IC3 and IC4 Trains Under Risk of Blocking their Wheels - A Big Data Case Story

Transcription

1 IC3 and IC4 Trains Under Risk of Blocking their Wheels - A Big Data Case Story Anders Stockmarr Associate Professor Section for Statistics and Data Analysis, DTU Compute. anst@dtu.dk Master of Global Management DTU Management March 21, 2017

2 Things to consider in Analysis when using Big Data What are the challenges in practical implementations of analysis using Big Data? Are the data the right data? Sufficient? Accurate? What can go wrong when using Big Data? How can one get around the problems? To what extent does the structure of the Big Data affect the conclusions that you may draw from analysis? 2 Dept. of Applied Mathematics and Computer Science,

3 The Marslev Incident On November 7th, 2011, an IC4 train passed an active stop sign at the Fuenen town Marslev. 3 Dept. of Applied Mathematics and Computer Science,

4 The Marslev Incident On November 7th, 2011, an IC4 train passed an active stop sign at the Fuenen town Marslev. The train attempted to break, but the wheels blocked. From an initial speed of 180 km/h, it took the train meters to come to a halt; it passed the stop sign with 651 meters, and stopped just 371 meters from an anterior freight train. Shortly after the Marslev Incident, the Accident Investigation Board was notified of an additional 3 incidents, where the train driver experienced problems with an IC4 train. 4 Dept. of Applied Mathematics and Computer Science,

5 The Marslev Incident Consequences After the Marslev Incident, the maximum train speed for IC4 trains was lowered to 140 km/h, in contrast to the 180 km/h for the older IC3 trains. An investigation was commenced, focusing on the mechanics and the electronics in the IC4 train, in order to clarify whether the braking system was working as it should. In August 2013, a report aquitted the braking system of the IC4 trains it was functioning as it was supposed to. What was then the cause of the Marslev Incident? 5 Dept. of Applied Mathematics and Computer Science,

6 The Marslev Incident Consequences In the autumn 2012, 30 test rides were performed on the route Copenhagen-Århus. Task: Find the cause of wheel blockings; Investigate if IC3 and IC4 trains differ wrt. wheel blockings. The suspicion gathered around leaf juice; leaf juice makes the tracks slippery. But no formal definition of slippery tracks exists! We must find a proxy for slipperyness. Method: Analysing the Adhesion Coefficient; a measure for the level of contact between wheels and track. However, it turned out the the Adhesion Coefficient could not be calculated from the available data. New plan. 6 Dept. of Applied Mathematics and Computer Science,

7 7 Dept. of Applied Mathematics and Computer Science,

8 Big Data Structure of Braking Data Volume: Medium/low Variety: Advanced Velocity: Medium Veracity: Advanced 8 Dept. of Applied Mathematics and Computer Science,

9 Train data; Train logs; GPS logs (position); Databases for Analysis Track data; Records of vegetation along the rail track; GIS shapes of forest, bushes and solitary trees. The Curve register; Curvature. The Route register (coordination of time and position); Distance to Copenhagen Central Station. The Track register to assess which track on the route that the train was running on. Register data for elevations/recesses. Environmental data; Meteorological data from the Danish Meteorological Institute. Local values of temperature, dew point, wind speed, wind direction, turbulence, precipitation og solar radiation. Within the last hour, and accumulated over 3, 4, 5, 6, 7, 8 and 24 hours. 9 Dept. of Applied Mathematics and Computer Science,

10 Databases for Analysis Elevation register Curve register Vegetation register Route register GPS log Track Register Train log Analysis Environme ntal data Indexed by time: Indexed by distance: 10 Dept. of Applied Mathematics and Computer Science,

11 New Analysis Plan The only response in the original train data that indicated anything about the functionality of the brakes was the so-called blocking flag; the train computer indicating that the wheels are blocking. Therefore, the item for the analysis was a given quantity. Hypothesis: H: The risk of a blocking flag is affected by:» slippery tracks;» train characteristics;» track characteristics; But we do not know what slippery tracks are! - we use indicators for leaf juice on the tracks and enviromental data for rain, moisture (dew point), turbulence, wind etc. 11 Dept. of Applied Mathematics and Computer Science,

