Abstract:In order to improve the grid-connected stability of wind power converter, the harmonic and resonance problems of the grid-connected interface of the converter are solved when the load fluctuates. This paper presents an improved active damping linear active disturbance rejection control (ADLADRC) strategy. Firstly, the mathematical model of LCL converter is derived. Traditional active disturbance rejection control technology is analyzed. On this basis, to improve the observation ability of traditional observer, an improved active disturbance rejection control based on series filter in the total disturbance channel of the observer is designed. And the introduction of active damping and its combination to complete the improved ADLADRC control strategy design. Then the frequency characteristics of the converter system under the improved ADLADRC are analyzed by frequency domain analysis method. The improved ADLADRC control has better grid-connection stability and harmonic resonance suppression. Finally, through simulation, the proposed control strategy is compared with the current waveform of traditional LADRC and traditional PI control. Under the steady-state condition, the simulation results show that the full-load harmonic rate of the improved ADLADRC is 2.89% lower than that of PI to 0.39%, and the half-load harmonic rate is 7.64% lower than that of PI to 0.60%. It indicates that the proposed control strategy not only has better grid-connected stability, but also has fast dynamic response and harmonic suppression when the load fluctuates.

Abstract:Aiming at the problems of high delay, high power consumption and less defect categories in lens defect detection using image processing method and neural network method, a software and hardware collaborative detection system based on FPGA and degraded YOLO is designed. In the system, the convolution layer is used to replace the reordering layer of the YOLO network for network degradation and mapped to FPGA; dynamic quantization, module fusion, double buffer pipeline, loop expansion and block segmentation optimization strategies are adopted to design dynamically configurable acceleration IP. The convolution calculation module implements fast convolution algorithms based on Winograd and GEMM respectively. The experimental results show that the acceleration IP of this system obtains the calculation performance of 51.89GOP/s on PYNQ-Z2, which is 0.76 times higher than that based on the convolution calculation method of typical sliding window. The time delay of the acceleration single image is 433ms, and the power consumption is 1.07W, compared with Core i5-10500 CPU, the energy efficiency is 365.27 times higher, which realizes the multi-defect detection of low delay and low power consumption of mobile phone lens by small equipment.

Abstract:In traditional thyroid surgery, it is difficult for clinicians to effectively determine the position of the recurrent laryngeal nerve by visual recognition. It is prone to recurrent laryngeal nerve injury, which lead to patients unable to speak normally and even endanger lives. Intraoperative nerve monitoring (IONM) techniques is useful for effectively identifying nerve signals and thus avoiding intraoperative nerve injury. This paper introduces the design and implementation of IONM in detail. The system consists of main control module, current stimulation module and EMG acquisition module. The main control module adopts ARM+FPGA dual-core architecture to realize the current stimulation module and signal acquisition module. At the same time, the data collected by the acquisition module is processed. The current stimulation module uses high-precision D/A conversion chip as the core to design the current source circuit, which is used to control the intensity and pulse width; The acquisition module adopts high-precision A/D analog conversion chip as the core and adapts low noise amplifier circuit to realize the acquisition of neural EMG signal. The recurrent laryngeal nerve experiment and clinical experiment were carried out in this system respectively. It is proved that this system has ability effectively to stimulate recurrent laryngeal nerve signal and recognize nerve EMG signal, which meets the needs of clinicians.

Abstract:To address the problems of difficult feature extraction and low classification accuracy in soft fault diagnosis of DC-DC circuits, a fault diagnosis method based on multi-strategy improved Harris optimization algorithm-back propagation (MHHO-BP) neural network is proposed. The method processes the fault signal by VMD, extracts its time-domain and frequency-domain features as the fault vector, and uses the MHHO algorithm to optimize the weights and thresholds of the BP neural network to establish a VMD-MHHO-BP soft fault diagnosis model of DC-DC circuits. The experimental results show that for soft faults of DC-DC circuits, the method has good diagnostic effect and high accuracy compared with the Whale Optimization Algorithm (WOA) and Butterfly Optimization Algorithm (BOA) optimized BP neural network.

