Abstract:The rule of time is a way of measurement and calculation time in order to unify time, it includes three elements: appointed unit of time, conventional origin of time and time calendar. The existing rules of time only apply to the earth and inside range of its gravitational potential, and not to the vast space out of the earth. For human beings to go to the space, it is an inevitable requirement to unify the time of space out of the earth. Here describes the definition and connotation of the time rule, and applies principle of general relativity to prove that the synchronicity should be only happened in the same coordinate system, and that does not have simultaneity between different coordinate systems. And then two expressions of time in the general relativity, proper time and coordinate time, were introduced. Coordinate time can be reproduced by pulsar looked like from infinite far away. Two modes to unify time were compared that one mode is “center time-keeping, local time service”, and another mode is “ local proper time, universal coordinate time”. The key point of this paper is to distinguish two kinds of viewpoint, absolute time view and relativity time view. And think that relative time viewpoint is the key to deal with the problem to unify time out of the earth. Three elements of time rules are explained by relativity time viewpoint again, and more, to put forward the concept of space time-keeping system.

Abstract:A UAV group can be introduced as a relay node in the base station that cannot provide a communication support area to build a UAV emergency communication network. To solve the problem of how to efficiently select the optimal relay node and maximize the system throughput, we propose a relay selection strategy for UAVs based on the SA-SARSA reinforcement learning algorithm. After all the relay nodes are forwarded after decoding retransmission (DF), the expression of the maximum average throughput of the user is obtained. The relay node with the maximum return value is selected by setting the SARSA algorithm’ s state, action, and reward function. At the same time, the annealing algorithm is introduced to make the source node explore more relay nodes so that the performance of the UAV swarm communication network can reach the optimal state. The simulation results show that compared with the previous SARSA relay selection strategy, the proposed SA-SARSA relay selection strategy increases the proportion of the ideal algorithm by 10%. At the same time, under the same total power condition, the throughput of relay nodes selected by the proposed strategy is 8% and 13% higher than that of the Q-learning relay selection strategy and SARSA relay selection strategy, respectively.

Abstract:Consumer-level UAVs have small scale, low fly speed and height, existing deep learning methods hardly achieve high detection accuracy and good robustness on detecting UAVs. In order to address this problem, this paper develops an improved YOLOv4 algorithm with feature enhanced module named as FEM-YOLOv4 for UAVs detection. Firstly, according to the characteristics of UAVs, this paper reduces the subsampling multiple of CSPDarkNet to improve the backbone network and make full use of shallow features containing detailed information. Secondly, this paper introduces the feature enhancement module to replace the SPP module. The feature enhancement module includes multiple branches and dilated convolution, and it obtains different levels of semantic information, which is beneficial to enhance the detailed semantic features and the detection capabilities of the network. Thirdly, delete the PAN module to improve the feature pyramid, and compress the depth of each detection layer to highlight the detailed and semantic information of the feature maps. Finally, the anchor box is initialized by the K-means++ algorithm to make the model more suitable for predicting the UAV targets. Compared with the six target detection algorithms, the experimental results show that the mAP and Recall of FEM-YOLOv4 algorithm reach 89. 48% and 97. 4% respectively, which are superior to other algorithms, and the average detection speed is 0. 042 s.

Abstract:A model predictive control (MPC) method based on extended state Kalman filter (ESKF) is proposed for the path tracking control problem of a quadrotor UAV that is susceptible to wind disturbance and measurement noise during flight. First, the Newton-Euler method is used to establish a four-rotor UAV dynamic model under the influence of wind disturbance; then, an MPC method based on the error model is used for position control, and an ESKF is proposed to estimate wind field disturbance to compensate the controller. The attitude dynamic model is linearized by the feedback linearization method, and an attitude controller based on ESKF-MPC is designed. Finally, the simulation results show that when the measurement noise variance is 0. 000 1, the position tracking mean square error of this method is 0. 013 meters which is smaller than that of the active disturbance rejection control method. When the variance is greater than 0. 000 1, the active disturbance rejection control method makes the system unstable, and the method in this paper can achieve better position tracking.

