Abstract:The 3D target mirror is a vital part of micro / nano coordinate measuring machine system. The orthogonality between its any two mirror planes is very important to ensure the measurement accuracy of the measuring machine system. In order to verify whether the 3D target mirror meets the accuracy requirement of the measuring machine, the orthogonal deviation angles of the 3D target mirror are measured and the measurement uncertainties are evaluated. On the basis of quantitative characteristic analysis and modeling, the traditional method (called as GUM method) given in the guide to the expression of uncertainty in measurement, Monte Carlo method (MCM) and adaptive Monte Carlo method (AMCM) are researched. The comparison of the measurement evaluation results shows that the 3D target mirror approximately meets the accuracy requirement of the measuring machine. The orthogonal deviation angles of Z-X mirror planes and Y-Z mirror planes are about 0. 5″, and the orthogonal deviation angle of X-Y mirror planes is about 3. 3″. The evaluation results obtained by the three methods are basically consistent. Moreover, MCM and AMCM are more reasonable than GUM, and AMCM method is more efficient than MCM method. This work provides reference methods for the task-oriented measurement uncertainty evaluation of micro / nano CMM.

Abstract:A method for evaluating the uncertainty of amplitude-frequency characteristics of pressure sensors is proposed. Firstly, the probability density distribution of the pressure sensor model parameters was calculated based on the kernel density estimation method, and the pseudo-random number conforming to the probability density distribution was generated by the acceptance-rejection method. Then, an adaptive Monte Carlo iteration convergence threshold optimization method is proposed to accurately estimate the optimal iteration number. Finally, based on the optimal number of iterations, the adaptive Monte Carlo method is used to evaluate the uncertainty of the amplitude-frequency characteristics of the pressure sensor, and the optimal estimate value, the standard uncertainty and the uncertainty interval under the given confidence probability are obtained. The performance of the proposed method is verified by the uncertainty simulation of the amplitude-frequency characteristics of the pressure sensor. The results show that the mean and maximum absolute errors of the uncertainty evaluation results of the pressure sensor amplitude-frequency characteristics obtained by the proposed method are 8. 837×10 -5 and 5. 103×10 -3 , respectively, reduced by about 55% and 76% compared with Monte Carlo method (100 000 tests), and reduced by about 67% and 79% compared with adaptive Monte Carlo method, respectively, indicating that the proposed method can effectively evaluate the uncertainty of the amplitude-frequency characteristics of pressure sensors.

Abstract:Length is one of the seven basic physical quantities, which is widely used in people′s work and life. In this paper, a two-arm laser ranging system composed of two laser ranging sensors is proposed. It can measure the distance between the instrument and any point in space and between any two points in space in a non-contact way. It also has the function of measuring internal angle, external angle and two-dimensional perpendicularity. The measurement error of the system is analyzed theoretically and simulated by MATLAB software. The actual experimental test shows that the absolute error can be controlled within 1 mm when the measurement distance is less than 1 m, and the relative error can be controlled within the order of 1‰ when the measurement distance is less than 50 m. And the repeatability measurement experiment and measurement uncertainty evaluation verify the error index of the system, and point out the direction for the system improvement. The instrument also has the advantages of low power consumption, low cost and easy portability, so it has broad application prospects.

Abstract:Due to the randomness of the initial phase of the vector network analyzer, there will be a large error when the polarization characteristic parameter H/ α — is measured by using it. In this paper, a method of measuring and calibrating the full polarization scattering matrix by using vector network analyzer is proposed to obtain the polarization characteristic parameter H/ α — . Firstly, the measurement system of the full polarization scattering matrix is constructed, then the polarization scattering matrix is calibrated by the disk, 0° and 22. 5°dihedrals. Finally, the error caused by the randomness of the initial phase of the vector network analyzer is eliminated, and the accurate measurement of the polarization characteristic parameter H/ α — is realized. The results show that the error of the scattering angle α — obtained by polarization calibration is less than 3 degrees compared with the theoretical value, which indicates that the polarization characteristic parameter H/ α — of the target can be accurately obtained by measuring and calibrating the amplitude and phase of the polarization scattering matrix.

