Abstract:Aiming at the problem that the pointing accuracy of star tracker bracket is reduced due to thermal deformation, a pointing inclination monitoring method based on neural network is proposed. Firstly, the structural characteristics of the star tracker bracket are analyzed, the pointing inclination prediction system of the star tracker bracket is built, the structural deformation and inclination change data of the star tracker bracket are collected, and the experimental data are preprocessed. Secondly, the depth neural network model is constructed, the strain information of each measurement point of the star tracker bracket model is taken as the input variable, and the network parameters are updated by Adam optimization algorithm. After training iteration, the pointing inclination prediction model is obtained. Then, aiming at the limitation of slow convergence and easy to produce local minimum value of traditional deep neural network, genetic algorithm is used to optimize the hyperparameters of deep neural network to improve the training efficiency. Finally, the test set data is used to predict the change of the pointing angle of the star tracker bracket, and the accuracy of the model in monitoring the pointing angle of the star tracker bracket under different temperature conditions is analyzed. The experimental results show that the average error of the directional inclination prediction method of the optimized depth neural network model is 0. 20" , and the accuracy of inclination prediction is significantly better than the traditional algorithm, which proves that the deep learning method is feasible to realize the directional inclination monitoring of star tracker bracket.

Abstract:In order to improve the diving efficiency of the penetrator, this paper uses the mass-spring system to establish the dynamic equation of the single-cycle diving process of the penetrator, obtains the optimal mass stiffness matching relationship for the efficient diving of the penetrator through the ergodic optimization method, and simulates its motion state in the working process based on ADAMS. Based on this, a high-performance lunar surface penetrator prototype was designed and tested. The test results show that the penetration efficiency of the new prototype is 63. 7% higher than the original one, and the stability is well. The motion of each component is coordinated, and the actual depth of the diving is in good agreement with theoretical and simulations, which is of good theoretical interest and important application value.

Abstract:The planar capacitance tomography (ECT) technique is usually used to reconstruct the image of bonding defects of composite materials by means of sensitivity matrix. The sensitive field of ECT sensor is distributed in a three-dimensional space, but the reconstructed image for defect detection is two-dimensional, so there is the problem of selecting the sensitivity matrix of different layers. The selection of sensitivity matrix in different layers is explored, and an improved method of sensitivity matrix is proposed. Through the threshold processing and absolute value processing of the negative sensitive region, the ill condition and non-uniformity of the sensitivity matrix are improved, and the modeling error of ECT forward problem is reduced. Simulation and experimental results show that the proposed method can significantly reduce the artifacts in the reconstructed image, improve the clarity of the reconstructed defects, and then reduce the probability of false or missed judgment.

Abstract:The hardened layer depth (HLD) of metal parts is one of the important factors to determine its wear resistance and fatigue strength. Ultrasonic backscatter method can realize nondestructive testing for HLD. However, for those induction hardened parts with large transition layer, it is difficult to accurately locate the limit hardness point, which leads to inaccurate testing results. In this work, HLD measurement experiments of induction quenched 35MnB alloy were carried out. The wavelet multi-resolution analysis (MRA) was employed to find the characteristics of the time-frequency domain of ultrasonic backscattering signals at different decompsosition scales. The energy distribution in each frequency band was analyzed, and the trend characteristics of ultrasonic backscattering signals were extracted to locate the limit hardness point in transition layer of the induction quenched samples. The result shows that the original backscattering signals contain obvious rapidly-oscillating components with low energy and irregular distribution. The signal trend varies slowly with the amplitude of the oscillation with energy portion as high as 96. 73%, which is significantly higher than other decomposition terms. The MRA process essentially filters out most of the clutter noise components irrelative to the depth information of the hardened layer. The actual measurement result shows that the maximum average error of this method is 0. 123 mm, the maximum repeatability error is 6. 24%, showing that the present method achieves high accuracy and repetitive reliability. Compared with metallography and microhardness, this method is more efficient and nondestructive, which shows a good practical application prospect.

