Talgo: Streaming two thousand events a second at three hundred kilometers an hour

About Talgo

Founded in 1942, Talgo is a Spanish manufacturer of intercity, standard, and high-speed passenger trains with a history of innovation. The company was listed on the Madrid Stock Exchange in 2005.

Industries: Manufacturing
Location: Spain

About OpenSistemas

Based in Madrid, Spain, OpenSistemas is an award-winning cloud consultancy specializing in open source solutions.

Train manufacturer Talgo revolutionizes train maintenance with Google Cloud Platform, by streaming sensor data in real time and applying machine learning to stop problems, without stopping service.

Google Cloud Results

  • Streams data in real time from 2,000 on-board sensors for immediate analysis, anywhere in the world
  • Collects 2,000% more data per train, per day than using a previous system
  • Uses machine learning for potential savings of over 200 hours a year per train in maintenance inspection time

Collects 2GB of data per train, per day

When is a high-speed train not a high-speed train? Answer: When it's standing in a depot, parked for maintenance. Talgo is one of the world's foremost manufacturers of high-speed trains, and a leading innovator in the use of new technologies to monitor trains while they're still on the move. The motivation is simple: by minimizing the need for maintenance stops, Talgo maximizes the time each train is available to move passengers from point A to point B.

Talgo has 80 trains in Spain, 36 in Saudi Arabia, and more than 30 additional trains scheduled for delivery by 2021, including vehicles for the United Kingdom's High Speed 2 project using real-time monitoring, and another 90 trains using event alarms monitoring. More recent contracts for trains include much higher expectations about maintenance, as José Antonio Marcos, Talgo's Chief Maintenance Engineer explains:

"The new contracts we have with operators are very, very demanding. In terms of reliability, we may have to guarantee more than 30,000 kilometers between basic inspection stops, and more than 1.5 million kilometers between failures. We also have to ensure 99% availability of the fleet, whether it's traveling through a Saudi Arabian sandstorm or winter in the UK. Talgo can provide that level of service, but the only way to do it is using cloud technology."

"Cloud technologies are essential for any predictive maintenance model on this scale. Today, companies that do not use cloud simply cannot meet the challenges imposed by the sector of high reliability and availability, alongside low maintenance costs. If you work with cloud, you are in. If you don't, you're out."

José Antonio Marcos, Chief Maintenance Engineer, Talgo

In order to meet these new standards, Talgo has greatly increased the volume of data it can collect, store, and analyze from each train. Previously, Talgo would collect 1MB of data per day from each train and trains would send alarm notifications in case of problems. Now, the company streams 2,000 times that amount of information: 2GB daily from each vehicle, including data from 2,000 sensors every second in near real time. Not only does that data monitor the trains in action, but it also feeds into high-precision machine learning models that predict maintenance needs, so Talgo can resolve them without disrupting service.

"Cloud technologies are essential for any predictive maintenance model on this scale," says José. "Today, companies that do not use cloud simply cannot meet the challenges imposed by the sector of high reliability and availability, alongside low maintenance costs. If you work with cloud, you are in. If you don't, you're out."

Creating an infrastructure to predict problems before they arise

Using sensor data from trains, maintenance teams can analyze operations and anticipate problems before they arise, greatly reducing the time trains have to spend stopped for routine checks or major repairs. Talgo streams 2,000 signals a second from each of its trains in real time, including information on voltage, hydraulics, acceleration, and temperature, as well as video of the driver's view. In addition, trains send batch files every four minutes, generated when a combination of signals creates an alarm for the attention of maintenance teams.

Each Talgo train generates 2GB of data per day, and with more than 100 trains in operation in several countries, it was clear to José and his team that on-premises infrastructure would require a prohibitively large investment in equipment alone. Instead, Talgo looked for highly scalable cloud infrastructure ready to connect to its trains and ingest, process and store data for machine learning models. They turned to Google Cloud Platform (GCP). "We chose Google Cloud Platform for its unlimited scalability, great connectivity, massive storage, data streaming, and its machine learning solutions," says José.

