AI Day Conference

Programme Committee

Date: June 26th, 2024, 09:30 a.m. to 05:00 p.m.

Location: Lecture Halls K001 to K002, Faculty of Computer Science and Mathematics, OTH Regensburg, Galgenbergstr. 32, 93053 Regensburg


Opening and Keynote

09:30 – 09:40

Welcome and Introduction

Prof. Dr. Ralph Schneider, President and Prof. Dr. Wolfgang Mauerer, Chairman Director of the Regensburg Center for Artificial Intelligence (RCAI) at OTH Regensburg

09:40 – 10:10


To be announced

Session 1 ‒ The Art of AI: Strategies for Optimisation and Automation

Session Chair: Andrea Stich, Director Frontend Academy at Infineon Technologies AG

10:10 – 10:30

Revolutionizing Design of Experiments Using Bayesian Optimization

Dr. Lavinia Israel, Data Scientist at ams-OSRAM International GmbH

Bayesian optimization is a sequential strategy to optimize objective functions that are unknown and expensive to evaluate. It builds a surrogate model – usually a Gaussian process – for the objective and quantifies the uncertainty in that surrogate. It then uses an acquisition function defined from this surrogate model to decide where to evaluate next. At ams OSRAM we have many processes where process parameters need to be optimized to improve quality and efficiency. These processes do not have a functional representation and in general evaluations (= measurements) are expensive and time-consuming. The well-established method Design of Experiments has some drawbacks which we try to overcome by using Bayesian Optimization. In first projects we showed that we achieve optimal results with very few measurements.

10:30 – 10:50

Automation is Key: MLOps for High Quality ML

Ron Krestel, Machine Learning Engineer and Dr. Thomas Beer, Senior Solution Architect at Continental Automotive Technologies GmbH

Even though ML allow novel software solutions, its models often remain opaque, which has a negative impact on the trustworthiness. Trust is needed in critical applications like autonomous or medical systems, where ML models must be especially reliable and of high quality. Aside from the model, also the code and the data that was used to create the model need to be of high quality to detect and minimize erroneous behavior. A combination of known and novel software engineering practices is needed to assess and improve the overall quality.
In recent years, there have been efforts to support engineers in ML-specific tasks with concepts such as ML Operations (MLOps). MLOps streamlines ML development, boosts productity through automation and promotes quality software. This presentation introduces concepts for software quality assessment that simultaneously fulfill the principles of MLOps and thereby address the requirements of projects in practice.

10:50 – 11:10

Exhibition Pitches

All poster presenters

11:10 – 12:00

Poster Session & Networking

Session 2 ‒ Driving Innovation: AI in the Automotive Industry

Session Chair: Christian Miedl, Segmentlead Safety, Security and AI at AVL Software and Functions GmbH

12:00 – 12:20

AI-Based Smartphone Localization for Keyless Car Access

Dr.-Ing. Patrick Philipp, Machine Learning Scientist at Continental Automotive Technologies GmbH

We present a real-world machine learning solution for enabling precise AI-based smartphone localization based on UWB sensors in cars, which was primarily developed in Regensburg.
UWB sensors provide time-of-flight and signal strength information for ranging devices, which can be used to localize smartphones to be inside or outside of a car. Here, data-driven methods for classification are especially useful as they reduce the effort for manual configuration and provide high predictive performance for complex data.
In this talk, we first motivate the complexity of realizing AI-based smartphone localization and then give a holistic overview of our iterative data collection, training and analysis process. We show interesting findings on how to meet diverse user requirements as well as how to transparently identify system boundaries. Finally, we emphasize fruitful research directions and possibilities to participate.

12:20 – 12:40

Application Examples of Using AI for Automotive Software Development

Christian Miedl, Segmentlead Safety, Security and AI at AVL Software and Functions GmbH

Artificial intelligence (AI) is not only a buzzword in the automotive industry, but also a key enabler for many advanced features and functions of self-driving cars. In this talk, we will explore how AI can be applied in different stages and aspects of automotive software development based on an example from control software development for safety monitoring functions. We will also discuss how the role of software developers will change as AI becomes more integrated and sophisticated in the development process.
AI for Control Software: Another challenge of cars is to ensure the safety and reliability of the control software that monitors and regulates the vehicle’s functions. AI can help to improve the quality and efficiency of control software development, by using pattern recognition to detect and diagnose malfunctions, and to provide safety support and feedback.
AI for Software Generation: A third challenge of self-driving cars is to cope with the increasing complexity and diversity of software requirements and specifications. AI can help to streamline and accelerate the software development process, by using large language models to generate individual development artifacts, such as requirements acceptance criteria, documentation, and test cases. AI can also help to enhance the creativity and productivity of software developers, by providing suggestions, recommendations, and corrections. However, AI cannot replace the human expertise and intuition of software developers, who still need to control and supervise the AI-generated software.

