Along the way, Nexans and Basefarm got a taste of a valuable side of machine learning.
While a great deal of routine maintenance occurs at fixed time intervals, predictive maintenance using machine learning can spur interventions when the need arises. For example, fully functional components might be replaced every three years, even though they only start showing signs of being worn out after five years. Even worse, components may be planned to be replaced in a couple of years, but break down before the next maintenance interval.
“We had a break in the work. When we started things up again, disturbing noises were recorded. The observation came from a component that was at the end of its life cycle. So there we got an extra bonus for our efforts,” says Johansen.
Machine learning creates different processes
Most businesses are used to being able to specify, assess, order and implement solutions of all kinds. The procurement of machine learning-based solutions works differently. Here, the solution is created in a fully finished form for the very specific local situation, with standard hardware, software library and storage from hyperscalers.
The very concept of “machine learning” is descriptive of this. The solution learns from the sensors that are used. With the massive powers of the hyperscalers and applications, this knowledge is used to build algorithms. The algorithms can then perform the correct action from the cloud or be inputted where actions should occur.
Technology and tasks
- Image analysis – machine vision/recording, both overviews and detail views (zoom). Evaluation with machine learning
- Vibration analysis where “no vibration” is central. Machine learning
- Audio recording with machine learning
- Data collection for 30-60 seconds every half hour
- Large amounts of data are sent to hyperscalers (GCP) for further analysis using a customized machine learning library