Track Degradation Prediction Models, Using Markov Chain, Artificial Neural and Neuro-Fuzzy Network

Transcription

1 Degradation Prediction Models, Using Markov Chain, Artificial Neural and Neuro-Fuzzy Network Abstract Y. Shafahi, P. Masoudi and R. Hakhamaneshi Sharif University of Technology,Tehran, Iran condition is one of the most important parameters affecting the track maintenance management. In order to obtain a good track maintenance management system, it is necessary to predict track condition through the time. In this study, the track s will be defined in terms of the Combine Record index (CTR) rating which can vary from 0 to 100 where 100 denotes the best possible track condition and the s are defined as five intervals of CTR, similar to those used in Iranian Railways. Four models are built and testes for the prediction of track quality, one conventional model suggested by ORE, and three new model, using markov chain, artificial neural network and neurofuzzy network. The data for our empirical application was collected from Iranian Railways network. Comparisons of the models show that all three proposed new models predict track deterioration better than the ORE model. Introduction The railway is a branch of the transportation system that is very expensive to construct but it has a long life and low operating costs. Therefore, the asset value is very high, which also leads to the possibility that maintenance might be expensive. Like other infrastructure with investment costs in the construction phase, maintenance plays a crucial role in the long-term cost effectiveness, and so maintenance management is one of the most important parts of the railway systems. condition is one of the most important parameters affecting the track maintenance management. In order to obtain a good track maintenance management system, it is necessary to predict track condition through the time. The cost of maintaining a track depends directly on its condition or "". Maintenance and rehabilitation funds are often allocated to tracks that are in the worst or that exhibit an accelerating rate of deterioration. Some of factors that affect the rate of deterioration of tracks include: traffic loads, weather, and construction materials. The track conditions vary considerably even when such contributing factors are similar. Therefore, it is important that a procedure that can accommodate such randomness be incorporated into a track management system. Several approaches and methods for predicting railways conditions have been proposed, and based on these, a considerable number of maintenance planning tools have been developed for railways systems in North America and Europe. A nonlinear regression model based on a product of the power functions has been proposed by the Office for Research and Experiments (ORE) of the International Union of Railways [9] to predict track deterioration. Having track degradation models, operations research techniques are commonly used to optimize track maintenance activity. Such approaches have been described by Esveld [4] and Zarembski [12]. Zhang [13] has proposed an Integrated Degradation Model (ITDM). ITDM simulates track degradation based on the interaction between different track components under varying traffic. It also considers several mechanistic characteristics, including train speed and axle load. Based on the ITDM, Simson et al. [11] have developed a Maintenance Planning Model (TMPM). It aims to deal with track maintenance planning in the medium to long term. TMPM outputs the net present value of the financial benefits of undertaking a given maintenance strategy compared with a base-case maintenance scenario. The Total Right-Of-Way Analysis and Costing System (TRACS) has been described by Martland et al. [7]. It is a system (software) developed by the Association of American Railroads (AAR) and Massachusetts Institute of Technology (MIT), in the USA. It is a computer-based tool developed to assist rail management to address change in the infrastructure. By combining engineering-based deterioration models with lifecycle costing techniques the model estimates track maintenance and renewal costs as a function of route geometry, track components, track condition, as well as traffic mix and volume. TRACS has been used by North American railroads as a tool for technology assessment costing in support of actions such as pricing, budgeting, and line consolidation. Both the ITDM and the TRACS models are based on an incremental approach where each event such as rail grinding, relining, and track renewal can be included. In recent years the application of soft computing techniques has been paid attention to predicting the future track conditions. Shafahi and Rasooli [10] have considered neuro-networks to predict future track conditions. A neuro-fuzzy decision support system for rail track maintenance planning has been described by Dell'Orco et al. [2]. During the 1990s, the International Railway Union

