ProRail’s story on getting machine learning ready!
WOE 12:45 - 13:15
Most companies that use machine learning are relatively young tech companies, being able to build these systems from their startup phase, working exclusively within a digital environment and having no direct connections or impact on their surroundings. But what if you are already a well-established company, working in an environment with high safety importance who wants to implement machine learning in their business? ProRail, owner and manager of the Dutch railways, took on the challenge and collects data (over 1.1 billion images!) on the ever-changing landscape that is the railway. They use machine learning (computer vision) to make real-life impact by identifying assets and its condition for a safe, sustainable and efficient railinfrastructure. But working in such a well-established, practical, less controllable environment does not come without its technical and organisational challenges. Job Wegman (ProRail) and Clint de Keizer (Vantage AI) will show you those challenges and the practical lessons to take home with you to make your organisation machine learning ready!
ProRail is the owner and manager of the Dutch railways. At ProRail we create an infrastructure for train operators to drive on. Therefore, it is our obligation to secure a safe and efficient railinfrastructure by maintaining the rail assets and managing the traffic.
The asset management department collects all sorts of information about our assets. What is the function of the asset? Which asset or which asset type is this? When and where is this asset installed? High quality of this information is crucial to secure a safe railnetwork. At ProRail we highly value safety and the high quality of our data is crucial to achieve this. Poor data means your maintenance plans could be based on inaccurate information, leading to higher cost in maintenance or even safety issues. Therefore, the contribution of this project is more than just financial benefits but also a way to improve the safety and reliability of the rail infrastructure.
To monitor the condition of the infrastructure, we retrieve images of the video inspection trains (videoschouwtrein). At this moment we have 1.1 billion images of our railnetwork! We use this data to apply computer vision to identify our assets to improve our configuration data.
This resulted in a successful project and DevOps team which is improving and developing new algorithms while also maintaining our models in production. Since the start we developed models to identify rail dampers (raildempers), Insulation-joints (ES lassen) and sweepers (dwarsliggers). Today we have improved the data on more than 14.000 assets and improved our data on more than 50% of our insulation-joints.
Doing this does not come without challenges: data drift, inaccurate data because of hardware problems, organizing when and where to collect your data and more. But also, challenges on the side of the DevOps team: like how do you manoeuvre within a big organization like ProRail, how do you transfer tacit knowledge from domain experts into programming and how do you maintain such a new and changing product with a likewise team?
We will show you how we came up with different ingenuities and solutions that has made our life easier and will hopefully help you on your next project! To make it even easier we will summarize this in different lessons to take hme with you.
Organisaties beschikken al over extreem veel data. De grote uitdaging zit erin om met de beschikbare data waarde toe te voegen aan een proces of klanttraject. Data science is daarvoor uitstekend geschikt. Finance en het voorspellen van consumentenbehoeften zijn twee gebieden waarop data science effectief is toe te passen. De investering loont sowieso, want beschikt een organisatie over meer inzicht, dan zijn ook slimmere keuzes te maken. Het thema data science neemt precies hierom een steeds prominentere rol in tijdens de Big Data Expo.