Prediction of turnout support deterioration through dynamic train- track interactions integrated with artificial intelligence Jessada Sresakoolchai 1 University of Birmingham Edgbaston, Birmingham, United Kingdom B15 2TT Mehmet Hamarat University of Birmingham Edgbaston, Birmingham, United Kingdom B15 2TT Sakdirat Kaewunruen 2 University of Birmingham Edgbaston, Birmingham, United Kingdom B15 2TT

ABSTRACT Due to the increase of rolling stocks’ speed and limited area for railway project construction which result in sharper curves. Railway turnouts are components that are highly impacted by this phenomenon. Therefore, the loads and vibrations applied to them are high and result in support deterioration. This study aims to use axle box accelerations of rolling stocks to predict railway turnout support deterioration. The key parameter used to measure the deterioration is support stiffness. Support stiffness deterioration can be occurred by different causes such as the exceeded load applying to rail infrastructure, regular application, or extreme events such as flooding. These causes all make the turnout support deterioration and the turnout support stiffness will decrease. Besides the rail infrastructure deteriorates and maintenance needs to be performed which results in cost, it also results in worse passenger comfort because the track infrastructure is less stable. The finite element method is applied to develop rolling stock models and generate numerical data showing the relationship between railway turnout support stiff- ness deterioration and axle box accelerations. Finite element models are verified with field data to ensure that the results from simulations are reliable. A predicted model is developed using a convolutional neural network model. Keywords: axle box acceleration, railway turnout, support deterioration, machine learning

1. INTRODUCTION

Turnout is one of the most important components of the railway structure. Turnout is a set of mechanical equipment that is used to guide rolling stock from one track to another track. Therefore, turnout will be found at railway junctions, railway spurs, siding tracks, or branches. Turnouts are movable and define the direction of rolling stocks. Turnout consists of a pair of switch rails or point 1 jss814@student.bham.ac.uk

2 s.kaewunruen@bham.ac.uk

i, orn inter.noise 21-24 AUGUST SCOTTISH EVENT CAMPUS ? O? ? GLASGOW

blades that lies between the stock rails. This part is laterally movable. If turnout is not locked, rolling stocks will pass through turnout. On the other hand, rolling stock will be directed to the defined direction according to the location of turnout. From the function, turnouts are affected by high load and impact during the operation [1-3]. However, railway turnouts are weak points in the railway structure [4-9]. In the case of the ballast structure, turnouts are lied on crushed rocks which gently deteriorate along with the operation. Turnouts are important causes of railway disturbance. More than 400 turnouts have to be replaced per year while two of them are urgently needed to be repaired per week [4]. In the Netherlands, the cost for replacement is about 6.4 million per year. It can be clearly seen that railway turnouts are railway components that need to be managed efficiently. Turnouts consist of different components such as railroad switch, connection part, frog, and guard rail. Each component also contains sub-components and they all have an effect on the functionality of turnout. Besides railway turnouts and their components’ conditions, the support that turnouts lie on also influent the functional readiness of turnouts. This study focuses on the prediction of turnout support deterioration. An objective is to develop a machine learning model to predict or estimate the turnout support deterioration using support stiffnesses as the key indicator. A machine learning tech- nique is used to develop a predicted model. Features of the machine learning model are axle box acceleration or vibration from the front wheels of rolling stocks. An expected contribution of this study is the developed machine learning model will have the ability to predict the turnout support deterioration which will be beneficial to railway maintenance planning and management. 2. LITERATURE REVIEW

