Abstract:In recent years, visual SLAM has attracted wide attention due to its advantages of simple structure, low cost and ability to integrate semantic information. Loop closure detection plays an important role on it. According to the loop information obtained, the visual SLAM back-end optimization algorithm can optimize the pose according to the loop constraint, eliminate the cumulative error generated after long-term work, and achieve accurate long-term positioning, in order to build a globally consistent motion track and map. First introduce the principle and function of loop closure detection in visual SLAM, then conduct an in-depth analysis of the traditional bag-of-words model from feature extraction, similarity judgment, and experimental evaluation, and outlines several improved algorithms based on the bag-of-words model and probability, and summarizes several loop closure detection methods based on deep learning, briefly summarize the loop closure detection methods combined with semantic information, and finally summarize and prospect the current problems and future development of loop closure detection.

Abstract:The acquisition of flexible tactile perception information is the key feedback of soft bionic manipulator to achieve autonomous control. The information of interaction force and surface characteristics is provided, which is generated between the bionic manipulator and grasping contact point. In recent years, although the flexible tactile sensing method applied to software manipulator has made some progress in sensitive materials, structural design, preparation methods and signal acquisition and analysis, there are some problems that limit the application and development of the technology. In view of this, this paper sorts out the concept and connotation of flexible tactile sensing technology for software bionic manipulator, and its related parameters and design principles are systematically analyzed. The design ideas, performance and applications of flexible tactile sensors with different structures are summarized, classified and discussed. Finally, the discussion and the prospect of development for the flexible tactile sensor for soft bionic manipulator are presented in the conclusion.

Abstract:Aiming at the problem of non-commutative errors in the large dynamic environment of MEMS inertial navigation system, an improved high-order iterative attitude optimization algorithm is proposed. In order to solve the influence of non-exchangeable errors on the entire inertial navigation system in a large dynamic environment, the traditional equivalent rotation vector algorithm is deduced. For this algorithm, it only relies on increasing the number of subsamples to improve the solution accuracy, ignoring the problem that high-order terms will cause large errors in a large dynamic environment. Using the method of fast and slow loops, the rotation vector solutions of different orders are obtained respectively, and then the iterative solutions of the fast and slow loops are obtained through the periodic iterative algorithm. Finally, through the large dynamic environment simulation experiment and the high-frequency swing dynamic experiment of the high-precision three-axis turntable, the performance advantage of the higher-order iterative algorithm is verified. The experimental results show that in a large dynamic environment, compared with the traditional algorithm, the improved high-order iterative attitude optimization algorithm improves the accuracy by two orders of magnitude.

Abstract:In order to improve the positioning accuracy of pedestrian dead reckoning ( PDR) for low-cost inertial measurement unit (IMU) arrays, this paper proposes for the first time the use of multi-layer perceptron (MLP) to achieve algorithms for low-cost IMU array data fusion. The measurement data of the IMU array ( including triaxial acceleration and triaxial angular velocity) and the measurement data of the high-precision IMU are obtained by synchronizing the motion of the self-designed IMU array and the highprecision IMU, and the measurement data of the high-precision IMU is used as a label. The MLP fuses the measurement data of the IMU array, predicts the actual acceleration and angular velocity of the object, and uses the positioning algorithm to verify it. In the localization experiment, the PDR localization accuracy using the prediction data fused by MLP is 33. 9% higher than the PDR localization accuracy using a single IMU measurement data; it is 20. 8% higher than the PDR localization accuracy using the simple average processing IMU array measurement data; it is 11. 6% higher than the PDR localization accuracy using the IMU array measurement data fused by the least square method, which proves the feasibility and effectiveness of the method proposed in this paper.

