Recently, there is a growing interest in the study of median-based algorithms for distributed non-convex optimization. Two prominent examples include signSGD with majority vote, an effective approach ...for communication reduction via 1-bit compression on the local gradients, and medianSGD, an algorithm recently proposed to ensure robustness against Byzantine workers. The convergence analyses for these algorithms critically rely on the assumption that all the distributed data are drawn iid from the same distribution. However, in applications such as Federated Learning, the data across different nodes or machines can be inherently heterogeneous, which violates such an iid assumption. This work analyzes signSGD and medianSGD in distributed settings with heterogeneous data. We show that these algorithms are non-convergent whenever there is some disparity between the expected median and mean over the local gradients. To overcome this gap, we provide a novel gradient correction mechanism that perturbs the local gradients with noise, which we show can provably close the gap between mean and median of the gradients. The proposed methods largely preserve nice properties of these median-based algorithms, such as the low per-iteration communication complexity of signSGD, and further enjoy global convergence to stationary solutions. Our perturbation technique can be of independent interest when one wishes to estimate mean through a median estimator.
The Reference Governor (RG) is a methodology based on predictive control for constraint management of pre-stablized closed-loop systems. This problem is motivated by the fact that control systems are ...usually subject to physical restrictions, hardware protection, and safety and efficiency considerations. The goal of RG is to optimize the tracking performance while ensuring that the constraints are satisfied. Due to structural limitations of RG, however, these requirements are difficult to meet for Multi-Input Multi-Output (MIMO) systems or systems with preview information. Hence, in this dissertation, three extensions of RG for constraint management of these classes of systems are developed. The first approach aims to solve constraint management problem for linear MIMO systems based on decoupling the input-output dynamics, followed by the deployment of a bank of RGs for each decoupled channel, namely Decoupled Reference Governor (DRG). This idea was originally developed in my previous work based on transfer function decoupling, namely DRG-tf. This dissertation improves the design of DRG-tf, analyzes the transient performance of DRG-tf, and extends the DRG formula to state space representations. The second scheme, which is called Preview Reference Governor, extends the applicability of RG to systems incorporated with the preview information of the reference and disturbance signals. The third subject focuses on enforcing constraints on nonlinear MIMO systems. To achieve this goal, three different methods are established. In the first approach, which is referred to as the Nonlinear Decoupled Reference Governor (NL-DRG), instead of employing the Maximal Admissible set and using the decoupling methods as the DRG does, numerical simulations are used to compute the constraint-admissible setpoints. Given the extensive numerical simulations required to implement NL-DRG, the second approach, namely Modified RG (M-RG), is proposed to reduce the computational burden of NL-DRG. This solution consists of the sequential application of different RGs based on linear prediction models, each robustified to account for the worst-case linearization error as well as coupling behavior. Due to this robustification, however, M-RG may lead to a conservative response. To lower the computation time of NL-DRG while improving the performance of M-RG, the third approach, which is referred to as Neural Network DRG (NN-DRG), is proposed. The main idea behinds NN-DRG is to approximate the input-output mapping of NL-DRG with a well-trained NN model. Afterwards, a Quadratic Program is solved to augment the results of NN such that the constraints are satisfied at the next timestep. Additionally, motivated by the broad utilization of quadcopter drones and the necessity to impose constraints on the angles and angle rates of drones, the simulation and experimental results of the proposed nonlinear RG-based methods on a real quadcopter are demonstrated.
