1
Faculty of Behavioural, Management and Social Sciences, Industrial Engineering and Management
Data driven wheel set usage
prediction and inventory control
Koopman, M.P.
M.Sc. Thesis May 2021
Supervisors from University of Twente:
dr.ir. E. Topan dr. I. Seyran Topan Supervisor from Arriva:
M. Melenhorst Faculty of Behavioural, Management and Social Sciences University of Twente P.O. Box 217 7500 AE Enschede
Title: Data Driven Wheel Set Usage Prediction
Date: June 6, 2021
Author: Martijn Koopman
m.p.koopman@student.utwente.nl s1601326
Study Program: MSc Industrial Engineering and Management Supply Chain and Transportation Management
Faculty of Behavioural Managment and Social Sciences University of Twente
Examination Committee:
University of Twente Dr.ir. E. Topan Dr. I. Seyran Topan
Arriva NL M. Melenhorst
Management Summary
This thesis was written at the train workshop of Arriva in Zutphen. Arriva is a public transporta- tion company which is active in a range of countries over Europe. Two types of trains receive maintenance in the workshop, DMU- and EMU trains. These trains have three different wheel types which are called: DMU MWST, EMU MWST and LWST. Several critical parts are present in a train make it ready for the operation. One of these parts are the wheels. Arriva is handling high safety stocks to make sure that trains are not down waiting for parts. This causes non-moving spare parts stock and locked-up cash. Arriva is curious how to optimize these spare part levels.
Wheel set usage prediction for train wheels is challenging due to the degradation process of wheel sets. Wheel degradation is dependent on multiple factors such as seasonal effects and reprofiling.
Despite the complexity, these predictions are crucial in maintaining the installed base. Poor fore- casts can result in not meeting the timetable (in Dutch: Dienstregeling) of trains, due to a stock out. Whereas overstocking wheels causes non-moving stock and locked up cash. Both have a direct impact on the train operators’ profitability and have a direct impact on traveller satisfaction.
Due to the importance of meeting the timetable, Arriva wants to have more than enough stock on hand to satisfy expected demand during the lead time. Arriva wonders how they can use degradation data to forecast future demand for preventive maintenance and use this information for planning inventory supply. Furthermore, the goal is to design a model which can be used for future decision making, it should be able to update itself based on new measurement data. The uncertainty regarding corrective replacements should be included such that those risks are taken into account as well. This leads to the following research objective:
"Develop and validate an analytical tool that is able to predict degradation of wheels of the installed base and integrate it with spare part planning to determine the optimal spare parts levels."
Method
After analysing the current situation and investigating relevant literature, we developed a model that is able to make a connection between the degradation process of wheel sets and demand that arises from that. Predictions are made until the lead time length, which is assumed to be ten months. Combining the MLR and reprofiling will result in an expected diameter during and at the end of the lead time. Based on historic data, we create a Multiple Linear Regression (MLR) using predictors: time of year (monthly), SKU, position and number of reprofiling instances received until now, which is needed for the diameter prediction as well. The MLR equation is trained over the course of four years. Reprofiling is only considered when the preventive or expected reprofiling date falls within the lead time. The cutting depth or material removed during a reprofiling instance is modelled as a random process and is around 5 cm per instance. This is around 8-10% of the useful life depending on the SKU. When a wheel reaches the minimum diameter, it is fully degraded and thus need to be replaced. For each Stock Keeping Unit (SKU), wheels are divided into five classes based on the current diameter. Demand arising from degradation and reprofiling for a certain wheel class is formulated as a Binomial Distribution. Here we estimate the failure probability of a class for each month during the lead time. From this, we are able to obtain demand rates for new wheels during the lead time.
The forecast quality of the proposed method is evaluated by comparing the predicted diameter values of the trained MLR in combination with the cutting depths if applicable with the historic data diameter values over the last two years of the data set. The performance of the degradation predictions are evaluated by the Mean Average Percentage Error (MAPE) for the degradation of the diameter over the lead time. The demand process can be seen as a classification problem.
Either we predict a demand correctly during a given month or not. We use the measures accuracy and kappa to test the classification performance.
When the predictions of the model turn out to be accurate, we can conduct some further exper- iments to illustrate the value of knowing the current state of the installed base. We make a new
MLR equation which is trained over all available data and use the same reprofiling distributions as before. Let’s define the actual policy which uses the current state of the installed base and inventory decisions made from the demand rates arising from this current state. To access the per- formance of the actual policy, we create four other policies and make comparisons. We create these policies, since we are lacking data regarding replacements, order arrivals and inventory levels over time. Thus, we define an actual, pessimistic, optimistic, even and random policy. This produces the following overview:
Table 0.1: Policy descriptions.
Policy Description
Actual Policy that uses the exact class information Optimistic Policy that sets all wheels class I
Pessimistic Policy that sets all wheels in class V
Even Policy that divides wheels evenly over all classes Random Policy that divides wheels randomly over all classes
To clarify, these policies will make their own replenishment decisions based on their own demand expectations. The optimistic policy will not expect any demand and will thus not order any additional stock. However, the actual demand (from the actual policy) will generate demand and thus shortages will occur in the optimistic policy. With measures such as Fill Rate (FR), Average Spare Part Value (ASPV) and Total Relevant Costs (TRC) (ASPV+shortage costs) we will compare the different policies with each other. The shortage costs are e10,000 per day.
This will be done with different starting inventories and different target FRs. Additionally, we compare the obtained safety stock levels to the expected safety stock levels of Arriva and make some comparisons.
Results
As mentioned before, we will first need to test the validity of the proposed method. With an aggregated MAPE of around 4.1% we can say that this method is a good approximation of reality.
Next to this, yielding an accuracy of around 91% and a kappa of 55% aggregated over the SKUs, we can say that the conversion from a degradation rate to a demand rate performs functions as well. From this, we can also say that we are able to predict the expected demand over the lead time and review period. This means that we are able to continue by experimenting with the different policies.
