Search

Even current OCR technology has made significant progress with the help of deep learning, rapid identification algorithms for specific information in documents still need elaborate design. In this paper, a lightweight operating framework based on deep neural network is established to deal with the extraction of date and time information from paper worksheets used in transformer station. In the framework, YOLOv4-Tiny is employed to detect the locations of the dates and the time in document images which are mixed with printing and handwritten fonts. Since the handwritten dates in pictures are difficult to segment, a novel lightweight CNN named MobileDateNet is proposed that can recognize the date serially without segmentation. The experimental results show that the identification accuracy of the whole framework reaches 92.1% at the average running speed of 36.5 FPS on an NVIDIA 3090 GPU, ensuring the recognition accuracy and the operating efficiency.

In order to realize the navigation for the solar L2 point halo orbiters, this paper proposes an autonomous navigation method based on LiAISON method. The method can determine the absolute positions of the geosynchronous orbiting spacecraft and the Halo orbiting spacecraft only by measuring the scalar distance between them. In this paper, the Unscented Kalman filter (UKF) is used. Under the condition of neglecting the process noise, the final positioning accuracy is in the same order with the measuring accuracy. This paper also studies the applicable conditions of the LiAISON method, and also studies the orbital configuration of the spacecraft constellation which can use this method. Fianlly, the paper analyzes and explains the results of experiment.

Human sensing technology such as human behavior recognition has gained tremendous attention in recent years due to the continuous development of artificial intelligence technology. With the development of wireless communication technology, WiFi based human sensing technology has received extensive attention due to its advantages of high popularity, low cost, difficulty in revealing user privacy, and being protected from light. In this paper, we proposed a human behavior recognition model based on DWT-SVM. Firstly, the channel state information (CSI) amplitude data is analyzed by wavelet transform, and the low-frequency wavelet coefficients after discrete wavelet transform (DWT) are selected as feature vectors. The low frequency wavelet coefficients of all channel data of CSI are extracted and spliced into a feature vector. Then a support vector machine (SVM) classifier is designed to realize human behavior recognition. Finally, the performance of the algorithm is verified on the experimental platform developed in this paper, and the performance of the proposed method outperforms other algorithms.

Object detection is a vitally important fundamental task in computer vision, and is applied in a number of scenarios including autonomous driving, robots, and industrial detection. Since 2014, object detection has become a research focus with the introduction of deep learning, which attracted a lot of attention. Even though object detection has developed rapidly in recent years with the power of deep learning, the ideas and motivations of many new models are still inspired by traditional algorithms. Therefore, it is necessary to understand the technical evolution of object detection to grasp its future direction. This paper provides a short survey on object detection, focusing on a comprehensive review of the technical development history in the past two decades. Secondly, the characteristics of each development period are summarized and some research hotspots in recent three years are given out. Then, the typical applications of object detection are illustrated. Finally, the paper discusses the possible challenges and development directions of object detection in the future.

As the most potential energy storage system at present, advanced adiabatic compressed air energy storage can better realize the regional comprehensive energy system’s absorption of new energy, efficient utilization of traditional energy and reduction of environmental pollution. In this paper, an energy supply optimization model of regional integrated energy system with advanced adiabatic compressed air energy storage is established to realize the economic optimization of the system’s electric heating energy in the daily operation cycle. By studying the energy storage characteristics of advanced adiabatic compressed air energy storage and its relationship with regional integrated energy system, the electric power balance, thermal power balance, operation constraints and environmental costs of various energy supply devices are analyzed. Gams optimization software is used to solve the operation costs of the traditional regional comprehensive energy system model and the model established in this paper in the daily operation cycle. Different models are compared from two aspects of environmental protection and economy, which verifies the effectiveness and feasibility of the model in reducing costs, promoting new energy consumption and energy conservation and emission reduction.

As a new energy vehicle, fuel cell hybrid electric vehicle (FCHEV) has a broad development prospect and provides a new way of thinking to solve the energy crisis. Energy management strategy is one of the key issues in the research of FCHEV. There are many power sources of the driveline system. Effective management of energy flows in power sources is the key to improving vehicle performance. Aiming at the FCHEV, an equivalent fuel consumption minimum control strategy (ECMS) was proposed to achieve the optimal fuel economy and real-time operation. Firstly, the fuel cell hybrid vehicle is modeled based on ADVISOR. Secondly, the online real-time energy management strategy ECMS is designed. Finally, the results of the NEDC and UDDS cycle are compared with the global optimal results of offline dynamic programming (DP). The results show that compared with the global optimal DP hydrogen consumption, the ECMS control strategy in this paper is only 0.59% and 2.10% different, but 13.31% and 12.28% higher than that of the rule-based CD-CS control strategy. In the following hardware loop experiment, the ECMS energy management strategy also improved by 5.00% and 5.64% compared to the CD-CS results. It is proved that the ECMS strategy in this paper has good control effects.