Abstract:Ultra-small hand-held multi-channel ultrasonic system is a highly integrated system. This system has some advantages in nondestructive testing. Such as the size of the system is small, the detection with the system is convenient and so on. As for the design of this system, the 64-channel ultrasonic detection system in this paper is established. The consequence of ultrasonic detection system design is delay precision as 5 ns, sampling frequency as 100 MHz, bandwidth as 0. 02~ 25 MHz, volume as 40×70×90 mm 3 . According to the results of the delay precision experiment, Synchronous sampling experiment, defect detection experiment and guided wave imaging experiment, the design dose work. The system can expand to more channels for application to medical ultrasonic detection and the volume will not increase too much. Meanwhile, its size is relatively small among the same type instruments.

Abstract:The dielectric constant is an important parameter that affects the performance of high-frequency substrate, and the accuracy and spatial distribution of the dielectric constant of the substrate will have an important impact on the performance of microwave circuits. In this paper, the resonant ring circuit of enhanced coupling ring, shielded through hole and coplanar waveguide technology is designed first, to enhance the anti-interference ability of the ring to the outside world, and to distribute S21 between -30 to 20 dB, and improve the transmission efficiency. Secondly, the influence of different ring average radius, ring width and coupling gap on the simulation results is analyzed. Then, the resonant ring circuit is processed and tested. The measured results show that the dielectric constant error is less than ±0. 007 at the frequency points of 2, 10 and 24 GHz, which proves that the method has high test accuracy. Finally, based on the test and analysis of the dielectric constant distribution of a domestic substrate based on the resonant ring circuit, the results show that the dielectric constant error caused by processing and the spatial distribution of the substrate is less than ± 0. 024, which proves that the substrate has good performance.

Abstract:Gait assessment is one of the most important part of motor assessment of stroke. The clinical gait function assessment is subjective and based on kinematic results rather than neuromuscular changes. Therefore, a muscle synergy analysis method based on surface electromyography (sEMG) signals was proposed for quantitative assessment of gait function. The non-negative matrix factorization (NMF) algorithm was used to decompose the pre-processed multichannel sEMG signals. The synergy stability index (SSI) was obtained to describe the similarity of synergy structure vectors under different tasks in different groups. The between-group comparison showed significant differences in SSI between the healthy control and the patient’ s unaffected leg ( p = 0. 010) and between the SSI of the patient’s affected leg (p = 0. 007) when the tested leg was turned medially; the SSI of the healthy control leg was significantly different from the patient’s affected leg when the tested leg was turned laterally (p = 0. 036). The between-task comparison showed that SSI of the patient’s unaffected leg was significantly higher in the lateral turn than in the medial turn (p = 0. 017); the SSI of the patient’s affected leg during the lateral turning was significantly correlated with the lower extremity portion of the clinical motor function scale FMA_LE (r = 0. 671, r 2 = 0. 451, p = 0. 033). The experimental results suggested that SSI under the turning task can provide an objective and quantitative means of analysis for the assessment of clinical gait function, providing a new way for quantitative assessment from a neuromuscular perspective.

Abstract:The use of channel simulators to reproduce the radio wave propagation environment in the laboratory is an important means to test ultra-wideband (UWB) communication equipment. Aiming at the frequency nonstationary characteristics of UWB channels, a UWB channel model based on stationary interval superposition in frequency domain is proposed in this paper, and a hardware real-time simulation system based on software radio platform is designed. The system uses a multi-channel data parallel processing architecture to realize real-time processing of high-rate data, and introduces an iterative algorithm and a time division multiplexing structure, which not only ensures real-time generation of channel fading but also reduces hardware resource consumption. The measured results show that the simulator can simulate and output the 600 MHz ultra-wideband signal after superposition non-stationary fading in real time. The measured results show that the simulator can support the simulation of 600 MHz UWB channel fading, the channel fading spectrum is consistent with the theoretical simulation results, and the output channel impulse response is basically consistent with the measured results. The system can provide an effective means for optimization, verification and evaluation of ultra-wideband communication systems.