The policies of Table 0.1 are tested with target Fill Rates of 0.95, 0.99 and 0.995 and different starting inventory positions. For a target Fill Rate of 0.995 aggregated over all six starting positions we come up with the following Table:
Table 0.2: Policy experiment overview, target Fill Rate 0.995.
Policy FR ASPV TRC
Actual .991 214 269
Optimistic .622 49.2 2814 Pessimistic .995 401.5 421.3
Even .912 178.7 1080
Random .942 197.8 599.2
From Table 0.2, we can conclude the following. The actual policy outperforms all other policies in terms of Fill Rate and TRC. The optimistic approach has a lower ASPV, but an unacceptable FR of around 62%, which results in a high TRC. The pessimistic approach reaches the target Fill Rate of 0.995, a 0.4% increase in Fill Rate with a TRC increase of 56.6%. The even policy does not perform well at all, this is mainly due to a high ratio on class IV and V for the actual policy, which resulted in shortages in the even policy. The random policy comes in terms of performance closest
to the actual policy. The main reason for this was that the class setup of the random policy was somewhat similar to the actual policy. Which is more than average amount of wheels in classes IV and V.
Lastly, we compare the obtained safety stock values with the approximated safety stock levels handled by Arriva (current system). For a target FR of 0.995, we obtain the following safety stock levels for the actual policy: (22, 20, 46) for the SKUs DMU MWST, EMU MWST and LWST respectively. The approximated safety stock levels handled by Arriva are: (34, 21, 64), resulting in an aggregated FR of 0.9959. To illustrate, a decrease of 30% of safety stock, will yield an aggregated fill rate loss of 0.0006 and a decrease of e130.000 in terms of ASPV. Thus, further increasing the safety stock from a FR of 0.995 and on wards will almost have no yield in terms of FR increase.
Recommendations
Based on the findings and limitations of the study, we make the following recommendations to Arriva:
• We advice Arriva to start tracking the mileage between maintenance instances. In this way, the precision of the MLR will increase, which benefits the forecasting performance.
• We recommend to start recording the re-order points and delivery dates of new wheels. This will give a better indication of the lead time. This data can be used to get a more precise indication of the lead time demand.
• Next to this, more attention should be given develop CBM approach on how the wheels interact with other bogie parts. Doing this can significantly decrease corrective maintenance costs and thus increase availability.
• We recommend that Arriva implements a time- or usage based maintenance strategy for their trains. Choosing this approach can contribute to the optimal level of risk versus maintenance costs given that the value lies within the constraints of the train maintenance manuals. We expect that such a maintenance strategy will also help motivating a choice of prolonging or shortening maintenance intervals.
• Future work can focus on the effect of having a variable lead time on the safety stock levels and the expected lead time demand.
Preface
It is my pleasure to present to you my master’s thesis, which means the end of almost seven years studying Industrial Engineering and Management at the University of Twente. This also means the end my period as a student, a time which I enjoyed thoroughly. I had the privilege of writing this thesis during my internship at Arriva. They provided me with an interesting challenge fitting my interests. I learned a lot during this period and it would not have been possible without the support, time and confidence of many people.
First of all, I would like to thank Marcel Melenhorst from Arriva for his supervision, explanations and insights. Furthermore, I would like to thank all colleagues of Arriva at location Zutphen, you made me feel welcome from the start.
Furthermore, I would like to thank Engin for his endless support and positivism. Even when progress was lacking, you were supportive and trying to motivate me to keep going. Also the discussions, proofreading and ideas were very helpful for the contents of the thesis. I also would like to thank Ipek for her second opinion and additional advice.
Lastly, I am grateful for all the support I have received from my friends and family. A very special thanks to my girlfriend, for putting up with me, her unconditional support and advise during my graduation period. For now, this thesis is the end of my student time, but I am more than ready to go beyond limits in the future!
Martijn June 2021
Contents
Management Summary i
Preface iv
Glossary vii
Acronyms vii
List of Figures ix
List of Tables x
1 Introduction 1
1.1 Arriva . . . . 1
1.2 Scientific Context . . . . 2
1.3 Problem Statement . . . . 3
1.4 Research Objective . . . . 3
1.4.1 Degradation Model . . . . 3
1.4.2 Inventory Model . . . . 4
1.5 Research Questions . . . . 4
2 Current Degradation, Maintenance and Inventory 6 2.1 Wheel set usage and degradation process . . . . 6
2.1.1 Train Layout . . . . 6
2.1.2 Bogie Parts . . . . 6
2.1.3 Wheel set degradation . . . . 7
2.1.4 Wheel Measures . . . . 9
2.1.5 MiniProf . . . . 11
2.1.6 Wheel-Rail Measurement Systems . . . . 12
2.2 Maintenance Policy . . . . 13
2.3 Inventory Policy . . . . 13
2.4 Cost Overview . . . . 15
3 Literature Review 16 3.1 Degradation Models . . . . 16
3.1.1 Gamma Process . . . . 16
3.1.2 Proportional Hazards model . . . . 17
3.1.3 Random Coefficient Model . . . . 17
3.1.4 Conclusion on Degradation process . . . . 18
3.2 Maintenance Strategies . . . . 18
3.2.1 Age- and Block Replacement . . . . 18
3.2.2 Control Limit Maintenance Policy . . . . 19
3.2.3 Conclusion on Maintenance Strategies . . . . 20
3.3 Prediction and Forecasting . . . . 21
3.3.1 Model Validation . . . . 21
3.3.2 Regression . . . . 22
3.3.3 Classification . . . . 22
3.3.4 Demand Forecasting . . . . 24
3.3.5 Advance Demand Information . . . . 25
3.3.6 Conclusion on Prediction Techniques . . . . 26
3.4 Inventory Policies . . . . 26
3.5 Spare Parts Management . . . . 27
3.5.1 Joint Maintenance- and Replenishment decisions . . . . 28
CONTENTS
4 Proposed method and experimental design 29
4.1 Data Import . . . . 29
4.2 Degradation Model . . . . 30
4.2.1 Multiple Linear Regression . . . . 30
4.2.2 Reprofiling . . . . 31
4.3 Demand Process . . . . 33
4.3.1 Method Validation . . . . 35
4.4 Inventory Model . . . . 36
4.5 Experimental Design . . . . 37
5 Experimental Results 39 5.1 Validating the method . . . . 39
5.1.1 Finding the best Multiple Regression- and Reprofiling factors . . . . 39
5.1.2 Performance of the method . . . . 41
5.1.3 Conclusion on method validation . . . . 41
5.2 Experimenting with scenarios . . . . 42
5.2.1 Policy Performance . . . . 42
5.3 Conclusion . . . . 44
6 Conclusion and Recommendations 46 6.1 Conclusion . . . . 46
6.1.1 Scientific Contribution . . . . 47
6.1.2 Limitations . . . . 47
6.2 Recommendations . . . . 48
6.3 Directions for future research . . . . 48
References 50 Appendices 53 A Example of Measurement data sheet 53 B Example of reprofiling sheet 54 C Reprofiling distributions 55 D Predictor factor table 56 E Experiment Tables 57 F Model Manual (in Dutch) 59 F.1 Uploaden nieuwe data . . . . 59
F.2 Treinen toevoegen of verwijderen . . . . 62
F.3 Inzoomen op één trein . . . . 62
F.4 Voorraad . . . . 64
Glossary
Bernoulli Distribution Discrete probability distribution of a random variable which takes value 1 with probability p and 0 with probability q=1-p. 34
Binomial Distribution A Bernoulli experiment that is executed multiple times.. i, 25, 34–36, 38, 41, 45–47
Bochumer Verein OEM of GTW train wheels located in Bochum (Germany). 14 Bonatrans OEM of LINT train wheels located in Bohumín (Czech Republic). 14
Cycle Service Level Expected probability of not hitting a stock out during the next replenish- ment cycle.
