Abstract:In order to improve the accuracy of arm fatigue model recognition, this study introduces time-frequency domain, nonlinearity and parametric model features based on common time-domain and frequency-domain features, and extracts 3-channel surface EMG signals to form features set. Feature dimensionality reduction is generally divided into feature extraction and feature selection. This research uses principal component analysis ( PCA) in feature extraction, kernel principal component analysis ( KPCA) and mutual information (MI) measurement methods in feature selection. Feature dimensionality reduction, using support vector machine ( SVM) and K-nearest neighbor (KNN) as the classifier; three dimensionality reduction methods and different combinations of SVM and KNN constitute a fatigue classification model. Results show that the correct recognition rate of KPCA and SVM is 99%, which is higher than other combination algorithms.

Abstract:In order to solve the shortcomings as halo and detail blur in the current infrared image enhancement methods, an infrared image enhancement algorithm using exponential homomorphic filter coupled with detail sharpening rule is proposed in this paper. Based on the homomorphic filtering method, an exponential homomorphic function is constructed by the distance between the pixel and the center pixel in the frequency domain to form an exponential homomorphic filter, which is used to denoise the image and enhance the contrast of the image. Based on the phase congruence method and the adaptive high lifting filtering method, detail sharpening rules are constructed. In the rule of detail sharpening, the phase consistency method is used to extract the detail features of the image accurately through Fourier transform. Then, the high lifting filtering method is introduced, which uses the mean value of the image to construct an adaptive sharpening factor to form an adaptive high lifting filtering method. The extracted image details are sharpened to enhance the details of the image. The experimental results show that the image enhanced by the proposed method has better contrast than the current infrared image enhancement algorithm, and the edge of the image is clearer, which shows that the proposed method has better enhancement effect.

Abstract:Aiming at the problem of low accuracy and slow speed of electronic components detection by manufacturing robots in the electronics industry, an electronic component detection method based on improved YOLOv4 is proposed. The network structure was improved by using depth-separable convolution instead of the traditional convolution in PAN networks to improve the detection speed. An inverse residual structure with a linear bottleneck was used instead of the CSP darknet53 backbone network to reduce the model parameters and further improve the detection efficiency. An attention mechanism was added before the YOLO head of the detection network to improve the detection accuracy. A data set of electronic components was established to simulate the industrial environment with conveyor belt and the data was enhanced. Compared with the original algorithm, the accuracy (mAP) is increased by 1. 31%, the speed is increased by 16. 34 fps, and the weight size is reduced from 245 to 41. 20 MB. The research can provide technical reference for the development of manufacturing robots in the electronics industry.

Abstract:Aiming at the problems of difficulty in detecting small targets in UAV inspection images, obstacles blocking targets and imbalance between positive and negative samples, a multi-target detection method based on improved Cascade R-CNN was proposed for transmission lines. The feature extraction network of the Cascade R-CNN was improved. Based on the basic network structure of ResNet101, a new 6-layer feature pyramid network structure was designed to achieve feature fusion, improving the detection ability of small targets and overlapping targets. The Gaussian Soft-NMS method was introduced to reduce the missed detection rate of the target with occlusion. The Focal loss was used to improve the loss function, alleviating the impact of imbalance between positive and negative samples on detection accuracy. During the training process, the data set was expanded based on data enhancement methods such as adding noise, brightness transformation and scaling, which improved the generalization performance of the training model. Experimental results show that the improved model can simultaneously detect three types of porcelain insulators, porcelain insulator defects, interphase rods, anti-vibration hammers and bird’ s nests under complex backgrounds. The average accuracy ( mAP) reaches 94. 1%, which provides a new idea for the intelligent inspection of transmission lines.

Abstract:Correlation Interferometer is a passive direction finding algorithm, which has the characteristics of low computational complexity, strong anti-interference ability and sample’s environment dependence. In order to reduce the dependence of the sample on the environment and improve the accuracy in actual working environment, an improved algorithm based on theoretical samples is proposed in this paper. By searching the local peak value of correlation matrix, establishing theoretical correlation peak system, calculating correlation peak evaluation coefficient ( including variance, kurtosis, skewness), and quadratic correlation matching, the algorithm can accurately detect the direction of arrival signal with theoretical samples. Field experiments show that in the frequency band 340 MHz~ 3 GHz and the signal-to-noise ratio 29 ~ 60 dB, the accuracy of the improved algorithm is 0. 8° higher than that of the traditional sampling algorithm in the urban environment. At the same time, the overall accuracy difference of many experiments is less than 1°. The improved algorithm has high computational complexity, and the cost of time is 3~ 4 times that of the traditional algorithm, and it has good practical value in scenarios without high timeliness requirements.

