Abstract:Aiming at the multipeak characteristics of photovoltaic arrays under complex illumination conditions and the problems that the conventional maximum power tracking algorithm cannot take into account the tracking accuracy and speed, and there are oscillations, an improved vulture search algorithm is proposed. In this algorithm, chaotic mapping is used to optimize the parameters of location control change when selecting search space, increase the diversity of population location and improve the global search ability of the algorithm. When searching space prey, Levy flight mechanism is introduced to optimize the new position of condor, which enhances the ability of the algorithm to jump out of local search. A new parallel Boost circuit is proposed to improve the stability of tracking and reduce the oscillation. Through Simulink modeling and simulation analysis, it can be seen that combined with the improved algorithm and the new parallel Boost circuit, the tracking success rate is 100%, 98%, 96% respectively in the three shadow cases, which are higher than the original algorithm, and can achieve the maximum power tracking faster and more stable.

Abstract:To fully exploit the eye movement pattern information, maximize the optimization model effect and improve the eye movement pattern recognition accuracy, this paper proposes an eye movement pattern recognition method based on fused features and optimized random forest. First, we extract three groups of feature parameters: Conventional eye movement features, eye movement sequence subpattern features, and gaze points gaussian distribution features, combine them with ReliefF to select important features and build a fused features matrix. Then we use the particle swarm algorithm to globally seek the model parameters based on random forest to build the optimized random forest eye movement pattern recognition model. We verify the effectiveness of the proposed method by using the open dataset of eye movement experiments of autistic patients, and the experimental results show that the proposed method can better distinguish the difference of eye movement patterns between normal and autistic patients, and the classification accuracy is improved by 957% compared with the random forest of Conventional eye movement features. Therefore, the fused features can better exploit the information contained in the eye movement patterns, and the particle swarm algorithm can effectively optimize the effect of the pattern recognition model, which provides a new idea and method for eye movement pattern recognition.

Abstract:At present, the hybrid active power filter (TCLCHAPF) with thyristor controlled LC reactor generally uses the equivalent impedance method to control the TCLC part and the hysteresis comparison method to control the active power filter part. This kind of controller ignores the connection between active and passive parts, so it can not guarantee the stable harmonic suppression capability in the whole range. In this regard, taking singlephase TCLCHAPF as the research object, the TCLCHAPF dualmode predictive control model is analyzed and obtained. On the basis of the predictive control model, the active part and passive part are jointly controlled. Then the historical waveform of the instruction current is input to the predictive controller, and the possible direction of performance optimization is obtained through multi period prediction and rolling optimization, thus improving the longterm operation performance with stable load. In order to make the optimization result of discrete output act on TCLCHAPF device, a terminal pulse PI controller is proposed. The simulation experiment under variable working conditions shows that the model predictive control scheme is feasible. Compared with the traditional scheme, the controllable range becomes wider and the control performance becomes better when it is stable.

Abstract:Based on the principle that the step impedance resonator (SIR) is smaller than the uniform impedance resonator (UIR) at the same frequency, a miniaturized interdigital filter is designed and fabricated by using thin film technology in this paper. The length of SIR resonator is 248% shorter than that of UIR resonator. A miniaturized interdigital filter is designed using this miniaturized resonator. The filter is fabricated on a 0254 mm thick ceramic substrate with a relative dielectric constant of 98 using a thin film process. The measurement results show that the passband of the filter is 355~435 GHz, the insertion loss at the center frequency is 3 dB, the flatness in the band is 17 dB, the return loss in the band is less than -186 dB, and the rejection at 28 and 51 GHz outside the band is -428 and -661 dB respectively. The size of the miniaturized filter is only 640 m×479 mm（021λg×016λg）.

