Abstract:In view of the problems that the existing deep learning repair methods are greatly affected by the structure and texture， resulting in structural disorder and texture blur in the repair results， a generative adversarial mural restoration model jointly guided by structure-gated fusion and texture is proposed. Firstly， the generation net-work composed of a structure-guided coding sub-network and texture-guided decoding sub-network is designed， which uses structure information to guide coding， and enhances edge contour information through gated features. Then， the texture guide and orientation attention module are used to extract layered texture features， guide the de-coder to repair， and improve the texture consistency of murals. Finally， skip connection is used to promote the feature complementarity of structure and texture， and the spectral-normalized PatchGAN discriminant model is used to com-plete mural restoration. The results from the digital restoration experiment of real Dunhuang murals show that the subjective and objective evaluation of the proposed method is better than the comparison algorithm， and the restora-tion results are clearer and more natural.
Abstract:To solve the problem that froth image is difficult to be accurately segmented due to its high complex-ity， this paper proposes a new I-Attention U-Net network for froth image segmentation. The algorithm uses the U-Net network as the backbone network and uses the Inception module to replace the first convolutional pooling layer so as to extract the multi-scale and multi-level shallow feature information of the froth image. A Pyramid pooling module is introduced to improve the segmentation effect by summing the feature maps of different scales. And the self-attention gating unit is improved to make it more suitable for the segmentation of flotation froth images， which strengthens the importance of deep features and performs reinforcement learning on froth edges of different sizes. The research results show that the Jaccard coefficient of the algorithm proposed in this paper is 91.73% and the Dice coefficient is 95.66%. Compared with the results of other similar segmentation algorithms， the Jaccard coefficient and Dice coefficient are increased by 1.59% and 0.88%， respectively. The model can better segment the zinc flotation froth image， and solve the problems of under-segmentation and over-segmentation， which is a good way for the follow-up. In addition， the method has less detection time and fewer model parameters and also has the ability to be deployed in industrial field computers， which has certain practical application value.
Abstract:Most of the commercially deployed unmanned mining systems use LiDAR and radar as sensors， which is difficult to identify object types， especially for distant objects. This affects the subsequent correct decision-making， as well as the safety and overall efficiency of the unmanned system. To solve these problems， this paper collects mine data from different scenes and proposes an image object detection algorithm based on YOLOv5S. The algorithm mainly improves in the following three aspects. Firstly， the sampling ability of the model is optimized by using different padding strategies and spatial attention modules. Secondly， it decouples the head prediction branches and makes each branch focus on its own task. Finally， the loss function is optimized to couple localization and classi-fication so as to realize the joint optimization of localization and classification tasks. Experiments show that the above three methods can improve the AP of YOLOv5S from 49.9% to 58.9% with real-time performance and realize object recognition in daytime and night scenes with different scales.
Abstract:To realize the tomato stem classification of irregular thin and long objects with a similar color to the background， an algorithm for tomato stem classification based on improved Mask R-CNN was proposed. First， the images of daytime and night tomato plants were collected， and stems in the images were labeled using the Labelme to produce data sets of tomato stem classification at daytime and night， separately. Then， the Mask R-CNN model was trained separately using the two data sets with the transfer learning method. The Mask branch was improved to generate its minimum external moment when the Mask was produced， an evaluation index Re was proposed to evaluate the accuracy of the Mask border， and pixel-level evaluation indexes were proposed to comprehensively evaluate the performance of the model. Test results show that the pixel F1 score and pixel mean average precision of the night and day stem classification model are 48.82%， 50.03%， and 57.76%， 56.06%， respectively. After the improvement of the Mask branch， the accuracy of the mask border is significantly improved. The average recognition time of each image is 0.31 s， indicating that the recognition model can meet the real-time requirements of the algorithm in practical applications. It can provide method support for the intelligence of plant pruning and other works.
Abstract:Aiming at the problem that the existing dehazing algorithms lack attention to the noise concentration in different regions of the hazy image and the distinction between far and near features， this paper proposes a new generative adversarial network model. In the model， two UNet3 + networks are used to realize the full-scale jump connection and depth supervision， and multi-scale feature fusion is used to extract the high and low-level semantics in different scale feature images. The addition of deep supervision can better learn the near-far level representation in the image. At the same time， the multi-scale pyramid feature fusion module integrating the self-attention mechanism is added to the generator structure to better retain the multi-scale structure information of the feature map and improve the attention to different haze concentration regions. The experimental results show that the algorithm network can obtain better visual effects than other advanced algorithms such as BPPNET on NTIRE 2020， NTIRE 2021， O-Haze datasets， and Dense-Haze datasets. The peak signal-to-noise ratio and structural similarity index on the Dense-Haze dataset are， respectively， 24.82 and 0.769.
