Failure Prediction on Railway Turnouts Using Time Delay Neural ...

Failure Prediction on Railway Turnouts Using Time Delay Neural Networks Halis Yilboga, Ömer Faruk Eker, Adem Güçlü

Fatih Camci

Computer Engineering Department Fatih University Istanbul, Turkey

Computer Engineering Department Meliksah University Kayseri, Turkey

Abstract— Turnout systems on railways play critical role on reliability of railway infrastructure. Identification and prediction of failures on mechanical systems have been attracting researchers and industry in recent years. Condition based maintenance focuses on failure identification and prediction using sensory information collected real-time from sensors embedded on electro-mechanical systems. This paper presents neural network based failure prediction algorithm on railway turnouts. Keywords- failure predictions; prognostics; condition based maintenance; railway turnouts; neural network; time series; forecasting

I.

II.

RAILWAY TURNOUT SYSTEMS

Railway turnout systems move the rails for trains to change their tracks. Fig 1 displays an example of a turnout system. Electro-mechanic, hydraulic, and pneumatic are some types of the turnout systems. Turnout systems are the one of the most important component of the railway infrastructure. Electromechanic turnout systems consisting of motor, reduction gear, several bearings, drive-detection rods, and switches are used in this paper.

INTRODUCTION

Turnout systems on railways play critical role on reliability of railway infrastructure. Identification and prediction of failures on mechanical systems have been attracting researchers and industry in recent years. Condition based maintenance focuses on failure identification and prediction using sensory information collected real-time from sensors embedded on electro-mechanical systems. Failure identification is relatively mature compared to failure prediction and several methods on failure identification exist on railway turnouts. [1]-[6]. Three main approaches exist in the literature for failure identification in turnout systems: feature-based, model-based and empirically-based methods. In feature based approach, special features are extracted to identify the failures [1]. In model-based approaches, failure is identified by the deviation amount of the collected signal from a pre-defined model [2], [3]. Empirically-based approaches analyze the difference of collected signal from a fault-free sample to identify the failure [4], [5]. Failure identification methods for turnout systems are summarized in [6]. Failure prediction is the next step of failure identification. In failure prediction, the failure time is estimated to be able to fully benefit from the parts or components to be replaced during the maintenance/repair. Several failure prediction methods have been presented for various systems [8]-[12]. This paper presents neural network based failure prediction algorithm on railway turnouts. Section II presents the railway turnout system. Section III gives neural network used for failure prediction. Section IV reports the results and section V concludes the paper.

978-1-4244-7230-7/10/$26.00 ©2010 IEEE

Fig 1: A Railway turnout system

III.

NEURAL NETWORK BASED FAILURE PREDICTION

Artificial Neural Networks (ANN) that are inspired from human brain are famous methods applied to various domains. ANNs are in general used for classification, clustering, function approximation and prediction different types of ANNs exist today. ANNs are good at function approximation. In fact, there is proof that a fairly simple neural network can fit any practical function. ANN is a multi-layered network with input layer, output layer and hidden neurons as shown in Fig 2.

Hidden Layer(s)

Input Layer

xt

Output Layer

xt-1 xt-n

Fig. 2: Describes the layered structure of an ANN

Learning in ANN can be considered as a function optimization problem, where the task is to determine the best network parameters (weights and biases) in order to minimize error. In this way, several function optimization techniques from numerical linear algebra can be applied to learning. Levenberg-Marquardt learning algorithm provides a numerical solution to the problem of minimizing a (generally nonlinear) function, over a space of parameters for the function. ANNs can be viewed as highly nonlinear functions with the basic form: (1) f x, w  y x is the input vector presented to the network, w are the weights of the network, and y is the corresponding output vector approximated by the network. In this manner, the problem of neural network training can be defined as finding the best w that gives the minimum difference between y and f(x,w). Levenberg-Marquardt is a simple and robust learning method solving the equation:



J

T



ot-1 ot-k Fig 3: An example of FDNN

IV.

EXPERIMENT & RESULTS

Data from an electro-mechanical railway turnout system is collected using several sensors. It is realized that force sensor includes the most valuable information about the health of the turnout system. Fig. 4 displays the sensors installed. Fig 5 and 6 displays the example of current and force data collected.



J  I   J T E

(2) J is the Jacobian matrix for the system, λ is the Levenberg's damping factor adjusted at each iteration and guides the optimization process, δ is the weight update vector and E is the error vector containing the output errors for each input vector used on training the network. δ is the change needed in weights to achieve a (possibly) better solution. The matrix can also be known as the approximated Hessian.

JT J

A. Time-delay Neural Network (TDNN) Fig. 4: Sensors installed in turnout systems 4

Failure progresses

(a)

3

2

Current

Prediction is the process of estimating future values of parameters based on current and past information. Time delay neural network (TDNN) is a neural network that incorporates the time information in the structure. The structure of TDNN is set so that the input of NN consists of parameters in different time units. This is important in time related problems such as prediction. Thus, the effect of past values of parameters can be incorporated in to the problem. The input can also include the past outputs of TDNN. TDNN is especially important in prediction and classification of time related patterns such as speech recognition. Fig. 3 illustrates an example of TDNN.