Abstract:In order to realize the integrated diagnosis of multiple faults in the working process of electro-mechanical actuators (EMA), a fault diagnosis method of EMA based on dual-stage attention-based long short term memory (DaLSTM) combined model was proposed. Firstly, the multi-source sensor signal of the EMA is used as the input. The long short term memory (LSTM) neural network based on input attention and time attention is used to adaptively extract the relevant features in the original multi-source sensor data, and the time series prediction of multi-source sensors is realized by the DaLSTM combination model. Secondly, in the fault diagnosis time window, the minimum difference between the predicted value and the sampled value of the DaLSTM combination model under different states is used as the decision function to diagnose the fault type of EMA. Finally, time series prediction and fault diagnosis experiments are conducted using the public National Aeronautics and Space Administration (NASA) dataset, and the average recognition rate of fault categories reaches 98. 76%, which proves the effectiveness of the proposed method.

Abstract:In order to solve the semiconductor manufacturing defect detection with low efficiency, the error rate is high, the result is not stable, imaging accuracy is low and cannot accurately detect the problem such as different kinds of defects. In this paper, by using a custom CCD industrial camera with a high ratio of optical microscope scan images on the surface of the wafer, combined with the improved YOLOv4 algorithm, a high precision wafer defect detection method based on deep learning is implemented. Experimental results show that the proposed model can identify different kinds of silicon carbide wafer defects under various complex conditions and has good robustness. The average accuracy of defect identification is 99. 24%, which is about 10. 08% and 1. 92% higher than that of YOLOV4-Tiny and original YOLOv4, respectively. Compared with the Halcon-based method and OpenCV template matching method, the average recognition time of defects per graph reaches 0. 028 3 s, which is about 93. 42% and 90. 52% higher than other conventional wafer defect detection methods and has realized stable operation in independently designed verification systems and application platform.

Abstract:In order to meet the needs of spraying robots for spraying quality detection, the improved padim modeling and ResNet network are fused by using transfer learning to build an integrated model of spray quality detection for the autonomous spray robot. The model can be used for defect location and classification at the same time by extracting image features once. At the defect location end, the Padim model is improved to reduce the network computing consumption caused by feature redundancy, and the patch embedding vector semantic layer obtained by the ResNet-18 network is first changed from the original three layers to a single second layer. Then the dimension of feature expression is reduced from 100 to 20 dimensions. Finally, the normal distribution model obtained by training the positive samples and the test image are used for defect location. At the defect classification end, the pre-trained ResNet-18 network is re-trained with negative samples, and the ResNet-18 classification model is obtained to classify the test images for defects. After experiments, the integrated model is transplanted into the Jetson nano mobile terminal. The parameter quantity is 11. 69 M, the positioning accuracy is 94. 5%, and the classification accuracy is as high as 99. 6%. The detection time is 0. 730 s when the robot displacement speed is 0. 02 m/ s, and there will be no missing frame detection, which meets the requirements of real-time detection.

Abstract:Long-term traffic flow prediction is a vital part of comprehensive transportation system planning, it is also an important basis for formulating macro traffic flow management policy. In order to improve the accuracy and efficiency of traffic sequence prediction, a hybrid prediction model based on singular spectrum analysis (SSA) is proposed, in which the problems of noise and unstable prediction of a single model in time series prediction are well solved. Firstly, raw data are reconstructed into tendency term, periodic term and residual term using SSA. Specifically, trend term is predicted using support vector regression ( SVR), and grey wolf optimization (GWO) algorithm is introduced to optimize parameters of the regression model. Moreover, periodic term is predicted using forgetting online sequential-extreme learning machine ( FOS-ELM). Finally, the predicted results are obtained by superimposing the above mentioned two parts. The experiment is carried out with real traffic flow data, and the mean absolute error (MAE) and root mean square error (RMSE) of the proposed hybrid prediction model are 215. 15 and 278. 51, respectively. The overall results show that the proposed model can solve the problems of large error fluctuation and unstable prediction of the single model. Furthermore, compared with empirical mode decomposition (EMD) and unprocessed time series, the prediction error of each model to the time series after singular spectrum analysis is reduced, which further illustrates the effectiveness of SSA in time series decomposition.

