Prediction of Influence of Doping of NaNO 3 on the Solid Phase Thermal Decomposition of Bitumen using neural networks

Transcription

1 Prediction of Influence of Doping of NaNO 3 on the Solid Phase Thermal Decomposition of Bitumen using neural networks Foad Qassemi Department of Civil and Survey Engineering Kerman Graduate University of Technology Highway Haft Bagh Alavi, Campus of Mahan Knowledge Iran zzfqassemi@yahoo.com Ali Reza Ahmadi Department of Civil and Survey Engineering International Center for Science, High Technology & Environmental Sciences Highway Haft Bagh Alavi, Campus of Mahan Knowledge Iran Iran.mahan@gmail.com Mostafa Alizadeh Department of Civil and Survey Engineering International Center for Science, High Technology & Environmental Sciences Highway Haft Bagh Alavi, Campus of Mahan Knowledge Iran Mostafa_alizadeh56@yahoo.com Ahmad Qassemi Department of Electrical and Computer Engineering University of Shiraz Zand Street Iran Ahamd.ghasemi@gmail.com Abstract: This study presents an application of artificial neural networks (ANN) for the prediction of the thermal stability of solidified NaNO3 salts in bitumen has been investigated for safety considerations in the field of solidification of radioactive waste. The thermal decomposition of bitumen in presence of NaNO3 in a temperature range C has been used. The proposed ANN model uses the NaNO3 concentration and temperature ( ). In this research data s are modelled and analyzed by Multi Layer Perceptron network (MLP). Predictions made by this method are compared with real data s. The results of parametric analyses were used to evaluate Mass loss of bitumen in a quite well manner. It can be concluded that MLP predictions are compatible with the actual data and they yield acceptable predictions. Key-Words: Neural networks, bitumen, solidification, thermal stability, MLP, thermal decomposition 1 Introduction The distribution and composition of organic matter in oil and bitumen containing rocks in deposits of different ages had been investigated. Several authors studied the thermal decomposition of the solidified salts in bitumen by using thermogravimetric method. Many studies on NaNO 3 and NaNO 2 incorporated in bitumen had been investigated [1]. The aim of the present work is to predict the thermal decomposition of bitumen in presence of different percentage of NaNO 3. Calculations of the kinetic parameters of the thermal decomposition of doping of NaNO 3 in bitumen were evaluated by using artificial neural network. In the context of knowledge based mathematical models, one needs to determine the relationship between input and output. In Artificial Neural Networks (ANNs) the transfer knowledge or the law, lies beyond the data; and it is established by processing experimental data through the network structure. Field ANNs are well suited for cases in which there are complex relationships between field variables, and the complexity is such that the ISBN:

