The Use of Artificial Neural Network and the Road Conditions - Essay Example

of South Australia School of Natural & Built Environments APPLICABILITY OF ARTIFICIAL NEURAL NETWORK (ANN) IN GEOTECHNICAL/PAVEMENT ENGINEERING TABLE OF CONTENT INTRODUCTION …………………………………………………………………3 RESEARCH OBJECTIVE………………………………………………………….4 SCOPE OF RESEARCH……………………………………………………………5 LITRATURE REVIEW…………………………………………………………… 5 1.1 RADIAL BASIS FUNCTION NETWORK (RBF)…………………………..6 1.2 Modular Neural Network (MNN)………………………………………………7 1.3 APPLICATION of ANN IN PAVEMENT ENGINEERING………………..7 METHODOLOGY…………………………………………………………………… 8 1.1 NEURON SOLUTION ………………………………………………………….9 1.2 PREDICTION OF OUTPUT MODEL.................................................................9 1.3 MODULUS OF FIRST LAYER (E1)....................................................................10 CONCLUTION………………………………………………………………………..11 RECOMMENDATION……………………………………………………………… 12 REFRENCES …………………………………………………………………………13 Introduction In the recent years, Pavement has been gaining popularity. It has been adopted by various types of transportation vehicles with different weights .This has greatly affected pavement surfaces on different levels. Because of the weight issues, the modulus of the pavement has taken center stage with many researchers trying to research on how to improve the conditions of the pavements and keeping them in a useful state. Modulus most important responsibility is to assist in the improvement of the pavement conditions and keeping it in a state that can useful. The cautious monitoring of the existing pavements has taken center stage in many countries, However determining the character of the pavement that consist of a variety of systems in a non linear position has became a difficult and always challenging task. Hence the most commonly used methods in determining the conditions are the non destructive (NDT) methods. This technique always assists the engineers when it comes to repairs and maintenance work. The data that is always obtained by the FDW, is always used to determine the fitness and stiffness of the road .The analysis of this data is used to determine the lifespan of the road by the engineers (Owusu-Ababio 1998). This paper seeks to address the use of Artificial Neural Network (ANN) in back calculation. Efficient use of ANN can be used by engineers in checking the road conditions and can go long way to enhance the engineers’ decision making process. There are three types of ANN which include: Basis Function Network(RBF),Modular Neural Network(MNN) and the Multi Layer Perceptro (MLP).All these three types will be placed under observation in order to determine there efficiency in capturing the inefficiency to back calculate and its ability to predict the strength of the pavement given the outcome from the back calculation of data from EFRROMD3 program. RESEARCH OBJECTIVES The project seeks to highlight and address issues that are involved with pavement engineering. It seeks to establish if the ANN program can be relied upon to predict the pavement condition with non destructive testing data. The main objectives that are stated in the proposal are: To investigate and a certain the application of ANN in pavement engineering, with the main focus being to establish the pavement modulus using the ANN . Its output is predicted using the in putted data from the field. To establish the data that form the filling weight deflector (FWD) e.g. the load, deflection or the thickness of the layer this affects the performance of the (ANN). To compare the effectiveness of the three types i.e. RBF, MNN AND MLP The out put results will be compared with the expected results of each type. The research question This project aims to analyze the application of ARTIFICIAL NEURAL NETWORK (ANN) in pavement engineering. The questions focused on predicting the modulus of flexible pavement layer. The main questions in the research were. 1 .If ANN can predict pavement modulus 2. How effective and accurate can it be in comparison with the others? 3. Which type of the ANN is the most effective? SCOPE OF RESEARCH The scope of his study involves assessing the conditions of flexible pavements that comprises of Asphalt over granular and sub grade layers Asphalt granular bae, granular sub base and sub grade layer Asphalt overlay granular layer. The data collection process will aim to collect pavement data for the interpretation procedures that can predict the layer condition parameters. These are the data that were used as the main input for the ANN LITERATURE REVIEW An Artificial Neural Network (ANN) is like an abstract computer model of the human brain although the operation of the ANN is much simpler. Just as the brain which depends entirely on neurons, the processing power of the ANN depends on the way the neurons are organized. Key in the design of the ANN ,is the suitability for modeling in prediction or classification tasks is Multi Layer Feed –Forward Neural MFNN(Flood & Kartam 1994).it is the non visible layer which extends from the