Search

In this work, we present CleanUNet, a causal speech denoising model on the raw waveform. The proposed model is based on an encoder-decoder architecture combined with several self-attention blocks to refine its bottleneck representations, which is crucial to obtain good results. The model is optimized through a set of losses defined over both waveform and multi-resolution spectrograms. The proposed method outperforms the state-of-the-art models in terms of denoised speech quality from various objective and subjective evaluation metrics. We release our code and models at https://github.com/nvidia/cleanunet., Published in ICASSP 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). Listen to audio samples from CleanUNet at: https://cleanunet.github.io/

This paper addresses the problem of multiple extended objects (EOs) tracking, which has attracted increasing attention recently due to the developments of high-resolution sensors. Specifically, the state of an EO is modeled by the Gamma-Gaussian-inverse-Wishart (GGIW) distribution, and the message passing inference method, which provides a high efficient implementation for multi-target tracking, is adopted to propagate the posterior of the distributions of multiple EOs. The performance of proposed method is verified via simulation experiments.

Beampattern synthesis is a research hotspot in the design of MIMO radar transmission waveforms. In addition, the probing performance of the transmission waveform on the target also needs to be taken into consideration. Specifically, the high autocorrelation sidelobes of the transmitted waveform can easily deteriorate the performance of the matched filtering at the receiving end. The sidelobes of a strong target can conceal the weak target at the near range unit, and it is also easy for the range deception interference to affect filtering performance of the echo of the real target. This paper creatively modeled the waveform design problem as minimizing the mean square error, and designed the transmit waveform based on the alternating direction method of multipliers (ADMM) algorithm to obtain better beampattern synthesis performance and spatial autocorrelation performance. Simulation and numerical results verify the effectiveness of the algorithm.

Space-time steering vector mismatch of targets usually happens in the space-time adaptive processing (STAP) technique, causing space-time beam distortion. Due to the effects of clutter spectrum broadening and jamming, performance of the conventional STAP method deteriorates dramatically. This paper proposes a novel robust STAP method to improve the performance of clutter suppression and moving target detection. The proposed method employs the alternating projection and clutter cancellation method, to reconstruct the clutter plus jamming plus noise covariance matrix (CJNCM) and re-estimate the target steering vector. Firstly, reconstruction of CJNCM consists of summing the eigenvectors of the projection operator, combining the sample covariance matrix and the integral covariance matrix (ICM) of specific region. Secondly, the re-estimated target steering vector is obtained by estimating the dominant eigenvector of the ICM of the Region of Interest (ROI), which contains the target signal purely because CJNCM is cancelled before estimation. Finally, the robust space-time adaptive filtering weight vector is calculated through the MVDR method with the reconstructed CJNCM and the re-estimated target steering vector. Simulation results indicate that the proposed algorithm shows robust performance against space-time steering vector mismatch, better output signal-to-noise ratio performance and better moving target detection performance than the traditional robust STAP method.

Face age transformation aims to generate a face image in the past or future and has been receiving increasing attention for its significance recently. This paper proposes a local-global interactive framework for long-span face age transformation. We divide a face image into five local parts and design a local generative network for each of them to learn the local changes of a face image. Meanwhile a global generative network is utilized to learn the global changes. We introduce an interactive network and an age classification network, which are respectively used to integrate the local and global features and maintain the corresponding age features in different age groups. Given a face image at a certain age, our network can produce a realistic image of face aging or rejuvenation. We test and evaluate the model on complex datasets. Extensive qualitative comparison experiments prove the effectiveness and potential of our proposed method.

Video-based person re-identification plays important roles in surveillance video analysis. This paper proposes a duplex spatiotemporal filtering network (DSFN) to re-identify persons in videos. A video sequence is represented as a duplex spatiotemporal matrix. DSFN model containing a group of filters performs filtering at feature levels in both temporal and spatial dimensions, by which the model focuses on feature-level semantic information rather than image-level information as in the traditional filters. We propose sparse-orthogonal constraints to enforce the model to extract more discriminative features. DSFN seeks to characterize not only the appearance features but also dynamic information such as gaits embedded in video sequences and obtains performance improvement as a result. Experimental results show that the proposed method outperforms other comparison approaches.

Recognizing time-varying object states in complex tasks is an important and challenging issue. In this paper, we propose a novel model to jointly infer object fluents and complex tasks in videos. A task is a complex human activity with specific goals and a fluent is defined as a time-varying object state. A hierarchical graph represents a task as a human action stream and multiple concurrent object fluents which vary as the human performs the actions. In this process, the human actions serve as the causes of object state changes which conversely reflect the effects of human actions. For a given input video, a causal sampling search algorithm is proposed to jointly infer the task category and the states of objects in each video frame. For model learning, a structural SVM framework is adopted to jointly train the task, fluent, cause, and effect parameters. We test the proposed method on a task and fluent dataset. Experimental results demonstrate the effectiveness of the proposed method.

Up to now, the cumulative installed capacity of renewable energy power generation has exceeded 400 million kilowatts, accounting for more than 20% of the total installed capacity in China. The increasing proportion of renewable energy occupied the space of conventional power generation. The impact of renewable energy transient performance on stability of regional power grid and UHV AC/DC transmission capacity is increasingly significant. This paper proposes the control strategy optimization scheme for renewable power generation from the perspective of operational demand of the power grid. Based on the key technologies such as cluster simulation modelling technology of different types of inverters, as well as extraction and optimization technology of key parameters, the transient performance of renewable energy is improved. The applications on actual equipment, power stations, and power grids are performed to verify the effectiveness of the scheme and good performance are achieved. The achievements have good value for wide application to enhance the stability and renewable energy accommodation capability of regional power grids.

Regardless of the generation shedding of renewable energy in emergency control, or the system failure induced disconnection of renewable energy from the grid, the risk of large-scale renewable energy disconnection is increasing in modern power systems. The traditional path recovery models take less consideration of the characteristic of renewable energy, and cannot adapt well to the problem of system restoration containing renewable energy. In response to this, it is important to address the influcence of renewable enrgy during the path recovery process. Firstly, the traditional path recovery model is introduced. Then the influence of the access of renewable energy wind turbines on the system frequency is analyzed. With the goal of minimizing the recovery time of both renewable energy and conventional energy units, a new path recovery model for power systems containing renewable energy is proposed, based on improved node weighting method. Finally, an actual power system is used as an example to verify the effectiveness of the proposed model.

The continuous increase in the proportion of new energy access has led to a decline in the voltage support capacity of the sending-end power grid, and transient voltage instability problem was prone to occur under the small startup mode of synchronous power sources. Taking the Northwest Power Grid as an example, the transient voltage stability problem of the typical new energy collection and transmission end power grid was revealed. Based on the advantages of the synchronous condenser, the effect of the synchronous condenser location on the transient voltage stability problem after the accident was analyzed. Furthermore, the optimal configuration method of synchronous condenser for the sending-end power grid with less voltage support has been proposed in this paper. By coordinating and optimizing the location and capacity of the synchronous condenser before the accident, and emergency control after the accident. With the goal of minimizing the control cost, and constraint of the power system safety. Finally, the effectiveness and rationality of the configuration method was validated by actual typical power grid.

A B-Spline implementation of the multi-target multi-Bernoulli (MeMBer) filter for nonlinear Gaussian/non-Gaussian models is proposed. Specifically, the spatial PDF (SPDF) of each Bernoulli component in the MeMBer density is represented by a B-Spline curve, which is characterized by the spline knots and control points. The spline knots and control points are then propagated via prediction and update steps of the MeMBer filter. Besides, a revised fitting algorithm is proposed so as to improve the implementation efficiency. The effectiveness of the proposed method is assessed via simulation experiments.

The one-bit massive multi-input multi-output (MIMO) with receiving channel state information is considered in this letter. This letter focuses on the constrained capacity, i.e., the mutual information with given distribution of the source, computation with given modulation. The complexity analysis shows that the complexity of the constrained capacity computation based on the definition grows exponentially with the number of antennas at the base station. We reduce part of the computation by deriving the definition formulas, and propose the Monte-Carlo method to calculate the remaining part. With the efforts, we reduce the complexity down to polynomial. The numerical results show that the proposed method provides the estimation of the constrained capacity with small error if the number of Monte-Carlo tests is large enough. Analysis based on the numerical results is shown, which provides the system designer a basis for choosing modulation type and signal-to-noise ration level. The result also indicates that the degradation is only 2dB due to the coarse quantization of one-bit system at the main signal-to-noise ration region.

Depth completion is becoming a particularly important yet challenging problem with the growingly rapid progress of depth sensing technologies. Depth completion aims to complete sparse and noisy depth images to generate dense depth images. In this paper, we propose a multiscale adaptation fusion network (MAFN) for depth completion. The depth features are fused with RGB features at multiple scales with adaptation modules, where a neighbour attention mechanism is designed to adapt the local structures of the RGB image and the depth image. The fusion and completion process are unified under the encoder-decoder framework which is learned in an end-to-end way. By exploiting the detailed structural relationships of RGB images and depth images, our MAFN model can accurately complete and restore the invalid depth values on the sparse depth images. We test the proposed method on the challenging KITTI depth completion benchmark. The experimental results prove the effectiveness and strength of the proposed method.

In order to improve the student’s ability under the Emerging Engineering Education, the Blended learning in college physics experiments has been explored. In this study, a new learning method was developed using the combination of online and face-to-face class for students in University. This new learning method is characterized by absolute autonomy in the classroom. The students could complete self e-learning through the process of the class assisted by multimedia and teachers.

Despite recent advances, low-voltage operation remains one of the key approaches for power reduction. However, the continuous scaling of the SRAM bit cell, in advanced technologies, based on the transistor's minimum geometry is accompanied by increased random threshold voltage variations that limit the SRAM's minimum operating voltage (V MIN ). The introduction of FinFET transistors has provided better short channel effects and less random dopant fluctuation, compared to prior bulk CMOS technology. However, the quantization of transistor sizing continues to be a major challenge for high-density 6T SRAM design. Figure 15.1.1(a) shows the layout of a high-density 6T SRAM bit cell in a 5nm EUV and high-mobility channel FinFET technology. In order to achieve a compact bit-cell area, using FinFET technology, all transistors have to consist of a single fin. Figure 15.1.1(b) shows the SRAM bit-cell area scaling trend since 2011: area continues to scale, even as technology evolved into the sub-20nm era. This work reports a 0.021 µm2 high-density SRAM bit cell for a leading edge 5nm technology [1]. Figure 15.1.2(a) shows the SRAM voltage scaling trend: where the blue line indicates the V MIN of the SRAM bit cell without circuit assist and the red line indicates the V MIN with circuit assist. Write assist is the key enabler for lowering V DD operation in 5nm technology. Figure 15.1.2(b) shows a contention between pull-up (PU) and pass-gate (PG) transistors during a write operation. A stronger PU transistor yields higher read stability, but it degrades the write margin significantly and results in a contention write V MIN issue. In order to improve the write V MIN , the negative BL (NBL) and lower cell V DD (LCV) techniques have been proposed [2], [6]. However, the write-assist circuitry results in area overhead. In order to reduce the area of the assist circuitry, this work proposes NBL [7] and LCV with metal coupling/charge sharing capacitors to reduce the area overhead of the NBL-boosted and LCV charge-sharing capacitors. Furthermore, the flying BL (FBL) architecture [8] is also implemented to further reduce area.

In the past localization methods, both mechanical and phased array radars require angle scanning because their beampatterns are only angle-dependent, resulting in the echo time difference only contains the range information of the target. It will cause some ambiguity problems and limitations. Since Frequency Diverse Array (FDA) was proposed, its range-angle dependent beampattern provided more degrees of freedom (DoFs) to obtain more information in array signal processing. However, how to use these increased DoFs to get more information still needs to be studied. In this paper, we propose a novel localization method using FDA beam intersection. In this proposed method, part of the FDA beampattern approximates a straight line and propagates throughout the range axis. This moving beam line can capture the target like a fast-moving net and the echo time difference contains both the angle and range information of the target. By a simple straight-line intersection, we can obtain the whole position information, where two FDA beam lines intersect at the target. A simulation diagram is provided to illustrate the proposed localization process.

System moment of inertia and frequency regulation ability is decreasing with the increasing scale of UHVDC and new energy integration in China. Therefore, the frequency stability of the system is becoming increasingly prominent. The cost of traditional security control measures such as generator tripping and load-shedding is high, the low-cost DC emergency power support, the new energy frequency modulation, load control and other continuous regulation control measures get more attention in daily operation. Aiming at different types of control measures, this paper puts forward a frequency security margin measurement index, which takes multi-type control measures into account. On the premise of ensuring frequency security and stability of power grid, the cost of frequency control after power grid failure can be reduced and the requirement of lean frequency stability control of modern power grid can be met by coordinating the continuous and discrete control measures. Based on the actual planning of UHVDC transmission network, the feasibility of the proposed method is verified.

An HVDC transmission system is prone to generate overvoltage at the sending end after the DC block fault, which will present a challenge of high voltage ride-through (HVRT) for generators at the sending end. To ensure safe operation of the hybrid power system with large-scale photovoltaic (PV) power stations, a nonlinear HVRT control strategy based on state-dependent Riccati equation (SDRE) technique is proposed. The optimal control law is obtained by solving an algebraic Riccati equation (ARE)-like problem. Through the selection of the weight matrix and optimization calculation of state dependent coefficient (SDC) matrices, the control objective to maintain terminal voltage is achieved during the transient period. Compared with PI control strategy, the proposed method can make full use of converter control ability to suppress overvoltage at sending end and have better transient control performance during fault period.

In current photovoltaic transmission systems, low voltage ride-though is a main challenge that restricts the development of photovoltaic power generation. In this study, a power system with large photovoltaic (PV) power penetration is considered, and nonlinear control strategies for PV power stations are designed to guarantee the operation of PV power system and provide enough reactive power for the grid. First, the dynamic model of PV stations is constructed, and the control objective is to enhance the robustness of the system against grid faults. Second, the state-dependent Riccati equation (SDRE) technique is used to solve the robust control problem. The SDRE technique reconstructs a nonlinear system to a linear-like structure, and controls the system through solving an algebraic Riccati equation (ARE)-like approach with state-dependent coefficient (SDC) matrices. Comparisons with conventional PI controllers are performed, and better control performance is achieved through the proposed strategy.

Direction-of-arrival (DOA) estimation has been an important research topic in the field of radar signal processing for a long time. Recently, joint estimation of DOA and frequency has been extensively studied. This paper considers joint DOA and frequency estimation of multiple narrow-band signals received by array radars at sub-Nyquist rates. A new sub-Nyquist array architecture (a sparse array radar whose the reference unit connects to a sparse multi-coset structure and the other units connect to 2 channel multi-coset structure) and corresponding efficient technique (a joint estimation algorithm with an augmenting method) are proposed. The estimation algorithm can match frequencies up with corresponding DOAs, and the augmenting method can increase the capability of the number of identifiable signals from $M-1$ to $Q-1(Q$ is extended estimating degrees of freedom). We further analyze the minimum total sampling rate of achieving joint estimation. Simulations results verify the effectiveness and performance of this technique in this sparse array structure.

This paper addresses the problem of multisource direction-of-arrival (DOA) tracking based on the labeled multi-Bernoulli (LMB) filter. In order to avoid information loss, we directly utilize the received signal at an array antenna. The normal superpositional measurement model can be used to describe the relationship between DOAs from multiple sources to the received signal. LMB is not conjugated under the multitarget likelihood function of this distribution, so no analytical solution can be obtained. In order to overcome such difficulty, the likelihood of superpositional measurement model is approximated as a complex Wishart random matrix, and the updated LMB distribution is computed by matching the probability hypothesis density (PHD) of posterior. The auxiliary particle filter is employed to implement the LMB filter, and the performance of proposed algorithm is verified via simulations.

Accurate traffic light detection is an important problem for intelligent vehicles. While most existing methods deal with classification of red, green and yellow signals, traffic lights in real world contain a diversity of specific semantic forms, such as red right-turn arrow, green left-turn arrow, and red U turn. In this paper, we propose a multi-scale attention (MSA) network for semantic traffic light detection. The task is to infer the bounding boxes of traffic lights in images and identify their specific semantic categories. Our MSA network combines multi-scale information with attention blocks overcome the ‘small object’ challenge and increase the computation efficiency. Since there was no a proper dataset containing specific semantic traffic lights, we collected a large scale semantic traffic light dataset. Our dataset contains 11 standard categories of specific semantic traffic lights and about 14800 sample images. We test the proposed approach on the new dataset, Bosch Small Traffic Lights Dataset, and LISA Dataset. Experiments show that the proposed method improves the performance of the traffic light detection.

With the growth in popularity and size of social media, there is an urgent need for systems that can recognize the coreference relation between two event mentions in texts from social media. In existing event coreference resolution research, a rich set of linguistic features derived from pre-existing NLP tools and various knowledge bases is often required. This kind of methods restricts domain scalability and leads to the propagation of errors. In this paper, we present a novel selective expression approach based on event trigger to explore the coreferential relationship in high-volume Twitter texts. Firstly, we exploit a bidirectional Long Short Term Memory (Bi-LSTM) to extract the sentence level and mention level features. Then, to selectively express the essential parts of generated features, we apply a gate on sentence level features. Next, to integrate the time information of event mention pairs, we design an auxiliary feature based on triggers and time attributes of the two event mentions. Finally, all these features are concatenated and fed into a classifier to predict the binary coreference relationship between the event mention pair. To evaluate our method, we publish a new dataset EventCoreOnTweet (ECT)1 that annotates the coreferential relationship between event mentions and event trigger of each event mention. The experimental results demonstrate that our approach achieves significant performance in the ECT dataset.

Face age transformation is an important issue in many applications. While unidirectional and short-span face ageing has achieved remarkable progress, it remains a challenging problem to generate both younger-look and older-look face images over long age span. In this paper, we present a progressive residual adversarial autoencoder (PRAA) model for bidirectional and long-span face age transformation. Given an input face image, our model aims to synthesize face images of its younger looks (face rejuvenation) and older looks (face ageing). The PRAA contains adversarial generators and discriminators where age information is encoded as latent features. It adopts a residual-encoding strategy by which the original face images and the residual face images are jointly encoded. We adopt a progressive multi-scale method to train the network, by which our model can capture both the global structure and the local detail changes in face age transformation. We test our model on challenging data and the experimental results prove the strength of our method.

The interest of this paper is to detect an unmanned aerial vehicle (UAV) and then initialize its trajectory, if it exists. The difficulties of detecting a UAV mainly lie in two aspects: a) small radar cross section (RCS), which causes extremely low signal-to-noise ratio (SNR); and b) low velocity, which results in weak Doppler effect. In this case, traditional track-before-detect (TBD) algorithms cannot achieve the satisfying probability of detection. In this paper, we propose to solve the problem of detecting a UAV based on micro-Doppler aided dynamic programming TBD (MA-DP-TBD) algorithm where the effect of micro-Doppler caused by the blades of UAV is taken into consideration to aid the detection process. The performance of proposed algorithm is examined via simulations.

Vehicle candidate generation is important for vehicle detection. Existing vehicle detection studies usually employ general-purpose region proposal methods to generate vehicle candidates, which do not consider the specificity of on-road vehicles in traffic scenes. In this paper, we propose a model to re-rank the candidates that are generated by general-purpose region proposal methods. Our model considers the specificity of on-road vehicle candidate generation in traffic scenes by encoding global-local semantic context and location-size geometric compatibility. In the experiments, we test our model on three art-of-the-state region proposal methods using two public datasets. The results show the significant performance improvement is gained after applying our model.