Abstract:When microwave RF chips are measured in the time domain, the broadband probe of digital oscilloscope is a key subsystem used to test the chip, and the transmission performance of the oscilloscope probe system must be clear for accurate measurement of the chip. So, the calibration system containing a grounded coplanar waveguide is designed to calibrate a broadband probe of digital oscilloscope. According to the signal transmission characteristics of the calibration system, the frequency response of the digital oscilloscope broadband probe system is obtained by deconvolution, and its bandwidth is calculated, which realizes the calibration of the probe system. The Keysight Infiniimax high-frequency differential active probe system is calibrated by this method, and the bandwidth value of about 12 GHz is in accordance with the nominal bandwidth of 12 GHz, indicating that the calibration system is reasonable and feasible.

Abstract:In this paper, a dual-stage curved surface scanning measuring device is developed. The purpose is to solve the contradiction between the high precision and the large range of the sensors, which makes it possible to measure the curved surface with a large height difference. A dual-stage measurement system is used to expand the range of the high-precision sensor so that a scanning measurement for curved surfaces with a large range and high-precision is realized. The dual-stage measurement system uses a voice coil motor to drive the flexure hinge to move the high-precision sensor in a straight line, so the measured surface can be kept within the measurement range of the high-precision sensor. The measurement data of the high-precision sensor is taken as the first-stage measurement data, and the measurement data of displacement of the flexure hinge by the high-precision grating encoder is taken as the second-stage measurement data. The information of the measurement point can be obtained by processing the data of the dual-stage measurement system, and the range of 5 mm can finally be realized. At the same time, the high-precision grating encoder is also used as the feedback device of the servo system. Then, the performance of the dual-stage measurement system is evaluated, and the measurement uncertainty is analyzed and calculated. Finally, experiments are carried out on the large range scanning capability of the measuring device, which proves that the device has the ability of large range scanning measurement.

Abstract:Monochromatic X-rays calibration facility is a device that generates monochromatic X-rays based on Bragg diffraction. The continuous adjustable monochromatic X-rays can provide detailed calibration experiments with optional energy for a detector. Monte Carlo calculation software is used to simulate the detection efficiency of two different types of silicon drift detectors, and the detection efficiency of the 3 ~ 50 keV energy band is obtained. The calibration experiment of the detection efficiency of the SDDs is completed on the monochromatic X-rays calibration facility, and the detection efficiency measured by the experiment is obtained and compared with the simulation efficiency. Moreover, the maximum error between the experimental results and the theoretical calculation results of MLC SDD is 4. 23%@ 15keV, and the maximum error between the experimental results and the theoretical calculation results of BEV 133 SDD is -6. 88%@ 12keV. The results show that the experimental calibration of the two detectors are consistent with the theoretical calculation results in the energy band of 7~ 16 keV, and the real detection efficiency is greater than 90% at 8 keV. The research results verify the superiority of using Bragg diffraction-based monochromatic X-rays for detector calibration, and also provide a reliable performance research and testing platform for domestic X-ray detectors.

Abstract:Aiming at the problem of low accuracy and overfitting in the unbalanced data of molecular pump of EAST all-superconducting tokamak device, a method of time-frequency analysis and improved LightGBM algorithm is proposed. Firstly, the normal and fault vibration data are collected by the molecular pump experimental platform. Then, extract the time and frequency domain features. Moreover, the cost-sensitive LightGBM fault detection framework was established by optimizing the misclassification cost function. Finally, the obtained features are used as the input of the cost-sensitive LightGBM algorithm for molecular pump fault detection. The experimental results show that the fault detection accuracy is 99. 4%. Meanwhile, the proposed method can consistently outperform traditional classifiers and LightGBM algorithms. This method can effectively solve the problem of overfitting and realize the detection of molecular pump fault with high accuracy.