Abstract:An improved RGB-D ORB-SLAM2 algorithm is proposed to address the problems that the traditional ORB-SLAM2 system cannot build a dense map and lack an octree map that can be applied to the navigation and path planning of mobile robots in specific scenarios. The algorithm uses depth information to calculate the 3D spatial position of the point clouds in key frames through the pinhole imaging model and uses outlier filtering to remove redundant clutter and voxel filtering to retain point clouds with feature information for stitching and reduce map redundancy. And through dense loopback processing， the point cloud poses are further optimized and updated in keyframes， and an accurate dense point cloud map is constructed and converted into an octree map. The experimental data shows that compared with the RGB-D SLAMV2 system， the global trajectory error and relative positional error of the RGB-D ORB-SLAM2 system are improved by more than 50%， the root mean square error is 0.89%， and the mean error is 0.76%； in terms of map-building performance， the number of point clouds is reduced by about 30% on average when compared with the same type of algorithms. In addition， the octree map occupies only 0.6% of its memory when compared with the point cloud map， which better meets the high precision and fast navigation demands.
Abstract:Most of the existing abnormal traffic methods are based on supervised learning. It is extremely difficult to obtain and mark abnormal traffic data samples in real life， and there are many limitations. In addition， due to the diversity and complexity of abnormal network data， the adaptability of various detection methods is poor， and it is difficult to judge the new abnormal traffic. Based on the above problems， this paper designs a semi-supervised abnormal flow detection framework， MeAEG-Net （Memory Augment Based on Generative Adversarial Network）， to detect anomalies by training only normal flow sample data and comparing the reconstruction errors of the underlying characteristics of input flow of generator module. A generative adversarial network is used in the model to better train the generator. The generator adopts the structure of an autoencoder and decoder to solve the problem that the autoencoder is susceptible to noise. The memory-augmented module is added to the sub-network of the autoencoder to weaken the generalization ability of the generator module and increase the reconstruction error of abnormal traffic. Experimental results show that the method proposed in this paper can achieve a good effect on abnormal traffic detection under the premise of learning only normal traffic data samples. Finally， the future research direction and challenges have been prospected.
Abstract:The text representation method based on graph structure has a better effect in news text deduplication. However， at present， this representation method cannot fully represent the complete information of the text， and ignores the semantic information of the graph， which reduces the deduplication effect of news text. To this end， this study proposes a text deduplication algorithm based on event heterogeneous graph representation. The algorithm first represents the global semantic and structural information of news text through event heterogeneous graph， and then proposes a dual-label graph kernel algorithm to represent event heterogeneous graph to realize the structure and semantic information of the deep representation graph. The experimental results show that the deduplication algorithm proposed improves the F1-score index by 10%， compared with the existing text representation deduplication method based on graph structure. Finally， the algorithm can improve the deduplication effect of news text.
Abstract:Entity resolution （ER） is a task to identify whether several records correspond to the same entity in the real world， which is a key problem in data cleaning and data integration. Recently， deep learning-based entity resolution is popular， which requires a large number of labeled data to achieve better results. However， a large number of high-quality labeled data are not always easily available in the real scenario. This paper proposes a deep transfer learning-based entity resolution model. The common features of the source domain and the target domain are extracted through a domain separation network. ER results are obtained by utilizing these common features. Therefore， the common features are transferred from the source domain to the target domain. The experimental results show that， on several datasets， the proposed method has a maximum improvement of about 40% in the F1 metric compared with the previous best method. Experiments show that the proposed method has superior performance and shorter training time.
Abstract:Aiming at solving the problem that it is easy to fall into the local optimum when using RGB-D data to perform 3D point cloud registration， a 3D point cloud registration method based on a multi-dimensional-feature PVDAC descriptor is proposed. Firstly， the key-points of the two-dimensional data are extracted through the ORB feature detection algorithm. Secondly， the gray features of the key-points in 2D， the local pixel value distances， point cloud normal angles， and curvature features of the key-points in 3D are calculated， respectively. Thirdly， the 2D feature and 3D feature are combined to generate a new PVDAC pixel descriptor， which is used to describe the key-points to achieve the coarse registration of the 3D point cloud. Finally， the fine registration of the 3D point cloud is completed based on the ICP algorithm. Experiment results show that overall mean square error of this method is about 0.05 m2 when registering a point cloud in a large scene， and it reaches a small error of 0.0002 m2 when registering a single-object point cloud in a small scene.