2 (UIC) in conjunction with the European Rail Research Institute (ERRI) developed an expert system for track maintenance and renewal (ECOTRACK). This model builds on the fact that rules can be specified for certain maintenance activities under certain conditions. A historical database containing infrastructure information on components and current condition is also a prerequisite to use this model. ECOTRACK solves the planning problem given the rules specified and points out the activities needed at a section at certain time. Recent work on ECOTRACK at ERRI [6] has developed the model further in order to improve its functionality. To describe the condition of the track, several indexes and criteria have been defined and used in different railway systems around the world. Commonly, the track quality index (TQI) is used to define the track quality. TQI is normally determined by track geometry parameters (TGP). The term track deterioration is used to describe any changes in track geometry. deteriorations are classified as: unevenness, twist, alignment, and gauge. TQI is a function of these four parameters. In this study, the track s will be defined in terms of the Combined Record index (CTR index) rating which can vary from 0 to 100 where 100 denotes the best possible track condition and the s are defined as five intervals of CTR index (Table 1). It is supposed that the track began its life at some time in the past in near-perfect condition. It was then a subject to a sequence of duty cycles that caused its condition to deteriorate. The duty cycle for the track in this study will be assumed to consist of one year s weather and traffic load. These discrete and discrete time unit definition let us to express the deterioration process as a Markov chain. CTR index quality Failed Medium Good Very good Excellent Table 1: classification by CTR index To have better models, in this study the tracks are categorized so that those with similar traffic loads and geographical location are collected into one class. A network of tracks based on the topography was divided into three groups of plain, hilly and mountainous areas; and based on traffic loads was divided into two groups of light and heavy traffic. We thus have six track classes (Table 2). Traffic condition Topography condition Plain areas Hilly areas Mountainous areas Light Heavy Table 2: Iranian Railways classification by topography and traffic conditions (K) The rest of the paper is organized as follows: First, we introduce the data bank which used for our empirical application. Then, the prediction of track condition by using conventional ORE model, markov model, artificial neural network model, and neuro-fuzzy model are discussed. Summary, comparisons and discussion on prediction models, and conclusions are the last part of this paper. Data bank The data for our empirical application was collected from the Iranian Railways network and incorporated four types of survey: topography, annual traffic and axle load, date of construction or reconstruction, and the track condition of the block in each year. In this study the block, which is the distance between two consecutive stations in the network of tracks, is the smallest maintenance unit, similar to those used by Iranian Railways. Nowadays, the conditions of railways are monitored by special diagnostic trains. MATISA, the diagnostic train of Iranian Railways, measures the geometric parameters of the railway track such as: alignment, gauge, cross-level, and profile to allow us to finally compute the average CTR index for the. The Iranian Railways' network consists of 7,203 km of track (in 2004). Topography, annual traffic load, axle load and the latest date of construction or reconstruction for 215 (4,228 km) were

3 collected from Iranian Railways and the average CTR indexes of these for three years ( ) were computed. These data were collected from different divisions of Iranian Railways including Shomaleshargh, Khorasan, Shomalegharb, Azarbayegan, Gonoob, Lorestan, Shomal, Arak, Gonoobeshargh and Hormozgan. On the basis of this data, 60% (122, 2,544 km) are in plain areas, 19% of these tracks (44, 797 km) are in hilly areas, and 21% (49, 887 km) in mountainous areas. 70% of these tracks have been designed for 20-tonne axle loads and the rest of them for 25-tonne axle loads. Furthermore, 36% of them (72, 1,525 km) have heavy traffic passing over them. Conventional suggested ORE model The ORE of the International Union of Railways proposed a simple model of track deterioration in the more general form of a product of the power functions in In this paper we rebuild the ORE model using our existing data bank and compare the results with the proposed markov, artificial neural and neuro-fuzzy models. Equation (1) shows the rebuilt ORE model. where: E T P = (1) E = a track degradation index (in this paper, CTR index), T = total accumulated tonnage since the track was new (million tones), P = design axle load (tone), the number of observations is 523, and all estimated parameters are within the 95% confidence intervals. Markov model The basic assumption for a Markov process is that the probabilities of going from a to any other depend only on the current, and not on the manner in which the current is reached. This property is often called the memoryless property of the process. In our case, this is a reasonable assumption. Let : p ij = Probability that a track will be in j in the next year when it is currently in i, and P = [p ij ] is the transition matrix which shows the probabilities of various transitions. Our data shows that the track condition will almost never decrease more than 10 CTR index in a year. Thus, we can assume that during each duty cycle (a year) a track will either stay in its current or jump to the next lowest. Thus, the nonzero elements of the matrix P consists of diagonal and adjacent lower diagonal elements. In other words: p ii = p i = The probability that a track will stay in its current s, i, in the next duty cycle, p i(i-1) = 1-p i = q i = The probability that a track will be in i-1 when it is currently in i, and p ij = 0 for all j i, i-1 and i-1>0. (2) Then, transition matrix becomes: q2 p P = 0 q 3 p3 0 0 (3) 0 0 q4 p q 5 p5 The of the track at any time, n, n=1,2, can be expressed in a probabilistic manner as a (1 5) row vector p(n):

