Abstract:In order to realize the miniaturization and digitization design of ground penetrating radar (GPR), this paper studies the design of pseudo-random (PRN) coded GPR system. A dual-channel pseudo-random coded ultra-wideband (UWB) GPR controller is presented in this paper. A dual-channel golay complementary pairs coded signal with a center frequency of 60MHz and 800MHz is transmitted and received by using the symbol balanced direct transmission signal source design based on the Virtex-5 and space-level ADC chip ADS5463-SP. The signal source test and the closed-loop test of the controller are carried out. The peak sidelobe ratios of the two-channel pulse compression results are greater than 25dB, and the distance resolution of 1.875m in low frequency and 11.72cm in high frequency is achieved. The controller is connected to the radar system, and the high frequency channel test is carried out in a sand pit. The thickness of the buried marble slab is measured to be about 17cm. The results show the performance of the radar controller designed in this paper is reliable, and it can be widely used in the PRN UWB GPR system.

Abstract:In order to achieve real-time and accurate fall prediction, and transplant the deep learning model to run on wearable devices, a lightweight convolutional neural network model is proposed. Drawing on the lightweight model idea of DSC network, the network structure is designed, and the number of channels and the size of convolution kernel are optimized, which greatly reduces the computational complexity of the model while keeping the accuracy rate basically unchanged. In order to deploy the algorithm in the wearable fall protection device, a real-time running framework of the model on the embedded side is proposed, and the algorithm is written as a C program and transplanted to the STM32 microcontroller. This model achieves 97.5% accuracy with 204.3ms lead time on the Sisfall dataset. The transplanted model is only 11.65KB in size, and the algorithm delay in the STM32 microcontroller is only 8.24ms. The experimental results show that the model has high prediction accuracy and good real-time performance, which provides a further reference for the development of fall prediction algorithms and fall protection devices.

Abstract:Quench protection system plays an important role in the operation of superconducting tokamak. When quench occurs, it quickly transfers the magnetic field energy stored in the coil to protect the superconducting magnet. Due to the high power of quench protection system, multiple solid-state switch modules must be connected in parallel, so the current sharing problem needs to be further solved. After selecting IGBT as the solid-state switch in the quench protection system, this paper analyzes the factors affecting its current sharing. The dynamic current sharing when IGBT is turned off is carried out by adjusting the grid resistance, and the static current sharing is realized by PI control of the grid voltage. The simulation and experimental results show the effectiveness of the control scheme at 220A current. Compared with the current sharing method of traditional branch series impedance, this scheme does not need to change the main circuit, with simple control mode and better flexibility. At the same time, it can effectively resist the influence of interference, and further promote the practical application of solid-state switch in quench protection system.

Abstract:Aiming at the problem that liquid surface features are few and discrimination is low, which is difficult to be recognized and detected effectively by machine vision, two laser light sources with different wavelengths are used to irradiate liquid at the same time to improve the discrimination between different liquids. An automatic collection device of data set is designed to provide a large number of effective samples for model training and then a visual recognition model based on EfficientNetV2 deep neural network was constructed. After introducing cosine learning rate dacay into the model and regulating super-parameters to the best, the optimal method was formed to realize efficient training, and the prediction accuracy was further improved. The results showed that the visual detection system could obtain 100% prediction accuracy, and successfully solved the problem of few features in liquid visual detection.

Abstract:In this paper, new designs of slot elements and a new configuration for array antennas for synthetic aperture radar at L- and Ku-band are proposed. The L-band are cavity-backed crossed slots for dual-polarization, and for Ku-band, waveguide slot arrays for single-polarization are used. The height of the L-band cavity is reduced by adopting an “inverted T” configuration. Each of the crossed slots is fed by a pair of differential probes. The Ku-band slot array is fed by a coax probe although waveguide divider. An initial prototype array comprising 2×2 L-band elements interleaved with Ku-band arrays was fabricated and measured. The impedance bandwidth (VSWR<2) of the antenna are dramatically broadened to 22% at L-band and 8.6% at Ku-band. This element design is applicable for larger Shared-Aperture arrays.