In the architecture Talgo created for real-time streaming, each train sends 2,000 events a second through a VPN and gateway to a UDP Bridge hosted on a Compute Engine instance. This streams the data through Cloud Pub/Sub to a UDP Decode on App Engine, which presents the data for analysis on dashboards, as well as connecting to Cloud Datastore.

"We can easily send results from our Google Cloud Platform machine learning model to mobile devices using App Engine. Predictions, recommendations, and evaluations of the condition of trains can be sent directly to engineers anywhere, and we can automatically generate reports using Google Data Studio, so maintenance can be more effective."

José Antonio Marcos, Chief Maintenance Engineer, Talgo

"Once the data is in Cloud Datastore, we conduct real-time monitoring using an application we developed with the consultancy OpenSistemas," says José. "Both our maintenance engineering department in Saudi Arabia and our corporate headquarters in Spain can use that to monitor that real-time data, and from Cloud Datastore we can easily add extra functionality, like tracking train locations with Google Maps."

A separate architecture collects information for historical analysis. Stopped at a station or depot, trains upload their entire 2GB database once a day to a Cloud Storage Bucket, where it is transformed in Cloud Functions ready for storage in BigQuery. That data is complemented with the alarms and events the trains send every four minutes, which upload to Cloud Storage and are queued in Cloud Pub/Sub before transformations in Cloud Dataflow and delivery to BigQuery, ready for use by analysts and in ML models.

Experimenting with automated machine learning models

Talgo maintenance experts interpret the data from trains, comparing variables to keep track of the condition of each vehicle, make predictions, and decide what needs to be repaired and how. Now the company plans to create machine learning models to assist in that predictive maintenance using condition-based monitoring.

"We are experimenting with sending data using Cloud IoT Core and Cloud IoT Edge, and ingesting it with the same Cloud Pub/Sub, Dataflow, BigQuery architecture that we already use," says José. "We are using Cloud IoT Edge to load SW libraries into on-board equipment. That means we can control every single sensor, to change on-board patterns and rules, filter the information that trains send to the ground, and send messages directly to the sensors in the train." The collected data can then all be used for machine learning in Tensorflow, Cloud Machine Learning Engine, and Cloud Datalab.

"We can easily send results from our Google Cloud Platform machine learning model to mobile devices using App Engine," says José. "Predictions, recommendations, and evaluations of the condition of trains can be sent directly to engineers anywhere, and we can automatically generate reports using Google Data Studio, so maintenance can be more effective."

"Google Cloud Platform is the core of our maintenance system for all our trains. We have no doubt that we will continue putting our trust in it as we develop our use of machine learning, artificial vision, and Cloud IoT Core. For us, this is a collaboration."

José Antonio Marcos, Chief Maintenance Engineer, Talgo

High-speed maintenance for high-speed trains

Talgo aims to use cloud technology to move from preventative maintenance to condition-based maintenance, based on machine learning predictions. "One of the neural networks we are developing can predict failure in a rolling bearing one to two weeks before it occurs," says José. "We predict the temperature of the bearing based on data from our sensors, and if our prediction is wrong by a certain margin, we can determine in advance that the bearing is going to fail." Soon, the company will complement the trains' on-board data sources with installations on the tracks, too, that will take measurements from the exterior of trains as they pass through at speeds of around 300 kilometers an hour.

"Our new automatic inspection equipment 'TALVI' uses cameras with artificial vision to detect if anything is loose or broken on the side of the train," says José. "It can see if the thickness of the contact plate is within acceptable limits, and inspect the entire rolling assembly as it passes through. This technology will mean that instead of needing to go to the depot every week, a train can go every three weeks. It takes six hours to inspect a high-speed train, so we would save more than 200 hours of depot inspection time per train, every year."

"Google Cloud Platform is the core of our maintenance system for all our trains," says José. "We have no doubt that we will continue putting our trust in it as we develop our use of machine learning, artificial vision, and Cloud IoT Core. For us, this is a collaboration."

About Talgo

Founded in 1942, Talgo is a Spanish manufacturer of intercity, standard, and high-speed passenger trains with a history of innovation. The company was listed on the Madrid Stock Exchange in 2005.

Industries: Manufacturing
Location: Spain

About OpenSistemas

Based in Madrid, Spain, OpenSistemas is an award-winning cloud consultancy specializing in open source solutions.