Predicting track quality using structural elements and relative track geometry data measured by RILA

Abstract

Track quality measures differ from standard to standard. Widely used indices for track quality are the standard deviations of the rail longitudinal level data and the standard deviations of alignment data. Researchers have opted for various learning methods as to relate track quality to different parameters. Some research has gone to the point of using track quality interchangeably with the notion of defect density. What is common in all researches is that a track quality index is used as label for predictions. The differences are inherent to the variables/features that were chosen to conduct such predictions. With the plethora of data at the hands of asset managers nowadays, it has become easier to opt for these learning methods. Machine Learning classification is a method wherein a number of features are used to classify a datapoint into a label. With thorough learning, by inputting the features, the label would then be predicted. The features of interest in this project are structural elements describing the track sections. This can go from the sleeper type and subsoil type all the way to the distance from the nearest switch or bridge. This will allow for the relationships between the structural elements and the track’s quality to be bolstered all the while creating a predictor based on data that need not be measured. The research question is thus formulated as follows: How to relate track quality in a section to its surrounding structural elements and form a predictor out of it? 1.What structural elements are to be studied? 2.How to automate the generation of the features? 3.Which classifier algorithm yields the most accurate predictions? 4.Which structural features are the most important in predicting track quality? The research question is divided into 4 sub-research questions each pertaining to a part of the research. 1.Decide on which structural elements are the most relevant for the study. Some elements may be redundant with respect to other elements. For instance, subsoil type and ballast settlement could form a redundancy. This will be decided on from the literature. This part will contribute to a section in the thesis in which a strong relation will be drawn with respect to the courses taken and studied at the TU Delft. 2.Data Gathering is a huge part of this research. Input data will be from Fugro’s relative track geometry data and KRDZ data. In addition, ProRail’s informatieportaal, which is a website containing all the railway assets and their locations on the map will also be of huge aid to gather the structural elements data. The answer to this sub-research question lies in automating the structural feature generation, making them ready for learning. 3.Running the Machine Learning classification code on sections of a real-life track and comparing the accuracies of each classification algorithm as to come up with recommendations for future work. 4.Some machine learning algorithms, namely the random forest classifier have in their output an importance of features measure. This will give us insight on which structural elements affect track quality the most, which would allow for better resource allocation in the long run. In the goal of answering all the research questions and to realize the research objective, an outline of the research is included in this report along with a workflow diagram. The project has shown that predicting track quality using structural elements and relative track geometry can be done using classification methods, on the condition that the predicted track and the trained track have similar feature importance values, and that structural features that are displayed in the track to predict are also available in the trained model.