Improving Cargo Train Availability with Predictive Maintenance: An Overview and Prototype Implementation
Sebastian Kauschke, TU Darmstadt
We use machine learning algorithms on the log-files of cargo locomotives to detect patterns which may allow us to predict failures of selected system components in advance.
In cargo transportation, reliability is a crucial issue. In the case of railway
traffic, the consequences of locomotive failure are not limited to the affected
machine. Beside the cost of the machine itself, delays are caused and ultimately
propagated through the railway network, rendering the accumulated cost of a
single incident unpredictable. In order to avoid failures, Predictive Maintenance
(PM) can be used. Predictive Maintenance targets the substitution of existing
maintenance processes (e.g. time-based preventive maintenance) by conveniently
scheduled corrective maintenance through exploitation of the underlying deterioration
processes. In an ideal PM scenario, constant monitoring of the machine
is available, measuring all relevant variables, e.g., temperatures or vibrations on
a regular basis. However, in the real world, this assumption is limited: The hardware
often does not deliver the required amount of data in the necessary precision.
Often the machines record only a log-file which provides all activities—useful or
not—that the various systems in the machine keep track of.
In this paper, we give a short overview on PM in general and on the various types
of systems that can be considered for PM.We elaborate on the differences in data
as well as the nature of the systems it is possible to predict failures upon.
In a prototypical example, we make use of machine learning methods to construct
a failure prediction model for cargo trains. This data-driven approach focuses on
a specific failure problem which is important to improve upon and aims at an easy
prototype implementation for the currently available system.
We train a classification model which uses the pattern structure of the diagnosticmessages
of the locomotive to recognize abnormal activities in the locomotives
behaviour. A meta-classification layer on top of this anomaly detection allows us
to build a predictor with little false positives.We evaluate our findings on the data
of 180 locomotive tours and elaborate on possible improvements of the method.
Association for European Transport