This thesis serves to finalise the bachelor graduation project on the topic of self-supervised federated learning, specifically the on-chip implementation of the algorithms. The goal of the project is to implement a self-supervised learning setup in a decentralised approach using Field-Programmable Gate Arrays (FPGAs) for the processing of data. In this thesis, we endeavour to illustrate the possibility of employing FPGAs to move the fairly compute-intensive self-supervised learning algorithms to the edge. We have developed a number of modules that can accelerate key algorithmic blocks that underlie the major bottlenecks of the classical application of the algorithms and showcase prospective results, which are extensively discussed afterwards to pave a clear path towards truly autonomous and efficient edge-intelligence.

Distributed feeding in photo-conducting antennas (PCAs) leads to simple optical designs, less prone to overheating, as well to terahertz (THz) antennas that can operate non-dispersively over wide bandwidths. However, the efficient analysis of pulsed PCAs was so far limited to architectures characterized by feeds small with respect to the THz wavelengths. In this thesis, an efficient and rigorous procedure for the time domain analysis of an infinitely long slot antenna printed on photo conductive material and excited by a distributed pulsed laser is presented. The procedure is electromagnetically rigorous, and relies on spectral representations of the fields in both the frequency and spatial domains.

One of the main problems with Instance-Level Image Retrieval in video data is that for query videos with multiple objects of the same instance, extracting features from keyframes of this query video is time consuming. This thesis aims to solve this problem by implementing a Convolutional Neural Network based approach, which significantly reduces the extraction time and increases the accuracy. After analysing multiple methods, Second-Order Loss and Attention for image Retrieval (SOLAR) was found to be the most promising method. SOLAR will be tested based on three performance metrics: the mean average precision, the recall and the extraction time per image. The performance will be evaluated based on a selection of videos and images from a dataset provided by Dr. Andrea Nanetti from the Engineering Historical Memory project. For this dataset SOLAR achieved a mean average precision of 83 %, a recall of 85 % and an extraction time of 0.73 seconds per image. To conclude, SOLAR is specialised in detecting and describing landmarks due to its pre-trained model, but for a more general case the backbone model should be trained differently which will increase the accuracy. Future work could also include speed improvement by looking at object detection methods.

In this report, the analysis and design of a system that extracts keyframes from videos is detailed. The need for such a sub-module stems from the similarity of frames in a video. To aid in reducing the computation time of the content based video search engine, the Keyframe Extraction Module reduces the amount of frames by discarding frames that are similar in information. Determining what frames can be considered similar is one of the main challenges, as there are many ways of assigning values to how much frames differ. In the past decades, many research has been done on keyframe extraction and video summarization and many methods are proposed to form keyframe selections, varying in what is considered salient information and varying in computation time. The most challenging part of the design is that there is a time constraint present, which called for a proper analysis in what methods are suitable. After all, this limitation when creating video summaries is often not a large topic in research papers.This report will cover Shot Detection techniques along with various Keyframe Extraction methods that can be categorized in clustering, visual content, fixed selection and uniform sampling. Furthermore, evaluation methods like the Fidelity measure for the performance of particular methods are also addressed, as determining how well a keyframe selection is is not trivial. It is concluded that out of the techniques analyzed, a combination of VSUMM clustering and histogram matching along with histogram-based shot based detection with a CFAR threshold and pre-sampling is most suitable for the general case under a time constraint. Future work could include looking at hierarchical clustering methods and optimizing the Shot Boundary Detection module following the most recent papers.

THz time-domain systems driven by photoconductive antennas (PCAs) promise a bandwidth of the order of hundreds of GHz. This characteristic can be utilized to build see-through radars with sub-mm resolution and a field of view comparable to the current mm-wave radars at a relatively low cost. These radars can solve the issues in the field of public safety and security. However, the analysis and design of a THz time-domain radar is a challenging and time-consuming task even at the preliminary level. In this work, a Matlab based GUI is built that serves as a one-stop solution for the THz time-domain radar design and analysis. The GUI has three interfaces, the first one is the “Transmitter interface” which implements the Norton equivalent circuit of the PCAs and quantifies its radiated THz power along with other important parameters of the PCA semiconductor gap such as the conductance and current across the gap, and the resistance of the gap. The second interface, known as the “QO Channel Analyzer interface”, of the GUI tool analyzes a pre-defined QO channel model to estimate the power loss in the channel and the power received at the receiver. The third interface is the “Main interface” which analyzes the detection and imaging-related parameters of the radar such as the resolution, field-of-view, detection bandwidth and integration time.

The effectiveness of ultraviolet (UV) irradiation for disinfection has been well-known for more than a century. Recent advances in light­-emitting diode (LED) technology have brought UV LEDs to the consumer market. These offer many advantages over traditional gas­-discharge lamps, allowing UV radiation to be generated with higher reliability in a much smaller form factor. In this thesis, the design of a portable battery­-powered sterilizer based on UV LEDs is described. The sterilizer is intended for the sterilization of personal items such as mobile phones, face masks, and keys.

