Determination of Boiling Range of Xylene Mixed in PX Device Using Artificial Neural Networks
DDetermination of Boiling Range of Xylene Mixed in PX Device UsingArtificial Neural Networks
Ting Zhu, Yuxuan Zhu, Hong Yang
College of Software Engineering, SichuanUniversity, Chengdu, Sichuan 610064, ChinaE-mail: [email protected]: [email protected]: [email protected]
Hao Li * College of Chemistry, Sichuan University,Chengdu, Sichuan 610064, ChinaE-mail: [email protected]
Abstract -Determination of boiling range of xylene mixed inPX device is currently a crucial topic in the practicalapplications because of the recent disputes of PX project inChina. In our study, instead of determining the boilingrange of xylene mixed by traditional approach in laboratoryor industry, we successfully established two Artificial NeuralNetworks (ANNs) models to determine the initial boilingpoint and final boiling point respectively. Results show thatthe Multilayer Feedforward Neural Networks (MLFN)model with 7 nodes (MLFN-7) is the best model to determinethe initial boiling point of xylene mixed, with the RMS error0.18; while the MLFN model with 4 nodes (MLFN-4) is thebest model to determine the final boiling point of xylenemixed, with the RMS error 0.75. The training and testingprocesses both indicate that the models we developed arerobust and precise. Our research can effectively avoid thedamage of the PX device to human body and environment.
Keywords -boiling range; determination; xylene mixed; PXdevice; artificial neural networks; multilayer feedforwardneural networks
I. I
NTRODUCTION
A. Background
Currently, the analytical approach of xylene mixed inPX device is mainly the distillation method [1]. It's aclassical approach of distillation analysis, which isconvenient and precise. Therefore, distillation method iswidely used in analyzing the correlate matter. However,aromatic hydrocarbon samples are harmful to human body[2-3], and impact the environment seriously [4-5]. Forresolving this problem, previous researches aimed at usingmathematical approaches to analyze the boiling range of correlate chemical products. L.P. Wang and herco-workers [6] have successfully proved that the boilingrange of the relevant chemical products. W. Cao and hisco-workers [7] has discovered the correlation approachesof gas phase chromatography distillation. As for themathematical modeling, Y.Y. Ou [8] developed a multiplelinear regression model to calculate the boiling range ofxylene mixed from PX device. Previous researchesindicate that the correlation between the different productcomponents can be described by mathematical method,such as linear prediction. In our study, we tried to usenon-linear function to plot this relationship. B. Principle of Artificial Neural Networks
Non-linear model is a powerful tool to describe thephysical and biological phenomenon [9-10]. ArtificialNeural Networks (ANN) [11] is one of the most usefulmodels to describe the non-linear relationships. It is amathematical or computational model which illustrates thepossibilities of improved understanding of neural systemsby applying concepts from an apparently disparate field,namely electric circuits and computer science [12].A neural network is made up of an interconnectedgroup of artificial neurons, using a connectionist approachto process information. Most of the time, an ANN modelis an adaptive system that has the ability of adaptingcontinuously to new data and learning from experience[13-15]. Besides, the system changes its structure based onexternal or internal information that flows through thenetwork during the learning phase. They are usually usedto abstract essential information from data or modelcomplex relationships between inputs and outputs. igure 1. A schematic view of artificial neural network structure.