The Solid-State Circuit Breaker (SSCB), as an emerging semiconductor-based circuit protection technology, is featured with its extremely fast fault interruption/isolation speed and regarded as a ...promising alternate to the electromechanical circuit breakers in the DC systems. However, in the conventional SSCBs, large surge voltages are clamped across their semiconductor switches when the breakers open and the dynamic voltage unbalance is incurred when the series-connected switches are used. With these technical defects, the efficiencies and reliabilities of the SSCBs are impaired and their wide adoption to the DC distribution systems is set back. To overcome these technical limits of conventional SSCBs, four types of Multilevel and Surgeless Solid-State Circuit Breakers have been proposed in this dissertation. By utilizing the fast switching speeds of the semiconductor switches, the proposed SSCBs can commutate the fault current to the different conduction paths of the circuit breakers and attain significant benefits on efficiency and fault isolation speeds in comparison with the conventional SSCBs. Particularly, for the proposed Multilevel Solid-State Circuit Breaker (MLSSCB), the series-connected switches are clamped to their voltage dividing capacitors during their switching transience and then the dynamic voltage unbalancing issues among the switches can be averted. For the proposed surgeless SSCBs, with surge voltage suppressed, the semiconductor switches do not need to be overdesigned for the voltage ratings and the conduction efficiencies of the SSCBs can be improved on the ground that the semiconductor device with higher voltage block capability has thicker drift regions and larger on-state resistance. Derived from the integration of the Ground-Clamped Surgeless SSCB and the Multilevel SSCB, the proposed Surgeless Multilevel SSCB (SMLSSCB) can solve both the surge voltage and dynamic voltage unbalancing issues in the medium voltage DC SSCBs and attain higher efficiency and an ultra-fast isolation speed prior to the other SSCBs. A fault-tolerant configuration of the SMLSSCB has also been proposed to improve the reliability of SMLSSCB and make it prior to that of the conventional SSCBs. In this dissertation, the operating principles of the proposed SSCBs have been presented. Besides, to demonstrate the proposed SSCBs’ advantages over the conventional SSCBs on fault isolation speeds, power efficiencies and reliability, the comparisons between the proposed SSCBs and their counterparts of the conventional SSCBs have been made in terms of several key parameters of the circuit breakers. Additionally, the simulation/experiment results and design considerations of the proposed circuit breakers have been introduced to validate their technical feasibilities and practical uses.
This thesis investigates design considerations and topologies for spin-qubit control circuits integrated with qubits and readout circuits into one- and two-lane quantum processor test vehicles in a ...production 22-nm FDSOI process. The control circuits produce millimeter-wave pulses through direct modulation of an externally generated carrier in the DC-to-70-GHz and 140-to-160-GHz frequency ranges. At 2 K, one design featuring a shunt-series-shunt switch achieves 31.1 dB to 77.7 dB of on/off isolation from DC to 27 GHz, while another featuring a distributed switch achieves > 30.2 dB of on/off isolation at 2-46 GHz and 50-70 GHz. The standalone distributed switch shows > 34 dB of on/off isolation over 40-218 GHz at 295 K. These measurements demonstrate not only the possibility of minimizing qubit idling errors with the adopted switch topologies, but also the feasibility of the proposed elevated-temperature (1-7 K) monolithic quantum processor and methodology for designing cryogenic control circuits.
The rapid growth of consumer and industrial electronic devices has created unprecedented demand for fast and reliable mobile wireless access. Since the rise of the fifth-generation (5G) wireless ...communications, the millimeter-wave (mmWave) frequency band (commonly considered between 30 GHz to 100 GHz in 5G) has been extensively investigated due to the tremendous amount of raw bandwidth (several GHz continuous spectrum resource available for cellular networks). The success of 5G system deployment since 2019 has motivated the wireless community to look into the Terahertz regime where tens of GHz unexplored bandwidth exists. THz communications are nominated as a vital technology for the next-generation (6G) wireless systems and beyond that are envisioned to support prospective applications such as wireless cognition, ultra-high definition (UHD) virtual reality and augmented reality (VR/AR), and spectroscopy. Sub-THz frequencies (100 GHz - 300 GHz) have been studied as the next frontier leading to THz communications over the past few years. The underlying propagation channel characteristics at sub-THz frequencies are fundamental for wireless system design. Therefore, accurate channel models that can faithfully