Thesis Topics

Title: Automatic Generation of Network Configuration in Simulated Time Sensitive Networking (TSN) Applications
Subject: Computer science, Embedded systems
Level: Advanced
Description:
IEEE Time-sensitive networking (TSN) standards are the extension of the IEEE Ethernet standards to support high-bandwidth and low-latency real-time communication. TSN standards are promising solutions to be applied in various real-time domains such as automotive, industrial and etc. Designing dependable real-time networks based on TSN standards requires various analysis in terms of behavior, timing and scheduling.
Various simulation tools exist for this purpose. The aim of this project is to automate and enhance simulation tools based on Omnet++ .
The goals of this master thesis are as follows: (a) Study and review the existing simulation tools for TSN based on Omnet++; (b) Investigation and comparison of methods to perform end to end timing analysis of TSN-based ethernet switches using these simulation tools; (c) Investigation on how to automatically configure a simulated TSN network; (d) Investigation on how the automatic network configuration for simulated TSN network can improve the overall performance

Title: Architectural mitigation for systems using Deep Neural Networks (DNNs) in safety-critical applications
Subject:
Level: Advanced
Description:
This thesis addresses architectural mitigation techniques for the use of DNNs in safety-critical systems. A DNN has to be trained with data sets of images (or other data) with objects it should be able to classify, but it cannot be trained with all possible inputs. Thus, misclassification of objects may appear. DNNs are also weak to adversarial inputs (the alteration of inputs which forces a trained DNN to misclassify) e.g. due to malicious attacks or external faults caused by the environment such as single event upsets. The focus of this thesis is on redundant architectures that can detect misleading errors. Simulink is the suggested tool for implementing DNNs and associated detection architectures.
Title: The implementation of a Recurrent Neural Network Model on Xilinx’s PYNQ board
Subject: Computer network engineering, Computer science, Embedded systems, Robotics
Level: Advanced, Basic
Description:
Deep learning has emerged as an important application area for Field-Programmable Gate Arrays (FPGAs). FPGA implementations of machine learning applications can often run much faster than software implementations and can consume significantly less power than Graphics Processing Unit (GPU) implementations. However, these applications mainly focus on large scale FPGA clusters that have an extreme processing power for executing massive matrix or convolution operations but are unsuitable for portable or mobile applications. This thesis will describe research on a single-FPGA platform to explore the applications of FPGAs in these fields.
This thesis will design a Recurrent Neural Network (RNN) for ECG signal classification and implement a hardware accelerator with the AXI Stream interface on a PYNQ board. The PYNQ has a flexible embedded operation system, which makes it suitable to be applied in deep learning applications.
Title: Patient-specific method for Classification of ECG signal using Recurrent Neural Networks (RNNs)
Subject:
Level: Advanced, Basic
Description:
An electrocardiogram (ECG) measures the electric activity of the heart and has been widely used for detecting heart diseases. By analyzing the electrical signal of each heartbeat, it is possible to detect some of its abnormalities. In the last decades, several works were developed to produce automatic ECG-based heartbeat classification methods. This thesis will develop an ECG classification algorithm based on LSTM recurrent neural networks (RNNs).
In this project, a patient-specific procedure will be employed. In other words, the model is trained for every patient individually. Once the model is trained for a patient, continuous ECG monitoring and heartbeat classification is performed in real-time based on the trained model of that patient. Another approach is to train only one model by feeding data from many patients, and then, use the trained model for classification of data from other patients. We do not employ this approach because the ECG waveform varies significantly among different patients.
In the target method, the training data for a patient is formed by combining two sets of data: local ECG data and global ECG data. The first part, i.e., local data, is specific to the patient and helps increase the classification accuracy due to existing similarities among the heartbeats of every patient. The second part, i.e., global data, is the same for all patients. It consists of a number of representative heartbeats from all arrhythmia classes. It helps the model learn other arrhythmia patterns that are not included in the local data.
This thesis will train an LSTM recurrent Neural Network, then test the algorithm on ECG signal data obtained from patients. For this purpose, we aim to process the patient’s ECG signal with python, convert them using Numpy such that it can be fed to the input of LSTM.