Abstract:In order to solve the problem of low recognition accuracy and poor positioning accuracy of steel bar binding robot, a target recognition and positioning method of steel bar binding robot based on deep learning is proposed. Firstly, YOLOv4 algorithm is used to identify and cut the target frame of binding point, and the initial positioning of binding point is completed. Secondly, the contour corner selection method is designed to calculate the image coordinates of the binding points by corner points. Then the feature extraction part of Monodepth algorithm is improved by integrating CBAM attention mechanism, and the decoding part introduces path aggregation network structure to improve the feature extraction ability of the model and further improve the stereo matching accuracy. Finally, the depth information of binding points is obtained by binocular stereo vision positioning technology, and the coordinate transformation is used to solve the mapping relationship between the hand-eye coordinate system of the steel bar binding robot, so as to realize the accurate identification and positioning of binding points. The experimental results show that the recognition accuracy of this method for binding point target frame reaches 99.75%, and the number of frames per second reaches 54.65. The Maximum error of positioning accuracy in space is 11.6mm. It can better identify and locate the binding point position, and provide strong support for automatic binding work.

Abstract:Aluminum metal is often used as thermocouple strips in infrared thermopiles because of its good Seebeck coefficient difference with polysilicon and low cost, the preparation of thermocouple strips is crucial in the preparation of MEMS process, and its structure and morphology have a great influence on the thermopile performance. In order to study the effect of different preparation methods on the morphology and performance of aluminum thermocouple strips, this experiment used metal etching process and stripping process to prepare thermocouple strips, adjusted sputtering power, photoresist thickness, exposure dose, ultrasonic power and other parameters to optimize the preparation process, and characterized the morphology by confocal microscope, SEM and step meter, and the resistance value by semiconductor analyzer. The experiments show that the high aspect ratio metal-aluminum thermocouple strips with width of 3 μm and thickness of 0.4 μm are prepared by ROL-7133 negative adhesive, pre-baking for 1 min30s, exposure dose of 85 mj/cm2, mid-baking for 1 min40s, and development for 48s by metal stripping process, and the overall morphology is good and the device resistance value meets the requirements.

Abstract:One type of military radar is a complex electronic equipment, the service time of a certain type of military radar has increased year by year, and the faults in the use process have increased significantly. Due to the weak maintenance capability at the grass-roots level, the maintenance support of the radar basically depends on factory support, and the timeliness and effectiveness can’t meet the needs of the army, which seriously restricts the formation of the combat effectiveness of the radar unit. In order to solve the problem of difficult maintenance of radar in grass-roots units and improve the comprehensive support ability of radar, a portable fault diagnosis system is developed. The system is composed of signal acquisition board, embedded motherboard, power management module and human-computer interaction module. Neural network technology is used to realize rapid fault diagnosis. It solves the problems of the existing radar maintenance equipment, such as huge volume, complex supporting equipment, high price, complex operation and so on. The experimental results show that the system realizes the miniaturization and intellectualization of fault diagnosis platform, has high fault diagnosis efficiency, and can quickly improve the maintenance and support ability of radar.

Abstract:In view of the trend of weak subsynchronous oscillation signal in the normal operation of power system, poor noise resistance and low reliability of identification results, a identification method of subsynchronous oscillation mode parameter based on deep residual network is proposed.A deep residual network model composed of convolutional layer, several residual layer and fully connected layer is established; the model training data set is generated according to the characteristics of SSO signal, all using simulation data; the parameter adjusted and optimized model can realize the blind identification of low SSO signal mode parameters measured in the field.Using ideal signal, noise simulation signal and field measured data three schemes of the model performance verification, the results show that the algorithm can effectively identify the weak SSO frequency and damping and other key parameters, compared with convolutional neural network (CNN) and random subspace (SSI) algorithm, higher accuracy, small noise interference, has the characteristics of blind identification, can be used for power system secondary synchronous oscillation risk warning.

