Abstract:The characteristics of Moir􀆧 signal and the positioning accuracy of turntable is sensitive to eccentricity and inclination. A research on the effect of eccentricity and inclination on the Moir􀆧 signal of grating of turntable in this paper. The working principle of turntable angular measurement system is described and the principle of separation method of Moir􀆧 signal characteristic is elaborated in detail. The error sources of Moir􀆧 signal are analyzed, the effect model of eccentricity and inclination are established respectively, and the relationship among the harmonic order of Moir􀆧 signal characteristic, eccentricity and inclination is deduced. The separation method of Moir􀆧 signal characteristic is implemented using a laboratory-made field-programmable gate array circuit. The result of experiment performed on the eccentricity and inclination test platform confirmed that eccentricity and inclination have linear relationship with first and second harmonic of Moir􀆧 signal characteristic, and nonlinear error is 3. 88% and 2. 08%.

Abstract:Facing the urgent need of dynamic measuring force sensor for heavy load, multi-dimensional, high-frequency and strong impact load in the field of high-performance bearing rolling element forming quality monitoring, a new structural form and measurement principle of piezoelectric film three dimensional heavy load force sensor are proposed. The sensing element of the sensor is composed of upper, middle and lower piezoelectric films. The change of three-dimensional force causes the change of charge of piezoelectric films cut in different regions of each group, so as to realize the calculation of three-dimensional force. The sensor support shell plays the role of load sharing, so as to realize heavy load measurement. The quasi-static calibration experiment of the sensor prototype is carried out. The experimental results show that the sensor has the advantages of good linearity. The nonlinear error is less than 2%. It can meet the dynamic measurement requirement under heavy load. Therefore, it provides an important reference basis for the design, development and application of piezoelectric thin film force sensor in the field of heavy load multi-dimensional load dynamic measurement.

Abstract:The rotor current signal of maglev gyro is highly sensitive to environmental changes, therefore, noise will inevitably be introduced in the process of signal sampling. To solve this problem, an algorithm based on local mean decomposition (LMD) and fusion of Hausdorff distance and threshold denoising (TD) is proposed to reduce noise interference. Firstly, the original signal is decomposed into several PF components and a margin. Then, the noise and signal components are determined according to the Hausdorff distance between each pf component and the original signal. Then, the noise component is denoised by threshold. Finally, the noise components, signal components and margin after thresholding are superimposed to obtain the reconstructed signal, so as to realize the reconstruction of the gyroscope rotor current signal Noise reduction. The simulation results show that the signal-to-noise ratio of the reconstructed signal is 12. 86 db higher than that of the original signal, and the root mean square error is 9. 25×10 -6 A lower than that of the original signal. The de-noising results of measured signals show that the filtering gains of the de-noising algorithm for four wire sides are 40. 0%, 93. 5%, 30. 8% and 50. 0% respectively.

Abstract:The end-effect of which is caused by the limited axial size of radial HTS magnetic bearing will affect axial levitation behavior of radial HTS magnetic bearing. A 2-D finite element modeling method based on H-formulation is proposed to analyze the end-effect of radial HTS magnetic bearing. Firstly, the proposed finite element modeling method is verified by experimental tests. Then, the finite element models of radial HTS magnetic bearings with different sizes of stator and rotor are established respectively. Finally, it is obtained that the relationship between axial levitation force and displacement of superconducting stator of different size and permanent magnet rotor with different number of poles. The results show that due to the end effect, the axial levitation behavior deteriorates with the increase of the number of superconducting stator bulks. Moreover, with the increase of the number of permanent magnet rotor poles, the maximum axial levitation force shows a trend from rise to decline and it eventually approaches the maximum axial levitation force of permanent magnet rotor for infinite length.

Abstract:The zero error is an important parameter of the moisture balance, which directly affects the weighting accuracy of measuring the moisture. A method for compensating the zero error of moisture balance based on long-short term memory ( LSTM) network was proposed. The mechanism of the zero error of moisture balance was analyzed, and the prediction model of zero error based on the double layer LSTM was established by using historical zero error data. This proposed model was used to compensate the zero error of the moisture balance with some loads and without load respectively, where the measuring range of the moisture balance is 200 g and its resolution is 1 mg. The experimental results show that the maximum error of the compensated moisture balance is about 6 mg, which is much smaller than before. The experiments have verified the effectiveness of this method.