Abstract:Aiming at the problem of accurate diagnosis of ball milling load state under complex grinding conditions， a load state diagnosis method of ball milling based on Deep Wide Residual Shrinkage Networks （DWRSNs） is proposed. Firstly， a wide-convolutional neural network is used to extract the short-term features of the vibration signal， a three-layer deep residual shrinkage network is established， and a soft threshold function is used for nonlinear transformation. Then， the advanced features of the load state oriented are extracted based on the self-learning threshold of the attention mechanism module. And discrimination of the load state of the ball milling is realized through the full-connection layer and the soft layer. The measured results prove that the DWRSNs method proposed in this paper is superior to the existing DCNN， ResNets， and DRSNs diagnostic methods in terms of fit， convergence speed， and learning ability. Meanwhile， the exacted vibration signal features are highly representative， the compactness within the cluster is high， and the boundary between the clusters is obvious after TSNE visualization. The accuracy of the DWRSNs diagnostic test set of the proposed method exceeds 99%， and the cross-entropy loss is 0.0772. Compared with the existing load state diagnosis method， it has higher accuracy and less time-consuming diagnosis and can achieve accurate identification of the load state of the ball milling and provide an effective and reliable criterion for optimizing the control of the process of beneficiation and grinding and improving the efficiency of grinding.
Abstract:The current human pose estimation networks are difficult to be widely used in mobile devices and embedded platforms due to the arithmetic power and memory limitations. To address this problem， this paper proposes a lightweight human pose estimation network X-HRNet with HRNet as the basic framework and uses the ResNeXt module to replace the common basic module to reduce the parameters and computational complexity of the network. The proposed model achieves 78.2% accuracy on the COCO validation set， which is 1.9% higher than that of the HRNet， the number of parameters decreases by 22.2M， and the computational effort decreases by 27.3 GFLOPs. The proposed X-HRNet is a method with the combination accuracy and lightweight， which proposes a new lightweight human pose estimation network for embedded platforms by reducing the computation and the number of parameters effectively while maintaining accuracy.
Abstract:A power amplifier （PA） with high outputpower based on the GaAs HBT process is proposed. The design adopts the cascade structure of three-stage amplifiers to improve the power gain of the power amplifier.And the RC lossy stabilization network is connected in series at the base of the power transistor to improve the stability and the gain flatness. Active bias is used to improve the output power， efficiency， and linearity performance of large signal output. A power detection circuit is added at the output stage amplifier to obtain a DC voltage signal that varies with the output power. The layout EM simulation results show that： the output frequency range of the PA is 5.1~ 6.5 GHz， and the gain is 33~33.7 dB. Return loss is less than -9.8 dB， the saturated output power is 32.8~ 34.9 dBm， and peak efficiency is 38.7 %~42%. In the case of meeting the wireless LAN standard 802.11ax and the modulation strategy of MCS7， the output power is 20~21 dBm when the EVM reaches -30 dB. The chip area is 1.69 mm×0.73 mm. The test results show that the S-parameter test results and the simulation results show good consistency， and the output power of the PA is 13.6~19.8dBm when it meets the aforementioned wireless local area network standards.
Abstract:To meet the requirements of power consumption and area for the output signal quantization of the flexible piezoresistive sensor， this paper presents a low-power successive approximation （SAR） analog-to-digital converter （ADC）.The monotonic switching method based on the ground sampling technique minimizes DAC switching energy， while a split-capacitor DAC achieves low power in an area efficient manner. In addition，a comparator using a two-stage dynamic preamplifier was proposed to diminish the offset and noise. And dynamic element matching （DEM） techniques are employed to enhance linearity.Circuit design and layout drawing of the proposed SAR ADC were realized in 0.18 μm 1P6M CMOS technology，which occupies an active area of 630 μm× 575 μm. The SAR ADC consumes 25.7 μW at 1.8 V supply voltage.The measurement results at a sampling rate of 250 kHz show that this 11-bit ADC achieves a signal-to-noise- and-distortion ratio （SNDR） of 65.0 dB and an efficient number of bits （ENOB） of 10.51 bit.