Abstract:Aiming at the problem of synchronous dynamic data acquisition of longdistance multiaxis laser interferometry, a highspeed dynamic measurement data acquisition system based on FPGA and a computer dynamic measurement software are developed. Using the hardware system scheme of DDR3 mounted on FPGA, the offline system is designed based on edge detection and capture of the external trigger signal. Read or write FIFO combined with MIG IP core was constructed to realize FPGA logic, which meets the requirement of reducing the realtime performance of measurement software. The hardware establishes serial communication with the computer to realize realtime sampling of multiparameter and multichannel environment sensors such as temperature, atmospheric pressure, and humidity, which achieve realtime correction of air refractive index. The difference is less than 8 nm with pulse repetition interval TPRI≤20 μs, while the dynamic measurement is compared with the interferometer product. The result shows that the system can achieve high precision measurement at the nanoscale at the sampling rate of 50 kHz without losing data or error code. The system can be widely used in highspeed and highprecision dynamic measurement data acquisition.

Abstract:Two-dimensional discrete fast Fourier transformation is widely used in digital image processing, which is of great significance in engineering field. Usually, 2DFFT is computed using column decomposition, that is, a rowwise 1DFFT followed by another columnwise one. Due to the limitation of data transmission bandwidth of field programmable gate array and the physical structure characteristics of related storage hardware, this method cannot meet the requirement of realtime processing of highresolution images. The scheme of row FFTtransposedrow FFT can reduce the waiting time of direct memory access controller in the computation process and improve the computational efficiency of 2DFFT, but the existing implementation of matrix transposition has significant limitations. Traditional design uses load and store instructions to complete the transposition of a matrix. This paper proposes a 2DFFT scheme based on fast block transposition. By building a transposition module and a fourway parallel 1DFFT module, the FPGA onchip resources are fully utilized, thus the delay is reduced. The experiment is based on Xilinx Kintex UltraScale FPGA, and under the same clock frequency and parallel conditions, different 2DFFT calculation schemes are compared. Within the experimental error range, the solution proposed in this paper improves the computational efficiency by about 15 times.

Abstract:In the location of thunderstorm clouds, the measurement of the relative dielectric constant of the ground surface is very important. In order to solve the problem of continuous monitoring of surface media, this paper analyzes the measurement principle of the same surface plate capacitance sensor, builds the design model of the same surface plate capacitance sensor, and completes the sensor plate design based on the actual surface monitoring application scenarios; The Debye relaxation relation of dielectrics is introduced to find the best driving signal frequency suitable for surface dielectric measurement; Design and build the signal measurement circuit, introduce temperature and humidity compensation, and improve the measurement accuracy; The least square method is used to linearly fit the best data matching function. The experimental results show that the measurement error of the system can be controlled within 5%, which can effectively monitor the relative permittivity of the ground in real time.

Abstract:To realize the precise positioning and quick recognition of insulators on a line that transmits electricity.A rotatingg target detection algorithm was proposed based on YOLOv5（You Only Look Once v5）.A rotatingg retangular frame containing three types of labels was proposed for selfexploding insulators,It can improve the detection effect of the model and the convergence rate of the model.Use a Hardswish activation function to accelerate the computation and improve the numerical stability of the model. replace the Conv structure with the Ghost module structure,to reduce the number of parameters in the model and improve the detection efficiency of the model.Replace the CIOU Loss(Complete IoU loss) function with the EIOU Loss(Efficient IoU Loss) function and add a SLL(Smooth L1 Loss) about angle for regression of rectangular box angle,It can position the insulator more accurately and improve the detection effect of the model. The experimental results show that the improved insulator fault detection algorithm reduce the calculation amount of YOLOv5s model by 48.7% and the model size by 44.5%,Inference speed increase 2.9%,The overall detection accuracy of the model is 97.7%.In addition,it can also meet the realtime requirements of mobile deployment

Abstract:Aiming at the problem of poor identification accuracy of bridge moving load by integer Tikhonov regularization method (Tik), a bridge moving load identification method based on improved fractional Tikhonov regularization method (IFTik) is proposed. According to the theory of time domain method, the bridge moving load identification model is established to simulate the driving process of two axle vehicles on the bridge. The moving load is expressed as the differential form of moment response and acceleration response kernel function by kernel function method. The differential equation is transformed into linear equations by discretization method, which is solved by improved fractional Tikhonov regularization method. The results show that compared with FTik and Tik methods, IFTik method has certain advantages in recognition accuracy and noise resistance, and IFTik method has higher recognition accuracy at low noise level, and the recognition error is only 15% of FTik method; The identification result is less affected by the moment response, and has good robustness, which is more suitable for the field identification of bridge moving load.