Abstract:To solve the problem of noise in the ceramic percussion detection signal, a denoising method combining VMD optimization and wavelet packet denoising (WPD) was proposed. Firstly, the energy starting point detection criterion extracted the effective information of the actual signal; Secondly, the genetic algorithm (GA) selected the VMD parameters to adaptively decomposed with the noise signal. Then, by calculating the correlation coefficient of each modal component with the original signal, they were divided into the main signal component and the noise component. Finally, the main signal component was de-noised by wavelet packet analysis. And the information is reconstructed to obtain the original signal. The simulation experiment proves that the method has the highest signal to noise ratio (23.81dB, 24.75dB) and the lowest mean square error (0.07, 0.01) when adding 10dB and 20dB noise, respectively, and the denoising effect is significantly improved compared with usual denoising methods. The test experiments of ceramic specimen percussion detection signals show that the method can remove the noise of different types of ceramic percussion detection signals effectively and has good denoising performance.

Abstract:Aiming at the problem that the pulse signal is nonlinear, non-stationary and difficult to denoise, a joint denoising method based on improved adaptive noise set empirical mode decomposition (ICEEMDAN) and wavelet packet decomposition (WPD) is proposed to denoise the collected pulse signal. Firstly, ICEEMDAN mode decomposition is performed on the noise signal to generate a series of intrinsic mode functions (IMF). Then these IMF components are calculated with the correlation coefficient of the original signal respectively, the value of the correlation coefficient is compared, and then the signal is reconstructed. Finally, the reconstructed signal is decomposed by wavelet packet to extract the denoised pulse signal. The simulation data and the actual pulse signals are used for experimental analysis. The method is compared with the ensemble empirical mode decomposition (EEMD), and the signal-to-noise ratio (SNR) and root mean square error (RMSE) of the two methods are compared. The experimental results show that the joint denoising method based on ICEEMDAN-WPD can remove the noise more effectively and preserve the characteristics of pulse signal better.

Abstract:Hydrophobicity Class (HC) is one of important indexes to measure the performance of composite insulators. The hydrophobicity of insulator shed is different on the part surface for the various factors in the natural environment. In order to judge the performance of insulator, this paper proposes a region adaptation method for insulator segmentation and hydrophobicity recognition based on deep learning. First, separate the insulator area and the background by the insulator segment module, which provides segment information for the later operators on the insulator area; then, the insulator area is cropped into several image blocks with the fixed resolution, which can reduce the resolution and the operational complexity while preserving the insulator surface details; finally, judge the hydrophobicity class of insulator by the hydrophobicity classification module. The experiment dataset from maintenance sites is used to build model in stages and evaluate separately the accuracy of the stage of segment and HC classification. The experiment results show that the segment stage module can identify the insulator regions and the background, whose accuracy on the cross-validation test dataset is greater than 97.21%, and the HC classification stage module can classify the HC of insulators, whose accuracy of 140 test images can reach 98.65%. The proposed model is proven to be an effective solution to checking insulators performance in complex natural environment by experiments.

Abstract:Reliability assessment of photoelectric conversion system can provide valuable reference for photovoltaic power station planning. Hence, this paper proposes a reliability assessment model of photoelectric conversion system based on Markov chain, which applies empirical criteria to classify temperature-irradiation state and construct its Markov chain, and then uses stress analysis method to calculate the failure rate of photoelectric conversion system, and then forms the Markov chain of photoelectric conversion system failure rate. Accordingly, the reliability indexes such as average failure rate of photoelectric conversion system are calculated, and the elastic coefficients of temperature and irradiation on the average failure rate of photoelectric conversion system are defined, which to quantify the sensitivity relationship between two meteorological factors on average the failure rate of photoelectric conversion system. Collecting the measured temperature and irradiation data of several observation stations in North Dakota, and evaluating the reliability index of a photoelectric conversion system at different observation stations. The results show that the average failure rate of the photoelectric conversion system at low latitude observation stations is higher, and irradiation elasticity coefficient for the average failure rate of the photoelectric conversion system is higher than temperature elasticity coefficient, the average failure rate of the photoelectric conversion system is more sensitive to irradiation than temperature.