Recently, industrial robots turn out to be an urgent need in 3C (Computer, Communication, and Consumer Electronics) products industries that are characterized by dense labor and increasingly high product quality. However, the high dexterity, high complexity, and high precision requirements of the 3C assembly process still keep the robots out. One of the key issues is how to obtain the high accuracy pose (position and orientation) estimation of the assembly parts, which possess few features. In terms of 3C assembly applications, a pose estimation algorithm that takes point clouds as the data source to estimate the pose of the assembly parts is proposed in this paper. Given the CAD model of the target object, a virtual 3D sensor method is used to generate a virtual 3D view from any viewpoint around the CAD model’s point cloud. Then an ICP based particle filtering method, which takes the generated virtual 3D view as a reference, is carried out to perform the point cloud registration process. The pose estimation method proposed in this paper is validated by simulation and experiments.

Deep convolutional networks have achieved great success in semantic segmentation. However, existing semantic segmentation networks are difficult to achieve an ideal balance between accuracy and speed. To tackle this dilemma, the key idea of this paper is to design a two-part knowledge distillation framework to improve the accuracy of real-time networks, including separated distillation and long-range distillation. Separated distillation treats semantic segmentation as an independent pixel classification task, aiming to align logits between a cumbersome teacher network and a lightweight student network. To extract the long-range knowledge from the scene, the similarity attention map is defined as knowledge, which is transfered from teacher to student through the long-range distillation. The framework proposed in this paper improves the accuracy of real-time networks on Cityscapes and Fisheye datasets by 2.7% and 1.7%, respectively. The distilled network is tested on NVIDIA Xavier embedded platform with speed of 15FPS. Code will be available here: https://github.com/WangYunnan/Real-Time-Semantic-Segmentation.

At present, many industrial applications urgently need a real-time high-accuracy automatic counting algorithm to realize the counting function of various objects. Some scholars are studying the use of deep learning algorithms to count objects in industrial applications. However, facing closely arranged and different objects, deep learning algorithms cannot achieve this function. Aiming at the application background of regular cuboids counting with a wide range of requirements, this paper provides a real-time high-accuracy automatic counting algorithm based on a deep understanding of the scene and a comprehensive use of multiple computer vision technologies. The algorithm can meet the stringent requirements of industrial applications and realize real-time high-accuracy automatic counting of closely arranged cuboids. And there is no requirement for whether the surfaces of the cuboids are the same, and there is no need to prepare the template pictures of the cuboids in advance. The algorithm mainly uses the characteristics of gaps between cuboids, and adopts improved Hough line detection to identify gaps, and finally realizes the counting function. The algorithm proposed in this paper has been verified on the counting of pillboxes in the industrial application of smart pharmacies. Compared with the deep learning algorithm, it has the characteristics of good interpretability, good generalization, fast speed and high accuracy.

Traditional laboratory provides students with physical places and equipment to conduct experiments. Under this circumstance, the record and analysis of experimental data lack convenience and completeness to a certain extent. Compared to the former laboratory, online laboratory is able to record and store more data during the process of experiment. Online laboratory has changed the way of assessing students’ laboratory work. Based on these online recorded data, judging criteria of laboratory work can be extended to a richer dimension like behavior, attendance, and online experimental reports. Meanwhile, modern large-scale data storage and analysis technologies can be used to carry out the assessment. In this paper, big data assessment methods for students’ laboratory work are proposed. The data generated by users’ experiments is stored by the method of big data through distributed storage system. The assessment emphasis on laboratory work lies in behavior data which is generated during experiments and the assessment is carried out based on fuzzy inference. This paper takes Wuhan University as a case study, where students major in automation are using Networked Control System Laboratory (NCSLab) as a tool to conduct experiments for their laboratory sessions.

In the process of chip wafer dicing, it is necessary to manually check the relative position of the cut mark and the cutting path in real time in order to adjust the cutting disc in time. During the process, the instability of manual calibration will affect production efficiency. Therefore, this paper proposes an improved U-Net convolutional neural network to locate the cutting path. It improves the network's ability to extract image features and generalization capabilities by adding the RFB receptive field module proposed in the target detection task to the U-Net encoder. On the image segmentation effect of the wafer cutting path, the network RFB-UNet designed in this paper improves 2% on the indicators of MIoU and FWIoU. It is verified that the RFB is a module that can improve the existing network in the task of semantic segmentation. Besides, this paper constructs a dataset for image segmentation of the wafer cutting path.

In order to make convolutional neural network (CNN) run more effectively on embedded devices, many studies about binary neural networks have appeared. The traditional binary neural networks adopt the STE-based binarization method. This method has the problem of gradient mismatch during the training process. It will cause an enormous loss of accuracy to the network. For this reason, this paper proposes an adaptive binarization method. The method selects hardtanh function as the basic binarization function. This function changes alternately as the training epoch increases. It will eventually approach the sign function. And the method adjusts the threshold of binarization adaptively by introducing a learnable parameter. Based on the two CNN models of VGG-Small and ResNet-20, this paper conducts training with the adaptive binarization method on the CIFAR-10 dataset. The results show that the proposed method is better than traditional methods (BC, BWN). And Its accuracy is very close to some advanced methods, while these methods have lower running speed. Therefore, the adaptive binarization method has higher accuracy, and it can achieve faster running speed in the application of binary neural network.