Abstract:In order to solve the problem that the existing ground segmentation methods are inaccurate in the complex road surface and sparse point cloud scene, proposes a ground segmentation algorithm based on concave bag algorithm. This method first generates concave packets according to the lidar point cloud, then selects the ground triangle according to the difference of the points in the triangle surface extracted by rough filtering and the scanning characteristics of the normal vector of the triangle surface, and then accurately extracts the interior points of the ground triangle surface. The ground segmentation can be completed accurately according to the distance from the interior point to the triangle surface. The experimental results show that this method can fully consider the geometric characteristics around the point cloud, and is sensitive to the geometric boundary of the object. It can finely segment small bumps, kerbstones and other small obstacles in the scene with sloping ground.

Abstract:Aiming at the problems of low detection efficiency and poor accuracy caused by the variety of steel surface defects, strong background interference and various scale changes, this paper proposes a steel surface defect detection algorithm: CBEYOLOv5. On the basis of YOLOv5 algorithm, improve the ability of feature extraction by using coordinate attention mechanism in the trunk to strengthen the focus on the target; BiFPN is used as a special extraction network, and effective feature corresponding weights are given to fully integrate features of different scales. The model loss is calculated through EIOU, so that the model can be regressed more accurately. The experimental results on the public dataset NEUDET show that the CBEYOLOv5 algorithm mAP is 755%, 38% higher than YOLOv5, and the detection speed is also higher than some common target detection algorithms.，it can detect the defects on the steel surface more accurately and quickly.

Abstract:In the multi-UAV coordinated power inspection task, the existing path planning methods generally ignore the impact of signal quality, which makes it difficult to be effectively applied in a wide range of power inspection tasks. Therefore, aiming at the problem that the detection effect is degraded due to the limited communication, this paper proposes a path planning method for multiUAV cooperative power inspection under 5G signal constraint. Firstly, based on the 5G signal transmission characteristics, a propagation loss model for power inspection in large scale space is established. Then, based on the genetic algorithm architecture, a path planning method for multiple UAVs is proposed, which integrates the constraints of 5G signal quality, flight mileage and patrol objectives. Finally, the path planning method of multiUAV cooperative power inspection based on 5G signal is simulated and verified. The results show that, compared with the traditional methods, the flight length of the path with poor signal quality after the constraint is reduced by 45.2%, and the vehicles will give priority to the tower with strong signal when the distance difference is small, thereby improving the detection effect of the patrol task, so as to ensure the use in a wide range of environments.

Abstract:A five-leg inverter based on model predictive control (MPC) strategy has a problem of excessive common leg current. This paper takes the dual permanent magnet synchronous motor (dualPMSM) system driven by a fiveleg inverter as the research object, and a common leg current suppression method on the basis of MPC strategy is studied. Combined with the topology of fiveleg dualPMSM system and the algorithm architecture of MPC strategy, the influence of the phase difference of twophase current corresponding to common leg on common leg current is discussed. The common leg current suppression method based on the rotor flux linkage angle difference is applied to MPC strategy, and corresponding experiments are carried out so as to verify and evaluated current suppression effect. Experimental results show that the common leg current suppression algorithm based on MPC strategy can achieve a fine control effect whether under the balanced load or not.

Abstract:The traditional runway intrusion alarm equipment has the problems of low automation level and high cost of installation and maintenance. In this paper, the airport scene image information is obtained through the airport video system, and YOLOv5 is used to detect the airport scene aircraft. A lightweight network ShuffleNetv2 is used to optimize the DeepSORT algorithm to track the airfield aircraft. Through the monocular video acquisition system, the coordinate transformation and ranging model is established to accurately measure the distance between the airport aircraft and the runway midline. According to the ground protection zone, runway intrusion alarms can be realized by setting an appropriate threshold. The experimental results show that the average processing time of the optimized model is reduced by 2564%, the average ranging errors of aircraft 11, 18 and 43 cm from the runway center line in the simulated environment are 002, 001 and 001 cm, respectively, and the accuracy of runway intrusion alarm is 95.86%. The model has good realtime performance and high accuracy. This method can effectively prevent the occurrence of runway intrusion events.