Abstract:To achieve person recognition in water environment, a person category identification algorithm based on vision sensors on unmanned surface ship (USV) is proposed. Firstly, base on the data acquisition and model update workflow, a person category dataset of 39 959 pictures and 7 categories is created after data cleaning and labeling on original videos. Secondly, YOLO v5, the mainstream object detection network in the field of deep learning method, is practiced, and an improved person category identification algorithm based on YOLO v5 is proposed according to the characteristics of water environment scenes. Thirdly, the algorithm is deployed to the edge computing platform to realize the real-time use of the algorithm on the unmanned ship. The algorithm achieves an average accuracy of 86% on our dataset and achieves real-time performance of processing 38 frames per second with accurate person recognition in the unmanned ship test.

Abstract:Aiming at the low efficiency of measuring the size of a manual pinch with a vernier caliper, an improved adaptive Canny operator is proposed to measure the size of a manual pinch. Firstly, the mean shift method is introduced to segment the image, and the image color becomes gradual and the fine grain texture is gentle. Secondly, the average pixel value of approximate contour is selectively used to replace the high threshold of edge detection, to achieve the purpose of better adaptive edge extraction. Thirdly, Scharr operator replaces Sobel operator to enhance the ability of weak edge extraction. Finally, the small connected domain is removed, and the principal direction of the contour is obtained by principal component analysis ( PCA), and the straight-line equation of the principal direction is obtained. The length of the pinch is obtained by the maximum projection distance of the contour point on the straight line, and the width of the pinch is obtained by the maximum projection distance of the contour point in the vertical direction of the straight line. Experiments show that this method has significantly improved the recognition rate of irregular objects more than other methods, with an accuracy of 97% and an error of about 2 mm from the real value. Through the comparative analysis of strong light, weak light and sidelight experiments, this method is less affected by light and meets the needs of industrial automation.

Abstract:The task scheduling work in the current mobile edge computing (MEC) environment often ignores the dependency between tasks, resulting in a long delay in completion. In response to this problem, first of all, with the goal of reducing the system completion delay, in the multi-user and multi-edge server scenario that takes cross-server collaboration into account, the breadth first search algorithm (BFS) is used to build a dependent task scheduling model. Then, according to the interaction between tasks and edge servers, the joint offloading and migration problem of each scheduling layer in the model are modeled as a Stackelberg game with multiple leaders and multiple followers. Finally, in order to achieve Stackelberg equilibrium, an offloading algorithm based on the Q value and a distributed iterative migration algorithm are proposed to solve the model. The simulation results show that compared with the baseline algorithms, the proposed algorithm reduces the system completion delay by 44. 1% and 63. 2% respectively in the scenarios of users and edge servers with different scales. Further experiments show that compared with the traditional solutions, the proposed model reduces the system completion delay by 20. 1% and 6. 7% respectively in the scenarios of users and edge servers with different scales, and effectively guarantees the quality of service.

Abstract:Hall sensors are active devices manufactured from semiconductor materials, their output voltage drifts with temperature change, which limits their application in high-precision magnetic field measurement. A new symmetrical closed loop Hall current sensor is proposed and systematically designed. The output voltage operation circuit is designed based on the sensitivity difference of Hall element, the working principle of symmetrical closed-loop current sensor is analyzed and the closed-loop transfer function is given. The results of circuit simulation and prototype experiment show that the symmetrical closed-loop Hall current sensor has good temperature characteristics, and the temperature coefficient is reduced to 0. 000 8% / ℃ . Compared with other temperature compensation methods, this method is simple and easy to implement, and can realize the complete compensation of temperature drift. Hall components can be selected arbitrarily according to the measurement needs, and are not limited by the driving mode.

Abstract:The current noise reduction methods have noise residue and inadequate adaptability, so that the abnormal detection index is greatly affected by noise, a sensor fault detection method based on double noise reduction and fuzzy index for central air conditioning is proposed. Complete EEMD with adaptive noise (CEEMDAN) is used to extract k-order modes and replace modal estimation to achieve initial noise reduction. For the false mode appearing in the early stage, firstly, the noise-containing components are screened by the correlation coefficient criterion to retain the effective information as much as possible. Then, singular value difference spectrum is calculated to determine the order of denoising and singular value decomposition ( SVD) to complete the secondary denoising. The experimental data of central air conditioning system are used to verify the proposed method, this method has good ability of noise reduction and sensitive feature screening, the SNR was improved by 20. 203 7 dB, the mean square error was reduced by 48. 75% on average, the fault detection accuracy was improved by 8. 67% on average, and the response speed was improved by 33. 3%.