This paper introduces a model for vital signs estimation, mainly, the heart rate, breathing rate and body movement, based on convolutional neural network (CNN). Due to the limitation of hardware acquisition system and detection method, the traditional measurement method with ballistocardiogram(BCG) signals or electrocardiogram(ECG) signals is difficult to meet the requirements of high accuracy and use convenience at the same time. Aiming at this problem, this paper, which uses high precision BCG data as label value, improves the data forecast model, including the use of deep learning approach for data prediction, and the appropriate pretreatment process data, using piezoelectric ceramic signal, which records the BCG signals, to carry on the forecast, building a forecasting model for vital signs using CNN. Low complexity and good precision of the model, making it accessible for both the convenience and data accuracy of the entire vital signs monitoring system.

With the popularization of smart screens, screen bonding technology has gradually developed. This paper uses the SIFT algorithm to recognize the LCD screen. Aiming at the problem of long matching time and large memory usage for its 128-dimensional floating-point descriptor, this paper proposes an improved binary descriptor, which can effectively increase the percentage of correct matching between key points and reduce the matching time by 1/3. After recognizing the object, this paper proposes a two-stage positioning method. We use the screen center obtained by the SIFT algorithm as the seed to perform Flood-fill processing, and obtain the four inner corners of the LCD screen through sub-pixel corner detection to accurately calculate the screen center. Finally, it is verified through experiments that the positioning error can be kept within 2.5 pixels, and the angle error can be kept within 0.9 degrees, which can meet the screen fitting accuracy requirements.

With the development of society and the progress of science and technology, the demand for electricity in all walks of life is also increasing. Transformer is the core of power system, plays a decisive role in power transmission and distribution, and coil winding is the main core technology of transformer manufacturing, the quality of winding coil directly determines the stability and reliability of transformer work. In this paper, under the background of low automation of transformer winding machine and higher and higher requirements for winding efficiency, the paper takes the winding system of insulation layer of transformer winding machine as the research object, improves the traditional fuzzy control algorithm, and designs a fuzzy PID control scheme combined with the tension control concept. Compared with the traditional PID control algorithm, the feasibility and advantages of the scheme are verified by the results of the Matlab simulation experiment.

In non-frontal situations, the image generally does not contain complete facial information, and skin color is the most important feature for detecting human faces. In order to improve the detection effect of non-frontal face skin color and applicable scenarios, this paper first analyzes and compares the YCbCr and YCgCr color spaces. Then, by comparing the fixed threshold method, the Gaussian skin color model method and the Otsu method, it is found that the Gaussian skin color model method is the most effective method, but still needs improvement. This paper proposes the Otsu method to optimize the threshold calculation in the Gaussian skin model, and the nonlinear segmented ellipse model reduces the impact of brightness. Experimental results show that this method can eliminate skin-like backgrounds more effectively, and can still maintain a better detection rate under dim lighting, and has a more ideal detection effect.

An integrated motion control strategy for spacecraft rendezvous and proximity operations with parametric uncertainties, disturbances and couplings is presented in this paper. Firstly, the characteristic model of relative motion is established. By virtue of the obtained characteristic model, a robust H ∞ relative translational and rotational controller is derived. Correspondingly, the controller design for spacecraft relative motion is formulated as a linear matrix inequality condition. Compared with previous integrated translational and rotational control schemes of spacecraft, the control strategy in this paper is easier to obtain without relying on precise mathematical model of the plant, and can guarantee the system performance subject to disturbances and parametric uncertainties. Numerical simulation verification indicates the effectiveness of the developed controller.

Tunnel LED lighting system is an important infrastructure to ensure the safety and comfort of driving in highway tunnels. In this paper, an automatic robust detection of lighting LED lamps in highway tunnels based on vehicle-mounted video processing is proposed. By analyzing the geometric structure of the interior road environment in tunnels, this method includes the following three aspects. Firstly, the front-vehicle vanishing point is detected by using the M-estimator sample and consensus (MSAC) algorithm, just with the set of significant parallel lines extracted using the LSD (Line Segment Detector) operator. Secondly, after the dark region on the top of the tunnel is accurately segmented, an image segmentation algorithm based on adaptive threshold is adopted, realizing the LED lamp candidate region detection. Thirdly, this paper proposes post-processing methods to ensure the accurate detection of LED lamps in complex scenes with highlight-interference in the tunnel. The experimental results show that this method can robustly detect the LED lamps in the tunnel under complex conditions with high accuracy, which can meet the needs of efficient acceptance inspection and daily maintenance of tunnel LED lighting system.