12:40 – 13:00

Self-Calibrating Perception Sensors for Autonomous Vehicles: The AI Approach

Arya Rachman, Expert Software & Functions Engineer at AVL Software and Functions GmbH

Prior to driving, sensors in an autonomous driving system need to be calibrated.
Calibration is crucial to ensure that safety-related perception functions can reliably perceive the environment.
Vehicle sensors are also exposed to mechanical perturbations requiring periodic re-calibration for regular use.
The current widely accepted calibration approaches are based on robust but potentially demanding target-based methods. Such methods require a car to be taken offline and rely on static infrastructure and operators.
The AI approach: we propose a deep-learning-based self-calibration strategy for the vehicular sensors that learns from driving scenarios – they make an inherently large-scale dataset – and is validated back-to-back against checkerboard reprojection for proven reliability

13:00 – 13:50

Poster Session & Networking

Session 3 ‒ Spotlight on Precision: AI-Driven Diagnosis and Quality Enhancement

Session Chair: Dr. Daniel Grünbaum, Team Lead Data Science at ams-OSRAM international GmbH

13:50 – 14:10

Unsupervised Anomaly Detection in Continuous Integration Pipelines

Daniel Gerber, Data Scientist at BSH Hausgeräte GmbH

Modern embedded systems comprise more and more software. This yields novel challenges in development and quality assurance. Complex software interactions may lead to serious performance issues that can have a crucial economic impact if they are not resolved during development. Henceforth, we decided to develop and evaluate a machine learning-based approach to identify performance issues. Our experiments using real-world data show the applicability of our methodology and outline the value of an integration into modern software processes such as continuous integration.

14:10 – 14:30

A New Approach for AI-Based DGA for Transformers and Tap Changers

Andreas Dreger, Head of Diagnostic Systems at the Competence Center for Measuring and Diagnostics at MESSKO GmbH / Maschinenfabrik Reinhausen GmbH

In the field of power transformer maintenance, Dissolved Gas Analysis (DGA) is a widely used technique to assess the health of transformers. It works by examining the concentrations of gases dissolved in the insulating fluid. A new DGA interpretation method is proposed to address the limitations of previous methods, which do not fully account for uncertainties and application data. The proposed method corrects DGA values for non-failure-related influences using application data and equipment-specific parameters. A machine learning approach is then employed to predict the probabilities of different failure states from the corrected DGA values, reflecting the inherent uncertainties in the overall DGA process. The method also explicitly handles missing data and considers the resulting increase in uncertainty. Both the failure state probabilities and uncertainty information are used to automatically generate a recommended course of action. The method’s response to clear-cut and ambiguous situations is demonstrated using a real case example.

14:30 – 14:50

Enhancing Manufacturing Precision through Automated Independent Diagnosis: An On-Demand System for Production Deviation Protocol (PDP), Classification and Error Analysis

Maximilian Deichsel, Business Analayst Artificial Intelligence at Infineon Technologies AG

The AID system is a cutting-edge tool designed to function autonomously and on-demand in the manufacturing environment. Triggered by the creation of a Production Deviation Protocol (PDP), which signals an error or anomaly on the shop floor, the AID system initiates a specialized workflow. PDPs are generated either manually by operators or automatically by shop floor control systems such as Statistical Process Control (SPC) or Advanced Process Control (APC). Upon activation, the AID system evaluates the situation and provides a probability vector as an outcome that categorizes the deviation, facilitating easy integration and documentation within the PDP. This result vector is crucial for the initial decision-making process, termed First Decision (FiDe), which relies heavily on the AID’s recommendations. To maintain efficiency and accuracy, the AID system offers a maximum of five classification suggestions per PDP, based on the existing Deviation Decision Help (DDH) and a predefined probability threshold. In addition to the classification suggestions, the AID tool is capable of evaluating the probability of a new defect, based on previously unknown patterns or anomalies. Operating specifically for lot deviations, the AID not only classifies the most probable defect but also enhances the monitoring of defect classifications, aiding in system oversight and root cause analysis, particularly important for new, previously unknown defect characteristics.