Abstract:In order to solve the problem that current electro-optic (EO) electric field sensor has no continuous power supply capability, a pulsed electric field sensor powered by laser photocell is designed and developed. The sensor consists of a laser photocell power supply circuit, an EO modulation circuit and a monopole antenna. The laser photocell power supply circuit which incorporated with a diode laser, a lithium battery, a photovoltaic power converter, a lithium battery charging, protection and discharge circuit are designed and fabricated. The experimental results show that the output voltage accuracy and ripple coefficient of the developed laser photovoltaic power supply circuit are 1. 04 % and 0. 3% respectively. Besides, the output voltage fluctuation is ±0. 035 V after continuous work for 48 hours. The EO modulation circuit of the sensor incorporated with a monopole antenna and a FET type operational amplifier is designed and fabricated. The measurement results show that the dynamic range of the developed sensor is from 0. 256 kV/ m to 13. 79 kV/ m.

Abstract:The actual working conditions of oil extraction are complex and changeable. When using the conductometric method to measure the water content online, the environmental factors, temperature, salinity and the electrode structure of the sensor itself will bring about measurement deviations. In this paper, three-electrode and four-electrode with different electrode ring inner diameter and excitation electrode spacing are designed for comparative experiments, and conductivity and temperature information were obtained by calibration experiments. Considering the nonlinearity caused by temperature and signal conditioning circuit, the comprehensive compensation function is established and optimized by the least square method to obtain the temperature and nonlinear comprehensive compensation parameters. The performance of the sensor is optimized by optimizing the electrode structure and compensating parameters. The water content testing system was established for simulation testing and verification. Experiments show that the measurement error of the sensor optimized and compensated for the electrode structure parameters is within 2%, and the stability is good, which can meet the measurement of crude oil water content.

Abstract:There’s an existing problem on linear structure that, the farther the array probe is located on coil’ s inner diameter, the smaller is the peak of differential signal of the same position defect. Aiming at solving this problem, this paper put forward an approach of placing the TMR sensor in the shape of semicircle with same horizontal spacing within the same excitation coil, and designed a new type of pulse vortex C-scanning imaging device. The findings suggest that, contour maps drawn after C-scanning and the conversion and threshold segmentation of scanning results could well reshape different forms of cavities on copper plate and successfully gained the 2D images of them, thus proving the effectiveness and feasibility of this approach.

Abstract:The thumb key-touching is a complex spatial movement which involves many factors. Its quantitative analysis is very important for the scientific guidance of piano teaching and the quantitative evaluation of teaching effect. This paper puts forward a measurement and analysis method for the training of thumb key-touching in the early stage of piano teaching. Firstly, the Leap Motion is used to build the measurement platform to collect the motion data of key-touching. Secondly, the hand joint coordinate systems are established, and the angle parameters representing the movement of thumb key-touching are set. Then the least square method is used to obtain the motion regression equation of the thumb key-touching. Finally, according to the regression equation, the change law of joints’ angle parameters in the process of thumb key-touching is predicted. This study can provide prediction curves of thumb key-touching motion for people with different thumb lengths, provide scientific guidance for piano playing and learning effect evaluation, and provide a theoretical basis for the follow-up design of piano teaching manipulator.

