In the world of railway maintenance, efficiency is paramount. From the tracks that guide the trains to the switches that direct their path, every component plays a crucial role in ensuring safe and smooth journeys. To help uphold this level of operational excellence, Railway Robotics has developed Railchap, an innovative rail-bound robot. Its primary function is to inspect and lubricate slide plates, essential components of railway switches that need to be consistently well-maintained to guarantee the smooth transition of trains from one track to another.

This article details a case study of transfer learning and MLOps for industrial applications, showing an example of how BearingPoint's Data & Analytics teams take theoretical concepts into practical solutions. 

The critical role of slide plates

A key function of Railchap is slide plate maintenance, and its routine involves both lubrication and inspection. These metal plates are positioned beneath the movable sections of the tracks during a switch. Regular lubrication of these plates is vital to guarantee the unimpeded movement of the steel parts over them. The frequency of this task varies with the seasons, requiring biweekly attention during winter, while in the summer, the interval stretches to about every five weeks. 

Historically, slide plate maintenance is an intensive manual task, which demands a significant workforce on the field. Not only is this process time-consuming, but it is also susceptible to human error. However, with the advent of Railchap, we are moving towards a future where this process is automated, significantly improving the efficiency and accuracy of slide plate maintenance, all while reducing the need for manual labor. 

The power of Computer Vision 

To achieve this level of automation, we turned to a powerful tool: computer vision. Computer vision involves capturing and processing visual data, like the way human eyes do. For Railchap, it plays a vital role by enabling the robot to "see" slide plates. This capability helps Railchap determine when it's near a slide plate and find the perfect moment to activate the lubrication system. 

Building an end-to-end MLOps system 

To operationalize our computer vision models, we embarked on the journey of building a comprehensive Machine Learning Operations (MLOps) system. This system was architected and implemented to label and process image data, train our models, and deploy the vision modules on the robot. 

As part of the Microsoft for Startups program, we leveraged Microsoft's Azure Cloud for our machine learning infrastructure. This powerful cloud platform gave us access to robust IoT capabilities and substantial data storage. 

The images taken by Railchap's onboard camera were moved to our cloud-based storage. We then created inspection scripts in a more user-friendly environment, also provided by Azure, to examine the contents of these image files. Finally, the image files were processed and ready to be used in model training. The system architecture is displayed below, for the most tech-savvy of our readers. 

Model selection and training: Choosing RetinaNet for object detection 

While researching, we came across several projects that used computer vision in the railway industry. Although none were specifically focused on identifying slide plates, the general focus on railway component detection was encouraging. 

We decided to use a model called RetinaNet, developed by Facebook AI Research, for our task. This model is known for its ability to accurately detect and locate objects. The architecture of RetinaNet is divided into two main parts - one learns the general features of the images, and the other uses these features to make predictions. 

Initial results 

Preliminary results have been promising! The trained model consistently identified the slide plates in our test set. 

Next Steps: Moving toward edge processing 

The next steps involve deploying our trained model at the edge, meaning that machine learning inference will be performed on Railchap's onboard hardware. This process includes optimizing the object detection model and setting up a pipeline for fast, effective image processing. These steps will ensure real-time processing and responsiveness, a critical aspect of automated systems like Railchap. Inference speed is essential for this application, as every minute the robot spends inspecting the railway is another minute without regular train traffic. Railways need to be booked for inspection, and the rail network is vast. 

Using transfer learning for industry-specific applications 

In this brief overview, we have shared our journey of automating a critical railway maintenance task using transfer learning and edge processing. As displayed here, the fusion of industry-specific knowledge with innovative machine learning techniques can lead to efficient and practical solutions. This powerful combination has allowed us to help push the boundaries of what is possible in railway maintenance, and we look forward to seeing where these advancements take Railway Robotics next. 


Markus Skagemo
Data Scientist


Christian Aalby Svalesen
Lead Data Scientist