Deutsche Bahn German railway mother company of Arriva. 1 Fill Rate Fraction of demand directly full filled from stock.
Gotcha Wheel measurement system placed on rails throughout the Netherlands. 12, 13 Infor ERP system used by Arriva. 14
MiniProf Device used to measure the wheel surface. ix, 6, 11, 12, 29 Mobiturn Device that is used to reprofile train wheels. ix, 10
NP-hard An optimisation problem that cannot be solved in Polynomial Time.
Polynomial Time An algorithm feature if for some k>0 the running time of size n is Onk. vii ProRail Government task organisation that takes care of the national railway network infrastruc-
ture. vii, 9, 12
Remaining Useful Life The time left before failure occurs given the current condition of the unit and the past and current operation profile.
Sd Flange diameter measured from the inner wheel side until the flange foot. 10, 11 Sh Flange height measured from the flange foot until the flange top. 10, 11
Stadler German manufacturer of the DMU and EMU trains used by Arriva. 1, 2
Strukton Rail Dutch railway maintenance company which is a subcontractor of ProRail. 1, 10, 32
Trubeka Wheel set wholesaler located in Hoorn (the Netherlands). 14, 15
Acronyms
Acronyms
ADI Advance Demand Information. 25, 26,
ASPV Average Spare Part Value. ii, 27, 38, 42–44, 46, CBM Condition Based Maintenance. iii, 19, 20, 48, CSL Cycle Service Level. 27, 36,
DMU Diesel Multiple Unit. iii, vii, 1, 2, 6, 9, 10, 15, 32, 33, 42, EBO Expected Backorders. 4, 27, 36, 37,
EMU Electric Multiple Unit. iii, vii, 1, 2, 6, 9, 10, 12, 15, 29, 33, ERP Enterprise Resource Planning. 14,
FR Fill Rate. ii, iii, 27, 36–38, 42–47, 49, GTW Gelenktriebwagen. vii, 1, 2, 6, 9, 11, 14, HW Holt-Winters. 24, 26,
ILT Inspectie Leefomgeving en Transport. 10, IPO Initial Public Offering. 1,
KPI Key Performance Indicator. 4, 21, 27, 35, 36, 38, 46,
LINT Leichter Innovativer Nahverkehrstriebwagen. vii, 1, 2, 6, 9–11, 14, 29, LTO Long Term Overhaul. 13,
MAE Mean Absolute Error. 22,
MAPE Mean Average Percentage Error. i, ii, 21, 22, 26, 35, 38, 41, 46, MLR Multiple Linear Regression. i–iii, 18, 22, 30, 31, 35, 37–42, 46–49, MSE Mean Squared Error. 21, 22, 24,
OEM Original Equipment Manufacturer. vii, 8, 10, 14, 48, PDF Probabilty Density Function. 16, 18,
PH Proportional Hazards. 17,
RCF Rolling Contact Fatigue. 2, 7–9, 16, RMSE Root Mean Square Error. 22, 35, 38, 46, RSU Rolling Stock Unit. 2, 6, 29,
RUL Remaining Useful Life. ix, 3, 9, 16, 17,
SKU Stock Keeping Unit. i–iii, ix, 14, 15, 24–27, 30–42, 46, sMAPE Symmetric Mean Average Percentage Error. 22, 26, 41, STO Short Term Overhaul. 13,
TRC Total Relevant Costs. ii, 44,
VBA Visual Basic for Applications. 38, 40,
LIST OF TABLES
List of Figures
1.1 Maintenance site in Zutphen. . . . 1
1.2 Different trains maintained in Zutphen. . . . . 2
1.3 Problem Cluster. . . . 4
2.1 Schematic overview per train type. . . . 6
2.2 Bogie parts. . . . 7
2.3 Wheel overviews. . . . 8
2.4 Mobiturn. . . . 10
2.5 Wheel profile measures. . . . 11
2.6 MiniProf figures. . . . 12
2.7 Part of a data sheet. . . . 12
2.8 Maintenance Policy overview. . . . 13
2.9 Inventory Policy overview. . . . 14
3.1 Taxomony of statistical data driven approaches for the RUL estimation (Si, Wang, Hu, & Zhou, 2011) . . . . 16
3.2 Cost Function Curve. . . . 19
3.3 Control Limit Maintenance Policy (Topan, 2018). . . . 20
3.4 The standard logistic function . . . . 23
4.1 Flow chart of modelling process. . . . . 29
4.2 Random process with cutting depth probabilities of a DMU engine wheel. . . . 32
4.3 Example of the degradation process of an individual LWST wheel. . . . 33
4.4 Example of an inventory process chart of a wheel SKU. . . . . 36
5.1 Fill rate vs ASPV with start inventory I and target FR 0.95 . . . . 43
C.1 Reprofiling Distributions. . . . 55
F.1 Dashboard van het model. . . . 59
F.2 Tekst naar kolommen interface. . . . 60
F.3 File picker voor upload. . . . 60
F.4 Foutmelding verkeerde ascode. . . . . 61
F.5 Foutmelding verkeerde diameter waarde. . . . 61
F.6 Archief van bestanden. . . . 61
F.7 Parameter sheet. . . . 62
F.8 Interfaces bij het toevoegen van een trein. . . . 62
F.9 Installed base sheet. . . . 63
F.10 Overzicht sheet van een individuele trein. . . . . 63
F.11 Slijtage process van een individueel wiel. . . . 64
F.12 Overzicht sheet van verschillende SKUs. . . . 64
F.13 Draaitabel van de status van een SKU. . . . 65
List of Tables 0.1 Policy descriptions. . . . ii
0.2 Policy experiment overview, target Fill Rate 0.995. . . . ii
1.1 Installed base information. . . . 2
2.1 Wheel set threshold information. . . . 10
2.2 Reprofiling measures. . . . 11
