Track Fixity Layer

Authors: Qian Fu, John M. Easton and Michael P. N. Burrow
Project support: Sarah Jordan

Introduction

This project aims to develop a data mining tool to assist engineers in analysing and predicting track movements over time. The analysis is based on recorded asset data, local operational and environmental factors and knowledge of track design in the area. In this prototype phase, the team focused on assembling a structured dataset for a representative track section to develop data-driven models of track movement, identify key track geometry parameters predicting future movements and generate fixity values for all reference curves used on the GB rail network.

Key achievements

The project has resulted in the following key achievements:

• Data integration and management: A data cleaning and pre-processing workflow has been developed using the open-source PostgreSQL database. This workflow enables smooth integration and management of the data corpus. Initially, six industry data streams were used (CARRS, CNM, OPAS, ballast type, track quality, and point cloud surveys). The workflow is easily extensible, allowing for the inclusion of additional data sets in the future, provided they are referenced linearly (e.g., ELR + track mileage) or using standard geographic coordinate systems.
• Predictive machine learning model: A machine learning model was trained on data from an 80-km section of the East Coast Main Line south of Edinburgh. This model predicts future track movements, with most predictions falling within the correct bin or within a single bin width of the true value. The model primarily used parameters such as curvature, cant, and maximum speed for its predictions.
• Python package development: A three-module Python package has been developed to facilitate the rapid implementation of the described functionalities in software workflows.

These achievements lay the groundwork for further development and refinement in future phases of the project.

Technical Report

• Fu, Q., Easton, J. M. and Burrow, M. P. N. (2021). Track Fixity Layer - Final Report. Approved by Network Rail Ltd. [Appendix - Technical Documentation]

Conference/Workshop Posters

• Clark Lecture 2024, University of Birmingham, Birmingham, 26 June 2024. [Poster]
• Research Software Engineering in Data & AI Workshop, 15-17 February 2023, University of Warwick, Coventry. [Poster]
• Transportation Research Board 102nd Annual Meeting, Washington, D.C., USA, 8–12 January 2023. [Poster]

Publication

• Fu, Q., Easton, J. M. and Burrow, M. P. N. (2024). Development of an Integrated Computing Platform for Measuring, Predicting, and Analyzing the Profile-Specific Fixity of Railway Tracks. Transportation Research Record, 2678(6), 1-13. doi:10.1177/03611981231191521.

Demo

Note: The following notebook code is for demonstration purposes only. Access to the project's database server will be necessary to successfully instantiate the prototype model.

>>> from src.modeller.prototype import TrackMovementEstimator

>>> # Specify element, direction and subsection length over which average movement is calculated
>>> element = 'Left Top'
>>> direction = 'Up'
>>> subsect_len = 10

>>> tme = TrackMovementEstimator(element=element, direction=direction, subsect_len=subsect_len)

Initialising the feature collator ... Done.
Initialising the estimator ... Done.

1. Create training and test data sets.

>>> tme.get_training_test_sets(test_size=0.2, random_state=1)

Calculating track movement ... Done.
Collating features: 
  ballasts (src. Ballast summary) ... Done.
  structures (src. CARRS and OPAS) ... Done.
Finalising the data integration ... Done.
Splitting into training and test sets ... Done.
The data is now ready for modelling.

>>> tme.X_train

      Curvature    Cant  Max speed  ...  Retaining walls  Tunnels  Stations
3306   0.000000    0.00       90.0  ...                0        0         0
1910  -0.000858 -110.00       95.0  ...                0        0         0
1192  -0.000545  -89.01       80.0  ...                0        0         0
981   -0.000767 -130.00       95.0  ...                0        0         0
5139  -0.000256  -63.00      110.0  ...                0        0         0
...         ...     ...        ...  ...              ...      ...       ...
1016   0.000000    0.00       95.0  ...                0        0         0
5907  -0.000504 -127.00      110.0  ...                0        0         0
4597  -0.001060 -125.00       85.0  ...                0        0         0
243   -0.000185  -21.42       95.0  ...                0        0         0
5869  -0.000394 -108.78      110.0  ...                0        0         0

[5433 rows x 9 columns]

>>> tme.y_train

2736    1
1578    3
1060    0
870     2
4512    3
       ..
905     2
5192    2
3980    2
235     1
5157    3
Name: lateral_displacement_mean, Length: 5433, dtype: int64

>>> tme.X_test

      Curvature    Cant  Max speed  ...  Retaining walls  Tunnels  Stations
5628   0.000731  103.38      100.0  ...                0        0         0
5696   0.000955  135.00      100.0  ...                0        0         0
5431   0.000960  142.00      100.0  ...                0        0         0
3471  -0.000784 -108.50       90.0  ...                0        0         0
2908  -0.001264 -155.00       85.0  ...                0        0         0
...         ...     ...        ...  ...              ...      ...       ...
247   -0.000337  -39.00       95.0  ...                0        0         0
3904   0.000000    0.00      125.0  ...                0        0         0
1054   0.000000    0.00       95.0  ...                0        0         0
3191  -0.000858 -156.00       90.0  ...                0        0         0
4026  -0.000271  -92.71      125.0  ...                0        0         0

[1359 rows x 9 columns]

>>> tme.y_test

4958    3
4999    0
4793    1
2901    2
2394    2
       ..
239     2
3309    2
943     0
2621    3
3431    3
Name: lateral_displacement_mean, Length: 1359, dtype: int64

2. Train an RF model

>>> tme.classifier(n_estimators=300, max_depth=15, oob_score=True, n_jobs=8)

Model training in process ... 
[Parallel(n_jobs=8)]: Using backend ThreadingBackend with 8 concurrent workers.
[Parallel(n_jobs=8)]: Done  34 tasks      | elapsed:    0.0s
[Parallel(n_jobs=8)]: Done 184 tasks      | elapsed:    0.1s
[Parallel(n_jobs=8)]: Done 300 out of 300 | elapsed:    0.3s finished
Done.

Testing accuracy ... 
[Parallel(n_jobs=8)]: Using backend ThreadingBackend with 8 concurrent workers.
[Parallel(n_jobs=8)]: Done  34 tasks      | elapsed:    0.0s
[Parallel(n_jobs=8)]: Done 184 tasks      | elapsed:    0.0s
[Parallel(n_jobs=8)]: Done 300 out of 300 | elapsed:    0.0s finished
Done.

Mean accuracy: 49.96%

                    Importance
Curvature            0.388747
Cant                 0.372477
Max speed            0.201602
Underline bridges    0.009513
Overline bridges     0.007261
Max axle load        0.006700
Retaining walls      0.006048
Tunnels              0.005787
Stations             0.001866

3. Create a confusion matrix

3.1 View the confusion matrix for the trained RF model

>>> tme.view_confusion_matrix()

[Parallel(n_jobs=8)]: Using backend ThreadingBackend with 8 concurrent workers.
[Parallel(n_jobs=8)]: Done  34 tasks      | elapsed:    0.0s
[Parallel(n_jobs=8)]: Done 184 tasks      | elapsed:    0.0s
[Parallel(n_jobs=8)]: Done 300 out of 300 | elapsed:    0.0s finished

3.2 Normalise the confusion matrix over the predicted labels

>>> tme.view_confusion_matrix(normalise='pred')

[Parallel(n_jobs=8)]: Using backend ThreadingBackend with 8 concurrent workers.
[Parallel(n_jobs=8)]: Done  34 tasks      | elapsed:    0.0s
[Parallel(n_jobs=8)]: Done 184 tasks      | elapsed:    0.0s
[Parallel(n_jobs=8)]: Done 300 out of 300 | elapsed:    0.0s finished