Abstract:As the medium of oil and gas transportation, the stress concentration in the defects of the metal pipeline will cause safety hazards. In order to realize the non-contact quantitative detection of metal pipeline defects, a magnetic memory detection method has been studied. Adopt the magnetic anomaly gradient matrix to locate the defects with stress concentration; use translation invariant wavelet denoising (TI) and feature extraction for signal processing. Sparrow search algorithm (SSA) optimizes the BP neural network to achieve defect size inversion. Experiments show that compared with wavelet threshold denoising, the translational invariant wavelet denoising can increase the signal-to-noise ratio by 1. 56% and reduce the mean square error by 4. 87%; the mean square error of SSA_BP neural network inversion is 67. 2% lower than that of BP neural network. The detection method can detect pipeline defects in real time in the lift-off state and invert the defect size.

Abstract:This paper takes the planetary gear transmission system with high incidence of faults as the object, a fault diagnosis method based on variational mode decomposition (VMD) and particle swarm optimization (PSO) to optimize support vector machine (SVM) is presented. Firstly, the signal is decomposed by VMD, the decomposed components are processed by improved wavelet method, and the processed components are reconstructed to highlight the signal. The weak information of SVM is extracted. Then, the sample entropy and root mean square error of the processed vibration signal are extracted, and the input matrix is formed. Finally, PSO is introduced to optimize the key parameters of SVM, and the extracted eigenvectors are input into PSO-SVM for training and recognition. The method is applied to the planetary gear crack fault, the solar gear tooth fault and the planetary gear bearing fault signal obtained by the planetary transmission test platform. The effectiveness of the method is verified by multi-dimensional comparison.

Abstract:Oil pollutants are one of the main factors that endanger the ecological environment. The accurate identification and measurement of mixed oil pollutants is of great significance for environmental monitoring. In view of the serious overlap of spectral components of mixed oil pollutants, it is difficult to distinguish them by chemical methods. In this study, three-dimensional fluorescence spectrum combined with alternating residuals trilinearization (ART) algorithm is adopted to detect oil pollutants. ART algorithm is an improvement of parallel factor algorithm without presetting factor number. A three-dimensional fluorescence spectrum detection system was built to study the three-dimensional fluorescence spectrum characteristics of 0 # diesel oil, 97 # gasoline and kerosene. 30 groups of SDS micelle solutions of 0 # diesel oil, 97 # gasoline and kerosene were configured, and three-dimensional fluorescence spectral data were obtained. The three-dimensional fluorescence spectrum data of mixed oil pollutants after scattering elimination were analyzed by ART algorithm. The experimental results show that the ART algorithm is effective for 0# diesel, 97# gasoline and kerosene. The average recovery rates of gasoline and kerosene are 103. 6%, 97. 32%, and 99. 86%, respectively. The spectra of the three components obtained by ART algorithm are in good agreement with the real spectrum. Both qualitative and quantitative analysis show that the threedimensional fluorescence spectrum combined with ART algorithm is an effective method for the detection of oil pollutants.

Abstract:Aiming at the non-stationary characteristics of the wireless transmission scene of vehicle-to-vehicle (V2V) communication system and the non-independent characteristics of elevation angles (EAs) and azimuth angles (AAs) in the three-dimensional( 3D) wireless transmission, a novel 3D regular-shaped geometry-based stochastic reference model and the corresponding simulation model for vehicle-to-vehicle (V2V) channels were proposed. In order to describe the non-stationary of the channel, the multiple-time-varying parameters caused by the transceiver vehicle moving in any direction and speed have been added in the proposed model. Some of the statistical characteristics of the model are derived for the case that the EAs and AAs Angle spectrum obey the joint distribution of von mises fisher ( VMF), and the effects of vehicle driving environment, non-stationarities and the driving state at the receiving and transmitting terminals on the space-time correlation function of the channel are investigated in depth. Simulation results show that the proposed channel model can capture the influence of vehicle driving direction changes on the space-time correlation of the channel, the theoretical value of Doppler power spectral density ( DPSD) is basically consistent with the measured value, and the statistical characteristics of the simulation model and the reference model are highly fitted, it demonstrates not only the utility of simulation models but also the correctness of the theoretical derivations and simulations.

