Abstract:In order to effectively cope with the change and growth of telemetry requirements of precision strike weapon flight test, this paper proposed a wireless transmission system design scheme of missile navigation information based on Zigbee technology. The node of wireless telemetry network was based on ARM+UZ2400 architecture. According to the requirements of telemetry test, a two-stage power amplifier structure was adopted to amplify the output RF signal. The missile transmitting antenna was designed with a miniaturized dipole antenna based on sleeve structure, which effectively solved the problem of insufficient available space on the missile. In the software part of the system, the data processing framework was designed according to IEEE 802.15.4 standard to complete the transmission of navigation information. Finally, the test framework of the telemetry network node data transmission system was built to test the quality of communication links and the accuracy of data transmission. The test results showed that the Received Signal Strength Indicator error stabilized at 0.505, and the consistency rates of horizontal coordinate accuracy, vertical coordinate accuracy, horizontal velocity accuracy and vertical velocity accuracy with the design indicators of 16 channels were 81.25%, 100%, 100% and 100%. Consequently, the system can accurately and stably transmit the real-time position information and velocity information of the telemetry target in 2.4 GHz band and met the design requirements of miniaturization and versatility of the missile system.

Abstract:The International Telecommunication Union (ITU) defines the uplink for ship-to-satellite communications for the VHF Data Exchange (VDE) subsystem of the VHF Data Exchange System (VDES). This paper studies a demodulation and decoding technology for VDE uplink signal to realize uplink physical layer communication. Firstly, the data-aided and non-data-aided methods are used to estimate and correct the frequency offset and phase offset, and the timing recovery algorithm is used to achieve symbol synchronization. Secondly, the ID segment and the data segment of signal are decoded by using the majority-logic decoding algorithm and the iterative decoding based on the Max-Log-Map algorithm respectively. Finally, the whole system is simulated. According to the results, under the frequency offset of 4 kHz, when Es/N0≥5 dB， the error probability of signal ID decoding is less than 0.01, and the frame error rate of data segment is less than 0.1, which shows the effectiveness of the algorithm under large frequency offset.

Abstract:Requirement for playback of baseband signals of different bandwidths to the target intermediate frequency, a variable bandwidth baseband signal playback system was designed. The system took FPGA as the core data processing unit, and completed the playback function of the baseband signal by configuring the high-speed digital-to-analog conversion chip AD9122 and the clock chip AD9516. In order to solve the problem that the sampling rate of baseband signal with different bandwidth does not match the transmission rate of DAC, a multi rate processing algorithm is designed, The cascade structure of multi-stage HB filter, CIC filter and Farrow filter is adopted to realize the function of arbitrary multiple sampling rate conversion. The algorithm simulation and actual test results show that the system can play back the 1 kHz~20 MHz variable bandwidth baseband signal to the target IF with less resource consumption The playback signal has spurious free dynamic range not less than 60 dBc, meeting the requirements of the actual communication system.

Abstract:The active power filter (APF) is a new type of power electronic device with the functions of dynamic harmonic suppression and reactive power compensation. The accurate and rapid detection of current harmonics in real time is an important part of determining the performance of APF. The fast Fourier transform (FFT) is a widely used harmonic detection method. However, the conventional FFT algorithm is complicated in calculation, has time delay, poor real-time performance, and is easily affected by grid voltage waveform distortion or frequency fluctuation, which affects the accuracy and efficiency of harmonic detection, thereby reducing the compensation effect and comprehensive performance of APF. Therefore, this paper proposes an APF harmonic detection and compensation strategy based on split-radix FFT algorithm. Through butterfly operation, the radix-2 algorithm is used for the input of even numbers, and the radix-4 algorithm is used for the input of odd numbers, effectively reducing the complexity of the FFT algorithm, enhancing the real-time detection of harmonics. The split-base FFT algorithm is optimized by using Kaiser window to improve the harmonic detection accuracy and anti-interference ability of harmonic detection, ensuring the harmonic detection and compensation effects and overall performance. The correctness and effectiveness of the proposed harmonic detection strategy are verified by the three-phase four-wire APF prototype experiment, and the conditioning time to re-reach the new steady-state can be shortened by 25% in the case of sudden load changes.