Abstract:The purpose of blind image deblurring is to recover a clear image from a blurred image by iterating in the case where the blur kernel is unknown. When the real image has fewer dark pixels, the dark channel a priori algorithm does not produce satisfactory results. It is found that the absolute value of the elements in the secondorder gradient (Hessian matrix) decreases as the image is gradually blurred. Using this feature, a model of dark channel a priori blind image deblurring algorithm based on regularized secondorder gradients is proposed. Firstly, the theoretical proofs related to the algorithm are shown, and the feasibility of the Hessian matrix in preserving edge details and image details is experimentally illustrated. Then, a semiquadratic splitting strategy is used to solve the nonconvex optimization problem, and finally, the fast Fourier transform is used to obtain the final clear image and the blurred kernel. The experimental results show that the algorithm can well preserve the edge details and eliminate ringing artifacts while suppressing noise, and it is more robust and performs well than existing image deblurring methods on both synthetic and natural images. The SSIM values are improved by more than 10% on average in the natural image dataset.

Abstract:Sea-land segmentation is an important basis for the application of remote sensing images, such as coastline change analysis and resource management. Due to the complex scene of remote sensing images and uneven distribution of land size and shape, sealand segmentation is faced with problems such as misclassification and unclear boundary segmentation. Aiming at the above problems, proposes a gated pyramid fusion network for sealand segmentation in remote sensing images. Firstly, two deep features are aggregated through the attentioninduced crosslayer aggregation module to capture the global context and accurately and roughly obtain the size and shape information of the land. Then, the aggregated global features are sent to the gated fusion module, guided by the global information, useful context information is selected from the multiscale features, optimizes boundary details layer by layer and highlights the entire land area. Finally, global supervision is performed on each side output. Two sets of remote sensing images from different data sources were selected for experiments, the accuracy was 99.13% and 98.98%, the F1 score was 99.03% and 98.89%, and the mIoU was 98.26% and 9797%, respectively. Experimental results show that this algorithm has better segmentation effect than other algorithms.

Abstract:To solve the problems of insufficient bimodal feature fusion and low quality of feature fusion in multispectral pedestrian detection, a multispectral pedestrian detection algorithm based on multistage cross information fusion is proposed. Firstly, the algorithm extracts the features of visible and infrared images through the dual stream backbone network; The cross information fusion module is designed and embedded in the dual stream backbone network in multiple stages to guide the bimodal feature fusion, so as to achieve full fusion of bimodal feature information; Conditional convolution is introduced to dynamically process the fused feature information to improve the quality of the fused information and ultimately improve the detection performance of the algorithm. Experimental results show that the missing rate of the algorithm is only 1041%, which is 10% lower than the original algorithm, and the detection performance of the algorithm is significantly improved.

Abstract:Pupil location plays a key role in HumanComputer Interaction. However, due to the noise of reflection, blink and eyelash occlusion, as well as the pupils at the edge of image and blurred due to movement, the accuracy and robustness of pupil center location are reduced, which leads to great difficulties in accurate pupil location. Therefore, this paper proposes a pupil detection and location algorithm based on attention mechanism and cavity convolution. The algorithm is based on the codingdecoding structure of UNet, VGG16 is used in the coding part and dilated convolution is introduced to fully extract features, Attention Gate is added in the decoding part to make the model more robust. Then, the least square method is used to fit the pupil segmentation map output by the network. Finally, the pupil center coordinates are obtained according to the fitted image. The algorithm is verified by 24 data sets that are publicly available. Experiments show that the algorithm can accurately locate the pupil position, and the average detection rate can reach 92.6%.