Abstract:Aiming at the difference of bearing fault data feature distribution caused by complex working conditions in industrial scenes and the difficulty of obtaining a large number of labeled data, an one-dimensional convolution subdomain adaptive adversarial neural network (SANN) based on Wasserstein distance and local maximum mean discrepancy ( LMMD) is proposed. Firstly, the network constructs a feature extractor based on CNN for pre-training and learning the domain feature representation. In the adversarial training stage, the adversarial layer introduces Wasserstein distance to measure the difference between the source domain and the target domain, realize the alignment of marginal distribution and solidify the training results. In the feature extraction layer, LMMD calculation module is introduced to capture the fine-grained information of each category to realize the alignment of conditional distribution. The performance of the model is verified by the bearing fault data sets under two different working conditions. The experimental results show that under unsupervised conditions, the proposed method improves the recognition accuracy by 5. 0% and 6. 9% respectively compared with the basic domain adversarial network on the target data set, and the performance is better than the existing migration algorithms.

Abstract:Series arc fault is an important cause of electrical fire, and effective detection can ensure the normal operation of lines and reliable work of electrical equipment. According to the difficulty of low voltage series arc fault detection, a recognition model based on asymmetric convolutional neural network is proposed to extract series arc fault information adaptively. To solve the problems of series arc faults with many types and hidden information, firstly, the time-domain data processing method of Gramian difference angular field is used to map the time-domain signals simulated by load into two-dimensional matrix after polar coordinate transformation and trigonometric transformation, so as to increase the space occupancy of fault data points and data association information. Then, in order not to increase the time cost and improve the recognition efficiency of the model, the residual neural network is improved by adaptive asymmetric convolution and multi-channel discrete attention mechanism as the series arc fault model in low-voltage lines. Finally, a container is used to encapsulate the trained fault identification model to realize the fast analysis of fault information. Verification shows that the recognition rate of series arc fault can reach 99. 95%, and it has good recognition effect.

Abstract:As an indispensable part of modern encryption system, true random number generator (TRNG) plays a very important role in information security. A configurable, lightweight and high throughput true random number generator is presented. The structure uses NAND gates and XOR gates to form a configurable asynchronous feedback ring oscillator. By increasing the phase noise in a short time, the time jitter range is expanded, so as to improve the randomness of entropy source. The structure is tested and verified on Xilinx Kintex-7 for many times. The experimental results show that the proposed TRNG has strong robustness under the environmental changes of different temperatures (0 ℃ ~ 80 ℃ ) and different output voltages ( 0. 8 ~ 1. 2 V), and only consumes 43 LUTS and 6 DFFs in hardware resources, and obtains a throughput of up to 300 Mb / s. At the same time, the generated random bit stream can pass the NIST SP800-22 and NIST SP800-90B tests with high P values.

Abstract:In mixed speech separation, the performance of signal time-domain features is better than that of frequency-domain features. However, the current speech separation methods based on time domain feature have poor robustness in real noise environment, and single time domain feature has limitations on the performance of the separation model. Therefore, a multi-feature speech separation method based on Conv-TasNet network is proposed, which integrates frequency domain features and time domain features to improve multidimensional information of data. In order to further improve the performance of separation network, multi-scale convolution block is introduced to improve the feature extraction ability of network. Compared with the Conv-TasNet model and the latest time-frequency fusion speech separation baseline model, the performance and robustness of the proposed method are improved by 0. 91 and 0. 52 dB respectively in the experimental environment containing real noise.