Abstract:In land drilling, the land drill column conveying robot can automatically grab drill columns, adjust the attitude of drill columns, and then transport drill columns to the rat hole. However, drill columns may be inclined when placed, causing the robot to fail to grasp. Therefore, this paper designs an automatic deviation correction algorithm for drill column transportation based on laser sensors. Firstly, two laser sensors are used to scan and acquire the distance information of any two sections of the drill column, and the position of the center of the section is obtained by least square fitting. Due to the fact that the soil adhered to the surface of drill strings will affect the collected data during actual field operations, the Grubbs criterion is used to screen the data through multiple iterations. Then, the deflection angle of the drill string is calculated by the center of the circle and sent to the manipulator for positioning and grabbing. Finally, the effectiveness of the algorithm is tested by simulating the different positions and thicknesses of the soil attached to the drill string in the drilling site. The experimental results show that the algorithm is feasible, highly portable in engineering, meets the requirements of engineering application, and improves the automation level of drill string conveying robots.

Abstract:In order to reduce the manual collection of fingerprint data and obtain high positioning accuracy, a fingerprint database reconstruction and node location of aircraft structure strength test algorithm based on WLAN fingerprint was proposed in this paper. The support vector regression method was used to reconstruct fingerprint data, K-means algorithm was used to reduce the workload of fingerprint collection, and the optimized DBN was used to extract the features of RSS information. Finally, the WLAN fingerprint location database of the aircraft body is established, and the algorithm performance and system were analyzed and evaluated through simulation experiments. The experimental results showed that the average positioning errors of IPDBN-54, IPDBN-41 and IPDBN-26 in KNN algorithm were 10.389 2, 10.786 3 and 11.117 7 respectively. In the WKNN algorithm, the average positioning errors of IPDBN-54, IPDBN-41 and IPDBN-26 were 10.290 4, 10.714 3 and 11.103 8, respectively. The average positioning error of IPDBN was the smallest and the positioning accuracy was relatively high. Compared with BPNN, the average training time of IPDBN was 166.2 s, with relatively low training time. The optimized depth belief network algorithm has strong adaptability to the establishment of WLAN fingerprint location database system, with short training time and high location accuracy. The research aims to achieve accurate spatial positioning of various parts of the aircraft fuselage in the test building and improve efficiency.

Abstract:Aiming at the problem that the existing underwater wireless sensor network positioning algorithm has insufficient positioning accuracy and cannot adapt to the underwater changeable network topology, an underwater wireless sensor network positioning algorithm based on mobile beacon is proposed. Firstly, RSSI ranging positioning and DV-Hop algorithm are used to obtain the approximate distribution of unknown nodes. Secondly, taking the positioning coverage rate of unknown nodes as the objective function, the improved bald eagle search algorithm optimized by adaptive inertia weight and Cauchy-t disturbance strategy is used to iteratively solve the optimal position of beacon node movement. Finally, the beacon node moves to the optimal position and then the unknown node is repositioned. The simulation results show that compared with the comparison algorithm, the mobile beacon node localization algorithm can effectively improve the positioning accuracy of unknown nodes, and can also maintain high positioning accuracy and stable positioning effect when the network topology changes.

Abstract:As the variation law of frequency with time is the most important difference between different modulated signals, a radio modulation classification and recognition method combining Choi-Williams distribution and improved convolutional neural network model is proposed. In the signal preprocessing stage, in order to better retain the time-frequency characteristics of the signal, the Choi-Williams transform is introduced to transform the original time series signal into time-frequency image, and then the modulation signal classification problem is transformed into an image recognition problem. In the signal recognition stage, the convolutional neural network model is introduced with residual dense blocks and global average pooling layer to overcome the shortcomings of poor generalization ability and long training time of convolutional neural network model. Experimental results show that the proposed method can effectively solve the problem of gradient disappearance, and has the advantages of high recognition rate and strong generalization ability. Especially in the case of low SNR, the performance is even better. When the SNR is -4 dB, the classification accuracy of 8 kinds of signals can reach 100%.

Abstract:Aiming at the problem of insufficient wind turbine gearbox fault feature extraction and low fault diagnosis rate, a wind turbine gearbox fault diagnosis method based on RF feature optimization combined with WOA-ELM feature identification is proposed. Firstly, the wind turbine gearbox time domain, frequency domain, and time-frequency domain features are extracted to construct a multi-domain high-dimensional feature set. Secondly, the RF is used to rank the feature importance and extract 10-dimensional preferred features. Finally, the input weights and implied layer thresholds of the ELM model are optimally adjusted using WOA to achieve wind turbine gearbox fault classification and identification. The experimental results show that the average diagnosis rate of this method can reach 99.81%, and the diagnosis accuracy is higher than that of the comparison methods and the diagnosis time is the least, which can effectively diagnose the wind power gear box faults.

