Electrical engineering

Permanent URI for this collectionhttps://digital.lib.washington.edu/handle/1773/4915

Browse

Now showing 1 - 20 of 422

Evaluating the Efficiency of Neural Network Implementations on AMD Versal AI Engines
(2025-01-23) Shen, Yilin; Hauck, Scott
The AI Engine (AIE) is an optional component of the AMD Versal Adaptive SoC. It is an innovative device that offers extensive parallelism to enhance compute density and reduce power consumption. However, the performance of the AIE, particularly for small models requiring low latency, remains uncertain.In this research, we mapped three neural network benchmarks to the AIE section of the Versal VCK190. We explored the best coding practices and characteristics of the AIE. Additionally, we mapped these models to the FPGA fabric portion of the VCK190 and compared the cost and performance with our AIE implementation. Based on six metrics, we found that the AIE's efficiency is slightly better than the FPGA fabric in terms of power and silicon area utilization, but worse than the FPGA in terms of performance, resource utilization and price. This discrepancy is due to limitations in interconnection and the inefficiency of hardware units when the vector data path cannot adapt to certain shapes of the input data.
Low-Cost 3-D Printed Helical Antenna with Dielectric Support
Kung, Cheryl; Charczenko, Walter
Satellite-based internet connection requires high directivity, millimeter wave phased array antennas to be able to receive and transmit signals effectively. Phased array antennas for millimeter waves have historically been very expensive to manufacture. Exploring low-cost methods for manufacturing high directivity antennas may bring down the costs of these systems, allowing more equitable access to internet. Helical antennas are a type of high directivity antenna that can be used for these purposes. However, helical antennas are difficult to manufacture and scale due to its three dimensional (3-D) shape of the helix conductor. New 3-D printing technology allows the creation of a dielectric support for the helical antenna element. This adds mechanical rigidity to the antenna and is feasible for high volume manufacturing at a lower cost. This thesis explores the design of a low-cost helical antenna using a 3-D printed dielectric core for mechanical support. The research in this thesis concludes that it is possible to design a helical antenna using low-cost dielectric materials with high relative permittivity at microwave frequencies. As a proof of principle, a 5 GHz helical antenna embedded in a solid dielectric was designed and modeled using electromagnetic field simulation software. At 5 GHz, the software simulations can be compared to helical antennas that are manufactured on conventional 3-D printers and commonly used resin dielectrics. The conclusion and results of the computer simulations show that helical antennas with dielectric support will radiate in the axial mode with high directivity and circular polarization.
Optimizing FPGA Resource Allocation in SDR Remote Laboratories via Partial Reconfiguration
(2024-09-09) Zhang, Zhiyun; Hussein, Rania RH; Chen, Tai-Chang TC
Software-Defined Radio (SDR) remote laboratories provide engineering students in wireless communications and radio frequency courses with hands-on experience using SDRs. These devices are renowned for their flexibility and reconfigurability via software. However, cost-effective SDRs often feature lower-end System-on-Chips (SoCs) with Field Programmable Gate Arrays (FPGAs) capable of parallel data processing but limited in hardware resources for running complex digital signal processing algorithms. This limitation restricts their use to simpler FPGA-based tasks, reducing their operational flexibility. Partial Reconfiguration (PR) offers a compelling solution to these limitations by dynamically allocating hardware resources based on the system’s operational mode, enhancing the FPGA's functionality. PR allows the execution of complex programs by enabling independent modifications, recompilation, and reconfiguration of specific FPGA regions without requiring a full project recompile. This process streamlines modifications, shortens development cycles, and enables rapid iteration, significantly improving conventional FPGA programming techniques. This thesis investigates the implementation of PR within SDRs, specifically on Red Pitaya’s SDR platforms which, to the best of our knowledge, have not been previously used for PR. It also explores the integration of this implementation into the existing structure of an SDR remote laboratory. Experimental results demonstrate notable improvements in hardware efficiency, including reductions in logic resource utilization and total power consumption, compared to traditional FPGA reconfiguration methods. These findings underscore the potential of PR in advancing the field of SDR.