12 Data Providers We obtained the described data when we realized that we needed them Train logs were provided as observational data; Register data was provided by BaneDanmark; Environmental data were provided by Danish Meterorological Institute; extracted from their interpolation model and specifically designed for our time and space needs. It is a challenge to make these three data sources aligned to support an analysis. And each of them have their challenges. But the value of the combined data was immense in this project. 12 Dept. of Applied Mathematics and Computer Science,

13 Database; Train logs: Data from train logs for the test rides contained: Speed; Time Stamp; Braking Power; Blocking Flag Status. Wheel blocking is a common event; the typical event doesn t create dangerous situations, but affects the material and is therefore pricey. And of course, once in a while the event can result in danger. 13 Dept. of Applied Mathematics and Computer Science,

14 Train Ride Data Train Data: GPS Data: 14 Dept. of Applied Mathematics and Computer Science,

15 Messy Data I Train log Data for train-log: Here plotted sequentially from. the dataset: Comment from the data provider: Anders is right. Time goes backwards. I have no explanation for it. 15 Dept. of Applied Mathematics and Computer Science,

16 Cleansing the Speed Profiles The only way to get around a problem is to work your way through it! 16 Dept. of Applied Mathematics and Computer Science,

17 Cleansing the Speed Profiles Differenced Time Stamps, Train 12 diff(time) Index 17 Dept. of Applied Mathematics and Computer Science,

18 Messy Data I Train log Cleansed Speed profile: 113 corrections for time shifts unaccounted for. 18 Dept. of Applied Mathematics and Computer Science,

19 Messy data I Train Log Speed :30 09:00 09:30 10:00 10:30 Time 19 Dept. of Applied Mathematics and Computer Science,

20 Messy Data II GPS Log The position, and thus the traveled distance, from the Train Log is unreliable, in particular because of the uncertainty about the time variable, but also in the case of braking/slipping (based on turnings of the wheels). Position should be derived from GPS data But typically there are only a few hundred GPS points from Copenhagen to Århus and 328 km. For some reason, the request for 2000 GPS points was forgotten/mislaid/ similar when the test rides were carried out. Impossible to re-create position from interpolation between GPS points already from Copenhagen to Glostrup, a shortfall of 500 meters is seen. 20 Dept. of Applied Mathematics and Computer Science,

21 Messy Data II GPS Log For positioning, we use that we know WHEN you are at a station 40 reference points on the Copenhagen-Århus route, for which coordinates exist in the Route register. Using these, we can fix the time stamps of the remaining GPS points, relative to the Train Logs, and calculate positions from time stamps in Train Logs (one per 7 seconds for IC3, more for IC4) using great circle distances. 21 Dept. of Applied Mathematics and Computer Science,

22 Messy data III - Track Lanes Necessary information to decide the curvature of the track. Always use the right track we were told; but look at this picture. Red trains are S-trains, local trains which in this instance run on separate tracks. 500m after this bridge the right track curves northward away from the main route, and therefore cannot be used by the trains bound for Århus. 22 Dept. of Applied Mathematics and Computer Science,

23 What explains variations in Blocking Flags? - Slippery tracks- - Vegetation; - Weather in general; - Track and Train characteristics; - INTERACTIONS between these. +? 23 Dept. of Applied Mathematics and Computer Science,

24 Vegetation data GIS Shapes 24 Dept. of Applied Mathematics and Computer Science,

25 Vegetation data GIS Shapes xxx 25 Dept. of Applied Mathematics and Computer Science,

26 Vegetation data GIS Shapes Handling: Extent and area of GIS shapes known. Conversion to rectangles with the same area and extent. Goal: Construction of a Leaf Fall Index by combining meteorlogical data with track data. 26 Dept. of Applied Mathematics and Computer Science,

27 Vegetation data GIS Shapes xxx 27 Dept. of Applied Mathematics and Computer Science,

28 Construction of a Local Leaf Fall Index: The Thickness Index T Forest Bushes 30m 10m Forest: 30 meters thickness: T=1. 15 meters: T= 3/4. 10 meters: T= 4/9. On site Average 1km backwards Average 2km backwards Copenhagen H 28 Dept. of Applied Mathematics and Computer Science,

29 Construction of the local Leaf Fall Index C θ T Track angle in the point B: South B θ North θ W angle between the immediate wind direction and the north-south axis; A South θ North T R, T L the Thickness Index to the left and the right of the track, respectively (Copenhagen-Aarhus direction). The Leaf Fall Index I is then constructed as I F = sin (θ W θ T ) T R 1 sin (θw θ T <0} + T L 1 {sin θw θ T >0} Full effect if the wind is perpendicular to the rail track, no effect if it is parallel. 29 Dept. of Applied Mathematics and Computer Science,