Abstract:The “bird strike” and the “black flight” disturbance incident of the rotary-wing UAVs have become the “two hidden dangers” threatening the flight safety of civil aviation. The different countermeasures against birds and rotary-wing UAVs will be taken in the airports. The identification of birds and rotary-wing UAVs is of great significance for improving the monitoring performance of noncooperative targets and ensuring flight safety. Aiming at the problem that the discrimination performance of the rotor-wing UAV with strong maneuverability for the discrimination method based on the motion feature extraction is degraded, considering that the timefrequency spectrum of the bird target is more complex relative to the rotary-wing UAV. Firstly, the micro-motion features of the spectrum energy entropy corresponding to the target echo spectrum and the peak symmetry pair are constructed, Secondly, K-means is used to fuse the extracted motion features and micro-motion features and the identification of birds and rotor-wing UAV targets can be realized. The experimental results verify the effectiveness of the proposed method.

Abstract:A wind turbine blade infrared image stitch method based on UAV speed information is proposed for the difficulties in wind turbine blade infrared image stitch. Firstly, the blade mask image is predicted by U-net network to remove the redundant background information. Secondly, the parameters of translation, rotation and scaling are calculated to register the stitched image. Finally, the multiband blend algorithm is used to fuse the stitched images to eliminate the stitching caused by the change of field of view and illumination. The experimental results show that the RMSE of the proposed method in the x gradient direction of the splicing area is less than that of traditional image stitching methods, the stitching success rate is 98. 26%, and the infrared panorama of wind turbine blade is successfully obtained. The multiband blend algorithm is applied to wind turbine blade infrared image fusion, and the results show that the RMSE at the image mosaic is significantly reduced and the transition is smoother.

Abstract:In order to solve the problem of low accuracy caused by false and missed detection in the corner detection methods of current visual measurement of thread key parameters, we proposed a new method based on contour corner detection. Firstly, bilateral filtering and iterative threshold method are used to improve the Canny operator to increase the accuracy of edge detection. Secondly, the doublethreshold DP algorithm and Hough transform were used for piecewise fitting; the edges are protected and Contour smoothness is kept. Furthermore, the corner points of the tip and bottom of threads were extracted with CTAR algorithm. Finally, thread major diameter and minor diameter is measured according to the position information of corner points. Experimental results show that, compared with the traditional Canny operator and Harris corner detection method, our new method can accurately and reliably detect the thread edges and corners. Therefore, the high precision measurement of screw thread parameters is realized. The average measurement accuracy of major and minor diameters of screw thread is 0. 003 3 mm and 0. 002 6 mm, respectively.

Abstract:For the characteristics of non-stationarity, spatial correlation and temporal dependence of short-term traffic flow time series, this paper proposes a combined prediction model of traffic flow based on improved variational mode decomposition ( VMD), graph attention (GAT) network and gated recurrent unit (GRU) network to improve its prediction accuracy and convergence speed. First, the variable mode decomposition algorithm improved by mutual information entropy (MI) is used to decompose the traffic flow time series into a series of amplitude modulation and frequency modulation signal sub-sequences, which reduces the non-stationarity of the time series signal and improves the prediction accuracy of the model. Then, they are sent to the graph attention network to capture the traffic flow of adjacent nodes of the road network to different degrees on the traffic flow of the central prediction node, so as to realize the spatial correlation modeling and further improve the prediction accuracy of the combined model. Next, the traffic flow component sub-sequences are sent to the gated recurrent unit network separately to capture the temporal dependence of the traffic flow sequence, and use the improved RMSPRop optimization algorithm to iteratively search for optimization, which not only improves the convergence speed of the optimization algorithm, but also improves the prediction accuracy of the model. Finally, the prediction values of each component subsequences are combined as the final output of the prediction model. The experiment used traffic data from the RTMC system, the results show that compared with LSTM, GCN and GAT baseline models, the mean absolute error (MAE) is reduced by 9. 35, 4. 12 and 4. 09, respectively, and the mean absolute percentage error ( MAPE) is reduced by 16. 42%, 7. 32%, and 8. 1%, respectively. The convergence speed of the optimization algorithm and the prediction accuracy of the combined model are effectively improved.