Abstract:Aiming at the problem of insufficient attack samples and traffic characteristics in network intrusion detection, a CNN-BiLSTM network intrusion detection method based on self-supervised feature enhancement was proposed to detect abnormal network traffic in traffic data. By analyzing the difference in the distribution of traffic characteristic, IQR outlier processing method was used for data preprocessing, and autoencoder was used to enhance the number of attack samples. A semi-self-supervised model composed of CNNBilSTM neural network and autoencoder was constructed to extract high-dimensional traffic characteristics and self-supervised features respectively. The combined features are input into the classification model as the final features for prediction and classification, so as to realize the function of network intrusion detection. The experimental results show that compared with other intrusion detection methods, the accuracy and F1 score of the proposed method are 85. 7% and 85. 1% respectively, which can effectively improve the detection accuracy of network intrusion and the detection ability of unknown attacks.

Abstract:The timing control system is an indispensable part of cold atom interference experiment, gravimeter development and application test. It is responsible for coordinating, driving and controlling the operation of each subsystem, and is the central unit of cold atom interference control. Aiming at the timing control requirements of high-precision regulation, multi-channel synchronization and flexible regulation in the process of cold atom interference, a set of programmable timing control system for cold atom interference is designed and implemented based on ARM+FPGA architecture, which is composed of timing master module and multiple slave device modules, and has the functions of synchronous trigger, analog modulation, RF drive, signal acquisition and data processing. The results of system test and cold atom interference experiment show that the timing adjustment accuracy of the system can reach 10 ns, and the control accuracy and multi-channel synchronization accuracy are better than 2 ns, which meets the application requirements of cold atom interference timing control and is successfully used in cold atom interference gravity measurement.

Abstract:As the important part of traction power supply system, pantograph and catenary are related to the safety and stability of highspeed train. It is of great significance to identify pantograph arc as soon as possible. By calculating the " Z" friction rate which is more in line with the actual train operation, the running speed, contact pressure and contact current during train operation are adjusted in single variable to simulate the pantograph arc experiment under four different working conditions. Based on the experimental data, the features of pantograph catenary current are compared and analyzed by D-score at first, and the arc identification features and their significant identification intervals are selected. At the same time, a method for finding the suitable number of samples containing sufficient feature information is designed. Finally, seagull optimization algorithm is used to optimize support vector machine to model and identify pantograph arc. The test results and comparative analysis show that SOA-SVM can quickly and effectively model and identify pantograph catenary arc with an average recognition level of 98. 5% and an overall recognition level of more than 97%.

Abstract:The existing defect detection methods for PCB bare boards have problems such as low accuracy, poor real-time performance, and difficulty in deploying on mobile terminals. Based on the YOLOv4 algorithm as the basic framework to improve it, this paper proposes a defect detection algorithm specifically for PCB bare boards. In response to the problem that the YOLOv4 algorithm is difficult to deploy on the mobile terminal, the proposed improved algorithm uses GhostNet instead of CSPDarknet53 to lighten the entire detection network. In order to make up for the lack of performance of YOLOv4 algorithm in multi-scale feature fusion, this paper proposes a bidirectional adaptive feature fusion network AF-BiFPN to replace the PANet network in YOLOv4 algorithm. In order to further improve the detection accuracy of the model, the m-ECANet channel attention mechanism is inserted in the sampling process of the AF-BiFPN feature fusion network. The experimental results show that the model size of the improved YOLOv4 algorithm is 18. 64 MB, the mean average precision (mAP) of detection is 98. 39%, and the detection speed is 62. 23 FPS, which can provide theoretical guidance for actual PCB bare board detection.

Abstract:To address the problem that low accuracy of rolling bearing remain useful life ( RUL) prediction caused by the limited information of single source domain and the insufficient granularity of domain, a new method of RUL for rolling bearing based on multisource subdomain adaption network is proposed. Firstly, fast fourier transform is applied to the collected raw vibration signals to obtain the frequency-domain signals and it takes the frequency-domain signals as the input of the model. Secondly, to reduce the distribution difference between multiple source domains and target domains, all domains are mapped to a common feature space by one-dimensional convolution, and the local maximum mean discrepancy is used to align the degradation stage of each source domain and target domain in an independent feature space. Finally, the RUL of rolling bearing is obtained by comprehensive output of the module in different domains. The results on PHM2012 data set show that the prediction accuracy of proposed method is higher than the comparison method, and can effectively predict the RUL of rolling bearing.