COVID-19 can be transmitted by respiratory droplets and other means. With the normalization of epidemic prevention and control, mask detection should be performed when entering public places with heavy human flow. A mask detection method for robot was proposed in this paper to prevent cross-infection of the detection personnel. This paper used an improved algorithm of YOLO. First, this article is based on an improved backbone Darknet-53, and add the EfficientNet attention mechanism as the basic network, reduces the consumption of calculation and improve the training speed, secondly the original IoU replaced by DIoU, improves the prediction precision of the border, finally combining with robot platform, according to more than 6000 pieces of VOC data format homemade model training and test data set. The robot detects the target based on the camera it carries, and then guides it to the next step. The experimental results show that the mAP of the algorithm is 86.92% on the self-made verification set, and the preset process can be realized in the actual detection process by using the robot.

Satellite cloud image data plays an important role in the ultra-short-term photovoltaic (PV) power forecasting. By calculating the cloud displacement vectors of two adjacent cloud images, it is possible to calculate the cloud cluster's occlusion of the PV power plant for a period of time in the future, thereby forecasting the PV power. The satellite cloud image data provided by FY-4A Satellite has large and unequal sampling time intervals, which brings difficulties to cloud displacement vector calculation. This paper proposes a method for calculating the displacement vector of the cloud edge block based on the cosine similarity and the adaptive adjustment of the matching spatial scale. The parameter equation is used to establish the relationship between the relative position of the cloud cluster and the PV plant over time, and to predict the future position. On this basis, this paper proposes an ultra-short-term forecasting method of PV power based on cloud displacement vector. The performance of the proposed model is verified through case analysis.

For the map building problem of laser and vision fusion, a novel fusion algorithm is proposed in this paper. At first, the algorithm preprocesses the original point cloud obtained by the depth camera with voxel filtering, coordinate transformation, direct-pass filtering, dimensionality reduction processing and conversion of the polar coordinates. Finally, two types of data fusion are performed. The paper designs three experiments on data fusion, map building, and robotics navigation. The experimental results show that the fused data of camera and LiDAR can reflect more useful information from the surrounding environment compared with the 2D LiDAR data, and the environmental map built by the robot with the fused data for SLAM presents more comprehensive environmental information. Similarly, when the robot navigates with the fused data, the actual effect of obstacle avoidance and navigation is satisfactory.

In spacecraft control, in order to achieve fixed-point flight, target tracking, formation control and other tasks, the spacecraft is often required to have the ability to avoid obstacles. Especially in recent years, with the increase of space junk, the flight environment of spacecraft becomes more and more complex. Therefore, this paper studies the problem of obstacle avoidance in spacecraft flight, and designs a second-order sliding mode controller. By constructing an artificial potential function, the spacecraft can plan a safe path in the flight mission. And in order to better observe the state of the spacecraft, this paper designs a monitoring platform based on smart phones to realize the real-time monitoring of the state of the spacecraft. Finally, the effectiveness of the controller and the monitoring platform is verified by the physical experiment based on the spacecraft simulator.

Aiming at the low detection accuracy problem of small samples in network intrusion detection, this paper researches the data expansion method based on the generative adversarial network algorithm. The data sample is generated to supplement the small data set sample, which is helpful to improve the accuracy of network intrusion detection. Firstly, the non-numerical data in the original data set is converted into numerical data by one-hot coding, and reconstructed them into two-dimensional data. Secondly, this paper proposes an S-DCGAN algorithm to generate data. Initially, we improve the network structure of deep convolutional generative adversarial networks (DCGAN). S_ReLU function is used instead of the ReLU function as the activation function after each convolutional layer of the adversarial network, which makes the quality of the generated data better and the effect of the generated data more stable. Furthermore, we improve the loss function of generation model and discriminant model of DCGAN. A simplified total variation loss function is added to the loss function of the generated model to reduce the noise of the generated data. The gradient penalty term is added to the loss function of the discriminant model to alleviate the problem of gradient explosion or gradient disappearance and improve the stability of network training. In order to verify its effectiveness, various classifiers are used to evaluate the performance of the proposed data augmentation method. The results show the accuracies for classifiers are increased after employing the S-DCGAN-based data augmentation techniques.

Myocardial infarction is one of the common cardiovascular diseases, and it also is a major cause of death and disability worldwide every year. Therefore, early detection of myocardial infarction is vital for patients. Deep learning algorithms have been used successfully for the recognition of myocardial infarction in recent years, but most of them are based on the segmentation of heartbeat signals and the single deep learning model. In this paper, we proposed a new myocardial infarction classification method based on an improved ECG feature extraction scheme and the multi-network stacking model. Firstly, the mean amplitude spectrum (MAS) is used in feature extraction to elaborate the frequency distribution difference of each wave among ECG signals. Secondly, due to the lack of time-domain features in the MAS map, we proposed a novel time-frequency feature fusion method. Finally, the time-frequency feature is fed to the multi-network stacking model for the classification of myocardial infarction. To improve the diversity of feature learning, this paper researches to stack different networks for learning the characteristic of the time-frequency feature. Meanwhile, we used the improved training methods to solve the problem of data imbalance. Experiments using the PTB database show that this feature extraction scheme and multi-network stacking model are effective for the classification of myocardial infarction.