Abstract:How to weaken the influence of high temperature radiation and imaging distortion on component measurement is of great significance in aerospace and automobile manufacturing. In this paper, the radiation characteristics of the object at different temperatures are analyzed based on the blackbody radiation theory, and the imaging spectral range is determined. The high temperature binocular stereo vision measurement model is established. The deflection law of light propagation in the high temperature field is analyzed by combining the finite element modeling simulation and the ray tracing algorithm, and the influence of temperature and light wavelength on the deflection of light is analyzed. Finally, through the establishment of binocular stereo vision measurement system, the interference of high-temperature radiation light is weakened by the measurement system composed of CMOS camera, narrow-band filter and parallel light source, and the analysis results are experimentally verified. The experimental results show that the measurement system can obtain clear image features at high temperature. The results of ray tracing algorithm analysis are in good agreement with the experimental data; compared with the normal temperature measurement results, the error of visual measurement results under 800 ℃ is less than 0. 28 mm, and the standard deviation of ball spacing measurement is less than 0. 11 mm, which can meet the measurement requirements of components in high temperature environment.

Abstract:In order to improve the visual navigation ability of mobile robots without maps and improve the success rate of visual navigation, a visual navigation model of mobile robots is proposed that integrates long short term memory (LSTM) and proximal policy optimization (PPO) algorithms. Firstly, the model integrates LSTM and PPO as a network model for visual navigation. Secondly, a new reward function is designed to train the target through factors such as the action of mobile robots, the distance between the robots and the target, and the running time of robots. Finally, the RGB-D image obtained from the first perspective of mobile robots and the polar coordinates of the target in mobile robots coordinate system are used as the model input, and the continuous motion of mobile robots is used as the model output to realize the task of end-to-end visual navigation without maps, and the new target that has not been trained is reached according to the model inference. Compared with the pre-order algorithms, the model has an average increase of 17. 7% in the navigation success rate of the old target and 23. 3% of the new target in simulated environments, which has better navigation performance.

Abstract:In view of the high noise of synthetic aperture radar images and inconspicuous imaging features, especially in complex scenes such as sea and land boundaries, ports, and coastal reefs, it is difficult for common detection algorithms to extract target features from SAR images, resulting in low detection accuracy and leak detection, etc. This paper designs a rotating target detection method based on YOLOv5, and proposes that the multi-branch attention module can be used for cross-dimensional information fusion, which can better extract the location information and semantic information in SAR image targets. In addition, the boundary discontinuity will be caused by rotating target detection, which will affect the regression of the bounding box. Therefore, the circular smooth label method is used to transform the angle parameter from regression problem to classification problem, thus improving the accuracy. Finally, experiments are carried out on HRSID and SSDD+ datasets, and the accuracy reaches 84. 98% and 90. 13%, respectively, which is 1. 29% and 2. 57% higher than the original YOLOv5 algorithm, respectively. Experimental results prove the effectiveness of the proposed algorithm.

Abstract:Strip steel surface defect detection has become one of the important links to guarantee the quality of strip steel production. Aiming at the problem of improving the detection accuracy of current strip steel defect detection algorithm, an improved MT-YOLOv5 algorithm based on YOLOv5 is proposed. Firstly, introducing Transformer self-attention mechanism in the backbone network to make the network more focused on the extraction of global image feature information. Secondly, combining the Transformer layer with the BiFPN structure, and the T-BiFPN network is used to further enhance the fusion of image shallow feature information and deep feature information. Then, an improved lightweight network RepVGG is introduced to replace part of the convolutional layers in the backbone network, which can enhance the feature extraction capability of backbone network. Finally, adding a prediction layer to detect objects of different scales. The experimental results show that the value of mean average precision (mAP) of the MT-YOLOv5 algorithm is 82. 4% on the NEU-DET dataset, which is 5. 3% higher than the original YOLOv5s algorithm, and the detection speed reaches 65. 4 fps, which achieves a better balance between detection speed and detection accuracy.

Abstract:To deal with the problem that the traditional phase-shifting (PS) method cannot be used to measure the three-dimension (3D) shape for a non-static object, a grating image projection-based method which combines the improved PS method with the Morlet wavelets is proposed. First, the phase of grating image is extracted by the Morlet complex wavelet ridge, the phase variation between image frames is estimated, and the calculation model of the phase-shifting method is modified according to the phase variation. Finally, the phase of the captured grating image is extracted using the modified calculation model, and the 3D shape of the non-static object is reconstructed. The root-mean-square (RMS) error of the proposed method is 0. 081 6 millimeter for a non-static measurement object in the field of view. The RMS error of the proposed method is 16. 92% and 44. 42% of the PS method and the single frame projection method, respectively, and is the smallest among the comparison methods. According to the measurement result, the proposed method expands the application scope of the PS method, and can measure the 3D shape for a non-static object accurately.

