Siemens believes Big Data analytics will enable forecasting component faults weeks in advance

Siemens tracks Big Data for trains that keep on running

German train maker Siemens is exploring Big Data analytics to make sure no train will ever again be left stranded on tracks due to an unforeseen technical failure.

A small team of data scientists and engineers based at the tech giant’s site in Allach - a suburb of the Bavarian capital, Munich - embarked on a mission three years ago to make trains so smart that they could effectively inform their operators weeks in advance that a component is going to fail.

Siemens, which won the contract in 2013 to deliver 115 trains for London’s Thameslink project, hopes the service will provide a crucial competitive edge in the increasingly competitive rolling stock market.

“My goal is very clear: for every relevant component, I want to have a warning at least a week in advance that this thing is going to fail,” said Gerhard Kress, head of the Siemens Mobility Data Services Center, when outlining his vision. “This gives the operator enough time to do maintenance at the normal time when the train is not in use.”

Other train makers are also flirting with trendy Big Data analytics to produce smarter trains. However, Kress is confident that his team is sufficiently ahead in this game, turning locomotives into machines smarter than modern airplanes.

“We are essentially building on the know-how that Siemens has developed over the years for other types of applications, namely in healthcare and gas turbine operations,” Kress said. “We have also massively invested in building our team. All our scientists not only have PhDs in data science, machine learning or mathematics, but also a background in mechanical engineering.”

The data platform, developed by the team, handles vast amounts of data. A single high-speed train, such as a Velaro serving on the Eurostar route through the Channel Tunnel, sends about 400 data points per second. For a fleet of 300 trains, that would be millions of data points per hour. The data includes sensor readings about temperature of all critical components, torque on gearboxes or error messages. In addition, the analysts receive weather data, geographical and track information, as well as reports about work carried out on the trains. The massive task is to make sense out of this never-ending flood of information, or better yet to develop systems that would do that automatically.

“We have an underlying data model that reflects the structure of the vehicle,” Kress explained. “But first we need to give meaning to different parts of the data and then we use machine learning techniques to find and identify patterns, which indicate abnormal behaviour occurring long before a component failure. When we find these patterns, we verify them with the engineers to make sure we understood the right thing and then we build automated models that look for indications of component failures.”

Until two years ago, technicians would only download sensor data during regular train inspections. Most of the data was usually wasted as there was no intelligent platform that would make sense out of the millions of readings – a task clearly superhuman.

The roll-out of 4G mobile networks eventually enabled transmitting data from the thousands of sensors on every train in real time or, at least, with very low latency.

Previously, technicians in train depots were faced with downloading massive data sets about individual trains to their laptops. Today, there is a neat application on the desktops of the data scientists in the Allach centre, visualising data for every fleet.

“Today, we can compare vehicle behaviour and fleet behaviour and that makes things a lot easier,” said Kress. “We can for example see what components suffer the most in trains operating in very hot climates compared to those in cooler ones.”

In the past three years, Kress believes, trains have overtaken planes in their ‘smartness’. And the quest doesn’t end here. 5G networks are already on the horizon and with them lower latencies and higher bandwidths.

“Bandwidth is obviously a problem,” said Kress. “When the train travels fast, bandwidth shrinks even more. At the moment, for the data that we are using, it’s fine, but we would like to go to vibration data as well and in this type of data, you are looking at the kilohertz range, where you are monitoring these movements and for that we don’t have the bandwidth to transfer the data. Right now, if we want to do something on vibration topics, we have to get that from the vehicles. That would change with the 5G networks.”

More bandwidth would also naturally mean more data. With more data, the failure forecasting may aim for even more ambitious goals. Instead of the current one week’s notice, the operators in future may be able to learn up to three weeks in advance that a particular train will need to have a component replaced or adjusted.

The ideal, Kress explained, is to have no faulty trains standing in the depots awaiting repairs or waiting for spare parts to be delivered. This is important for rail operators in busy global cities where passengers are complaining about delays and service disruptions.

“We want to help our customers to provide more service with the same assets,” said Kress. “If you only learn one day in advance that a train needs to have work done, you need to carry out emergency action. If you know three weeks ahead, you can plan to do that during the next regular inspection so there is no unplanned downtime for the train.”

Recent articles

Info Message

Our sites use cookies to support some functionality, and to collect anonymous user data.

Learn more about IET cookies and how to control them