Project: Audio signal processing of MED-EL's cochlear implant
By electrically stimulating nerve fibres in the cochlea, cochlear implants allow recipients to perceive sound. The MAESTRO Cochlear Implant System is an effective, high performance solution for individuals with severe to profound sensorineural hearing loss. Most users can enjoy music or successfully participate in conversation, even in the most challenging listening situations.
The external sound processortransfers the microphone input signalsinto stimulation patterns, which are then transmitted to the implant. The signal processing in the sound processor consists of a front-end part and a coding strategy. The primary goal of the front-end processing is to automatically identify, extract and enhance important features of the microphone signals, in particular speech features. Examples for that are beamforming algorithms, wind-noise reduction, or signal compression.
The aim of the project is to investigate further speech improvementmethods for the front-end signal processing part.
Barracuda Networks offers network security and storage solutions. One of our internal, medium sized 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 8000 clients, with a theoretical maximum of about 400 billion connections per week. With this vast amount of network traffic it is difficult to identify suspicious or dangerous network traffic, like systematic network intrusion attempts or immanent hacker attacks, which intentionally try to go unnoticed.
Most of the network traffic follows certain patterns. For example: Many periodic tasks – like network backups or server synchronizations – always occur at approximately the same time of day, or only on certain weekdays, usually have similar data volume every time, and have internal network addresses as source and destination. As another example typical traffic from internet surfing occurs mostly between Monday and Friday during office hours, and usually has internal source and external destination address.
The goal of this project is twofold:
1. In a first stage an offline classification of collected network traffic should be performed. For each handled connection the firewall logs certain properties, like start and end time, source and destination network address, protocol, data volume and some more. From these collected data a set of descriptors should be derived, that can be used as feature vectors for classification by either a clustering method, a support vector machine, or similar algorithms. The result of the classification should reflect the most prominent classes of traffic. Further it should identify connections which do not fit well into any of the prominent classes, since these connections might be the suspicious or dangerous ones.
2. Using the results from the first stage, an online classification system (for example an artificial neuronal network) shall be trained such that it can then perform real-time (or almost real-time) classification of newly created network connections.
The evaluation of medical treatments relies increasingly on Patient-reported Outcomes (PROs), i.e. patients' self-reports on somatic and psychosocial symptoms and problems. Advanced assessment methods for PROs (e.g. physical function, depression, pain, fatigue) are based on Item Response Theory (IRT) measurement models allowing computer-adaptive assessments. Traditional, unidimensional IRT models describe the probabilistic relationship between a latent construct (the PRO domain) and patients' responses to questions measuring that construct. Model parameters are estimated using Maximum Likelihood methods.
Within the Modelling Week we would like to explore the use of multidimensional IRT models and implement them into a software package for PRO data capture. Multidimensional IRT models relate a patient's response to a single question to various PRO domains. This allows to obtain estimates for a patient's symptom level on different domains at the same time and thus enables more efficient and more precise PRO assessments.
Optimization of micro heat networks In 2014 the first Tyrolean micro heat network was commissioned in the municipality of Erl. It is based on two groundwater wells which supply 13 houses with groundwater. To use the geothermal energy of groundwater, in each house a heat pump is installed, which transfers the temperature of the groundwater to a higher level. Even though such micro heat networks are a sustainable and economic method of energy production, there is still some potential of improvement. For that purpose several constellations of micro heat networks will be modelled and analyzed concerning their energy saving potential – based on input data and experiences from the micro heat network in Erl.
ZAMG as the National Weather and Geophysical Service of Austria provides weather forecasts and warnings, conducts research and operates a meteorological and a seismic monitoring network (~280 stations). The ZAMG customer service center in Innsbruck is specialized in mountain meteorology and among a bunch of other tasks conducts expertises e.g. for snow-load and wind-load. This requires the estimation of a snow pack and wind speed occurring only once within 50 years. This is straightforward for observation sites, where measurements are available for a sufficiently long period of time. When observations are missing, e.g. between stations, appropriate spatial modelling of the underlying extreme value distribution leads to satisfying results, including the possibility to specify uncertainties. Such spatial extreme value models work well for spatially smoothly distributed parameters like snow depth. Applying such a model to wind speed, which is spatially very inhomogeneous, requires additional covariates like wind direction. To use the direction of extreme wind speeds as a covariate, it has to be known at arbitrary points. So, the development of a spatial model for the direction of extreme wind speeds is a goal of this project. As wind direction is largely determined by topographical features (channel, mountain top, plain,...). A high resolution digital elevation model can be used to relate wind direction and landform. Along with measurements of wind speed and direction from ~280 Austrian weather stations, it should be possible to develop a statistical model for (the frequency distribution of) the direction of extreme wind speeds over the complex terrain of Austria.
Contact persons University of Innsbruck: Hell, Tobias
Project: Prediction of short term costumer behaviour
Predicting whether a customer is going to purchase products or services of an enterprise is an increasingly important problem for efficient product management. Nowadays, enterprises can collect various information about customer behaviour, such as frequency of purchasing, times of purchases, amounts of purchases, and so on. The goal of this project is to develop automatic tools of using such data for predicting whether a given customer is going to make a purchase or not at the enterprise in the next few months.