Facial expression recognition technology is significant to psychological analysis, polygraph technology, clinical medicine, and safe driving. However, traditional facial expression recognition technologies mainly recognize frontal facial expressions, and few studies focus on the recognition of facial expressions under deflection angles. In this situation, this paper combines Adaboost cascaded detectors, Ensemble of Regression Trees algorithm, and Random Forest algorithm to recognize facial expressions under deflection angles. After detecting faces by Adaboost cascaded detectors, the facial expression features under deflection angles are extracted by the Ensemble of Regression Trees algorithm. Finally, the Random Forest algorithm is used to output the recognition results based on expression features. Experiments show that the enhanced facial expression recognition algorithm proposed in this paper realizes the recognition of non-frontal facial expressions, and it has good real-time performance and recognition accuracy.

As modern engineering systems tend to be more complicated and integrated, prognostics and health management (PHM) has become an important and practical technique to predict remaining useful life (RUL) and ensure reliability. However, there are still challenges for RUL prediction via multi-sensor data. Therefore, this paper focuses on the RUL prediction problem for complex engineering systems. In order to make full use of multi-sensor monitoring data and enhance prediction accuracy, a feature selection method is developed based on Relief algorithm, then a practical RUL prediction approach is proposed based on deep long and short term memory (LSTM) network in this paper. The proposed approach can selectively focus on certain important historical data without any prior knowledge. Experimental results and comparison with other conventional methods are presented to demonstrate the effectiveness and superiority of the proposed RUL prediction approach.

In order to predict heart failure disease in advance and keep heart failure patients away from pain, this paper combines genetic algorithms (GA) to optimize Extreme Learning Machine (ELM) and builds a network model that predicts heart failure disease based on some of the characteristic factors that may induce heart failure. The ELM model has the characteristics of fast training speed and strong generalization ability, but its randomly generated input weights and hidden layer bias will cause instability in the prediction results. In this paper, we use GA’s principle of survival of the fittest to find the most suitable parameter configuration of the ELM algorithm, and analyzes the simulation results, and compares them with the prediction results of the original ELM model. The results show that the GA-optimized ELM model can better predict the occurrence of heart failure disease and has better predictive performance. The ELM algorithm, which has undergone parameter optimization has made up for its original defects, which may affect clinical practice and become a new assistant for doctors to predict patients with heart failure.

This paper presents a new location detection technology to find the bar code under complex background. Firstly, the contour of the grayscale image is extracted by the optimized Sobel operator and differential algorithm, and then we use the morphological algorithm to make extracted contours more complete and clear. Secondly, we select the area where the bar code may exist, with the area of contours as a standard. Then the paper uses the Hough transform to detect and filter the contour edges that meet the requirements. Then aiming at the area where the edge of the selected contour was located, we calculate the gradient in the relevant direction and statistic results. Furthermore, based on statistics, we filtered out the area where the bar code is located. Experiments show that the bar code location detection algorithm proposed in the paper is not only superior to the traditional gradient detection and morphological detection methods, but also has the advantage such as good robustness, small amount of processed data, fast processing speed, and high accuracy, which is fit for the image larger than 4 million pixels.

With the intention of obtaining the room temperature and economic cost control strategy of an ice storage air conditioning system in a small office building in Shanghai, the ice storage air conditioning system is established in this paper as a Markov decision process model and deep reinforcement learning algorithms are adopted to optimize its operation. In order to avoid the problem of over-dimension and over-estimation of value function caused by reinforcement learning, the DDQN (Double Deep Q-Network) algorithm with dual neural network structure is adopted to optimize the operation of ice storage air-conditioning system. Aiming at overcoming the shortcoming of slow convergence of DDQN algorithm, the action space of DDQN is taken into consideration in this paper. Firstly, the appropriate action set is selected according to the convergence speed of different actions. Secondly, the exponential function is addressed in the reward function. Based on the exponential function, the reward function can adjust the penalty value according to the difference between the expected room temperature and the actual room temperature, thus speeding up the convergence of the DDQN algorithm. Finally, Python is adopted to model and simulate buildings and ice storage air conditioning systems. The simulation results show that the operating cost and the proportion of uncomfortable time are both reduced by DDQN control method proposed in this paper with better control performance.

With the development of artificial intelligence technology, computer vision has gradually entered people's daily lives. In recent years, target detection as an important branch of computer vision has attracted the attention of many scholars and has made some progress. In various application scenarios with target detection and recognition, such as night, haze and other severe weather conditions, the demand and application of target detection for infrared images are becoming more and more extensive. There are obvious differences between infrared images and visible light images. Infrared images have low imaging contrast, unobvious texture features, and more noise. These physical characteristics make infrared target detection always challenging. This paper is based on the one-stage target detection algorithm YOLOv3. In order to improve the detection ability of small targets, we improve the network structure and expand the network to 4 feature scales; by introducing GIOU, the loss function is improved, and the accuracy of network detection is improved; by merging the Batch Normalization layer and the convolutional layer, the speed of network inference is speeded up. Experimental results show that compared with the original YOLOv3 network, the improved YOLOv3-4GB network has improved detection accuracy and enhanced detection capabilities for small targets; this paper deploys the improved algorithm on the embedded platform to meet the requirements of real-time detection.