4 ( p{ X ( n) = 1 }, p{ X ( n) = 2},..., p{ X ( ) 5} ) p ( n) = n = (4) where X(n) is the track at time n and p{x(n) =j} is the probability that a track is in j at time n. Obviously, the elements of the vector p(n) must sum to 1.0. The initial of the track is given by p(0). For example, if we suppose that the track was in excellent condition when it was new, then p ( 0) = (0,0,0,0,1) (5) Other initial vectors may be defined to reflect the uncertainty in the initial condition of the track. It can be shown [1] that the vector at some future time may be calculated from the transition matrix and the initial vector, as p(1) = p(0) P p(2) = p(1) P = p(0) P... p( n) = p(0) P n 2 (6) Or, more generally, p ( m = ) n + n) p( m P (7) The collected data for the s of the tracks during their lives may be used to estimate the transition matrix P. It is necessary to transform the track survey data into CTR index versus age in order to use the data for estimating the transition matrix. To do so, the age of the track is defined as the difference between current date (date of geometric parameters recording) and the latest date of reconstruction or construction. CTR index, geographical location, annual traffic load, and the latest date of reconstruction or construction of the must be specified to identify the transition matrix P. CTR index versus age data was specified by computing the average CTR index of the that are at the same age in each class. Thus, the track was specified on the basis of the observed condition of the track. Table 3 summarizes the CTR index versus age data. On the basis of Table 3, there are some that are more than 20 years old in light traffic. This means that there are some that have not been repaired for 20 years in the light traffic tracks in Iranian Railways, but in heavy traffic tracks there are no that are more than 9 years old. The sum of the squared differences between the expected predicted from the Markov process (Equations 6) and the observed track condition at each age for which data were available is used as the objective function to estimate the transition matrix. Minimizing this objective function produces the elements of the transition matrix. The Quasi-Newton method [3] used to minimize this objective function. Using an initial guess for the transition matrix elements and excellent condition for the initial vector, the minimization algorithm obtained the transition matrices. An independent transition matrix (Table 4) was computed for each track class. The descending process of the probability of staying in the current in all six classes represents the fact that the track deterioration rate accelerates when the tracks are badly deteriorated. It means that in badly deteriorated s the inclination to jump to a lower increases (note the low differences between the respective figures in columns p 5 and p 4 and the significant differences between those in columns p 3 and p 2 in Table 4). Furthermore, Table 4 also shows that tracks with light traffic have little differences in the deterioration rates across track classes. Moreover, for heavy traffic, tracks in mountainous areas have rates of deterioration faster than those in plain or hilly areas.

5 age Plain areas with light traffic (K = 1) Hilly areas with light traffic (K = 2) Mountainous areas with light traffic (K =3) Average CTR Averag e CTR Averag e CTR age Plain areas with heavy traffic (K = 5) Hilly areas with heavy traffic (K = 4) Mountainous areas with heavy traffic (K=6) Average CTR Averag e CTR Averag e CTR Table 3: Summary of CTR index versus age data for different classes of tracks (K=1 to 6)

