Menu

3 key components that are essential for implementing image recognition

Jasper Derikx
3 juni , 2020

The Netherlands has the busiest railway network of Europe, so maintaining the tracks comes with interesting challenges and is of great importance. This article shows how artificial intelligence has made its introduction in this field. After showing three real life use cases of image recognition, this article will touch upon the three key components these examples have in common and which form the basis of a successful implementation.

For several years, every single centimeter of track has been photographed by a custom build train equipped with ten cameras, all with different directions and level of detail. From that point in time the track could stay in operation during inspection since the special train can be scheduled in between the regular train traffic or even at night. These images were checked manually for irregularities and defects, but the number of images made it impossible to inspect all of the Netherlands. Needless to say, that these images are also a gold mine for data scientists and with the current developments in AI the inspection of the track can be automated even further.

Image for post

What can AI do?

With the use of artificial intelligence, more information can be retrieved from the images. Here we will show you 3 examples:

Registration - Keeping track of 7000 kms of railway is rather difficult, especially because every day hundreds of construction workers from different parties maintain the track. Luckily, image recognition can identify objects in pictures, and even classify them based on (for example) width, type, etc. This process increases the amount of descriptive data significantly. Moreover, by analysing the images with regular intervals, changes can be determined and automatically registered.

Image for post

Monitoring - To keep passengers save, critical components in the track are monitored 24/7. But this is not possible for all components. Take for example the roughly 100 million spring clamps, which keep the track in the right position.

Image for post

One broken spring clamp is not an issue at all. However, more than 3 broken spring clamps within 10 meters is. Manually checking this would be heaps of work and prone to error. That is where machine learning can come to the rescue. With a model that can distinguish broken clamps this is an easy task to automate!

Prevention - Recognizing broken assets is useful but preventing failures would be even better. Replacing the whole network out of precaution is of course way to expensive, so how can we detect the weak components? Using failure data of the past, a model can be trained to indicate the degradation of an asset. Here we see an example of insulation joints of which there are 40.000 throughout the Netherlands.

Image for post

With this model the state of an asset is automatically determined, which enables us to adapt the maintenance schedule based on the current situation.

The key components

These use cases have similarities but also subtle differences. We will now focus on three main components which are essential for implementing all of them and which are the first topics to consider when starting an AI project.

Image for post

Solid data architecture - As you can imagine, frequently capturing the whole network with ten different angles piles up to an enormous amount of data. Although storage expenses are usually manageable, in this case storage is not straightforward. Not to mention processing and analyzing these images. Therefore, it is essential that this work is done by specialists. The main reason that this investment is easily justified is that there are lots of use cases (like the three mentioned above) that all rely on the same data. Hence, a solid data architecture pays off in multiple projects simultaneously.

Understandable model performance - The models represent the intelligence of the system, so a proper understanding of how to train and maintain them is required. Besides the usual error-measurements that are monitored, it is key to identify the model’s mistakes. Did the model falsely identify certain objects? Or did it miss any? Especially the latter can be quite difficult to measure.

Business driven models - By now you have seen just a glimpse of the endless possibilities in this project. Creating a model alone, however, is not enough. If you want to effectively apply AI, you need to understand the business case. Take for example the degradation of insulation joints which are discussed above. Maintenance is done by contractors who plan their work months in advance. They do not have the flexibility to reschedule their work for the coming week. A complicated model that predicts the failure with a precision of one week might be great achievement from a data science viewpoint, but superfluous from the business perspective. Instead, indicating the 100 insulation joints with the highest risk of failure in the coming 6 months fits the current needs much better. In general, you want to adapt your model given the possible actions, in stead of acting on your model’s predictions.

We hope that this article will help you in defining your own image recognition use cases. But before you start working on it, think carefully about the three main components to every project (data, models, and business) to ensure your implementation becomes a success.

Thanks to Richard Bartels, Guido Tournois, and Lieke Kools.