recreate the channel statistics for frequencies above 100 GHz are imperatively needed. This work explores the use case of sub-THz communications in industrial scenarios to enable 6G intelligent factories that are envisioned to support Tbps networks and centimeter-level sensing/positioning. This work investigates several fundamental aspects of channel modeling and performance evaluation of sub-THz communication systems based on a 142 GHz propagation channel measurement campaign conducted over 80 locations in four factories with different topologies and production facilities in 2021-2022. The propagation measurements are conducted using a sliding-correlation-based channel sounding system equipped with steerable directional horn antennas with 27 dBi gain and 8-degree HPBW. The measurements are conducted over a wide range of distances from 5 m to 85 m in both line-of-sight (LOS) and non-LOS (NLOS) environments, yielding over 75,000 directional power delay profile (PDP) measurements for statistical channel modeling. In addition, a novel multipath parameter estimation algorithm, antenna de-embedded multipath extraction (ADME), is proposed to resolve multipath components (MPCs) from measured directional PDPs with 1-degree spatial resolution and two ns~temporal resolution. The factory propagation channel characteristics at 142 GHz, such as path loss, delay spread, and angular spread, are extensively analyzed and compared with early measurement campaigns in indoor offices and urban open square scenarios. For example, the omnidirectional path loss measurements yield path loss exponents (PLE) of 1.8 and 3.1 for the industrial LOS and NLOS scenarios at 142 GHz. The NLOS PLE of factories is higher than the values in offices (PLE=2.8) and urban scenarios (PLE=2.9), suggesting that the factories can be harsh for sub-THz signals, especially in the NLOS environments, due to the heavy clutter such as machinery, workstations, and large shelves. Furthermore, the propagation measurements with a near-the-ground receiver (RX) (i.e., 0.5 m height) are compared with the channel characteristics at 1.5 m RX height. As the RX height drops from 1.5 m to 0.5 m, the LOS probability reduces from 42% to 8%, and the omnidirectional PLE increases from 2.5 to 2.9. The results show that the antenna height dramatically influences the channel characteristics, where low-RX channels tend to have higher path loss and larger azimuth angular spread of arrival (ASA) caused by the local clutter surrounding the low RX. Channel enhancement methods are investigated to improve the factory channel condition, particularly for the near-the-ground RXs. We leverage a 1 m x 1 m flat aluminum plate as a perfect passive reflecting surface to increase the reflecting path powers. The metal plate is carefully rotated in the horizontal plane to identify the best reflecting orientation with the highest received power. The rotation of the metal plate emulates the dynamic beam steering of reconfigurable intelligent surfaces, which are expected to be a key technology for 6G wireless communications and beyond. The measurements show that a metal plate significantly increases the received power, where the omnidirectional PLE for NLOS locations approaches two (i.e., close to the free space propagation). Overall, the metal plates are shown to create (additional) strong paths, which can counter the random dynamic blockage, improve spatial multiplexing, and provide a critical communication link for near-outage locations. A complete channel model for the sub-THz indoor factory (InF) scenario, including large-scale path loss models and small-scale spatial statistical multipath channel models, is derived from the 75,000 directional PDP measurements for LOS and NLOS scenarios. The proposed sub-THz InF statistical channel model is compared with the standard reference channel model in the 3rd Generation Partnership Program (3GPP) TR 38.901 Release 16 for the InF scenario that is currently applicable for frequencies below 100 GHz and provides critical insights on upcoming channel modeling development above 100 GHz in standardization bodies. The derived statistical channel models are implemented in a mmWave and sub-THz channel simulator, NYUSIM. The root-mean-square (RMS) delay spread and angular spread of channel impulse responses generated by NYUSIM show satisfactory agreement with the empirical values from the measurements. NYUSIM now generates life-like drop-based and spatially consistent channel impulse responses for outdoor, indoor, and factory scenarios for mmWave and sub-THz frequencies to facilitate communication algorithm designs such as hybrid beamforming and link-level simulations. This work presents the first extensive channel measurements for factory environments at sub-THz frequencies with co-polarized and cross-polarized antenna configurations at various antenna heights with the aid of passive reflecting surfaces, which will facilitate the sub-THz communication algorithm and system development for future intelligent factories in 6G and beyond.
