back to list

Project: Visual Analytics of Latent Representations in Deep Learning Models for Industrial Time-Series Data

Description

Internship Title
Visual Analytics of Latent Representations in Deep Learning Models for Industrial Time-Series Data

Introduction
ASML is developing deep learning models that learn abstract representations from thousands of sensor signals collected across its worldwide machine fleet. These latent representations capture complex machine behavior but are difficult to interpret. This project focuses on developing visual analytics methods that make these learned representations understandable and actionable for engineers and data scientists.

Objective
Develop visualization and analysis techniques that reveal how high-dimensional time-series data are represented in the latent space of deep learning sequence models, supporting model understanding, root-cause analysis, and prediction interpretation.

Background and Scope
Sequence models such as GRUs and Temporal Fusion Transformers learn compact latent representations that summarize machine behavior over time. These representations are critical for prediction tasks but are often treated as black boxes. The student will investigate dimensionality reduction, clustering, and visual analytics techniques to explore latent spaces and identify meaningful patterns, machine states, anomalies, and temporal trajectories. The work will focus on linking latent representations back to original signals and operational events to improve interpretability and support engineering decision-making.

Deliverables

·         Literature review of latent space analysis and visualization techniques for deep learning models.

·         Exploration of latent representations generated by industrial sequence models.

·         Evaluation of dimensionality reduction and clustering approaches.

·         Interactive visual analytics prototype for latent space exploration and temporal pattern analysis.

·         Methods for linking latent features to machine behavior and input signals.

·         Final report, thesis, and presentation summarizing methodology and results.

Details

Student

CK

Cansu Kars

Supervisor

Stef van den Elzen

Secondary supervisor

Nicola Pezzotti

External location

ASML