The design of the sterilizer consisted of two parts: the LED array and the LED driver. For their design, a model­-based approach was taken to ensure high performance and low cost. For the LED array, a radiometric model was developed and applied in an optimization procedure. The design of the LED driver was based on an analytical loss model in MATLAB and LTspice simulations.

The UV LED array provides a radiant power of 113 mW to the disinfection area, resulting in an irradiation dose of 195 μW/cm2. This allows for more than 99.9% disinfection in less than 15 min. The LED driver is a boost converter operating at 750 kHz in the discontinuous conduction mode. This converter supplies the array with a nominal input power of 4.3 W at a simulated efficiency of 91%.

The design of the LED driver is verified with a prototype. Measurements show efficiencies around 83% at 4.1 W output power. Incorporating several improvements over the prototype, efficiencies between 85–87% can be expected for the LED driver in the sterilizer.

This thesis is concerned with a first time demonstration of a high frequency scanning lens phased­-array with dynamic steering. Such phased­-array could provide the high gain, wide bandwidth, and steering capability that the next generation of wireless ap­plications requires. The phased-­array development was divided into three distinct parts. First, the electronic phase steering system was discussed and developed, and its stability was characterized over time. Results obtained over 24 hours indicated that the system would be stable to within 6° of high frequency output phase. The drifts in output amplitude remained to within 0.3dB. Next, the lens antenna elements were developed with a novel corrugated leaky ­wave feed that provides very high and broadband aperture efficiency. Antenna measurements have validated its ability to be a very aperture efficient radiator (>80%) with wide bandwidth (1:2). Scanning measurements have shown a 3dB scan loss at around 15°. Finally, the phase steer­ing and antenna elements were combined into a 4×1 array concept. Its behaviour was simulated and indicated again good performance with wide bandwidth (1:2) and high directivity (31.5dBi). Simulated scan loss was approximately 3dB at 20°. The amplitude and phase errors resulted in a SLL standard deviation of 0.63dB. The array prototype is currently awaiting completion of fabrication by DEMO at the TU Delft.

Photoconductive antennas have been used extensively for THz radiation the last few years. In this thesis, we propose a photoconductive connected dipole array consisting of 36x36 elements that is used as a feed in a THz silicon lens and radiates in a band ranging from 100 GHz to 5 THz. Specifically, we compute theoretically the radiated field patterns of the array as well as the secondary field beyond the lens. Furthermore, we investigate the feasibility of the fabrication of such a photoconductive connected array, given the challenges in 3D printing of a um-sized microlens array that is used to focus the laser power to the excitation gaps of the dipoles. We conclude that microlenses of sufficient accuracy can be fabricated in the premises of TU Delft without compromising the efficiency expected in theory. Lastly, we build a Matlab GUI that computes the far field radiated by various types and sizes of lens antennas in transmission, provided that the field radiated by the feed is already known. This tool has been successfully validated and supported the work in the first part of the thesis at the calculation of the far field of our proposed THz silicon lens antenna.

In recent years, the demand for the wireless connectivity is increasing and leads to the research into the above 100GHz design. Developments have been made with the circuit and technologies to make the circuit operate at 100GHz. The current availability of silicon-based technologies providing devices with ft/fmax frequencies exceeding 300-GHz allows targeting above 100GHz for commercial telecom links.
The project presents an automated design approach that utilizes “Ocean” and “Skill” code to automate parts of the design flow of the mm-wave power amplifier. The Ocean script, the scripting language used by the cadence environment to control the simulations, together with Matlab is applied for data handling and to automate lots of the recurring simulations. The Skill-code, which under the cadence’s environment is applied to automatically generate the tunable layouts to realize a DRC conform circuit design.
The design of a 170-GHz Power Amplifier in 0.13-μm SiGe BiCMOS applied the automation design flow is presented in this thesis. The gain-stage characteristics such as gain, stability, power-added-efficiency, saturated output power or 1dB-Compression are automatically generated and plotted. Specifically, this 170-GHz PA achieves a 15dB power gain and 11.7dBm output 1dB compression point.