Abstract:In order to improve the recognition rate of the detection drone, the flight sound of the drone is analyzed in the time and frequency domain. For the sound characteristics of the drone, so as to better characterize the UAV sound signal, the Mel cepstrum coefficient (MFCC) and the inverted Mel cepstrum coefficient (IMFCC) are combined. The new feature parameters are used for feature dimensionality reduction based on Fisher criterion, construct the UAV “voiceprint library”, optimize the parameters in the support vector machine (SVM) through the gray wolf optimization algorithm (GWO), and establish the UAV audio Classification model. Experimental results show that the new feature parameters can make up for the low resolution of a single feature in the entire sound and audio domain. The GWO-SVM audio classification model can achieve drone detection within a distance of 50m, which has significant advantages over traditional detection methods.

Abstract:In order to solve the problems of slow convergence and low accuracy of the algorithm for thread pitch diameter measurement based on image processing at this stage, we proposed a thread steel wire head pitch diameter measurement method based on improved Harris hawk optimization algorithm. Firstly, the cubic spline interpolation method is used for sub-pixel edge detection, and the parameters such as the peak and valley of the thread are extracted accurately, then the pitch diameter fitness function is constructed. Finally, the spiral updating mechanism and nonlinear energy decreasing strategy are introduced into Harris hawk optimization algorithm (HHO) to solve the pitch diameter fitness function. The experimental results show that the improved Harris hawk optimization algorithm has better stability and higher accuracy, in which the standard deviation of diameter measurement is 59. 33% lower than the traditional HHO algorithm, the absolute mean error of pitch diameter measurement is 5. 08% lower than the three needle measurement method, and 37. 78% lower than the HHO algorithm.

Abstract:Accurate extraction of flame image features for converter steelmaking is the key to predicting end point carbon content. Aiming at the high similarity of flame images, it is difficult to distinguish flame images with similar carbon content, which leads to the problem that the carbon content cannot be accurately predicted. In this paper, an improved complete local binary pattern (ICLBP) color texture feature extraction method is proposed to extract more differentiated flame features at furnace mouth under different carbon contents and predict the endpoint carbon content. Firstly, local phase quantization ( LPQ) is used to extract image frequency domain phase information under different color channels, and the fusion feature ICLBP _ MP is combined with image spatial domain amplitude information extracted by CLBP to enhance the robustness of CLBP algorithm structure. Then, it is weighted by an improved color information weighting strategy to enhance the color contrast information of the flame image. Finally, the K nearest neighbor regression model is used to predict the carbon content. The experimental results show that the accuracy rate of carbon content prediction is 83. 9% within the error range of 0. 02%.

Abstract:In order to meet the demand of high data transmission bandwidth of spectral imaging method using fiber Bragg grating (FBG) in engineering applications for transient temperature and high frequency vibration signal measurement in the gun bore, a high-speed spectral data transfer method based on direct memory access ( DMA) is proposed. A hardware logic of high-speed spectral data transmission based on Zynq is designed, and the FBG demodulation system is built with the principle of spectral imaging, which has realized the synchronous transmission of spectral data by DMA and FBG wavelength demodulation. The results of data transmission simulation experiment and FBG center wavelength demodulation show that the demodulation system can transmit FBG spectrum data at a high speed, the data transmission bandwidth is 320 Mbit / s, the demodulation rate reaches 34 kHz, and the spectrum data transmission has a high stability.

Abstract:In order to improve the prediction accuracy and speed of short-term traffic flow, based on the instability and randomness of the traffic flow sequence, fast ensemble empirical mode decomposition ( FEEMD) and natural selection adaptive mutation particle swarm optimization algorithm (SAPSO) are proposed to optimize the two-way Predictive model combined with long and short-term memory network (BiLSTM). Firstly, using FEEMD to decompose the original unsteady traffic flow sequence into multiple stable intrinsic modal components (IMF) and residual components (Res), and filter out the noise part to improve modeling accuracy; secondly, introducing composite Multi-scale permutation entropy (CMPE) to detect the randomness of traffic flow sub-sequences and regroups them to simplify model construction and improve prediction accuracy; then, using BiLSTM to predict the reorganized subsequences, and use SAPSO to optimize the weights and thresholds of BiLSTM to further improve the prediction accuracy and prediction speed of the combined model; finally, the prediction values of each sub-sequence are superimposed to obtain the final prediction value. The experimental results show that the root mean square error of the FEEMD-SAPSO-BiLSTM combined model is 22. 9% and 54. 3% lower than the FEEMD-PSOBiLSTM combined model and the SAPSO-BiLSTM combined model, respectively. In terms of convergence speed, the FEEMD-SAPSOBiLSTM model is obviously faster than FEEMD-PSO-BiLSTM model. Therefore, in predicting short-term traffic flow, the proposed combined model improves the prediction accuracy and speed and achieves the desired prediction effect.