2.3 Wheel price information. . . . 15
3.1 Confusion Matrix for a two-class problem . . . . 23
3.2 Inventory Policies. . . . 26
4.1 Wheel replacement conditions. . . . . 33
4.2 Wheel classes. . . . 34
4.3 Starting inventory positions. . . . 38
4.4 Experiment Overview. . . . 38
5.1 Reprofiling cutting depths example. . . . . 39
5.2 Multiple linear regression factors for testing. . . . 40
5.3 Performance metrics of method testing. . . . . 41
5.4 Performance metrics of classification. . . . 41
5.5 Exact distribution of conditions of wheels. . . . 42
5.6 Results experiments, target fill rate 0.95. . . . 42
5.7 Results experiments, target fill rate 0.99. . . . 43
5.8 Results experiments, target fill rate 0.995. . . . 43
5.9 Results experiments, TRC with a target fill rate of 0.99. . . . 44
5.10 Results experiments, TRC with target fill rate of 0.995. . . . . 44
5.11 Safety stock levels of actual policy. . . . 44
D.1 Predictor factors of MLR. . . . 56
E.1 Random class policy. . . . 57
E.2 Demand rate of EMU MWST per policy. . . . 57
E.3 Demand rate of DMU MWST per policy. . . . . 58
E.4 Demand rate of LWST per policy. . . . . 58
1 Introduction
This thesis is written for the completion of the Master Industrial Engineering and Management within the Supply Chain and Transportation Management track at the University of Twente. In this context, research has been conducted at an Arriva train workshop located in Zutphen. This research is conducted at the technical department of Arriva. The aim of this research is to forecast wheel degradation of the trains in combination with determining a spare parts level needed to replace worn out wheels.
This chapter introduces the research carried out at Arriva. Section 1.1 gives background informa- tion about Arriva and some specifics about the workshop in Zutphen and the trains maintained.
Section 1.2 gives the scientific context about the research subject. From this, a problem statement is formed in Section 1.3. Following the problem statement, a research objective and -questions are elaborated in Sections 1.4 and 1.5.
1.1 Arriva
Arriva is a multinational public transportation company. It was founded in 1938 as T Cowie and was rebranded to Arriva in 1997. Its headquarters is located in Sutherland, England, whereas the Dutch headquarters is located in Heerenveen. Arriva operates in 14 countries in Europe, such as England, Poland and the Netherlands. Around 55,000 people were employed by Arriva in September 2018 (Arriva, n.d.). Next to this, Arriva operates 17,000 busses and 1,100 trains. In 2010, Arriva was acquired by Deutsche Bahn. Due to high infrastructure and train investments in 2018 pushed Deutsche Bahn’s debt close toe20 billion (Topham, 2019). In order to reduce its debt, Deutsche Bahn opted to sell Arriva, however no suitable offers came in. As an alternative, Deutsche Bahn has plans for an Initial Public Offering (IPO), however it is certain that this will not happen in the near future due to the amid concerning COVID-19 (Gazzette, 2017).
In October 2012, Arriva and Strukton Rail moved into a new workshop close to the train station in Zutphen (van Gompel, 2012). Before this, Arriva outsourced all maintenance activities for their trains in this region. The goal of starting this site was to increase availability and lower the cost of maintenance. By increasing the number of maintenance activities performed by Arriva throughout the years, more operational gains were realised. From Figure 1.1a, we can see that five tracks are present at this site. Three of which are used by Strukton Rail and two are used by Arriva. The two tracks used by Arriva can be seen in Figure 1.1b. The track on the right side is called the E-track.
Engineers can walk beneath and on top of the train on this track and perform various maintenance activities. The track on the left side of Figure 1.1b is called the D-track. On this track, a train can be lifted up with the help of the yellow beams on the side of the track. When the train is lifted, engineers can perform maintenance tasks as well underneath the train. In comparison to the D-track, additional activities can be performed on this track, such as reprofiling the wheels and replacing bogies.
(a) Outside view. (b) Inside view.
Figure 1.1: Maintenance site in Zutphen.
On this maintenance site in Zutphen, Arriva is able to maintain 42 trains, which can be divided into three types: 24 Stadler GTW-DMU, 14 Stadler GTW-EMU and 4 LINT trains. The different types are illustrated in Figure 1.2.
1 INTRODUCTION
(a) Stadler GTW-DMU. (b) Stadler GTW-EMU. (c) LINT.
Figure 1.2: Different trains maintained in Zutphen.