Abstract:Effective prediction of the degree of ash in the heated area of the boiler can provide an important basis for boiler production efficiency and fault early warning. The cleaning factor is used to evaluate the ash deposition status of the heated surface, and according to the characteristics of nonlinearity and non-stationariness of the sequence, a method of predicting the heated area ash based on the empirical modal decomposition and time convolutional network of complementary sets is proposed. Firstly, the original sequence after wavelet threshold denoising is decomposed into a set of sub-sequence components by complementary set empirical mode decomposition, then the time series prediction model based on the time convolutional network is constructed for different sub-sequences, and the network hyperparameters are optimized to improve the prediction accuracy; finally, the prediction results of each IMF component are superimposed to obtain the prediction values of the cleaning factor. Compared with the other two models, the prediction accuracy is improved by 62. 1% and 57. 1%, respectively, and the CEEMD-TCN model has the highest prediction accuracy for the ash condition of the heated area, which verifies the accuracy and reliability of the model.

Abstract:For the problem of early fatigue damage detection and damage degree assessment in aluminum plates, this paper proposes a damage quantitative assessment method based on anomaly index ( AI). In view of the nonlinear nonstationary and chaotic dynamic characteristics of the structural system response caused by fatigue damage of aluminum plate, the signal time-frequency transformation and phase space reconstruction method are introduced to extract multidimensional damage features of aluminum plate, and the damage sensitive features are selected according to the monotonicity and the correlation between the features and damage degree. The aluminum plate damage detection problem is converted into a binary classification problem with a set of damage-sensitive features in the state description space, and a self-organizing feature mapping (SOM) network is used to identify the health status of aluminum plate. In order to further quantitatively characterize the damage degree of the aluminum plate, the SOM is used to fuse the damage sensitive features, and the AI values are used to quantitatively evaluate the damage state of the aluminum plate. The results of simulations and experiments showed that the SOM-based anomaly index proposed in this paper has high sensitivity and good dynamic tracking capability for fatigue damage evolution of aluminum plates, and has both good application prospects in the health monitoring and management of aluminum plate structures.

Abstract:The determination of the current rice grade mostly relies on manual picking and weighing, the discrimination process has defects such as strong manual subjectivity, slow detection speed and low efficiency. Therefore, it is an inevitable trend in the rice industry to realize rapid and automatic determination of rice grades. Based on machine vision technology, this paper designs and develops a rapid automatic determination system for rice grades. This system obtains images of rice grains with high-resolution through imaging technology, uses Watershed algorithm and adaptive threshold function to process the images, marks different grains and uses convolutional neural network training, selects the optimal training model to classify brown rice, Use linear regression to analyze the data to realize the judgment of rice grade. It has been proved by experiments that the similarity between the system and the artificial judging results of the same batch of rice can reach 91. 4%. The system designed by this method not only eliminates the human subjectivity in the process of judging the rice grading, but also detects the speed that has been significantly improved, thereby improving the efficiency of rice grading and judgment.

Abstract:In order to extract feature information from spectrogram more accurately, this paper proposes a speech enhancement method based on A-DResUnet ( attention-dilated ResUnet). The A-DResUnet model incorporates dilated convolution on the basis of ResUnet model to improve the ability to capture the contextual information of speech; at the same time, the convolution block attention module (CBAM) is added into the ResUnet encoder to improve the attention to the features of the noise spectrogram. The experimental results show that when the noise spectrum is used as the output target of the model, the model has a stronger ability to separate unknown noise than when the output target of the model is clean speech spectrum; compared with the ResUnet model, the proposed A-DResUnet model reduces the loss of speech detail information; compared with the speech enhancement methods based on DNN and GAN, PESQ increased by an average of 22. 81%, 33. 11%, STOI increased by an average of 9. 62%, 15. 33%, which is a more effective method for speech enhancement in complex environments.