Abstract:The steam ejector is a key component of the low-temperature multi-effect distilled seawater desalination system (MED-TVC). To improve the efficiency of the system, the steam ejector has been optimized and analyzed for its structure. The influence of mixing chamber diameter on ejector ratio is studied using the computational fluid dynamics model (CFD) technology during the process of the ejector structure optimization, and the quadratic function relation between the mixing chamber diameter, the entrainment ratio, and the critical back pressure are obtained. The results show that the increase of mixing chamber diameter will lead to the increase of entrainment ratio and the decrease of critical back pressure. When the primary flow pressure and the secondary flow pressure are fixed at 600 and 15 kPa, respectively, by increasing the diameter of the mixing chamber from 15. 8 to 20. 35 mm, the ejection ratio can be increased from 0. 485 to about 0. 746, and the performance can be increased by 53. 8%. The water production ratio can be increased from 7. 425 to 8. 73, and the percentage increase is about 17. 58%. The optimized steam ejector structure can effectively increase the daily water output of the seawater desalination system, improve system operation efficiency, and reduce energy consumption.

Abstract:3D lung information can help doctors making faster and more accurate diagnosis. However, 3D lung reconstruction is derived from stacking of two-dimensional tomographic images. The quantity of two-dimensional images restricts the quality of 3D reconstruction. Increasing the number of electrode layers could achieve more two-dimensional images. However, the calculation complexity is increased and imaging speed is down. The inter-layer interpolation can solve this problem, however, the specificity and irregularity of human body geometry make electrical impedance tomography (EIT) lung images irregular. As a result, the traditional interpolation algorithm suitable only for regular images which cannot be directly applied to EIT images. Thus, a three-dimensional interlayer interpolation algorithm suitable for pulmonary electrical impedance is proposed. Firstly, the contour of the interpolated image is obtained, and then the pixel value is interpolated with the corresponding point. Simulation and experiment verify that the relative error of the new interpolation algorithm is 5. 69%, 3. 3% lower than that of linear interpolation and spline interpolation respectively. The three-dimensional interpolation method based on contour shape can further improve the quality of three-dimensional electrical impedance imaging and better reflect the real shape of the lung under the condition of limited measurement data.

Abstract:To study the mechanism and characteristics of abnormal vibration caused by partial discharge of GIL equipment, and predict the trend of partial discharge of GIL. Firstly, the tip discharge model and test platform of GIL equipment are established by taking the tip discharge as an example. Secondly, the characteristics and mechanism of GIL abnormal vibration behavior are studied by measuring the pulse current signal and abnormal vibration signal of tip discharge and combining with electromechanical signal. Finally, WOA-ELM model is established to predict the trend of GIL partial discharge and is compared with ELM model. The results show that the repetition frequency of discharge pulse current at GIL tip is consistent with the frequency of abnormal vibration signal, the partial discharge of GIL is directly related to the abnormal vibration of GIL shell, and the abnormal vibration energy is mainly concentrated in 1 600~ 2 800 Hz, and the frequency doubling distribution is mainly 20 Hz; Compared with the traditional ELM model, WOA-ELM has better prediction ability, which can realize the accurate prediction of GIL partial discharge trend, and provides a new method for GIL partial discharge trend prediction based on vibration signal.

Abstract:Aiming at the problem of low accuracy and slow training speed in complex circuits with multiple electrical faults, a method of window division combined with wavelet decomposition and empirical mode decomposition ( EMD) is proposed to extract current characteristic quantities respectively from multiple dimensions in time domain, frequency domain and time scale, identifying arc fault by using machine learning classification models. Firstly, the fault and normal current data are collected by the electrical fault experimental platform, and the current data is segmented by window. Then, the wavelet transforming and EMD methods are used to decompose the current signal and calculate the characteristic quantities in different dimensions. The characteristic information collected is used as the input of the classification algorithm for arc fault diagnosis. The experimental results show that the arc fault detection accuracy of the feature extraction method on the gradient boosting decision tree (GBDT) is as high as 98%, which is 1. 87% higher than that of the current without segmentation. It can effectively obtain the arc fault characteristics and realize the detection of arc fault with high efficiency and high accuracy.

