Abstract:As an important part of the clock circuit, the crystal oscillator is the module with the largest power consumption in the clock circuit, in view of the problem of excessive power consumption of the traditional crystal oscillator, according to the characteristics of the average value of the output voltage signal in the case of starting and stopping the crystal oscillator, a method is proposed to obtain the minimum current required for the oscillator to maintain the oscillation state, which greatly reduces the power consumption of the crystal oscillator circuit. A Pierce oscillation circuit was designed based on the 28 nm CMOS process. The simulation results showed that at a supply voltage of 1.8 V, the circuit can start up quickly within 200 ms, the oscillation frequency is 32.768 kHz, and the oscillator operates at only 270 nA after the output clock signal is stable.

Abstract:A terahertz graphene patch-type high-temperature sensor is designed to solve the problems of the small temperature measurement range, large size, and low sensitivity. The sensor is composed of a hexagonal graphene film, an alumina substrate, and a three-layer gold structure. The characteristic size of the terahertz MEMS high temperature sensor with large range and high sensitivity is obtained by optimizing the graphene structure size and alumina substrate thickness. The results show that the temperature sensor can measure temperatures ranging from 25 ℃ to 1 050 ℃ in a frequency range of 3.0~5.0 THz, with a sensitivity of 0.259 GHz/℃ and an overall volume of 20 μm×21 μm×14.5 μm. The sensor has excellent sensing performance and a wide temperature range, allowing it to provide a technical method for measuring the temperature of an aero-engine, rocket launch system, or harsh environment monitoring system in a high temperature environment.

Abstract:To address the problem that the traditional LLC resonant circuit still needs to convert AC to DC before DC conversion in the process of power conversion, this paper proposes a type of topology of LLC resonant converter based on AC-Link technology with corresponding control method. The topology can convert three-phase AC power to DC power by the control strategy of frequency conversion, and the input current has high power factor and low harmonic current, while the output voltage has excellent regulation ability. Finally, a Matlab/Simulink simulation model is established and a 1 kW experimental prototype with 380 V three-phase AC input and 500 V output is built for verification. The results show that the power factor of this topology can reach 0.99, the total harmonics distortion of the input current is 2.31%, and the output voltage error is less than 1%. Finally, the experimental prototype is constructed to test, and the experimental results prove the feasibility of the topology and control algorithm.

Abstract:Aiming at the problems of less training data and high labeling cost of named entity recognition in the military aircraft maintenance field. The paper proposed an improved named entity recognition method based on pre-training BERT. Firstly, learn from the idea of remote supervision, we fuse the boundary features of remote Tag word on token to get the feature fusion vector. Then the vector is sent to BERT to generate a dynamic word vector representation. Finally, the CRF model is connected to get the global optimal result of the sequence. Experiments are carried out on the self-built dataset, and the F1 value reaches 0.861. In order to compress the model parameters, the trained BERT-CRF model is used to generate pseudo label data, and the student model BiGRU-CRF with less parameters is trained in combination with knowledge distillation technology. The experimental results show that compared with the teacher model, the student model reduces 95.2% of the parameters and 47% of the reasoning time at the cost of losing 2% of the F1 value.

Abstract:In order to solve the energy management problem of fuel cell sightseeing vehicle, the model of fuel cell sightseeing vehicle and its components are established, and the thermostat control strategy is designed. The accuracy of the model is verified by the bench test. In order to improve the economy and durability of the vehicle, a novel strategy based on quadratic utility function is proposed, and the real-time maximum benefit is obtained by KKT condition. Finally, a multi-objective lightning search algorithm is designed, and the optimal parameters of the novel strategy are solved by the algorithm. Simulation results show that compared with thermostat control, the proposed strategy can improve the driving range by 1.7% and durability by 11.2%. In addition, the strategy takes into account the SOC and combines the historical output power information of the vehicle and its components. The novel strategy also has a strong adaptability to working conditions.