Abstract:In order to solve the problem that it is difficult to accurately estimate the multipath time delay of linear frequency modulation (LFM) signals in impulse noise environment, this paper designs a new nonlinear amplitude transformation function P-NAT (piecewiseNAT) function, and proves that any random variable has finite second-order moment after transformation by this function, a multipath time delay estimation method of LFM signals based on P-NAT function is then proposed. The optimal order FRFT is performed on the transmitted signal and the noisy multipath received signal transformed by P-NAT function respectively. According to the relationship between peak position offset in FRFT domain and the time delay, the multipath time delay estimation of the LFM signal is realized. Simulation results show that this method is superior to fractional lower order statistics method and myriad filter method in impulse noise suppression, when the generalized signal-to-noise ratio is 0 dB, the normalized root mean square error of time delay estimation is less than 10 -4 . It is suitable for multipath time delay estimation of LFM signals in low signal-to-noise ratio and strong impulse noise environment.

Abstract:Using the multi-beam optical stress sensor (MOSS) to find the curvature of the film requires measuring the distance between the beams before and after reflection, and the radius of curvature of the film can be calculated based on the distance between adjacent incident (refracted) beams. The spacing of the beams can be determined by the center point of the received light spot. Extracting the center point of the light spot mainly uses Canny edge detection algorithm and ellipse fitting algorithm. The spot array image obtained with the CMOS sensor has a lot of salt and pepper noise mixed around the spot. For the traditional Canny algorithm, the edge of the spot cannot be accurately detected. An improved Canny edge detection algorithm is used, which can effectively eliminate the noise around the light spot and retain the edge of the light spot. Finally, an ellipse fitting algorithm is used to fit the edges to obtain all the center points of the spot array. At the end of the experiment, standard parts were selected for verification, and the radius of curvature calculated by the measured distance was compared with the actual one. The error result was 2% ~ 4%, which proved that the method was feasible.

Abstract:Multi-sensor integrated navigation system is a typical nonlinear system, a federated UKF algorithm is proposed to improve its filtering accuracy in this paper. Firstly, the standard UKF is simplified on the basis of establishing nonlinear state equation and linear measurement equation of multi-sensor integrated navigation system. Then, based on this simplified UKF, the federated UKF algorithm of multi-sensor integrated navigation system is proposed, the attitude fusion algorithm is designed, and the fault detection function is designed simply in order to verify the fault-tolerant performance of the algorithm. Finally, the GNSS / CNS / SINS multi-sensor integrated navigation system is taken as an example for simulation verification. The simulation results show that the federated UKF algorithm can improve the position and attitude accuracy by 25. 8% and 22. 2% when compared with the federated linear Kalman filter, and inherit the fault-tolerant performance of the federated linear Kalman filter.

Abstract:As a parameter to characterize the creep state of metal materials, it usually requires the emission ultrasonic probe and the receiving probe to be in coaxial position in penetration detecting. It is difficult to align the position of the emission probe and the receiving probe accurately when the penetration method is used for detecting and acquisition. A set of reflective nonlinear ultrasonic experiment system is designed to solve the problem. The detection information is extracted by analyzing the reflection signal of a bottom wave. The creep properties of P91 specimen with weld structure and creep time of 0 h, 120 h and 250 h were detected by the experiment system. The results show that the ultrasonic second harmonic nonlinear parameters after the 120 h and 250 h creep time of the P91 steel test block, the relative change of the base metal area is 2. 9%, 17. 4%, the weld zone is 2%, 23. 6%, the heat zone is 5. 6%, 34%.