Image target recognition based on strapdown seeker is the key to extracting guidance information and realizing precise guidance. This paper proposes a new deep neural network target recognition framework. The framework extracts features from target template images and strapdown images at the front end and inputs the extracted feature maps into the deep neural network for target region recognition, thereby realizing the objects given by the template are accurately identified in the strapdown visible light seeker image. Aiming at the application scenario of strapdown image target recognition, this paper produced a target dataset, and tested typical algorithms such as SSD, YOLOv3, YOLOv4 and YOLOv5 on the proposed framework. The results show that the proposed framework can effectively recognize the objects given by the template in strapdown images.

This paper studies the mismatch of the overlap areas during the stitching process of Unmanned Aerial Vehicle (UAV) remote sensing images. For solving this problem, the feature points are filtered first by combining double matching with the Random sampling consensus (RANSAC) method. Secondly, to ensure the consistency of the projection relationship between each image and panoramic photograph, we propose a local stitching method. In order to avoid the mistake of image tilt caused by perspective changes accumulate as the number of images increases, we have established an image rotation coordinate system and limited the relationship between images to translation and rotation. The relative position of the coordinate origin is used to represent the translation distance, and the optimal rotation angle is solved by iteration. Finally, the overlapping parts of the image are fused by linear weighting. It is verified through the experimental results that the method proposed in this paper can guarantee faster processing speed and higher processing accuracy under the condition of a large number of images, thereby achieving an ideal stitching effect.

CP-ABE (ciphertext-policy attribute-based encryption) is a promising encryption scheme. In this paper, a highly expressive revocable scheme based on the key encryption keys (KEK) tree is proposed. In this method, the cloud server realizes the cancellation of attribute-level users and effectively reduces the computational burden of the data owner and attribute authority. This scheme embeds a unique random value associated with the user in the attribute group keys. The attribute group keys of each user are different, and it is impossible to initiate a collusion attack. Computing outsourcing makes most of the decryption work done by the cloud server, and the data user only need to perform an exponential operation; in terms of security, the security proof is completed under the standard model based on simple assumptions. Under the premise of ensuring security, the scheme in this paper has the functions of revocation and traceability, and the speed of decryption calculation is also improved.

This paper discusses the problem of avoiding threats during the cruising flight of hypersonic vehicles (HV). Considering the constraints on kinematics of HV and changing environments, this paper proposes two methods of trajectory planning that taking the overload or the rotational angular velocity of the ballistic deflection angle as actions of agent. Meanwhile, the agent’s policy is optimized with policy gradient method, and Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC) are used for comparison. Experimental results show that PPO and SAC have similar performance in penetration missions. Moreover, in the complicated flight environment, the method of taking overload and exploration distance as actions has a higher penetration success ratio.

For the self-supervised image depth estimation algorithm based on deep learning, the characteristics of self-supervised methods cause some problems such as visual shadow and infinity estimation. This paper tackles these problems by adopting the spatial consistency loss and multiple loss constraints. In this paper, the depth estimated network and pose estimated network are trained by binocular image sequence. And the depth estimated problem is transformed into image reconstruction and scene reconstruction to realize self-supervised training. Based on the Deep Convolutional Neural Networks(DCNN), the depth estimated network introduces the Deep Atrous Convolution (DAC) and the Atrous Spatial Pyramid Pooling (ASPP) modules to extract the multi-scale feature and increase the receptive field of the network. The spatial consistency of scene reconstruction strategy greatly reduces the estimated error of pose network. The quantitative and qualitative evaluation on the KITTI dataset shows that the proposed algorithm outperforms the existing self-supervised depth estimation algorithms in accuracy and error, even exceeds most of the supervised algorithms.

The parasitic inductance of transmitting coil in transient electromagnetic transmitter causes some serious problems like the overlong falling edge time, overshoot and oscillation of the emission current. In order to solve these above problems, this paper proposes a design scheme of passive constant voltage clamping circuit of highly speeding shutoff based on TVS (Transient Voltage Suppressor) and switch. In this design scheme, resistance, TVS and switch are configured in parallel and then connected into the main launch bridge circuit. During the earlier stage of the shutoff of the emission current, TVS is utilized to form the high voltage clamping and thus realize the highly speeding shutoff of the emission current. At the later stage of the emission current decline, the energy of the load inductance is set free by the resistance to prevent the overshoot and oscillation of the emission current. In this paper, the operating procedure and principle of the circuit, the influence of the parameters of core devices on the circuit performance, and the effectiveness of this design scheme is verified by simulation and experiment. The results of simulation and experiment show that this circuit is effective to reduce the shutoff time of the emission current, restrain the overshoot and oscillation meanwhile, and hence improve the wave quality of the emission current.