Abstract:With the rapid development of my country′s smart grid, live robots are gradually being used to replace manual tasks for routine maintenance and accident repairs. However, there are problems of low control precision and jitter in live work robots, which affect the quality and efficiency of work tasks. In order to solve this problem, a fuzzy PID (proportion intergration differentiation) controller is designed to realize the precise control of the output displacement of the live working robot. Firstly, the structure of the live working robot and the principle of the hydraulic system are analyzed, the mathematical model of the main components of the hydraulic system is established, and the transfer function model of the system is deduced. Secondly, the closed-loop fuzzy PID controller of the output displacement of the live working robot is designed based on fuzzy rules. Finally, the simulation models of the original system, PID and fuzzy PID control schemes are established in the Simulink environment for simulation analysis, and the prototype is tested experimentally. The simulation results show that: compared with the original system and the PID control system, the adjustment time of the fuzzy PID control system is relatively shortened by 81.3% and 30.8%, and the overshoot is reduced by 0.064, achieving no overshoot and having better position tracking ability. In addition, the experimental data results of the prototype are consistent with the simulation results, indicating that the fuzzy PID controller has better application effect.

Abstract:In order to meet the needs of calibration and verification of AC voltage parameters, a high precision AC voltage standard source circuit is designed in this paper. Based on the requirements of high precision and high stability, the overall circuit adopts a closed-loop feedback structure, uses AD9850 chip to generate AC signal, and then modulates the amplitude of the sine wave through the DC reference voltage. The microcontroller of the circuit uses AVR chip. At the same time, the signal analog circuit and the digital control circuit are electrically isolated, and the accuracy is improved through hardware selection, software compensation and other measures. The output range of the designed circuit is 1~10 V, and the output frequency band is 40 Hz~100 kHz. The AC voltage standard source is repeatedly measured at typical frequency points. The results show that the permissible error of the output full-scale voltage is less than 0.006% in 40 Hz~50 kHz, the permissible error of the full-scale voltage at 100 kHz is kept within 0.025%. The circuit has the advantages of high precision and good stability.

Abstract:Non-line-of-sight link visible light communication has the advantages of wide signal coverage and not easy to block the link due to occlusion. However, there are some problems, such as weak link signal and small field of view of the receiving system using a single detector. Aiming at the above-mentioned problems, a receiving mode of multi-detector fusion is proposed. The influence of receiving azimuth angle on the received signal is studied experimentally, the signal-to-noise ratio distribution of the received signal based on the single-detector receiving system and the four detectors receiving system is measured respectively, and the communication performance of the non-line-of-sight link communication system based on four detectors is tested. Experimental results show that the multi-detector fusion receiving method can effectively improve the performance of the non-line-of-sight link communication system, the maximum signal-to-noise ratio of the system with non-line-of-sight link communication can be increased by 10.71 dB by using four detectors, and the maximum communication rate of the system can reach 1 Mbps and the bit error rate is less than 10-6 when the line-of-sight link is completely blocked.

Abstract:A high-precision positioning method was proposed to meet the demands for automatic positioning in complex backgrounds. First, sub-pixel edge sequences are derived from discrete edge points of a preprocessed image by applying an edge tracing algorithm and interpolation algorithm. Next, the non-circular edges were excluded by salient point segmentation and center constraints based on geometric features. Then arcs belonging to the same circle are reorganized by more efficient relative position constraints and inscribed triangle constraints. And finally the arc segment group was subjected to least squares fitting to achieve accurate positioning of the target center. In this paper, by extracting sub-pixel arc sequence and a more efficient arc recombination method, the positioning accuracy is improved and the time of arc recombination is reduced. The experimental results demonstrate that the positioning accuracy of the simulated image reaches 0.007 pixel, and the accuracy of the dynamic displacement experiment reaches 0.05 mm, which can meet the needs of visual measurement real-time positioning under adverse conditions such as illumination changes, noise, and occlusion.