Post-Silicon validation is a fundamental step in integrated circuits fabrication, and designing the setup is an essential step to the process. The setup requires a lot of design effort, especially ...under the specification and constraint of a many-core processor, which has a large state space and requires high speed and bandwidth of data communication. In this paper, a custom test station design is proposed for the Kilocore2 many-core processor. It consists of a Linux test station that controls the instruments via GPIB, and communicates with an FPGA via UART and PCIe that act as the test driver and monitor for the DUT. Tests are performed using this setup and the results are presented. This setup shall establish the foundation for future validation efforts of the Kilocore2 chip.
Sensor arrays play an instrumental role in a variety of applications, including radar, sonar, radio astronomy, and wireless communications. Employing an array of sensors permits direction-of-arrival ...(DOA) estimation, interference suppression, and imaging of spatial distributions of sources or scatterers. Linear and planar array geometries can have sensors with uniform or non-uniform spacings. Non-uniform arrays require much fewer sensors to achieve comparable performance to uniform arrays in terms of the spatial resolution and the number of resolvable sources or scatterers. This dissertation proposes novel signal processing methods for narrowband passive (receive-only) sensing and imaging. The focus is on source estimation using linear and planar passive arrays with uniform and non-uniform geometries. Algorithm development for the non-uniform arrays is facilitated by a virtual array structure, called the difference coarray, which comprises pairwise differences of physical sensor positions. The difference coarray naturally arises from the passive sensing signal model. High-resolution DOA estimation techniques, such as the subspace-based methods, are computationally expensive, especially for arrays that span large apertures. Further, performance of such methods deteriorates for coherent sources. We propose efficient and effective Fourier-based iterative techniques for DOA estimation of coherent and uncorrelated sources using linear and planar arrays with both uniform and non-uniform geometries. The considered non-uniform arrays include those with uniform and non-uniform difference coarrays. The proposed DOA estimation techniques build on the iterative interpolated beamformer, which employs an estimate-and-subtract strategy to successively extract the sources and refines the estimates via an interpolation and spectral leakage subtraction scheme. We enable iterative beamforming in the coarray domain for linear and rectangular arrays, specifically compensating for non-uniformity of difference coarrays to yield asymptotically unbiased DOA estimates. We also design the iterative interpolated beamformer for oversampled and undersampled uniform circular arrays under the manifold separation framework, which permits the application of DOA estimation techniques that were developed for uniform linear arrays to arbitrary array geometries, such as circular arrays. The proposed iterative beamforming techniques not only estimate the source DOAs, but also provide source power/amplitude estimates. As such, these Fourier-based methods are applicable to narrowband passive imaging systems for providing an accurate estimate of the distribution of source intensity or amplitude as a function of angle.
