL of chemical modifier[100 g/L (NH4)2HPO4] were added respectively and diluted to a volume of 100 mL with 2%(volume fraction) HNO3, respectively. 2.4

Experiment Process

Soil sample of 0.2 g was weighed accurately and put into a 100 mL Teflon beaker and 10 mL of HNO3 was added into it. The beaker with cover was heated for 1 h on the electric heater. After cooling down, 10 mL of HF was added into the beaker and heated to boiling for 20 min. After cooling the sample, 5 mL of H2O2 was added to it and the sample was heated until the solid residue was dried approximately, to which 3 mL of H2O2 was then added and the sample was heated until the solid residue was dried approximately, and it was heated again until the residue of sample was offwhite. The residue was diverted into a volumetric flask. After 5 μL of 100 g/L (NH4)2HPO4 was added to the volumetric flask, the sample was diluted to 50 mL with 2% HNO3. Then 10 μL of the sample solution was injected into the graphite cuvette of GF-AAS for determining the total content of cadmium. The measurement was repeated three times for every sample. 2.5

Tessier-GF-AAS Method

The Tessier sequential extraction is an effectively chemical procedure for separating heavy metals speciation. By this method, five species contents of an unknown sample could be obtained, which will be used to train the network parameter. The procedure of Tessier-GF-AAS is shown as

No.6

WANG Lin-lin et al.

follows. Step 1(exchangeable species): sample of 1.0 g was put into a 50 mL centirfugal tube. Then 8 mL of 1.0 mol/L MgCl2(pH=7 adjusted by HCl) was added to the tube. It was shaken continuously for 1.5 h at (25±2) °C and centrifugalized for 10 min at 5000 r/min. The upper solution was transferred into a breaker. The residue was washed twice with 5 mL of secondary distilled water and centrifugalized for 10 min at 5000 r/min. The upper clear solution obtained was merged into the first one and they were digested with 5 mL of HNO3(68%, volume fraction) and 5 mL of H2O2(30%, volume fraction). After 5 μL of 100 g/L (NH4)2HPO4 was added to the solution, the solution was diluted to 50 mL with 2% (volume fraction) HNO3 for the determination of the content of exchangeable species by GF-AAS. Step 2(carbonates species): 8 mL of 1.0 mol/L NaAc (pH=5.0 adjusted by HAc) was added into the residue obtained from Step 1. It was shaken continuously for 3 h at (25±2) °C and centrifugalized for 10 min at 5000 r/min. The following procedure was the same as the procedure in Step 1. The result obtained in this step was the content of carbonate species. Step 3(Fe-Mn oxide species): 20 mL of 0.04 mol/L NH2OH·HCl(pH=2.0 adjusted by HAc) was added into the residue obtained from Step 2. It was shaken intermittently for 3 h at (96±2) °C and centrifugalized 10 min at 5000 r/min. The followed procedure was the same as the procedure in Step 1. The result obtained in this step was the content of Fe-Mn oxides species . Step 4(organic species): 3 mL of 0.02 mol/L of HNO3 and 5 mL 30% H2O2(pH=2.0 adjusted by HNO3) were added into the residue obtained from Step 3. It was shaken intermittently for 1.5 h at (85±2) °C and 3 mL of 30% H2O2(pH=2.0 adjusted by HNO3) was added to it. After cooling, 5 mL of 3.2 mol/L NH4Ac in 20% HNO3 was added to the solution. The mixture was shaken intermittently for 0.5 h at (25±2) °C and centrifugalized for 10 min at 5000 r/min. The followed procedure was the same as the procedure in Step 1. The result obtained in this step was the content of organic species. Step 5(residual species): the residue obtained from Step 4 was digested with 10 mL of HNO3(68%), 5 mL of HF(30%) and 5 mL of H2O2(30%). After 5 μL of 100 g/L (NH4)2HPO4 was added to the solution, the

901

solution was diluted to 50 mL with 2% HNO3 for the determination of the content of residual species by GF-AAS.

3

Results and Discussion

3.1 Calibration Curve and Detection Limit for Cadmium The calibration curve about absorbance(A) vs. the concentration of the work solution of cadmium(ρ) was made. The regression equation is A=0.1647ρ+0.0658 and the correlation coefficient(r) is 0.9997. The linear range is 0.1000—4.000 µg/L. The detection limit(3σ, n=9) is 0.024 µg/L and the limit of qualification(10σ) is 0.080 µg/L. 3.2 Architecture of Back-propagation Artificial Neural Network BP-ANN model is made up of input layer, hidden layer and output layer. In general, there are many nodes in the input layer, so every sample needs many determination points. Because only one determination point can be obtained by GF-AAS in this method, the BP-ANN of single input nodes was applied for determining the various species contents of cadmium. The architecture of BP-ANN model is shown in the Fig.1. It is made up of one input node, one hidden layer with many nodes and five output nodes. Though there is only one input node in this BP-ANN model, the BP-ANN will distribute the signal to every node of hidden layer. The nodes of hidden layer will redistribute the information to the output layer by transfer function(tansig, sigmoid tangent function). So, the multi-aim prediction of BP-ANN is obtained by the single point input information.

Fig. 1

BP-ANN of single point input

mij is the mass which is the connect parameter between nodes from two layers.

3.3

Establishing of Learning Set The establishment of learning set is the key of

902

CHEM. RES. CHINESE UNIVERSITIES

accurate prediction of the five species contents of cadmium in soil samples. For the establishment of the learning set, the contents of the five species of cadmium and the total content of cadmium in 40 soil samples from different sections were obtained by Tessier-GF-AAS. Every learning set was consisted of 10 samples at the different species concentrations of cadmium, which were selected discretionarily from 40 soil samples. The effect of learning set on predicting result was tested. Four learning sets were selected discretionarily and the five species of cadmium in four

Vol.26

samples(No.4 and No.6 soil samples were from the Nanhu square and No.7 and No.10 soil samples were from the Nanhu park) were forecasted for four learning sets respectively. The 80 forecasted results of four samples for the four learning sets are shown in Fig.2. It can be seen from Fig.2 that there is not great difference among the forecast contents for different learning sets. Compared with those of the results obtained by Tessier-GF-AAS, all the relative errors are less than 10%. It indicates that the learning set can be assembled optionally in this method.

Fig.2 Comparison of forecast results by the different learning sets Species: 1. Exchangeable; 2. Carbonate; 3. Fe-Mn oxide; 4. organic; 5. residual. No.4(A) and No.6(B) soil samples are from the Nanhu square; No.7(C) and No.10(D) soil samples are from the Nanhu park. represent four cadmium contents of the one sample, which were forecasted by four learning sets.

3.4 Selection of Hidden Layer Nodes Number Usually, the node number of input layer and output layer are decided by the determination method and the number of predicted aim. The node number of the hidden layer(n) needes to be selected. If the hidden nodes are too few, the convergence speed of the network is slow and the ability of identification decreases. Too many hidden nodes will result in long learning time and the worse ability of tolerance error. Generally, a rough scope of hidden nodes is defined according to the deviser’s experience at first. And then the optimal number of the nodes in hidden layer was decided by mean-squared error(MSE) of network. Fig.3 shows the relation between the number of the nodes of hidden layer and MSE. It can be seen from Fig.3

that the MSE of network is minimum at 27 nodes. So, 27 is selected as the number of hidden layer nodes.

Fig.3

3.5

Connection between the number of hidden layer nodes(n) and MSE

Selection of Learning Rate Learning rate, η, is a learning step, where

No.6

WANG Lin-lin et al.

η(0