Abstract:This study presents a design of a non-zero speed start-stop and low-power stepper motor controller. To achieve such design, the acceleration-deceleration curve algorithm for the realization of non-zero speed start-stop, and the hardware-logic design for the realization of low-power dissipation, have been proposed to improve the performance of the stepper motor open-loop control method. The acceleration-deceleration curve algorithm is able to divide the stepper motor rotation process to 4 different modes so that to establish any desired linear motor velocity profile, to tackle the problem of achieving the performance of non-zero speed start-stop. In this paper, the control pulse periods of the 4 different linear acceleration-deceleration process are theoretically derived; Then the hardware-logic model of the acceleration-deceleration process is optimized with the assistane of the Pipeline Design concept. The IP core of the stepper motor controller is built on a Field Programmable Gate Array (FPGA) , and ultilize the low-power ICs (e.g., Clock-Gating Technique) to achieve the low-power performance of the controller. Finally, an experimental platform is setup to demonstrate that, the IP core on FPGA is capable for the non-zero speed start-stop with 4 modes, and is able to drive the motor timely and precisely. The empirical results with statistical analysis show the proposed design is feasible and efficient, and successfully achieve 20% better control performance, 30% less electronic system circuit area, and 53% less power dissipation.

Abstract:In view of the lack of ore data set and ore classification and recognition model, a lightweight ore classification model algorithm based on improved mobilenet V2 is proposed, which takes the ore image imaged by X-ray irradiation as the data set and mobilenet V2 as the main network. Firstly, by adjusting the expansion factor and width factor, the amount of model parameters is greatly reduced to realize the purpose of model lightweight; Secondly, the efficient channel attention mechanism is embedded in some inverse residual modules and the original model classifier, and the residual inverse residual module is replaced by a parallel feature extraction network with deep hole convolution, so as to enhance the ability of model feature information extraction and improve the accuracy of model recognition; Finally, the training method of transfer learning is used to initialize the weight and accelerate the model training. After improvement, the ore recognition accuracy of the algorithm is improved to 96.720%. Compared with vgg16, googlenet, xception, shufflenet and mobilenet V2, the accuracy and ore detection speed have been improved. In general, compared with other algorithms in this experiment, the improved algorithm has better performance for ore recognition.

Abstract:Spiral search mechanism has strong global search ability and is widely used in firefly and whale search algorithms, but its convergence speed is slow, the convergence accuracy is low, and the local search ability is poor. By changing the search mode with small convergence range, a local spiral search is proposed to improve its local search ability, and a mutation operation is introduced to improve its local search ability, and a planet search algorithm is proposed. The algorithm is verified by single - peak and multi - peak test functions. The results show that the planet search algorithm is better than particle swarm optimization, firefly algorithm and whale search algorithm in convergence speed, search accuracy and local search ability.

Abstract:In order to carry out underwater detection, an underwater robotic fish with two caudal fins is designed. Due to the uncertainties of model parameters and the disturbances of complex water waves, a dynamic model with uncertain parameter factors is established to obtain a better control effect. A radial basis function (RBF) neural network sliding mode controller is designed according to the dynamic model. The stability of the system is proved by using the Lyapunov function. The simulation results show that the designed controller is insensitive to the changes of the parameters of the dual-fin underwater robotic fish and has higher control accuracy and stronger robustness than the traditional PID controller. This method provides a reference for the design and application of fin-actuated underwater robot in the future.