Examination of Drone Localization Performance with Commercially Available Embedded GPS Sensors
Nathan, Gokul; Makhsous, Sep
With the commercial drone market worth 24 billion USD and project to grow, accurate 3D localization in urban air mobility is critical. Existing GPS systems in these environments are plagued by multipath propagation and signal obstruction, resulting in typical urban GPS errors ranging from 15 to 100 meters, far below the precision needed for safe operations. Using a combination of experimental testing and data-driven modeling, this study quantitatively demonstrates that a composite machine learning approach using commercially available embedded GPS sensors, incorporating a Windowed Inverse Variance Weighted Filter combined with LSTM Recurrent Neural Networks, can significantly reduce localization errors. Testing across varied urban settings, GPS sensors improved location accuracy by up to 47% compared to conventional filter methods in literature, with the model effectively reducing the average error in urban environments to less than 1.8 meters. Dynamic Accuracy Index (DAI), a metric quantifying the balance between positional accuracy and data processing speed, is introduced. The experimental results reveal our system's proficiency, particularly evident in its DAI of 0.18 meter/Hz, surpassing other conventional filter methods operating at approximately 1 Hz with a DAI of 5.11 meter/Hz and proposed filter + neural network methods presented in the literature with DAI of 2.17. The findings affirm the potential of using recurrent neural network-based machine learning algorithms in enhancing GPS localization systems, presenting a viable pathway for integrating these technologies into commercial UAV operations. This contribution is not only pivotal for the advancement of urban air mobility but also enhances the safety and operational efficiency of UAVs in complex environments.
Enabling Vector Load and Store Instructions on HammerBlade Architecture
Ramstad, Robert; Taylor, Michael
Traditionally, computer architecture has been dominated by overly complex instruction sets that created a ''solution" to every problem by adding another instruction. If these complex instructions sets are one side of a coin, the Reduced Instruction Set Computer (RISC)-V architecture is the other. RISC-V processors consist of 47 base instructions. Having such a low amount of instructions is both the biggest strength and the biggest weakness of the new era of RISC-V processors. Currently, there is a remarkable lack of high performance RISC-V processors. The Hammerblade architecture is one of the few. The main difference between Hammerblade and other RISC-V processors is its leveraging of parallel computer architecture as a multi-core system. However, while Hammerblade consists of potentially thousands of cores, it does not perform any data-level parallelism.The primary intent of this thesis is to further understand how to increase the local memory throughput by extending the RISC-V core to include Single Instruction Multiple Data (SIMD) loads and stores. This will add the capability to locally load and store four words of data on top of the existing singular word loads and stores. A single Vanilla RISC-V core can now have the potential of a 4x speedup in loads and stores. This will allow Hammerblade to not only leverage the parallelism of the manycore architecture, but also the data-level parallelism on each individual core.
Development and Application of an Automated “Scan to Plan” System for Precision Paint Ablation Using LiDAR and Laser Cameras on Robotic Arms
(2024-02-12) Guo, James; Shapiro, Linda
The thesis presents an innovative approach to automating aircraft maintenance processes,particularly in paint and coating removal. It introduces a “Scan to Plan” system that combines LiDAR scanning and laser cameras with robotic arms for precision paint ablation. The work emphasizes the development of a modular system design, allowing for flexible adaptation to various industrial applications. The prototype demonstrates significant accuracy in scanning and paint ablation on curved aluminum surfaces, highlighting its applicability in laser surface operations beyond paint removal, such as micromachining and cladding. The thesis also explores potential enhancements to the system and its broader applications in industrial automation, paving the way for more efficient and precise maintenance procedures.