In 2022, Hamarat et al. [10] studied the deterioration of railway turnout support. They developed a finite element model and validated the model using the field data. They found that the railway turnout support deterioration could occur by different incidents. In their study, they focused on the flooding events. They found that flooding significantly accelerates the deterioration of turnout sup- port. Then, defects in terms of overall track geometry and each track geometry parameter. Ishak et al. [11] developed a safety-based maintenance approach for geometry restoration of turnouts. They applied the probability and severity of each failure to consider the maintenance approach. The probability and severity of failures were used to calculate the damage of failures and then they were tried to be minimized to the reasonably acceptable level. They used failures and fault tree analysis to analyze the probability and severity of failures. Then, they applied risk management to prioritize each failure. Kaewunruen [12] studied the structural deterioration of railway turnout using the dynamic wheel and rail interaction. In that study, the average peak acceleration was used to identify the deterioration of turnouts. However, only one value from axle box accelerations was used to identify the deterioration. in 2019, Sysyn et al. [13] developed a machine learning model to detect railway turnout defects. the developed machine learning model is the binary-class model in which the model would detect whether that turnout has a defect. They used axle box acceleration as the key feature to train the machine learning model. They achieved 90% accuracy using machine learning to detect defects in railway turnouts. In 2021, Sresakoolchai and Kaewunruen [14] applied machine learning models to detect wheelflat and classify the severity of wheelflat. They found that the accuracy could be achieved by more than 95%. In that study, they used axles box accelerations as the raw data to develop the machine learning models. Sresakoolchai and Kaewunruen [15] also applied axle box accelerations to detect and classify the combined railway defects. In that study, they were interested in detecting and classifying the

i, orn inter.noise 21-24 AUGUST SCOTTISH EVENT CAMPUS ? O? ? GLASGOW

severity of the combined defect between dipped joint and settlement that is usually found in the rail- way infrastructure, especially at railway joints. They also found that the accuracies were high up to 90% when applying machine learning to fulfill this purpose. Other than that, there have been many studies identifying that machine learning can be used in the railway industry for many applications such as defect detection [16-19], defect severity classification [20, 21], or operation [22, 23]. From the previous studies, it can be seen that the capability to estimate or detect turnout support is crucial. However, the machine learning approach has not been used for this purpose. This study has an aim to demonstrate the potential of machine learning to use axle box acceleration to detect turnout support deterioration. 3. METHODOLOGY

The machine learning model in this study uses numerical data to train the model. The numerical data are generated using finite element models developed by using LS-DYNA which is a popular finite element method software. The finite element models are developed based on field data and validated using the field data [24]. The railway turnout support deterioration is imitated by adjusting the stiffness and damping coefficient of ballast that supports turnouts. Different parameters are varied to create data variation. An example of the developed finite element model is shown in Figure 1

i, orn inter.noise 21-24 AUGUST SCOTTISH EVENT CAMPUS ? O? ? GLASGOW

Figure 1: An example of a finite element model.

From the figure, the finite element model contains different components of the railway infrastruc- ture and turnout. The main component is the turnout that a 1:9 crossing angle turnout is used in this study. Sleepers are imitated using beams supported by springs and dampers to imitate rail pads. Rail pads are supported by ballasts which are imitated by using a series of springs and dampers. Other boundaries are defined to imitate the realistic behavior of turnouts. To mimic the turnout support deterioration, adjusting of ballast’s stiffnesses and damping coefficient can be done by assuming different classes of deterioration [24, 25]. The data variation can be shown in Table 1. Moreover, other parameters are also varied which consist of weights and speeds of rolling stocks. The summarized table is shown in Table 1. The machine learning model will be developed as a classification model when classes are based on the classes shown in Table 1.

Table 1: Variation of stiffnesses and damping coefficient of ballast.