Abstract:In the process of oil and gas pipeline leak detection, the leak signal contains a lot of noise and the feature extraction is difficult. An improved total average empirical mode decomposition combined with Kalman filter algorithm is proposed to denoise the pipeline signal. First, the improved overall average empirical mode algorithm is used to decompose the collected pipeline negative pressure wave signal. The permutation entropy and Kalman filter algorithm are used to filter and process the decomposed inherent modal components, and finally the reconstructed Cut the noise signal. Furthermore, a feature extraction method based on diffusion entropy and kurtosis is proposed, the extracted feature parameters are used as the input of support vector machine to classify and recognize the working conditions of oil pipelines. The collected data verify that the improved overall average empirical mode decomposition, Kalman filter, spread entropy and kurtosis combined recognition method can more accurately classify and recognize pipeline signals, and the results show that the total average recognition accuracy is 98.89. %, it provides a new way for the research of pipeline working condition identification.

Abstract:The generation of Surface Electromyography is ahead of the occurrence of body movement and has the ability to predict body movement, which often assists patients in rehabilitation training. To solve the problem that single channel sEMG signal is difficult to predict people' joint angles effectively, this paper proposed a maximum Relevance minimum Redundancy based on multi-channel EMG feature acquisition and Particle Swarm Optimization feature selection algorithm. The performance of mRMR-PSO algorithm was verified by comparing with that of mRMR algorithm and Principal Component Analysis algorithm for joint Angle prediction accuracy. Experimental results show that the joint angle prediction accuracy of mRMR-PSO based on multi-channel feature selection algorithm is 32.6% and 14.9% higher than that of mRMR and PCA, respectively, which verifies the effectiveness of the algorithm,and the algorithm is applied to actual scenarios.

Abstract:A dynamic terminal sliding mode control method based on disturbance observer is presented to solve the problem of speed and position error jump when a mobile robot is subject to external irregular disturbance. Based on the kinematics model, a virtual controller is designed using Lyapunov method, and a non-singular dynamic terminal sliding mode track tracking controller is further designed. To reduce the influence of external disturbance on the system, a nonlinear disturbance observer is designed to compensate the disturbance of the controller. Finally, the method proposed in this paper is compared with the adaptive sliding mode control method by simulation. The results show that the line speed and angular speed of the adaptive sliding mode control method change by 1.4 m/s and 1.24 rad/s respectively when the 15s disturbance changes step by step. The speed jump amplitude of the method presented in this paper is less than 1/10 of that of the adaptive sliding mode control method. The simulation results show that the proposed method can effectively suppress the influence of disturbance on the system and reduce the jump amplitude of the speed and posture errors of the mobile robot.

Abstract:In order to solve the problem of decreasing detection performance of traditional adaptive constant false alarm detector (CFAR) in multi-target environment, the selection strategy of traditional exponential transform-based CFAR (VI-CFAR) is improved. An improved adaptive constant false alarm detector VIHCES-CFAR is proposed. Heterogeneous Clutter Estimating CFAR (HCE-CFAR) and Switching CFAR (S-CFAR) are selected in Clutter edge environment and multi-object interference environment respectively, which improves the ability of target detection in multi-target environment and effectively avoids the problem of target masking. Experimental results show that the SNR of VIHCES-CFAR is 0.05dB lower than that of SVI-CFAR at the detection probability of 0.5, and the detection probability of VIHCES-CFAR is 99.78% in the multi-target environment, and the detection probability of VIHCES-CFAR is about 10-4 in the clutter edge environment. It has stable anti- interference ability and good false alarm control ability. The effectiveness of VIHCES-CFAR detector in solving the problem of target masking is verified by the measured data.