Abstract:Aiming at the problem that the existing power vehicle cannot maintain constant power due to the change of pedaling rate in the test and training process, a new magnetoresistance generation device is proposed. Firstly, the magnetoresistance generation device is designed based on the principle of permanent magnet synchronous motor. The basic structural dimensions of the device are determined through the overall structure of the power car. Considering the maximum power required for the test, the electromagnetic load and select permanent magnets are determined. Then the stator and rotor structure size and winding connection form are designed, and the radial air gap parameter is specified. Finally, the system constant power control principle is analyzed. The system constant power control effect is simulated and experimented by fuzzy PID control and iterative learning control ( ILC ). The results show that the power error of the new magnetoresistive power vehicle can be controlled within ±5 W when the pedaling speed varies in the range of 60(±10) r/ min, which is acceptable. The power vehicle with the new magnetoresistance generation device improves the effect of constant power control and verifies its advancement in maintaining constant power.

Abstract:Allocating computing tasks to appropriate edge computing resources to meet the computing needs of users and improve the quality of service for user task requests is a key problem in edge computing. This paper proposes an edge computing task scheduling method (AFETSA) based on artificial fish swarm search. For improving the global search ability of the heuristic task scheduling algorithm and reducing the computation time delay, the artificial fish search algorithm was combined with the edge computing task scheduling model, and the field of view and step size of the artificial fish were dynamically adjusted by the nonlinear decreasing function. At the same time, for improving the optimization ability of the algorithm, the tabu search algorithm is fused, and the tabu list is introduced to prevent the algorithm from falling into local optimal in each iteration. The experimental evaluation results on the CloudSim3. 0 simulation platform, show that compared with the existing task scheduling algorithms AFSA, ACO and PSO, the proposed task scheduling method in this paper has significant improvement in task execution time, algorithm stability and load balance. It can make full use of the computing resources of edge servers to improve the computing performance of computing tasks, and effectively solve the problem of high delay and load imbalance caused by uneven task scheduling in edge computing.

Abstract:Accurate estimation of the state of charge ( SOC) of lithium-ion batteries is one of the key technologies in the battery management system, which has a vital impact on the service efficiency and safety of power battery pack. Lithium-ion batteries have complicated characteristics and SOC cannot be directly measured which are greatly affected by the current and temperature. Therefore, combining a gated recurrent unit (GRU) neural network with an unscented Kalman filter (UKF) algorithm is presented. The method uses GRU neural network to obtain the nonlinear relationship between the SOC and measurements, including the current, voltage, temperature. The relationship is used as the observation equation of UKF, and the SOC is estimated by the UKF to improve the accuracy and stability of estimation algorithm. Experimental results show that under different temperatures and different working conditions the root mean square error and the mean absolute error of the SOC estimate are less than 0. 51% and 0. 46%, respectively, which can improve the accuracy of SOC estimation.

Abstract:It is difficult to identify the fault features of rolling bearings under multiple working conditions. In this paper, one-dimensional multi-scale dense network (MSDNet) was applied to fault diagnosis of rolling bearings from the perspective of data-driven. Firstly, the time domain signal is used as the direct input of MSDNet to maintain the inherent characteristics of the signal. Secondly, three parallel convolution operations were used to extract multi-scale information inside the bearing fault signals. The addition of dense network prevented the loss of features in the process of information transmission, and alleviated the gradient disappearance problem in the model appropriately. Then, the Adabelief optimization algorithm is used to optimize the model parameters during the training process, which makes the model converge quickly and improve its generalization performance. Finally, confusion matrix and feature visualization were used to demonstrate the classification performance of the model. Several experiments have been carried out on Case Western Reserve University bearing datasets and Xi′an Jiaotong University datasets, and the fault recognition rate of the proposed algorithm can reach more than 98%, which proves the effectiveness of the proposed method.