Photoconductive antennas (PCAs) have been extensively utilized for the generation of broadband pulses over very large bandwidths. PCAs rely on a semiconductor (e.g. LT- GaAs) gap pumped by a laser and coupled to a passive structure biased at a certain voltage level. When the laser impinges on the semiconductor gap with an appropriate carrier frequency, enough energy is provided such that free electron-hole pairs are generated from the electrons that move from the valence band to the conduction band. As a result, the resistivity of the material decreases to a few ohms which in turns allows a time-varying current to flow across the gap. In recent year different hybrid equivalent circuits [1], [2], [3] have been developed in order to take into account all these complex phenomena although none of these models account for the frequency dependence of the impedance of the antenna, being formulated in the time domain. This approximation works for non-dispersive antennas such as the bow-tie, but fails in the characterization of more diverse and complex structures. The Norton equivalent circuit’s aim proposed in [4] was to fill the aforementioned gap by introducing an analytical model completely in frequency domain, although the difficulty in the characterization of the generator impedance obstructed the way for a wide acceptance in the community. In this thesis a novel approach based on a commercially available electromagnetic simulator [5] to characterise the biasing of the passive structure, the optical laser excitation and the impulse response of the photoconductor is proposed. The accuracy of the model is verified by calculating the average power radiated by a bow-tie and the results are compared to the measurements in [6]. Moreover, a revised version of the Norton equivalent circuit [4] which describes more accurately the effective generator impedance is presented. While the computer-aided model offers great introspection in the characterisation of voltages and currents and thus in the maximisation of the power radiated, the revised Norton equivalent circuit offers an even better accuracy and reduces significantly the computational time.

Problem Definition
This bachelor final project was an assignment given by the start up Momo Medical named: "Improving health care with technology". This project was to use their existing technology and ideas to create a system to prevent pressure ulcers. Pressure ulcer are the medical condition that happens when a patient is left too long in the same position in bed. In order to combat this the nurses turn the patient onto a new posture every 3 to 4 hours. To improve upon this Momo Medical decided to built a sensor pad that has to be placed under the mattress in order to track the patients bed posture and to alarm the nurses when a patient has been on a side too long. The project group was divided into three parts: the sensor group who would look at the sensor pad, the algorithm group who would use the data from the sensor pad to determine the bed posture and the test group that had to built a system that could verify the algorithm. This report will talk about the test group. Based upon the data that Momo Medical acquired it was also possible to acquire the heart rate and the respiratory rate from the sensor data. A way to acquire this data was also looked at. The design requirement that were set were 95% accuracy for bed posture and 90% for respiratory rate and heart rate.
Design Description
In order achieve these requirements a few different types of reference signals were looked at that could be used in order to detect the bed posture, respiratory rate and heart rate. In the end an accelerometer and an commercial grade ECG chest strap were chosen. The accelerometer could measure the patients orientation and thus his bed posture. When the accelerometer is attached to the chest it is also able to measure the chest compressing as a result of breathing. The ECG chest strap was chosen as this gave a good platform upon which the accelerometer and the micro controller could be attached.
Evaluation
A prototype was built upon our design and it was tested in different situations. In the end the accuracy for the bed posture detection was 99%, the accuracy of the respiratory rate was 91% and of the heart rate it was 95%. With these accuracies it meant that the design requirements were met. The disadvantage was that the system influences the sensor pad while the patient is wearing the prototype. Another disadvantage was that due to the low sampling speed it was not possible to attach the heart rate receiver to the testing group's prototype, but had to be attached to the sensor group's prototype. These two disadvantages could be improved in the future by making the prototype smaller and faster.

The founders of Momo Medical envisioned a health care product that would help nurses worldwide with pressure ulcer prevention. Pressure ulcers are a chronic wound that affects the skin of patients who do not regularly change bed posture. As it currently stands, nurses lack the manpower and time to change the postureof patients in a timely manor. Thus, a goal was set to create a product that would reduce this strain on nurses.The goal was to design a non-intrusive product that keeps track of the posture of a patient on a hospital bed, and send an alarm if this has not changed in a set amount of time. A constraint with this was that the product has to lay directly underneath the mattress of the patient. Furthermore, the product has to make use of as few sensors as possible. With these constrains, the project group was divided in three subgroups: the sensor, the algorithm, and the testing subgroup.This report will focus on the design, implementation, and evaluation of the algorithm. The algorithm has to be designed to process the sensor data and return, with 90% accuracy, the posture of the patient. It also has to be able to determine with 99% accuracy whether or not there is a patient on the bed. An optional objective was to design an algorithm to determine the heart and respiration rate of the patient, with 80% accuracy, but this was of lower priority. It was soon decided to give priority to posture detection and not further investigate detection of the heart and respiration rate.In the final product, the information about the patient’s posture will be displayed in a graphical user interface (GUI). This GUI has to be user-friendly and intuitive to use by the nurses. During the prototyping phase, however, a different GUI was used, one that displayed various debugging information. This prototype GUI was necessary to test the various algorithms.Several different mathematical models have been investigated, implemented, and tested. These models range from taking the variance of the sensor data, to deploying a Fourier transform on a polynomial curve based on the sensor data. With preliminary test data, an accuracy of 93% was achieved using a neural network. With the final testing data however, only the neural network reached the required accuracy of 90%, though only just, at 90.8%. With future iterations of the algorithm, more research and data collection is advised, as neural networks work better with larger data sets.