Abstract:For the specific emitter identification method, fingerprint feature extraction needs complex formula calculus reasoning, the feature difference is small, the extraction is difficult, and the accuracy of specific emitter identification after extraction is low. In order to better extract fingerprint features, a specific emitter identification algorithm based on dense connection structure and attention mechanism is proposed, which is called specific emitter identification network ( SEIN). First, the envelope extraction algorithm is used to extract the envelope of the radiation source signal with less noise, and an envelope map with rich fingerprint features is obtained, then the SEIN fingerprint feature extraction and individual recognition are performed. The experimental results show that SEIN can achieve a classification and recognition effect of 95. 12%, has the characteristics of high accuracy and automatic fingerprint feature extraction, and finally achieves better specific emitter identification in complex environments.

Abstract:Aiming at the problems of few samples and difficult labeling in the remaining useful life prediction modeling of rolling bearings in high-speed and high-precision machining processes such as precision electronics and plastic shaping, this paper introduces the deep belief network that integrates the unsupervised and supervised fine-tuning learning methods to carry out the research on the prediction of the residual service life of rolling bearings. The vibration data features of rolling bearing are taken as input and the remaining useful life as output. The probability distribution of features accuracy quantified by energy function is taken as the basic component, and the feature output of the previous layer of the components is taken as the input of the next layer. The remaining useful life prediction model of rolling bearing is constructed by connecting multiple such components head to tail. The initial parameters of each unit in the model are obtained by unsupervised pre training of the original data, then the supervised fine-tuning of the model is carried out by using the remaining useful life label data to further improve the accuracy of the model prediction. The experimental results show that the method proposed in this paper can predict the remaining service life of rolling bearing. Compared with SVR and PCA-DBN, the prediction error is reduced by 28. 48% and 5. 57% respectively. This method has higher prediction accuracy in field prediction, and it can reduce the dependence on expert knowledge as well as improve generalization ability.

Abstract:Based on the pressure potential differential (PPD) type gas laminar flow technology, a new type of laminar flow transducer is designed and tested. The flow transducer is composed of two components, i. e. main part and integrated cover plate. The shell of the transducer is machined from a block material, and capillary components are fixed in the two channels of the shell. There are internal embedded channels in the cover plate for pressure acquisition, which can replace outside pressure tubes. Other components, such as capillary components/ bundles, mesh filter, and pipe connector, are designed with the idea of modularity. The differential, absolute pressure sensors, and temperature sensor are integrated on the pressure acquisition cover plate. The principle of the PPD gas flow transducer is analyzed and the correction formula is given. Based on a piston gas flow standard device, the PPD laminar flow transducer was calibrated with air, and tested with air and nitrogen gases. The maximum flow rate is about 50 L/ min. The results show that, the flowrate measurement errors for the two kinds of gas are within ±0. 8%, and the turndown ratio is about 250. It is predicted that the PPD laminar flow transducer will be used in the micro & small gas flow measurement fields.

Abstract:Aiming at the low detection accuracy of traditional traffic sign recognition algorithms,a traffic sign recognition method with improved YOLOv5 algorithm is proposed. First,improve the loss function of the YOLOv5 algorithm,use the EIOU loss function instead of the GIOU loss function used by the YOLOv5 algorithm to optimize the training model,improve the accuracy of the algorithm, and achieve faster identification of the target,then use the weighted Cluster NMS to improve the YOLOv5 itself. The weighted NMS algorithm improves the accuracy of generating the detection frame. The experimental results show that the mAP value of the model trained on the CCTSDB traffic sign dataset produced by Changsha University of Science and Technology by the improved YOLOv5 algorithm reaches 84. 35%, which is 6. 23% higher than the original YOLOv5 algorithm. Therefore,the improved YOLOv5 algorithm has higher accuracy in traffic sign recognition and can be better applied to practice.