Abstract:Aiming at the problems of many parameters, complex data structure, high frequency of acquisition but difficult to guarantee data quality in industrial data acquisition. A time window data acquisition method based on multi-domain concurrency is proposed. This method build a data domain model based on the division of sliding time windows. The multi-source heterogeneous data is acquired by the method of multi-threaded data domain concurrency and processed by flattening and connecting the byte data. Finally, the acquisition and integration of multi-source heterogeneous data is accomplished. The effectiveness of this method for real-time data acquisition and integrated storage is verified by simulation experiments. The method is applied in the multi-dimensional health monitoring system of metallurgical cranes for data acquisition. The actual operation results show that the method can improve the real-time performance of multi-dimensional health data acquisition of metallurgical cranes, ensure the quality of the acquired data and increase the increase integration of multi-dimensional health data. This method is an effective method of real-time data acquisition and integration.

Abstract:Aiming at the problem of measurement error caused by signal drift of MEMS acceleration sensor in inertial measurement system, the measured data of MEMS acceleration sensor are analyzed by time series analysis method. After reading the data measured by MEMS acceleration sensor through DSP, the stability is tested by ADF criterion. The sensor data meets the stationary time series conditions. According to the characteristics of autocorrelation function and partial autocorrelation function of sensor data, it is judged that the sequence satisfies AR (P) model. Through AIC criterion for randomness test, time series model identification and parameter estimation, the sensor data is optimized by using AR (1) model. The signal drift AR (1) model of MEMS acceleration sensor is established, and the Kalman filter is designed according to the model. The results show that the zero bias stability of the acceleration sensor before filtering is 0.3032mg, and the zero bias stability of the acceleration sensor after Kalman filtering is 0.0247mg. The measurement stability is effectively improved, and the operation order is low, which can be well applied to the embedded system.

Abstract:In distributed Raman temperature measuring system, the temperature of the contained information of anti stokes light is weak, easily masked by noise, and the traditional denoising method is easy to filter to remove the original characteristics of the system, thus aiming at these problems, this paper USES the EEMD combined with wavelet threshold denoising method to reduce the noise of engine, under the condition of retain the original information feature, Greatly improve the system's signal to noise ratio (SNR) and temperature measurement accuracy. Simulation results show that the SNR of the algorithm is improved by 2.3 dB when the noise is 1dB. Experimental results show that the measurement accuracy of the system is improved by 76.11%. Through a separate wavelet threshold de-noising process, the temperature curve of the fusion point of the optical fiber is more than 1℃ fluctuation range, and the demodulation temperature fluctuation range of the system is about 0.5℃ after the algorithm is adopted. Finally, the experiment proves that the algorithm does not affect the spatial resolution of the system.

Abstract:In the launch vehicle parameter measurement system, the traditional sensor network is connected by cables. With the increase of test points, the use of a large number of cables not only brings excess load and power consumption, but also takes up a lot of cabin space , causing huge trouble for the layout and installation of other equipment in the rocket cabin. In response to the above problems, we propose a CC1310-based wireless sensor network (WSN) design scheme to realize the cableless test system. In this design, the ultra-low-power RF chip CC1310 with a communication frequency band below 1GHz (Sub-GHz) is selected as the sensor network controller, the low-power K-type thermocouple AD8495 is used to collect temperature data, and the lithium battery is used for power supply. The software design implements the periodic sleep and wake-up listening mechanism of sensor nodes, and the master node polls and accesses sensor nodes to obtain the function of status and data. The node low power consumption design is completed and the reliability of the data can be guaranteed. The experimental test shows that the system has a reasonable design, low power consumption, stable and reliable data transmission, and the average energy consumption of the system in working state is 88μA, which can realize long-term measurement of environmental parameters in the rocket cabin.