Abstract:Aiming at the problem of the natural frequency of the piezoelectric ultrasonic transducer in the ultrasonic therapy apparatus that causes the factory error of the natural frequency due to the manufacturing process,and the frequency drift caused by temperature changes and other factors, it will work in a non-resonant state, which will greatly reduce the electro-acoustic conversion efficiency of the piezoelectric ultrasonic transducer, and the treatment waveform will change and the output power will be unstable, which will affect treatment effect of the ultrasonic therapeutic apparatus. An automatic tracking frequency locking circuit of ultrasonic therapeutic apparatus is designed.The system uses STM32F103ZET6 as a microcontroller. Using the principle of inverse piezoelectric effect of piezoelectric ultrasonic transducer and the maximum voltage feedback method, the maximum voltage value and the resonant frequency are fed back in real time through frequency sweep sampling, so that the piezoelectric ultrasonic transducer always works at the resonance point, so that the To achieve automatic tracking frequency locking. The experimental test proves that the ultrasonic therapeutic apparatus based on the automatic tracking frequency locking circuit can maintain a relatively stable output power, a good treatment waveform, accurate frequency tracking and low cost, and can achieve a good treatment effect. Therefore, this subject has better innovation, application prospect and broad applicable value.

Abstract:For the study of wave spectrum wind speed inversion, although a preliminary inversion model with a slope of -4 power law in the wave spectrum equilibrium range has been established, however, for the complex and variable measured wave spectrum, the inversion of a single slope model is not effective. Aiming at the problem of variable spectral slope, an adaptive wave spectrum slope wind speed inversion algorithm is proposed by quantifying the spectral coefficients corresponding to the wave spectrum function. In order to verify the accuracy of the algorithm for wind estimation, this paper investigates the inversion of sea surface wind speed by using a large amount of wind and wave data observed by NDBC buoys, and the results show that the correlation coefficient R=0.83 between the inversion wind speed of the adaptive wave slope algorithm and the measured wind speed, which is significantly correlated, and the mean deviation BIAS and root mean square error RMSE are 0.75 m/s and 2.48 m/s, respectively. The offshore sea trial experiments show that the algorithm is still effective and universal. In general, the error between the adaptive wave spectrum slope inversion wind speed algorithm and the measured wind speed is relatively small and the inversion accuracy is better, which provides a new idea for wind and wave observation and wind and wave research, and the algorithm can be integrated and applied to the ocean observation application platform, combining theoretical research with engineering practice.

Abstract:In order to generate high resolution and wide dynamic delay sampling pulses, a coarse + fine two-stage delay framework is introduced. The external 0.1~12 GHz clock signal is divided by the phase-locked loop to generate a synchronous clock of about 100~250 MHz to drive the counter to count. When the counter reaches the preset value, a synchronous carry pulse signal with a frequency of 50 kHz is generated, and the resolution is 10 bit and 10 ps. The coarse delay chip and the fine delay chip with a resolution of 0.1 ps start to work and output sampling pulses with a certain delay amount, and the sampling pulse drives the sampler to precisely sample the synchronous radio frequency signal. The test results show that the resolution and dynamic delay range of the digital programmable precision delay trigger can reach 1ps and 10 ns, respectively.

Abstract:The attitude Angle information of aircraft is an important parameter information in the process of aircraft rescue, but the existence of magnetic interference will affect the accuracy of attitude solution. This article in the attitude Angle algorithm proposed based on extended kalman and the improving method of the linear combination of the kalman, extended kalman filter to update of pitch Angle and roll Angle, while the linear kalman separately for the course Angle is calculated, and join the adjustment coefficient, when the optimal estimate by judging measuring magnetic field and magnetic field Angle to adjust course Angle, the optimal estimate, In order to reduce the influence of external magnetic interference on the heading Angle. Experimental results show that compared with the traditional extended Kalman attitude Angle settlement algorithm, the improved algorithm is more close to the real aircraft attitude information, and has stronger anti-interference ability to the external magnetic field changes.

