Perspectives in Stochastic Optimization and Applications 

Abstract

In powder-bed additive manufacturing, a laser beam melts a thin layer of metallic powder, which solidifies during cooling and bonds to the already processed solid material. After finishing one layer, the surface is covered with a new thin powder layer, and the melting process iterates, until an entire three-dimensional object emerges. Due to the uneven heat distribution created by the laser beam, the part is usually under stress during the production which can lead to significant warping after cooling. Here, we study whether the order in which each layer is processed can be used to reduce warping in the finished product. As warping is difficult to model and predict, we use several approaches to approximate this effect, such as computing heat gradients and expansion through heat. These approximations are used to generate different objective functions for the mixed-integer linear program.

• In a first step, we developed a numerical simulation of the SLM printing process. This model was supplemented with machine-learning techniques to evaluate the geometries of the printed shapes, as well as the properties of the cooling process.
• Building on these results we devised automated methods for optimizing the printing orders using mixed-integer programming.
• Furthermore, we designed new algorithms which can be used for automated quality control in an on-line setting.
• These results were tested in practical experiments by our project partners from engineering using state-of-the-art printing machines at the BTU Cottbus - Senftenberg.

https://www.b-tu.de/en/fg-ingenieurmathematik-optimierung/forschung/projects/maledif

01/2020 - 12/2021, 24 months

SENSOR-AKTOR-KI

Abstract

The aim of the project is to evaluate the use of customized sensor-related or actuator-related signal preprocessing using AI-based methods (focus: TinyAI/Embedded AI) using the example of IPMS sensors and actuators in combination with the IPMS RISC-V computing platform EMSA5. Existing IPMS sensors and actuators are to be analyzed and categorized with respect to local preprocessing. Likewise, suitable AI methods are to be defined, practically implemented and analyzed. Furthermore, an AI library for the RISC-V platform will be ported. Finally, the complete AI process chain for embedded systems can be implemented for selected sensors/actuators on RISC-V basis. Demonstrator setups will exemplify AI-based signal processing with real-time data from sensors and actuators.

• Approach validated for embedded AI projects with IPMS sensors
  
  Use cases: predictive maintenance and composition of liquids investigated
• Tensorflow Lite library ported to EMSA5 RISC-V processor
• Implementation of demonstrator hardware and software

01/2021-12/2021, 12 months

New methods of optical early detection and pest-free control of insects harmful to stocks using laser beams (insect laser)

Abstract

In this joint project, the optical early detection of insects that are harmful to stocks and their pest-free control with laser beams was investigated. To detect the pests, a camera image scanning a grain surface is evaluated by an artificial neural network. The coordinates of the detected pests are then transferred to a laser control unit, which moves a laser to the coordinates and kills the insects using a laser beam. For detection, a YOLOv4-Tiny network was trained to traverse the 2048 x 2048 pix camera image sequentially, at 512 x 512 pix. The subframes are evaluated at 25 fps, achieving a detection accuracy of 93%.  

• We have successfully developed an AI system to detect the grain weevil on a grain surface.
• Furthermore, a low-cost laser was determined to kill the insects without permanently affecting the stored material.
• Detected insects can be saved as sample images with coordinates, time and object class to build a database for further training data for the artificial neural network.

09/2017-8/2020, 36 months

AI Summarizer

Abstract

The AI summarizer is useful for summarizing longer texts to get a quick overview. Be it long legal texts, documentation or manuals, the possible applications are diverse.
The AI ​​method used comes from unsupervised learning and is called "Unsupervised Multilingual Sentence Boundary Detection". This AI method is part of the so-called "Natural Language Processing (NLP)", i.e. the processing of natural language by a computer.

• Public usable demonstrator for the application of the AI method NLP

https://www.kompetenzzentrum-cottbus.digital/Textzusammenfassungsdemonstrator

05/2021 - 08/2021

Quantum Information Access and Retrieval Theory

Abstract

QUARTZ (Quantum Information Access and Retrieval Theory) is an Innovative Training Network (ITN) that aims to educate its Early Stage Researchers to adopt a novel theoretically and empirically motivated approach to Information Access and Retrieval (IAR) based on the quantum mechanical framework that gives up the notions of unimodal features and classical ranking models disconnected from context.
The current state of the art of IAR is not sufficient to address the challenges of a dynamic, adaptive and context-aware user-machine interaction and to make a major breakthrough in the overall effectiveness of retrieval systems, and that a genuine theoretical breakthrough is on the contrary necessary.

• see www.quartz-itn.eu/dissemination-outreach

http://www.quartz-itn.eu/

01/2017-12/2021