Aleksanteri Hämäläinen: Development of Data-Efficient and Highly Generalizable Deep Learning Models for Condition Monitoring of Rotating Machines

Development of Data-Efficient and Highly Generalizable Deep Learning Models for Condition Monitoring of Rotating Machines

By Aleksanteri Hämäläinen

Overview

Rotating machines, such as electric motors and turbines, consume and produce a significant share of the world’s electricity. Effective condition monitoring is therefore key both for industrial cost savings and environmental goals.

In my publication supported by the Walter Ahlström Foundation, ”Rapid Deployment of Deep Learning-Based Condition Monitoring for Rotating Machines” (IEEE Access, 2025), we have developed a new method based on deep learning and prototypical networks, called Rapid-FSCM. The method is an important step towards intelligent condition monitoring that better takes into account industrial needs and realities than research focused purely on laboratory conditions.

Challenge

One of the most significant practical challenges is that deep learning requires substantial amounts of training data on identifiable faults. This data is naturally not available for new or relatively new machines, and may not be available for years.

However, common examples of similar faults, such as bearing and gear faults, are almost certainly available, as bearings in particular are used in nearly all rotating machines.

Innovation & Approach

The Rapid-FSCM method we developed enables the combination of extensive training data from other machines with a small number of examples from the target machine. This allows intelligent fault diagnostics to be deployed significantly earlier.

In addition, the method can be seamlessly used for anomaly detection even before any fault examples exist for the target machine, enabling partial benefits to be achieved for any new machine almost immediately.

Next Phase & Impact

The goal of the ongoing next phase of this research is to completely eliminate the need to collect fault examples separately for each machine.

By combining the often-overlooked understanding of physics from traditional mechanical engineering with data-driven learning, we can move from machine-specific solutions to machine-independent diagnostics. In this way, the same model can identify a crack in a gear tooth in any gearbox, rather than only in a specific gear of a specific machine.