One of the major design goals of modern communication systems is to make the system green, i.e., power efficient. Since power amplifiers (PAs) occupy a major portion of the communication system’s ...power expenditure, it is critical to enhance the power efficiency of PAs. There are various approaches to optimizing PA efficiencies, like using PAs from more efficient PA classes such as the Doherty PA to achieve design requirements for lower power consumption. Unfortunately, making the PA more efficient usually comes at a cost of higher nonlinear distortion (NLD) that degrades communication receiver performance. The NLD problem has become more challenging today as communication systems continue to adopt denser quadrature amplitude modulation (QAM) constellations, like 4,096-QAM, as well as orthogonal frequency division multiplexing (OFDM) signaling techniques that are extremely sensitive to NLD. The impact NLD imposes on communication systems can be seen across multiple industries. Effective techniques are highly needed to mitigate the effects of NLD.This dissertation gives a systematic study of the PA NLD problem with a special focus on the application of artificial intelligence (AI), or machine learning (ML), to model NLD and to mitigate NLD. After reviewing existing NLD literature and practices, this dissertation showed that almost all the existing works addresses mild to moderate NLD only, not severe NLD. The review also showed the advantage of neural networks to deal with the NLD problem. As severe NLD will play a significant role in future power-efficient system design, this dissertation develops novel strategies that adopt deep neural networks to mitigate severe NLD.In Chapter 1, the application of PA and the associated NLD are introduced in cable TV communication system applications. In Chapter 2, since many existing works have shown that AI techniques consistently outperform traditional techniques in mitigating NLD, we provide a brief introduction to AI techniques and some ML models of NLD. We focus on specifically the Bayesian Network and Neural Network (NN) models.In Chapters 3 through 5, some NLD mitigation practices adopted in today’s practical systems are described. These chapters show that existing practices can mitigate mild to moderate NLD only, not severe NLD. Chapter 3 provides several conventional mathematical/analytical models of NLD. Digital pre-distortion (DPD) is discussed in Chapter 4. Post-distortion is then discussed in Chapter 5.In Chapter 6, we focus on severe NLD and develop a deep learning strategy to mitigating severe NLD at the receiver. Based on the Volterra model, deep neural networks (DNNs) are developed to learn the best NLD equalizer. A critical advantage of this strategy hinges on the capability of equalizing severe NLD, whereas conventional approaches were not competitive. We demonstrate that our proposed method outperforms traditional methods like the Volterra equalizer or neural network-based equalizer with both simulated data and real experiment data. This fact is both enlightening and motivating to further investigate the deep learning techniques for severe NLD equalization in the future.In the concluding Chapter 7, all the concepts of this dissertation, e.g., NLD, AI, Transmitter-side DPD and Receiver-side equalization, are proposed to integrate into one cohesive story that will lead to a robust NLD mitigation strategy that enables more efficient communications systems design and a better understanding of the application of deep learning.
Conventional control methods include voltage-mode and current-mode control for power supply engineers to learn and study. Hysteretic control topologies and their advantages are less well known. A ...pure hysteretic control may not be possible for some sensitive engineering applications, such as medical applications. The main practical disadvantage of hysteretic topology in its fundamental form is that switching frequency varies. This is because the switching frequency is not determined by a clock or synchronization signal. Rather, switching frequency is determined by the amount of hysteresis, system delays, and external components of the system. The immediate objective of this research project is to analyze and combine two fundamental control topologies, pulse-width modulation (PWM) and pure hysteretic control, in order to develop a hysteretic-based controller with the speed of pure hysteretic mode and fixed switching frequency.This hysteretic-based controller achieves correction speed equal to pure hysteretic control in part by transferring compensation components from the feedback loop to the power converter stage. The error amplifier was removed as well from the feedback path. As a result, a comparator would perform the functions of an error amplifier and pulse width modulator. Through a passive adder, a synthesized signal was coupled to the minus terminal of the comparator along with the feedback signal from the converter power stage in order to achieve power stage output regulation and fixed switching frequency. On the positive terminal of the comparator, the reference generator was applied. The basic hysteretic-based controller was enhanced with additional components to achieve a solution that can be used in any DC-DC converter application that requires complete control of the power converter.For verification, mathematical modelling and SPICE simulations were performed on the hysteretic-based controller. Finally, a functional prototype was built from standard off-the-shelf components. Key measurements were conducted and compared with simulation results.
This thesis investigates time-interleaved analog to digital converter front end circuits for use in next-generation fiber optic receivers capable of 200 GS/s with analog bandwidths reaching up to the ...Nyquist frequency. Measurements were conducted on an existing time-interleaved ADC front end which displayed record-breaking performance while sampling at 200 GS/s. A full four-way 200GS/s time-interleaved ADC front end employing a novel architecture was designed and fabricated in a 55nm SiGe BiCMOS process technology. The constituent track and hold amplifier was measured and achieved a record-breaking 101 GHz of track mode bandwidth. Finally, a 60GHz digitally-controlled oscillator was also designed and fabricated in 22nm FDSOI-CMOS technology with the goal of producing on-chip high-frequency clock signals to enable a full time- interleaved ADC front end system on chip.