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
SCIENCE PASSION TECHNOLOGY

WP2 Prototypes

Julian Rakuschek

16.01.2025

Our Main Quest

D2.2 Visualizations for AI results and for AI explainability

Task 2.4 Development of a web-based visualization including domain-specific visualizations

Task 2.5 Development of visual explanations for AI results

Currently four prototypes:

Cluster and Search

Anomalies

Forecasting

Vibrations

All are meant for Task 2.4, Task 2.5 is WIP

The Prototypes are Task-Oriented

Cluster & Search

Cluster & Search Prototype Goals

  1. Find pattern groups and compare across seasons
  2. Find anomalies
  3. Compare channels

Clustering Recorded Noise Level from a Seminar Room

Sound time series - each segment (day) is colored according to its clustering group
Each cluster is represented as the average of all its members
The calendar view enables the user to find recurrences - such as a lecture on each monday.

PV Energy Production of Households

Another dataset - PV energy production of households in southern germany. The daily segmentation shows the shift of charge and decharge over the day and the season of the year.

Anomalies

Anomalies in time series (Selection)

Check every time series by hand?

Let algorithms do the work!

Introducing AnoScout

A "sandbox" to check which algorithms work well and for exploring anomalies in the dataset.

Anomalies are represented as cards:

Manual Inspection

In the manual inspection interface, users may explore how different algorithms performed to detect a particular anomaly. Each algorithm output is a scoring - a time series where higher values correspond to anomalies in the data.

Exploring Anomalies

Exploring anomalies is achieved through a linked view, users may further provide feedback on the importance of an anomaly to satisfy a specific information need.

How do we gain an overview of all clusters?

Clustering!

Clustering Anomalies

The anomalies can be arranged by similarity in the scatterplot, such that similar anomalies are grouped together. This enables the user to discover patterns in the dataset.

Main features of AnoScout summarized

  1. Exploration pipeline for anomalies in time-oriented data.
  2. 7 algorithms for computing anomalies.
  3. "Playground" for testing various algorithms.
  4. Using user labels to fine-tune the system.

Towards XAI

AutoEncoders learn normal patterns.

Show which normal labels are the most important ones.

Application Scenario

  • A company wants to install a new machine.
  • The machine conducts an etching process (semiconductor manufacturing).
  • Each etching process is recorded through a sensor (e.g. pressure, temperature, and gas)
  • We want to use AnoScout to:
    1. Find possible anomaly patterns.
    2. Check which algorithms work well.

Forecasting

How can we build a "sandbox" to explore forecasting models?

Introducing PredictPal

Just like AnoScout, but for Forecasting.

Prediction Models

ARIMA

AutoRegressive Integrated Moving Average

VAR

Vector Autoregression

Analysis View

The user selects a subset of the time series for training the prediction models and can immediately verify the model based on the visualization.

History of Models

The history keeps track of models found in the past.

A municipal office worker John Doe needs to predict the traffic load at Intersection X

The video shows the workflow of using PredictPal - first, we upload a dataset, then run statistical tests to ensure applicability of ARIMA and VAR. Next, the user tries various subsets of the time series and seasonilty configurations to arrive at a suitable model.

Possible Use Case: Solgenium

Caution: This is a different prototype for a forecasting visualization!

The Solgenium prototype is built to forecast the workload of a hospital. The dataset might be a possible application scenario for PredictPal.
  • Solgenium Prototype = Specific Use Case
  • PredictPal = Generic

Therefore: Merge Predictpal and Solgenium

Towards XAI: ShapTime

Vibrations

The Problem with Vibrations

Can you tell the difference?

A hidden signal

Linecharts are useless!

The Time Delay Embedding (TDE)

Noise is not exciting ...

... but oscillations result in circles!

What can we do with this?

  1. Visualise change points
  2. Cluster signals
  3. Find labels

Our Prototype

Vibrana

EuroVis 2025 Submission

In Review 🤞

Can we find a change point in the signal?

Vibrationens of a hydropower plant:

The TDE is a fingerprint evolving over time!

Applications

Engines

When can we detect wear?

Bearings

Which are faulty?

While the spectrogram is sometimes better ...

... what about some added noise?

Which are faulty?

Automatic Clustering

What about Explainable AI?

Task 2.5 needs more thinking

Task 2.4 Development of a web-based visualization including domain-specific visualizations

Task 2.5 Development of visual explanations for AI results

Task 2.4 (Vis) Task 2.5 (XAI)
Cluster ✅ ❌
AnoScout (Anomalies) ✅ 🤔
Predictpal (Forecasting) ✅ ✅
Vibrana (Vibrations) ✅ ❌

Some Ideas

Clustering
  • AI not reasonable in this task
AnoScout
  • While AI can used, it is not advisable!
  • Matrix Profile methods outperform AI methods
  • AI only useful when classifying with user-labels
  • Show label influence for AI classifier (XAI)
PredictPal
  • Merge Solgenium Prototype and PredictPal: Show error rate (XAI?)
  • Implement LSTM Forecasting with XAI
Vibrana
  • User-labels for the similarity search
  • Show which labels have which amount of influence on findings
  • Strictly speaking not AI!

Contact

Julian Rakuschek

julian.rakuschek@tugraz.at

Prof. Dr. Tobias Schreck

tobias.schreck@tugraz.at

Slides

https://presentations.rakuschek.at/2025-01-16-present-konsortialmeeting

Questions?