Abstract:Aiming at the disadvantages of wired transmission of hydraulic parameters of drilling pump and pumping unit by pressure transmitter in oil field, as well as the low efficiency and poor real-time performance of manual regular inspection, a wireless pressure acquisition system using ZigBee communication technology is designed from the actual needs of the field. Its hardware circuit is mainly composed of power supply management module circuit, USB to serial port circuit, MSP430 microprocessor and AD conversion circuit, pt2407 flush film pressure transmitter interface circuit and ZigBee communication module circuit. The design concept of the system adopts star network topology, which improves the anti-interference ability of the system. The laboratory and oilfield field tests show that the relative error of the pressure measurement value of the wireless pressure acquisition system is 0.01% ~ 0.5%, and the effective transmission distance is up to 100 meters, which overcomes the shortcomings of traditional physical wiring. The wireless mode can accurately collect and transmit data. The wireless pressure acquisition system has the advantages of low power consumption, easy networking and wireless, and greatly improves the production efficiency.

Abstract:Under the background of large amount of data support, how to use a large number of SAR images efficiently and improve the accuracy of ship target detection is the current problem of ship target detection. This paper focuses on how to improve the accuracy of YOLOv4 algorithm for SAR ship target detection, and presents a YOLOv4 enhancement algorithm that combines multiscale and attention enhancement. The Attention Module (CBAM) is added to the PANet of the original YOLOv4, and the enhanced K-means clustering algorithm is used to cluster the ship target real frame in the dataset, and the result of the anchor frame is transformed linearly to make the algorithm anchor frame more suitable for the training set. Experiments show that the average accuracy () of the proposed algorithm in SAR ship detection is 94.05%, which is 0.7% higher than that of the original YOLOv4. The experimental results fully demonstrate that the proposed algorithm can improve the accuracy of SAR ship image detection and provide technical support for the accuracy of sea activities judgment.

Abstract:Color ring resistors are commonly used electronic components, and their resistance value is mainly represented by the color ring. The color ring relies on manual judgment, which is inefficient and has a high false detection rate. Traditional color ring judgment is based on image processing, which is not robust and is greatly affected by physical factors such as illumination. Based on this, the paper proposed a color ring detection and resistance value interpretation method based on deep learning. Firstly, the proposed object detection algorithm was used to realize the color ring detection and the resistance body detection. Secondly, the proposed color relationship matching method was used to combine the detection results to judge the subordination relationship between the color ring and the resistance body and sort the color ring. Finally, using the proposed resistance inference method, combined with the color code table, the real-time detection and interpretation of the color ring resistance was completed. The experimental results show that the algorithm has better performance in the accuracy of color ring detection compared with other detection algorithms, reaching 98.71%. The parameter amount is only 10.61M and the calculation amount is only 31.68GMAC. Randomly select 20 pictures on the test set for verification, and the accuracy of resistor interpretation is as high as 98.59%.

Abstract:In order to meet the requirement of micron scale measurement for many small industrial parts, a measuring method combining point cloud multiple filtering and plane fitting was proposed.A 3D line laser sensor was used to obtain the point cloud model of the regular triangular prism, which was parallel to the upper and lower planes, and the point cloud model was transferred to the computer for processing. Firstly, the noise and outliers were removed by statistical filtering.Secondly, the number of point clouds is reduced by voxel filtering and down-sampling.Then the upper and lower surfaces of the workpiece point cloud are separated by straight-through filtering.Then, the point clouds on the upper and lower surfaces are fitted with the plane equation by RANSAC algorithm.Finally, the distance between the upper and lower planes is calculated as the height information of the measured workpiece.The height measured by this method is compared with that measured by laser triangulation. The results show that the accuracy of this method is improved by 72.33%.At the same time, for different point cloud densities, the proposed method is used to measure, and the measurement error is minimum when the side length of the voxel cube in the sample is 15cm (when the number of point clouds is reduced by 98.3%), and the minimum can reach 5.1𝜇m.This method greatly improves the measuring accuracy of workpiece and can be widely used in industrial measurement.

Abstract:Aiming at the problem that the recurrent neural network has a single feature extraction and insufficient processing of spatial information of the feature, a two-branch fusion human behavior recognition model based on bone is proposed. The model is extracted by the two-branched network of two-way cyclic gate network and multi-scale residual network, which obtains rich feature information in time and space, and increases the attention mechanism in the bidirectional cyclic gate network to further improve the performance of the whole network, and finally the feature information is classified through the classifier to obtain action. Experiments were conducted using the UCF101 and HMDB51 datasets, respectively, with an accuracy rate of 98.0% and 67.8%, respectively. Through experimental tests, it is proved that the model can obtain more complete feature information and has good performance indicators.

