Good Vibrations: Finding a Solution during a Hackathon

For Tim Klaas of the “TUBOT” hacker team from Berlin, it was precisely this problem that held a special appeal: “We’re all absolutely enthusiastic about machine learning and knew right away that we wanted to get to grips with it.” TUBOT is made up of “TUB,” the initial letters of Technical University Berlin and “BOT” for robotics. The five-member team – consisting of four electrical engineers and one computer scientist – tested four different methods and ultimately relied on an autoencoder to solve the task. This artificial neural network first compresses the input data and then reconstructs it. This produces a reconstruction error, which is used to train the network. The particular difficulty was that it was not obvious what an anomaly looked like, since only error-free training data was available. “We therefore built the algorithm using the training data so that error-free data would be reconstructed correctly, but in the anomaly case there would be a reconstruction error. We were also able to implement the same principle much more efficiently in terms of training time using the principal component analysis,” says Klaas, a master's student. And the effort paid off: Although it was the very first hackathon for the TUBOT students, they were able to convince the Leadec jury with their solution. “We are really impressed by the precision and technical know-how with which work was done here. Now we want to implement an even further optimized variant on our installation as well,” says Andreas Zeller.