Abstract:Efficient and accurate detection of small targets in dense scenes is a key problem in the field of target detection. In order to solve the problems of diversity of environments and complexity of small targets, such as difficult feature extraction and low detection accuracy, a small target detection method for dense scenes combined with TCYOLOX is proposed. Firstly, by introducing Transformer Encode module into CSPNet, the target weight is continuously updated to enhance the target feature information and improve the network feature extraction capability. Secondly, the convolutional attention mechanism module is added to the feature pyramid network to focus on important features and suppress unnecessary features, so as to improve the detection accuracy of targets of different scales. Then, CIoU is used to replace IoU as the regression loss function, which makes the network converge faster and has better performance in the process of model training. Finally, it is verified on PASCAL VOC 2007 dataset. The experimental results show that the designed TCYOLOX model can effectively detect small target objects under normal, dense, sparse and dark conditions in diversified scenes. The mAP and detection speed can reach 946% and 38 fps, which is 109% and 1 fps higher than the original model. It has good applicability to small target detection tasks in multiple dense scenes.

Abstract:Aiming at the problems of low measuring efficiency and large error in the current process of measuring loader loading and unloading action time, a measuring system of loader loading and unloading action time based on machine vision is designed. Firstly, the video image sequence of loader loading and unloading motion is preprocessed, and then the shape template matching and motion prediction are used to find the position of the bucket in the image and obtain the center of mass pixel coordinates of the bucket. The motion track of loader bucket is established by using the pixel coordinates of the bucket′s center of mass. Then, the trajectory dividing points of the four loading and unloading motion stages are located, and the starting and ending frames and their frame numbers of the four motion stages are determined by using the trajectory dividing points. Finally, the frame number of the start frame and the end frame is converted into time by frame rate for calculation, so as to realize the automatic measurement of the loading and unloading action time of the loader. The experimental results show that the measured deviations of the measuring system for the loading and unloading action time of the loader is less than 05 s, the measuring results is stable and effective. On the other hand, because motion prediction is used to estimate the bucket position, the time consumption for shape template matching is reduced by 1645%.

Abstract:Traditional methods for remaining useful life（RUL）prediction of mechanical equipment require manual intervention processes such as multisource data fusion and establishment of health indicators, and the prediction accuracy is limited by the ability of health indicators to characterize the degradation process of the equipment. To achieve endtoend RUL prediction and improve the prediction accuracy, a RUL method based on a combination of attention mechanism and residual depth separation convolutional network is proposed, and the effectiveness of the method is tested by using the CMAPSS aeroengine simulation data set. A sliding window is used to intercept multivariate sequences from the engine multisource state parameters as samples to characterize the engine state, and a RUL prediction model is built based on a onedimensional separable convolutional network, and an attention mechanism and residual network are introduced into the network to improve the prediction accuracy of the model. The final mean root mean square error values of the proposed method for the four test sets of C-MAPSS are 11.28, 14.12, 11.57 and 15.61, respectively, and it also has good generalization capability for RUL prediction during engine operation. The comparison results with various RUL prediction methods show that the overall prediction accuracy of the proposed method is high for all four test sets, indicating that the method is an effective RUL prediction method for mechanical equipment and can be used for early fault warning of equipment.

Abstract:The damage of the outer sheath of the cable at the industrial site mainly relies on manual inspection, which consumes manpower, is subject to high subjectivity, and is prone to blind spots. The realtime performance is poor and the manual inspection of some industrial sites is more dangerous. Aiming at a series of problems caused by manual inspection, this paper proposes an improved Faster RCNN cable sheath damage detection method. In order to improve the generalization ability of the model, grayscale, flip, pan, and sharpen the collected training set are used for data enhancement; use the feature extraction network RseNet50 with fewer parameters and deeper layers to replace the original VGG16 as the backbone feature extraction network; use migration learning to use the weights trained on the ImageNET dataset as the initial weights of the model; use bilinear interpolation to replace the ROI Pooling operation; use the Kmeans clustering algorithm to analyze the original data Cluster analysis was performed on the collection, the Silhouette method was used as the evaluation standard, and the anchor frame of the outer sheath damage detection was customized based on the clustering results. Experimental results show that the improved Faster RCNN has an average accuracy (mAP) of 8833% for the detection of damage to the outer sheath of the cable, which is 549% higher than the original Faster RCNN, and is better than the classic SSD algorithm and YOLOv3 algorithm. The improved detection speed achieve 036 frame/s to meet the testing requirements. This model can be subsequently equipped with various mobile detection platforms and has high engineering value.

