Kapil R. Dandekar

Indoor wireless systems are becoming increasingly important in healthcare, automation, asset tracking, and next-generation communication networks. In these environments, radio frequency (RF) signals are strongly shaped by walls, furniture, human presence, device placement, antenna behavior, blockage, and multipath. These interactions create both an opportunity and a challenge: they can provide useful information for sensing, but they also make reliable communication difficult in confined indoor spaces. This dissertation addresses these two needs by developing RF sensing and evaluation methods for indoor wireless systems, with emphasis on machine-learning-based sensing, adaptive link optimization, and reconfigurable propagation control. The sensing portion of this work demonstrates how ultra-high-frequency (UHF) radio-frequency identification (RFID) technology can be extended beyond identification and tracking to support practical indoor sensing applications. Two systems are developed: an RFID-based indoor localization method for position-aware indoor environments and an RFID-based intravenous fluid-level sensor for real-time healthcare monitoring. In both cases, machine learning is used to map complex RFID measurements to useful physical quantities, enabling low-cost and battery-free sensing with simple passive tags. The evaluation portion of this dissertation develops end-to-end frameworks for reconfigurable transceivers, including pattern-reconfigurable antennas and reconfigurable intelligent surfaces (RIS), for indoor millimeter-wave systems. These frameworks span three stages: full-wave EM simulation (HFSS, CST, openEMS), ray-tracing-based channel modeling with machine-learning-driven state selection and channel-tap reduction, and hardware-in-the-loop emulation. This dissertation also develops a fully automated Python-based RIS design and simulation workflow that streamlines parameterized unit-cell modeling, EM characterization, and integration with site-specific channel modeling, enabling repeated full-wave simulations, geometry updates, and far-field radiation-pattern extraction for distinct RIS on-off configurations. The emulation platform combines a wireless channel emulator, software-defined radios, and DragonRadio to support packet-level wireless experiments under repeatable channel conditions. This simulation-to-emulation pipeline enables realistic and repeatable evaluation of next-generation wireless systems under site-specific indoor propagation conditions.

The convergence of advancements in antenna technology with Machine Learning (ML) is envisioned to enhance coverage and capacity for wireless communication systems in complex and dynamic millimeter-wave (mmWave) indoor environments. These environments often experience significant performance variability due to user movement and obstacles. Our study highlights the potential benefits of combining reconfigurable antenna (RA) systems with ML to address mmWave propagation challenges in indoor environments. However, rigorous verification and validation are essential to ensure accurate modeling of mmWave propagation, which is inherently complex and challenging to evaluate experimentally. To circumvent costly, time-intensive, and non-repeatable real-world measurements, we introduce a hardware emulation framework. It enables realistic evaluation of non-stationary, ray-traced channel models with a large number of propagation paths. This framework integrates realistic channel coefficients from site-specific 3D raytracing scenarios with RA-equipped access points (APs) and user mobility features. It incorporates them into a softwaredefined radio (SDR)-based full-mesh wireless channel emulation system, enabling the coexistence of virtual and real nodes. We present experimental results from transceiver hardware-in-theloop testing in this testbed. These results feature repeatable and controllable path loss and delays between communicating nodes. Experimental evaluations confirmed that intelligent state selection algorithms, particularly Thompson Sampling and UCB1-Tuned, significantly enhance system performance in terms of throughput and packet error rate, outperforming traditional omni-directional antenna configurations.

Learning the underlying dynamics of periodic and aperiodic systems evolving over space and time provides substantial challenges across scientific fields. Despite recent advances in machine learning (ML), our understanding of its limits to learning, identifying, and emulating real physical systems remains incomplete. With the increasing complexity of dynamical systems and higher power consumption, ML systems take inspiration from the complex learning mechanisms of the mammalian brain, for biologically plausible learning. As we delve into the designing of brain- inspired ML systems, we deal with the challenges of the learning procedures, network sizes, energy efficiency and training complexities involved with it. It's crucial to characterize their performance in learning and controlling unknown dynamical systems to address them. We address these issues by designing sparsity aware learning mechanisms in feedback driven recurrent neural networks (RNNs), incorporate biologically plausible neuron models, and integrate computational neuroscience to build data driven AI systems. We explore brain-inspired neuronal dynamics to build AI systems that can mimic and classify dynamical systems. We propose full-FORCE (FF) SNN, Reservoir-in-reservoir (R-i-R) architecture, and SNN based applications learning, We pro- pose sparsity aware learning in RNNs, full-FORCE based sparsity regularization in recurrent spiking neural networks (RSNN) and an RSNN based BLE application for dynamical systems learning. Our proposed learning methods and architectures exhibit strong performance, noise robustness and energy efficiency in simulated and real-world adaptive system emulation. We evaluate bio signals, motion trajectories, Lorenz systems, real-world dynamic variable datasets, and other dynamical systems in this dissertation.