Abstract:Aiming at improving the deficiency of positioning accuracy of moving targets such as underground personnel, vehicles and equipment, this paper studies the location algorithm and fingerprint location model of mine moving target and a fusion location method based on SVR model optimized by improved particle swarm optimization and Chan distance fingerprint is proposed. Firstly, an ultra wideband (UWB) core node model based on STM32 ARM main controller and DWM1000 is designed, and the transmission distance data are analyzed through bilateral bidirectional ranging and time of flight ( TOF). On this basis, the moving path of the target is predicted by successively collecting distance fingerprints at specific points and the moving target route fitting within the improved PSOSVR model. Then it is combined with the Chan algorithm fingerprint, and expand the optimized distance fingerprint fusion location method. The experimental results show that the optimized distance fingerprint fusion location method can correctly predict the moving path, with the maximum error of no more than 20 cm and the average error of no more than 1 cm. The study is of great significance to mine intelligent construction and safety production.

Abstract:In the process of real-time monitoring of fatigue crack growth by adopting direct current potential drop method (DCPD), multiple noises interference make the potential drop signal of fatigue crack propagation inaccurate. To improve the accuracy and stability of signal, the optimal smoothing noise reduction model based on variational mode decomposition (VMD) is established. The sample entropy, correlation coefficient, and mean square error of each intrinsic mode function (IMF) are calculated to eliminate the interference source in the original signal, and the effective components are selected to reconstruct new potential drop signal. Comparing different signal reconstruction schemes to select the optimal reconstructed signal. Finally, different smoothing noise reduction models are established for the optimal reconstructed signal, and the optimal smoothing noise reduction model is obtained by comparing smoothness, mean square error, signal-to-noise ratio and other indicators. The analysis results show that the smooth noise reduction model has excellent noise reduction effect, the noise reduction error ratio of the optimal smooth model is 0. 122 050, and improves the smoothness and accuracy of the monitoring signal.

Abstract:Manufacturing of iron ore green pellet is a significant step in metallurgy industry. Crack detection for green pellet is a key step in the measuring process of the important pellet quality metric (drop strength). However, current image-based methods are mainly used to detect cracks on the flat surface of bridges, roads and solar cells. Therefore, their crack detection ability is limited on pellet with curved surface, and is easily affected by the raw material stains, pellet edge contour, strong light reflection or other interferences in pellet images, resulting in the false detection or missed detention. To solve the problem, a crack detection method for green pellet based on steerable evidence filter (SEF) is proposed. Firstly, the target area of green pellet is segmented by the active contour model, which is used to eliminate the raw material stains in the image background. In order to overcome the interfaces of pellet edge contour and strong light reflection, steerable evidence filter is used to generate the response map of pellet crack, followed by the morphological processing and connected domain analysis method used to eliminate the pellet edge response and noise in response map of pellet crack, so that more accurate crack segmentation results can be obtained. Finally, the connectivity domain method is used to detect cracks and calculate the number of cracks. In order to verify the proposed method, experimental platform was built to establish dataset of green pellet, and about 300 green pellet images with different backgrounds and the number of cracks were captured. Results show that our method outperforms five crack detection methods in crack segmentation metrics including accuracy, precious, F1. Accuracy of detecting cracks in pellet is 96%, and the accuracy of detecting the number of cracks is 90%. The crack detection results lay a foundation for the automatic and intelligent detection of drop strength quality metric of green pellet.

Abstract:Degradation prediction is an important technical approach for equipment health management. In recent years, a large number of time series prediction methods have been applied in degradation prediction. However, due to the complex structure and diverse functions of many large equipment, there are obvious stages in the degradation process, and the application of a single model to predict the degradation at different stages will significantly reduce the accuracy, and the retraining of the model for different stages will also bring the loss of time and computing power. To solve the problem of multi-stage degradation, this paper introduced the idea of transfer learning and proposed a multi-stage degradation prediction method combining degradation pattern recognition and LSTM-fine-tune. The LSTM model was trained with degradation data, and then part of network parameters was frozen. After identifying the new degradation stage of equipment, the model is fine-tuned with the degraded data of the new stage to quickly match the data of different stages. In order to verify the validity of the model, this paper takes oxygen concentrator as an example to apply the model. The results show that the proposed method can effectively identify the degradation of oxygen concentrator at three stages, and the mean square error of prediction for each stage is 0. 507, 8. 976 and 0. 375 respectively, which is far lower than the mean square error of direct prediction without segmentation of 76. 87. In terms of training time, compared with the retraining time of each stage, the training accuracy is obviously superior to the traditional methods such as Wiener process and Lstar.