14:50 – 15:10

Autoencoder for X-Ray Anomaly Detection

Dr. Johannes Oberpriller, Data Scientist at ams-OSRAM International GmbH

Autoencoders are neural networks that can learn to reconstruct their inputs with minimal loss of information. They have been widely used for anomaly detection in various domains, such as fraud detection or health monitoring, where abnormalities need to be identified. In this talk/poster, we show the usefulness of an autoencoder for X-ray diffraction anomaly detection for the epitaxy in semiconductor production. Our autoencoder is trained on samples of X-ray diffraction patterns from epitaxial layers and can detect anomalies by measuring the reconstruction error or the latent representation. We evaluate our method on real datasets at the ams OSRAM group. We discuss its applicability, robustness and challenges for different epitaxy structures and demonstrate its effectiveness in automated production setting.

15:10 – 16:00

Poster Session & Networking

Session 4 ‒ Beyond Words: Success Stories of Large Language Models

Session Chair: Prof. Dr. Timo Baumann, Professor for Natural Language Processing at OTH Regensburg

16:00 – 16:20

LLM-Powered Spoken Dialogue Systems – The Future of Voice Interaction for Home Appliances

Dr. Manuel Kirschner, Speech Technology Specialist at BSH Hausgeräte GmbH

BSH Hausgeräte GmbH is the largest manufacturer of home appliances in Europe and one of the leading companies in the sector worldwide. In this talk, after a brief introduction of our company and our location in Regensburg, we will provide an overview of the current status of Voice Interaction (and Spoken Dialogue Systems) for our global home appliance portfolio. We will present an overview of the different technical solutions (on-the-edge, cloud-based, and hybrid) available today and give you a sneak preview of prototypes that we are currently working on, including a first system that harnesses Large Language Models to demonstrate our vision of the next step towards an enhanced User Experience.

16:20 – 16:40

HORSCH Translation Platform

Dr. Maximilian Braun, Data Scientist at HORSCH Maschinen GmbH

Language: German

Faced with the challenge of decentralized translation processes reliant on external providers who produced inadequate quality, HORSCH embarked on a project aimed at enhancing translation quality and reducing costs through the application of AI technology.
Initially, a specialized dictionary was created by compiling existing translations from multiple sources. Utilizing a transformer-based architecture, bespoke models for specific language pairs were refined to achieve optimal instantaneous translation accuracy.
The unveiling of the initial use case set the momentum for the development of a comprehensive enterprise architecture to manage all of HORSCH’s translation needs.
Fast forward: In the six months following its introduction, the platform has processed an impressive total of approximately 110,000 translations, with an automation rate exceeding 75%. Continuous rapid development of new features for the overall system and improvements to the neural networks ensures further increases in automation.

16:40 – 17:00

Using ChatGPT to Support Intelligence Analysis in Strategic Forecasting

Omololu Adu, Julie Hoffmann, Samuele Minelli and Elena Potitò, Research Trainees at OTH Regensburg

ChatGPT analyses information and generates content in seconds. This poster presents research on how generative AI tools such as ChatGPT can be used to support intelligence analysis, focusing on the Analysis of Competing Hypothesis (ACH) as one of the most commonly used methods in scenario development and forecasting. The research presented in the poster examines two aspects of ACH more in depth: (1) the plausibility and consistency of generated hypotheses and (2) the soundness and appropriateness of generated indicators to measure and observe the hypotheses created in (1). To do so, the research draws on the case study of attempting to predict internally displaced persons as a leading cause of Boko Haram violence. It concludes with a set of recommendations for using ChatGPT as a supportive tool in intelligence analysis and outlines fields of further research.