Abstract:Aiming at the problems of inaccurate features extracted and low prediction accuracy in the life prediction of rolling bearings,a method of remaining life prediction of rolling bearings based on hierarchical dispersion entropy (HDE) and gated recurrent unit (GRU) was proposed.Firstly,the time series of vibration signals were analyzed by hierarchical analysis,and the dispersion entropy of each node was calculated,and the dispersion entropy was reconstructed and fused to obtain the HDE.Secondly,correlation,monotonicity and robustness are combined to form a comprehensive index to verify the superiority of HDE.Finally,the training set and test set of HDE are divided,and the life prediction test is carried out by GRU network.The results show that the comprehensive index value of HDE is the best,and the prediction error of the proposed method HDE-GRU is 42.77%,39.57% and 20.24% lower than that of RMS-GRU, DE-GRU and MDE-GRU,respectively.It has the shortest running time and higher prediction accuracy,which provides practical value for rolling bearing health management.

Abstract:In view of the complex working environment of rolling bearings, the difficulty of extracting fault features from bearing vibration signals due to noise interference, and the low accuracy of traditional fault diagnosis algorithms, a rolling bearing fault diagnosis method based on complete ensemble empirical mode decomposition with adaptive noise analysis algorithm (CEEMDAN) combined with convolution neural network (CNN) and embedded long short-term memory neural network (LSTM) is proposed. Firstly, the original vibration signal of the bearing is decomposed by CEEMDAN algorithm to obtain the intrinsic mode function (IMF); Then the permutation entropy of the reconstructed signal is calculated and normalized as the eigenvector; Finally, the eigenvector is input into the deep learning model established by CNN-LSTM for diagnosis and recognition. The results show that the proposed method has faster fitting speed and higher accuracy, and the average fault diagnosis accuracy rate reaches 98.63%.

Abstract:Electrical Impedance Tomography was widely used in medical imaging, two-phase flow industrial inspection and special material inspection due to its non-invasive measurement characteristics, intuitive results visualization and convenient measurement methods. However, the inverse process of image reconstruction is inherently under-determined and ill-conditioned, resulting in some deviations in the results. An improved deep neural network based on Resnet34 is designed to solve the inverse problem of electrical impedance tomography for it. The training and test data sets were established, by setting the pixel point as the center in the field, the random radius and resistivity distribution change intensity in a small range, and the boundary voltage at each electrode in the case of 32 electrodes was simulated forwardly. The method was proved to be able to converge quickly, and can obtain better judgment performance compared with Gauss-Newton iteration method, total variation method and Tikhonov regularization algorithm after parameter adjustment and training.

Abstract:In view of the problems of PVC moisture content in the PVC drying section, such as nonlinearity, large lag, complex correlation with other variables and difficult to predict, a multiple attention mechanism and weight correction long-term and short-term memory network (LSTM) model are proposed for the prediction of PVC moisture content. In the encoder part: use the correlation between the input sequences related to water content to correct the variable weight of spatial attention mechanism training, so as to avoid the large weight difference between the input variables with strong correlation due to simple data training, and then the actual drying process is inconsistent. At the same time, due to the hysteresis of water content prediction, in order to reduce the loss of cell state information of LSTM unit in the eldest son time window, an information compensation mechanism is proposed to compensate the cell state information at the previous time. In the decoder part, we use the time attention mechanism to update the weight of the hidden layer state of the encoder, and remove the limitation of the fixed length vector on the performance of the model. Finally, the DCS data of the drying section of a chemical company were selected for verification. Compared with RNN, VA-LSTM and STA-LSTM, the correlation coefficient (R2) were increased by 571%, 122.6% and 82.6% respectively. The results showed that the model in this paper had certain advantages.

Abstract:Aiming at the problem of low search efficiency of F-RRT* algorithm in narrow environment and complex environment with multiple obstacles, a F-RRT* algorithm based on bidirectional search (BF-RRT*) is proposed. Based on the F-RRT* algorithm, firstly, a two-way search structure is adopted, the double tree is expanded from the starting point and the ending point in turn, and the greedy heuristic is used to guide the random tree growth. Secondly, the redundant points generated in the continuous expansion process are eliminated. A low-cost path is obtained, which effectively improves the planning speed; then a heuristic function is introduced, and the connection points are optimized to improve the overall quality of the path. Finally, the improved algorithm is compared based on MATLAB and Gazebo simulation platforms. The results show that compared with the original algorithm, the algorithm reduces the number of iterations by 63.5% on average and the planning time by more than 88.41% in different environments. Improved planning efficiency.