Evaluating the Quality of HLS4ML’s Basic Neural Network Implementations on FPGAs
(2024-02-12) Johnson, Caroline; Hauck, Scott
Field-programmable gate arrays (FPGAs) offer an attractive platform for machine learning models; however, creating these models is a time-consuming process that demands specialized knowledge. High-level synthesis (HLS) offers the potential to streamline this hardware development, but with the trade-off of not being able to hand tune the design to take advantage of the specific resources and performance capabilities available. The popular open-source Python library HLS4ML offers low latency HLS machine learning models. However, it is unclear how much quality reduction is caused by adopting an HLS design flow. In this thesis we create carefully optimized SystemVerilog versions of identical HLS4ML designs, allowing us to evaluate the quality of the designs produced by HLS4ML. Our research reveals that HLS designs are highly competitive with hand-optimization techniques for basic layers with standard parameters, especially in terms of the capabilities of the Vivado HLS tools. However, when advanced parameters like reuse factor and stride are altered, our hand-written design is able to achieve a 2x-3x increase in performance over HLS4ML’s. Through this evaluation we were able to identify weaknesses in the existing tools and develop workarounds to help provide significant quality improvements.
FAIR Modeling for Perovskite Solar Cells: An Open Source Machine Learning Pipeline
Roberts, Nicholas; Lin, Lih Y
Perovskite solar cells (PSCs) show great promise for commercialization, rivaling traditional silicon-crystal solar cell efficiency despite their comparatively short research lifetime. This efficiency is achieved while being manufactured at low temperatures and in ambient conditions, lowering fabrication costs dramatically. Machine learning (ML) promises to significantly expedite further optimization by recommending novel configurations based on insight from existing literature. This paper utilizes the Perovskite Database Project (PDP), an open source PSC database consisting of over 43,000 entries from published literature, to train three ML architectures with short circuit current density (J$_{sc}$) as a target. Using the XGBoost architecture, an RMSE of 3.73 $\frac{mA}{cm^2}$, R-value of 0.63, and MPE of 10.35% were achieved. This performance is comparable to the results reported in literature and through further investigation could likely be improved. To overcome the challenges of manual database creation, an open-sourced data cleaning-pipeline was created to leverage the PDP. Through the creation of these tools this research aims to increase the availability of ML as a tool to promote improvement in novel device configurations for PSC while showing the already promising performance achieved.
On the Feasibility of Laser Inter-satellite Links for Low-latency High Frequency Trading
Singh, Vaibhav; Gadre, Dr. Akshay
Internet satellite constellations are proliferating and enabling new applications in long-range and large-scale connectivity. One such promising application is providing low-latency communication for high frequency trading by leveraging their global coverage. While existing approaches leverage higher orbit satellites as low-latency relays for private clients, there is a new opportunity that can provide low-latency connectivity to every customer via the satellite mesh – laser inter-satellite links (LISLs).While past research has explored the capability of LISLs for enabling physical deployment and impact on the constellation topology design, this work presents a holistic empirical study on the feasibility of leveraging LISLs for low-latency high frequency trading. We evaluate the feasibility, utility and practical limits of the latency benefits of LISL-based constellations for ten large and geographically diverse financial markets across over 150 million satellite-seconds of simulation. Our empirical evaluation across two of the largest proposed satellite constellations, Starlink and Project Kuiper, demonstrates that LISLs can provide full coverage with roughly 1000 satellites and improve latency for 78.76% of link-pairs. However, these benefits decline drastically with processing delays of 4 ms to roughly 19.05% of link-pairs.
MiniMeta-Miniature Imaging System with Meta-optics for Biomedical Applications
Kumari, Khushboo; Majumdar, Arka
This thesis delves into the world of meta-optics for biomedical applications imaging. Mission to improve the diagnosis, prevention and treatment of diseases by developing novel imaging technology is booming in the research globe. Medical imaging in hospitals and laboratories has shown growing benefits around the globe. In recent years plenty of research articles have been published highlighting the need of medical imaging in solving problems of healthcare institutions. Size of the optics limits the imaging functionalities. For example neurons are inherently distributed in 3D, which leads to a need of probing upto an extended depth of focus. This is the first motivation we are moving towards meta-optics in biomedical. The work begins with the foundation in metasurfaces, examining issues such as narrow depth of field in metalenses and the development of EDoF called extended depth of focus in metalenses with larger working distances. Alternative solution for miniature imaging system called Miniscope developed by UCLA researchers. It further delves into the integration of meta-optics with different versions of Miniscopes. Subsequent chapters explore the extension of different metalens capabilities with miniscope, so called MiniMeta. It further delves into the more imaging features required for microscopic images and how the size of the optics limits the imaging features like, resolution, wide field of view and simultaneous multiple imaging functionalities with minimal distortion of metalenses. The thesis concludes with multiple resulting solutions in microscopic images for biomedical applications. Overall this thesis looks at the future opportunities to share discoveries and advanced imaging research with metalenses for multiple medical applications.
