Implantable medical devices, especially brain implants, show great promise for applications such as epilepsy detection, Parkinson’s disease regulation, and brain–computer interfaces. However, battery-based power supplies suffer from degradation, surgical replacement risks, and limited compatibility with miniaturization, motivating wireless power transfer (WPT) as a promising alternative. Among various approaches, ultrasonic WPT is particularly attractive due to its strong penetration capability, no electromagnetic interference, and favorable safety profile. This thesis presents the design and implementation of a high-efficiency receiver (RX) circuit for ultrasonic WPT applications, with a focus on biomedical implants. Several rectifier topologies are investigated. Although the current-mode rectifier (CMR) demonstrates promising characteristics, its performance is significantly limited in ultrasonic WPT due to the low quality factor of the transducer. Consequently, the cross-coupled rectifier is ultimately selected for its simplicity and superior conversion efficiency. A boost converter is integrated to further increase the rectified voltage, with its duty cycle serving as the sole tunable parameter for impedance matching. Based on the maximum power transfer theorem, an adapted perturb and observe (P\&O) maximum power point tracking (MPPT) algorithm is developed to dynamically adjust the duty cycle, thereby achieving automatic impedance matching and maximizing the power delivered to the load. The start-up issue is resolved through the implementation of a maximum-voltage selector, ensuring stable operation from system initialization. Simulation results demonstrate a peak overall system efficiency of 83.55\%, with the cross-coupled rectifier and boost converter achieving efficiencies of 91.55\% and 91.26\%, respectively. The proposed system maintains high efficiency under varying load and input conditions, validating its feasibility for reliable and efficient power delivery in ultrasonic WPT systems for brain implants. ... Read more

Aphasia is a language disorder caused by brain damage. It often behaves in the form of disfluent speech, fragmented grammar, and irregular word usage. These characteristics make it difficult for current automatic speech recognition (ASR) systems to produce accurate transcriptions. While models like Whisper perform well on fluent speech, they often struggle with aphasic speech, especially in low-resource clinical settings. This thesis presents AS-ASR, an aphasia-specific ASR system built on the lightweight Whisper-tiny model and optimized for real-time use on edge devices. To improve transcription accuracy, we created a mixed dataset of fluent and aphasic speech and used the GPT-4 large language model (LLM) to clean and refine transcripts from disfluent recordings. We tested different ratios of aphasic to typical data during training to find a balance that supports both accuracy and generalization. In addition, we applied mixed-precision quantization and a simple energy-based voice activity detection method to reduce model size and inference time. The fine-tuned model achieved up to 40% lower word error rates on aphasic speech and also kept accurate performance on fluent speech. Our findings suggest that targeted fine-tuning and lightweight deployment strategies can make ASR systems more accessible and effective for clinical and assistive use. ... Read more

The globe is being challenged by growing population and accelerated urbanization, which demands sustainable usage and sharing of resources to create more livable and smarter environments at all scales. To answer those demands, the Internet of Things (IoT) is ultimately a powerful enabler to share resources on a large scale. IoT requires pervasive sensing, which demands a large number of distributed nodes with integrated circuits to collect information. Therefore, Photovoltaic Energy Harvesters (PVEH) have been widely chosen to power IoT nodes, which minimize the maintenance costs. PVEHs are usually managed by Power Management Integrated Circuits (PMIC), also referred to as Photovoltaic cell based PMIC (PV-PMIC). One of the most important feature of PV-PMIC is Maximum Power Point Tracking (MPPT), which ensures the system can output power at Maximum Power Point (MPP). Among various MPPT algorithms, Perturb and Observe (P\&O) is the most popular, which requires measurements of a combination of current and voltage of different nodes or branches, which are analog signals and need to be converted into digital signals. To achieve this analog-to-digital conversion, many ADC and TDC architectures have been proposed, but they are all challenged by power limitation and mismatch and Process, Voltage, and Temperature (PVT) variation. This thesis aims to present the design procedure of a new type of ADC for MPPT structure in a PV-PMIC system. To achieve the objectives, a literature research is carried out, with the research question being 'Which Analog-to-Digital Conversion techniques are suitable for MPPT in a low-power, small-area chip?'. After the literature research, a set of requirements is established and an alternative ADC design is proposed and tested. The proposed design combines capacitive current sampling and a Non-Linearity Canceling TDC. The proposed design is further verified at different levels, including the system level and the circuit level. The final verification includes three tests, namely the Linear Time Response, the Static Response, and the Frequency Response. The obtained results show that the proposed system at the circuit level achieves ENOB=3.97, FoM=4.17 pJ/conv-step, DNL bounded to +-0.8 LSB, and INL bounded to 2.5 LSB. It is further concluded that the proposed system generally meets the requirements. Moreover, the proposed system has unique advantages compared to conventional ADCs for the dedicated MPPT application. Possible future improvements include further improving linearities, designing calibration and trimming methods, and replacing the hard-coded logic with a storage component.