6 class (K) Terrain Traffic p 2 p 3 p 4 p 5 1 Plain Light Hilly Light Mountainous Light Plain Heavy Hilly Heavy Mountainous Heavy Table 4: Diagonal elements of transition matrices for the 6 classes of tracks Artificial neural network model Artificial Neural Networks (ANNs) known as connectionist models, are systems that try to make use of some of the known or expected organizing principles of the human brain. They consist of a number of independent, simple processors - the neurons. These neurons communicate with each other via weighted connections - the synaptic weights [8]. An ANN is such a mapping device, which assigns to each input exactly one output (both input and output may be vectors of values). The function is determined by the network's weights which are set while training the network. A neural network is trained, roughly, as follows. The network is shown a set of examples, each consisting of inputs and outputs. It learns the connections among them by assigning weights to connections. This is done by continuously changing weights to get closer to the desired outputs [14]. For the problem at hand, track condition data were available in the form of surveys in which many tracks of the same type were inspected at a single point in time. In this study, track began its life at some time in the past in near-perfect condition and then is subjected to a sequence of duty cycles that cause its condition to deteriorate. The duty cycle for track will be assumed to consist of one year s duration of weather and traffic load. The use of an index (here CTR) is the method to describe the track condition. As mentioned, many items impact the deterioration of CTR index. Factors used as effective items of track degradation icons in this study are: 1. CTR index of a year, previous year, and two and three years before, which was classified from 1 to 5 (Table 1). 2. Traffic volume or traffic load which was divided to two groups, light and heavy traffic (Table 2) 3. Maximum allowable speed which was classified into 5 classes (Table 5). 4. Geographical location which was classified into three classes, plain, hilly, and mountainous areas (table 2). 5. The maximum gradient of the block which was classified into five classes (Table 6). 6. Minimum radius of curves in the block which was classified into 7 classes.(table 7). Maximum Speed intervals Less than 60 Kilometer per hour (k/h) 60 to 80 K/h 80 to 100 k/h 100 to 120 k/h More than 120 k/h Speed classes Table 5: classification by maximum allowable speed Maximum gradient of the block Gradient classes 0% to 0.5% 0.5% to 1% 1% to 1.5% 1.5 % to 2% 2% to 3% Table 6: classification by maximum gradient

7 Minimum radius of curves Radius classes Less than 250 meters (m) 250 to 400 m 400 to 750 m 750 to 1000 m 1000 to m 2000 to 4000 m More than 4000 m Table 7: classification by minimum radius of curves To build a neural network model, some steps should be done. At first the topology of the network should be created, number of layers and nodes of the network, type of network, initial and activation functions are some effective parameters that should be specified. Second, on the basis of training process in a network, the weight of parameters should be corrected, and the data of every situation, as training data for many times, should be shown to the network. Finally, as the neural network, used in this article was a supervised learning one, after training of the network was finished, it will be examined by some known data, so probable errors could be corrected. Testing different topology in this study a network with 3 layers and 5 neurons in internal layer is selected as the optimal network. The data was randomly divided in two sets: training set with 82% of data and test set with 18% of the data. Observed data in terms of 5 groups of CTR classification shows the CTR index of each year never descend more than one level. The results of modeling the neural network mention that all of the predictions of the next year CTR indexes are in the level of the CTR of previous year or one level below that. Table 8 shows results of comparison of the model predictions and observed data for a sample set. Percent of correctness of neural network results With one level wrong correct 33% 67% Table 8: Comparison of the neural network model predictions and observations for a sample set Neuro-fuzzy model The aim of Fuzzy Inference System (FIS) is to take account of the fact that connects of a proposition may be imprecise. Vague predicates induce fuzzy sets. A fuzzy set is defined by its membership function, which usually ranges on the interval [0,1]. A proposition (P) which restricts the possible values of a variable (x) is represented by means of a membership function ( μ p ( x) ) [14]. The combinations of ANNs with FIS, so-called neural fuzzy or neuro-fuzzy systems, offer a great number of interesting applications. Hybrid neuro-fuzzy models are one of modern neuro-fuzzy approaches. A neural network and fuzzy system are combined in a homogenous architecture. The system may be interpreted as a special neural network with fuzzy parameters, or as a fuzzy system implemented in a parallel distributed form. One of the first hybrid neuro-fuzzy systems for function approximation was Jang's ANFIS model. By using a hybrid learning procedure, the proposed ANFIS can construct an input-output mapping based on both human knowledge (in the form of if-then fuzzy rules) and stipulated input-output data pairs [5]. In ANFIS structure membership function parameters, if-then rule exertion and output parameters are calculated by train data set. The training algorithm is usually hybrid or back propagation. With all of the condition mentioned before, the neuro-fuzzy network model was built. Table 9 shows results of comparison of the model predictions and observed data for a sample set. As it shown, by using neuro-fuzzy network, around 73 percent of the estimations are similar to real outcomes exactly. On the other hand, more than 6 percent improved from the previous model. With one level wrong Percent of correctness of neural network results correct 27% 73% Table 9: Comparison of the neuro-fuzzy model predictions and observations for a sample set

