How to read these slides?

Navigate via arrow keys between slides Overview

Click on the menu bar items to navigate to chapters

Click here for PDF version
Note: No videos in PDF version!
SCIENCE PASSION TECHNOLOGY

Progressive Visual Segmentation of
Vibration Signals to Detect Change Points

Julian Rakuschek, Johanna Schmidt, Udo Schlegel, Michèle Posch, Jutta Isopp, Tobias Schreck

VMV 2026 - 25.09.2026

All about vibrations

Engine

+

Sensor

=

Vibration signal

Why analyze vibrations?

Find long-term deviations

Why analyze vibrations?

Find long-term deviations

Why analyze vibrations?

Find anomalies (short-term, unexpected)

Why analyze vibrations?

Find anomalies (short-term, unexpected)

Why analyze vibrations?

Predict faults

Why analyze vibrations?

Predict faults

A good segmentation can help solve these tasks!

Condition Monitoring

Predictive Maintenance

Anomaly Detection

The transitions between segments mark change points

Related Work (1)

Ali, Mohammed et al. “TimeCluster: dimension reduction applied to temporal data for visual analytics.”
The Visual Computer 35 (2019): 1013 - 1026.

Related Work (2)

Bernard, J. et al. “TimeSeriesPaths : Projection-Based Explorative Analysis of Multivariate Time Series Data.”
International Conference in Central Europe on Computer Graphics and Visualization (2012).

Related Work (3)

Schmidl, Sebastian et al. “DendroTime: Progressive Hierarchical Clustering for Variable-Length Time Series.” International Conference on Extending Database Technology (2026).

This is hard to read

We transform it to the frequency domain

The big problem with vibrations:
Large data volumes

High sampling rate = large data volumes

Example: 30 kHz (30 000 samples per second)

Assumption: 8 byte per sample

After one minute: 14,4 MB
After one hour: 864 MB
After one day: 20,7 GB
After a year: 7,5 TB

How do domain experts approach this?

Signal Analysis via Manual Window Positioning

How about 10TB?

Manual scanning is tedious!

Our approach: Sampling

We iteratively construct an overview

In many cases a small fraction of the data suffices!

Do we need to save the full signal?

No!

We only save a FFT histogram per sample into the database.

Further representations possible

Time Series Projections

We start with an empty timeline

Next we randomly draw samples

The subsequences are groupes by similarity

The background is colored by closest label

The height is adjusted based on the density

For a subset of samples we can show representations as thumbnails.

We obtain a segmentation

With more samples, the segmentation becomes more accurate.

FFT Feature Descriptors

5 samples

FFT Feature Descriptors

10 samples

FFT Feature Descriptors

25 samples

FFT Feature Descriptors

50 samples

Projection-Based Feature Descriptors

5 samples

Projection-Based Feature Descriptors

10 samples

Projection-Based Feature Descriptors

25 samples

Projection-Based Feature Descriptors

50 samples

How can users steer the sampling?

Defining intervals to constrain the sampling

(Accelerated video footage)

How can users explore the result?

Classic zooming and panning

Inspecting samples in detail

Other sampling strategies

Binary Sampling (1)

Binary Sampling (2)

The result: Sharp boundaries

Gap Filling (1)

Gap Filling (2)

The result is a uniform distribution of samples

Linear Sampling

Pick samples in chronological order

(Accelerated video footage)

Formative Evaluation (1)

With two domain experts from the field of vibration analysis

Round 1 / 3

  • Discusssion of initial idea.
  • Formulation of requirements.

Round 2 / 3

  • Experts used the system to solve a task.
  • The task was to identify one change point in a vibration signal.
  • We identified several points for improvement.

Formative Evaluation (2)

Round 3 / 3

  • Experts used own dataset.
  • Tried to replicate a finding.
  • Successfully used system to identify leakage in a machine.

Early analysis results:

Final analysis results:

Where does it not work?

Small patterns (e.g. Anomalies) can be hard to detect

Chirp signals are challenging

Future Work

  • Multi-resolution visualization to reveal small, localized patterns
  • Uncertainty-aware segment boundaries with gradual transitions
  • Broader user studies to validate findings and segment discovery
  • Transient Visual Analytics to remove redundant samples and reduce storage
  • Scale to longer signals and larger analysis windows
Progressive Visual Segmentation of
Vibration Signals to Detect Change Points

Thank you!

Julian
Rakuschek
Johanna
Schmidt
Udo
Schlegel
Michèle
Posch
Jutta
Isopp
Tobias
Schreck

Open Source

PRESENT and ENVIRON-HYDRO

Slides