Abstract:In order to better distinguish the pedestrian positioning information based on micro inertial measurement unit ( IMU), this paper deeply studies the traditional pedestrian dead reckoning (PDR) algorithm model, and finds that the traditional algorithm has single discrimination conditions, low accuracy and is not suitable for scenes with a variety of terrain. Aiming at the problem of inaccurate step estimation model in traditional algorithms, this study first proposes an error compensation optimization algorithm based on extended Kalman filter (EKF) to realize the error compensation of accelerometer, gyroscope and other sensors integrated in IMU. The study put the optimized original data into BP neural network algorithm to train the single parameter step estimation empirical model. The experimental results show that the step size algorithm based on BP neural network fusion basic model can improve the closed-loop accuracy by more than 0. 3% and reduce the open-loop error by 8. 5 times compared with the simple basic step size model. The improved PDR algorithm based on BP neural network can effectively suppress the error dispersion of inertial algorithm.

Abstract:The 2D shape sensing based on optical frequency domain Reflectometry (OFDR) is limited by the coherent noise and the spectral mismatch, and the shape sensing accuracy is seriously restricted with the chaotic and ghost peaks. In this work, we proposed an accuracy enhancement method for a 2D shape sensing based on OFDR system utilizing a median filter. The measurement principle of the OFDR is analyzed. Besides, the proposed method and the Frenet-Serret framework are demonstrated. The 2D shape sensing experiment configuration is set up. And a distributed static strain experiment is conducted to calibrate the strain-wavelength shift coefficient. Then, the proposed method is used for noise reduction of the Rayleigh backscattering wavelength offset. The experimental results show that the strain-wavelength offset coefficient is 1. 20 με / pm. At the length of 0. 5 m, the end errors decreased from 3. 08%, 0. 94%, and 0. 82% to 0. 80%, 0. 66%, and 0. 48%, respectively. It is shown that the 2D OFDR shape sensing accuracy can be improved by proposed method.

Abstract:Aiming at the problems of S700K switch machine fault diagnosis, which is difficult to extract effective features and signal processing and classification algorithms, a fault diagnosis method for switch machine combining one-dimensional convolutional neural network (1DCNN) and bidirectional long short-term memory neural network (BiLSTM) is proposed. Firstly, the power curve of the switch machine collected by the microcomputer monitoring system is processed. Secondly, the fault features are extracted adaptively from the processed data by the convolution layer and pool layer of CNN. Then through Flatten, the extracted fault features are taken as the input of BiLSTM layer to further mine the deep-level features. Finally, the Softmax function is used to implement intelligent fault diagnosis. The model is validated by the real data provided by a railway bureau. The results show that the accuracy, recall and F1 value of the proposed model reach 98. 99%, 98. 89% and 98. 89% respectively, which are better than other classical fault diagnosis models, 1DCNN-BiLSTM model improves the accuracy of fault diagnosis by at least 1. 08% when the training speed is fast

Abstract:In order to realize the accurate measurement of drilling depth of directional drilling rig under coal mine, a lightweight drilling rig target detection network integrating attention mechanism and inverse residual structure ( GCI-YOLOv4) is proposed. Through automatic, rapid and accurate detection, the movement track of drilling rig, the number of driven drill rods and the drilling depth are obtained. Aiming at the problem of low color gamut discrimination in coal mine, GhostNet is used as the feature extraction network to remove the redundant features of complex background, lighten the model and accelerate the speed of model reasoning. Aiming at the problem of low target saliency of drilling rig caused by uneven illumination in coal mine, the attention module is introduced to enhance the saliency of drilling rig in complex background. Aiming at the problem that it is difficult to detect accurately when the drilling rig moves at high speed, the inverse residual structure is introduced to extract richer semantic features while maintaining the balance between speed and accuracy. In order to ensure the accuracy and reliability of the model, the proposed detection algorithm is compared with five classical target detection algorithms. The experimental results show that the proposed detection algorithm can better solve the problems of low background gamut discrimination, high-speed movement of drilling rig and uneven illumination under coal mine. The average detection accuracy is 99. 49% and the detection speed is 58. 10 FPS. The performance is better than the classical target detection algorithm. The proposed detection algorithm is deployed in the field of the working face for testing, which can accurately obtain the motion trajectory of the drilling rig. The number of drill pipes is calculated by filtering and counting the rising edge. The counting accuracy of drill pipes is 99. 4%. The drilling depth is accurately calculated, which verifies the feasibility and practicability of this method.