Abstract:At present, the development of performances has evolved from factors such as auxiliary failures, severe signal interference, and poor signal interference in various control systems such as performing arts, stage machinery, sound, and equipment development. The development of stable and reliable stage equipment is developed for the evolution of performance and performance, and the actual operation of stage equipment After environmental research, an adaptive frequency hopping anti-jamming technology for stage carrier communication based on deep network is proposed. First, the model is used to obtain a large number of stage power line transmission channel parameters, and then a deep neural network is used to train an adaptive frequency hopping anti-jamming model, which can adaptively select the communication frequency band according to the stage power line communication environment. Experiments have proved that this technology can quickly and effectively select a stable communication frequency band according to the interference situation of the stage communication power line environment. When the signal-to-noise ratio is close to - 18 dB, the bit error rate is reduced by nearly 10 times compared with the traditional method, which improves the communication efficiency while further improving the reliability and stability of communication between stage equipment.

Abstract:In order to overcome the disadvantages of slow convergence speed and poor separation accuracy when solving blind source separation (BSS) problems in traditional swarm intelligence algorithms, a BSS method based on improved elephant herding optimization (IEHO) is proposed. The method uses the principle of independence to construct the objective function, combining the kurtosis and the mutual information of separating signals. In the clan update stage, the scale factor of algorithm is modified as well as exploiting the neighborhood search to improve the diversity of the search. In the separation stage, the quantum-behaved particle swarm optimization strategy is introduced to improve the global search ability of the algorithm. The simulation results show that the IEHO algorithm has a better optimization effect compared with the traditional elephant herding optimization algorithm and the particle swarm optimization algorithm. Meanwhile, the proposed algorithm can also realize blind source separation of images and speech successfully, with higher separation accuracy and faster convergence speed.

Abstract:Aiming at the low tracking accuracy of target work-piece in complex environment, an improved tracking method for target workpiece based on ECO is presented. Firstly, based on the framework of eco correlation filter, VGG features and traditional manual features are weighted and fused to improve the tracking accuracy. Then, the fast discriminant scale space tracker is used to track the target workpiece adaptively. Finally, a high confidence update index is introduced to determine the sparse update strategy of the tracking model to improve the robustness of the algorithm. Tested on the OTB - 2015 standard dataset and compared with other mainstream tracking algorithms, the experimental results show that the average tracking accuracy and the average overlap accuracy of the algorithm are both optimal, reaching 89. 2% and 68. 6%, This algorithm also has good tracking performance for target work-piece dataset taken with CCD industrial camera, which further verifies the validity of the algorithm.

Abstract:In the process of using electromagnetic induction method to diagnose the breakpoint of the grounding grid of the tower, aiming at the error caused by manual diagnosis, this paper proposes a diagnosis model based on one dimensional-convolutional neural network (1D-CNN), the diagnosis model takes the one-dimensional magnetic field data directly above the grounding grid as input, and outputs the number and location of breakpoint faults through a deep neural network. This paper firstly verified the effectiveness of electromagnetic induction method in the diagnosis of tower grounding grid breakpoints through experiment, then a magnetic field breakpoint fault dataset was established and a 1D-CNN diagnosis model was trained. In the diagnostic accuracy verification experiment, the diagnostic model reached 97. 50% diagnostic accuracy on 40 faulty magnetic field samples, showing good generalization. The comparison experiment of the diagnosis effect shows that the AUC value of the 1D-CNN diagnosis model reaches 0. 951, the average recognition rate of various faults in three random trainings reaches 92. 08%, and the average test set accuracy in 15 trainings reaches 94. 30%. and the average training time per generation is 0. 875 0 s, which has obvious advantages over DNN and RNN in various indicators.

Abstract:In order to solve the problem of pattern recognition of gas pressure detection in natural gas pipeline, the original signal is preprocessed to remove redundant information, and then the signal is decomposed by variational mode decomposition to extract the optimal Intrinsic mode function and reconstruct the signal. Then, the processed signal is transformed into a high-resolution twodimensional image in time and frequency domain by continuous wavelet transform. Finally, the image is extracted by deep convolution neural network, and the output of part of the network is connected with support vector machine to realize supervised learning and training. The trained support vector machine is used for unsupervised pattern recognition of the remaining data. Experiments show that the accuracy of vmd-cnn-svm is 90. 66%, which is the highest compared with other methods.