Abstract:Aiming at the problems of fuzzy target recognition, poor classification efficiency and slow response speed of parallel robot in the field of machine vision, a positioning and detection technology of parallel robot based on deep learning is proposed. Firstly, put the parallel robot into the image collection set to improve the image recognition accuracy and improve the object recognition efficiency; Secondly, improve the training mode, improve the reliability and loss strategy through pre-training and actual training; Then, the base coordinate system and camera coordinate system of the parallel robot are established. Combined with the hand eye calibration and camera calibration methods, the transformation relationship between the actual coordinates of the target and the base coordinate system of the robot is obtained; Finally, the target calibration results are verified on the parallel robot experimental platform. Compared with the relative error between the network positioning and actual positioning of the parallel mechanism obtained by the mainstream deep learning algorithms YOLOv3, YOLOv4 and Faster-RCNN, the results show that the positioning accuracy error of YOLOX is about 3.992-5.061mm, and the average accuracy is about 91%. This method can provide a certain reference value for the detection and positioning of parallel robot combined with deep learning.

Abstract:Aiming at the problem of insufficient robustness of existing target detection algorithms in surface garbage detection due to the interference of illumination, water ripple and reflection in images, a surface garbage significance detection method combining spatial prior information and frequency-domain phase spectrum was proposed. Based on background prior, local contrast prior and dark region prior information, the minimum obstacle distance map, contrast map and background map were fused in spatial domain to obtain the initial saliency map of surface garbage. In the frequency domain, the phase spectrum of the image is reweighted by low rank decomposition to obtain a significant target with less redundancy. Experimental results show that the accuracy of this method can reach 96.4%, and the interference of ripple, light and reflection can be effectively suppressed.

Abstract:In order to solve the problem of the wind turbine blade surface defect defection which has low detection recognition rate and can be easily affected by the light, this paper puts forward a wind turbine blade surface defect detection method based on convolutional neural network which combines local binary patterns with core extreme learning machine. The convolutional neural network introducing attention mechanism is used to extract deep information of images. Then, local binary patterns characteristics which can describe shallow texture information of images are also exacted. Besides, the principal component analysis can reduce local binary patterns characteristic dimension. Serial combination is then done to these two complementary characteristics which can describe images from different levels and the improved sparrow search algorithm is used to optimize Kernel extreme learning machine parameters. Besides, the syncretic feature training model is used to obtain the optimal model for defect recognition. The experiment shows that the classification accuracy rate after the training of self-built data sets can reach 97.5% and that the kappa coefficient can reach 95.1. Compared with single feature detection, the classification accuracy is significantly improved. The actual verification of the wind farm shows that the average classification accuracy of the model is 96.3%, the kappa coefficient is 94.5, and the missing rate is significantly reduced.

Abstract:With the continuous progress of deep learning theory, end-to-end stereo matching network has achieved remarkable results in the fields of automatic driving and depth sensing. However, the most advanced stereo matching algorithm still have trouble in accurately recover the edge contour information of the object. In order to improve the accuracy of disparity prediction, in this study, we propose a stereo matching algorithm based on edge detection and attention mechanism. The algorithm learns parallax information from stereo image pairs and supports end-to-end multi task prediction of parallax map and edge map. In order to make full use of the edge information learned by the two-dimensional feature extraction network, we propose a new edge detection branch and multi feature fusion matching cost volume. The results show that the edge detection scheme based on the model helps to improve the accuracy of parallax estimation. The error matching rate of the obtained parallax map on KITTI 2015 test platform is 1.75%. Compared with pyramid stereo matching network, the accuracy of parallax map is improved by 12% and the running time is reduced by 20%.