Knowledge Transfer Group “Artificial Intelligence”

The speaker of July’s meeting was Noah Klarmann, from TH Rosenheim, presenting “AI in Industry: Opportunities, Risks and Challenges". 

The talk was about "AI for Industrial Applications," which seeks to enhance your understanding of AI and its potential in the industrial sector and beyond. The discussion commenced with a comprehensive introduction to AI, including the specific challenges Small and Medium Enterprises (SMEs) encounter when implementing and adopting AI models. A key segment of the talk delved into the different paradigms of AI, with a particular focus on Machine Learning. The session concluded with an exploration of potential risks and challenges that companies might face in adopting - or not adopting - AI methodologies.

Noah Klarmann is a Research Professor of Artificial Intelligence at Technische Hochschule Rosenheim since 2021. His teaching and research areas include topics such as the use of Artificial Intelligence for technical systems and digital ethics.

Physical inspection of microelectronics can turn data into value for owners of intellectual property. In addition to competitor benchmarking, the inspection process enables increased supply chain security through physical hardware verification. This talk elaborates three use cases how advanced computer vision methods are utilized for physical inspection of semiconductor devices. First, an application for automated detection and classification of competitor ESD protection diodes in PCB images is presented. Second, an approach for semantic segmentation of microscopic cross-sectional images of integrated circuits is proposed. Finally, we dive deep into the speaker's actual research topic: The physical layout recovery of highly integrated circuits. The computer vision methods used, from thresholding to vision transformers, will be introduced and elaborated in terms of their (dis)advantages. In addition, the main challenges and the importance of Deep Learning in hardware inspection will be discussed.

About the speaker

Ann-Christin Bette is working as a Research Scientist for Advanced Vision Methodologies in IC Analysis at Infineon Technologies AG in Munich since January 2020, investigating methods to ensure the applicability of the IC layout recovery process in advanced technologies. Currently, she is pursuing an external Ph.D. Program at the TUM School of Computation, Information and Technology at the Chair for Robotics, Artificial Intelligence and Embedded Systems supervised by Prof. Dr.-Ing. Alois C. Knoll at the Technical University of Munich.

Drawing on her educational background in computer science, mathematics, and physics, coupled with research and industry experience, she is eager to contribute to the establishment of value-added applications of AI methods in the industry.

Manufacturing industries have efficient and voluminous production of finished goods. To ensure high quality control and prevent bottlenecks, a keen interest is in research and development of anomaly detection. Recent advancements in artificial intelligence for defect localization have resulted in precise and accurate predictions. However, challenge remains to have datasets of balanced feature set and also possible defects occurring in the production line. This project is about the implementation of a defect localizing autoencoder by selecting classes in an unsupervised fashion to maintain class balance. The selected classes then have an overlay of artificial defects. Initial tests are promising with good generalization and future work is to have it implemented on a prototype system.

About the speaker

Devang is from India and is doing his Master in Ingenieurwissenschaften at Technische Hochschule Rosenheim. This meeting is the defense of his master thesis.

In addition to short delivery times, many customers demand the option of adapting the products to their needs. To enable the necessary, flexible production processes, the industry is increasingly relying on the use of driverless transport systems (DTS). In the project, a fleet management system is to be developed that will enable SMEs to independently integrate their special requirements for the first time.

About the speaker

The speaker is Deepak Vivekanaandan, from TH Rosenheim, an enthusiastic engineer with Master and Bachelor engineering degree in Mechatronics and Automation.

FAST AI Movies offers an AI-based SaaS solution that quickly generates interactive training videos from text input for medium-sized businesses, e.g. for product communication. The founder and CEO of Fast AI Movies goes through the technical side of the generative transformer models, used in the start-up, and brings the discussion about ChatGPT as well, showing the weakness of the technology along with its impact in society.

About the speaker

Florian Scherl is the founder and CEO of Fast AI Movies, and have won numerous startup competitions, such as the Startup Teens Award.

Randomization is widely used approach in Sim2Real transfer for data-driven learning algorithms in robotics. Still, it is difficult to evaluate different randomization approaches systematically. A professor from Technische Universität München explained the study developed to address this problem where it was defined an easy-to-reproduce experimental setup for a robotic reach-and-balance manipulator task, which can serve as a benchmark for comparison.

About the speaker

Josip Josifovski is affiliated with the Chair for Artificial Intelligence, Robotics and Real-Time Systems of TUM. He works on simulation environments and learning algorithms that help robots in performing industry-related tasks, in the scope of the Artificial Intelligence for Digitizing Industry (AI4DI) research project.

Databricks is a cloud-based platform for data engineering, data science, and machine learning and offers a range of tools and services, including data integration and machine learning capabilities. An expert in the technology showed how it is to work with Data Bricks and its possibilities.

About the speaker

Bernardo Cambruzzi is a Brazilian engineer that, at the time of the meeting, was the Lead Data Engineer of Bix Tecnlogia, in Brazil. He is a specialist in Data Bricks, having founded a company to teach on the subject.

The meeting shed light on the various challenges encountered in the domain of embedded artificial intelligence (AI) and Tiny Machine Learning (TinyML). Embedded AI and TinyML present unique challenges due to limited computational resources and energy constraints. Overcoming these obstacles requires the development of efficient algorithms and models that can effectively operate on resource-constrained devices. The meeting aimed to deepen the understanding of the obstacles faced in deploying AI on small-scale devices and fostered conversations around potential solutions and strategies to overcome these challenges effectively.