Smartphone-based 3D Scanning and Image Processing for Ostomy Wafer Customization
(2023-09-27) Luquin Monroy, Francisco; Hussein, Rania; Mamishev, Alexander
Millions of ostomy patients rely on outdated tools to adjust their wafers, which must fit around their stoma. The stoma can be described as a surgically created hole in a patient's abdomen, from which the patient’s intestinal or urinary tract will divert bodily waste. One of the purposes of ostomy wafers is to protect the surrounding skin from bodily waste, and incorrectly fitting the wafers can lead to many complications. For example, a tightly fitting wafer can cut into the nerveless intestine, leading to unnoticed excessive bleeding. Conversely, a loosely fitting wafer exposes skin to bodily waste, which can result in skin erosion and infection. Recognizing the need for better tools for ostomy patients, we are developing Osto-Mate, a system utilizing smartphones with depth-sensing capabilities, to provide ostomy patients with customized wafers. This thesis explores the characterization of the iPhone’s TrueDepth sensor to confirm its suitability as a 3D scanning device for the system and introduces a method for extracting stoma model contours from 2.5D images. It further discusses the design and functionality of the application, illustrating its potential to improve the process of ostomy wafer fitting. Preliminary tests show promising results for enhancing the accuracy of wafer adjustment, paving the way for improved patient safety and convenience.
Characterization of Nafion-based Resistive Switching Devices
Bhandari, Pritam; Choi, Seungkeun
The development of computers in the modern era has escalated the race towards the development of powerful and efficient memory devices. By using advanced miniaturization techniques and new materials, we have been able to dramatically reduce the size of the memory devices while increasing the storage capacity and computing performance. However, we are reaching a point of slower growth in the computing performance of MOSFET-based nonvolatile memory devices. It becomes increasingly difficult to further decrease the size of memory devices. Hence, the next generation memory technology must have the following features to meet the high computing performance in the era of artificial intelligence: low-power consumption, fast switching, non-volatile, high-density fabrication. Resistive Random-Access Memory Devices (ReRAM) meets all those requirements; hence, is considered as one of the promising candidates for the next generation memory technologies. In this research, a ReRAM device with Nafion as a switching layer was fabricated. To characterize the resistive switching performance, Nafion was annealed at three different temperatures: 30Â°C, 60Â°C, and 90Â°C. In order to study the effect of different electrode, we used two different bottom electrodes (Au and Cu) and Al as a top electrode. The devices with Cu as a bottom electrode exhibited good resistive switching properties while the device with Au as a bottom electrode showed little or negligible switching performance. We found that the performance of switching was best when Nafion was annealed at 60Â°C. However, the experiment shows a wide variation of device performance even in the same substrate, indicating the importance of uniform film thickness and quality of Nafion.
RHL-Butterfly: A Scalable IoT-Based Breadboard Platform for Embedded Systems and Remote Laboratories
(2023-08-14) Guo, Matthew; Hussein, Rania
The RHL-Butterfly is a research to practice virtualized breadboard solution for FPGAs and ARM microcontrollers in remote laboratories and engineering education curriculum. The COVID-19 pandemic brought challenges to traditional engineering education practices, particularly with hardware, hands-on engineering practices without compromising creativity and instruction. The RHL-Butterfly aims to address traditional engineering education shortcomings and provide equitable access for all students interested in the engineering curriculum. This work improves upon an existing virtual breadboard model by using virtualization to interface a virtual breadboard with physical, remote hardware from a website user interface and presents a solution to support FPGAs and ARM microcontrollers and supporting intermediate logic gate integrated circuits. The new iteration of the virtualized breadboard uses a custom protocol that converts the graphically represented breadboard layout into a 1D string representation for network communication. The 1D string representation is then parsed in a custom designed, open-source, and scalable breadboard parser for embedded systems. This balance between a virtualized interface and physical hardware implementation preserves a hardware curriculum embedded systems engineering education and brings a promising solution to expand the scalability and accessibility of engineering labs.
