How Hitachi Uses AI to Detect Railway Overhead Line Defects in Near Real Time

Hitachi's railway division builds the infrastructure that keeps train networks running—including systems for monitoring the overhead lines that power electric trains. Keeping those lines in safe condition is a constant, costly challenge. Traditionally, rail network operators maintained infrastructure by walking tracks on foot, watching from moving trains, or sending dedicated measurement trains that ran infrequently and could only detect certain fault types. When a line broke, the result was unplanned delays and disruption costs. "We set out to innovate and disrupt traditional railway maintenance, leveraging existing train fleets and empowering them with state-of-the-art AI technology," said Andreas Herman, Lead Data and AI Architect at Hitachi.

Hitachi's solution was to install cameras on existing trains and upload the footage to the cloud for analysis—but this created a new problem. The cameras captured footage across 40,000 km of overhead lines, the distance equivalent to one lap around the world. Processing and analyzing that volume of video data efficiently, cost-effectively, and with the ability to generate near-real-time alerts required a platform that traditional cloud data warehouses couldn't provide. "We needed a digital solution that was flexible, scalable and cost-efficient," Herman said.

Hitachi chose the Databricks Lakehouse Platform on Azure. Databricks was a fully managed platform that removed the need to maintain Apache Spark infrastructure, addressed all AI model requirements, and could scale to handle the large volumes of video data streaming from trains. Delta Lake became the default storage format for all data pipelines. MLflow manages model experimentation, training, deployment, and tracking. Databricks SQL powers monitoring dashboards that track alert volumes, model performance, and system health. Shared workspaces and notebooks let the team collaborate on code, reuse models, and work more efficiently with a smaller team.

The system now ingests video data from trains, processes it through ML models, and generates automated alerts on defects or displacement of overhead lines. A worker gets dispatched to the area to fix the issue before it causes a service disruption. What used to take days—engineers traveling to sites to look for faults—now happens in near real time, remotely. "Our clients can now be much more proactive instead of reactive," said Avinash Kumar, Data Scientist at Hitachi. "Now the rail network owners we support can identify issues much faster to prevent costly disruptions to their services."

Hitachi has discovered thousands of equipment risks and line displacements through the system, proactively resolved before they caused outages. The estimated savings for Hitachi's customers run to millions of pounds. The engineering team now operates with fewer people focused on data infrastructure and more focused on solving customer problems. "Using Databricks has made us more efficient with a smaller team, which is incredibly powerful," said Herman. "Now, a single team can create impactful innovations that benefit thousands of customers every day."