Abstract:Aiming at the problem of low accuracy of transformer fault diagnosis, this paper proposes a power transformer fault diagnosis model based on improved parasitic predation algorithm ( IPPA) and optimized probabilistic neural network (PNN). Firstly, principal component analysis (PCA) is used to reduce the dimensionality of fault data to reduce invalid features, then use multiple strategies such as chaotic reverse learning, Cauchy mutation operator and the weight of linear decreasing function fused with beta distribution to improve the hunt-prey algorithm (IPPA) and its optimization ability, and use the improved IPPA algorithm to optimize the smoothing factor of the PNN network to improve the classification accuracy and robustness of the PNN. Finally, the PCA-processed data is input into the IPPAPNN model for fault diagnosis and testing based on the transformer data. The test results show that the accuracy of the IPPA-PNN model reaches 93%, which is 7% and 10% higher than that of the PPA-PNN and PSO-PNN models, and can effectively improve the fault diagnosis performance of the transformer.

Abstract:Most of the existing unsupervised domain adaptive fault diagnosis methods are only implemented based on a single domain signal, and the extracted fault information is not comprehensive enough. Only focus on realizing the edge distribution alignment of source and target domain features, ignoring the conditional distribution differences of samples, which limits the improvement of diagnostic accuracy. To overcome the above problems, a cross-domain fault diagnosis method of rolling bearings based on joint distribution offset differences (JDOD) is proposed. Two structurally consistent CNNs are used to extract the time-domain and frequency-domain features of the signal respectively to obtain more complete fault information. Joint distribution offset difference is proposed to realize edge distribution alignment and conditional distribution alignment of different domain features. Comparing experiments with various advanced methods on two multi-condition bearing datasets, the average diagnostic accuracy of more than 99% is obtained. The experimental results show that the joint distribution offset difference effectively improves the cross-domain fault diagnostic accuracy.

Abstract:The Bluetooth SIG announced support for angle of arrival ( AoA) and angle of departure ( AoD) in the Bluetooth 5. 1 specification, which provides a new solution to greatly improve the precision of Bluetooth positioning. However, the Bluetooth 5. 1 does not provide a unified algorithm for angle calculation. An improved two-dimensional MUSIC algorithm based on rectangular array was proposed to solve the problems of angle calculation caused by serious multipath interference in indoor environment. The antenna array is divided twice, and the covariance matrix of one-dimensional subarray is corrected by the method of forward and backward smoothing, and the covariance matrix of the whole array is corrected by the generalized inverse of the received signals matrix. This algorithm restores the rank of the signals covariance matrix without reducing the dimension of the signals covariance matrix, interference caused by indoor multipath sources can be effectively suppressed. Through the simulation experiment, it is proved that the proposed algorithm can greatly improve the accuracy of angle estimation under the condition of low fast beat number and low SNR. Positioning system is designed for indoor and outdoor experiments. The results show that the improvement of location accuracy is more obvious in indoor location. The circular probability error of stationary location is reduced by 26. 75% ~ 60. 25% in the indoor environment. Simulation and indoor positioning experiments show the effectiveness of the angle estimation algorithm proposed in this paper, this algorithm also has important reference significance for direction of arrival estimation in other application fields.

Abstract:To solve the problems of the existing remaining useful lifetime prediction methods for lithium battery with low prediction accuracy and long training time, a prediction model based on GBDT algorithm with grid search method is proposed. Firstly, analyze the charge-discharge cycle of lithium battery and select voltage, current and temperature as useful health index. Secondly, process the outliers of historical data and average useful health index data as feature input. Finally, establish the remaining useful lifetime prediction model for lithium battery by GBDT algorithm and optimize parameters by grid search method. Based on the capacity decay data of NASA lithium battery, the results show that the prediction model is superior to other methods about tenfold in RMSE, MAE, MAPE. The remaining useful lifetime prediction error is within 0. 05 and the training time reduces to 4. 5 s.

Abstract:Aiming at the difficulty of series arc fault detection and the difficulty of detection method based on decomposition strategy to capture sensitive discriminant components, a series fault arc detection method combining complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) decomposition and sensitive intrinsic mode function( IMF) selection was proposed. In this paper, the CEEMDAN algorithm was first applied to complete decomposition of arc current in series faults. Then, 12 feature indicators of arc current were defined, and the frequency band division of IMF component was realized according to the kurtosis index and energy feature which were more sensitive. On this basis, a feature calculation method based on time window was proposed to obtain the local features of the time scale of each high-frequency IMF component. Accurate selection of sensitive IMF components was realized by comparing feature indexes such as variance and root mean square value. Finally, for the current feature set, the second dimension reduction was realized by principal component analysis, and the series fault arc detection was implemented based on SVM. The feasibility of the proposed method and the validity of fault arc detection were proved by practical experiments.

Abstract:Aiming at the problem of misdetection caused by complex texture during the defect detection process of mesh fabric, a structured matrix decomposition method for mesh fabric defect detection is proposed. First, the image is enhanced by the Retinex algorithm, the feature matrix is generated using the extracted underlying image features, and it is decomposed into a low-rank matrix containing fabric image background information and a sparse matrix containing defect information. Secondly, the enhanced image is used to obtain Advanced priori matrix and index tree to achieve significant enhancement of defects. By calculating the value of the sparse matrix, the saliency of the defect is obtained. Finally, the defect saliency map is segmented by the optimal threshold segmentation algorithm to obtain the defect detection result. The performance of the algorithm is verified by using the defect images of the mesh fabric collected by the public data set TILDA and the CCD industrial camera. The results show that compared with other algorithms, the recognition accuracy of this algorithm reaches 94. 25%, the recall rate reaches 92. 48%, and the classification accuracy rate reaches 90. 12%.

Abstract:In order to improve the milling surface quality of screw rotor and other parts with spiral surface. According to the machining characteristics of screw rotor, the single factor rotation milling experiment is carried out according to the spindle speed, feed rate and intermittent feed. The improved particle swarm optimization algorithm is used to determine the optimal value of the initial weight and threshold of BP neural network. The trained improved BP neural network algorithm is used to predict the surface roughness of the milled screw rotor, and compared with the traditional BP neural network. The results show that the training accuracy of traditional BP neural network for surface roughness is the lowest, and the average relative error of 2000 iterations of particle swarm optimization in the improved algorithm is the lowest, which is 1. 21%. Using the model to predict the influence law of process parameters on surface roughness, it can be seen that under the premise of other process parameters unchanged, the surface roughness shows a decreasing trend with the increase of spindle speed; With the increase of intermittent feed rate, the surface roughness first decreases then increases; With the increase of feed rate, the surface roughness decreases first then increases. Conclusion: The improved neural network algorithm can accurately predict the surface roughness of spiral surface after milling, and provide theoretical guidance for the selection of process parameters in screw rotor milling.

Abstract:To solve the problem of large harmonics of output voltage and current of three-phase inverters with unbalanced and non-linear loads, an optimal sliding mode preview repetitive control strategy is proposed to achieve fast response to abnormal conditions such as unbalanced loads and non-linear loads and high precision tracking of reference voltage. Firstly, a preview controller is introduced in the feed-forward compensation link, and an extended state error system including target signal and feedback from lagging link is designed by using difference operator, which converts the control problem of the inverter into the regulation problem of linear discrete system. Furthermore, by using Lyapunov method, linear matrix inequality and design method of optimal controller, an optimal sliding mode preview repetitive controller including sliding mode control, state feedback, repetitive control and preview compensation is obtained. The experimental results show that the distortion rate of voltage waveform decreases from 3. 12% to 0. 78% and the response time decreases from 10 ms to 5 ms when the optimal sliding mode preview repetitive control with linear load. The distortion rate of voltage waveform with non-linear load is reduced from 6. 72% to 0. 92%, which verifies the effectiveness of the proposed method.

Abstract:Fish segmentation in underwater environment is the key technology to realize intelligent measurement such as body length measurement, weight estimation and population counting. In order to improve the accuracy of fish segmentation, an improved Deeplabv3 + fish segmentation method combined with edge supervision is proposed. In the encoder part, fewer down sampling times are used, and convolutional block attention module (CBAM) is added in the shallow layer to reduce information loss and enhance the shallow semantic information; By designing hybrid dilated convolution (HDC) to improve atrous spatial pyramid pooling(ASPP) module, deep features are extracted. In the decoder output part, Canny edge detection operator is combined to introduce edge supervision, and the edge prediction and edge label errors are obtained through the edge loss function to better learn edge features. Finally, the optimized loss function is introduced according to different pixel ratios to further improve the semantic segmentation performance of the model. This method achieves 84. 56% mIoU on VOC2012 dataset, which is 3. 27% higher than Deeplabv3+ method, and verifies its generalization ability. In the ablation experiment on DeepFish dataset, mIoU is as high as 93. 66%, which is higher than common methods such as Deeplabv3+, Unet and PSPNet. This research improves the accuracy of fish segmentation in underwater environment and can provide support for intelligent aquaculture.

Abstract:To effectively improve the maintenance efficiency and fault location accuracy of high-speed railway turnouts, a combined model for multi-level fault diagnosis of high-speed rail turnouts based on character and word fusion was proposed. Firstly, a professional thesaurus of high-speed rail turnout equipment was established, and fault texts were represented as character vectors and word vectors and the character vectors and word vectors were deeply fused. Secondly, considering the problem of imbalanced categories in fault texts, the Borderline-SMOTE algorithm was used to process the imbalanced text data to optimize the fault text data distribution. Then, a combination of Bi-directional long short-term memory ( BiLSTM) and convolutional neural network ( CNN) was used to extract deep features of the fault text. Finally, an intelligent diagnosis of faults was achieved by means of a classifier. The model performance was validated using fault text data of China high-speed railway turnout faults. The test results show that the accuracy of the proposed model reaches 95. 62% for the primary fault diagnosis and 93. 81% for the secondary fault diagnosis, which proves that the multi-level fault diagnosis accuracy can reach the desired effect.

Abstract:The task of lane line detection includes difficult samples such as road wear, shadow occlusion and curves. The line information in these samples can be missing with different levels, which results in missed or false detection of the detection results. The detection scheme based on deep learning extracts feature information through convolution operation. Convolution operation discards a series of tedious operations of traditional image processing, such as manually designing filters, and benefits from weight sharing and inductive bias, which greatly reduces the workload of feature extraction. This operation not only reduces the image resolution, but also obtains long-distance information, resulting in the loss of regional edge and other details of the small resolution feature map, which affects the quality of the detection results. In deep learning, the segmentation model processes more detailed information than the detection model. Based on the segmentation model, this paper introduces transformer to improve the sampling method and improve the lack of convolution operation in obtaining global information. After the model is improved, the test accuracy on Tusimple is improved by 0. 4%, the pixel accuracy is improved by 0. 3, and the amount of multiplication and accumulation operation is increased by 36. 09 G. The results show that the transformer’s unique sampling method can improve the lack of convolution operation sampling, and improve the situation of missing detection of lane line difficult samples in semantic segmentation network.

Abstract:Aiming at the prediction accuracy of transformer hot spot temperature, the ant colony algorithm ( ACO) combined with improved principal component analysis ( IPCA) was proposed to optimize BP neural network model to predict hot spot temperature. Firstly, IPCA is used to remove data redundancy information and solve the correlation between parameters to improve the ability of network generalization. In order to avoid BP neural network that is easily falling into local optimum and slow convergence speed, ACO was used to optimize the weights and thresholds of the network to speed up the algorithm and improve the prediction accuracy. Verified by the measured transformer temperature data, the mae, mse and mape indexes in the predicted results are 0. 065 7, 0. 006 7 and 0. 44%, respectively. The prediction accuracy and network performance are better than those of IEEE, BP and IPCA-BP models, thus verifying the validity and feasibility of the proposed model.