Cross-Turbine Transfer Learning for Acoustic Fault Detection

Motivation

A machine about to fail usually sounds different first. Acoustic monitoring turns that into early warning: a microphone on a turbine, a model that flags the anomaly before the failure [1]. It is cheap, non-invasive, and it works, on the turbine you trained it on.

The catch is that every turbine sounds different. New geometry, new room acoustics, new background noise, and a model that scored well collapses. Retraining from scratch needs labelled fault data from every machine, which almost nobody has. If a model could transfer from one turbine to another, condition monitoring would scale from a single demo to a whole fleet.

This thesis studies whether it can, on real hydropower data.

Problem statement

Acoustic fault-detection models do not transfer well across machines, and it is not clear which technique fixes that best. Fine-tuning needs target-machine labels that are scarce; domain-adversarial training [3] and feature normalisation promise label-free transfer but are rarely compared head to head on real industrial recordings [4]. Without that comparison, a deployment has no principled way to move a model to a new turbine.

That is the gap. This thesis measures cross-turbine transfer on real pumped-storage recordings plus public benchmarks, comparing the main transfer strategies under the same conditions.

Research questions

How much does acoustic fault-detection accuracy drop when a model trained on one turbine is applied to another, and what drives the drop (geometry, noise, fault type)?
Which transfer strategy (fine-tuning, domain-adversarial training [3], feature normalisation) recovers the most performance, and how much target data does each need?
Can transfer be validated with little or no labelled fault data on the target machine?

Suggested methodology

Assemble the data: real recordings from our pumped-storage collaboration, public benchmarks (MIMII [1], DCASE [2]), and a controllable fault simulator.
Train a baseline detector per source turbine and quantify the cross-turbine performance drop [4].
Apply and compare transfer strategies: fine-tuning, domain-adversarial training [3], and feature normalisation, holding the model and evaluation fixed.
Vary the amount of target-machine data to trace each strategy’s data-efficiency curve.
Deliver a practical recommendation: which strategy to use, and how much target data it takes, to move a model to a new turbine.

What you need: machine-learning basics and Python. Signal-processing interest helps, we bring the acoustic pipeline and the plant data.

Previous theses in this line

Acoustic Anomaly Detection in Hydropower Plants, Dominic Tobler.
Acoustic Sensing for Anomaly Detection: Spectral-Temporal Anomaly Detection for Remote, High-Altitude Hydro-Power Plants, Nicolas Keller.
Distributed Acoustic-Vibration System for Industrial Predictive Maintenance, Maximilian Huwyler.
Multimodal Sensing for Acoustic-Vibration Based Fault Localization in Hydropower, Zakaria Omarar, and the paper Multi-Modal Acoustic-Vibration Array Monitoring for Anomaly Detection and Source Localization in a Pumped-Storage Turbine (Zhang et al., Viennahydro 2026). The group’s From Noise to Knowledge study compares the anomaly-detection models this topic transfers.

References

[1] H. Purohit, R. Tanabe, K. Ichige, T. Endo, Y. Nikaido, K. Suefusa, and Y. Kawaguchi. “MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection.” DCASE Workshop, 2019.

[2] Y. Koizumi, S. Saito, H. Uematsu, Y. Kawaguchi, et al. “Description and Discussion on DCASE Challenge Task: Unsupervised Detection of Anomalous Sounds for Machine Condition Monitoring.” DCASE Workshop, 2020.

[3] Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. Marchand, and V. Lempitsky. “Domain-Adversarial Training of Neural Networks.” Journal of Machine Learning Research, 17(59):1-35, 2016.

[4] S. J. Pan and Q. Yang. “A Survey on Transfer Learning.” IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2010.