Abstract:To solve technical problems of small flow measurement and automatic flow regulating in micro fan aerodynamic performance testing system, a micro fan aerodynamic performance test technology based on the flow balancing method is proposed. Based on the principle of the flow balancing method, the characteristics of aerodynamic and flow regulating of the pipe system are analyzed. Because of adopting flow balancing method, the overall impedance of the pipeline system is basically unchanged when switching on different working channels, and the intended flow rate can be obtained by combining different measuring branches. Theoretically, according to the valve switching on or off, the flow rate of one of the parallel measuring branches is either “1” or “0”, thus it is very convenient to obtain the required flow rate by combining different branches. Experiments were carried out, including flow calibration, flow regulation performance tests and aerodynamic performance test for a micro fan. The results show that flow regulation performance of the combination method meets the expectation, and the combining flow regulation method works well in the micro fan testing experiment, which proves the practicability of the flow balancing method for the micro fan aerodynamic performance testing technique.

Abstract:To solve the capacitive accelerometer sensitivity low, cross sensitivity higher performance problems, we design a new quality of three pieces of three axis MEMS capacitive accelerometer, the quality of three piece of accelerometer to measure the X, Y, Z three orthogonal axes and acceleration, separate horizontal (X, Y axis) and vertical axis (Z axis) accelerometer integrated on the same SOI wafer, Low cross sensitivity and high displacement sensitivity are achieved by parallel bonding on the glass substrate. The mathematical model of the accelerometer structure was analyzed and established, and the structure and size of the accelerometer were optimized through COMSOL simulation analysis. The displacement sensitivity of the three mass blocks accelerometer in the vertical axis direction was 1. 536 69 μm/ g, and the displacement sensitivity in the horizontal axis was 6. 78 μm/ g, and the cross sensitivity was all less than 1%.

Abstract:Aiming at the problem of low segmentation accuracy due to small blood vessels and retinopathy in fundus blood vessel images, a U-Net retinal blood vessel image segmentation model that introduces residual blocks, cascaded cavity convolution, and embedded attention mechanism is proposed. First, increase the resolution of the retinal image, crop the data set with point noise as the center and 512 as the side length, and then introduce residual blocks in the U-Net model to increase the utilization of pixel features and avoid the degradation of deep networks; And replace the bottom of the U-Net network with a cascaded hole convolution module to expand the receptive field of the feature map and extract richer pixel features; finally, the attention mechanism is embedded in the decoder to increase the weight of the target feature and slow down useless information Interference. The experimental results based on the CHASE data set show that the accuracy of the proposed model reaches 98. 2%, the sensitivity reaches 81. 72%, and the singular value reaches 98. 90%. Compared with other multi-scale neural network methods, it embodies better segmentation results, and fully verifies that the improved U-Net network model can effectively improve the accuracy of blood vessel segmentation and assist in the diagnosis of vascular disease.

Abstract:As the terminal equipment of the power grid, the degradation of smart meters is closely related to factors such as working environment and running time. Aiming at the problem that the degradation of smart meters under complex variable conditions is difficult to predict, a smart meter basic error prediction method based on the composite core support vector machine (SVM) is proposed. First, analyze the degradation data of smart meters, and use Pearson correlation analysis to find environmental variables that are highly correlated with the basic errors of smart meters. Then, in order to further extract the data degradation features, the fuzzy C-means clustering algorithm is used to cluster the smart meter degradation data and determine the degradation feature vector. Finally, based on the Gaussian radial basis kernel function and polynomial kernel function, a new composite kernel SVM model is constructed to predict the basic error of smart meters. The performance of the composite core SVM model is verified by combining the degradation data of smart meters in Xinjiang. The experimental results show that the composite core SVM model proposed in this paper can accurately predict the basic errors of smart meters in complex environments, and its prediction accuracy is higher than that of Bayesian methods. Neural network method and classic SVM method.

Abstract:In order to solove the problem of slow detection speed and low recognition rate of common commodity recognition algorithms on the intelligent vending cabinet embedded system platform, an improved commodity recognition algorithm DS_YOLOv3 is proposed on the basis of YOLOv3. The traditional YOLOv3 neural network algorithm is improved by obtaining a prior bounding box suited for the image data of beverages sold in the vending cabinet by using k-means++ clustering algorithm, using the deep separable convolution to replace the standard convolution and adding the inverted residual module to reconstruct the YOLOv3 algorithm, which could reduce the computational complexity and enable real-time detection on the embedded platform, and introducing CIoU as the bounding box regression loss function to enhance the accuracy of target image positioning. The commodity testing experiment under typical scenarios is performed on a computer workstation and Jeston Xavier NX embedded platform. The results show that the accuracy of DS _YOLOv3 algorithm reaches 96. 73%, and the actual detection rate on the Jeston Xavier NX platform is 20. 34 fps, which meet the real-time and commodity detection recognition accuracy requirements of the intelligent vending cabinet based on the embedded system platform.