processing pore of the network. The figure below shows a typical 1 non visible multi layer feed forward neural net work. The processing unit’s number depends on the number of predictors and predictions .However the selection in number of the node in each hidden layer depends on the nature of the application. Hence larger architecture with many invisible layers can handle more complex modeling operations but in the scarifies the training time and level of generalization(Owusu-Ababio 1998).An experimental neural network has three or four hidden layers that utilize millions of neurons, (Negnevitsky 2011)But it is suggested that one hidden layer is the most suitable()for predicting and classifying task of practical application. Owusu-Ababio also stated that maximum of four hidden nodes per unit variable were recommended as a rule of thumb in choosing number of hidden nodes. 1.1 RADIAL BASIS FUNCTION NETWORK (RBF) This is nonlinear hybrid network typically containing a single hidden layer of processing elements. It is embedded into a two –layer feed forward Neural Network (Bors 2001)the layer uses Gaussian transfer rather than the standard sigmoidal functions employed by MLP networks. Figure showing RBF model The Gaussian function is set by both the supervise and unsupervised learning rules. These networks tend to learn much faster than the MPLS. 1.2 Modular Neural Network (MNN) Modular Neural Network (MNN) is especial type of MLP .Normally it processes its input using several parallel MLP. This tends to create some structure within the topogy, which will foster specialization of an action in each sub module. However unlike the RBF and ANN , it does not have full interconnectivity between the layers. Hence a smaller number of weights are required for the same size network. This tend to speed up training and reduce the number of required training exemplar(Neural Dimension)they are mostly used as neural monolithic neural networks to solve a specific problem. 1.3 APPLICATION of ANN IN PAVEMENT ENGINEERING Pavement rehabilitation is a major activity for the entire highways agency and has several implications on the resources of the agency and traffic disruptions because of extended lane closures. There has been an increase in traffic on major roads all over the country hence this has lead to cases of pavement failures and this has resulted into larger expenditure in terms of pavement rehabilitation.Several equipments like falling weight deflectometer, road rater, and the dynaflect are used(Sharma, S & Das 2008) by state highway agencies to apply patterns of loading and record deflection data along the pavement. METHODOLOGY The method developed in this study to predict pavement stiffness is based on the ARTIFICIAL NEURAL NETWORK (ANN) Deflection measurements of the pavement surface are important for the evaluation of the performance of a flexible pavement structure and load transfer. The surface deflection is a function of many factors used in the design of the pavement such as traffic, pavement structural section, temperature, and moisture. Thus many characteristics of a flexible pavement can be determined by measuring its deflection in response to load. The advantage is that no empirical transfer functions are used in this comparison. Non Destructive Testing (NDT) methods are very important tools in determining the mechanical properties of the flexible pavement layers.NDT methodology for the structural performance of the pavement system is inversely proportional with the a mount of surface deflections emerged by the applied load. Based on different loading details (type , duration, and magnitude) and deflection measurement locations ,the NDT can arrange into three categories which are static steady-state vibratoy, and the time domain impulse. Static loading is the first and easiest case and cannot represent the actual traffic loads, data were collected under time domain impulse dynamic load((Saltan & Terzi 2008)).The steady state dynamic case is similar to the effect of the vehicle passing over the pavement section, and the loading period depends on the vehicle speed ()in the time domain impulse, loading and impulse load. This is applied on pavement surface and deflection data is recorded in the time domain.FWD is used to find the pavement surface deflection by applying the impulse load. 1.1 NEURON SOLUTION This helps in providing an alternative way to predict the material behavior, and it only accepts the comma separated values-formatted file. Hence the user needs to identify the model that can suit the project. A network is then chosen, such as multilayer perception (MPL) networks or radial basis functions (RBF) networks. Cross validation is highly recommended method for stopping network training. It monitors the error on an independent set of data and stops the training when the errors begin to increase. In this project, the MLP, RBF and modular neural (MNN) networks are selected .the result from our analysis would be used to determine the better method in prediction. 