Abstract:At present, the spatial target position detection system based on optical measurement method mainly adopts monocular vision and binocular stereo vision, of which binocular vision benefits from the use of bionic parallax principle, and the spatial relative position of the target can be calculated by optical triangulation, which is widely used. In addition, in view of the complicated image processing process of traditional CCD and CMOS vision cameras, a detection method that combines the principle of binocular vision and PSD sensor to apply to the position of the light spot in space is proposed. The image processing is converted into the signal processing process of the photodetector, and a specific position detection system of the cooperative light spot is built. Experimental results show that the PSD visual phase X and Y direction coordinate fluctuations are within ± 130 μm, and the detection system accurately restores the motion trajectory of the spatial light point in the 160 mm×160 mm×200 mm cubic space at a distance of 1 000 mm, and its positioning error is 4. 35 mm, with each 50 mm increase in depth, the error increases by about 27%, and the average error of the system is 7. 30 mm, with good stability.

Abstract:Fault diagnosis for gearbox of wind turbine plays an important role in the normal operation of WT. Current studies commonly focus on diagnosis of fault types, nevertheless, in addition to identifying the fault type, the severity of the fault is also instructive for maintenance and repair for wind turbine. Thus, a novel optimized stacked diagnosis structure (OSDS) is proposed for identification of fault type and severity. Compressed sensing is adopted to implement compressed sampling of original vibration signals. Then, compressed samples are input into first and second layer deep belief networks ( DBNs) for identification of fault type and severity, separately. In addition, every single DBN in the OSDS is optimized with chaotic quantum particle swarm optimization ( CQPSO) algorithm. Comparison experiments based on bench mark gearbox fault data and working planetary gearbox show that the fault type diagnosis accuracy of this method reaches 99. 24% and 97. 21%, while the fault severity accuracy reaches 99. 06%. Meanwhile, the testing times are only 1. 493 and 2. 176 s.

Abstract:Power quality data compression is an important step in the detection and identification of power quality problems. Its essence is the process of exploring power quality sparse features. Aiming at the problem that the matching pursuit algorithm commonly used in sparse decomposition cannot meet the real-time requirement of power signal analysis because of its high computational complexity and long time-consuming in matching the best atom, a lightning search matching and tracing algorithm is proposed by applying the lightning search algorithm with high convergence precision, fast convergence speed and strong global search ability to search the best atom. The proposed algorithm is used to decompose the power quality signal into atoms and extract the characteristic parameters of power quality, and the extracted parameters are used as the compressed power quality data, realization of power quality data compression. The experimental results show that the time of the proposed algorithm to match the best atom is reduced to 1 / 98 of the original algorithm, and the power quality data compression method based on the proposed algorithm meets the real-time requirement of power signal analysis when the best atom is matched, with high compression ratio and low reconstruction error, the performance of data compression is improved.

Abstract:Aiming at the problem that the classification accuracy of nonlinear support vector machine is susceptible to kernel function, a multi-scale kernel support vector machine (MSK-SVM) classification model is proposed and applied to rolling bearing fault diagnosis. In this model, Morlet, Marr and DOG wavelet kernel functions are introduced on the basis of Polynomial, Gaussian and Sigmoid kernels. Using the global and local characteristics of various kernel functions, as well as the characteristics that kernel functions with different scale parameters have distinct influence range, kernel functions with different characteristics and scale parameters are combined as multiscale kernel. Based on the gradient descent method, the weights of multi-scale kernel function are adaptively determined, and the MSK-SVM rolling bearing fault diagnosis models are obtained. In order to illustrate the effectiveness of the algorithm, the rolling bearing fault data set and life cycle data set are selected for experimental verification, respectively. The classification performance of MSKSVM models based on different characteristic kernel functions and the same characteristic kernel function are analyzed. The results show that the proposed algorithm can achieve higher classification accuracy and better generalization ability than the traditional single kernel SVM.

Abstract:In order to solve the problem that traditional radar breathing identification relies on artificial predefined features, an ultrawideband (UWB) radar identification algorithm based on breath sample space (BSS) is proposed. The algorithm uses singular value decomposition (SVD) to filter out the clutter in the UWB radar human respiratory echo; the target cross-range respiratory signal is constructed as a BSS sequence containing time-distance information according to the echo; the convolutional neural network (CNN) is used to model the BSS to obtain the target classification results. In the indoor scene experiment, the identification accuracy of the four persons was 84. 64%. The comparison results show that the proposed algorithm has a good ability to distinguish the unique breathing characteristics of different individuals.