Abstract:The traditional network clock protocol supported by the current train communication network can only achieve sub-microsecond time synchronization accuracy, which cannot meet the needs of the time synchronization accuracy of each node of the current train. Aiming at the above problems, this paper proposes to apply the precise clock protocol to the train communication network. In order to achieve high synchronization accuracy under the condition of low offset range, based on the traditional PI control algorithm, this paper proposes a multi-model PI control optimization algorithm in the improved clock servo system, and sets the threshold limit values of the proportional coefficient KP and the integral coefficient KI, The quantitative relationship between the threshold limit value and the PI output compensation value is derived to compensate for the offset of the slave node. Finally, taking a train communication network scene as the research object, modeling and simulation is carried out on the OMNeT++ simulation platform, and the offset value of the master and slave nodes is analyzed. Compared with the offset value obtained by the traditional PI control algorithm, the improvement value of the offset of each node in the train communication network is all within 30ns, and the offset can be as low as 1.38ns, which verifies the superiority of the proposed algorithm.

Abstract:Epilepsy is one of the most common life-threatening neurological diseases. Epileptic EEG signals are complex and diverse, and manually scanning long-time EEG signals is commonly time consuming, error prone with low consistency between physicians, while the raw clinical EEG data involving noise and artifacts reduce the performance. Thus, it is important to develop a reliable, effective and stable automatic seizure detection technology based on EEG signals, to reduce physicians’ burden. Here, using the raw clinical EEG data from Chinese 301 Hospital, we introduced a novel one-dimensional convolutional neural network with a successive double-convolutional structure, to achieve high performance on seizure detection, with the sensitivity, specificity, accuracy and F1-score reaching 96.8%, 99.8%, 99.6% and 96.1%, and only using a third or half GPU time for training. The results show that the introduced neural network model based on one-dimensional convolutional neural network is superior to the existing methods, and achieves reliability, efficiency and stability on seizure detection, which is of great significance to the auxiliary diagnosis of epilepsy.

Abstract:Using multiple UAVs to cooperate in reconnaissance missions can effectively improve the accuracy of reconnaissance. The importance of different reconnaissance mission targets is often different, and their mission value is also different. Therefore, it is necessary to reasonably allocate the UAV resources for cooperative reconnaissance and improve the efficiency of cooperative reconnaissance. This paper focuses on the allocation of UAV reconnaissance time resources. Firstly, an autonomous collaborative resource allocation mechanism is constructed. Taking the assisted UAV as the leader and the auxiliary UAV as the follower, the Stackelberg game model is established. Then, by solving the lower level game equilibrium and the upper level game equilibrium, the closed expression of the optimal assistance time of the auxiliary UAV is deduced. The Nash equilibrium solution of the Stackelberg game model is obtained. Finally, the proposed model and method are verified by simulation. The simulation results show that the proposed method makes full use of the time resources of auxiliary UAV. The effectiveness of cooperative reconnaissance has been effectively improved.

Abstract:In order to realize the automatic classification of skin burns of the wounded after suffering a major fire and other disasters and speed up the diagnosis efficiency, a lightweight model BI-YOLOv5 algorithm for skin burn classification was proposed. Replace the Swish activation function to improve the convergence ability and detection efficiency of the model; use the K-means++ algorithm to perform cluster analysis on anchors to enhance the adaptability to targets of different scales; modify the feature extraction network to extract feature information of multiple scales and establish multi-scale features The fusion network improves the utilization rate of the deep feature information by the model and improves the recognition accuracy of small-area burns. The experimental results show that the BI-YOLOv5 algorithm has high accuracy and efficiency in detecting and distinguishing different burn types and environmental disturbances, and the mAP reaches 97.6, which is 8.4 percentage points higher than that of YOLOv5.

Abstract:In order to solve the problem of detecting the text information of traffic signs by the external environment perception system of unmanned vehicles, a two-stage method for detecting and recognizing the text information of traffic signs in an autonomous driving scenario is proposed, which realizes the refined collection of autonomous driving information. First, use the YOLO detector to detect traffic signs. At the same time, use the improved DB detection network in this article to detect the text in the scene. The intersection of the traffic sign detection results and the scene text detection results to get the text area to be recognized; finally, the lightweight CRNN network is used to treat Recognize area text for recognition. Use CSCT-1600 data set and MTWI-2018 data set for training and testing respectively. The experimental results show that the accuracy of the traffic sign information positioning algorithm is 94.95% when the recall rate is 92.98, and the recognition speed of the traffic sign information recognition algorithm is 25 frames when the F1 is 77.2%.