Abstract:In order to improve the accuracy of rolling bearing fault diagnosis effectively, a method of bearing fault diagnosis based on the combination of complete ensemble empirical model decomposition with adaptive noise, bubble entropy and support vector machine is proposed. Firstly, a series of intrinsic modal function components were obtained by CEEMDAN. Then, the important IMF components was chose through the chart and calculate it. Fault feature vectors were constructed and input into the SVM optimized by arithmetic optimization algorithm to train for bearing fault classification. The results show that the accuracy of this method is up to 992% which is 28% higher than that of GASVM. It can also successfully identify the single fault and compound fault of rolling bearing, so it can be used for bearing fault classification.

Abstract:For the ultra-low frequency dielectric loss factor measurement method, due to the low frequency of the measured signal, the sampling time is long and the amount of data collected is large. In addition, in the case of asynchronous sampling, the FFT has spectrum leakage and fence effect, which affects the accurate measurement of the dielectric loss factor. In order to reduce the sampling time and the amount of data collected, as well as the spectrum leakage and fence effect during asynchronous sampling, an ultralow frequency dielectric loss measurement method based on Prony algorithm quasi synchronous sequence is proposed. Prony algorithm and data identification method are used to estimate the fundamental frequency of the sampled voltage signal, and Newton interpolation algorithm is used to realize the quasi synchronous interpolation reconstruction of voltage and current signals, The quasi synchronous sequence of the sampled signal is obtained, and the quasi synchronous sequence is solved by FFT and dielectric loss equivalent circuit model to obtain the ultralow frequency dielectric loss factor. The dielectric loss factor is measured under the conditions of frequency fluctuation, harmonic content change, dielectric loss angle change and noise with different signaltonoise ratio. The simulation results show that the method realizes quasi synchronous sampling in software, effectively reduces the influence of fence effect and spectrum leakage on dielectric loss factor measurement, and is suitable for accurate measurement of ultralow frequency dielectric loss factor due to short sampling time, small amount of collected data and high measurement accuracy.

Abstract:In order to solve the timing sequence and multistage problems in industrial production process, a fault detection method based on orthogonal local preserving projection(OLPP) and adaptive timing sequence window weighted k nearest neighbor (ATSWKNN) was proposed. Firstly, basing on maintaining the sample nearest neighbor relationship, the original data are projected into the lowdimensional feature space by using OLPP. Secondly, a certain kind of timing window is selected in the feature space, and the timing sequence square distance is calculated. Then, the reciprocal of the average cumulative square distance between the sample in the window and its spatial nearest neighbor set is taken as the weight. Finally, statistics are constructed to monitor the process. OLPPATSWKNN reduces the autocorrelation of process and solve the problem of multistage statistical difference by extracting time series information and introducing weight within the window. In addition, the problem of abnormal statistical indicators during phase switching is solved by adaptive window switching strategy. The monitoring performance of OLPPATSWKNN is verified by monitoring the numerical simulation process and penicillin fermentation process, and the monitoring results are significantly better than the classical methods.

Abstract:Aiming at the problems of motor bearings being prone to failure, the traditional fault diagnosis method has long time, low diagnostic accuracy and many adjustment parameters, and this paper proposes a bearing fault diagnosis method for support vector machine SVM optimized by improving sparrow algorithm ISSA. The classification algorithm introduces improved Tent chaos mapping, flock algorithm random following strategy, adaptive t distribution and dynamic selection strategy in the traditional sparrow optimization algorithm, and first uses CEEMDANenergy entropy to decompose the vibration signal, selects the energy entropy values of the five IMF components with the greatest correlation with the original signal as the eigenvector, and then inputs it to the ISSASVM classifier for bearing fault diagnosis. Experimental comparison with PSOSVM、GWOSVM and SSASVM classification models shows that the diagnostic accuracy of the ISSASVM diagnostic model can reach up to 100%.