How to read these slides:

  1. Use arrow keys to navigate.
  2. Press O for switching to overview mode.
  3. Press O again to zoom in.

Vibrana

Analyzing Vibration Signals

Julian Rakuschek
Prof. Dr. Tobias Schreck

The Problem

One of them bears a secret

The Problem

One of them bears a secret

A Gravitational Wave

[Video Source: https://www.youtube.com/watch?v=Y3eR49ogsF0]

Preliminary: Time Delay Embedding

Exposing the Secret

Exposing the Secret

Why vibrations?

Strong Frequency Components!

Research Question:

Is the projection-based view effective to guide users towards interesting patterns?

Related Work

Vibration Analysis Framework

Vibration Analysis for Machine Monitoring and Diagnosis: A Systematic Review - Ghazali et al. (2021), Shock and Vibration

Using TDA to detect GWs

  1. Embed Signal in lower dimensional space
  2. Extract features through persistent homology
  3. Train a CNN
Detection of gravitational waves using topological data analysis and convolutional neural network: An improved approach - Bresten et al. (2019), arXiv.org

Time Series Paths

TimeSeriesPaths: Projection-Based Explorative Analysis of Multivariate Time Series Data - Bernard et al. (2012), Journal of WSCG

Exploring Embeddings

Visual Exploration of Relationships and Structure in Low-Dimensional Embeddings - Eckelt et al. (2023), IEEE Transactions on Visualization and Computer Graphics

Vibrana Showcase

Interaction Design and Workflow

Our Starting Point

Analyzing One Signal

Selecting a Subset
Marking annotations
Explore through 2D Projection

Selecting a Subset

Intermezzo: Advanced Brush

Using the Brush to add Annotations

Using the labels

Perform similaritiy search to find anomalies

How to decide anomaly threshold?

Use an anomaly-free sample as reference

Selecting Anomaly Free

Example for an anomalous sample

Everything below the lower red line is considered abnormal.

One good label is all you need!

Workflow Summary

Our Vision

Evaluation - WIP

Partnering up with domain experts

Open Problems

  1. Performance
  2. Information Loss when projecting
  3. Benchmark Datasets
  4. Evaluation

Conclusion

  1. Projection-based visualizations help
  2. Similarity search is often enough
  3. A little help from the user avoids complicated deep learning

Thank you!

Prof. Dr. Tobias Schreck

tobias.schreck@cgv.tugraz.at

Julian Rakuschek

julian.rakuschek@tugraz.at

http://cgv.cgv.tugraz.at/present/presentations/vibrana/