Abstract:To alleviate the shortage of health care workers under the novel coronavirus pneumonia (COVID-19) and to achieve intelligent monitoring of inpatients, this paper proposes a new behavior recognition method based on enhanced micro-Doppler spectrograms in the space-time domain using frequency modulated continuous wave (FMCW) radar. Firstly, constructing a micro-Doppler spectrum of the human behavior acquired by the radar. Then, a new time-space domain enhancement algorithm combining histogram equalization and homomorphic filtering is used for the enhancement of spectrogram information. Finally, an improved convolutional long short term memory network (ConvLSTM) is proposed to extract the time and space features of the spectrum, which effectively identifies seven common inpatient behaviors, such as drinking and falling. The experimental results show that the method in this paper can effectively monitor the patient's behavior, and the recognition accuracy of the seven actions can reach 94%.

Abstract:When the dual-axis inclinometer is used in combination with other position and orientation measurement sensors, since the true direction of its own coordinate axis cannot be accurately mapped by the mechanical enclosure, it is difficult to achieve accurate calibration of the installation position error, which leads to the problem of measurement error. The object of the study is the dual-camera target of the linear pipe jacking machine guidance. The theoretical value of the attitude angle is obtained by three points pose transformation algorithm, and then combined with the angle measurement value of the dual-axis inclinometer to establish an error compensation matrix, which is used to correct the pose parameters. Through simulation analysis and actual measurement experiments, the proposed method can realize the accurate compensation of the measurement angle information of the dual-axis inclinometer, and the absolute measurement accuracy of the attitude angle is better than 0. 02°, which meets the accuracy requirements of the linear pipe jacking guidance. It can be widely used in other combined measurement systems based on dual-axis inclinometer.

Abstract:In order to suppress the serious influences of environmental temperature and humidity on the measurement accuracy of the nondispersive infrared ( NDIR) CO2 gas sensor, the low humidity and constant temperature control modules were designed from the perspective of hardware compensation. The gas-drying tube constructed by the polyvinylidene fluoride ( PVDF) hydrophobic filter membrane and 3A molecular sieve is used to reduce the humidity of gas chamber, and the proportion integration differentiation (PID) algorithm is applied to adjust the power of the heating plate wrapped the outside of optical gas chamber to achieve a constant temperature effect. The influence of humidity and temperature on the accuracy of gas concentration measurement is studied. At the same time, the anti-interference design of the low-humidity and constant-temperature is tested and verified. Experimental results show that the low humidity control module based on the gas-drying tube can reduce the humidity of CO2 gas to about 8%, and the constant temperature control module can stabilize the temperature of the gas chamber at 40 ℃ . In a complex temperature and humidity environment, the average relative error of the concentration measurement of CO2 gas sensor including the anti-temperature and humidity interference design is 8. 38% within the gas concentration range of 0 ~ 2 000×10 -6 , which significantly reduces the temperature and humidity drift of the detection system. The research results have certain reference value for the development of high-performance NDIR gas sensors.

Abstract:Infrared focal plane array (IRFPA) non-uniformity correction (NUC) problem of staring infrared imaging system is the key issue that affects the imaging quality. To assess NUC performance based on self-adaption scenes, a joint test method was developed successfully, and was analyzed using three typical scenes as a test scenario. A joint test system was built further. The convergence time was respectively 16, 18 and 15 s at the frame frequency of 50 Hz; the correction accuracy was 0. 28%, respectively lower than 20 s and 0. 6% by the traditional algorithm, and the effective temperature was in the range of 19 ℃ and 31 ℃ with the input of sea scene; the corrected noise-equivalent temperature difference ( NETD) was 87 mK. The experimental results show that the proposed joint test method, compared with the traditional self-adaption correction algorithm, can evaluate the NUC performance of IRFPA availably, and the NUC algorithm of IRFPA can improve the imaging quality of infrared imaging system.