Abstract:Applying the inspection robot to the automatic reading of the meter in the substation can reduce the labor cost and improve the work efficiency, but because the instrument is usually suspended in the high place, the posture of the robot is severely restricted in order to obtain accurate reading. In addition, the time consumption of PTZ adjustment during robot reading greatly reduces the inspection efficiency. In order to solve this problem, by analyzing the observation window constraint, the route map constraint, and the time consumption of PTZ adjustment, at the cost of the total time for the robot to complete the inspection task, a robot path planning model for meter reading is established. Then, an ant colony optimization algorithm based on pheromone reuse is proposed to solve the inspection path and parking scheme of the robot. Simulation results show that the time consumption of the inspection path obtained by this method is 66% less than the initial time cost, thus verifying the effectiveness of the model and the feasibility of the algorithm.

Abstract:In response to the requirements of large optical remote sensing satellites for high reliability, high performance and ease-touse, an integrated electronic system with high fault tolerance is presented. The design of system architecture and information flow are showed, and the key design technologies such as satellite operation management and system fault tolerance mechanism are discussed exhaustively. The on-orbit application of the designed integrated electronic system on a certain satellite is introduced. The satellite fully adopts the multi-redundant backup architecture, fault-tolerant technology and operation management technology proposed in the paper, and completes 21 imaging and 11 transmissions per day on average. The average daily shooting area is as high as 1. 1 million square kilometers, and the longest continuous trouble-free operation time is more than 6 months, which fully verifies that the designed integrated electronic system has the ability to operate efficiently and stably in orbit.

Abstract:Aiming at the problem of the extremely low visibility of water scene low at night, which results in the difficulty in detecting and locating personnel targets, the author combines infrared thermal imaging technology with deep learning object detection algorithm to study an object detection method for people in dark water area. After multi-scene field collection, a set of human target data set IR-YZ in thermal imaging water scene was independently constructed. On the basis of the performance of the IR-YZ data set and compared with the classical object detection methods, environmental characteristics, an enhanced lightweight water object detection network infrared water person target-YOLO is proposed, featuring the characteristics of thermal imaging and water areas. The experimental results show that the IWPT-YOLO algorithm has the advantages of being more accurate, faster and more concise than those of the classical algorithm. The model size is 93 MB, the average precision mAP reaches 85. 34%, and the detection speed reaches 20. 975 FPS. Compared with the classic algorithm YOLOv3 network and SSD network, the model size, average precision and detection speed are all improved. It verifies that the IWPT-YOLO algorithm has better detection performance and more obvious advantages for the characteristics of thermal imaging and water areas.

Abstract:Aiming at the variety of building change types in the process of urban development, this paper proposed a high-resolution remote sensing building change detection method based on D-S evidence theory. Based on the results of multi-scale image segmentation, a non-building index NBI is designed by combining multiple factors at first. On this basis, the multi-temporal NBI, traditional building index MBI and differential features are combined to construct the building change evidence set. Finally, an evidence confidence index combined with shadow detection is proposed, then a complete set of D-S evidence theory change detection model is constructed. Thus, the buildings can be divided into new, demolished and rebuilt categories. The experimental results of images from different regions show that the change detection accuracy and Kappa coefficient of the proposed model can reach more than 80% and above 0. 7 respectively, which are better than contrast method in both visual analysis and quantitative evaluation.