17:00 – 17:05

Closing Words

Prof. Dr. Wolfgang Mauerer, Chairman Director of the Regensburg Center for Artificial Intelligence (RCAI) at OTH Regensburg


Production Optimisation

Applying Non-Negative Matrix Factorization for Efficient Crystal Phase Identification in White LEDs

Dr. Sebastian Imhof, ams-OSRAM International GmbH

Revolutionizing Optical Inspection in Manufacturing: A Global Deployment of Serverless AI Technology

Johannes Brunner and Magdalena Listl, Continental Automotive Technologies GmbH

AI-Enabled Precision: Defect Detection in Semiconductor Scanning Acoustic Microscope Images

Saad Al-Baddai, Infineon Technologies AG

Uncharted Bonds: Pioneering Scalable Virtual Metrology in AI-Driven Wirebonding

Dr. Jan Papadoudis, Infineon Technologies AG

Leveraging Convolutional Neural Networks to Streamline Root Cause Analysis

Singla Sahil, Infineon Technologies AG

Fast Defect Density Wafermap Classification

Dr. Wu Shu-Chun, Infineon Technologies AG

Optimizing Manufacturing Throughput via Digital Twin Technologies

Alexander Swinarew, Infineon Technologies AG

Predictive Maintenance im Sondermaschinenbau

Amritha Kalluvettukuzhiyil¹, Robert Halladay², Jonathan Epp², Markus Franke¹, Markus Schwab², Michael Kronfeld², Sebastian Bock² und Prof. Dr. Martin Weiß¹, ¹OTH Regensburg und ²Baumann Automation

Transient Surrogate Modeling of Modally Reduced Structures

Markus Franke and Prof. Dr. Marcus Wagner, OTH Regensburg

Transfer Methodology of Data-Independent Knowledge Enhancing Yield Prediction Accuracy on Ordinal Delivery Grid

Stefan M. Stroka, ams-Osram International GmbH and Ludwig-Maximilians-Universität

Smart Cities and Environments for Sustainability

AI-Enhanced Façade Planning

Simon K. Höng and Prof. Dr. Mathias Obergrießer, OTH Regensburg

A Slim Digital Twin for a Smart City and its Residents

Simon Thelen, Friedrich Eder, Matthias Melzer, Danilo Weber Nunes, Michael Stadler, Prof. Dr. Christian Rechenauer, Prof. Dr. Mathias Obergrießer, Prof. Dr. Ruben Jubeh, Prof. Dr. Klaus Volbert and Prof. Dr. Jan Dünnweber, OTH Regensburg

Determinants of Electricity Expenditure by Private Households. Analysis of the 2018 Survey of Income and Expenditure for Germany

Dr. Caroline Dotter, Miriam Vetter, Prof. Dr. Sonja Haug and Prof. Dr. Karsten Weber, OTH Regensburg

Acceptance and Concerns About Smart Meters and Web-Based Apps: Results of the Project EVEKT

Miriam Vetter, Dr. Caroline Dotter, Prof. Dr. Sonja Haug and Prof. Dr. Karsten Weber, OTH Regensburg

Potential of AI-Based Analysis of Hyperspectral Images for Plant Health Monitoring

Lukas Bauer, Thomas Vitzthumecker, Prof. Dr. Matthias Ehrnsperger and Prof. Dr. Rudolf Bierl, OTH Regensburg

Healthcare and Assistance

Hardware-Aware Neural Architecture Search for Optimizing EEG Seizure Detection in Closed-Loop Neurostimulator

Jonathan Larochelle, Prof. Dr. Peter Woias, Albert-Ludwigs-Universität Freiburg and Prof. Dr. Laura Maria Comella, Hochschule Karlsruhe

Data Augmentation for Images of Chronic Foot Wounds

Max Gutbrod, Benedikt Geisler, David Rauber and Prof. Dr. Christoph Palm, OTH Regensburg

Smoke Classification in Laparoscopic Cholecystectomy Videos Incorporating Spatio-Temporal Information

Tobias Rückert¹, Maximilian Rieder¹, Hubertus Feußner², Dirk Wilhelm², Daniel Rueckert²·³ and Prof. Dr. Christoph Palm¹, ¹OTH Regensburg, ²Klinikum rechts der Isar, Technical University of Munich and ³Imperial College London

Climbing Hold Detection with AI for a Selective Assistance System for Visually Impaired Climbers

Viola Schneider, Matthias Schlauderer and Prof. Dr. Armin Sehr, OTH Regensburg


Introducing a Machine Learning based Voting Technique for Comparison of Learning Path Recommendations

Flemming Bugert and Prof. Dr. Jürgen Mottok, OTH Regensburg

More to be announced…

Beitrag veröffentlicht