Impact of Data Reduction on Acoustic Fault Detection

Motivation

High-fidelity acoustic monitoring produces a firehose of data: around 3.4 GB per hour per sensor. A remote hydropower plant on a thin uplink cannot ship that, so the data has to be reduced before it leaves the site. The question is what you lose when you do.

Reduce too little and you cannot transmit; reduce too much and you throw away the faint signal that a fault detector depends on [1]. Between those extremes is a budget every real deployment has to pick, usually by guesswork. There is no map of how much reduction acoustic fault detection can actually tolerate.

This thesis builds that map.

Problem statement

The trade-off between data reduction and fault-detection accuracy is undocumented. General audio codecs [2] and learned embeddings [3] are tuned for perceptual quality or classification, not for preserving the subtle cues of an early fault, and their effect on anomaly detection is untested across fault types and severities [1]. A deployment therefore has no principled way to choose a reduction level.

That is the gap. This thesis produces a compression sensitivity map: for each family of data reduction, how far you can compress before detection degrades, broken down by fault type and severity.

Research questions

1. How does detection accuracy fall as a function of reduction, for each family (codecs, downsampling, statistical summaries, spectral features, learned embeddings [2], [3])?
2. Does the safe reduction level depend on fault type and severity, and how much?
3. Which family gives the best accuracy per transmitted byte for a remote-plant budget?

Suggested methodology

1. Take a labelled acoustic fault dataset (our pumped-storage recordings plus MIMII [1]) spanning fault types and severity levels.
2. Implement the five reduction families: codecs (e.g. Opus [2]), downsampling, statistical summaries, spectral features, and learned embeddings [3], each at several levels.
3. Run a fixed fault-detection model on each reduced stream and measure accuracy vs bitrate.
4. Assemble the compression sensitivity map: accuracy-vs-reduction curves per family, split by fault type and severity.
5. Turn it into a deployment guide: how much reduction a given uplink and fault profile can afford.

What you need: Python, and an interest in signal processing and embedded constraints. We bring the acoustic data and the detection pipeline.

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 provides the detection models this topic stresses under compression.

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] J.-M. Valin, K. Vos, and T. B. Terriberry. “Definition of the Opus Audio Codec.” RFC 6716, IETF, 2012.

[3] S. Hershey, S. Chaudhuri, D. P. W. Ellis, et al. “CNN Architectures for Large-Scale Audio Classification.” IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017.