Abstract:A sliding-mode decoupling control strategy based on an extended state observer （ESO） is proposed for the problem that the output branch of a single-inductor dual-output Buck converter suffers from severe crossover effects and slow transient response when load disturbances occur. Firstly， considering the load disturbance problem， the system model is transformed into the output voltage deviation model.And the main circuit ESO is designed to estimate the load disturbance and compensate the disturbance estimation information to the improved reaching-law backstepping sliding-mode controller of the main circuit switching tubes. Next， considering the branch circuit coupling problem， a branch circuit is fitted as an independent system based on the self-anti-disturbance paradigm， where the branch circuit coupling term and the external disturbance are regarded as the total disturbance.And the branch circuit ESO is designed to estimate it， and a sliding mode and active disturbance rejection controller is constructed in the branch circuit switching tube based on the disturbance estimation information and the sliding-mode control algorithm. Finally， the closed-loop stability of the main and branch circuit controllers is demonstrated using the Lyapunov theory. The simulation results show that the control strategy significantly reduces the cross-influence between the branches and improves the transient response speed of the system.
Abstract:Most current distribution network fault diagnosis schemes need to be supported by a large number of fault simulations. With the continuous expansion of the scale of the distribution network， the fault probability increases year by year. This kind of method can easily be limited by different fault types and numbers， resulting in a sharp increase in the amount of simulation calculation and difficulty in diagnosing quickly. Therefore， this paper proposed a fault diagnosis method based on unified features. Firstly， it used the voltage increment relationship of sparse measuring points to deduce the unified fault characteristics of the distribution network and introduced a neural network to build the fault diagnosis model. Combined with an example， the unified feature diagnosis method is tested， and its computational advantage is analyzed. After that， it extended the unified feature diagnosis method to a large-scale distribution network and realized the parallel diagnosis of each sub-network through the partition method. The results of several diagnosis examples show that the proposed method can diagnose effectively by using the sparse voltage increment value. The simulation times are independent of the fault type and number but only depend on the number of branches， which greatly reduces the amount of calculation， and has no strict synchronization requirements for the measured data.
Abstract:Lightning seriously threatens the safe and stable operation of power system. With the continuous development of power system， the traditional line lightning protection assessment technology has been unable to meet the requirements of lightning risk assessment. Based on this， this paper uses the coefficient of variation method to analyze the data of lightning falling in recent years within 10 km of ten important transmission channels of 500 kV inter-provincial connection lines in central China， and finds that subsequent lightning return strokes have a great impact on lightning strike trip. At the same time， many circuit actual lightning trip faults are related to subsequent lightning return strokes. Accordingly， this paper proposes a flashover criterion model of insulator string considering subsequent return stroke. The influence of subsequent lightning return stroke on lightning resistance level of 500 kV transmission lines is analyzed. The results show that the lightning resistance level of 500 kV transmission line will be reduced by the follow-up return stroke， which is related to the power frequency voltage cycle， and the impact of the subsequent lightning return stroke on the lightning back-flashover resistance level is more significant than the lightning shielding resistance level. The conclusion of this paper can help improve the accuracy of lightning risk assessment of high-voltage transmission lines and provide help for lightning protection and early warning of high-voltage transmission lines.
Abstract:This paper proposes a method to characterize the spatial electric field distributions of long rod-plane and sphere-plane gaps. The finite element simulation result of the electric field was post-processed， and 66 feature quantities were extracted from the interelectrode path and a conical region. A prediction model was established based on the least squares support vector machine （LS-SVM）， taking the electric field distribution feature set and the discharge voltage as input and output parameters. The improved grey wolf optimizer was used to optimize the model parameters， and the feature dimension was reduced by the maximal information coefficient method. This model was applied to predict the standard switching impulse discharge voltages of long rod-plane and sphere-plane gaps. The results of the case study indicate that the predicted values of test samples are in good agreement with the experimental values， with a maximum relative error of 8.3% and the mean absolute percentage error of 3.2%. This study can provide references for air gap insulation calculation.
Abstract:The current mechanical pressure-holding ball valve is limited by the narrow space in the well， the pressure-holding ability is weak， the pressure-holding success rate is low， and the flexibility is poor. In view of the above problems， a design concept of a "semi-floating" coring and the pressure-maintaining ball valve is proposed. Based on the technical requirements of pressure-holding coring， numerical simulation and theoretical calculation are used to design and develop a "semi-floating" ball valve pressure-holding control technology. The innovatively designed "semi-floating" pressure-bearing structure changes the pressure-bearing position of the ball valve， effectively enhancing the ball valve pressure capacity. A new pneumatic control method is proposed， and the opening and closing of the ball valve can be controlled， which can effectively solve the problem of closing the ball valve in the operation of highly deviated wells and horizontal wells. The finite element comparative analysis of the "semi-floating" and fixed ball valve bodies and valve seats were carried out， respectively. The pressure-bearing capacity of the "semi-floating" ball valve reaches 70 MPa， and the pressure-holding capacity is significantly improved. The torque and minimum gas pressure of the ball valve are determined by theoretical calculation. The research results can provide certain technical support for the development of pressure-holding coring technology.