These trains are all operational in the region of Zutphen. The trains maintained in Zutphen operate on the following tracks:
• Zwolle-Emmen
• Arnhem-Winterswijk
• Zutphen-Apeldoorn
• Zutphen-Winterswijk
• Almelo-Hardenberg
• Arnhem-Tiel
The LINT trains are only active on the Almelo-Hardenberg track, whereas the EMU trains are only active between Zwolle and Emmen. The remaining four tracks are covered by the Stadler GTW-DMU trains.
However, the GTW trains can differ in length. The length of a Rolling Stock Unit (RSU) or train is determined by the number of axis present in the train. A DMU or EMU can have six or eight axis. This difference is indicated by 2/6 or 2/8, where 2/6 stands for 2 out of 6 axis are powered.
The same holds for 2/8. A LINT has 6 axis, 4 of which are powered (4/6). An overview of the installed base is given in the Table 1.1. The RSU number is used as identification for the trains by Arriva.
Table 1.1: Installed base information.
Name Type # of trains RSU number
DMU 2/6 13 252-264
2/8 11 365-375
EMU 2/6 3 411-413
2/8 12 514-525
LINT 4/6 4 33,34,37
1.2 Scientific Context
The maintenance and renewal activities of wheel sets account for a large proportion of the whole-life costs for railway rolling stock. These activities are influenced by a large number of factors including depot constraints, wheel surface damage, fleet availability and vehicle design. If these factors are not managed efficiently it can have significant implications on a vehicle’s service provision, track damage, environmental and whole-life costs (Bevan, Molyneux-Berry, Mills, Rhodes, & Ling, 2013a). Next to the significant costs for the system, wheel set failures can in an extreme case lead to train derailment, which has obvious consequences for the train condition, but also for the surrounding infrastructure (Brabie & Andersson, 2006). Combining these two leads to the conclusion that wheel set management is essential for effective train maintenance.
In order to execute effective wheel set life cycle management, it is essential to know the relationship between the degradation process and the reliability of the wheel set (Freitas, de Toledo, Colosimo,
& Pires, 2009). Wheel degradation occurs in multiple forms such as crack formation, ovality and flat spots (Ekberg & Kabo, 2005). An important factor that influences the wheel degradation is the wheel-rail interaction which causes Rolling Contact Fatigue (RCF). Extensive research has been conducted on this subject (Johansson & Andersson, 2005; Sladkowski & Sitarz, 2005). This
1 INTRODUCTION
also applies to the wheel-rail interaction on crossings and switches (Wei, Shen, Li, & Dollevoet, 2017; Nicklisch, Kassa, Nielsen, Ekh, & Iwnicki, 2010). Other factors that could influence the wheel degradation are (extreme) climates (Asplund et al., 2014), mileage and wheel slipping of a train (Ekberg & Kabo, 2005).
From the paragraphs above, it can be concluded that extensive research has been conducted on the behaviour of wheel degradation in several contexts. However, no research can be found on the link between inventory control and predicting wheel set degradation. This highlights the relevance and importance of this research. In the next section, we zoom in on how Arriva copes with the wheel set degradation of their installed base in Zutphen.
1.3 Problem Statement
From section 1.2, we conclude that the wheels are a critical part of a train. Maintaining them well is crucial for passenger nuisance and train reliability. Therefore, we need more insight on the current- and future state of the wheels. An analysis of the current situation at Arriva Zutphen is made in Section 2.2 to indicate how they cope with wheel degradation. Wheel set degradation has an impact on the wheel diameter. Scratches, ovality and flat spots cause the need of reprofiling (Lin, Asplund,
& Nordmark, 2015). There are two cases of reprofiling, either corrective or preventive. Preventive reprofiling can be seen as a regular reprofiling instance, where the wheel is reprofiled after a fixed amount of kilometers driven. Material is removed from the wheel, clearing any inequalities on the surface. After receiving reprofiling, the wheel set can be considered as good as new. Doing this results in an average wheel set life of 48% and an average decrease in unplanned maintenance costs of 45% (Bevan, Molyneux-Berry, Mills, Rhodes, & Ling, 2013b). Arriva already uses this approach for the wheel set maintenance of the trains in Zutphen. Corrective reprofiling is needed when severe wear outs occur on the wheel surface. When this happens, the relevant wheel set needs to receive reprofiling quickly in order to mitigate the damage on the wheels. In terms of wheel set replacements, a wheel needs to have a minimum diameter, when this is reached, both wheels on the axis will be discarded and new wheels are installed. Currently, expected degradation is considered as a linear function over the kilometers driven. Using this, the number of wheels to be replaced within the next lead time cycle are determined.
From past experience, the planner of Arriva has found out that ordering the expected amount of new wheels is insufficient and resulted in wheel shortages and thus reduced train availability. In order to tackle this problem, the planner decided to order with a margin on top of the expected demand. Since wheel shortages inevitably result in downtime, which is considered expensive, a rather high margin order margin of around 50-60% is added to the expected wheel demand. This causes the reorder point to come sooner than actually needed. Another result of this is non-moving stock, which can be considered as locked up cash. It is expected that inventory can be reduced when a reliable prediction of wheel set demand is made. The demand of wheel sets is twofold:
The first part is due to worn out wheels, the second one is due to bogie replacements. When this happens, wheels with Remaining Useful Life (RUL) could be discarded due to high replacement costs. The decision between discarding or re-installing (axis including wheels is removed from train A and installed under train B) these wheels is currently based on past experience. A standardized decision support including cost references is needed.
1.4 Research Objective
The objective of this research is to develop a forecasting tool to support decision making for spare part planners of Arriva. The tool should consist of two parts: A degradation based model and an inventory or spare parts model using the the degradation model as input. We will not focus on an optimal replacement strategy or decision, but we will use a degradation model to predict the future replacements. Together with the tool, a user manual will be developed and recommendations will be made to Arriva with clear steps to follow up on this research.