Parameters Units Classes Ranges Stiffnesses MN/m 1 More than 12.8 2 10.5-12.8 3 8.1-10.5 4 Less than 8.1 Damping coefficient kNs/m 1 Less than 1.2 2 1.2-1.5 3 1.5-2.0 4 More than 2.0 Weights of rolling stocks Tons 32-48 Speeds of rolling stocks Km/h 135-225

The total number of simulations is 1,936. As mentioned in the previous section, outputs from the finite element model simulation that are used to train the machine learning model are axle box accel- erations or vibrations. Two series of axle box accelerations from two front axles are used as the main features to train the machine learning model. Axle box accelerations are in form of time-series data with the frequency of 1,000 Hz. The purpose of the machine learning model is to classify the classes of turnout support deterioration as shown in the table. In this study, a convolutional neural network (CNN) is used to develop the machine learning model. The data splitting is done in that 70% of data are used to train the model while another 30% are used to test the model. Hyperparameter tuning is also conducted to find the combination of hyperparameters that provide the best performance. A tech- nique used for hyperparameter tuning is grid search. The list of combinations of hyperparameters can be shown in Table 2.

Table 2: The list of tuned hyperparameters.

Tunned hyperparameters The number of convolutional layers Kernel The number of filters The number of pooling layers Pool size The proportion of dropout The number of dense layers The number of nodes Activation function Optimizer Batch size 4. RESULT AND DISCUSSION

From the finite element model development, different parameters are varied to create the data variation as mentioned in the previous section. Examples of axle box accelerations when turnout support stiffnesses are varied are shown in Figure 2. In the figure, the weight and speed of rolling stocks are 48 tons and 135 km/h respectively. However, the turnout support stiffnesses are varied from the minimum and maximum values which are 8.1 and 12.8 MN/m respectively. it can be seen that the lower stiffness results in higher vibration or accelerations. The figure shows the differences between the minimum and maximum which might be able to classify easily. However, when the stiffnesses are close, it is quite difficult to recognize so the use of a machine learning model will be useful. From the development of the machine learning model, the confusion matrix and model per- formance are shown in Table 3and Table 4 respectively.

i, orn inter.noise 21-24 AUGUST SCOTTISH EVENT CAMPUS ? O? ? GLASGOW

250

12.8 MN/m 8.1 MN/m

200

Axle box accelerations (m/s 2 )

150

100

50

0

-50

-100

-150

0.0 0.2 0.4 0.6 0.8 1.0 1.2

Time (s)

Figure 2: Example of axle box accelerations when turnout support stiffnesses are varied.

Table 3: Confusion matrix.

Classes Predicted classes 1 2 3 4

1 130 0 1 1 2 8 125 7 0 3 0 0 160 0 4 0 0 0 149 Table 4: Model performance. Classes Precision Recall F1-score Support 1 0.94 0.98 0.96 132 2 1.00 0.89 0.94 140 3 0.95 1.00 0.98 160 4 0.99 1.00 1.00 149 Accuracy 0.97 From Table 3and Table 4, it can be seen that the developed machine learning model can classify the turnout support deterioration with high reliability. For class 1 or perfect-condition turnout support, the model almost achieves the recall of 1.00. For classes 3 and 4 where the turnout support deterio- rations are severe, the recalls are 1.00 which demonstrates that the model can detect the severe dete- rioration completely. This is important to railway maintenance because the ability to detect severe deterioration is necessary. After all, the maintenance can be done promptly and minimize the damage from the severe deterioration. For class 2, the precision is 1.00 which demonstrates that it can be completely reliable when the model predicts the deterioration and the result is class 2. In addition, the performance of the model is supported by high precisions, recalls, and F1-score. The overall ac- curacy of the model is 0.97 which clearly demonstrates the high potential of the developed machine learning model to use the axle box acceleration to estimate the turnout support deterioration. As mentioned, the performance of the machine learning model is ensured by using hyperparameter tuning through grid search. The combination of hyperparameters that provide the best performance is shown in Table 5.

classes

Actual

i, orn inter.noise 21-24 AUGUST SCOTTISH EVENT CAMPUS ? O? ? GLASGOW

Table 5: Tunned hyperparameters. Hyperparameters Tuned values The number of convolutional layers 2 Kernel 1 The number of filters 64 (Conv1) and 32 (Conv2) The number of pooling layers 0 Pool size N/A The proportion of dropout 0.5 (Conv2) The number of dense layers 3 The number of nodes 100 (Dense1), 100 (Dense2), 1 (Dense3) Activation function Relu Optimizer Adam Batch size 64 5. CONCLUSION

