Parameter quantization for Large Language Models (LLMs) has gained significant attention for reducing memory costs and improving computational efficiency. However, existing methods struggle with performance degradation in low-bit scenarios. In this paper, we propose Outlier-Oriented Quantization (OOQ), a novel framework designed to address these challenges through three key innovations. First, we design an outlieroriented metric to determine quantization precision based on outlier percentages in channel parameters. Second, we dynamically allocate varying quantization precision to different parts of the model according to the outlier distribution. Finally, guided by the outlier-oriented metric, we preserve some high-precision outliers during the quantization process. Experiments on LLaMA models demonstrate that OOQ achieves state-of-the-art results across various bit settings, particularly in extremely low-bit regimes. Additionally, OOQ improves inference speed by up to 24%.

With the development of the Internet of Things (IoT), wireless pushbuttons are being increasingly used to control devices such as lights, fans, and air conditioners in smart homes and offices. In contrast to conventional systems, self-powered pushbuttons eliminate the inconvenience of cable arrangement and the cost of battery replacement. However, existing self-powered wireless pushbuttons can only send commands unidirectionally and lack an instant feedback mechanism. This limitation reduces system reliability and impairs the user experience. To tackle this issue, we propose ViPSN-button, a motion-powered wireless pushbutton that offers instant feedback. When the ViPSN-button is pressed, it can deliver instant feedback to the user. This feedback mechanism enables the user to confirm whether the host unit has successfully acknowledged their command. ViPSN-button is powered by a quasi-static-toggling energy harvester (QST harvester). The entire communication process of the ViPSN-button includes three steps: sending commands, receiving acknowledgments (ACK), and displaying an indication. All of these actions are carried out solely by utilizing the energy generated from a single press action. Experiments demonstrate that the ViPSN-button can achieve low-power, fast, private, and reliable bidirectional communication under the Enhanced ShockBurst (ESB) communication protocol. Field tests have been conducted, showing that the ViPSN-button can achieve reliable communication at a distance of 50 meters. ViPSN-button offers an innovative design concept for self-powered sensing nodes, facilitating bidirectional communication between host units and sensing nodes. This feature renders it highly suitable for a wide range of applications, including smart homes, smart offices, smart cities, and industrial IoT.

As an energy storage technology, supercapacitors feature a high power density. In particular, supercapacitors can be charged or discharged by a relatively large pulse current for a limited period of time. This paper takes advantage of this characteristic and develops a dual-mode fast charger for supercapacitors that employs both continuous and fine-tuned pulse currents. Based on the conventional forward converter, the proposed charger introduces an energy storage capacitor and a branch resistor to tune the rising and falling edges of the pulse current, respectively. Consequently, the transitions between the continuous and pulse current modes are significantly accelerated, which ultimately shortens the supercapacitor charging time. A prototype is built and tested using a 3 V/6 F supercapacitor. To charge the supercapacitor from 2 to 2.5 V, the proposed charger takes 0.87 and 1 s when it is configured to operate in the dual-mode and the continuous current mode only, respectively, which leads to a 13% reduction in the charging time. Moreover, compared to the forward converter, the pulse characteristics of the proposed charger are significantly improved in that the pulse rising and falling times are dramatically reduced: 2.1 versus$147~\mu $s and 7.2 versus$103~\mu $s, respectively.

Existing methods for predicting lithium-ion battery remaining useful lifetime (RUL) rely on complete capacity degradation data or extensive historical profiles. However, such sufficient conditions are usually unavailable in practical battery usage. To cope with this issue, a framework for RUL estimation with fragment data is proposed. The framework utilizes a small amount of prior knowledge as benchmark data to create an empirical model-based predictive method for estimating RUL by fragment historical data during nonlinear degradation stage. A more specified parameter initialization is obtained by trend identification of the fragment. Particle filter (PF) algorithm is utilized for model parameter update with proposed improved resampling strategy. RUL predictions using two different datasets demonstrate the effectiveness of the proposed method. An error margin of less than ten cycles in RUL predictions is consistently achieved in CS2 dataset when employing fragments ranging from 50 to 60 cycles. And an error margin of around 20 cycles is achieved in CX2 dataset by fragments ranging from 60 to 80 cycles. The proposed method renders a more precise and stable predictive result with high confident level.

Energy harvesting (EH) provides a promising solution for powering distributed Internet of Things (IoT) devices. Due to the low-level and sporadic ambient energy supply, an EH-powered device should operate in an intermittent and energy-driven mode. Commercial voltage supervisors were not optimized for the EH scenario, making it difficult to satisfy all new demands. The conventional energy management (EM) circuit has a risk of locking up during the turn-ON transient; therefore, it might fail to power the IoT load device. Previous technologies have used a relatively large circuit to solve this problem. In this article, a concise discrete three-transistor EM (3T-EM) circuit is proposed. It can track stored energy, switch ON/OFF to the load device, and provide a regulated voltage output. These key functions are realized by utilizing a minimum number of components; therefore, power consumption and manufacturing cost are largely cut. The voltage thresholds and minimum input current are theoretically derived. In experiments, the ON/OFF thresholds can be adjusted accurately, as predicted by the theory. The 3T-EM circuit can ensure the correct operation when the input current is as low as$0.4~ \mu \text{A}$. Control experiments also prove the effectiveness and performance of the 3T-EM circuit. The proposed 3T-EM circuit shows the characteristics of low cost, low power, inherent regulation, high voltage rating, and good predictability. It is a good candidate to perform the EM task in widely distributed EH-powered IoT devices.

Modern speaker verification (SV) systems typically demand expensive storage and computing resources, thereby hindering their deployment on mobile devices. In this paper, we explore adaptive neural network quantization for lightweight speaker verification. Firstly, we propose a novel adaptive uniform precision quantization method which enables the dynamic generation of quantization centroids customized for each network layer based on k-means clustering. By applying it to the pre-trained SV systems, we obtain a series of quantized variants with different bit widths. To enhance low-bit quantized models, a mixed precision quantization algorithm along with a multi-stage fine-tuning (MSFT) strategy is further introduced. This approach assigns varying bit widths to different network layers. When bit combinations are determined, MSFT progressively quantizes and fine-tunes the network in a specific order. Finally, we design two distinct binary quantization schemes to mitigate performance degradation of 1-bit quantized models: the static and adaptive quantizers. Experiments on VoxCeleb demonstrate that lossless 4-bit uniform precision quantization is achieved on both ResNets and DF-ResNets, yielding a promising compression ratio of$\sim$8. Moreover, compared to uniform precision approach, mixed precision quantization not only obtains additional performance improvements with a similar model size but also offers the flexibility to generate bit combination for any desirable model size. In addition, our suggested 1-bit quantization schemes remarkably boost the performance of binarized models. Finally, a thorough comparison with existing lightweight SV systems reveals that our proposed models outperform all previous methods by a large margin across various model size ranges.

Synchronous switch (SS) technique has been extensively studied in piezoelectric energy harvesting (PEH). The SS circuits can significantly enhance the output power under the same vibration excitation. Some SS solutions have also been developed for an inductive electromagnetic (EM) source by referring to its capacitive PEH counterpart and taking a reciprocal design. This paper proposes a synchronized current inversion and energy extraction (SCIEE) circuit for EM energy harvesting (EMEH). SCIEE utilizes two switched capacitive branches to carry out the synchronized current inversion at the electromotive voltage negative-to-positive zero-crossing instants and energy extraction at the voltage positive-to-negative zero-crossing instants. By inverting the transducer current, SCIEE increases the torque/force inside the transducer to extract more energy from the relative movement between magnets and coils. Theoretical analysis shows that the proposed circuit is suitable for use with an EM transducer, whose quality factor is relatively large. Experiments compared the output power of three harvesting schemes: SCIEE, synchronized switch energy extraction (SSEE), and conventional pulse-width modulation (PWM)-based harvesting scheme. When using the same prototyped EM harvester under the same mechanical excitation, SCIEE can harvest 38% more power, compared with the cutting-edge SSEE circuit for EMEH; and 900% more power, compared with the PWM-based harvesting scheme.

Due to the compatibility considerations, it is not attractive for the commercial wireless charger to modify the Qi-standard coils for charging multiple loads. This paper would explore the potential of a power relay module (PX) to address this issue. The multiple-coil PX would enhance the effective coupling when the standard coupler fails. In this paper, different types of PXs are developed for various applications, including a two-coil PX using series compensation, a two-coil PX using high-order compensation, and a three-coil PX using high-order compensation. Their power and efficiency characteristics are analyzed in a uniform manner, and the benefits of different PXs are justified through a planar charger and a bowl-shape charger in the experiment. The implemented bowl-shape charger is able to offer one fast-charging channel for a single device (30 W) and one multiple-load channel for at most four devices (each 10 W) simultaneously. The peak efficiency is 88% when the overall delivered power of PX is 70W.

As the most widely used technique, deep speaker embedding learning has become predominant in speaker verification task recently. Very large neural networks such as ECAPA-TDNN and ResNet can achieve the state-of-the-art performance. However, large models are computationally unfriendly in general, which require massive storage and computation resources. Model compression has been a hot research topic. Parameter quantization usually results in significant performance degradation. Knowledge distillation demands a pretrained complex teacher model. In this paper, we introduce a novel self-knowledge distillation method, namely Self-Knowledge Distillation via Feature Enhancement (SKDFE). It utilizes an auxiliary self-teacher network to distill its own refined knowledge without the need of a pretrained teacher network. Additionally, we apply the self-knowledge distillation at two different levels: label level and feature level. Experiments on Voxceleb dataset show that our proposed self-knowledge distillation method can make small models have comparable or even better performance than large ones. Large models can also be further improved when applying our method.

To improve the efficiency of hybrid energy storage systems composed of fuel cells and supercapacitors used in high-power applications such as electrified transportation systems and renewable energy systems, the interfacing power converters need to be carefully designed. This paper proposes an active clamping current-fed three port converter for an application scenario in which modular converters are required to aggregate distributed energy storage resources. The converter topology is conceived to implement four operation modes: single input dual output, dual input single output, and two scenarios of single input single output. These operation modes are combinations of three types of power flows: fuel cell to load, fuel cell to supercapacitor, and supercapacitor to load. Simulation results verify the functionality of the proposed converter.

In this article, we introduce a vibration-powered sensing node (ViPSN), a programmable Internet-of-Things (IoT) platform for the development of vibration-powered or motion-powered sensing and transmitting systems. It leverages the exploitation and utilization of ambient vibration energy by using a piezoelectric transducer. The roles and relations of six necessary modules, including energy generation unit (EGU), energy transduction unit (ETU), energy enhancement unit (EEU), energy management unit (EMU), energy user unit (EUU), and edge demonstration unit (EDU) are discussed in detail. In particular, an enhanced EMU is proposed by making necessary complements to an extensively used off-the-shelf integrated circuit (IC) solution for piezoelectric transducers. It provides more comprehensive energy storage indicating signals, such that the sensing, computing, and transmitting tasks can be carried out more robustly by keeping a good awareness of the remaining energy. Owing to the enhanced EMU design, vibration energy in various forms, such as intermittent and transient ones, can be more effectively harvested and utilized. The performance of ViPSN is evaluated, in terms of its lifetime and Quality of Service (QoS), under different vibration scenarios. The inclusive design and affiliated opensource project of ViPSN help build a new ecosystem for the research and development of vibration- or motion-powered IoT systems.

In battery/ultracapacitor (UC) hybrid energy storage systems (HESS), sizing and energy management strategies are crucial, which determine the system cost and performance. However, research on these two problems in a coupled manner for plug-in electric vehicles is still immature. This article aims at resolving this issue in the perspective of minimizing the average operating cost. Both manufacturing cost and system end-of-life timing are incorporated. A quantitative battery degradation model is employed to evaluate the battery dynamic capacity loss and cycle life. Dynamic programming algorithm is then deployed to achieve optimal power distribution between battery and UC. Furthermore, the power management and HESS optimal sizing strategies are unified into a single cost-minimization problem. Combining those efforts, the optimal size of the HESS with minimized average operating cost is solved by simulated annealing method. Optimization results illustrate that a minimum cost of 15.52 USD is achieved with 72 UC cells and 7100 battery cells. A large set of simulation data has proved the optimality of the optimization results. Compared with the battery-only solution, the proposed solution demonstrates 11.9% cost reduction and 21.7% battery cycle life extension under the Urban Dynamometer Driving Schedule. Moreover, the temperature rise of the battery is reduced by 31.1%. Finally, based on the optimal results, the energy management strategy is extended to fit real-time applications by utilizing Markov chain and stochastic dynamic programming.

This paper introduces the series synchronized triple bias-flip (S-S3BF) interface circuit for piezoelectric energy harvesting (PEH) enhancement. The S-S3BF topology is derived from its parallel counterpart (P-S3BF), which was proved to offer the maximum energy harvesting capability when three bias-flip actions are implemented. The S-S3BF is a little bit less capable than P-S3BF. However, it eliminates the circuit components by removing the bridge rectifier connected in parallel, such that can use only one capacitor to realize the bias voltages and energy storage simultaneously. Compared to the other single capacitor designs, the S-S3BF makes the best energy harvesting capability so far. Moreover, S-S3BF can automatically shift among single, double, and triple bias-flip operations under heavy, medium, and light load conditions, respectively. Therefore, it is more adaptive than previous designs. Theoretical and experimental results show that the harvested power can always follow the envelope (best case) of the single, double, and triple bias-flip operations during an entire charging process.

In wide output applications, the switching frequency of conventional LLC converter needs to swing in a wide range, which jeopardizes the conversion efficiency. Moreover, the high output voltage brings high voltage stresses to the secondary side diodes. To cope with those issues, a novel LLC topology with reconfigurable voltage multiplier rectifier is proposed. This converter is suitable for both high output voltage and wide output range applications. Its rectifier may switch between voltage quadrupler and sixfolder configurations, depending on the desired output voltage. Benefits include narrow switching frequency range, low circulating current, and reduced conduction loss. Zero-voltage-switching (ZVS) and zero-current-switching (ZCS) are realized among all power MOSFETs and all power diodes, respectively. The voltage stresses on secondary side diodes are constrained. Detailed circuit operation principles and modeling method are presented. A 1.3 kW converter prototype, generating 500 V-840 V output from 390 V dc link is designed and tested. Both the circuit functionality and the theoretical analysis are verified by the experimental results.