. Analysis of different algorithms for identifying defects on the carrier line.
2. Strategy regarding the type of algorithm that should be used to identify defects on the carrier line and a hypothesis for unsupervised learning will be applicable.
3. Create a ranked list of the challenges that can arise in developing the machine learning algorithm and describe how the challenges should be handled.
The data used consists of 1,000,000 images collected on a 6.8 kilometer long distance between Insjön and Leksand. From the images, clippings have been extracted on only the carrier line, which was then used to train two different unsupervised learning models. The best model for identifying injuries was the model based on CutoutNet.
The results showed that a suitable strategy for continuing this feasibility study should be to use a combination of unsupervised and supervised learning. This is because in many cases unsupervised learning assesses that images on the carrier line that do not contain injuries are more aberrant than images that contain injuries and that no cluster with injuries alone was created. In continuing with the feasibility study, we recommend starting the project by using an unsupervised learning model to reduce the amount of data that needs to be reviewed. Subsequently, images can be annotated as damage or not damage to create a new algorithm based on supervised learning. When analysing the results and tests of algorithms during the feasibility study the following challenges were identified:
1. Unsupervised learning identifies many non-injury deviations
2. Parts of the railway infrastructure obscure the carrier line
3. Pictures on the carrier line are taken from below, which means that some damage may be obscured by the carrying line
4. The extraction of the carrier line from the image data set is not 100% effective
5. The amount of surface next to the carrier line that is extracted
6. Image quality
Some of these challenges are more difficult than others to deal with, but they will all be able to be handled. They can also be managed by combining supervised with unsupervised learning mentioned above and via the new measurement system that Latronix plans to set up in 2020.
With regard to continued work, our recommendation is to start a new project according to our identified machine learning strategy as soon as Latronix has upgraded its measurement system. The goal of the project should be to create an algorithm to identify the occurrence and position of strand fractures. The result of the project should be utilised by Trafikverket in production to avoid train stops due to broken carrier lines.
Stakeholders: Luleå tekniska universitet, InfraNord, Latronix AB, Aiwizio och Trafikverket.
Project leader: Jacob Tersander, Aiwizio