Die folgenden Success Stories spiegeln unser Forschungsspektrum wider, das von der Grundlagenforschung (Nr. 1) über herausragende Preise bei internationalen Wettbewerben (Nr. 2) bis hin zu anwendungsorientierten Forschungsprojekten mit hoher industrieller Relevanz (Nr. 3, 4 und 5) als Beitrag zur Smart Factory und zur Produktivitätssteigerung in der Softwareentwicklung reicht.
Habilitation in mathematics
of B. A. Moser on novel similarity and distance measures with applications in machine learning and image and signal processing (April, 2017)
This habilitation thesis is mainly the result of strategic research at SCCH over the last seven years, supplemented by a fundamental research project (Austrian Science Fund, FWF, P21496 N23). It covers research on similarity and distance measures mainly motivated by application-oriented problems in the field of pattern matching and image and signal processing. One source of motivation was the observation that standard (dis)similarity measures show oscillations of the corresponding auto-misalignment function when applied to oscillating signals, as illustrated by Figure 1 below. It is interesting to observe that for Hermann Weyl’s concept of discrepancy for probability measures, we obtain a unique and distinct minimum for signals with nonnegative intensities independently of its spectral properties.
The mathematical analysis of this observation has led to novel research directions of Weyl‘s discrepancy measure in discrete mathematics and in image and signal processing:
- Related to discrete mathematics, e.g., a lattice path enumeration approach [Mos14b] for determining analytically the distribution of the range of a simple random walk, which provides the first elementary solution to this combinatorial problem stated by Feller 1951.
- Related to image processing by extending Weyl’s discrepancy norm for vector spaces to tensor spaces representing image and volumetric data, which allows robust rigid registration and template matching of low contrast images as encountered, e.g., in the application of magnet resonance imaging (MRI) on soft tissues. As shown by (best paper award at AARP/AWR 2016) this method can be efficiently vectorized for implementation on embedded platforms.
- Related to signal processing by the discovery of the fundamental quasi-isometry relation for threshold-based sampling based on Weyl’s discrepancy and its equivalent variants such as the Alexiewicz norm.
1st and 2nd prizes of IEEE Detection and Classification competition (DCASE 2016)
This research results from a collaboration of SCCH with JKU/CP. The first author of, H. Eghbal-Zadeh, participates in SCCH’s PhD study support program.
Machine Vision and Machine Listening are two major fields in the topic of Machine Learning that try to have a better understanding of our surrounding world. Thanks to the recent advances in Deep Learning during the recent years, Machine Vision has had a considerable progress and nowadays many commercial vision products have been developed and are integrated in our daily life in applications such as autonomous cars, video surveillance systems and many more. Although Machine Vision has had many success stories, this cannot be said for the field of Machine Listening. Compared to the field of vision, Machine Listening is a much younger field with many complex challenges that have not been yet well studied and the recent studies show that the same solutions proposed for Machine Vision are not well suited to address the complexities in the field of Machine Listening.
As a step forward to the next level in the Machine Listening field, we have developed a method for Audio Scene Classification (ASC) — one of the major tasks in Machine Listening — which combines state-of-the-art techniques from machine vision and machine listening and combines the best of both worlds. Using the power of feature learning in deep Convolutional Neural Networks and the efficiency of Factor Analysis in estimating specialized hidden factors via the Probabilistic Modeling, we developed a novel method for ASC that outperforms the state-of-the-art. To demonstrate the abilities of our method, we attended the IEEE Detection and Classification of Acoustic Scenes and Events (DCASE) challenge in 2016 in the task of ASC and won both Ranks 1 and 2 among 49 submissions from many international research centers and companies. Our work has been cited so far by many companies and research centers active in machine audition and has gained increasing attention in the field.
Smart factory for products with complex quality dependencies
Data analytics, machine learning and artificial intelligence are at the core of Industry 4.0 (the smart factory). Significant aspects of a smart factory, e.g., flexibility and resource efficiency, can be improved by leveraging the power of machine learning and optimization techniques. In the COMET project moFOCS, this is exemplified for the production of power transformer cores at our company partner Siemens Österreich. To optimize the production of such a core, which consists of hundreds of individual metal sheet layers, a system was developed that combines machine learning based predictions for critical quality attributes and a fast and efficient optimization core. Base on the quality of the raw materials (transformer metal sheets) and geometric features, the machine learning model is able to predict the resulting product quality in terms of dissipation losses and noise generation. These predictions are used inside the optimization to minimize overall production costs (which includes raw material prices, scrap costs, setup time and more). The research has led to several publications that describe the methods and algorithms; the results obtained with the prototypical implementation of the system at Siemens (Weiz) show significant savings in material costs while meeting customer constraints. Power Transformer Furthermore the flexibility of the production is increased due to the possibility of the system to optimize the production of several jobs at once and to take partially processed jobs into account. Primary success factors for the project were the integration of all related data, e.g., raw material attributes (a dedicated measurement system was developed), purchase data, and quality assurance data; this enabled the successful development of reliable machine learning models that incorporate domain knowledge and a scalable parallelized optimization core.
Software quality assurance by test case generation
The project AutoTest is the largest COMET project of the Center (3 Mio € budget) and very likely one of the largest projects in Austria focusing on software test case generation and automation.
AutoTest has led to numerous awards, e.g., the ACM SIGSOFT Distinguished Paper Award at ASE 2015, best paper at Profes 2015 in Bozen, and best papers at industry oriented conferences such as Software Quality Days in Vienna 2015 and 2017. As a result of a close collaboration with A. Egyed (JKU ISSE), M. Felderer (University Innsbruck), C. Artho (AIST, Japan) and others, in total over 30 papers and posters were published in FP1.
Selected references of successful use case as a result of the current COMET project AutoTest:
- TRUMPF Maschinen Austria on guided random tests to assure stable software for their machinery and closing the gap between software and system/hardware testing as a further step toward continuous deployment in industry
- mutation testing in a difficult technical setting, showing that our approach helps to improve test suites already covering 100% of the tested software
- Omicron on testing complex UIs in 16 languages
This research project proves to support sustainable software solutions.
Software understanding based on knowledge extraction from source code
Since 2011 SCCH has been researching software analysis methods based on static code analysis to gain knowledge about the software under consideration. This knowledge is used to generate documentation and to provide different views on software (e.g., mathematical view with formulas; behavioral view regarding paths and their conditions; structural view with metrics and relations between code structures).
This research has led to the framework eKnows (extracting knowledge from software), which significantly eases the implementation of new analysis support.
Selected references of successful use case as result of the current COMET project Next:
- Siemens Österreich (Weiz) with a legacy code base for transformer design in C++ and Fortran in order to extract technical documentation largely automatically. This ensures synchronization between program code and technical documentation and thus their consistency. Complete program documentation can be generated automatically. Achieved degree of automation: 70% of printable documentation is generated without additional changes; the rest can be covered by enhancements in the program code.
- voestalpine Stahl with software for complex computational models used for optimizing their production processes. With eKnows we analyzed one such software application that was hardly understood by the developers and experts; this prohibited the reengineering of the software or even adding new features. The analysis result was an interactive documentation that showed the behavior of the software depending on the parameters used.
- TRUMPF Maschinen Austria on software quality assurance and software understanding with a focus on detecting coding guideline violations, bad code smells, multi-tasking problems and architecture problems. Software understanding is improved by documentation support that extracts a state chart from the source code.