Abstract:Aiming at the problem that the influence of PEMFC characteristics on the life prediction method of the proton exchange membrane fuel cell (PEMFC) is unknown and the low prediction accuracy of the model, a PEMFC remaining life prediction method based on XGBoost-RFECV algorithm and LSTM neural network is proposed. First of all, the PEMFC original data is reconstructed and smoothed by equal interval sampling and SG convolution smoothing method, which effectively retains the original data degradation trend. Then the XGBoost-RFECV algorithm is used to calculate the importance of different PEMFC features, and the 10 PEMFC features with the smallest mean square error of average cross-validation are selected to form the optimal feature subset. Finally, the optimal feature subset is input into the constructed two-layer LSTM neural network to realize the remaining life prediction of PEMFC. The experimental results show that the average absolute error and root mean square error of the method are 0. 001 9 and 0. 002 5, respectively, and the coefficient of determination R 2 is 0. 974. Compared with the XGBoost-RNN, XGBoost-LSTM and XGBoost-RFECV-RNN model, the prediction accuracy is higher and it can effectively predict the remaining life of PEMFC.

Abstract:In order to solve the problem of long time and high cost of traditional detection methods for falling block in railway tunnel lining. Based on the acoustic signal recognition technology, this paper proposes a sound detection method for lining falling block in railway tunnels based on GA-SVM. After extracting the MFCC characteristic coefficient matrix of the lining falling block and other event sounds in the railway tunnel, this method uses the optimization ability of genetic algorithm to optimize the two parameters C and σ that affect the accuracy of the prediction model in support vector machine, so as to construct the railway tunnel lining dropped block detection model. The test results show that compared with the traditional SVM model and the PSO-SVM model, the GA-SVM model can more accurately detect the size of fallen lining blocks on the basis of a small number of training samples, and the detection accuracy reaches 96. 67%, which verifies the feasibility of the application of acoustic signal recognition technology in the detection of lining falling block of railway tunnels

Abstract:In order to meet the requirements of high acoustic source level emission and multi-channel low noise acquisition in multibeam sonar system, and to improve the scalability of transmitter and receiver, this paper introduces the design method of a high voltage pulse transmitter module and a multi-channel data acquisition module with TVG function, and uses FPGA as the logic control chip to plan the trigger timing sequence of the transmitter module and the acquisition module. According to the transmission bandwidth of Gigabit Ethernet, the data format of IP packet is designed, the data transmission path is analyzed, and the real-time data acquisition is realized. The experimental test shows that at the operating frequency of 100 kHz, the sound source level of the transmitter module reaches 200 dB, the background noise of the acquisition module is less than 4 μVrms, the amplitude consistency deviation is -6. 94 dB, and the phase consistency deviation is 0. 25°.

Abstract:In order to meet the requirements of miniaturization and high isolation MIMO antenna in wireless communication system, a compact UWB multiple-input multiple-output (MIMO) antenna with high isolation between units is proposed. The antenna size is only 41 mm×25 mm×1. 6 mm. The compact UWB MIMO antenna using a spanner microstrip line to widen bandwidth, Serrated grounding structure and comb-shaped electromagnetic band gap ( EBG) structure on the ground plane of the antenna helps to obtain higher isolation, introducing C-shaped branches on the antenna discards worldwide interoperability for X-Band satellite downlink communication band. The optimization of various parameters proves that the proposed UWB-MIMO antenna satisfies the port reflection coefficient S11 < -10 dB in the 3. 1~ 12. 0 GHz band, and the notch band is 7. 0 ~ 7. 9 GHz, the isolation is greater than 20 dB in the entire working bandwidth. The UWB-MIMO antenna designed in this paper has good radiation characteristics, stable gain and low envelope correlation coefficient (ECC<0. 006), which is suitable for UWB-MIMO system applications.

Abstract:Given the existing signal denoising or reconstruction methods can not completely remove the noise, and the time-frequency representation has the problem of energy ambiguity, a method for rolling bearing fault diagnosis based on element analysis was proposed. Firstly, the proposed method constructs an elemental model to characterize the signal, then the Morse wavelet transform is applied to the elemental model and the impact point of the signal is calculated from the wavelet transform to obtain the characteristic defect frequency of the signal. Based on a small number of solitary points in the time or scale plane of the wavelet transform, the method is used to reconstruct the signal. In this paper, a set of simulated signal data and two sets of experimental data are used to estimate the performance of the method and compare it with other signal reconstruction methods and time-frequency analysis methods. The results demonstrate that the proposed method has a good performance in the identification and reconstruction of rolling bearing fault signals.