Abstract:Aiming at the problems of low positioning accuracy and slow speed of target objects in the magnetic levitation control system, a novel YOLOv5 (you only look once v5) algorithm was proposed to identify and locate the magnetic levitation ball. Firstly, by using the Mish loss function to replace the SiLU ( sigmoid-weighted linear units) activation function of YOLOv5 model, the higher accuracy and stronger generalization network model could be obtained. Then fusing the coordinate attention module into YOLOv5, the feature extraction capability of the model could be improved. On this basis, the CIOU ( complete-intersection over union) loss function was selected to replace the GIOU ( generalized intersection over union) loss function to improve the identification accuracy. Finally, the simulation verification was carried out. The results showed that the improved YOLOv5 algorithm could improve the target recognition accuracy of the magnetic levitation ball from 92. 4% to 96. 2%, and the MAP (mean average precision) from the original 88. 8% to 94. 3%. Therefore, the effectiveness and feasibility of the proposed method could be verified.

Abstract:In fringe projection three-dimensional measurement, double N-step phase-shifting profilometry compensates the phase error caused by the nonlinear response of the measurement system by expand twice the number of projection fringes, but the measurement efficiency is also reduced. To address the above concerns, an improved double N-step phase-shifting method is proposed in this paper. Compared with the traditional double N-step phase-shifting method, the proposed method can reduce the number of phase-shifting fringes while maintaining the measurement accuracy. The original and additional phase are calculated by using the deleted fringes, and the two phases are fused to realize the three-dimensional reconstruction of the object. Experimental results demonstrate that the proposed method has the same phase error compensation accuracy as the traditional double N-step phase-shifting method, and the measurement efficiency of the proposed method is improved by 16. 7%. At the same time, it is confirmed that the reliability of the three-frequency hierarchical phase unwrapping used in this paper is better than that of the three-frequency heterodyne method.

Abstract:Aiming at the problems of slow speed and high false detection rate of traditional tunnel fire detection methods, a real-time flame detection algorithm based on YOLOv5s was proposed, the size of anchorage frame was recalculated by K-means. In this paper, a fusion algorithm of YOLOv5s-SRGAN is proposed. The recall rate of 1 326 tunnel flame images is 94%, 1. 7 times that of YOLOv5s. CBAM attention mechanism module and gradient equalization mechanism were introduced to improve the performance of the model through feature extraction network and loss function respectively. Compared with YOLOv5s, the average accuracy of flame detection (IOU= 0. 5) is increased by 44%, the average detection speed of the test set reached 32 FPS. The results show that the improved flame detection algorithm has better recognition effect on small flame targets.

Abstract:As the main power source for aircrafts, the reliability of an aeroengine is critical for ensuring the safety of aircrafts. remaining useful life (RUL) prediction is of great importance for improving the availability of an aero engine and reducing its life cycle cost. For the problem of the shortcomings of existing estimation algorithms in the extraction of multi-dimensional data features, this paper proposes an attention-based CNN-BiLSTM model for RUL estimation. This model using CNN layers to extra feature and BILSTM network can capture the short-term and long-term dependencies of the extracted feature. Afterwards, attention mechanism layer is used to highlight the important features in order to improve model performance. To evaluate the effectiveness of our approach, experiments are carried out on CMAPSS datasets and its result shows that the performance of the proposed approach is superior to other traditional approaches.

Abstract:The intrusion of metal foreign objects will lower the efficiency and stability of wireless power transfer systems, even causing safety issues, thus it is extremely essential to achieve metal object detection. Aiming at the problem that existing technologies are subjected to blind-zone and cannot detect small foreign objects, a metal object detection method that combines deep learning-based object segmentation and support vector machine (SVM)-based object classification is proposed. First, object segmentation is performed using a YOLO v3 neural network based on RGB image of the charging area. Then the corresponding hyperspectral images of each object region are classified by the SVM. Finally, an experimental platform is built to verify the effectiveness of the proposed method. Results show that the proposed method not only detects tiny metal object such as a screw nut and a paper clip, but also has the potential to detect metal objects wrapped by non-metal material. Compared with the pixel-by-pixel detection using SVM alone, the proposed method improves the detection speed by about 38. 9%.