1.4.1 Degradation Model
The first model should forecast a distribution of expected demand during the lead time of new wheels. The expected demand during the lead time should be based on the current state of the
1 INTRODUCTION
installed base and estimating possible factors that can influence the wheel degradation. Associated parts or incidents could also have an effect on the total demand of new wheels, so this should be taken into account as well. The performance or accuracy of the model should be tested against a number of KPIs, which will be introduced in chapter 4. Wheel measurement data from the past will be used to train, validate and test this model.
1.4.2 Inventory Model
Second, once we assure that the degradation model is accurate, an expected amount of wheels during a cycle can be determined with some confidence. This can have a big impact on wheel set spare part levels. Modelling degradation of wheels and better predicting replacement times can explain some of the uncertainty of the spare part demand and this can eventually lower spare part inventories. This would mean more space in the warehouse and gains in terms of inventory value reduction. When ordering, there are a couple of trade-offs, since ordering more would mean less Expected Backorders (EBO), but higher inventory costs. It is up to the planner to make a decision between the two. The planner should also be able to change certain parameters to make a founded decision. In the context of decision support, the inventory model could give a recommendation on an optimal order reorder point.
Summarizing all this information into one problem cluster gives the following overview:
Lack of insight in expected demand during lead time of
wheel sets
Lack of insight in the expected diameter decrease of a wheel set during lead time No clear inventory
policy regarding the wheel sets No predetermined
safety stock level for the SKUs
No expectations regarding cutting depth of reprofiling are considered No data available regarding corrective
replacements
Figure 1.3: Problem Cluster.
Combining the two aspects and Figure lead to the following research objective:
"Develop and validate an analytical tool that is able to predict degradation of wheels of the installed base and integrate it with spare part planning to determine the optimal spare parts levels."
We divide the objective into several parts in order to reach it. For each part, we define one or possibly more research questions. These questions are stated in section 1.5.
1.5 Research Questions
The following research questions are defined to answer the main research question. First, we would like to get familiar with the current practises of Arriva in terms of maintenance- and inventory policies.
1. What does the degradation regarding wheel sets at Arriva look like?
(a) Which factors can influence wheel degradation?
(b) What is the current maintenance policy for the wheels?
(c) What is the current inventory policy for the wheels and associated parts?
(d) What are the costs associated with the current policies?
Second, we would like to familiarize ourselves with research conducted relating to this topic.
For this, we would like to investigate literature related to our research.
1 INTRODUCTION
2. What can be found in literature about the research objective?
(a) Which factors can influence wheel degradation?
(b) Which modelling approaches for wheel set degradation are present in the literature?
(c) What is known about the relationship between predictive maintenance and spare part levels?
(d) What metrics can be used to evaluate the performance of the prediction- and inventory model?
After the literature review, we would like to combine the findings of the current situation and the literature review to come up with a modelling approach. This modelling approach should fit the research objective stated in Section 1.4.
3. What modelling approach do we propose to map the degradation process?
(a) What steps are taken within the models?
The steps needed are as follows:
i) Make predictions of wheel set degradation based upon historical data. Use this information to ii) make replenishment of spare part decisions.
(b) How should the (statistical) performance of both models be evaluated?
(c) How do the models fit the current situation?
Lastly, we would like to validate and test the performance of the degradation model with historic data. This will be done in three steps: First the degradation model will be trained with a part of the data set. Subsequently, other parts of the data set will be used to validate and test the degradation model. The prediction of the degradation model can be used in the inventory model, when the results of the degradation model are promising. Following from this, the inventory model is made based upon the current and possibly other maintenance policies. The inventory model does not need to be tested against some historic data, since it does not provide a forecast.
4. What is the potential value of using wheel set degradation predictions to make replenishment decisions for spare parts?
(a) What is the (statistical) performance of the degradation- and inventory model?
(b) What are the possible savings of implementing the proposed methods from the models?
(c) What are the limitations of implementing the models in the current situation?
(d) Which future steps are needed to fully implement the model into the current situation?
Answering these research questions should collectively lead to the answer on the research objec- tive. The remainder of this thesis is structured as follows: In Chapter 2, the current situation with regard to a maintenance- and inventory policy is analysed. In Chapter 3, we review litera- ture with regard to data-driven inventory control, factors influencing wheel set degradation and environment dependent simulation or prediction techniques. Subsequently, we propose a modelling approach in Chapter 4. From this, the results are discussed in Chapter 5. Lastly, conclusions and recommendations to Arriva are provided in Chapter 6.
2 Current Degradation, Maintenance and Inventory
In this chapter, the current situation with regard to wheel set maintenance and inventory man- agement is explored. First, a closer look to the wheel set degradation is needed to understand the current maintenance policy better. This is elaborated in section 2.1. Following this, the current maintenance policy is investigated. Subsequently, the inventory policy with regard to wheel sets is elaborated in section 2.3. At last in section 2.4, a cost overview of the current situation with the used policies is given.
2.1 Wheel set usage and degradation process
In this subsection, we will first elaborate on the different train layouts in section 2.1.1. After this, a more concise picture of all the parts present in a bogie are described in section 2.1.2. Furthermore, a deeper understanding of wheel set degradation and its wear outs is needed. This is explained in section 2.1.3. Next to this, in section 2.1.4 the current wheel measures and thresholds are set out in section 2.1.4. Lastly, in sections 2.1.5 and 2.1.6 two measurement systems are elaborated: The MiniProf and Gotcha respectively.
2.1.1 Train Layout
Most of the trains maintained by Arriva in Zutphen have different layouts. Three train layouts are displayed in Figure 2.1. The numbers underneath the correspond to the numbering handled by Arriva. The train direction or the arrow in Figure 2.1 determines the left and right side of the train. In case of Figure 2.1, the bottom wheels are left and the upper are right.
Figure 2.1: Schematic overview per train type.
Each axle has a unique identifiable number. Two axles together form a bogie. Indicated with letters A, B, C or D in Figure 2.1. The A bogie is always the first one in the driving direction of the train, whereas the B bogie is the last one. As shortly mentioned in Section 1.1, some bogies are powered by an engine or driving and some are not. A driving bogie can be identified by the double connection between the wheels in Figure 2.1. Driving bogies are much more complex than trailing ones, since connections to the traction motor and brake discs are present. This also means that replacing an axis connected to a driving bogie takes more time than replacing a trailing one.