This study aims to use axle box accelerations of rolling stocks to predict railway turnout support deterioration using the machine learning technique. Turnout support stiffnesses are used as indicators to identify the deterioration of turnout support. The numerical data are generated using the validated finite element models. the total number of samples is 1,936 which are separated into the proportion of 70/30 for training and testing the developed machine learning model respectively. Axle box accel- erations are exported from the finite element simulations and used to train the machine learning model. in this study, CNN is used to develop the machine learning model. From the development, the machine learning model can achieve 0.97 accuracies. From the performance of the model, the devel- oped machine learning model has the potential to use axle box accelerations to estimate the turnout support deterioration. The benefit of the study is that railway maintenance can be done more efficiently because the turnout support deterioration can be estimated more correctly and easily using the axle box accelerations from regular operations. In addition, the operations are likely to be less disturbed because unforeseen failures are prevented by continuous condition monitoring and measurement can be done with the regular operation speed. The future study is to include additional features to improve the machine learning model performance. Field data will be collected to train the developed machine learning model. 6. ACKNOWLEDGMENT

The authors also wish to thank the European Commission for the financial sponsorship of the H2020-RISE Project no.691135 “RISEN: Rail Infrastructure Systems Engineering Network”, which enables a global research network that addresses the grand challenge of railway infrastructure resili- ence and advanced sensing in extreme environments (www.risen2rail.eu). 7. REFERENCES

1. Kassa, E., C. Andersson, and J.C.O. Nielsen, "Simulation of dynamic interaction between train

and railway turnout ," Veh. Syst. Dyn. , 2006. 44 (3): p. 247-258. 2. Andersson, C. and T. Dahlberg, "Wheel/rail impacts at a railway turnout crossing ," Proc Inst

Mech Eng F J , 1998. 212 (2): p. 123-134.

i, orn inter.noise 21-24 AUGUST SCOTTISH EVENT CAMPUS ? O? ? GLASGOW

3. Dindar, S., S. Kaewunruen, and M. An, "Identification of appropriate risk analysis techniques for

railway turnout systems ," J. Risk Res. , 2018. 21 (8): p. 974-995. 4. Xin, L., V.L. Markine, and I.Y. Shevtsov, Analysis approach of turnout crossing performance by

field measurements and finite element modeling, in The Dynamics of Vehicles on Roads and Tracks . 2016, CRC Press. p. 1576-1583. 5. Ren, Z., S. Sun, and W. Zhai, "Study on lateral dynamic characteristics of vehicle/turnout

system ," Veh. Syst. Dyn. , 2005. 43 (4): p. 285-303. 6. Kaewunruen, S., R. You, and M. Ishida, "Composites for timber-replacement bearers in railway

switches and crossings ," Infrastructures , 2017. 2 (4): p. 13. 7. Kaewunruen, S., Acoustic and dynamic characteristics of a complex urban turnout using fibre-

reinforced foamed urethane (FFU) bearers, in Noise and Vibration Mitigation for Rail Transportation Systems . 2015, Springer. p. 377-384. 8. Zhou, L., H. Li, Y. Xu, X. Xiao, and X. Yao, Turnout Components Distribution Model and its

Application in Urban Rail Transit, in ICTE 2013: Safety, Speediness, Intelligence, Low-Carbon, Innovation . 2013. p. 2650-2653. 9. Min, B.-C., Y.-T. Heo, D.-Y. Hong, W.-J. Lee, S.-Y. Jo, and N.-C. Jeong. The Development of