FPGA Design Upgrades for the ATLAS Pixel Readout System in the Large Hadron Collider
(2023-08-14) Roychoudhury, Sanjukta; Hauck, Scott
The ATLAS Pixel Detector in the Large Hadron Collider uses a system of FPGAs in the off-detector readout system. The Read Out Driver (ROD) is a major component of the system that handles processing of front end data and sending it to the next stages of filtering and storage. This thesis starts with a background on the detector and readout system and moves on to discuss efforts in support and development of ROD firmware. The first project investigated anomalous behavior in the ROD Histogrammer. This was understood and resolved after testing in the SR1 facility, which has a mockup of the detector hardware. The second project, called Smart L1A Forwarding, was restarted in 2022 with an objective to mitigate desynchronization of data in Pixel during the challenging conditions of Run-3. Initial development and tests in SR1 and the detector have shown promising results. Further modifications have been made in an attempt to resolve issues found during testing; these modifications are being tested in the detector.
A Multi-Domain Trojan Detector for Deep Neural Networks
(2023-04-17) Asokraj, Surudhi; Poovendran, Radha
Backdoor attacks have been demonstrated to compromise the functioning of machine learning models that utilize deep neural networks (DNNs). An adversary carrying out a backdoor attack embeds a predefined perturbation called a Trojan trigger into a small subset of input samples. The DNN can then be trained in a manner such that the presence of the trigger in the input results in an output label that is different from the correct label. At the same time, outputs of the DNN corresponding to inputs without the trigger remain unaffected. Backdoor attacks, where an attacker can negatively affect the DNN's behavior, might have severe repercussions in safety-critical applications. Existing defenses in the literature against backdoor attacks involve pruning or retraining DNN models, which can be computationally expensive. In addition, researchers have demonstrated the success of these solutions on input domains based on images. The performance of such defenses on other inputs needs to be understood better. In this thesis, we propose and develop MDTD, a multi-domain Trojan detector. MDTD for DNNs has several distinguishing characteristics, including (i) not requiring retraining DNN models (ii) not requiring knowledge of the trigger or the embedding strategy of the attacker, (iii) is computationally inexpensive (iv) capable of being applied to image and graph-based inputs. To the best of our knowledge, MDTD is the first Trojan detection mechanism proposed for graph-based inputs. MDTD uses the insight that input samples containing a Trojan trigger are located relatively further away from a decision boundary than clean input samples. Initially, MDTD estimates the distance to a decision boundary using adversarial learning methods. These methods estimate the smallest magnitude of noise required for the model to misclassify a sample. MDTD uses this information to infer whether a given sample is Trojaned or not. More precisely MDTD learns a threshold for the distance to the decision boundary using a small set of clean labeled samples and uses this threshold to flag a sample as possibly Trojaned. We evaluate MDTD against state-of-the-art (SOTA) Trojan detection methods across five image-based datasets - CIFAR100, CIFAR10, GTSRB, SVHN and Flowers102- and four graph-based datasets - AIDS, WinMal, Toxicant and COLLAB. Our results show that MDTD effectively identifies samples that contain different types of Trojan triggers. We also show that an adversary who trains robust DNN models using a combination of clean and Trojaned samples does not cause a significant deterioration in MDTD performance without significantly reducing the classification accuracy of the DNN model.
Observation Centric and Central Distance Recovery for Tracking of Sports Athletes
(2023-04-17) Huang, Hsiang-Wei; Hwang, Jenq-Neng
Multi-Object Tracking on humans has improved rapidly with the development of object detection and re-identification algorithms. However, multi-actor tracking over humans with similar appearance and non-linear movement can still be very challenging even for the state-of-the-art tracking algorithm. Current motion-based tracking algorithms often use Kalman Filter to predict the motion of an object, however, its linear movement assumption can cause failure in tracking when the target is not moving linearly. And for multi-player tracking over the sports field, because the players on the same team are usually wearing the same color of jersey, making re-identification even harder both in the short term and long term in the tracking process. In this work, we proposed a motion-based tracking algorithm and three post-processing pipelines for three sports including basketball, football, and volleyball, we successfully handle the tracking of the non-linear movement of players on the sports fields. Experimental results achieved a HOTA of 73.968 on the testing set of ECCV DeeperAction Challenge SportsMOT Dataset, showing the effectiveness of the proposed framework and its robustness in different sports including basketball, football, and volleyball.