Abstract:A strong tracking robust extended Kalman filter is proposed for attitude control systems of satellite with state delay and unknown uncertainty to implement concurrent fault estimation of actuator and sensor. Firstly, considering the system noise, taking the faults as the auxiliary variables of the system, an augmented time-delay nonlinear system is established. Then, a robust extended Kalman filter is proposed, and a robust upper bound is introduced to decrease the linearization error. Further, for low accuracy of prediction covariance caused by system process uncertainties, a strong tracking algorithm based on multiple sub-optimal fading factors is introduced to decrease the influence of uncertainty on filtering accuracy. Finally, a simulation example is given to compare the proposed method with the robust extended Kalman algorithm as well as the extended Kalman algorithm. The simulation results show that, compared with the other two algorithms, the average value of the root mean square error of the state estimation and fault estimation by the proposed method is reduced by 69.2%, 60.6% and 88.1%, 78.9%, respectively. The simulation results verify the effectiveness of the design scheme.

Abstract:In view of the complexity and diversity of the disaster causing factors that cause debris flow disasters, resulting in the excessive dimension of the model input data and the problem that extreme gradient boosting is easy to fall into local optimization, resulting in the low accuracy of the prediction model. A debris flow disaster prediction method based on GWO-XGBoost model is proposed. First, the original data collected by the sensor is preprocessed to obtain the standard data, and then the dimension of the data is reduced by linear discriminant analysis, and the disaster causing factors with low coupling and high contribution rate are obtained as the model input to predict whether the debris flow disaster occurs; Secondly, grey wolf optimizer is used to optimize the super parameters of the model; Finally, Mozigou monitoring data are used for simulation verification. The results show that the normalized data after preprocessing and dimensionality reduction by linear discriminant analysis algorithm solves the problem of dimensionality disaster of model input. The prediction accuracy of GWO-XGBoost debris flow disaster prediction model is 96.64%, which is 6.69%, 5.13% and 3.86% higher than that of random forest model, support vector machine model and xgboost model respectively, It enriches the prediction methods of debris flow disasters and provides new ideas for relevant decisionmaking departments.

Abstract:The generation of test data is an important part of achieving software test automation. In order to improve the quality and efficiency of test data generation in unit testing, a test data generation algorithm based on shuffled frog leaping algorithm is proposed. The algorithm introduces a dynamic threshold to control the moving step size of individuals, so as to balance the global exploration and local exploitation abilities. At the same time, the worst individual random jump strategy in the standard algorithm is transformed into learning from random individual to enhance the information exchange between populations to improve the algorithm′s global search capability. Apply the improved algorithm to test data generation. The experimental results show that the improved shuffled frog leaping algorithm is more stable than the standard shuffled frog leaping algorithm, cuckoo search algorithm and particle swarm optimization algorithm under the condition of changing population size. The improved shuffled frog leaping algorithm is better than the comparison algorithm in the evaluation index of the average number of iterations generated by the test data.

Abstract:When the predictive control system has a large response time or a non-self-balancing object system with integral element, it is necessary to increase the prediction time domain to improve the control effect. Since optimization exists in the prediction time domain, the numerical solution process may fall into a pathological state as the prediction time domain is too long. Therefore, an exponential weighted asymptotic stability optimization strategy is proposed. Firstly, the state space model is reconstructed by embedding integral function to simplify the feedback correction link, and Laguerre function is introduced to further improve the MPC execution efficiency; Then, in the predictive control design, exponential weighting is used to specify the model eigenvalue in the unit circle to ensure the stability of the model; Finally, the closed-loop asymptotic stability of long predictive time domain optimal control is achieved by modifying the weight matrix. The simulation results of single variable and multivariable systems with integral objects show that the proposed MPC algorithm can effectively avoid ill conditioned numerical solutions and improve the dynamic and steady-state performance of the system.