1.2 PREDICTION OF OUTPUT MODEL For three layered pavement that has one output, the prediction of the elastic moduli E1 is influenced by the number of inputs. Hence as the number of inputs increases in MNN network, the root mean square error (RMSE) is relatively stable and slightly changes in dissenting manner. Also the coefficient correlation of these networks is relatively stale .the coefficient of correlation of RBF and MLP networks are lower than the coefficient of correlation of MNN network. Inputs RMS Coefficient of correlation RBF MNN MLP RBF MNN MLP 5 2472 2600 0.902 0.862 6 2613 2427 0.886 0.895 7 2890 4067 0.841 0.733 8 2748 2629 0.875 0.891 9 2848 1976 2033 0.886 0.926 0.922 10 3477 1884 3716 0.878 0.931 0.848 Figure showing coefficient of correlation. No of output for three layered pavement one out put. Prediction of two output model 1.3 MODULUS OF FIRST LAYER (E1) If the network has two outputs, the prediction of the elastic moduli E1 is influenced by the number of in puts. As indicated in the figure below and table below. An increase in the number of inputs in MNN network, leads to a reduction in the mean square error (RMSE) which reduces from 2306 to 1448 in 7 and 8 inputs model and 1448 to 3329 in the 9 and 10 input models. The chat and the table below shows that RMSE in RBF network increase when the number of inputs increases but the RMSE of MLP networks is decreased with increase in the number of inputs .The table below show that RMSE in RBF network is increased from 1908 to 5314; and the RMSE in MLP network is reduced from 4658 to 4037. Also the coefficients of correlation of these works are in a descending trend. Hence this means that the correlation become weaker, when increase the number of input Inputs RMS Coefficient of Correlation RBF MNN MLP RBF MNN MLP 7 1908 2306 0.930 0.902 8 2106 1448 0.922 0.959 9 2797 2915 4658 0.866 0.843 0.800 10 5314 3329 4037 0.626 0.786 0.776 Table 4. 1 RMSE and Coefficient of correlation of the networks (3layered pavement 2 output (E1)) CONCLUSION Generally, MNN perform well on the given data set. One possible explanation the pavement data has highly dimensional problems and complex (Sharma, D & Dhar 2010). MNN is well suit the task as network is divided into subtask to perform each one in parallel. Fewer inputs are required for better prediction of more outputs in MNN model. About 8 to 9 inputs are sufficient enough to produce results with few errors for the MNN model. RBF model prediction depends on the number of outputs predicted. More inputs are needed for predicting the modulus of a 3 layers pavement with more outputs. RECOMMENDATIONS FOR FUTURE RESEARCH Due to a lot of constraints during the project schedule, a lot of other useful parameters were not covered. The time frame for this project was not ample to cover all of the other important information. The shortage of project programs and software is the other limiting factor. Only one license was available for the neurosolution. This made it hard to keep checking and running the program anytime. Due to these constraints, recommendations for future research were made to further analyze this area of study. These are as follows: The remaining life of the pavement Shortage of the project sofwares and programs Comparison of the modulus should be done using the equations. During this research, too many different types of formulas for calculating modulus were found. It was found in the literature that each equation can be used for a specific location or region. REFERENCES Munakata, T 2008, Fundamentals of the new artificial intelligence [electronic resource] : neural, evolutionary, fuzzy and more, 2nd edn, Springer, London. Negnevitsky, M 2011, Artificial intelligence : a guide to intelligent systems, 3rd edn, Addison Wesley/Pearson, Harlow, England ;New York. NeuroDimension, Neurosolutions 6.0, NeuroDimension, Orr, MJL 1996, Introduction to Radial Basis Function Networks, Centre for Cognitive Science, Edinburgh EH8 9 LW Scotland. Owusu-Ababio, S 1998, Effect of Neural Network Topology on Flexible Pavement Cracking Prediction, Computer-Aided Civil & Infrastructure Engineering, vol. 13, no. 5, p. 349. Saltan, M & Terzi, S 2008, Modeling deflection basin using artificial neural networks with cross-validation technique in backcalculating flexible pavement layer moduli, Advances in Engineering Software, vol. 39, no. 7, pp. 588-592 Shahin, MA, Jaksa, MB & Maier, HR 2002, Artificial Neural Network based Settlement Prediction Formula for Shallow Foundations on Granular Soils, Australian Geomechanics, vol. 37, no. 4, September 2002. Sharma, D & Dhar, J 2010, Face recognition using modular Neural Networks, paper presented at the Software Technology and Engineering (ICSTE), 2010 2nd International Conference on, 3-5 Oct. 2010. Smith, M 1993, Neural network for statistical modelling, Van Nostrand-Reinhold, New York. Terzi, S 2007, Modeling the pavement serviceability ratio of flexible highway pavements by artificial neural networks, Construction and Building Materials, vol. 21, no. 3, pp. 590-593 Read More