For the DMU and EMU trains, the C bogie is always powered, whereas the rest of the bogies are trailing. In addition, the B bogie has magnetic brakes, whereas the A and optional D bogie do not have this. In Section 1.1, the terminology regarding the train layouts was shortly discussed. From the layout shown above, it can be concluded that the GTW 2/8 is the same as a 2/6 with an extra (D) bogie placed between bogies A and C. In the past, there was demand for larger RSUs, which resulted in this train layout. In the LINT, bogies A and B are powered, whereas the C bogie is trailing.
2.1.2 Bogie Parts
As mentioned in the previous section, there exist different bogies per train type, namely trailing and driving bogies. These bogies are complex and are therefore expensive to replace. An illustration of a typical bogie can be found in Figure 2.2. A bogie consists of several key parts: four wheels,
2 CURRENT DEGRADATION, MAINTENANCE AND INVENTORY
two axis, two breaks and four bearings. The driving bogies are more costly to replace than trailing bogies due to the fact that they are connected to a traction engine. These connections need to be removed before the bogie can be replaced safely.
Figure 2.2: Bogie parts.
2.1.3 Wheel set degradation
Before we can understand how a train wheel degrades more information is needed about the dimensions. An overview of a train wheel can be found in Figure 2.3b. A train wheel has some typical characteristics and consists roughly of four parts: The flange, the flange root, the wheel thread and the field side.
One can imagine that due to the steel to steel interaction, the wheels as well as the track degrade.
The difference between a new and a worn wheel is shown in Figure 2.3. In the ideal situation, there is only one interaction point between the wheel and the rail. All the energy generated from the engine create a vertical force which presses upon the interaction point with the rails. The size of this point can be compared with a fingernail. To clarify, large forces are put upon a small surface when the train is accelerating. This frequently causes slipping and sliding of the wheels on the track. Approximately the same holds for braking. A train driver can brake in two ways: Dynamic- and mechanical braking. Dynamic braking is using the traction motor as a generator to slow down the train. This also means, less sliding over the rails. Thus, dynamic braking has a lower impact on the wheel and brake wear, but a higher impact on traction motor wear. Mechanical braking can be compared to the usual braking process. Bogies connected to a traction motor have brake discs, which can block the rolling motion of the wheels. Like the accelerating process, mechanical breaking can cause slipping and sliding of the wheel over the track. Next to the brake discs, trains also have magnetic brakes. These can be activated when heavy breaking is needed. When this happens, the brakes are lowered on the track and the magnetic force between the brake and the rail forces the train to slow down.
When the wheel degrades, several wear can start to exist. These wear outs are explained later in this section. But first, we need to understand the different parts of a wheel to fully understand where these wear outs occur on the wheel. The first part is the wheel flange. This is the part of the wheel that is next to the track. It is located on the inside of the wheel and makes sure that the train stays on the track. The wheel flange can be seen on the left side of Figure 2.3b. A train typically does not move in a straight line over the track. There is often some difference between the distance between two wheels on an axis and the distance between the two rails. As a consequence, a train touches the flange face on each side which causes a Z-like direction. Over time this will decrease the width and increases in height of the flange. Next to the flange, the wheel tread also degrades over time. The degradation of the wheel surface can express itself in different wear outs (Cantini, Cervello, & Gallo, 2016). Most of these wear outs are illustrated in Figure 2.3b. All these wear outs are a result of RCF between rails and the wheels.
• Hollow wear: Concentrated wear in the centre of the thread.
• Flat spots: Singularity in wheel tread circle related to slip or skid movement between wheel set and rail. Often an effect of breaking faults.
2 CURRENT DEGRADATION, MAINTENANCE AND INVENTORY
(a) Wheel-rail interaction. (b) Worn and unworn wheel (Coenraad et al., 2006).
Figure 2.3: Wheel overviews.
• Shelling: Important loss of material.
• Flange wear: Is described as a decrease of the flange thickness and an increase of flange height.
• False flange: A groove worn into the tread of the wheel.
• Ovality: Unroundness of the wheel, this is often a consequence of flat spots.
These failure modes are visible on the surface of a wheel. They will be only noticed when the wheels are inspected in the workshop or on one of the wheel measurement systems throughout the Netherlands. Next to visible failure modes, hidden- or unexpected failures can occur as well. A hidden failure may not cause the system to break down, but may cause performance loss during operation (Liu, Yeh, Xie, & Kuo, 2017). In this context, a hidden failure can be expressed as a sudden loss of material due to chipping. The cause of this can be a manufacturing defect, which is often an air bubble in the wheel. This bubble can cause loss of material on the wheel surface upon which the wheel fails immediately. When this happens, a train will have a higher probability to further damage the wheel and other wheels or even derail. Reprofiling will often not be carried out in this context, since the wheel can be send back to the OEM it can be seen as a manufacturing mistake in this context. Next to a manufacturing defect, chipping can also be caused by a worn out wheel. In this context, corrective reprofiling needs to be carried out. Since all these wear outs are caused by RCF, this will be the main focus of this study.
Multiple factors can have a negative impact on the RCF, or wheel set degradation:
• Rail quality. Several studies implicate that worn out rails and crossings have an increased impact on the degradation of the wheels itself (Sainz-Aja et al., 2020). Worn out rails can cause increased "searching" for the track by the train wheels as is explained above. This can even increase when there is a high number of turns across the track.
• The load of the rolling stock can influence the degradation as well (Kuka, Verardi, Ariaudo,
& Pombo, 2018). The study was carried out on a heavy haul train, so it is less relevant in this problem context.
• Mileage of the wheels. As in many problem contexts operational time is a significant factor of degradation.
• Wheel quality can also influence the degradation. Wheel sets are typically made of steel, whereas there is a significant trade-off between the stiffness and the ability of the wheel to deform or to cope with sudden impacts (Mädler & Bannasch, 2006). This paper also considers different types of steel and compares them with each other.