Abstract:The paste attenuation mentioned in this paper is a conductive coating flaw on flexible printed circuit board. For avoiding harmful effect caused by paste attenuation on FPC trace, which would cause economical losses, it is a reasonable way to locate the position of paste attenuation for the later repair or get rid of the unqualified board directly. However, the size of paste attenuation flaw is uncertain. And the non-uniform illumination results in the non-uniform of image grayscale. Also trace region crisscrosses with the background. The above reasons increase the difficulty of trace paste attenuation detection. Aiming at this problem, a detection scheme is proposed which made up of adaptive directional template and gray difference, based on the detailed analysis of paste attenuation texture and trace morphological feature. First of all, the trace region after preprocessing is traversed with ADT template according to skeleton tracking process, then the large paste attenuation regions, diminutive paste attenuation regions and the whole candidate pixels relative to the ADT template size, thus, the preprocessed image is segmented into several diminutive regions. Secondly, the noise regions which are not flaws are eliminated utilizing the numerical relationship of diminutive regions energy tendency value and default threshold. At last, the non-deficit regions gained after the above two steps processing are screened again via method of gray difference, then the middle region of large paste attenuation texture is extracted, thus, the whole paste attenuation texture detection is completed. Results show that the paste attenuation flaw detection EER of suggested method is 3. 93 percent on the testing of self-built image storage SUT-F2, which decreases at least 5. 28 percent compared to the other typical methods of textural features extraction and paste attenuation flaw detection. The above facts prove the high efficiency and actual value of suggested method.

Abstract:Aiming at the problem that it is difficult to apply the fault diagnosis model based on deep neural network in the industrial environment with limited resources, a bearing fault diagnosis method based on compressed deep neural network was proposed. Firstly, the filter corresponding to the output low-rank feature graph in the convolution layer is removed by structural pruning. Then unstructured pruning was used to remove non-important connections in the whole connection layer. Finally, the number of bits required for parameter representation is reduced by quantizing the parameters of the weight matrix, and the storage of parameters is further reduced by using the compression storage method of the weight matrix. Experimental results show that the proposed compression method can greatly reduce the parameter storage and floating point computation of the network, shorten the training time of the network and speed up the response of the network on the premise of high diagnostic accuracy, which provides a beneficial exploration for the industrial application of the deep neural network method.

Abstract:In GPS / BDS combined system, the regression period of MEO orbit satellite of BDS system is different from that of BDS (GEO, IGSO) / GPS orbit satellite, which shows that the corresponding sidereal daily filter model is not unique, and the multipath error suppression effect is also different. In order to achieve better deformation monitoring effect, CEEMDAN-WT joint filtering was used to eliminate the noise effect of the baseline coordinate sequence, and then a sidereal daily filtering model suitable for GPS / BDS combination was constructed. Based on the multipath information extracted from the model, the multipath error of the baseline sequence of the next day was corrected. The measured results show that the horizontal positioning accuracy of baseline is less than 5 mm, the elevation positioning accuracy is less than 1. 33 cm, and the overall positioning accuracy is improved over 40% after the multipath error is suppressed by the sidereal daily filtering model combined with BDS (GEO, IGSO) and GPS.

Abstract:According to the requirements of water entry pressure test of large amphibious aircraft independently developed in China, the pressure field and test environment of the ship bottom structure are analyzed firstly. The ship bottom pressure is numerically simulated by ALE algorithm, the pressure distribution at the time of landing the water at the bottom of the ship is obtained. Combined with the theoretical calculation based on the improved Wanger model, the optimal measuring point of the water inlet pressure is determined, and the signal sensing scheme based on the nonlinear vibration model is designed. Secondly, analyze the practical engineering problems of aircraft modification, and determine the sensor refitting process. Finally, through the flight test, the feasibility of the scheme and the reliability of the simulation model are verified. The maximum deviation between the test results and the simulation results is 3. 4%, and provide a basis for the optimization and improvement of the test scheme in the next stage.

