Federated Multi-Task Learning on Acoustic Signals for Predictive Maintenance

Studenteropgave: Kandidatspeciale og HD afgangsprojekt

  • Kristian Juul Tilsted
4. semester, Matematik-teknologi (cand.polyt.), Kandidat (Kandidatuddannelse)
This Master's thesis investigates the possibility of developing a federated multi-task learning algorithm with focus on minimal computational complexity and low communications requirements, for use in a distributed predictive maintenance setting. This Master's thesis is a collaboration with Grundfos A/S on predictive maintenance on acoustic data, from their pumping systems. The proposed solution - developed in this work - is compared to the current state-of-the-art federated multi-task algorithm, the MOCHA algorithm. We compare the two solution using ROC curves, which yielded an average decreased performance, in terms of AUC, for 100 runs of each algorithms, across 9 tasks, of 0.095. Though the proposed solution performs worse, the computational complexity is about 150 times less, in terms of FLOPs, compared to that of the MOCHA algorithm.
Udgivelsesdato4 jun. 2021
Antal sider95
Ekstern samarbejdspartnerGrundfos DK AS
Senior Data Scientist Rasmus Engholm rengholm@grundfos.com
ID: 413353905