Link prediction is a network analysis method to solve practical problems and has important research value in many fields. In the field of traditional Chinese medicine, according to the different needs of the disease and the different characteristics of the drugs, the combination of the two drugs is called drug pairing. Based on traditional Chinese medicine network, this study provides a new research perspective based on complex network and link prediction. Aiming at the problem of strong randomness in the traditional random walk, a new link prediction method based on graph embedding method is proposed in this paper. Compared with the traditional random walk, the index proposed in this paper improves the performance greatly. Through link prediction, it is verified that the drug pair without compatibility in the existing prescription may appear in the future prescription, as a supplement to the new drug pair.

Under Emerging Engineering Education circumstances, this paper aims to cultivate the learning interest and strengthen the engineering practice ability of first-year students majoring in automation and control by making a self-leveling quadrotor. Firstly, the attitude calculation based on complementary filtering and the PID control algorithm based on a quadrotor altitude control case is intuitively explained. Furthermore, this paper guides students to utilize Arduino Uno and MPU-6050 to implement the PID control program and attitude calculation program step by step. Finally, the essential hardware and the circuit diagram are given to build a self-leveling quadrotor. This paper can be used as an introductory course for automation and control majors, mainly to stimulate students’ interests and establish students’ systematic design thinking and hardware and software coordination ability.

In order to solve the problems of inaccurate non-contact heart rate measurement, some methods in the process are optimized in this paper. Firstly, the KCF is used to track the points of facial feature detected by SSD method. Secondly, each of the tracked facial images are segmented by Pixel2Pixel to extract the mean of RGB channels from skin region. Thirdly, the extracted RGB signals are processed by the combination of POS and EEMD to obtain purer BVP signals. Finally, the heart rate value is calculated by two methods of time domain and frequency domain, which is compared and analyzed with the truths provided by the UBFC-RPPG database to verify the feasibility of this method. The experimental results show that the method proposed in this paper can measure the heart rate value accurately, which meets the relevant standards of medical physiological measurement fully.

Since the new century, high-precision sensors and high-performance controllers have ushered in high-speed development, which provides the necessary conditions for the realization of tractor self-driving system. In practical control algorithms, the tractor’s actual trajectories often show a one-sided bias in a straight-line tracking task. This phenomenon relates to asymmetry in tractor structure, in-run bias in sensors and actuators, and uncertainties in tractor model. In this paper, a control system aims to eliminate the system-bias of tractor is proposed. On the basis of ordinary tracking algorithms, the bias of heading angle is compensated and an additional PID controller is integrated to eliminate the lateral bias. Based on the Dongfeng tractor DF1004-2, this paper establishes experimental platform, equipped with controllers, sensors, actuators and other ancillary equipment. The control algorithm proposed in this paper is verified. The experimental results in a real farm field show that the tractor control system designed in this paper has robust stability and high precision, the steady-state lateral absolutely error is 85% within 5cm and 60% within 3cm, and the system bias is reduced close to zero which meets the normal operation requirements of autonomous tractors.

An automatic quality inspection system, aiming to recognize defective products that would bring great damage, is essential for the production of solar cells. However, few papers focus on the detection of back electrodes in solar cells. Segmenting electrodes accurately is the first step of defect detection for them, for which a complete algorithm is proposed in this paper to segment back electrodes. Firstly, coarse and fine positions of electrodes are obtained using the edge intensity projection and template matching method. And then the edge of electrodes is extracted by the threshold method and seed growing method. Thirdly, an edge-to-region method is adopted to generate a closed region with edge points. To verify the performance of the proposed method, an experiment is carried out. From the experiment result, the segmentation and detection scheme for electrodes proposed in this paper outperforms the three famous methods implemented in the experiment.

This paper considers a decentralized formation tracking with disturbance suppression control problem to achieve unmanned aerial vehicle (UAV) collaborative transportation. The main purpose of the paper is to develop a scheme of the UAV collaborative transportation independent of the globally positioning system (GPS). We focus on a scenario that every UAV is equipped with a tether sensor linked to a specified rigid load. Suppose that, the wire angles measured by IMU together with the load attitude and position, relative to the global frame, are all accessible to all the collaborative lifting UAVs. Then we present a control algorithm using an internal model principle approach to achieving the decentralized formation tracking With disturbance suppression. We demonstrate the effectiveness of the proposed positioning scheme and algorithm via simulation and initial experiment.