Abstract:Sparse decomposition method usually shows poor performance in terms of matching with the fault signal due to the personalized vibration behavior of the bearing, and has some drawbacks especially in practical applications due to the improper setting and selection of dictionary parameters. To address these issues, a novel personalized sparse diagnosis method based on dynamical wavelet dictionary was presented. It lies in the foundation for the idea of finite element model (FEM) technology and sparse decomposition. In order to obtain dictionary atoms, according to the different operating conditions, the FEM is built to generate the vibration signals which accord with the bearings features of faults, and the fault transient shock extracting from vibration signal will be regarded as dictionary atom. The dynamical wavelet analysis dictionary can be constructed via atomic Toplitz transformation. The bearing fault feature frequencies can be extracted by performing sparse decomposition and reconstruction of the signal with the help of orthogonal matching pursuit (OMP). The FEM simulation signal and experiment signal results show that the presented scheme can extract the fault features more effectively than the popular parametric dictionary based on a correlation filtering algorithm ( CFA), fast-kurtogram and the K-SVD self-learning dictionary and has a stability and scalability.

Abstract:Traditional junction temperature estimation methods cannot be adjusted according to the health status of IGBT module in real time, which leads to inaccurate junction temperature estimation when the module is degraded. Therefore, to solve the problem of junction temperature estimation error caused by module package degradation in actual conditions, this paper established a multi-data-driven IGBT junction temperature online estimation model with artificial neural network as main body. Firstly, the saturation voltage drop was determined as a thermoelectric parameter and its composition was studied. The coupling relationship between the saturation voltage drop, collector current, chip junction temperature and package degradation are analyzed. Then, to solve the problem of temperature characteristic change of saturation voltage drop caused by package degradation, a junction temperature estimation model was constructed by combining the advantages of Miller voltage temperature characteristic and the artificial neural network algorithm driven by saturation voltage drop and collector current. And the data were extracted by building an experimental platform to complete the training of the model. Finally, by comparing the estimation error with the traditional junction temperature estimation method, the new model reduces the estimation error from 20% to about 5%.

Abstract:To address the problems that the gradient descent method is easy to converge to the local optimum and the convergence speed is slow under large sample data sets, a dynamic attenuation network and a dynamic attenuation gradient descent algorithm are proposed by changing the network structure and gradient descent process in the paper. On the basis of the existing network, an attenuation weight is added between each neuron of each two layers, while an attenuation weight term is introduced in the gradient descent process. The attenuation weight value decreases continuously with iteration, and eventually converges to 0. Due to the addition of the attenuation weight term, the gradient descent speed and convergence speed can be accelerated in the early stage of gradient descent. At the same time, it can avoid crossing over the optimal solution and oscillating around the optimal solution. At the last, it can also improve the probability of the network to obtain the optimal solution. The experimental results on MNIST, CIFAR-10 and CIFAR-100 datasets show that the proposed dynamic attenuation network and dynamic attenuation gradient descent algorithm, compared with the original network that used Adam optimizer and stochastic gradient descent with momentum, improve the test accuracy by 0. 2% ~ 1. 89% and 0. 75% ~ 2. 34%, respectively, while having a faster convergence speed.

Abstract:Magnetic field measurement plays an extremely important role in material science, electronic engineering, power system and even industrial fields. In particular, magnetic field measurement provides a safe and reliable tool for industrial non-destructive testing. The sensitivity of magnetic field measurement determines the highest level of detection. The diamond nitrogen-vacancy (NV) color center is a new type of quantum sensor developed in recent years. The external magnetic field will cause Zeeman splitting of the ground state energy level of the diamond NV color center. Optical detection magnetic resonance (ODMR), using a microwave source and a lock-in amplifier to detect the resonant frequency of the NV color center, and finally the change of the resonant frequency can accurately calculate the size of the external magnetic field and the sensitivity of the external magnetic field change. In the experiment, a diamond containing a high concentration of NV color centers is coupled with an optical fiber to realize the preparation of a magnetic field scanning probe. Then, the surface cracks of the magnetized iron plate weld are scanned, and the scanning results are drawn into a twodimensional magnetic force distribution map, according to the magnetic field gradient change of the magnetic force distribution map, the position and size of the crack can be judged accurately, which provides a very effective diagnostic tool for industrial safety.