• Sudden slipping and sliding of the wheels on the track. This can be caused by two phenomena.
Heavy breaking of the train driver is one of them. As mentioned above, a train can be put to standstill in two ways: Dynamic- and mechanical breaking. Whereas the latter one should be avoided as much as possible to decrease RCF (Tran, Ang, Luong, & Dai, 2016). The other phenomenon is the operating conditions. One can imagine that certain weather conditions can increase slipping and sliding of the train on the track. Operating conditions which have
2 CURRENT DEGRADATION, MAINTENANCE AND INVENTORY
a negative influence on the friction is when objects such as leaves are on the track and harsh winter or hot conditions, which can cause deformation of the rails. It has been proven that slipping and sliding depends on the so called coefficient of friction, which can be expressed by the following formula (Lundmark, 2007):
µ= M ∗1000
Fn∗d2 (2.1)
where,
µ is the coefficient of friction (-) M is the torque (Nm)
Fn is the Normal Force (N) d is the diameter (mm)
One would like to have a low coefficient of friction, since this implicates a low probability of slipping and sliding and this lower wheel set wear. Operating conditions which might significantly lower the Normal Force, are leaves on the track and drizzle (Baker, 2018). This mainly happens in the fall, which suggests an increase in wear during this time of year.
In contrary to the negative impacts of the above mentioned factors, there is one factor which can decrease the RCF. Several studies have been carried out in the field of using lubricants to decrease this (Everitt & Alfredsson, 2020; Borgaonkar & Syed, 2020). These lubricants are sprayed from underneath the train on the rails. The goal of this is to decrease slipping of the wheels on the rails when accelerating or braking. Another function of the lubricant could be sound reduction (Nam, Baek, Do, & Kang, 2017). It has been proven that using a variety of lubricants can reduce wheel as well as rail wear. However there is one significant setback of using lubricants, namely the environmental damage around the track (Milovanova & Egorova, 2020). A study on this subject has been carried out on three EMU trains of Arriva on the Zwolle-Emmen track. This was a temporary study, which lasted approximately a year. Arriva is still negotiating with the province, ProRail and other parties to see if these trains can be operational with these lubricants permanently. Since there is no decision on this yet, the positive and thus temporary effects of these lubricants will be out of the scope of this study.
Next to the wheels, these other parts degrade as well. Crack forming can occur in an axis which can have serious consequences. Brakes can loose functionality when worn out and bearings can overheat. All these wear outs can have an effect on the wheel set demand. It could be the case that any of the above mentioned parts have been worn out so that replacement of this part is needed.
This can cause an increase in wheel set demand, because the choice can be made to replace the current wheels as well with these maintenance action. In other words, wheel sets which have RUL left can be replaced due to high replacement costs of bogies. We need to make an assumption regarding this situation, since these replacements will create demand, either sooner than expected or the axis is removed and placed underneath another train. This will be taken into account and will be explained further in section 4.4. The costs associated to this are further elaborated in section 2.4.
2.1.4 Wheel Measures
As mentioned throughout this report, the diameter of train wheels decrease over time. First, we will discuss which different diameters are handled for the installed base. Adding to this, certain rules need to be met to ensure a safe environment for passengers and staff. For this end, reprofiling is used reshape the wheels, which has some requirements. The specifics about reprofiling are discusses after the different diameters of the installed base.
For GTWs, driving wheel sets have a larger diameter than trailing ones. In addition, driving wheels of DMUs have a larger diameter then driving wheels of an EMU. For LINT trains, the same wheels are used for driving- and trailing wheels. As mentioned before, over time, the wheel sets degrade and need to be reprofiled to remove inefficiencies from the wheel surface. There are however certain thresholds which need to be followed which are shown in Tables 2.1 and 2.2. A wheel set needs to be replaced when the minimum threshold is reached. These thresholds are determined by the
2 CURRENT DEGRADATION, MAINTENANCE AND INVENTORY
OEM in collaboration with Inspectie Leefomgeving en Transport (ILT), which is the supervisor of the Dutch Government regarding environment and transport.
Table 2.1: Wheel set threshold information.
Name Bogie New ∅ (mm) Min ∅ (mm)
DMU/EMU A-B-(D) 750 700
DMU C 870 810
EMU C 860 810
LINT A-B-C 770 710
Reprofiling of the wheels is carried out in the workshop in Zutphen on the D-track. This is the left track in Figure 1.1b. Since 2013, Strukton Rail has got a Mobiturn device to its proposal in Zutphen. This device is able to reprofile the wheels. The Mobiturn is placed underneath the lifted train on the D-track. Some kind of chisel is placed against the wheel surface. Subsequently, the wheel will start rotating. The chisel will remove material from the wheel and adds some kind of profile to the wheel set. A certain amount of material is removed per rotation. A rotation can be seen as scraping the whole wheel once with the chisel. Material is removed until there are no more inefficiencies present on the wheel surface. This can vary from a couple of millimeters to a couple of centimeters if serious damage is present on the wheels. We will call the total amount of material removed during a reprofiling instance the cutting depth or∆r. The reprofiling is executed by Strukton Rail against a fee.
Figure 2.4: Mobiturn.
A distinction between preventive and corrective reprofiling is handled by Arriva. Preventive re- profiling is planned upon expected mileage of a train. Every 200,000 kilometres, wheels need to be reprofiled. This is also in line with findings in literature (Cantini et al., 2016). Corrective reprofiling needs to be carried out when inefficiencies are found before the mileage goal is reached.
Inefficiencies can be found by visual inspections on the train wheels. When these are severe, the train cannot be approved to get operational. Quick re-scheduling is then needed to either reprofile the damaged wheels or replace the entire bogie. Strict measures are handled when reprofiling wheel sets. The following measures need to be satisfied before the reprofiling can be approved (Stadler, 2014; Alstom, 2005). It is assumed that the wheels are in an as-good-as-new state after the reprofiling is applied. The hf and tf measures in Figure 2.5 correspond to Sh and Sd in Table 2.2.