Third-Rail System Applied to Turn-out Section for Urban Maglev. in Proceedings of the KSR Conference . 2011. Korea: The Korean Society for Railway. 10. Hamarat, M., M. Papaelias, and S. Kaewunruen, Vulnerability of railway switches and crossings

exposed to flooding conditions, in Resilient Infrastructure . 2022, Springer. p. 337-348. 11. Ishak, M.F., S. Dindar, and S. Kaewunruen. Safety-based maintenance for geometry restoration

of railway turnout systems in various operational environments. in The 21st National Convention on Civil Engineering . 2016. Songkhla, THAILAND. 12. Kaewunruen, S., "Monitoring structural deterioration of railway turnout systems via dynamic

wheel/rail interaction ," Case Stud. Nondestruct. Test. Evaluation , 2014. 1 : p. 19-24. 13. Sysyn, M., D. Gruen, U. Gerber, O. Nabochenko, and V. Kovalchuk, "Turnout monitoring with

vehicle based inertial measurements of operational trains: a machine learning approach ," Communications – Scientific Letters of the University of Zilina , 2019. 21 (1): p. 42-48. 14. Sresakoolchai, J. and S. Kaewunruen, "Wheel flat detection and severity classification using deep

learning techniques ," Insight: Non-Destr. Test. Cond. Monit. , 2021. 63 (7): p. 393-402. 15. Sresakoolchai, J. and S. Kaewunruen, "Detection and Severity Evaluation of Combined Rail

Defects Using Deep Learning ," Vibration , 2021. 4 (2). 16. Sresakoolchai, J. and S. Kaewunruen, "Railway defect detection based on track geometry using

supervised and unsupervised machine learning ," Struct. Health Monit. , 2021. 17. Kaewunruen, S., J. Sresakoolchai, and A. Thamba, "Machine learning-aided identification of train

weights from railway sleeper vibration ," Insight: Non-Destr. Test. Cond. Monit. , 2021. 63 (3): p. 151-159. 18. Yang, C., Y. Sun, C. Ladubec, and Y. Liu, "Developing machine learning-based models for

railway inspection ," Appl. Sci. , 2020. 11 (1): p. 13. 19. Krummenacher, G., C.S. Ong, S. Koller, S. Kobayashi, and J.M. Buhmann, "Wheel defect

detection with machine learning ," IEEE trans Intell Transp Syst 2017. 19 (4): p. 1176-1187. 20. Kaewunruen, S., J. Sresakoolchai, and H. Stittle, "Machine Learning to Identify Dynamic

Properties of Railway Track Components ," Int. J. Struct. Stab. Dyn. , 2022: p. 2250109. 21. Kaewunruen, S., J. Sresakoolchai, and G. Zhu, "Machine learning aided rail corrugation

monitoring for railway track maintenance ," Struct. Monit. Maint. , 2021. 8 (2): p. 151-166.

i, orn inter.noise 21-24 AUGUST SCOTTISH EVENT CAMPUS ? O? ? GLASGOW

22. Fink, O., E. Zio, and U. Weidmann, "Predicting time series of railway speed restrictions with

time-dependent machine learning techniques ," Expert Syst. Appl , 2013. 40 (15): p. 6033-6040. 23. Pineda-Jaramillo, J., P. Martínez-Fernández, I. Villalba-Sanchis, P. Salvador-Zuriaga, and R.

Insa-Franco, "Predicting the traction power of metropolitan railway lines using different machine learning models ," International Journal of Rail Transportation , 2021. 9 (5): p. 461-478. 24. Hamarat, M., S. Kaewunruen, M. Papaelias, and M. Silvast, "New insights from multibody

dynamic analyses of a turnout system under impact loads ," Appl. Sci. , 2019. 9 (19): p. 4080. 25. Kaewunruen, S. and T. Tang, "Idealisations of dynamic modelling for railway ballast in flood

conditions ," Appl. Sci. , 2019. 9 (9): p. 1785.

i, orn inter.noise 21-24 AUGUST SCOTTISH EVENT CAMPUS ? O? ? GLASGOW