Using an off-the-shelf rigid gripper to grasp objects of different strengths and compliances
(2023-04-17) Sullivan, Kate; Chizeck, Howard
Robotic grippers are being increasingly used in everyday tasks and complex missions that require the ability to perform a range of grasps. Rigid grippers have been used for decades since their start in industrial applications, but their inability to grasp delicate objects has led to the development of soft grippers. While soft grippers are showing promising results in many applications, they often are limited by low grasp forces and availability. This paper presents an algorithm for using an off-the-shelf rigid gripper to grasp objects of different strengths and compliances, without the need for tactile sensors or modifications. The algorithm was used to successfully grasp a variety of objects, most notably a potato chip and a highly compliant paper cup.
Low-Power Communication for Environmental Sensing Systems
(2023-01-21) Kapetanovic, Zerina; Smith, Joshua R.
Over the last decade, the Internet of Things (IoT) has been changing the world, from enabling connected electronics, smart homes, to smart agriculture. Today, IoT systems have the potential to make a significant impact when it comes to environmental monitoring, which has become increasingly relevant in the times of the climate change crisis and the need to achieve biodiversity conservation. Imagine being able to use passive wireless communication and deploying battery-less sensors for remote environmental monitoring. This dissertation aims to advance and empower these efforts and presents new methods of low-power wireless communication and sensing systems. First, I introduce FarmBeats, an IoT system for data driven agriculture that solves key challenges related to power, connectivity, and cost. Next, I discuss low-power downlink solutions for ambient backscatter systems. In particular, Wireless Quantization Index Modulation and 'Glaze', which build upon data-hiding techniques to enable downlink communication using existing infrastructure and occupied RF spectrum. Lastly, I present Modulated Johnson noise, a new wireless communication system that uses Johnson noise to enable very low-power wireless communication.
Multilevel Resistive Switching in a Metal Oxide Semiconductor based on MoO3
(2022-09-23) Raza, Moosa; Choi, Seungkeun
Over the years a resistive random-access memory (ReRAM) has received great attention due to its simple structure, CMOS compatible fabrication process, low-power consumption. Among other attracting memory characteristics, multilevel switching is considered as a very important feature since a single memory cell can store more than one bit of information, thereby increasing memory density. While molybdenum trioxide (MoO3) has been widely used for many optoelectronic devices as a charge transport layer, it has not been extensively investigated as a resistive switching layer.In this research, I have used MoO3 as a switching layer and demonstrated a multilevel resistive switching by controlling the compliance current. In a novel lateral device architecture, I have also demonstrated a self-compliance resistive switching behavior. However, devices need to be further optimized to reduce the operating voltage in a lateral device architecture.
Power and Energy Implications for Electrification of the UW Transportation Vehicle Fleet
(2022-07-14) Foster, Kelsey; Kirschen, Daniel
As climate change and greenhouse gas emissions become an increasing concern, there is a push for phasing out traditional combustion engine vehicles and replacing them with electric vehicles. The University of Washington has ambitious carbon emissions reduction plans including electrification of the UW Transportation vehicle fleet of over 500 vehicles. To successfully accommodate an electrified fleet with minimal cost implications and infrastructure upgrades, UW Transportation must deploy charge management and charge scheduling techniques to minimize energy, power, and charger requirements for the fleet. This thesis analyzes several strategies for when and where to charge electric vehicles in the UW Transportation fleet. As a result, charging every fleet vehicle every weekday and splitting charging between weekdays and weekends are the best options. The approach used in this analysis can be expanded to apply to other fleets using the UW Transportation fleet as a case study.

Browse

Recent Submissions