... Read more

Integrated temperature sensors are widely used in various applications, including thermal monitoring, medical devices, and voltage/frequency references. An emerging class of temperature sensors operates by measuring the thermal diffusivity (TD) of silicon. This can be done by measuring the time it takes for heat to diffuse between a heater and a relative temperature sensor, both implemented in a silicon substrate. The key advantage of such TD-based sensors is that their accuracy improves linearly with lithographic precision.

Increasing the length of the heat path and/or moving to a more advanced technology node will enhance measurement accuracy. However, since silicon is an excellent heat conductor, the output of such sensors is at the millivolt level, which makes the design of accurate readout circuitry quite challenging.

In this work, an accurate readout circuit was designed for TD-based sensors realized in a 65nm process. It consists of a low-noise pre-amplifier, followed by a Sigma-Delta ADC. Based on simulation results, the designed circuit achieves a 23 mK (3σ) inaccuracy at 27◦C, which is negligible compared to the expected inaccuracy of the TD-based sensors. ... Read more

As a crucial component, one common challenge of power amplifiers (PA) is the nonlinearity in the wireless communication system. Digital predistortion (DPD) is essential for mitigating nonlinearity in radio frequency (RF) power amplifiers, particularly for wideband applications. This thesis work aims to present the TCN-DPD model, a novel parameter-efficient architecture based on temporal convolutional networks (TCNs) to enhance the performance of PA. Therefore, the main problem should focused on the TCN-DPD model implementation with parameter efficiency. When TCN architecture was designed with several noncausal and dilated depthwise convolution layers and 1*1 convolution layers, optimized activation functions should be explored to complete the TCN architecture. By evaluating on the OpenDPD framework with the DPA_200MHz dataset, Hardswish, Tanh, SiLU, and GELU were considered by benchmarking different activations of TCN-DPD based on SIM- NMSE and SIM-ACLR metrics on average. Hardswish was confirmed in the later experiments as the optimized activation function in TCN architecture based on simulated results ACLR of -51.54 dBc and
NMSE of -44.61 dB. Since the TCN-DPD architecture was completed, this proposed model’s performance in PA and DPD benchmarks is desirable to be tested, and later experiments will use the same dataset and framework as the benchmark of activation function did. In PA benchmarking, the TCN model achieves SIM-NMSE -34.99 dB on average compared to other models, LSTM, GRU, RVTDCNN, VDLSTM, PNTDNN, and DGRU. This achievement shows the TCN architecture has a high potential to handle a range of dependencies efficiently in the DPD application system. Furthermore, DPD benchmarking is the main experiment in this thesis. Two architectures were selected as the pre-trained PA model: the DGRU and TCN models. When the pre-trained PA is fixed as the DGRU model, TCN-DPD demonstrates superior linearization performance with only 500 real-valued parameters, achieving averaged and simulated ACPRs of -51.58/-49.26 dBc (L/R), EVM of -47.52 dB, and NMSE of -44.61dB. The results are simulated ACPRs of -50.39/-50.01 dBc (L/R), EVM of -47.88 dB, and NMSE of -45.51 dB in average when the pre-trained PA model is TCN. Both DPD benchmarks include different DNN-DPD models, and TCN-DPD has superior performance in the comparison, especially the SIM-NMSE and SIM-EVM performance is significantly higher than other models when the pre-trained model is TCN. These results establish TCN-DPD as a promising solution for efficient wideband PA linearization. Moreover, the evaluation extended to DNN-DPD performance with various numbers of parameters ranging from 200 to 1000, where the TCN-200 model highlighted its effectiveness by showing impressive results in SIM-NMSE -41.27dB/-43.51dB(DGRU/TCN PA), achieving superior linearization performance while using significantly fewer parameters than existing deep neural network solutions, proving the TCN-DPD model’s parameters efficiency.
The research in this thesis conclusively demonstrates that TCNs can be implemented in DPD applications, providing more parameters efficiency, better performance, and robust PA linearization solutions, potentially setting a new alternative in DPD technology. ... Read more

Keyword spotting (KWS) is an essential component of voice recognition services on smart devices. Its always-on characteristic requires high accuracy and real-time response. Also, low power consumption is another key demand for KWS devices. In previous research, neural networks have become popular for KWS tasks for their accuracy compared to traditional machine learning technologies. Among classical neural networks like recurrent neural networks (RNNs) and convolutional neural networks (CNNs), temporal convolutional networks (TCNs) have begun to catch attention recently. Moreover, studies related to sparsity are always an efficient method to deal with the growing model size issue for modern neural network designs. As a potential solution, in this work, a TCN model is trained for KWS on the Google Speech Command V2 dataset and achieves an accuracy of 94.1\%. Based on that, two different sparsity are applied to the TCN model. One is temporal sparsity. By creating a Delta convolution layer, the Delta temporal convolutional network (DeltaTCN) achieves an accuracy of 93.6\% with a 72\% reduction in floating-point operations (FLOPS) compared to the original TCN model. Another is structural weight sparsity. By creating sparsity on the weight matrix of each convolution layer, the structural sparse temporal convolutional network (SSPTCN) achieves 93.6\% accuracy with a 70\% reduction in FLOPs and a 39\% reduction in parameters. ... Read more

Bidirectional neural recording ICs faces the challenge of simultaneous stimulation and recording. The recording IC should endure large stimulation artifacts while capturing weak neural signals excited by the stimuli. The stimulation artifacts can be as large as hundreds of millivolts, which can saturate the recording front end. Conventional neural-recording ICs use a low-noise-amplifier (LNA), a programmable-gain amplifier (PGA) followed by an analog-to-digital converter (ADC), leading to low power consumptions and great noise performances. However, the dynamic range (DR) is usually limited to 50 dB. State-of-the-art recording ICs using direct-conversion ADCs have been intro- duced to increase the DR. However, the typical power and area consumption for these architectures are exceeding the requirements for next-generation brain-computer interfaces. This work proposes a novel neural-recording IC system architecture with satu- ration prevention in presence of large stimulation artifacts. The proposed architecture consists of an AC-coupled boxcar sampler, switched-capacitor low-pass filter followed by a 10-bit asynchronous SAR ADC. The integrated voltage at the output of the boxcar sampler is sampled by the ADC and monitored by a level-cross detection block. Based on the output of the level-cross detection block, the integration time, and thus the gain, can be tuned to different configurations pre-defined in a look-up table (LUT). The additional noise penalty due to noise-folding from decreasing the integration time is compensated by oversampling and averaging. The proposed system architecture is partially imple- mented at transistor level while the ADC and digital blocks are modeled using verilog-A as a proof-of-concept. The analog front-end achieves a DR of 69.5 dB with a peak-to-peak maximum input amplitude of 180 mVPP and a typical ENOB of 8.01 bits. ... Read more

This work proposes an energy harvesting DC-DC converter that is able to efficiently process a wide input power range from 25 mW up to 250 mW, to charge a storage device used in Internet-of-Things (IoT) applications. An interleaved boost-converter topology with two inductors is used to divide the high currents over the two branches, in order to reduce the conduction losses in the MOSFET power switches at high input power.

A Maximum Power Point Tracking (MPPT) strategy is employed to find the optimal matching between the harvesting sensor, which in this case is a solar cell, and the input of the converter. The Maximum Power Point (MPP) is found by measuring the output current of the converter, decreasing the duty cycle of the switching power MOSFETs, and again measuring the output current to check if it increases. This process is repeated until the output current no longer increases, which is where the MPP will be. The duty-cycle pulses that control the power switches are generated by comparing a reference voltage representing the duty cycle with two sawtooth waves with a phase difference of 180 degrees generated on-chip.

Usually, for converters that handle large currents, large power switches are required to keep the on-resistance of the switches minimal such that the conduction losses do not dominate. Since the converter also needs to be able to efficiently convert smaller powers, the size of the power switches is made configurable. This is necessary since otherwise the gate-charge losses will become dominant. The total size of the segmented power switches are controlled digitally by a logic unit that calculates the expected input current using 2-bit representations of the measured output current and the two most significant bits of the 5-bit representation for the duty cycle set by the MPPT.

Schematic simulation results show a conversion efficiency of up to 92%. The proposed system is designed and simulated using 180 nm TSMC CMOS technology. The chip covers a total silicon area of 2.20 mm2, with an active area of 1.12 mm2. ... Read more

Time-of-Flight (ToF) is an optical range detection technique that derives range from the round-trip time of light pulse between the imager and the target. Due to the simple hardware and distance extraction algorithm, ToF imagers are widely adopted in existing (automotive, consumer electronics, etc.) and emerging (AR/VR, etc.) applications requiring depth information. ToF is divided into Direct-ToF (D-ToF) and Indirect-ToF (I-ToF). Compared with the D-ToF system which directly measures the round-trip time of the laser beam, the I-ToF system indirectly measures the time from a phase measurement of the received modulated pulse. I-ToF provides a smaller detection range but it is less sensitive to jitter in the system and can achieve higher accuracy. Hence, I-ToF is used in short-range high precision applications such as face recognition.

This thesis describes the readout ADC design for the next-generation I-ToF imagers used for smartphone face recognition in collaboration with Infineon. A compact 11-bit 2 MS/s column-level ADC is developed in 65nm CMOS image technology. To balance between the available area and required speed, a hybrid SAR-RAMP ADC concept is proposed, which also embeds a threshold comparison phase to relax requirements on successive phases. The ADC has been verified with the post-layout extracted view of the column-level analog circuits and the schematic view of the central/digital circuits with an ideal ramp buffer, achieving a 1.1 V - 3.3 V input range with a 1.1 V reference, a signal-to-noise and distortion ratio (SNDR) of 64.3 dB, 45 µW power consumption, a Walden figure of merit (FOM$_W$) of 107 fJ/conv-step and an area of 7 µm $\times$ 815 µm. ... Read more

This work proposes an energy harvesting platform that is able to convert power from both DC sources (photo-voltaic cells and TEGs) as well as piezo element sources. It does so only using a single input channel to which a single harvester can be connected. The proposed system is able to differentiate between the two source types and adjust the power converter configuration accordingly.For the DC sources, a novel switched-capacitor power converter (SCPC) is proposed, that is able to convert the energy from a harvester that has a maximum power point (MPP) output voltage of 170mV to 5V and a maximum power point output power of 10uW to 50mW. This DC-DC converter offers 119 different positive voltage conversion ratios, with a maximum voltage conversion ratio of 16, using four in-package capacitors. As a result of this high number of conversion ratios, the MPP output voltage of the harvester and the input voltage of the power converter are matched accurately, causing the harvesting efficiency to be very high. A maximum harvesting efficiency of 96.2% is found in simulations. For the piezo element sources, the concept of a flipping-capacitor rectifier (FCR) has been adjusted to work in harmony with the designed SCPC. In a steady-state condition, the capacitors of the SCPC reach specific voltages, such that they can create evenly spaced voltage steps for the flipping operation. With this technique, a voltage flipping efficiency of 0.9375 and a theoretical maximum output power improvement rate (MOPIR) of 32 can be reached. Due to losses in the system, simulation results show a MOPIR of up to 20.0, which is still significantly higher than the state-of-art. The system is designed to work with harvesters with a piezo capacitance of up to 100nF, an excitation frequency of 1Hz to 200Hz and an equivalent FBR maximum power point output power of 1uW to 50mW. An implementation of the proposed system is discussed and simulated. The total active silicon area for the designed system is 2.12mm2 in a 0.18 um TSMC technology.

... Read more