Abstract:In order to detect malicious PDF and DOCX format documents more accurately and quickly, a visual detection method of malicious documents based on deep learning is proposed. This method converts the byte stream of the document into a three-channel color image through the Markov model, so as to obtain a visual representation that can better distinguish between malicious documents and benign documents, and uses the current mainstream EfficientNet-B0 model to extract visual features to classify. Combined with the fine-tuning technology in the field of transfer learning, the classification weights on ImageNet are applied to the training of the EfficientNet-B0 model, which speeds up the convergence of the detection model and shortens the training time of the model. Experiments show that on two datasets, the convergence speed of the model is faster than the pre-training of random initialization weights, and the detection accuracy of the model for malicious PDF documents and malicious DOCX documents reaches 99.80% and 98.14%, respectively, which is better than models such as ResNet34 and MobileNetV2.Compared with the mainstream malicious document detection tools Wepawet and PJScan, the proposed method has better comprehensive detection performance, which further verifies the effectiveness of the proposed method for malicious document detection.

Abstract:Aiming at the problems of low accuracy and low efficiency of complex human motion recognition in video, a dense connection network model for spatio-temporal feature extraction is proposed. Firstly, two dense connected networks are used to extract spatiotemporal features; Secondly, the dense connection between spatiotemporal networks is constructed, and the feature information extracted from the spatiotemporal network is input into the spatial flow network layer by layer to improve the spatiotemporal interaction between the two flows; Then the LSTM network is used to process the characteristics of the two stream network respectively, and the prediction results of the two streams are obtained; Finally, the prediction results of dual stream network are fused to realize the recognition of complex behaviors in video. The comparative experiments on ucf101 and hmdb51 benchmark data sets show that the accuracy rates of 94.69% and 68.87% are better than other algorithms. Experiments show that this model can increase the interaction between spatiotemporal networks and is conducive to the recognition of complex human actions.

Abstract:Leaf area index (LAI) is an important parameter for studying forest ecosystem and vegetation canopy structure. It is an important work in forestry engineering to measure LAI efficiently and accurately. The traditional LAI measurement method requires manual hand-held instruments to make on-site measurement, which is time consuming and laborious. In recent years, with the development of Internet of Things, the technology of using wireless sensors to measure LAI has gradually become mature, but some problems still need to be solved. This paper proposed a LAI measurement method based on digital infrared hemispherical photography (DIHP), and designed an adaptive segmentation algorithm for the color space of infrared photography, which was deployed on the edge computing platform "Raspberry Pi" to solve the problem that traditional digital hemispherical photography (DHP) methods are prone to environmental interference. The measurement results of the sensor designed in this paper are significantly correlated with the hand-held vegetation canopy analyzer HM-G20, with an R value of 0.99691 and an average measurement accuracy of 93.57%, which is 13.85% higher than that of DHP. The DIHP sensor has low operating power consumption, which meets the requirements of long-term field deployment of forestry Internet of Things and has great application prospects.

Abstract:Aiming at the problems of excessive parameters, insufficient real-time performance and weak detection ability of small targets in the existing algorithms for bird's nest detection on high-voltage tower, an improved YOLOv4 algorithm is proposed. Firstly, Mobilenetv2 network is used to replace CSPDarknet53 network as the backbone network, which reduces the amount of parameters of the algorithm and improves the detection speed. At the same time, the Coordinate Attention module is embedded in the inverse residual network of Mobilenetv2 network, which enhance the ability of the network to extract target features. Then, the PANet network is improved to obtain more detailed feature information and improve the detection ability of small target bird's nest. Finally, the Focal Loss function is used to optimize the loss function, reduce the weight of a large number of simple background samples, and improve the focus on the difficult sample training of small target bird's nest, which further improves the detection ability of small target bird's nest. The experimental results show that compared with the original YOLOv4 algorithm, the parameters of the improved YOLOv4 algorithm are reduced by 48.1%, and the detection speed and accuracy are improved by 12.9fps and 2.33% respectively. That is, the improved YOLOv4 algorithm greatly reduces the amount of algorithm parameters, and has better detection performance for bird's nest detection.