Known for its expertise in the electronics industry, DH Electronics specializes in designing and manufacturing innovative electronic components and systems. The company has gained a reputation for delivering high-quality products and solutions to various sectors, including automotive, industrial automation, and consumer electronics.

The meeting provided an insightful overview of the innovative startup, Fast AI Movies, based in Bavaria, Germany. Participants gained valuable insights into the company's mission to revolutionize the film industry through the application of artificial intelligence (AI) technology. Fast AI Movies specializes in developing AI-powered solutions that streamline various aspects of movie production, ranging from scriptwriting to post-production tasks like video editing and visual effects. By leveraging cutting-edge machine learning algorithms, the startup aims to enhance efficiency, reduce costs, and unlock new creative possibilities for filmmakers.

The meeting hosted by Technische Hochschule Rosenheim offered valuable insights into the practical implementation of data analysis techniques. The focus was on showcasing specific products utilized at Landesforsten RLP and Landesbetrieb Wald und Holz NRW, two forestry organizations. Participants gained a deeper understanding of the real-world applications of data analysis in the forestry sector, exploring how these organizations leverage data-driven insights to optimize forest management, monitor environmental impact, and make informed decisions.

The meeting with speaker from Brückner Maschinenbau focused on the crucial topic of accurately predicting film temperature during the film production process within an industrial oven. Participants explored innovative techniques and methodologies to model and forecast the temperature of the film as it undergoes heat treatment within the oven. This knowledge is vital for ensuring optimal film quality, performance, and consistency in industrial film production.

Brückner Maschinenbau is a renowned company in the field of industrial machinery, particularly in the film and packaging industry. With a rich history spanning several decades, Brückner Maschinenbau has established itself as a global leader in providing cutting-edge solutions for film production, including high-performance ovens and production lines. The company's expertise lies in designing and manufacturing advanced machinery that enables efficient and precise film processing.

The meeting hosted by Technische Hochschule Rosenheim delved into the fundamental concepts and intuitions behind reinforcement learning (RL). Participants gained insights into the key principles and algorithms that form the foundation of RL, a subfield of artificial intelligence. RL is focused on teaching agents to make optimal decisions in dynamic environments. Inspired by human and animal learning, RL involves an agent interacting with an environment to maximize rewards. Through trial and error, the agent learns which actions lead to positive outcomes. RL algorithms utilize techniques like value functions, policies, and exploration-exploitation strategies to improve decision-making. RL has practical applications in robotics, gaming, autonomous systems, and recommendation systems.

The meeting focusesed on utilizing convolutional auto-encoders, a type of deep learning model, to detect anomalies in various types of data. By leveraging the power of convolutional neural networks, these models can effectively learn representations of normal patterns and identify deviations that may indicate anomalies. 

DH Electronics, a leading company in the electronics industry, was the speaker of the meeting to foster knowledge sharing and explore innovative approaches to anomaly detection using advanced neural network architectures.

This meeting was hosted by TH Rosenheim, and the discussion was about Artificial Intelligence (AI) in production scheduling, that involves the application of advanced algorithms and techniques to optimize and automate the scheduling processes in manufacturing and production environments. By harnessing the power of AI, production scheduling systems can analyse complex data, such as order information, machine capabilities, and resource availability, to generate efficient and optimal schedules. These AIpowered systems can handle dynamic and real-time changes, consider multiple constraints, and make intelligent decisions to optimize production throughput, minimize downtime, reduce costs, and improve overall operational efficiency. Implementing AI in production scheduling enables companies to streamline their production processes, increase productivity, and achieve better resource utilization, ultimately leading to improved customer satisfaction and competitive advantage in the industry.

The meeting featured a speaker from DH Electronics and focused on the theme of using autoencoders for cluster analysis. Participants explored how autoencoders, a type of neural network, can be used for unsupervised learning tasks like clustering. Autoencoders are powerful unsupervised learning models that can capture meaningful representations of data by learning to encode and decode it. In the context of cluster analysis, autoencoders can be trained to group similar instances together, forming clusters based on underlying patterns and structures in the data. By leveraging the ability of autoencoders to reconstruct data while minimizing reconstruction errors, this approach allows for effective data exploration, identifying hidden relationships, and discovering clusters within complex datasets. Cluster Analysis with Autoencoders provides valuable insights and techniques for applying this approach to unsupervised learning tasks, enabling data scientists and researchers to gain deeper insights and extract valuable knowledge from their data.

Known for its expertise in the electronics industry, DH Electronics specializes in designing and manufacturing innovative electronic components and systems. The company has gained a reputation for delivering high-quality products and solutions to various sectors, including automotive, industrial automation, and consumer electronics.

The project ALPKI was developed at Innovationslabor, Campus Chiemgau. The focus of this project is to utilize Data Science techniques to analyze collected data from the production line and provide valuable information regarding process capability. The project used engine shafts as the parts to be studied, which were measured using inductive probes, as shown in the picture. During this meeting, the project author explained the work conducted and the main discoveries, as well as the concept of unsupervised learning. Unsupervised learning enables the ability to cluster data by searching for patterns in the database and grouping the most similar data points into clusters. This allows a human being to examine the similarities within each cluster's data points and extract insights from them.