Firstly, different types of faults and principles of fault propagation are identified and existing literature is reviewed. Relevant key figures, classes and patterns are worked out or, if necessary, specified.

By reviewing existing technologies, possibilities for fault modelling and fault propagation are worked out. The driving function is then specified and executed under the influence of selected faults and fault types. The open-source simulator CARLA is used for this purpose. A propagation of the fault in the simulator is then identified and further analysed. This forms the knowledge base to provide an estimate of the fault propagation for further simulation runs at runtime.

The final evaluation provides information about the possibilities for detecting a fault and estimating its propagation.

Firstly, principles and relevant aspects of a digital twin are identified and existing literature is reviewed. Possible types, their functionalities and specialised applications and the type of use are specified and critically reflected with regard to the nomenclature. By reviewing existing technologies, possibilities for modelling digital twins are identified.

In a second step, a scheme for the representation of a modular and expandable set of a federated digital twin is conceptualised. The focus here is on the normalisation of digital twins in the federation. At a high level of abstraction, this can support, for example, the structured integration
of a digital twin into the federation as well as formalised methods for its removal. The standardisation of units of measurement or the use of a common ontology are further basic building blocks of normalisation.

The implementation of a simple data structure and the associated management services demonstrate the feasibility of the concept. The final evaluation provides statements on the use of a normalised federated digital twin and discusses normalisation in terms of the level of quality gained.

This master thesis addresses a key challenge of AI functions for autonomous driving system (ADS). Robust deep neural networks (DNN), for instance for pedestrian detection, require redundancy to avoid wrong decisions and enable a safety argumentation. This is especially important for safety critical tasks such as pedestrian detection in our case. The thesis is part of ongoing research for the application of DNNs in ADS.

Regarding the perception of an ADS, we see a multitude of AI-driven implementations. The field of computer vision experienced a shift from traditional approaches towards machine learning methods in the last years, applying DNNs with convolutional operations for challenges such as object detection and semantic segmentation. Due to the complexity of the perception task and the environment of the ADS itself, safety argumentations heavily rely on multiple redundant functions. Thereby, redundancy does not have to be achieved through multi-modal approaches alone. A promising field of research is in the fusion of methods that uses the same type of sensor to improve robustness.

The goal of the offered master thesis is the fusion of methods for 2d object detection and semantic segmentation. A starting point is given by the work regarding the late fusion of a SSD, focussing on the binary pedestrian detection [1]. For the start, we provide code for 2d object detection and semantic segmentation. We exclusively use PyTorch as our deep learning framework.

Throughout the work, new methods should be investigated and evaluated with State-of-the-Art pedestrian datasets. The new fusion methods are not only expected to improve performance, but also to enhance robustness as needed for critical safety applications.

Start is possible immediately. If you are interested, please feel free to contact patrick.feifel@uni-oldenburg.de.

[1] Du et al, Fused DNN: A deep neural network fusion approach to fast and robust pedestrian detection, IEEE winter conference on applications of computer vision (WACV), 2017.

Nowadays, software is increasingly being used in sensitive or security-critical areas. This master's thesis addresses the AI protection of software. Specifically, a multi-agent simulation is implemented using the Python framework Mesa in order to evaluate various aspects of AI security. To this end, game theory approaches are applied to parcel delivery simulations.

In the designed game, various last-mile logistics approaches are pursued by the players, for example the use of electric cargo bikes from the Sustainable Crowd Logistics project. Players receive income for delivering parcels, which is to be secured over the entire course of the game using AI. In addition, the detection of fraudulent external interference in the artificial generation of customer orders is addressed. Based on this, algorithms can be used to compare and analyse different gaming options. For example, a neural network is used and evaluated to improve guaranteed delivery times.

The demonstrated research into the use of AI safeguarding enabled the profitable use of a neural network to increase the number of parcels actually delivered on time.

Due to the constant development of new software that increases the driving range, safety and comfort of a vehicle, there is an increasing need to organise the distribution of this software appropriately. In my work, a concept is created that enables software collections customised to the vehicle to be installed. This saves resources, for example in the form of storage capacity.

Firstly, an overview of the market for recognising and providing new scopes of driving is created using a business model canvas. Based on the knowledge gained, the most important building blocks of the value chain are identified, explained and linked together. In addition, the possible sales process of a software is modelled.

A concept for the classification of software is presented according to the type of different access rights to the vehicle's systems. A communication protocol is developed for the provision of software, which enables secure interaction between server, vehicle and service providers.

By implementing the prototype, the created sales and value creation process as well as the technical concepts are exemplified. A vehicle is simulated that can install and use software from a software store. The store is operated via an Android interface and the selected use case of automated parking is visualised using a CARLA simulation.

There are many different sensors, all of which are described differently. This not only makes it difficult to find them, but also to compare their properties. In the bachelor thesis, a document schema for standardisation is developed, which can describe sensors in the field of autonomous driving. In addition, a draft program is being developed that will make it possible to search for sensor properties in these standardised documents. Specifically, the document is used as a basic document that can be enriched with further information, such as test results, and also supports the saving of a sensor model. The standard is defined in XML and, in addition to the technical data of the sensor, should also contain other relevant information such as the manufacturer or details of existing models and test data. In order to develop the standard, various data sheets of different sensor types will be analysed and compared. In the course of the work, a programme is also designed that enables a search for the specified properties to be carried out on the standardised documents. The following procedure is proposed.

Standardised XML documents for the corresponding sensors are stored on several servers. The programme has a list of the servers and the storage locations of these XML documents. If the user of the programme carries out a search for specific sensor properties, the programme updates its local data and outputs the result documents. For example, when searching for a sensor with a range of 100 to 120 metres, the corresponding result documents are displayed.

The Step-UP!CPS project looks at concepts for updates to improve the functions of cyber-physical systems (CPS). The task of the bachelor thesis is to create a simulation of an intelligent vehicle that drives automatically and whose driving performance can be recorded, stored and analysed using a monitor. The simulation is to be created in VIRES Virtual Test Drive (VTD). The simulated vehicle is to drive through a junction and, for example, occasionally come too close to the sides of the carriageway - both to the outer edge and to oncoming traffic. It is also conceivable that the new software version could be improved with regard to other criteria, such as long waiting periods. Such situations are to be improved by updating the vehicle's software so that the vehicle drives through problematic areas less frequently or waiting periods can be reduced. The two different software versions are to be analysed using the monitor so that a statement can be made as to whether the update has actually brought about an improvement. The monitor should be able to make such a statement automatically after an update. Specific evaluation metrics were identified in the research seminar. Among other things, Time to Collision (TTC) and Distance to Line Crossing (DLC) are to be used. The analysed driving performance is then to be classified using these metrics. A categorisation into very good, good and bad journeys would be conceivable. Finally, these classified driving performances should be further analysed with the help of statistics so that a comparison can be drawn up.

At this point we can also offer the following external work.

As part of our cooperation with the German Aerospace Centre, we would like to draw your attention to the following job advertisement:

Investigation and prediction of fault propagation