The main structure of the artificial neural network(ANN) consists of the input layer and the output layer(Figure 1). The input variables was introduced to thenetwork by the input layer [16], while the networkprovides the response variables with predictions whichstand for the output of the nodes in this certain layer.Apart from that, it includes the hidden layers. The typeand the complexity of the process or experimentationusually iteratively determine the optimal number of theneurons in the hidden layers [16].II. M
ODEL D EVELOPMENT
A. Training process
22 samples were used to develop the predictionmodels. Thereinto, 8 groups were used as the testing set,14 groups were used as the training set. Independentvariables are made up of different main constituents ofchemical product, including nonaromatic substance,toluene, ethyl benzene, p-xylene, m-xylene, isopropylbenzene, o-xylene, n-proplbenzene and C9 aromatics.Dependent variables are the boiling range of xylene mixedin PX device, containing initial boiling point and finalboiling point.To ensure the robustness of the models, two groupsof experiments were done respectively, one is thedetermination of initial boiling point, the other is thedetermination of final boiling point. To find out thesuitable and reasonable model of the determination, theRMS error of the testing is the main factors forconsideration. In each group of experiments, we utilizedlinear regression, General Regression Neural Networks[17-19] (GRNN) and Multilayer Feedforward NeuralNetworks [19-21] (MLFN) to the establishment. Thereinto,MLFN models were experimented with different nodes, sothat, the best situation of nodes could be found out.Training results are shown as follows:
TABLE I. R
ESULTS OF DIFFERENT MATHEMATICAL MODEL INDETERMINING INITIAL BOILING POINT . ANN model Trained samples Tested samples RMS error
Linear predictionGRNN 1414 88 0.660.23MLFN 2 Nodes 14 8 0.37MLFN 3 Nodes 14 8 0.37MLFN 4 Nodes 14 8 0.55MLFN 5 Nodes 14 8 0.54MLFN 6 Nodes 14 8 0.25MLFN 7 Nodes 14 8 0.18MLFN 8 Nodes 14 8 0.32MLFN 9 Nodes 14 8 0.21MLFN 10 Nodes 14 8 0.21MLFN 11 Nodes 14 8 0.39MLFN 12 Nodes 14 8 0.20MLFN 13 Nodes 14 8 0.52MLFN 14 Nodes 14 8 0.50MLFN 15 Nodes 14 8 0.56MLFN 16 NodesMLFN 17 NodesMLFN 18 NodesMLFN 19 NodesMLFN 20 NodesMLFN 21 NodesMLFN 22 NodesMLFN 23 NodesMLFN 24 NodesMLFN 25 NodesMLFN 26 NodesMLFN 27 NodesMLFN 28 NodesMLFN 29 NodesMLFN 30 Nodes 141414141414141414141414141414 888888888888888 4.012.422.242.632.490.982.191.552.181.702.492.312.596.683.17
Table 1 implies that the MLFN model with 7 nodes(MLFN-7) is the best model during the training process,with the RMS error 0.18.Results of different mathematical model indetermining final boiling point are shown in Table 2:
TABLE II. R
ESULTS OF DIFFERENT MATHEMATICAL MODEL INDETERMINING FINAL BOILING POINT . NN model Trained samples Tested samples RMS error
Linear predictionGRNN 1414 88 6.191.38MLFN 2 Nodes 14 8 2.13MLFN 3 Nodes 14 8 1.76MLFN 4 Nodes 14 8 0.75MLFN 5 Nodes 14 8 1.89MLFN 6 Nodes 14 8 2.36MLFN 7 Nodes 14 8 2.29MLFN 8 Nodes 14 8 1.53MLFN 9 Nodes 14 8 1.04MLFN 10 Nodes 14 8 2.15MLFN 11 Nodes 14 8 1830.25MLFN 12 Nodes 14 8 1.19MLFN 13 Nodes 14 8 1.05MLFN 14 Nodes 14 8 1.52MLFN 15 Nodes 14 8 11.26MLFN 16 NodesMLFN 17 NodesMLFN 18 NodesMLFN 19 NodesMLFN 20 NodesMLFN 21 NodesMLFN 22 NodesMLFN 23 NodesMLFN 24 NodesMLFN 25 NodesMLFN 26 NodesMLFN 27 NodesMLFN 28 NodesMLFN 29 NodesMLFN 30 Nodes 141414141414141414141414141414 888888888888888 9.5816.305.617.906.3710.785.1519.959.685.8015.316.489.595.607.03
According to the results presented in table 2, MLFNmodel with 4 nodes (MLFN-4) possesses the lowest RMSerror (0.75), which is obviously much lower than othermodels. Therefore, we considered that the MLFN-4 modelis a reasonable model in determining final boiling point.III. R
ESULTS AND D ISCUSSION A. Determination of initial boiling point
To describe the training process of initial boilingpoint, figure 2 is used to present the accuracy of theMLFN-7 model, which is shown as follows:
Figure 2. Comparison between predicted values and actual values ofinitial boiling point during training process.