Abstract:Aiming at the shortcomings that the single injection dose of the current single-hole needle-free injector is small and requires repeated injections， a large-dose multi-hole needle-free injector is proposed. ANSYS Workbench simulation platform was used to analyze the flow field of the needle-free injector， and the working pressure range was gained not less than 13 MPa. Meanwhile， the stress variation law of the needle-free injector under different diameters was obtained by fluid-structure interaction analysis. The relationship polynomial between the allowable stress and the minimum ampoule diameter under various pressures is fitted. According to the simulation results， an orthogonal test was designed to optimize the structure of the contraction section for the needle-free injector. The optimized scheme was obtained as follows： the contraction angle was 20°， the aspect ratio was 1.4， the diameter of the distribution circle was 2 mm， the contraction section length was 4.05 mm， and the diameter of the micro-hole was 0.165 mm. Through the stress analysis of the ampoule， it is verified that the optimized scheme can meet the mechanical property requirements of the material. The optimization results show that a larger dose is achieved than the traditional single-hole injector， the dose can be up to 5 mL， and the injection speed is as high as 150 m/s， which can penetrate the skin without causing too much damage. The feasibility of the optimization scheme was further verified through experiments and kinetic analysis.
Abstract:In order to obtain the performance attenuation characteristics of the fuel cell air compressor and im-rove its service life， the durability test for a self-developed fuel cell air compressor was carried out. Combined with the actual road cycle characteristics of fuel cell vehicles， an air compressor life test condition close to the real road cycle was constructed. On this basis， the durability test of the air compressor was carried out for 4 000 hours， and the performance and operation parameters of the air compressor in different running-in periods were collected. Through the analysis of durability test data， the attenuation law of the air compressor performance and its influencing factors were clarified. In addition to running-in time， the performance attenuation of the air compressor is mainly affected by speed and pressure， where the pressure attenuation becomes larger with the increase of speed， the attenuation rate increases first and then slows down with the running-in time， the attenuation of airflow is sensitive to both speed and pressure， and the attenuation rate remains unchanged with running-in time. After 4 000 hours of running-in， the maximum attenuation of air compressor outlet pressure and airflow reaches 7.5% and 29.7%， respectively， at rated operating conditions. The above results provide data basis and reference for the optimization of reliability and design of accelerated endurance conditions for the air compressor.
Abstract:In this study， a notchback SAE model with a 20° back angle is used as a reference model. The unsteady flow field structure of vortex generators （VGS） and riblets （RTS） on the roof and rear of the inclined plane is studied based on the improved detached eddy simulation （IDDES） turbulence model. The drag reduction mechanism of the two passive control devices is analyzed in detail. In addition， the sound source information of the flow field is extracted， and the noise of the automobile is analyzed by Computational Aeroacoustics （CAA） and Ffowcs Williams-Hawkings （FW-H） equation， and the effect of the two devices on the acoustic characteristics is investigated. The results show that the additional devices can reduce the rear-vortex structure strength， aerodynamic resistance， and pulsating pressure， which not only achieves the effect of drag reduction but also has the effect of noise reduction. The drag reduction rates are 2.41% and 2.76%， respectively， and the maximum reduction values of total sound pressure levels are 9.55 dB and 5.46 dB， respectively.
Abstract:This study aimed to investigate the effect of bipolar plate geometry on the contact behavior of PEMFC. The bipolar plate compression test and the corresponding finite element （FE） model were developed taking a metal bipolar plate as an objective. The effectiveness of the FE model was validated against the experimental data collected by pressure-sensitive films. Then， the FE models of bipolar plates with trapezoidal， rectangular， and wavy section shapes and the corresponding compression tests were established， respectively. The effect of bipolar plate section shape on contact behavior under different assembly pressure was investigated by FE simulations. The results demonstrated that the assembly pressure and the section shape of the bipolar plate had a significant effect on the contact behavior between the bipolar plate and the gas diffusion layer （GDL）. Under different assembly pressure， the amplitudes of GDL surface contact pressure beneath the ribs of three kinds of the bipolar plate were between 0.5-2 MPa， which can meet the assembly requirements. The rib width of the bipolar plate is the critical factor that determines the amplitude of the overall average contact pressure between the bipolar plate and GDL as well as the average intrusion area of GDL. Among those three kinds of bipolar plates， the contact behavior under the rectangular bipolar plate is more sensitive than others， while the uniformity of average contact pressure distribution is the worst.