Abstract:When patients with hand dysfunction practice rehabilitation movements by themselves at home, the training steps are confusing due to the lack of scientific guidance, and the precision and intensity of the movements are difficult to be guaranteed, thus affecting the rehabilitation effect. In this paper, an active hand training system based on Arduino and STM32 is designed. The system is divided into three major parts: data glove, movement guiding palm and upper computer, The data glove obtains and processes the rotation angle data of the fingers and wrist through sensors, and then transmits the data wirelessly to the movement guidance palm, and the controller STM32 in the movement guidance palm compares the received data with the standard movement database, analyzes the movement standard, and then instructs the patient to make adjustments by voice.STM32 drives six digital servos to drive the movement of the bionic palm according to the movement data, thus imitating the human movement. The system uses the LD3320 to recognize patient commands and perform human-computer interaction. The system downloads the standard movement database to the STM32 external flash memory through the upper computer, which is used to compare the patient's hand movement data. The experimental results show that the system can effectively guide patients to complete the whole set of rehabilitation training movements with accurate data reading, precise and reliable guidance, and strong interactivity. It can help about 6 million stroke patients with hand motor dysfunction in China for rehabilitation training, which has strong application value.

Abstract:Electroencephalography (EEG) contains rich information about brain function, which is very important for the detection and diagnosis of different types of neurological diseases. In view of the fact that a single feature cannot fully express the EEG signal, this paper combines frequency domain features and spatiotemporal information to better represent the signal, and proposes an attention network based on spatio-temporal and frequency domain features (STFACN) for automatic detection of Parkinson′s disease (PD). From the perspective of frequency domain, the average power characteristics of Delta, Theta and Alpha frequency bands were obtained from the multi-channel EEG using the fast Fourier transform method. In terms of spatiotemporal feature extraction, a compact convolutional neural network based on spatiotemporal features is constructed, and the channel attention mechanism is embedded into the network, and the adaptive extraction can characterize the effective features of PD. Finally, the model based on frequency domain features is fused with the compact convolutional neural network model based on spatiotemporal features, and experiments are carried out on the University of New Mexico (UNM) dataset. The specificity, sensitivity and accuracy reach 87.97%,84.39% and 86.89% respectively. Cross-dataset experiments are performed on the University of Iowa (UI) dataset, and the accuracy rate reaches 77.33%. The experimental results show that compared with the existing methods, the method proposed in this paper can mine effective features from the original EEG, and has high accuracy and strong generalization ability in the EEG-based Parkinson′s identification problem.

Abstract:The state of health (SOH) of lithium-ion batteries is an important reference indicator for the reliable operation of lithium-ion batteries. To improve the accuracy of the battery state of health detection, a method for the lithium batteries state of health detection based on the CNN-BiLSTM network is proposed. This method uses CALCE lithium-ion battery capacity decay data set, extracts battery health indicator (HI) as the model input data, and uses grey relational analysis (GRA) to verify the rationality of HI selection. Convolutional neural networks (CNN) and bi-directional long short-term memory (BiLSTM) are used to construct network models to predict battery capacity and to detect the health status of lithium-ion batteries. The results show that the method has 1.79% RMSE and 1.3% MAE for SOH detection, with high accuracy and reliability.

Abstract:The partial discharge caused by the insulation fault of the transformer winding endangers the safe operation of the power system. At present, the ultrasonic external detection method for partial discharge is greatly affected by factors such as noise interference and propagation medium. Therefore, in this paper, eight sensors are built in the transformer winding. Internal detection of ultrasonic signals generated by partial discharge. Because the direct measurement method of the detected signal cannot reflect the actual time delay information, and the traditional intelligent algorithm has low positioning accuracy and poor stability, this paper proposes an ultrasonic method that combines the generalized crosscorrelation method (GCC) and the mixed strategy sparrow algorithm (MSSA). The transformer winding model is built in COMSOL and the partial discharge phenomenon is simulated, and the accurate time delay information is extracted by the GCC algorithm. According to the time difference information, the optimal constraint equation of ultrasonic positioning is established, and the MSSA algorithm is used to obtain the position information of the PD source. The simulation results show that the MSSA algorithm has a smaller positioning error than other traditional intelligent algorithms. Finally, a new type of EFPI ultrasonic sensor is used to conduct partial discharge experiments on transformer windings. The results show that the average positioning error of the method proposed in this paper is 4 cm, which is 2.9 cm larger than the simulated partial discharge point，which has practical value for engineering application.

Abstract:The traditional cloud detection methods are less effective in recognizing special scenes, which cause problems such as edge information loss and thin and broken clouds misjudgment. In this study, MSHA-DeepLab algorithm based on multi-scale feature fusion and hybrid attention is proposed for high-precision cloud detection. First, the attention module is introduced based on the original algorithm, which to increase the weight of important features and improve the sensibility of local features. Second, depthwise separable convolutions are used to extract the multiscale semantic information and reduce the amount of network parameters. Finally, continuous up-sampling and feature fusion are performed to reduce the loss of feature information. After testing and comparing the datasets with different scenes and different band combinations using various methods and the improved algorithm, it can be seen that the precision of the algorithm reaches 86.376 9%, the recall reaches 85.895 9%, the sepecificity reaches 96.915 6%, the IoU reaches 82.846 7%, the accuracy reaches 94.600 8%， which is a significant improvement compared with the original algorithm and other methods. It is verified that the proposed algorithm can achieve high accuracy cloud detection under different conditions.

Abstract:To solve the problem that existing graph convolution network (GCN) need to pre-define human skeleton topology and the model is large, a fall detection algorithm based on spatiotemporal adaptive graph convolutional network (ST-AGCN) is proposed. The network consists of three parts: firstly, HRNet, a human pose estimation algorithm, is used to extract human skeleton points from video and preprocess them into four-dimensional tensor. Secondly, the normalized embedded Gaussian function is introduced to obtain the human body topology by learning (without manual pre-definition), and the human body correlation features are obtained by spatial adaptive graph convolution. Thirdly, multi-scale convolution is used to extract temporal motion features to improve the model′s ability to obtain dynamic information. Simulations are carried out on public and self-built dataset, and the accuracy rates are 95.45% and 99.55%, respectively. The results show that the proposed algorithm is better than the current GCN methods, and the number of parameters is only a quarter of the latter, or even less. Another advantage of our algorithm is that it can be applied to different datasets.

Abstract:Bird′s nest encroachment is a frequent fault of transmission lines. Birds nesting on the tower will affect the insulation performance of the tower, resulting in tripping accidents. Traditional bird′s nest identification methods for transmission lines are inefficient and lack of security. Therefore, this paper proposes a bird′s nest detection algorithm for transmission lines based on improved YOLOv5 model. By adding CBAM attention module to the backbone network, the feature extraction ability of the backbone network can be improved with less computational cost. The adaptive feature fusion module is introduced into the neck network to replace the original structure and enhance the multi-scale feature fusion effect. The more stable and smooth Mish activation function is used as the activation function to improve the classification accuracy and generalization ability. Experimental results show that, compared with the original YOLOv5s model, the recall rate and average precision of the improved method are improved by 4.4% and 2.3% respectively. It shows good performance for occlusion targets and near-far targets, which verifies the effectiveness of the improved method.

Abstract:To improve the effect of plant segmentation and achieve accurate acquisition of plant phenotypic parameters, this study proposes a plant segmentation network to fuse spatial-frequency domain feature representation. Based on the U-Net network architecture, the frequency domain transform module is built for down-sampling, the frequency domain feature representation is introduced in the convolutional neural network to replace the pooling layer, and the frequency domain transform method of 2D-DCT and 2D-IDCT is used to perceive the global semantic features of the plant. The multi-scale feature fusion module is constructed by adding six up-sampling nodes to extract and connect the fine-grained feature information of the plant image. The channel attention module is modified to learn branch features and the hybrid loss function is employed to optimize the network. Experiments are conducted on the 2017 CVPPP public dataset, and the results show that the intersection over union, mean intersection over union, pixel accuracy, precision and F1 of the plant segmentation network reach 97.07%, 98.04%, 99.53%, 99.68% and 99.74%, respectively. Compared with the FCN-8s, FCN-ResNet, DeepLabV3+, SegNet and U-Net models, the intersection over union and mean intersection over union of the network were improved by up to 23.32% and 12.43%. The proposed method can improve the segmentation accuracy of plant at smaller scales in detail processing, and it can provide a useful idea for applied research in the plant phenotype.

Abstract:The metal shielding cavity can ensure the normal operation of the on-board recorder in the strong electromagnetic environment, but the heat dissipation and cable openings will reduce the electromagnetic shielding efficiency, so the structure needs to be optimized to improve the shielding efficiency. The influence of different opening schemes and structural parameters on shielding effectiveness is analyzed, and the calculation formula of electromagnetic shielding effectiveness is derived. Combined with the advantages of simple calculation by analytical method and accurate calculation by numerical method, the adaptive weighted PSO optimization algorithm is used to quickly find the optimization value of the analytical formula, and the numerical method is used to simulate the optimization result to calculate the fitness value, so as to obtain the final optimization result. The results show that compared with the pure numerical method, the calculation cost is reduced by 88.54%. Under the same shielding efficiency, the thickness of the shielding structure is reduced from 3.4 cm to 2.99 cm, and the volume is reduced by 9.03%; When the volume of the shielding structure is about the same, the shielding efficiency is increased from 98.79 dB to 102.65 dB. It can be seen that the proposed optimization method can greatly improve the design efficiency while improving the optimization effect.

Abstract:An intelligent fault diagnosis method for spiral bevel gear box based on continuous wavelet transform (CWT) and coordinate attention mechanism residual network (CooAtten-Resnet) is proposed. Firstly, a large number of signal samples are obtained by overlapping sampling of vibration signals. These samples are converted into time-frequency maps by continuous wavelet transform, and time-frequency data sets under different faults are constructed. At the same time, noise samples are added manually to verify the impact of noise on such diagnostic methods; Then the time-frequency map data set is used for CooAtten-Resnet training; Finally, the fault is classified and the diagnosis results are output. The results show that this method can accurately identify the fault of spiral bevel gear box, and the accuracy rate of diagnosis can reach 100% when no one adds noise, and the accuracy rate is still above 93% when no noise reduction is conducted after adding noise. Compared with other methods, this method has higher accuracy, stronger anti-noise ability, faster network convergence and more stable diagnosis results.

Abstract:Aiming at the problem that it is difficult to extract the features of planetary gearboxes under harsh conditions and difficult to classify accurately under various fault states. Based on the Empirical Wavelet Transform, the Improved Empirical Wavelet Transform is proposed, which replaces the original spectrum decomposition with the scale-spectrum decomposition which is more stable under noise interference. A fault diagnosis method combining Improved Empirical Wavelet Transform and Deep Extreme learning machine. Firstly, the signals of the planetary gearbox under different fault conditions are denoised by IEWT respectively and the FM-AM components of each order are extracted. Then, Multiscale sample entropy of the first six components with higher Envelope spectrum kurtosis was selected as the fault feature set and input into DELM for fault diagnosis and classification. The results of planetary gearbox fault diagnosis test show that compared with the fault diagnosis accuracy of EWT, EMD and DELM, the average fault recognition rate of this method can reach 97.6%, which has certain effectiveness.