On a daily basis, Swiss Federal Railway (SBB) manages 10,671 trains, serving 1.26 million riders on 3,232 km of track. To ensure the safety of this infrastructure, as well as the safety of its passengers, SBB uses “diagnosis” trains equipped with multiple high-resolution cameras and other sensors to obtain images of railway tracks while traveling at speeds of up to 160 kmh.
SBB’s traditional data processing approach was conducted in real-time on the diagnosis train, which produces false positives and negatives — to the extent that railway experts still need to go onto the track to physically inspect anomalies. This requires a lot of human labor, with people operating in dangerous environments that are sometimes impractical to access.
SBB partnered with CSEM (Swiss research and development center) to address this challenge by initiating the “Railcheck” project, which aims to use deep learning technologies to:
improve the detection and classification of faults in railway tracks,
minimize the number of onsite visual inspections, and
minimize the time experts spend on false positives.
The first and very important step is to have the right data, which serves as the “new source code” for deep learning problems. CSEM went through the following process to make sure their data pipeline would enable a successful deployment:
Data Gathering. SBB is planning to inspect 3,800 km of track per month by the year 2020. Images capture different weather conditions (rain, snow, ice) and artifacts (leaves, dirt, etc.).
Preprocessing. A model is trained to detect regions of interest (for example, railway beams, clamps) in the images.
Anomaly Detection. CSEM used generative adversarial networks (GANs) — which consist of deep neural net architectures composed of two networks, pitting one against the other — to define clusters within datasets. These clusters were then used to identify anomalies in railway components.
Fault Detection/Classification. Classification requires supervised learning. There are, in principle, about 20 different fault categories, but SBB decided to condense these to five categories (that is, welding, joint, surface defect, squat and wheel slip) since determining the appropriate fault category is not always trivial.
Data Labeling. As is common with many businesses starting with deep learning, very little labeled data was available when the project started. Additionally, certain fault categories require the judgment of well-trained railway experts who may not always agree. As a solution, CSEM decided to apply a crowd-sourced approach to decision-making:
Leveraged transfer learning, a machine learning technique that focuses on storing knowledge gained while solving one problem and applying it to a different but related problem.
Performed inference, the deployment of trained deep learning models.
Uploaded annotated images to a web server, where multiple railway experts could give feedback. Having the feedback of multiple experts allowed for different opinions to be taken into account.
Labeled data based on the feedback, fed the data back into the training set, and then retrained the models.
By doing the above steps iteratively, the model accuracy improves over time, in what is known as lifelong machine learning.
NVIDIA DGX Station, the purpose-built AI workstation, has played a significant role behind the scenes, helping accelerate deep learning training and improving detection accuracy.
“DGX Station just works! It was up and ready within a few hours,” says Nathalie Rauschmayr, a machine learning engineer at CSEM. “The best part about the DGX Station is it’s a complete system with the latest and most optimized deep learning frameworks. Instead of worrying about how to configure many low-level components on the system, I can focus on gathering the right data, training the AI workloads, and working with experts to identify faults accurately.”