Automated guided vehicles are widely used in industry, agriculture and service industry. To achieve the precise positioning of the vehicle, this paper designed and modeled a UWB-based automated guided vehicle positioning system, and proposed a positioning method based on trilateration characteristics and Kalman filtering. The base stations measure the distance to the tag on the vehicle, and then the coordinates of the automated guided vehicle are initially calculated through trilateration. To reduce the influence of noise signals and hardware errors on positioning accuracy, Kalman filtering is used to optimize the positioning system. The indoor positioning experiments show that the positioning ranges of the automated guided vehicle in uniform linear motion and uniform random motion are (-0.0307m, 0.0552m) and (-0.0531m, 0.0670m), respectively, and the standard deviations are 0.0242m and 0.0379m, respectively. The method proposed in this paper has good accuracy and practicability, and can meet the positioning needs of automated guided vehicles.

Integration of large-scale renewable energy in the power system helps reduce carbon dioxide emission. However, their volatile nature also challenges the security operation of power systems, which boils down to the feasibility of network constrained power flow under a given nodal injection pattern. This paper studies the solvability of alternating-current (AC) power flow with volatile renewable generation and operating constraints. The set of all admissible injection patterns of renewable power is defined as the dispatchable region, which contains essential and visualizable information on system security. This paper characterizes the dispatchable region through several multi-objective optimization problems; the Pareto frontiers constitute the boundary of the dispatchable region. A e-constrained method is developed to compute and visualize the dispatchable region. Case studies on the IEEE 9 bus system validate the proposed method.

This paper studies the platoon control of automatic vehicles based on deep deterministic policy gradient (DDPG). The main control goal of the platoon control of automatic vehicles is to control the acceleration of the following vehicles. In this paper to meet the requirement of the convergence quickly for the following automatic vehicles. Therefore this paper introduces the platoon control based on the DDPG algorithm. Through the reasonable design of the reward function in the DDPG algorithm can solve the speed following optimization problem, and modular design of reward function which avoids the existing optimization problems some dangerous driving behaviors such as collision and reversing do not occur during the train. The simulation results show that following vehicles can achieve a quickly learning process under the deep deterministic policy gradient method, and can also track the speed of the leading vehicle with security.

This paper proposes a set of automated garbage sorting system and designs the garbage sorting algorithm based on the neural network and YOLO algorithm. To realize the detection of books, cans, plastic bottles, and glass bottles, we design a feature extraction, two detection layers and corresponding anchors and a loss function in turn, which are based on a neural network, the FPN idea and the detection output characteristics, respectively. The results show that the network model designed in this paper is light, with an accuracy rate of 86.9% and a speed of 15ms. When detecting multi-scale targets, the detection accuracy is above 86.5%, which can meet the real-time and accuracy requirements of the system.

The leakage location for water supply system is a key point for the efficient operation and maintenance of urban water supply system. However, the limited source of leakage dataset, high cost of leakage simulation experiments and the limited prediction accuracy of traditional models increase the difficulty of the leakage location for urban water supply system. This paper proposes a new method based on ensemble learning method to address these problems. EPANET software is employed in this paper for the generation of experimental dataset. The leakage is located and predicted based on the hydraulic model of the water supply system through the design of the improved multi-model weighted voting mechanism for ensemble deep learning method. The experiments show that the proposed method has higher precision and lower discretization compared to traditional models.

Turbofan engines are known as the heart of the aircraft, as important equipment of the aircraft, the health state of the engine determines the aircraft’s operational status. Therefore, the equipment monitoring and maintenance of the engine is an important part of ensuring the healthy and stable operation of the aircraft, and the remaining useful life (RUL) prediction of the engine is an important part of it. The monitoring data of turbofan engines have a high dimension and a long time span, which brings difficulties to predicting the remaining useful life of the engine. This paper proposes a residual life prediction model based on Autoencoder and temporal convolutional network (TCN). Among them, Autoencoder is used to reduce the dimension of the data and extract features from the engine monitoring data. The obtained low-dimensional data is trained in the TCN network to predict the remaining useful life. The model mentioned in this article is verified on the NASA public dataset(C-MAPSS) and compared with common machine learning methods and other deep neural networks. The experimental results show that the model proposed in this paper performs best in the evaluation methods, and this conclusion has important implications for engine health.

Based on the e-commerce background of fresh agricultural products and the theory of supply chain procurement management, this paper builds the optimization model and evaluation system of fresh agricultural e-commerce supply chain procurement by the analysis of the existing problems in the procurement management of fresh agricultural products e-commerce supply chain and Case study of self-operated fresh produce in Yihaodian, Finally this paper optimizes the model of fresh agricultural e-commerce supply chain procurement management of Yihaodian by using cross-analysis, JIT procurement management, TOPSIS analysis method and balanced scorecard model.

Uncertainties of renewable energy sources and load demand have a significant effect on the microgrid operation. In this paper, a robust optimal energy management approach is proposed for a grid-connected microgrid where the adjustable uncertain sets are used to represent uncertainties. A novel statistic index is developed to evaluate the effectiveness of the robust optimization and obtain reasonable model parameters. Moreover, the optimal economic cost and the least environmental pollution are achieved simultaneously. Simulation results in different scenarios are given to verify the effectiveness of the robust optimization approach.