Industrial Projects


Project: Sensor delay effects in High-Resolution FluoRespirometry

Company: Oroboros

380px Logo OROBOROS INSTRUMENTSOroboros Instruments GmbH is a leading laboratory equipment manufacturer, continuously working to improve its products. The company distributes the gold standard O2k-technology for High-Resolution FluoRespirometry (HRFR) world-wide. The devices now achieve high time resolution, which has been unreachable before. At the same time, sensor delay effects become relevant for many applications. Advanced mathematical models and methods are therefore essential to account for these delay effects in the measurement signals.

The aim of this project is to develop models and improved algorithms accounting for delay effects in the analyzers of Oroboros Instruments GmbH. In particular, the aim is to establish accurate mathematical models representing delay effects in the HRFR for mitochondria and cell research. Tools and from signal processing algorithms are established and implemented for model-based deburring of measurements signals. The developed tools should be accurate and stable, and example implementation will be established and implemented.

MitoEAGLE: The objective of the MitoEAGLE network is to improve our knowledge on mitochondrial function in health and disease related to Evolution, Age, Gender, Lifestyle and Environment.

  • Academic experts:
    Markus Haltmeier (This email address is being protected from spambots. You need JavaScript enabled to view it.)
    Peter Burgholzer (This email address is being protected from spambots. You need JavaScript enabled to view it.)
  • Industry representative:
    Markus Haider (This email address is being protected from spambots. You need JavaScript enabled to view it.)
    Erich Gnaiger (This email address is being protected from spambots. You need JavaScript enabled to view it.)

 

Project: Classification of Network Traffic

Company: Barracuda

barracuda logoBarracuda Networks offers cloud-enabled solutions that empower customers to protect their networks by providing end-to-end network and data security. Our internal firewall handles approximately 40 million connections from about 30 clients during a typical working week. Our largest firewall models are designed to serve up to 15,000 clients, with a theoretical maximum of several billion connections per week.

Common network intrusion detection systems are based on pattern recognition or on pre-defined rules and heuristics to identify abnormal behavior within this high amount of data. Applying machine-learning techniques should improve this process to reduce the need of manual interactions and provide almost real-time classification of the throughput traffic. In detail, this task is divided into two stages as follows.

The goal of this project is twofold:

1. Most network traffic exhibits certain patterns due to repeated user behavior or automated tasks, ranging from higher network capacity during business hours to the occurrence of certain parameter values such as the ratio of internal to external destination addresses.
During the first stage, those patterns should be identified to be able to separate suspicious connections from regular traffic during the training phase. For this purpose, a set of descriptors should be derived to be used as feature vectors for later classification by either a clustering method, a support vector machine, or similar algorithms. Barracuda Networks provides pre-collected firewall logs, containing properties like source and destination network address, protocol, data volume etc.

2. The second stage is to build an (almost) real-time classification system e.g. based on a neural network like RNN, CNN or others. Therefore, a model has to be created, which should then be trained by using the prior created feature set. The goal is to determine the system’s behavior during the concluding testing phase.

  • Academic experts:
    Felix Krahmer (This email address is being protected from spambots. You need JavaScript enabled to view it.)
    Jürgen Frikel (This email address is being protected from spambots. You need JavaScript enabled to view it.)

 

Project: Optimizing a parallel picking strategy on a component wafer to maximize the machine efficiency.

Company: BE Semiconductor Industries N.V.

besiBesi develops leading edge assembly processes and equipment for lead frame, substrate and wafer level packaging applications in a wide range of end-user markets including electronics, mobile internet, cloud server, computing, automotive, industrial, LED and solar energy.
Besi offers a wide range of die attach systems based on leading edge technology. The offering includes multi-chip bonders for advanced packaging, epoxy and soft solder bonders, high precision flip chip bonders for mass production, stacked die bonders and die sorting equipment.
Target: Optimize a parallel picking strategy on a component wafer to maximize the machine efficiency.

Details: On the component wafer are rectangular dies positioned on a regular grid. The dies Di,j for 0 ≤ i < M, 0 ≤ j < N can have the state ok (shall be picked) or bad (must not be picked). The number of good dies varies between 10% and 90% with a random structure. Typically, a component wafer consists of hundreds or thousands of components M > 20, N > 20. A picking device has the capability to pick more than one component in parallel but only in a fixed rectangular grid matching the grid of the component wafer.
This device can pick k x l neighbouring components, where k ≥ 1 and l ≥ 1, but k M and l ≪ N, typically 1 x 3 or 2 x 2. All  Di,j marked good must be picked. The algorithm has to run on a standard PC (roughly i5 using a single CPU core and less than 4GB RAM) and needs to provide the first pick position(s) within 2-3 seconds. It is allowed to have a "stream" of pick positions, where every <0.5 s a new optimized position is provided. The target is to minimize the number of picking steps.

Stretch goals:
- minimize the distance between the pick positions
- interleaved picking device: only every n-th component is picked (checker board type)

  • Academic experts:
    Alexander Ostermann (This email address is being protected from spambots. You need JavaScript enabled to view it.)
    Michael Günther (This email address is being protected from spambots. You need JavaScript enabled to view it.)
  • Industry representatives:
    Felix Schwitzer (This email address is being protected from spambots. You need JavaScript enabled to view it.)
    Hubert Selhofer (This email address is being protected from spambots. You need JavaScript enabled to view it.)

 

Project: Speech intelligibility rating model

Company: Med-El

med el

In order to complement the speech rehabilitation of our patients, this project aims to create a model that rates the intelligibility of speech. The model has to be generic to accept anything from single sounds to full sentences.

The inputs to the model are as follows: Text representation, Phonetic representation (IPA),Audio recording.
The desired output is: Annotation of text and phonetic representation with quality of the individual sounds, Segmentation and alignment of the textual representation with the audio (implicitly)

As such the model is conceptually similar to ASR (automatic speech recognition), but rather than generating a text transcription the model shall detect pronunciation anomalies. We will provide pointers to existing data corpora and suggestions for the model, however, the project is open for any model definition. Preference for models of English language, but German could also be interesting (language agnosticism may be too complex for this workshop).

  • Academic expert:
    Linh Nguyen (This email address is being protected from spambots. You need JavaScript enabled to view it.
  • Industry representatives:
    Daniel Winkler (This email address is being protected from spambots. You need JavaScript enabled to view it.)
    Michael Zaremba (This email address is being protected from spambots. You need JavaScript enabled to view it.)

 

Project: Product Portofolio Optimization

Company: Fraunhofer institute for Industrial Mathematics ITWM

ITWM

The Fraunhofer ITWM is a research institute whose focus is on the transfer of mathematical methods and technology into applications. The customers in industry come from a broad variety of sectors: from the automotive, mechanical engineering and textile industries to the energy and financial branches.

Let us assume a producer of fluid pumps is thinking about his product folio consisting of several pumps in different sizes ready to serve different customer needs. Traditionally, customer specifications cover in particular delivery height and delivery volume, which be seen as the nominal process point of the pump. Of course, each pump will be able to cope with a variety of process points in the neighborhood of the nominal point. But, deviations from nominal points will come with a loss of energy efficiency or with a loss of lifetime due to mechanical stress. On the other hand, a manufacturer will not offer individually produced pumps for any nominal process point nor will it be practically possible to stay close to nominal process points.

Given a cloud of orders from the past with specified process points how should a manufacturer set up a pump portfolio balancing pump species, customer needs and technical excellence ?

  • Academic expert:
    Aviv Gibali (This email address is being protected from spambots. You need JavaScript enabled to view it.)