30 Local Leaf Fall Index 30 Dept. of Applied Mathematics and Computer Science,

31 Local Leaf Fall Index Should be combined with a continuous index that measures the rate that leaves fall off the trees as a function of calendar time. But such knowledge does not exist (Forest & Landscape, University of Copenhagen 2013, personal communication). To rectify this, we used polynomial regression on calendar time. 31 Dept. of Applied Mathematics and Computer Science,

32 Modeling: Blocking Sequences If first a wheel blocks, the the system reacts slowly to changed circumstances. A Blocking Flag is likely to be followed by another. The problem is handled by introducing a Blocking Sequence indicator, which indicates that the train had issued a Blocking Flag at the previous recorded position. This introduction is necessary. IC4 trains record far more data points than the IC3 trains, and therefore also far more Blocking Flags. But the amount of Blocking Sequences are compatible. 32 Dept. of Applied Mathematics and Computer Science,

33 Analysis Model the probability of initiating a Blocking Sequence. Generalized linear model: logit(p(initiating Blocking Sequence)) ~ β T X Temporal Autocorrelation, Residua ACF Lag 33 Dept. of Applied Mathematics and Computer Science,

34 IC4 vs. IC3 trains Predict probabilities for all test rides, assuming a constant speed of 140 km/h, and IC3/IC4 status. Quantile-quantile plot for inversed linear predictors: 34 Dept. of Applied Mathematics and Computer Science,

35 Smoothed Risks, IC3 140 km/h 35 Dept. of Applied Mathematics and Computer Science,

36 Constant Speed 180 km/h, Braking Power Dept. of Applied Mathematics and Computer Science,

37 Effect of Leaf Juice Both Leaf Fall Index for forest and bushes are statistically significant; on site, 1km back and 2km back. Solitary trees are not significant. The effect declines with calendar time; the amount of leaves becomes less. Leaf Fall Index interacts with meteorological covariates at large. 37 Dept. of Applied Mathematics and Computer Science,

38 Effects of Speed and Braking Power 38 Dept. of Applied Mathematics and Computer Science,

39 How about the Marslev Incident? We managed to obtain information on the Marslev Incident. However, the environmental circumstances were so extreme compared to our test rides, that the probability of a initiating a Blocking Flag Sequence was just 1, 1, 1 and always 1. Thus, we cannot claim to have modeled the event, but it is clear that the conditions were indicating extremely slippery tracks, in the sense of the proxies for slippery tracks that we have developed. 39 Dept. of Applied Mathematics and Computer Science,

40 Conclusion 1. A. The low number of GPS positions relative to the number of data points constitutes a problem. Since position is based on GPS data, position, and thus track characteristics, vegetation and meteorological data are subject to uncertainty. This lowers the validity of the study results. B. The Speed Profiles are faulty. While we do not believe that this plays a role, we cannot say for sure. A large amount of missing data also adds to the picture of a deficient data base. C. Data are collected on a limited number of days. In fact, the applied data are from only 11 different days, and this limits the possibility for generalizing the results. Additionally, we cannot generalize the result to outside the leaf fall period. 40 Dept. of Applied Mathematics and Computer Science,

41 Conclusion 2. A. The probability of initiating a Blocking Flag Sequence is affected by slippery tracks. B. One of the factors that contributes to tracks slipperiness in the trial period is leaf juice. C. Data are based on many approximations. The results lacks sufficient evidential weight. D. There is solid evidence that IC3 and IC4 trains have different probabilities of initiating Blocking Flag sequences. But we have not been able to obtain information on the sensitivity for blocking flags in IC3 and IC4 train computers. We *may* have shown that the train computers have different sensitivity to the same risk. E. Should the problems with data resolve themselves, it is our opinion that the method may be generalized beyond the leaf fall period. F. For further development, one of the perspectives is an early warning system in the train traffic. 41 Dept. of Applied Mathematics and Computer Science,

42 Acknowledgements - Cooperation on analysis and hypotheses formulations: - Bjarne Kjær Ersbøll - Data management: - Ewelina Kotwa - Camilla Thyregod 42 Dept. of Applied Mathematics and Computer Science,

43 Thank You For Your Attention 43 Dept. of Applied Mathematics and Computer Science,