1

Close to Failure State

Fault Free State

0

-1

-2 0

A sample 50

100

150

200

250

300

Life of the turnout

Fig 6: Current signals with failure progression from fault free to close to failure states

Fig 8: Failure Progression Forecasting 3

(b) 2.5

Failure progresses

2

Force

Fault free 1.5 state 1 0.5 0 -0.5 A sample

-1 -1.5 0

50

Close to Failure State

100

150

200

250

300

350

Life of the turnout

Fig 6: Force signals with failure progression from fault free to close to failure states

The degradation path of the turnout systems are given in Fig. 7. The degradation path is converted discrete degradation level. The signal collected from a turnout system is used for identification of the degradation level of the system. Then, TDNN is used to predict the future degradation levels of the system using the past degradation levels. Fig. 8 illustrates the failure progression forecasting. %25 of the dataset is left for testing; the remaining is used for training. Training data is used in the construction of the ANN and test data is given as input. Learning in ANN is controlled to avoid memorization as shown in Fig 9. Classification of discrete degradation levels are performed with a success rate of %99.3. Then prediction is performed and results of prediction are discussed below.

Degradation Level Failure Progression States

22 20 18 16 14 12 10 8 6 4

Fig 9: Performance of the ANN

When the predicted failure degradation is reaches to the predifed failure threshold, the number of predictions made is defined as the Remaing Useful Life (RUL). RUL prediction results with real RUL for ten turnout systems are shown in Fig 10. As seen from the figure, prediction gets better as the system approaches to failure. RMSE and r-square are used for evaluation of RUL prediction in the literature. Fig 11 shows the r-square values of real and estimated RUL of each turnout. The average of rsquare values is around 0.92. RMSE values represent the error rate of real and estimated RUL and shown in Fig 12. The average of RMSE values is around 2.75.

Degradation State 10 - Failure State State 9 State 8 State 7 State 6 State 5 State 4 State 3 State 2 State 1 - Fault Free State

2 0

5

10 Life

15

20

Fig 7: Degradation path for turnout systems

Failure Threshold Progression Real Progression Degradation Level

Forecasted Progression Current Time Time

Fig 10: Real and estimated RUL for 10 turnout systems with TDNN

ACKNOWLEDGMENT This research was supported by The Scientific and Technological Research Council of Turkey (TUBITAK) under project number 108M275. REFERENCES [1]

Fig 11: r-square values of real and estimated RUL

Fig 12: RMSE values of real and estimated RUL

V.

CONCLUSION

Time-delay Neural Network is used for prediction of failure on railway turnout systems. Data collected from turnout systems is used to identify the discrete degradation level of the system. Then, the future degradation levels are predicted using TDNN. The results are presented in the paper.

Marquez F. P. G., Schmid F., Collado J. C., A Reliability centered approach to remote condition monitoring: A railway points case study, Reliability Engineering and System Safety, 80, (2003), 33-40 [2] Marquez F. P. G., D. J. P. Tercero, Schmid F., Unobserved component models applied to the assessment of wear in railway points: A case study., European Journal of Operation Research, 176 (2007), 1703-1712 [3] C. Roberts, H.P.B. Dassanayake, N. Lehrasab, C.J. Goodman: Distributed quantitative and qualitative fault diagnosis: Railway junction case study, Control Engineering Practice, 10(4), 2002, 419-429 [4] Marquez F. P. G., Schmid F., A Digital Filter Based Approach to the Remote Condition Monitoring of Railway Turnouts, Reliability Engineering and System Safety, 92, (2007), 830-840 [5] V. Atamuradov, F. Camci, S. Baskan, M. Sevkli, "Failure Diagnostics for Railway Point Machines Using Expert Systems ", The 7th IEEE International Symposium on Diagnostics for Electric Machines, Power Electronics & Drive, Cargese, France, 2009 [6] Marquez F. P. G., Roberts C., Tobias A. M, Railway point Mechanisms: condition monitoring and fault detection, Proc. IMechE Vol. 223 Part F: Journal of Rail and Rapid Transit, doi: 10.1243/09544097JRRT289 [7] M. Dong, D. He, A segmental hidden semi-Markov model (HSMM) based diagnostics and prognostics framework and methodology, Mechanical Systems and Signal Processing, 21, 2007, 2248-2266 [8] P. Baruah, P and R.B. Chinnam, R.B., HMMs for diagnostics and prognostics in machining processes, International Journal of Production Research, 43(6), 2005, 1275-1293 [9] R. B. Chinnam, P. Baruah, Autonomous diagnostics and prognostics in machining processes through competitive learning-driven HMM-based clustering, International Journal of Production Research, 47 (23), 2009, 6739 – 6758 [10] N. Gebraeel, A. Elwany J. Pan, “Residual Life Predictions in the Absence of Prior Degradation Knowledge”, IEEE Transactions on Reliability, 58(1), 2009, 106-116 [11] Camci F., Chinnam R. B., "Health-State Estimation and Prognostics in Machining Processes", IEEE Transactions on Automation Science and Engineering, doi: 10.1109/TASE.2009.2038170 [12] F. Camci, R. B. Chinnam, Process Monitoring, Diagnostics and Prognostics in Machining Processes, LAP Lambert Academic Publishing, 978-3838335667, Jan. 2010