Abstract:The number of layers and the depth of the network are gradually increasing as the research on deep learning networks deepens and the accuracy of the network model improves, leading to an increase in computation. The lightweight, efficient and accurate network model becomes the key to research based on the need of deep learning model facial landmark detection for deployment on embedded devices. Therefore, an ultra-lightweight facial landmark detection network based on Ghost Model and Ghost Bottleneck is designed in this thesis to ensure the network accuracy while minimizing the network model size and reducing the computational effort. With a network width factor of 1X, the normalized mean error is reduced by 7% and the number of parameters is reduced by 36% compared to the best performing lightweight network model PFLD 1X; with a width factor of 0.25X, the proposed network model is only 420 KB in size, and the normalized mean error is reduced by 6.6% and the number of parameters is reduced by the average normalized error is reduced by 6.6% and the number of parameters is reduced by 25%.

Abstract:In order to solve the problem of long running time of single frame and single trajectory evaluation criteria existing in the optimal trajectory generation algorithm of driverless vehicles in urban road scenarios, a driverless vehicle trajectory planning method based on driving risk field to optimize sampling area was proposed. In this method, static and dynamic obstacle risk field models were established respectively through two-dimensional Gaussian distribution to quantify the target point of the sampling area on road. The sampling target points in the areas with low driving risks were selected by convolution to generate the optimal trajectory. The simulation results show that this optimization method only selects part of the sampling target points for trajectory generation in each planning period, which improves the running efficiency of the algorithm and makes the running time of each frame of the algorithm less than 0.1 s. The addition of driving risk field makes the sampling area of algorithm more consistent with the drivers′ behavior and habits, improves the degree of anthropomorphism of the algorithm planning results, and ensures the high driving efficiency of the driverless vehicle.

Abstract:In order to improve the delivery rate of automated pharmacies and realize the shortest delivery time of drugs with multiple delivery documents. Based on the structure of a hospital automated pharmacy delivery system, a time scheduling model of multiple delivery task documents is established. An improved hybrid whale optimization algorithm (H-WOA) is designed for this model. Firstly, multiple drug delivery documents are integrated and grouped, and the crossover and mutation of genetic algorithm are integrated to improve the diversity of population exploration. Secondly, the affinity determination principle of antibody and antibody of immune cloning algorithm is introduced to help construct the best execution sequence. Finally, the encirclement contraction and spiral update mechanism of the whale algorithm are combined to ensure that the population finally converges to the best whale position. The implementation results show that the efficiency of H-WOA optimization algorithm is improved by 6.11% and 18.11% respectively compared with the original algorithm and the algorithm without optimization. It has the same optimization effect on the scheduling optimization of 4 or 5 delivery documents. It is proved that the improved algorithm has good optimization ability.

Abstract:Aiming at the problems of long path length and more turning points of mobile robots in the flexible assembly process of circuit breaker, a path optimization method combining BAS algorithm and PSO algorithm is proposed. By combining the individual update method of beetles with group learning, the adaptive step size decay strategy and the dynamic weight change strategy are adopted to realize the optimization of global path planning. In order to verify the effectiveness of the BSO algorithm, three different test functions are used to compare the performance and the simulation map, and finally the algorithm is applied to the actual map through ROS. Experimental results show that compared with GA-PSO algorithm, AIW-PSO algorithm and BAS algorithm, the optimization efficiency of path length is increased by 7.7%, 14.8% and 12.5%, and the optimization efficiency of the number of turns is increased by 25%, 57.1% and 40%, respectively. In summary, the fusion algorithm proposed in this paper can effectively solve the efficiency problem in the assembly process and improve the efficiency of the flexible assembly line of the circuit breaker.

