Helsinki Data Science Meet-up gets a Demonstration: Analytics for Transport.
The VR Group is a Finnish State-owned enterprise that primarily operates in Finland, but also has operations abroad, including in Russia and Sweden. The VR group employs 7,500 professionals and prides itself on providing its customers with high-quality, environmentally-friendly passenger and logistics services. The VR Group has three main business operations revolving around customer groups. VR’s passenger services offer public transport services in commuter and long-distance trains and buses. VR Transport offers road and railway logistics services, while VR Track is focused on infrastructure, maintenance and supplies railway materials.
Annika Nordbo, Head of Data Science at VR Passenger Traffic, opened the night with an overview of VR Data Science Architecture in AWS. The audience was presented with a summary of what VR Data Science had achieved in less than a year – and the audience was convinced – the stellar team of seven have gained complex and useful insights using data science in just the short period of less than a year. Here is a summary of what was learnt…
Heikki Pulkkinen, Data Scientist at VR Passenger Traffic, started his presentation with an introduction to CausalImpact, which is an R package for causal inference using Bayesian structural time-series models. The R package implements an approach to estimating the causal effect of an intervention on a time-series, such as the effect of an advertisement campaign on daily clicks. Reference connections were discussed extensively with an active audience, and became a topic of interest. They are about representing the baseline sales for the rail network when part of the network is affected by an anomaly, e.g. a rock concert. Specifically, the presentation reviewed how such an event can increase sales and how the increase in sales can be taken into account in e.g. train scheduling and composition planning.
The second part of the presentation discussed predicting passenger counts for trains departing in the future. XGBoost was selected as the model and an XGBoost algorithm was implemented for the first selection and resulted in 464 features. Reservations by predication date had the highest correlations with passenger counts. The model has been implemented in production and results were demonstrated and visualized in PowerBI.
Tuomas Karavirta, Data Scientist at VR Maintenance, focused his presentation on laser scanning of wheels to simulate rolling dynamics giving the example of condition-based maintenance in practice. The presentation focused on the wheel wear of the SR2 locomotive. By implementing a new algorithm for the calculation of a wheelset observable conicity, which is used to trigger the reprofiling calls for the SR2 locomotives, the reprofiling strategy has improved dramatically. This change in operation corresponds to:
- 20% increase in wheel set lifetime
- 30% reduction time in reprofiling
- 17% cost reduction saved in materials and labor
- 17% reduction in bogie maintenance
The results provide a compelling use case for using advanced analytics for predictive maintenance.