Abstract:Leakage current is an important indicator to characterize the performance of GIL internal insulator. In practical engineering, the acquisition of leakage current signal is often interfered by narrow-band signal and white noise signal, which affects the accurate evaluation of insulator performance. The common methods of leakage current denoising rely on empirical parameters and manual settings. In order to solve the above problems, singular value decomposition ( SVD) is improved by using singular value curvature spectrum to remove narrowband signal interference. Then, the positive and negative white noise groups are introduced, and the leakage current signal with white noise is decomposed by empirical mode decomposition ( EMD). In the decomposition process, the modal component is denoised, and the final modal component is the noiseless leakage current signal. The results of signal simulation and field measurement show that the proposed method can effectively denoise the leakage current of GIL insulator.

Abstract:Single pixel imaging is the key technology to solve the problem of imaging in non-visible wavelengths because it only needs a single point detector without spatial resolution to obtain the spatial information of the target object and reconstruct the image of the object. The development and application of this technology in high resolution space have the problems of large data collection and slow imaging. Therefore, this paper designs a Fourier single-pixel microscopy super resolution imaging system to improve the efficiency and quality of existing single-pixel microscopy imaging. Firstly, a single-pixel microscopic super resolution imaging system was built, and a deep learning-based super resolution model was introduced to improve its imaging resolution, so as to quickly obtain high-resolution object reconstruction images. Experimental results show that under the same light intensity signal acquisition time, the imaging resolution can be improved nearly 9 times, and its peak signal-to-noise ratio reaches 28 dB, which effectively improves the imaging efficiency of single pixel microscopic imaging under the condition of high resolution.

Abstract:Aiming at the forbidden state problem of Petri nets with uncontrollable transitions, a controller synthesis method based on integer linear programming is proposed, which is suitable for any ordinary Petri nets model. Firstly, according to the structural characteristics of Petri nets, a set of constraint conditions that all uncontrollable transitions should satisfy is constructed. Secondly, the given linear constraints are converted into admissible linear constraints by solving the integer linear programming problem. Finally, the invariant method of the library is used to design the controller and integrate the converted constraints into the Petri net. The experimental results show that the proposed method is simple and efficient, and can be used as a reference for the forbidden state monitoring in the actual automatic manufacturing system.

Abstract:Since the catenary of high-speed railway will produce pantograph arc, it is harmful to pantograph system, in order to reduce pantograph damage. A current time series coding technology, namely, the Gram angle summation / differential field (GASF/ GADF) is proposed. Because the current signals of different current receiving states are different, the images formed by their time series coding are also different, which makes computer vision technology can be used for time series classification to identify pantograph arcs. A total of five groups of pantograph receiving experiments were carried out under different conditions to measure the current data in pantograph system under different conditions, and the current data obtained from pantograph experiments were divided into normal receiving state and arc receiving state. By constructing a neural network and extracting the arc current signal, it visually demonstrates the abstract feature extraction of the CNN from the arch-net arc data in the form of a Gram angle field (GAF) image. The experimental results show that the method in this paper can accurately identify pantograph and network arcs under different conditions, avoiding the problem of video image background changes, and provides an idea for pantograph and network arc fault identification.

Abstract:A non-contact measurement system based on high-speed camera is designed for deformation measurement of flexible hydrofoil in oscillating hydrofoil tidal current power generation device, which can continuously collect images and calculate hydrofoil deformation without affecting hydrofoil hydrodynamic characteristics. The coordinate transformation model of image coordinate system and hydrofoil world coordinate system is established, and the calibration method of relevant parameters and the formula of deformation calculation are given. A new method of edge detection of the upstream surface of hydrofoil based on Canny operator is proposed and the polynomial fitting formula of the boundary is given. In this way the deformation of any point on the edge of the upstream surface of the wing can be calculated. Different experimental conditions are set to verify the results. The total length error of the identified hydrofoil is less than 2%, and the deformation is in line with the actual law. The experimental results show that the method can effectively measure the deformation of any point on the edge of the upstream surface of the hydrofoil, and has the advantages of simple field arrangement and no influence on the hydrodynamic characteristics.