Abstract:This paper proposed a solder joint and wire bonding segmentation approach for detecting the bonding effect of RF chip based on AOI. According to the characteristics of solder joint and wire bonding segmentation task, this method improves the prior frame generation mechanism of the feature pyramid layer in Mask R-CNN. Also, it introduces a data enhancement method based on collision detection, which reduces manual annotation cost. The results show that the improved Mask R-CNN model can obtain the accurate segmentation positions of the solder joint and wire bonding in RF chips with mAP of 85.23% and mIoU of 71.27%. Meanwhile, this method speeds up to 1.168 per image, which basically meets the requirements for RF chip speed in production. Overall, the proposed method achieves high segmentation accuracy and meets the industrial production requirements for timeliness to a certain extent in the solder joint and wire bonding segmentation task.

Abstract:The artifact noise present in the CT image of a solid rocket engine will seriously affect the extraction of the initial combustion surface boundary and defects, and the extraction of defect information in the actual CT image is difficult.The algorithm that effectively removes CT image artifacts and automatically extracts the burn surface and defect data has important engineering practical value. Aiming at the problems of CT image de-artifacting and defect extraction, an IBM3D filtering algorithm is proposed, which uses the a priori information of edge detection to find similar blocks in the block matching stage. In addition, an adaptive Canny edge detection algorithm combined with seed eight connected labeling method is proposed to orderly separate the initial burning surface and defect data in the image. The experimental results show that the peak signal-to-noise ratio and structural similarity of the IBM3D algorithm are higher than other algorithms, and the fire surface defect information extracted by the adaptive edge detection algorithm is more complete than other algorithms.The CT image quality of solid rocket motor with defects is improved, and the initial burning surface and defect information are extracted accurately.

Abstract:The abnormal posture detection of industrial robots is an important means to ensure the safe operation of industrial robots. Aiming at the problems of low detection accuracy and insufficient timeliness of existing methods, a method for abnormal posture detection of six-axis industrial robots based on Kinect camera was proposed. The method uses the Kinect camera to collect the color image and depth image of the industrial robot, obtains the information of the joint axis of the industrial robot in the color image through the YOLOF target detection algorithm, converts the depth image into the corresponding three-dimensional coordinates, refers to the structural characteristics of the industrial robot, and constructs the robot joint vector. The angle feature is extracted, the attitude feature representation of the industrial robot is performed, and the attitude matching is performed based on the Euclidean distance and the cosine similarity to detect the abnormal attitude of the industrial robot. The method in this paper combines the three-dimensional information of the joint axis of the industrial robot to match the pose more accurately. A six-axis industrial robot working video dataset is constructed and abnormal posture detection is carried out. The experimental results show that the accuracy of the abnormal posture detection method of industrial robots in this paper is 96.6%, and the detection time of a single frame image is 43 ms, which meets the practical application requirements of robot safety monitoring.

Abstract:Pointer-type meters are widely applied into the petrochemical, industrial manufacturing, and the power sector of the tobacco industry. Given the low frequency of manual inspection and the hostile environment of part installed meters, the security risks exist in the procedures of the factories production, and personal safety of inspectors is difficult to be guaranteed. Based on the existing surveillance camera system of the procedures of the factories production, this paper proposes an automatic identification method of reading of pointer-type meters based on YOLO V3 object detection and DeepLab V3+image segmentation techniques. YOLO V3 is introduced to detect and incise the sub-images of the gauge dial. According to the characteristics of the meter images and actual needs, this paper improves the structure of DeepLab V3+and add corrosion treatment. So the scale lines and pointers of the sub-images are located effectively. The gauge range is extracted from the sub-images by OCR technology. And in combination with the relative positional relationship of the scale lines and pointers, the identification of meters reading can be computed. It is proved by experiment that the average relative error of reading of the meter image identified is 2.17%, and the proposed method can meet the applications requirements.

Abstract:Tire defect detection has important reference significance for tire grading, and it is particularly important to study the high performance tire anomaly detection method. Based on reinforcement learning algorithm, an automatic image classification algorithm based on abnormal loss value is proposed. This method firstly by a large number of positive samples input to reduce the loss value of data after gradient update, with a small amount of the loss of the abnormal sample input values form the obvious difference, introducing neural sampler, enlarge abnormal loss of contour difference between samples and the positive samples and provide training to sVAE batch, then put the training result as input of abnormal classifier, Finally, the classification and location of anomaly detection are completed. Through comparative study, it is found that the anomaly detection algorithm proposed in this paper is obviously superior to other traditional image anomaly detection methods in tire defect sample sets.

Abstract:The appearance characteristics of different models are highly similar and different from those of the same model, which poses a great challenge to the feature extraction network.Existing vehicle type classification schemes only rely on vehicle appearance feature recognition, and the overall recognition accuracy is not high.Therefore,firstly, this paper designs a multi level attention mechanism in backbone network to improve the ability of main network to extract and recognize vehicle features. Secondly, according to the changes of vehicle appearance characteristics at different vehicle locations in the bayonet environment, a feature fusion structure of vehicle location and appearance features is proposed, which extracts the composite image features of the fusion location, reduces the feature distance within the class, and enhances the expressiveness and robustness of the features extracted by the main network.Finally, based on the analysis of the attention heat map of difficult samples, the attention area of difficult samples is intervened to make the network focus on the local area of small differences between vehicles. The experimental results show that the overall performance of the vehicle type recognition method proposed in this paper is significantly improved than the existing scheme.

Abstract:For the lightning overvoltage on-line monitoring device of the transmission line, it is very important to accurately identify the lightning current signal. At present, whether it is hardware triggering or software triggering, there are triggering problems such as false triggering and leakage triggering, and the current monitoring system has obvious disadvantages in terms of economy and convenience. This paper proposes an embedded lightning strike online monitoring system, which can realize efficient and convenient real-time monitoring of transmission lines, and greatly improve the convenience and economy of lightning strike monitoring devices. At the same time, a triggering algorithm of first detection and identification is proposed to accurately identify the lightning current transient signal. First, the lightning current signal (bad data) is quickly detected by the method of amplitude difference, then the bad data is accurately identified by the method of period difference. Through analysis experiments, it is proved that the method can reduce the number of false triggers and missed triggers of the monitoring system. And through laboratory tests, it is proved that the acquisition accuracy of the monitoring system is 98.49%, which meets the needs of real-time monitoring and provides an economical and convenient monitoring system for online monitoring of transmission lines.

Abstract:Aiming at the problems of cumbersome pre-processing and high false alarm rate of pipeline leakage signals under multiple working conditions,an ensemble empirical mode decomposition (EEMD) combined with improved convolutional neural network (ICNN) is proposed as a leakage identification model. The proposed identification method uses EEMD to decompose the leak signal into several intrinsic modal components (IMFs) with steady-state performance, and the noise dominant vectors are divided and removed by correlation coefficients to achieve signal reconstruction. A series of indicator features of the reconstructed signal are extracted as the input of the ICNN mode for feature extraction to achieve the multi-condition classification of pipeline. The batch normalization layer was added by ICNN between each convolutional layer and the pooling layer to accelerate the network training. In accordance with the results, it indicates that the proposed model can quickly and accurately identify pump shutdown, valve adjustment, leakage and normal operating conditions, which can reach 98.25% of the average recognition accuracy under less training data.This technique significantly raises the accuracy of recognition when compared to the unimproved CNN and SVM classification recognition models.

Abstract:Aiming at the problem that the interference of components with low correlation with fault features in the bearing vibration signal reduces the fault diagnosis accuracy, a bearing fault diagnosis method based on improved spectral kurtosis map and multi-dimensional fusion CNN is proposed. To improve the correlation between vibration signals and fault features and reduce interference components, an improved spectral kurtosis graph model was constructed based on DTCWPT to enhance the expression of multi-resolution differential fault features. Then, considering the rich feature dimension, a multi-dimensional fusion CNN model is constructed, and the original signal and the improved spectral kurtosis map are used as input together. The experimental results show that the method can extract different fault features in the vibration signals of various types of bearings, and can accurately identify bearing faults under multiple working conditions, with good diagnostic accuracy.

Abstract:In order to realize the measurement of the horizontal ultra-high acceleration generated by the impact of penetration weapon on hard target, an in-plane high-range piezoresistive acceleration sensor with a single axis is proposed. The sensor has a range of 150 000 gn and consists of two support beams, proof mass and four microbeams. When axial acceleration is applied to the sensor, the micro-beams on both sides of the support beam are stretched and compressed respectively, and the piezoresistive signal is detected by a Wheatstone full bridge. The natural frequency, maximum equivalent stress and geometric structure parameters of the accelerometer are used as multi-objective functions to optimize the accelerometer with OSF sampling and Kriging proxy model to obtain a sensitive structure with high resonant frequency, high sensitivity and low cross-sensitivity. The finite element simulation and electrical simulation results show that the resonant frequency of the sensor is 0.84 MHz, the y-axis sensitivity is 0.594 μV/gn, the nonlinearity is 3.26%, and the cross-sensitivity of both x-axis and z-axis is less than 1%.