A number of control network applications necessitate data transmission in environments where metallic structures radio frequency network connectivity. In a majority of these applications, it is also highly undesirable to physically penetrate the impeding structure to enable communication through wires or cabling. One such example regards the deployment of wireless sensing and control networks on U.S. Navy vessels for vital automated ship operation maintenance. The ship hull obstructs electromagnetic wave propa- gaLion and limits network connectivity. However, mechanically penetratmg the bulkhead compromises the structural integrity of the ship compartments and the requirement for the vessel to remain watertight. In such applications, ultrasonic signaling has been proposed to achieve reliable data communication without penetration of the metallic structure. The reverberant nature of the acoustic channel generated between the interface of the transducers and metal wall becomes a network throughput bottleneck, bounding narrowband communication techniques to maximum data rates of roughly 5 Mbps in practice. Orthogonal Frequency-Division Multiplexing (OFDM) has smce been shown as a promising technique to mitigate the fre- qucncy selectivity of the ultrasonic channel and increase achievable data rates without the need for complex equalizers. However, OFDM is prone to high peak-to-avemge power ratios (PAPR) in the transmitted signal which can lead to decreased system performance. This thesis presents a link adaptive OFDM-based ultrasonic physical layer created to enhance through-metal communication by exploiting the slow-varying nature of the ultrasonic channel and employing a combined rate adaptive and PAPR reduction algorithm. Simulated and measured results obtained using this novel Joint Adaptive Algorithm suggest that an increase in throughput rates, reduction in PAPR, and adherence to a desired quality of service constraint can be simultaneously achieved in the ultrasonic channel. These enhancements provide throughput and reliability to support a variety of high-data-rate network applications in environments where metallic structures would otherwise compromise data transmission.

Multi-carrier systems have become the core of many communication systems. For the last 50 years, researchers have constantly been refining such solutions to convey large amounts of information. In this thesis, two problems are addressed; i) high peak to average power ratio (PAPR) in multi-carrier systems and ii) throughput improvement in multi-carrier systems by PAPR-aware rate adaptive bit loading. The first issue is one of the biggest drawbacks of Orthogonal Frequency Division Multiplexing (OFDM) systems and the second aims to improve the system data rate in scenarios where the communication system adapts to environmental channel conditions. Towards this goal, two symbol rotation-inversion algorithms that reduce the peak to average power ratio in multi carrier OFDM systems are implemented on a software defined radio platform. One is evaluated in a single input single output (SISO) OFDM system and the second in a multiple input multiple output (MIMO) OFDM communication system. Results from both algorithms show the benefit of PAPR reduction and motivate the proposed novel solution. A new scheme that combines the benefits of bit allocation and symbol rotation to reduce the PAPR is presented, simulated and evaluated. It is demonstrated in simulations by means of post processing signal to noise ratio (PPSNR), throughput and bit error rates, how this new framework succeeds in achieving higher data rates in a more robust way. PAPR remediation in OFDM communication systems can dramatically improve PPSNR and thus improve system range and robustness to noise. When coupled with adaptive bit loading techniques, these PAPR remediation strategies can substantially increase link throughput. For a fixed number of permutations, measurements showed PPSNR improvements of 2dB in SISO OFDM and almost 1dB in MIMO OFDM. Simulations showed that symbol rotation results in more than one order of magnitude BER reduction for SISO OFDM and one order of magnitude reduction in MIMO OFDM. These techniques thus have the potential to significantly improve future multi-carrier communication systems.

This paper proposes a heterogeneous carrier-sense multiple access (CSMA) protocol named LoHEC as the first research attempt to improve energy fairness when applying CSMA to long-range wide area network (LoRaWAN). LoHEC is enabled by Channel Activity Detection (CAD), a recently introduced carrier-sensing technique to detect LoRaWAN signals even below the noise floor. The design of LoHEC is inspired by the fact that existing CAD-based CSMA proposals are in a homogeneous manner. In other words, they require LoRaWAN end devices to perform identical CAD regardless of the differences of their used network parameter – spreading factor (SF). This causes energy consumption imbalance among end devices since the consumed energy during CAD is significantly affected by SF. By considering the heterogeneity of LoRaWAN in terms of SF, LoHEC requires end devices to perform different numbers of CAD operations with different CAD intervals during channel access. Particularly, the number of needed CADs and CAD interval are determined based on the CAD energy consumption under different SFs. We conduct extensive experiments regarding LoHEC with a practical LoRaWAN testbed including 60 commercial off-the-shelf end devices. Experimental results show that in comparison with the existing solutions, LoHEC can achieve up to 0.85 $\times$ improvement of the energy fairness on average.

Ultrahigh frequency (UHF) passive radio frequency identification (RFID) tag-based sensors are proposed for intravenous (IV) fluid level monitoring in medical Internet of Things (IoT) applications. Two versions of the sensor are proposed: a binary sensor (i.e., full versus empty state sensing) and a real-time (i.e., continuous level) sensor. The operating principle is demonstrated using full-wave electromagnetic simulation at 910 MHz and validated with experimental results. Generalized Additive Model (GAM) and random forest algorithms are employed for each interrogation dataset. Real-time sensing is accomplished with small deviations across the models. A minimum of 72% and a maximum of 97% of cases are within a 20% error for the GAM model and 62% to 98% for the random forest model. The proposed sensor is battery-free, lightweight, low-cost, and highly reliable. The read range of the proposed sensor is 4.6 m.