Abstract:State of health (SOH) prediction, as one of the key functions of lithium ion battery management system (BMS), is of great significance to ensure the safe and reliable operation of batteries and reduce the maintenance cost of battery system. In order to improve the prediction accuracy of lithium battery SOH, a SOH prediction method based on improved ant-lion optimization algorithm and support vector regression (IALO-SVR) is proposed. Firstly, the characteristic factors related to battery capacity are extracted from the battery charging data, and the correlation analysis is carried out. The three features with high correlation are selected as the model feature inputs, and then the sample data is imported. The key parameters of SVR model are optimized by the IALO algorithm, and the final prediction model is established. Compared with the existing GA-SVR and IPSO-SVR, the results show that IALO-SVR method NASA has higher prediction accuracy and fitting degree, and the prediction error is basically kept within 1%, which verifies the feasibility of the prediction method.

Abstract:In order to eliminate the influence of grounding mode, fault type and fault location on the accuracy of ground fault diagnosis in low current system. By analyzing the zero sequence current of all kinds of single-phase ground faults in this system, a single-phase ground fault detection method was proposed on the basis of the improved Hilbert-Huang transform (HHT) and Extreme learning machine (ELM). This method firstly used wavelet transform (WT) for multiband signal. Then HHT was performed on the characteristic signal that was selected by the charging and discharging characteristics of the ground capacitance to obtain the instantaneous energy of the zero sequence current of each line. Finally, gray wolf optimization (GWO) and particle swarm optimization (PSO) were used to optimize the ELM model to obtain the GWO-PSO-ELM model with fault type recognition and line selection functions. A fault detection system based on digital fault indicator (DFI) acquisition platform and master station data processor is designed. The test results show that this method can accurately judge the fault type and complete line selection, and the accuracy reaches more than 90%.

Abstract:Vibration signals of fault rolling bearings contain interference signals, which makes it difficult to extract fault information accurately. In this paper, a fault diagnosis method for fault rolling bearings was proposed based on singular value decomposition (SVD), variational mode decomposition ( VMD) and support vector machine ( SVM). First, the singular value decomposition was used to process the signal, and the effective order of the reconstructed matrix was determined according to the kurtosis difference spectrum of the singular value. Then, the reconstructed signal was reconstructed according to the effective order, and the VMD decomposition was performed on the reconstructed signal. The number of the decomposed intrinsic mode function ( IMF) components was determined according to the above effective order. From the IMF component of the decomposed to extract the fault characteristic parameters, as the input parameters of support vector machine ( SVM) to fault diagnosis. Finally validated bearing tester adopts Hefei university of technology, and directly into the decomposition of VMD and band-pass filter signal denoising based fault diagnosis method is compared, the results show that the method can effectively identify roller bearing fault type and can also be used for rolling bearing fault diagnosis.

Abstract:In order to improve the energy consumption management efficiency of the embedded real-time system and reduce the impact of traditional dynamic voltage scaling technology on system stability, a dynamic energy consumption optimization scheme supported by whale algorithm based on historical cognition is proposed. Firstly, a nonlinear dynamic convergence factor control strategy is proposed, which can effectively accelerate the convergence speed of the algorithm. Secondly, using the historical optimum solutions as interference factors, a hybrid guided strategy is designed in the constriction and envelopment mechanism to balance the local development and global search capability of the algorithm. Finally, the frequency characteristics of the processor can be changed in real time according to the dynamic voltage scaling technology, the tasks 10, 30 and 50 are optimized by the algorithm, so as to verify the effectiveness of the improved algorithm.

Abstract:In order to solve the problem of difficult detection due to the existence of dense workpiece with high similarity and disorderly arrangement, an attention mechanism is proposed to optimize RetinaNet’ s dense workpiece detection method. Firstly, the attention mechanism is introduced into the RetinaNet backbone feature extraction network to reduce the influence of interfering objects on the detection effect and improve the feature extraction ability of the neural network. then a new predictive box is constructed using Soft-NMS to improve the overlap localization accuracy. Finally, the dataset is trained by transfer learning method to improve the model training efficiency. The method effectiveness is verified on the produced dense workpiece dataset; Experimental results show that the detection accuracy of the improved method reaches 98. 11%, which is 2. 59% higher in comparison with that before the improvement. The detection speed of a single picture is up to 0. 026s. The proposed method can meet the purpose of accurate detection of workpiece in actual industrial production process, which can reduce the rate of missed and false detection and assure the speed simultaneously.