Abstract:In modulation recognition, the feature information of a single image is insufficient, the degree of discrimination is not high enough, and the recognition range is limited. In this paper, a modulation recognition feature enhancement method based on the feature fusion of time-frequency map and constellation map is proposed. The deep learning neural network is used to extract the features of signal image and construct the feature space. Through multi-dimensional feature fusion, the advantages of different features are mined and integrated to enhance the robustness of the model algorithm. In addition, the method of model migration is used, which only needs to train the classifier, which greatly saves the training time and resources, and has strong real-time and practicability. The simulation results show that under the condition of about 0 dB, compared with a single feature image, the average recognition rate of the signal can be improved by about 25% by using the feature fusion enhancement method. Through the model migration, the training of convolutional neural network is omitted, and the training time required is about 10% of that before migration, and the memory consumption is about 7.3% of that before migration. At the same time, the loss of recognition rate of the model is controlled within 5%.

Abstract:In this paper, a windmill shaped pattern reconfigurable element and its two-dimentional (2-D) wideband wide-angular scanning phased array antenna as an array element are proposed. The proposed single feed pattern reconfigurable unit antenna is composed of radiation patch, DC bias circuits and broadband artificial magnetic conductor (AMC) reflector. The radiation patch composed of four Vivaldi slots is a windmill-shaped patch with a reconfigurable feeding structure. By electrically controlling the PIN diodes integrated in the feeding network, the radiation beam of the element can be switched towards four endfire directions. Besides, the fan-shaped AMC surface is loaded backward the radiation patch. In this situation, the maximum radiation direction is titlted from the original endfire direction into a quasi-endfire direction, which benefits that the main beam of the planar phased array covers the broadside direction. The antenna unit and its construction 8 × 8 uniform planar phased array is simulated and analyzed. The simulation results show that the designed planar array has the performance of wide-band and wide-angle two-dimensional beam scanning,which supports its maxmum scanning angle of ±60° in two main planes from 5.4 to 6.1 GHz. Meanwhile, the gain fluctuation of the array is less than 4.3 dB, and the sidelobe level is quite low.

Abstract:In the actual direction finding system, due to the inconsistency of the transmission characteristics of the radio frequency parts such as multi-channel microwave components, antenna elements and connecting cables, the mismatch of the receiving channels is inevitable. When there is channel mismatch, the received data of the antenna array contains amplitude and phase errors, which eventually lead to a sharp decline or even failure of the estimation performance of the direction of arrival (DOA) of the spatial spectrum. Aiming at this problem, a channel correction algorithm based on a single auxiliary source is used to study the application of this algorithm in channel data pre-compensation and steering vector correction, and combines the multiple signal classification algorithm (MUSIC) for DOA estimation. The effectiveness of the correction algorithm is verified by the simulation parameters such as signal-to-noise ratio, number of snapshots and root mean square error. The measured data of the hardware platform is used to verify that the correction algorithm can effectively compensate for channel mismatch and obtain high DOA estimation performance.

Abstract:Wireless sensor network (WSN) is a distributed self-organizing network composed of a large number of sensor nodes. The sensor nodes are powered by batteries, and their limited energy seriously affects the life cycle of the network. In this paper, we introduced a new heterogeneous awareness routing protocol to elect a more reasonable cluster head, which comprehensively considers the residual energy of the node, the density of neighbor nodes and the relative distance to the base station within the coverage radius in clustering stage. We used the MATLAB R2016b simulator to analyze the performance of ED-SEP protocol under the condition of changing the base station location and the initial energy of the node, and it is compared with the stable election protocols. Simulation results show that the network’s stable period of ED-SEP protocol is increased by 29.6% and 25.7% respectively compared with SEP protocol and E-SEP protocol, and the data receiving capacity of base station is nearly three times that of E-SEP protocol. Therefore, the ED-SEP algorithm can effectively improve the energy utilization rate of the network and greatly improve the performance of the wireless sensor network.