Abstract:Aiming at the problems of transmission time delay and network congestion caused by load imbalance in highly “ central” connected multi-campus networks, a dynamic routing optimization algorithm based on adaptive multi-sampling Dueling deep Q-Network (AMD-DQN) is proposed. Firstly, the idea of Dueling DQN is introduced into the network model, and the structure of the multilayer perceptron is improved by centralized processing to prevent high estimation of value function. Then, the experience playback mechanism adopts an adaptive multisampling mechanism, which combines random, nearest and priority sampling methods, adjusts adaptively according to the load situation, and randomly selects the sampling mode according to the weighted probability. Finally, the AMD-DQN network structure is combined with reinforcement learning signal and random gradient descent to train the neural network, and the maximum value action of each step is selected till the transmission is successful. The experimental results show that compared with the traditional DQN and Dueling DQN algorithms, the average delay of the AMD-DQN algorithm is 128. 046 ms, and the throughput reaches 5. 726 / s, which effectively reduces the transmission delay of packets and improves the throughput. At the same time, the congestion degree is evaluated from five directions, and good experimental results are obtained, which further alleviates the congestion of the network.

Abstract:In the process of microwave heating, the accuracy of temperature will directly affect the chemical reaction rate and the performance of the prepared samples, so the accurate measurement of temperature in the microwave field is of great significance. The traditional temperature measurement methods in microwave field are reviewed, including the limitations of thermocouple, optical fiber and infrared in the application of microwave field. In view of the limitations of traditional methods, a new type of temperature measurement method: metal-organic frameworks (MOFs) thermometry. It not only has a high degree of adjustability, but also nanoscale MOFs particles can achieve temperature measurement at the micron or submicron level, which has attracted extensive attention. Analyzing the research results of dual emission MOFs thermometry and membrane thermometry in recent years, and reviewing the methods of optimizing MOFs thermometers, the future development of microwave temperature measurement technology was discussed.

Abstract:In order to improve the target detection ability of CenterNet Ancor-free target detection network, an improved CenterNet target detection network based on attention feature fusion and multi-scale feature extraction network was proposed. Firstly, in order to improve the expression ability of the network for multi-scale targets, an adaptive multi-scale feature extraction network was designed. The feature map is resampled by cavity convolution to obtain multi-scale feature information, and the fusion was carried out on the spatial dimension. Secondly, in order to better integrate semantic and scale inconsistent features, a feature fusion module based on channel local attention was proposed. the fusion weight between shallow features and deep features was adaptively learned, and the key feature information of different perceptual domains was retained. Finally, the algorithm was verified on VOC 2007 test set. The experimental results showed that the detection accuracy of the final algorithm reaches 80. 94%, which was 3. 82% higher than the baseline algorithm CenterNet, and effectively improves the final performance of the Ancor-free target detection algorithm.

Abstract:The method of using image and video technology to monitor the transmission channel in real time and using intelligent target detection algorithm to realize the identification and early warning of potential damage caused by external force with high accuracy and has been gradually popularized in recent years. However, in the real conditions, due to the scene change, weather change ( such as fog, rain, etc. ) and other factors, training data cannot cover all conditions, and the algorithm model generalization ability is weak, and there are often missed and false positives in the practical application. Based on these problems, we use the YOLOv5 as based algorithm. This article through the data amplification to simulate different weather, with the domain adaptive network to combat training of the training set, strengthen model generalization ability of the different weather, different scene, citing the attention mechanism ( CBAM) at the same time, strengthen model’s ability to extract features from data. Experiments prove that the Recall obtained by the algorithm in this paper reaches 86. 9%, which is significantly improved compared with the original algorithm, and the average accuracy (MAP) is 92. 2% higher than that of the original YOLOv5 algorithm, which can accurately detect the target to be detected and reduce missed and false detection.