2 process cannot be modeled properly. In such cases the classical modeling techniques have large errors in their predictions if incorrect or incomplete input is given, while ANNs' results remain acceptable. Very detailed information about the applications of traffic engineering can be found in the relevant literature. At this point, it is important to state out, one by one, the relevant important neural network applications in the pavement engineering area [1]. In a study by Ritchie, Kaseko, and Bavarian a system that integrates three artificial intelligence technologies: computer version, neural networks and knowledge-based system in addition to conventional algorithmic and modeling techniques were presented [2]. Kaseko and Ritchie used neural network models in image processing and pavement crack detection [3]. Gagarin, Flood, and Albrecht discuss the use of a radial-gaussianbased neural network for determining truck attributes such as axle loads, axle spacing and velocity from strainresponse readings taken from the bridges over which the truck is traveling [4]. Eldin and Senouci describe the use of a BP algorithm for condition rating of roadway pavements. They report very low average error when compared with a human expert determination [5]. Cal uses the backpropagation algorithm for soil classification based on three primary factors: plastic index, liquid limit, water capacity, and clay content [6]. Razaqpur, Abd El Halim, and Mohamed present a combined dynamic programming and Hopfield neural network bridgemanagement model for efficient allocation of a limited budget to bridge projects over a given period of time. The time dimension is modeled by dynamic programming, and the bridge network is simulated by the neural network [7]. Roberts and Attoh-Okine use a combination of supervised and self-organizing neural networks to predict the performance of pavements as defined by the International Roughness Index [8]. Tutumluer and Seyhan investigated neural network modeling of anisotropic aggregate behavior from repeated load triaxial tests [9]. The BP algorithm is used by Owusu-Ababia for predicting flexible pavement cracking and by Alsugair and Al-Qudrah to develop a pavementmanagement decision support system for selecting an appropriate maintenance and repair action for a damaged pavement [10, 11]. Kim and Kim used artificial neural networks for prediction of layer module from falling weight deflectometer (FWD) and surface wave measurements [12]. Shekharan studied the effect of noisy data on pavement performance prediction by an artificial neural network with genetic algorithm [13]. Attoh-Okine uses the self-organizing map or competitive unsupervised learning model of Kohonen for grouping of pavement condition variables (such as the thickness and age of pavement, average annual daily traffic, alligator cracking, wide cracking, potholing, and rut depth) to develop a model for evaluation of pavement conditions [14]. Lee and Lee presents an integrated neural network-based crack imaging system to classify crack types of digital pavement images which includes three types of neural networks: image-based neural network, histogram-based neural network and proximitybased neural network [15]. In an article by Mei, Gunaratne, Lu, and Dietrich it is presented a computer- based methodology with which one can estimate the actual depths of shallow, surfaceinitiated fatigue cracks in asphalt pavements based on rapid measurement of their surface characteristics [16]. Ceylan, Guclu, Tutumluer, and Thompson has investigated the use of artificial neural networks as pavement structural analysis tools for the rapid and accurate prediction of critical responses and deflection profiles of full-depth flexible pavements subjected to typical highway loadings [17]. Bosurgi and Trifiro has described a procedure that has been defined to make use of the available economic resources in the best way possible for resurfacing interventions on flexible pavements by using artificial neural networks and genetic algorithms [18]. Attoh-Okine in his paper, presents the application of functional equations and networks to incremental roughness prediction of flexible pavement [13]. 2 Artificial neural networks Neural networks are composed of simple elements operating in parallel. These elements are inspired by biological nervous systems. As in nature, the network function is determined largely by the connections between elements. A neural network can be trained to perform a particular function by adjusting the values of the connections (weights) between the elements. Commonly, neural networks are adjusted, or trained, so that a particular input leads to a specific target output. Such a situation is shown in Fig. 1. Here, the network is adjusted, based on a comparison of the output and the target, until the network output matches the target. Typically, many such input/target output pairs are used to train a network. Batch training of a network proceeds by making weight and bias changes based on an entire set (batch) of input vectors. Incremental training changes the weights and biases of a network ISBN:

3 as needed after presentation of each individual input vector. Incremental training is sometimes referred to as on line or adaptive training. Neural networks have been trained to perform complex functions in various fields of application including pattern recognition, identification, classification, speech, and vision and control systems. Today neural networks can be trained to solve problems that are difficult for conventional computers or human beings. The human brain, which can think, remember and solve the problem, is a very complicated system to understand. Many attempts have been made to simulate the human brain functions with computers. Neural networks are one of them. A neuron is the fundamental unit of the brain nervous system. It is a kind of processing element which can receive and combine the signals from other neurons by the way of dendrites. If the signal is strong enough, the neuron fibres produce the output signal along the axon. Each signal coming into the neuron passes through a synapse or synaptic junction [19]. Fig. 2. Typical neural network architecture. Theoretically, the number of hidden neurons is increased so that the ANN can approximate more complicated non-linear function if it is piecewise continuous [20]. However, it is not the case for all problems. More than one hidden layer can also be used in order to model a complicated system. Training consists of ) calculating outputs from input data, and ) comparing the measured and calculated outputs. The accuracy of the predictions of a network was quantified by the root of the mean squared error difference (RMSE), between the measured and the predicted values [21] Fig. 1. Basic principles of artificial neural networks [19]. The ANN elements are connected to each other with a sequence of layers. A typical neural network consists of three layers: input layer, hidden layer and output layer (Fig. 2). In Fig. 2, are input layer nodes; are hidden layer nodes; and are output layer nodes. In a typical neural network each node is fully connected to the next layer nodes. Input layer feeds data to the network, therefore it is not a computing layer since it has no weights and activation function. Output layer represents the output response to a given input. Generally, logistic function is used in the hidden layer. For simplicity, activation function of the hidden layer to output layer may be chosen linear. (1) 2.1 Multi Layer Perceptron Network (MLP) In this type of network every neuron in each layer is completely connected to neurons in the layer before it, in the sense of data processing. The output of each layer, after applying function effects, becomes the input for the next layer, this process continues until network s output is obtained. Network s behavior is expressed based on the following equations: (2) = input vector, output vector, number of layers, and superscript indicates the layer number The learning methods of MLP are based on Back Propagation (BP) algorithms. There are two calculation directions in the BP algorithm: first, is the feed forward path and second is the feedback ISBN:

4 path. In the feed forward direction, network's parameters will not change during calculations and stimulus functions act on each neuron as data proceeds through the layer. In feedback direction, beginning with the last layer (output layer), where error vector is available, the error is distributed from right to left (last layer to first layer) and local gradient of error distribution is through back propagation [22]. Here the utilized stimulus function is a sigmoid function and the equation is presented in Fig. 3. Fig. 3. sigmoid stimulus function 3. Database The model s success in predicting depends on comprehensiveness of the training data. Availability of large variety of experimental data was required to develop the relationship between the input and output data s. The basic parameters considered in this study were NaNO3 concentration and temperature ( ). A database of 114 samples from the literature was retrieved. The ANNs were designed using 90 pairs of input and output vectors for mass lose prediction, collected from studies. The predicted results obtained from neural network were compared with the experimental values obtained experimentally. The training of ANNs was carried out using pair of input vector and output vector. The summary list of data is given in Table 1 [23]. Table 1. Summary details of the experimental data NaNO3 concentration/% Mass loss/% Results and Discussion 4.1 Artificial Neural Network Model All phases of networks' modeling were performed in MATLAB )Version , The Math Works Inc.(. To train MLP network, back propagation algorithms are used and different types of back propagation algorithms can be seen in Table 2: MATLAB function name trainbfg traincgf traincgp traingd traingda traingdx trainlm trainoss trainrp trainscg Algorithm BFGS quasi-newton back propagation Fletcher-Powell conjugate gradient back propagation Polak-Riere conjugate gradient back propagation Gradient descent back propagation Gradient descent with adaptive linear back propagation Gradient descent w/momentum and adaptive linear back propagation Levenberg-Marquardt back propagation One step secant back propagation Resilient back propagation (Rprop) Scaled conjugate gradient back propagation Table 2. different types of Back Propagation algorithms In order to obtain an optimum architecture, different hidden neurons were assigned. Each architecture was trained with same training set. After each training, root of the mean square errors (RMSE) were determined. The lowest RMSE belongs to the best architecture [24, 25]. The best architecture for deflection basin has two input neurons, one hidden layer with eight neurons, and one output neuron. Learning rate and momentum parameters affect the speed of the convergence of the back propagation algorithm. Learning rate of and momentum of 0.1 were chosen Epoch (iteration number) was limited with 1,000 during training. Generally, while the training data set covers approximately 80% (four in fifth) of all data, the rest of the data (one in fifth) is used as testing data set. Then, the full data set was divided into five parts as randomly. One of the data parts was selected for testing, and the others were used for training. Since cross-validation technique was used, this process was repeated five times. Further, regression and RMSE between experimental and simulated data were determined. Table 1 shows root of the mean square error (RMSE) of each data parts. Table 3. Mean Square Error (RMSE) of each data parts Data group RMSE Data set Data set Data set ISBN:

5 Data set Data set Back Propagation algorithm is in fact the generalized least squares method applied to the multi-layer networks with nonlinear functions. For network training, back Propagation algorithm error with methods of momentum (Fig. 4) and Levenberg-Marquardt were used, Fig. 5. Fig. 6. Regression of network training, Levenberg- Marquardt method As seen from Fig. 5 after the 40 th cycle, with network training error of the process converges. Fig. 4. Sample of momentum method training Figure 2 shows that after 1000 cycles of computations, network is not converged, so the process is not successful. An example of network training by Levenberg-Marquardt method is shown in Fig. 5 and its regression of network training is shown in Fig Comparison of artificial neural network approach and the experimental data In Fig. 7, MLP models' results are presented and compared to real data. Fig. 5. Sample of network training, Levenberg- Marquardt method ISBN:

6 Fig. 7. Comparison of MLP models' and experimental data for NaNO3 concentration/%: a) 0%, b) 5%, c) 10%, d) 20%, e) 30% and f) 40% Comparison of the results yields that MLP is effective in modeling the problem and its predictions is close to real data. If the input data increase, we hope to obtain more successful results than this model and the new predicted data will be better. But this results data show that if we use the ANNs models for predict the effects of material on the bitumen, we can obtain the acceptable results to economize in experiments. However its maybe we use another artificial neural network like RBF to obtain the different and better results. 6. Conclusions In this study, we used artificial neural network for prediction of influence of doping of NaNO 3 on the solid phase thermal decomposition of bitumen. First we remember of history of application of ANN in pavement engineering. Second we introduce artificial neural network and multi layer perceptron. And then we used multi layer perceptron to prediction of influence of doping of NaNO 3. MLP showed the remarkable results that we hope to use this model for other problems and use another artificial neural network in this and other problems. At the end we conclude that usage of numerical modeling such as artificial neural network (MLP, RBF ) helps us to modeling problems and resolve them to cost saving to be. We suggest using this neural network for other problems and using other artificial neural networks like RBF for this and other problems. 7. References [1] S. Tapkin, A recommended neural trip distribution model. PhD thesis, Civil Engineering Department, Middle East ISBN:

7 Technical University, Ankara, Turkey, [2] S. G. Ritchie, M. Kaseko, B. Bavarian, Development of an intelligent system for automated pavement evaluation, Transportation Research Record, Washington, DC, USA, Vol. 1311, 1991, pp [3] M. S. Kaseko, S. G. Ritchie, A neural network based methodology for pavement crack detection, Transportation Research Part C, Vol. 1, No. 4, 1993, pp [4] N. Gagarin, I. Flood, P. Albrecht, Computing truck attributes with artificial neural networks, Journal of Computing in Civil Engineering, ASCE, Vol. 8, No. 2, 1994, pp [5] N. N. Eldin, A. B. Senouci, A pavement condition rating model using backpropagation neural network, Microcomputers in Civil Engineering, Vol. 10, No. 6, 1995, pp [6] Y. Cal, (1995). Soil classification by neural network, Advances in Engineering Software, 22, [7] A. G. Razaqpur, A. O. Abd El Halim, H. A. Mohamed, Bridge management by dynamic programming and neural networks, Canadian Journal of Civil Engineering, Vol. 23, 1996, pp [8] C. A. Roberts, N. O. Attoh-Okine, A comparative analysis of two artificial neural networks using pavement performance prediction, Computer-Aided Civil and Infrastructure Engineering, Vol. 13, No. 5, 1998, pp [9] E. Tutumluer, U. Seyhan, Neural network modeling of anistropic aggregate behavior from repeated load triaxial tests, In: A paper presented at transportation research board 77th annual meeting, Washington, DC, USA, [10] S. Owusu-Ababia, Effect of neural network topology on flexible pavement cracking prediction, Computer-Aided Civil and Infrastructure Engineering, Vol. 13, No. 5, 1998, pp [11] A. M. Alsugair, A. A. Al-Qudrah, Artificial neural network approach for pavement maintenance, Journal of Computing in Civil Engineering, ASCE, Vol. 12, No. 4, 1998, pp [12] Y. Kim, Y. R. Kim, Prediction of layer moduli from FWD and surface wave measurements using artificial neural network, In: A paper presented at transportation research board, 77th annual meeting, Washington, DC, USA, [13] A. R. Shekharan, Effect of noisy data on pavement performance prediction by artificial neural networks, In: A paper presented at transportation research board, 77th annual meeting, Washington, DC, USA, [14] N. O. Attoh-Okine, Grouping pavement condition variables for performance modeling using self-organizing maps, Computer-Aided Civil and Infrastructure Engineering, Vol. 16, No. 2, 2001, pp [15] B. J. Lee, H. D. Lee, Position-invariant neural network for digital pavement crack analysis, Computer-Aided Civil and Infrastructure Engineering, Vol. 19, No. 2, 2004, pp [17] X. Mei, M. Gunaratne, J. J. Lu, B. Dietrich, Neural network for rapid evaluation of shallow cracks in asphalt pavements, Computer-Aided Civil and Infrastructure Engineering, Vol. 19, No. 3, 2004, pp [18] H. Ceylan, A. Guclu, E. Tutumluer, M. R. Thompson, Backcalculation of full-depth asphalt pavement layer moduli considering nonlinear stressdependent subgrade behavior, The International Journal of Pavement Engineering, Vol. 6, No. 3, 2005, pp [19] G. Bosurgi, F. Trifiro, A model based on artificial neural networks and genetic algorithms for pavement maintenance management, The International Journal of Pavement Engineering, Vol. 6, No. 3, 2005, pp [20] H. Demuth, M. Beale, Neural network toolbox. User guide, version 4. The MathWorks Inc.; [21] L. H. Tsoukalas, E. R. Uhrig, Fuzzy and neural approaches in engineering, New York (USA): John Wiley & Sons Inc., [22] M. Saltan, H. Sezgin, Hybrid neural network ISBN:

8 and finite element modeling of sub-base layer material properties in flexible pavements, Materials and Design, Vol. 28, 2007, pp [23] M. B. Menhag, Basics neural networks Amir-Kabir University of Technology Press, [24] N. El-Said, M. S. Sayed, A. S. Mikhail, Influence of doping of NaNO 3 on the solid phase thermal decomposition of bitumen and cement, Journal of Thermal Analysis and Calorimetry, Vol. 63, 2001, pp [25] M. Saltan, S. Terzi, Modeling deflection basin using artificial neural networks with cross-validation technique in backcaculating flexible pavement layer moduli, Advances in Engineering Software, Vol. 39, 2008, pp ISBN: