Abstract:Aiming at the existing depth model that fails to take into account both pixel-level features and semantic-level features at the same time when repairing nurals， resulting in problems such as lack of texture fineness and structural distortion， a progressive mural inpaining algorithm that combines kernel prediction and feature reasoning is proposed. Firstly， the regional progressive module is designed to realize the progressive mapping of mural features through partial convolution. Then， a dual-branch repair module is proposed， in which the kernel predicts the volume integral branch to realize the pixel-level repair of the damaged area. The semantic feature reasoning branch introduces gated deformable convolution and combines the semantic consistency attention mechanism to realize feature reasoning to complete the semantic-level repair of damaged murals. Finally， the two-branch repair results are fused into the output to minimize the reconstruction error and improve the repair accuracy. Through the digital restoration experiment of Dunhuang murals， the results show that the restored murals by the proposed method have better structural texture characteristics， which are better than the comparison algorithm in terms of evaluation indicators.

Abstract:To address the issues of high model complexity and poor feature extraction performance in Convolutional neural network-based dehazing algorithms， this paper proposes an end-to-end image dehazing algorithm based on joint mapping of two-branch features. Firstly， the atmospheric scattering model is transformed to separate the mixed-parameter feature and the single-parameter feature model. Then two feature extraction networks， MPFEM and SPFEM are designed according to the two-branch features and the outputs are weighted by two attention mechanisms. Finally， the extracted two-branch features are sent to the restoration module to restore the clear image and perform color-enhancing to obtain the final restored effect. To avoid the loss of texture details caused by using a single loss function in the model training process， multi-scale structure similarity and mean absolute error weighting are used as the loss function. Experimental results show that the proposed algorithm has a simple network structure， obvious dehazing effect， accurate color brightness restoration， and strong edge preservation.

Abstract:To solve the problem of the training instability of the Generative Adversarial Network （GAN） in generating ground penetrating radar （GPR） images， the Wasserstein GAN with Gradient Penalty is used to generate the GPR images. Moreover， a new method for constructing the GPR dataset is proposed base on the Finite-difference time-domain method and the measured images. Compared with the original GAN and Wasserstein GAN methods， WGAN-GP has better stability and the generated GPR images are more similar to the actual images. On this basis， the Dense Residual Block （DRB） and the U-Net are combined to propose a Dense Residual Denoising U-Net （DRDU-Net） suitable for GPR images. It uses the coding and decoding process of U-Net to improve the denoising performance. In addition， the introduction of DRB enhances the feature reuse of GPR image and makes U-Net training more stable. The performance of the proposal is evaluated by simulation experiments and compared with the BM3D （Block-matching and 3D） and U-Net. The results show that our proposal has better denoising performance than BM3D and U-Net. When the variance is 20， the peak signal-to-noise ratio increases by about 6.5 dB and 2.4 dB and the structural similarity increases by 0.09 and 0.04， respectively.

Abstract:The images collected by fish eye cameras in autonomous driving scenarios have severe distortion， complex scenes， drastic scale changes， and many small targets， which lead to low detection accuracy of traditional object detection models. Therefore， YOLOv5s-R， an improved fish eye image detection model based on YOLOv5s， is proposed. Firstly， to solve the problem of difficult recognition of minor targets， the RCMS （Random Crop Muti Scale） data augmentation method is proposed， which performs better than the optimal data augmentation method obtained from ablation experiments. Secondly， to improve the detection accuracy of the model， SA （Shuffle Attention） and LDH （Light Decouple Head） modules are added to the network header to enhance the model’s feature extraction and recognition capabilities， suppress noise interference. Finally， an additional angle prediction branch is added to realize the rotating box object detection， a circular label is constructed to solve the PoA （Periodicity of Angular） problem， and the label is smoothed with the Gaussian function. The RIOU is proposed to optimize the loss function by adding an angle penalty term on the basis of CIOU， which improves the regression accuracy and speeds up the convergence of the model. The experimental results show that the proposed YOLOv5s-R model achieves good detection performance on the Woodscape dataset. Compared to the original YOLOv5s model， mAP@0.5 mAP@0.5 is 0.95 increased by 6.8% and 5.6%， respectively， reaching 82.6% and 49.5%.

Abstract:Automatic segmentation of polyp images usually results in low segmentation accuracy due to the various sizes of lesion regions and blurry boundaries. Based on these two perspectives， a novel Progressive Reduction Network （PRNet） is proposed， which first locates polyps and then gradually refines their boundaries. The network utilizes Res2Net to extract features from the lesion region and leverages the multi-scale cross-level fusion module to improve localization accuracy. By combining the attention fusion mechanism with cross-level features in this module， the network can effectively solve the issue of multi-scale lesion areas. Furthermore， PRNet combines an uncertain region processing module and a multi-scale context-aware module when restoring image resolution from top to bottom. The former gradually mines polyp edge information by setting decreasing thresholds to enhance the recognition of edge detail features， while the latter， to improve the overall representation capability of the model， further explores the inherent potential contextual semantics of lesion regions. In addition， a simple feature filtering module is designed in this algorithm to filter the valid information in the encoder features. Experimental results on the Kvasir-SEG， CVC-Clinic， and ETIS datasets show that the Dice coefficients of the algorithm reach 92.09%， 93.05%， and 74.19%， respectively. Compared with other existing polyp segmentation algorithms， PRNet outperforms them and demonstrates its superior robustness and generalization.

Abstract:Timely and accurately capturing the tiny cracks in the shaft lining is of great significance for shaft safety. Lightweight detection models are the key to realizing the automatic detection of shaft lining cracks. Departing from existing traditional methods that focus on extracting deep semantic information， the application of geometric structure information represented by shallow features should be paid attention to and a lightweight detection model E-YOLOv5s for shaft lining cracks is proposed. Firstly， the lightweight convolution module， ECAConv， is designed， which integrates traditional convolution， depth-separable convolution， and an attention mechanism called ECA. Then， thefeature extraction capabilities are further enhanced by incorporating skip connections to construct the feature comprehensive extraction unit， E-C3. Thereby， the backbone network ECSP-Darknet53 is obtained， which significantly reduces network parameters and enhances the ability to extract deep fracture features of cracks. Finally， the feature fusion module ECACSP is proposed and the thin neck feature fusion module E-Neck is built by using multiple groups of ECAConv and ECACSP modules. The purpose of E-Neck is to fully fuse the geometric information of small crack targets and the semantic information of crack cracking degrees while accelerating the network reasoning. Experimental results show that the detection accuracy of E-YOLOv5s on the self-made shaft lining dataset is improved by 3.3% compared to YOLOv5s while the number of model parameters and GFLOPs are reduced by 44.9% and 43.7%， respectively. E-YOLOv5s can help promote the application of automatic detection of shaft lining cracks.

Abstract:Existing image segmentation algorithms face challenges related to low detection accuracy and a lack of specificity in crack detection. To address these challenges， this paper proposes an extended LG-Block module Extend-LG Block， which leverages a multi-scale feature fusion method. This module consists of multiple parallel dilated convolutions with different expansion rates. The number of branches and the expansion rate of dilated convolutions can be adjusted by parameters to change the size of its receptive field， and then extract and fuse crack features of different scales. By comparing the advantages and disadvantages of the network using a multi-scale feature fusion module in the deep layer and the network using a fixed scale structure for multi-scale feature fusion， a U-Net model with a variable scale structure named VS-UNet is proposed. The basic convolution Block in the UNet network is replaced by multiple Extend-LG blocks with different parameters. This structure performs multi-scale feature fusion in the shallow layer of the network， and the scale extracted by the multi-scale feature fusion module gradually decreases with the deepening of the network layer. This structure not only strengthens the detail feature extraction ability of the image while maintaining the original abstract feature extraction ability but also avoids the problem of increasing network parameters caused by the increase of convolution. Experiments are carried out on the DeepCrack dataset and CFD dataset. The results show that compared with the other two structures and methods， the proposed network with variable scale structure has higher detection accuracy and better segmentation effect for cracks of various sizes in visual experimental comparison. Finally， compared with other image segmentation algorithms， all indicators are improved to a certain extent compared with UNet， which proves the effectiveness of the improved network.

Abstract:To solve the problems faced by the existing edge computing task scheduling based on deep reinforcement learning， such as fixed action space exploration， low sample efficiency， large memory demand and poor stability and to better carry out effective task scheduling in the edge computing system with relatively limited computing resources， an adaptive edge computing task scheduling method D3DQN-CAA is proposed based on the improved deep reinforcement learning model D3DQN （Dueling Double DQN）. In the task offloading decision， the corresponding relationship between the task and processor is regarded as a multidimensional knapsack problem， and the computing node with the highest matching degree is selected for task processing according to the state information of the current scheduled task and the computing node； For improving the parameters updating efficiency of the evaluation network and reducing the influence of overestimation， a comprehensive Q-value calculation method is proposed； For accelerating the convergence speed of neural networks， an adaptive dynamic exploration degree of action space adjustment strategy is proposed； For reducing the storage resources required and improving the sample efficiency， an adaptive lightweight prioritized playback mechanism is proposed. Experimental results show that compared with multiple benchmark algorithms， the D3DQN-CAA algorithm can effectively reduce the number of training steps of deep reinforcement learning networks and make full use of edge computing resources to improve the real-time performance of task processing and reduce the system energy consumption.

Abstract:A user dynamic clustering and power alloction scheme is proposed for maximizing the sum rate in a downlink communication system employing non-orthogonal multiple access （NOMA） with unmanned aerial vehicles （UAVs） assistance. Considering the user quality of service and UAV position constraints， an optimization problem is formulated to maximize the sum rate.Due to the non-convexity of the objective function，the original problem is decoupled into three sub-problems to enhance system performance： UAV position deployment and user association， user dynamic clustering， and power allocation to improve system performance. Firstly， a UAV position deployment and user association scheme are designed based on the K-means algorithm， the objective is to minimize path loss， determining the optimal deployment position of the UAV with the simultaneous selection of the optimal user group to be served. Secondly， the multi-density stream clustering （MDSC） algorithm is improved， and a static and dynamic clustering scheme for users under a single UAV is proposed. The static clustering scheme can adaptively balance the number of clusters and the number of cluster users， and obtain a large difference in user channel gain within the cluster. The dynamic clustering scheme formulates an instant update strategy for user mobility attributes.Finally， by applying fractional programming （FP） and quadratic transformation， an auxiliary variable is introduced to transform the original non-convex problem into a convex problem. The auxiliary variable and power allocation are alternately optimized to obtain a suboptimal solution for the original non-convex problem. The simulation results show that compared with other algorithms， the clustering scheme in this paper can obtain larger intra-cluster channel difference and smaller standard deviation of the number of users in the cluster， and the performance of the user system is also significantly improved.

Abstract:Network traffic recognition is the foundation of network management and security services. With the continuous expansion and increasing complexity of the Internet， traditional rule-based recognition methods or based on flow behavior characteristics are facing great challenges. Inspired by natural language processing （NLP）， this paper proposes a fast classification method for encrypted traffic based on multi-feature fusion. The method completes the feature representation of network flows by combining the packet characteristics of data packets and byte sequences， expands the selected features into a double-byte sequence using binary byte encoding， and adds contextual semantic features of the bytes. By using pooling methods that are suitable for packet feature processing， the proposed model can preserve the feature information of packets to the greatest extent possible， thereby enhancing its noise resistance and more accurate classification ability. The method is validated on the Information Security Center of Excellence-2016 （ISCX-2016） and a private dataset containing Encrypted Traffic Datasets for 66 popular applications（ETD66）. The results show that the proposed method has significantly better accuracy and performance than other models in ISCX-2016 and ETD66， achieving accuracy of 98.2% and 98.6%， respectively， and thus proving the strong feature extraction ability and the model generalization ability.

Abstract:As a huge knowledge network diagram， the knowledge graph contains entity concepts， relationships， and other information. Although the semantic representation based on deep learning has strong generalization， it is not sensitive to some proprietary knowledge， so many researchers try to combine knowledge graphs with neural network. At present， most of the methods of semantic representation of knowledge graphs are based on general domain knowledge graphs， and there is no research on the semantic representation of knowledge graphs in the academic field. In this paper， the full-text data of academic literature is taken as the research object， and the semantic representation method based on an academic knowledge graphs is studied. On the basis of constructing academic knowledge graph， the research method of the general field （K-BERT） is improved （KEBERT）， and entity knowledge is further used to enhance the semantic information of the text. By conducting comparative experiments on downstream tasks， the performance of KEBERT， K-BERT， and ERNIE is verified on academic retrieval datasets. The experiment uses the NDCG evaluation index commonly used in the retrieval task to evaluate the results. The experimental results show that the improved KEBERT is superior to other models in the retrieval task.

Abstract:The heterogeneous system on chip has the characteristics of customization to meet the specific requirements of applications and has become the mainstream solution in many fields. However， users need to face program errors brought by various computing resources when developing on heterogeneous system on chip， and it is also a great challenge to build a unified debugger framework for different heterogeneous system on chip. To solve the above problems， a debugging framework for heterogeneous system on chip is proposed in this paper. A general interface of the debugging framework for heterogeneous processor is designed in this framework， which enables developers to quickly build heterogeneous debuggers through the framework functional interfaces. The debugger framework is rich in functions. It realizes debugging of heterogeneous multicore programs through thread switching and performances analysis of heterogeneous programs. Compared with traditional hardware debugging， the debugger generated by the framework loads heterogeneous programs faster， which is 5.5 times the read memory rate and 16.5 times the write memory efficiency， and the debugging speed is greatly improved.

Abstract:According to the idea of FFT（Fast Fourier Transform）→ optimization window → IFFT（Inverse Fast Fourier Transform）， the performance constraints of window function integration in linear time-frequency transformation are broken through， achieve the application of high-performance optimization window function in linear time-frequency transformation is achieved， and establish a new time-frequency transformation algorithm， K-S transformation is establisged. After frequency shifting for the FFT spectrum vector of signal x（t） is carried out， Hadamard multiplication with the spectrum vector of Kaiser optimization window at the frequency shift point， and then inverse FFT transformation （IFFT） on the product result is performed， the K-S transform complex time-frequency matrix is constructed to obtain three-dimensional time-frequency amplitude and time-frequency phase information of x（t）. The mathematical derivation， local properties， linear properties， and variable resolution properties of inverse transformation are given； simulation experiments on steady state and time-varying superharmonic signals from 0~150 kHz power grids show that the K-S transform has better resolution in both time and frequency domains than popular short-time Fourier transform and S-transform， and has excellent variable resolution performance. The actual measurement of 0~40 kHz superharmonic signals proves that the absolute error of superharmonic voltage amplitude measurement based on K-S transformation is less than 0.032 3 V.

Abstract:The optimal adjustment of reactive power compensation device such as capacitors can not only reduce the network losses， but also present influences on harmonic power flow and harmonic power losses. Wind generators can produce harmonic pollution and present impacts on harmonic power flow and harmonic power losses. However， the effects of the harmonic characteristic and output uncertainty of wind power on harmonic power flow and harmonic power losses are not considered simultaneously in the traditional reactive power optimization methods， which may result in the violation of harmonic standard and is adverse to network losses reduction. In this regard， this paper proposes the reactive power stochastic optimization model for power systems considering the impact of the wind power harmonics. In this model， the uncertainty of wind power is modeled by the scenario method， and the base-frequency network losses and the harmonic losses are considered in the objective function. Also， the constraints such as base-frequency power flow equations， the harmonic power flow equations and the total harmonic voltage distortion constraint are incorporated into the proposed model. After that， focusing on the proposed reactive power stochastic optimization model， the highly efficient method combining the adjustable driving force-based particle swarm optimization and a fully-connected deep neural network is proposed in this paper. Finally， the effectiveness of the proposed model and method is validated by three modified IEEE test systems.

Abstract:The fault characteristics of inverter AC transmission lines in AC-DC interconnection networks are quite different from those of pure AC systems. The protection scheme of traditional AC transmission lines is no longer suitable for inverter AC transmission lines. First， based on the different topological structures of the inverter AC transmission line when the fault occurs inside and outside， the expressions of the first traveling wave of the down-mode fault component current are derived， respectively， and the difference of the first traveling wave characteristics of the AC transmission line when the fault occurs inside and outside is clarified theoretically. Then， wavelet transform mode maximum method and Levenberg-Marquardt algorithm are used to extract the curvature radius of the first traveling wave， and a high-reliability single-terminal protection scheme is constructed by using the curvature radius of the first traveling wave to identify the fault region. Finally， the AC-DC interconnection network model is built in PSCAD/EMTDC， and the protection scheme is verified by MATLAB. The simulation results show that the proposed protection scheme is fast and effective， has high reliability， strong resistance to transition resistance， and good anti-noise ability.

Abstract:Because the traditional synchronized phasor algorithm is greatly affected by the intermediate harmonics of the power grid， this paper proposes a synchronized phasor measurement algorithm based on a newly improved matrix bundle. The power grid signal is constructed into a Hankel matrix， which is decomposed into singular values， and the noise interference components are filtered out using an adaptive order determination method. The accurate voltage amplitude， frequency and phase angle are obtained through matrix operation， and the simulation results show that this algorithm is superior to the traditional recursive Fourier transform algorithm in measuring power grid signals containing interharmonic components. A portable power grid synchronous phasor measurement device is designed to solve the problem of large volume and high cost of traditional synchronous phasor measurement equipment. The actual frequency of the crystal oscillator is monitored by high-precision satellite synchronous signal， and the sampling control parameters are dynamically adjusted to reduce the sampling time error. The synchronously sampled power grid waveform parameters and time tags are wirelessly transmitted to mobile terminals to complete the calculation of power grid signal amplitude， frequency and phase angle. The actual test shows that the device has high measurement accuracy， in which the measurement errors of voltage amplitude， frequency and phase angle are 0.038 5%， 0.000 51 Hz and 0.053 7°，respectively， which meets the requirrement of "Test specification for synchrophasor measurement unit for power systems"（GB/T 26862—2011）， and has practical application value.

Abstract:A 2.4 GHz integrated single pole three throw （SP3T） radio frequency （RF） switch and low noise amplifier based on a 90 nm SOI CMOS process is designed for wireless local area network （WLAN） applications. The RF switch adopts a low-power equivalent negative voltage biasing method， which can make the off-state transistors obtain an equivalent negative voltage bias state without using negative voltage， thereby improving the linearity of the RF switch. The low noise amplifier uses the negative feedback technique and derivative superposition technique to improve its linearity， and the derivative superposition technique is used to reduce the third-order non-linearity of the low noise amplifier， which further improves the linearity of the negative feedback low noise amplifier. The low noise amplifier is integrated with the RF switch and has a Bypass attenuation path. The measurement results show that the transmitting branch of the RF switch has an insertion loss of 0.95 dB and an input 1 dB compression point of 34 dBm， and the Bluetooth branch of the RF switch has an insertion loss of 1.68 dB and an input 1 dB compression point of 30 dBm. Under 2 V power supply， in the high gain mode， the receiving branch has a gain of 15.8 dB， a noise figure of 1.7 dB and an input third-order intercept point of 7.6 dBm， and a power consumption of 28.6 mW， while in the bypass mode， it has 7.2 dB insertion loss and an input third-order intercept point of 22 dBm.

Abstract:Voltage reference plays a crucial role in influencing the performance and accuracy of analog systems. General curvature compensation techniques focus solely on eliminating second-order temperature-related terms， making it hard to meet the high precision requirements of certain circuits. The existing circuit has a high-temperature coefficient issue that requires urgent compensation for higher-order terms. This paper proposes a novel high-order curvature compensation method， successfully implementing a low-temperature coefficient voltage reference circuit by leveraging the subthreshold characteristics of CMOS transistors. Initially， two currents with different temperature coefficients flow through the same subthreshold CMOS transistor， generating two gate-source voltages with unique temperature characteristics. Subsequently， the subtraction of these voltages produces a logarithmic voltage， and the logarithmic voltage is weighted and superimposed with the first-order compensation voltage to realize the high-order compensation. To enhance the power supply rejection ratio （PSRR）， the circuit employs a high-gain negative feedback loop， eliminating the need for an amplifier in traditional voltage reference circuits and further reducing power consumption. This design is based on the 0.18 μm CMOS process and is implemented using Cadence software for circuit design， layout， and simulation verification. Simulation results indicate that the circuit operates within a normal voltage range of 1.6 V~3 V， with a reference voltage output of 295 mV at 2 V operating voltage. The temperature coefficient within the range of -45 ℃ to 125 ℃ is 1.26 ppm/℃， and the PSRR is 51.1 dB@1 kHz， with a maximum static current of 8.9 μA. These results show that the voltage reference circuit can meet the needs of high-precision integrated circuit systems.

Abstract:In the realm of semiconductor technology control， achieving complete autonomous chip control has emerged as a focal point in today's semiconductor technology advancement. Given its features of open source and widespread adoption， the study of RISC-V architecture holds significant importance for enabling microprocessor autonomous controllability. Within microprocessor systems， limitations on physical resources and potential risks associated with direct storage access necessitate restrictions on DMA access to I/O devices， thereby impacting access performance. The prevailing approach involves virtualizing I/O transactions to effectively address this issue. This article firstly proposes a I/O virtualization architecture based on RISC-V， which greatly accelerates the I/O access process， this architectrue consums a few clock period to complete DMA requests from I/O devices to memory. This design will be integrated into RISC-V architecture CPU as an IP， accelerating the access of I/O devices to memory.

Abstract:As for the problems of the difficulty of heat dissipation and uneven temperature distribution in the rotor area of large-scale permanent magnet synchronous traction motors for high-speed trains， a novel cooling structure is proposed， which involves adding axial rectangular air passages on the outer surface of the stator core of the water-cooled motor housing， and forms an internal and external dual-cycle cooling structure in conjunction with the air gap and rotor lightweight holes. The purpose is to investigate the impact low of reducing the temperature rise in the stator and winding areas and improving the uniformity of internal motor cooling. Firstly， simulations are conducted using the Ansoft Maxwell platform to obtain the losses of various components in the dual-cycle cooling structure under rated operating conditions. To better simulate the airflow in the air gap driven by the rotation of the rotor， the air gap is treated in layers， and a fluid-structure coupled finite element analysis method is used to study the airflow characteristics and temperature rise patterns inside the motor under both single and dual-cycle cooling structures. The results indicate that the internal circulation air-cooling structure significantly increases the airflow velocity inside the motor and markedly improves the average heat transfer coefficient on the surface. As a result， more heat in the rotor area is transferred to the relatively lower temperature stator area and housing， while reducing the heat transferred to the rotor， thereby reducing the temperature rise of the rotor and permanent magnets. Furthermore， the orthogonal analysis method is used to optimize the structural parameters of the rectangular ventilation holes， including the cross-sectional area， quantity， and aspect ratio. The temperature rise uniformity coefficient is used to evaluate the temperature rise of the motor， and the optimal solution results in a 12.1 K reduction in the maximum temperature rise when compared to the single-cycle cooling structure and a 16.54% improvement in the overall temperature rise uniformity of the motor.

Abstract:A method for measuring and calculating the radiation emission index of linear oblique polarization radar is proposed in this paper. The actual amplitude of each frequency of the oblique polarization radar at any unknown angle can be quickly measured and calculated by using a single combination of the receiving antenna orthogonal measurement， and then the harmonic and spurious suppression degree of the oblique polarization radar can be calculated. The accurate polarization angle of each frequency point of an arbitrary oblique polarization radar can be quickly measured and calculated with the three frequency-domain combined measurement results of a set of X-axis symmetric angles and three polarization directions of 0 degrees. The proposed measurement and calculation method can effectively avoid the shortcomings of the current radar radiation emission measurement method for the oblique polarization radar. The measurement efficiency is high and the error is small. The test results prove the correctness and accuracy of the proposed method.

Abstract:Aiming at the high cost and difficulty of obtaining labeled data for power grid companies， and the difficulty of training an effective electricity theft detection model with unlabeled data， this paper proposes a method based on CT-GAN （Co-training Generative Adversarial Networks） semi-supervised electricity theft detection method. Firstly， the principles and structures of generative adversarial networks and semi-supervised generative adversarial networks are explored. Secondly， it is proposed to replace the JS （Jensen-Shannon） divergence and KL （Kullback-Leibler） divergence distance with the Wasserstein distance to solve the problem of unstable model training and low quality of generated data caused by the gradient disappearance and mode collapse of the generative confrontation network problem， and built a multi-discriminator Co-training model to avoid the problem of high distribution error of a single discriminator. At the same time， it enhanced the ability of GAN to generate label sample data. By expanding the label sample data set， the model detection accuracy and generalization ability were improved. Finally， the accuracy and effectiveness of the method are verified using the Irish power grid dataset.