8 Summary and conclusions A model to portray track deterioration is required in track maintenance scheduling procedures, since the prediction of track conditions is a necessary part of deciding how to allocate maintenance funds. In this study four models are built and testes for the prediction of track quality, one conventional model suggested by ORE, and three new model, using markov chain, artificial neural network and neuro-fuzzy network. For comparison of four models liner regression between observation and estimation of each model is used. Table 10 shows results. Obviously, when coefficient a is close to one and coefficient b is close to zero, the model is a good predictor of the actual observations. R 2 values that have been shown in table 10 also show that all three proposed new models; Markov, Neural network, and neurofuzzy models; predict track deterioration better than ORE model. Model Observation-Estimation Relation a b R 2 ORE Markov chain Neural network Neuro-fuzzy network NOTE: Observation-Estimation Relation Form : y = ax + b ( x = Observation, y = Estimation) Table 10: Comparisons of markov chain, neural network, neuro-fuzzy network and ORE models. Acknowledgements The authors would like to thank The Iranian Railways for the provision of the data in support of this study. References [1] B. Ayyub, Uncertainty modeling and analysis in civil engineering. Carol Royal, New York, (1998). [2] M. Dell'Orco, M. Ottomanelli and D. Sassanelli, Neuro-Fuzzy decision support system for rail-tracks maintenance planning. Proceedings of the 10th World Conference on Transport Research. WCTR'04. July 4-8, (CD-ROM version). Istanbul, Turkey, (2004). [3] M. Davidian and D. M. Giltinan, Nonlinear Models for Repeated Measurement Data, Chapman & Hall, First Edition, (1995). [4] C. Esveld. Computer-Aided Maintenance and Renewal of. Proceedings of the 4th International Heavy Haul Conference, September, Brisbane, Australia, International Heavy Haul Association Inc., Virginia Beach, VA, USA, pp , (1989). [5] J. Jang and, R. Shing R, "ANFIS: Adaptive-Network-Based Fuzzy Inference System", IEEE Transactions on Systems, Man, and Cybernetics, (1993). [6] S. Jovanovic, and C. Esveld, ECOTRACK: An objective condition-based decision support system for long-term track M&R planning directed towards reduction of life cycle costs. Proceedings of the 7th International Heavy Haul Conference, International Heavy Haul Association Inc., Virginia Beach, VA, USA, pp , (2001). [7] C. D. Martland, M. B. Hargrove and A. R. Auzmendi, TRACS: A tool for managing change. Railway and Structures, vol. 90, pp , (1994). [8] D. Nauck, F. Klawonn and R. Kruse, Neuro-Fuzzy Systems, John Wiley & Sons, (1997).

9 [9] Office for Research and Experiments (ORE) Question D161, Dynamic vehicle/track phenomena, from the point of view of track maintenance, report no. 3, final report, of the International Union of Railways(1988). [10] Y. Shafahi and M. Rasooli, A neuro networks model to predict future track condition. 6th International Conference on Civil Engineering, Isfahan University of Technology, Isfahan, Iran, (2001). [11] S. A. Simson,, L. Ferreira and M.H. Murray Rail Maintenance Planning: An Assessment Model. Transportation Research Record 1713, pp , Transportation Research Board, National Research Council, National Academy Press, Washington D.C., USA, (2000). [12] A. M. Zarembski, Rail Life analysis and its use in planning track maintenance. Rail Technology International, pp , Sterling Publications Limited, (1993). [13] Y. J. Zhang An integrated rail track degradation model. PhD Thesis, Physical Infrastructure Centre, School of Civil Engineering, Queensland University of Technology, Brisbane, Australia,(1999). [14] H. J. Zimmermann, Fuzzy Set Theory and Its Applications, Third Edition, Kluwer Academic Publishers, (1996).