Figure 2 indicates that using MLFN-7 model candescribe the relationship between initial boiling point andthe product components primely. The good fitting resultshow that the training process is stable and robust.
B. Determination of final boiling point
Uniformly, figure 3 is used to depict the trainingprocess of the determination of final boiling point, whichis shown as follows:
Figure 3. Comparison between predicted values and actual values offinal boiling point during training process.
Figure 3 depicts the determination process of finalboiling point using MLFN-4 model. Results also show thatthe MLFN-4 model can fit the relationship between finalboiling point and product components admirably,indicating that the training process is robust and precise.
C. Discussion
Figure 2 to 3 depict the robustness of the ANNmodel in determining the boiling range of xylene mixed inPX device, indicating that the training process is accurateand reasonable. It's axiomatic that the boiling range ofxylene mixed in PX device can be determined by artificialneural networks model without the complex operation ofractical experiments. In addition, it's worth mentioningthat Multilayer Feedforward Neural Networks (MLFN) isparticularly accurate in calculating the boiling range indifferent nodes of networks.IV. C
ONCLUSION
Determination of boiling range for xylene mixed inPX device is currently a crucial topic in the practicalapplications because of the recent disputes of PX projectin China. In our study, we successfully established twoartificial neural networks models to determine the initialboiling point and final boiling point respectively. Resultsshow that the Multilayer Feedforward Neural Networks(MLFN) model with 7 nodes (MLFN-7) is the best modelto determine the initial boiling point of xylene mixed, withthe RMS error 0.18, while the MLFN model with 4 nodes(MLFN-4) is the best model to determine the final boilingpoint of xylene mixed, with the RMS error 0.75. Thetraining and testing processes both indicate that the modelswe developed are robust and precise. Our research canavoid the damage of the PX device to human body andenvironment. V. A
CKNOWLEDGEMENTS
The corresponding author of this work is Mr. Hao Li,College of Chemistry, Sichuan University, Chengdu,610064, China. E-mail: [email protected]
EFERENCES[1] E. W. Dean, D. D. Stark, "A Convenient Method for theDetermination of Water in Petroleum and Other OrganicEmulsions," Industrial & Engineering Chemistry, vol. 12,pp.486-490. May 1920.[2] M. Van den Berg, L. Birnbaum, A. T. Bosveld, et al, "Toxicequivalency factors (TEFs) for PCBs, PCDDs, PCDFs for humansand wildlife," Environmental health perspectives, vol. 106, pp.775. December 1998.[3] A. Brouwer, U. G. Ahlborg, M. Van den Berg, et al, "Functionalaspects of developmental toxicity of polyhalogenated aromatichydrocarbons in experimental animals and human infants".European Journal of Pharmacology: Environmental Toxicology andPharmacology, vol. 293, pp. 1-40.Janunary 1995.[4] S. K. Samanta, O. V. Singh, R. K. Jain, "Polycyclic aromatichydrocarbons: environmental pollution and bioremediation,"TRENDS in Biotechnology," vol. 20,pp. 243-248, June 2002.[5] S. R. Wild, K. C. Jones, "Polynuclear aromatic hydrocarbons inthe United Kingdom environment: a preliminary source inventory and budget," Environmental Pollution, vol. 88, pp. 91-108, January1995.[6] L. P. Wang , R. L. Wang , X. S. Meng , Y. Q. Liu , J. D. Liu , "TheSignificance of Relativity Between the Beneze Component and theDistillation Range in the Analysis Application (Chinese Article),"Science & Technology in Chemical Industry , vol. 7, pp. 58-61,February 1999.[7] W. Cao , A. J. Chen , Y. K. Zhao ,,