Abstract:This paper uses a bi-directional long short-term memory (BiLSTM) neural network model to solve the state of health (SOH) and remaining useful life (RUL) traditional prediction methods of lithium batteries. The accuracy of traditional prediction methods is low. The problem. Firstly, extract the capacity data of the national aeronautics and space administration (NASA) lithium battery, convert the capacity data into SOH data and use it as input data; secondly, establish a two-layer BiLSTM neural network and use the Nadam optimization function to dynamically adjust learning rate; Then, the lithium battery data is analyzed through the two-way long and short-term memory neural network model to establish the connection between battery capacity, SOH and RUL; finally, the fully connected layer outputs the estimated curve of the battery SOH to predict its remaining life. Prediction experiments with NASA data show that the RUL prediction error of the BiLSTM neural network is stable within 3, and the fit of the SOH prediction curve is stable at 94.211%-95.839%. The BiLSTM neural network has higher robustness and accuracy.

Abstract:Aiming at the problem that the range and accuracy of low-power laser ranging radar are difficult to balance, a high-performance laser radar ranging system based on FPGA is designed. The system first uses the 14-order M-sequence pseudo-random code to modulate the emission laser at a frequency of 100MHz, and uses dual-port BRAM ping-pong read and write to realize the accumulation of echo photons and restore the echo signal. Then the distance measurement information is obtained through correlation calculation to establish the echo probability model, and the centroid algorithm is used to improve the accuracy of the distance measurement. Finally, a prototype of the radar system is built. The range error is stable at ±2cm when the distance is measured in a 100-meter space environment; when the optical fiber delay method is used to measure the equivalent target of 14.8km, the range accuracy is better than 0.1m. The experimental results show that the system can improve the ranging accuracy when the laser modulation frequency is lower, effectively reduce the difficulty of data processing, and finally realize long-distance high-precision lidar ranging.

Abstract:Aiming at the problem that the output current of the wireless power transfer system was changed due to the easy migration of coupling mechanism, a dual topology wireless power transfer system with anti-migration characteristics was studied. In this paper, the output characteristics of SS and LCC-LCC topological circuits were analyzed firstly, and DDQ coil structure was selected to realize the decoupling between coils. Then, the parameters of compensation network were optimized based on particle swarm optimization algorithm to make it have better anti-migration characteristics. Finally, the system circuit was built in PSIM simulation software to verify the feasibility of theoretical analysis.The simulation results showed that the offset range of the system coupling mechanism was 0 mm~160 mm, the output current fluctuation range was 8.52 A~9.36 A, the output current fluctuation ratio was ±4.7%, and the system efficiency was up to 91.47%.

Abstract:In the field of GNSS-RTK deformation detection, the observation time of dynamic relative positioning is short and the error elimination is not sufficient, which leads to the phenomenon of gross error and large data noise in baseline. To solve the above problems, an improved unconstrained adjustment algorithm is proposed in this paper, which can reduce the error of the system solution and restrain the gross error in the baseline vector. Aiming at the residual gross errors in the baseline vector, the anti-error Kalman filter algorithm is adopted to further eliminate the residual gross errors and obtain more accurate state estimation. Compared with traditional moving average filtering, wavelet threshold denoising and EEMD-wavelet combined denoising, the new method achieves outstanding denoising effect without adding hardware resources, and retains more details in the baseline coordinate domain. The baseline root mean square error is reduced from 2.36cm to 1.00cm, and the accuracy is improved by 57.6%.

Abstract:Aiming at the problems of large computing resource consumption and lack of real label data in the process of partial discharge mode intelligent identification in GIS. This paper uses the MobileNet-V2 model whose activation function is Leaky ReLU to extract a large amount of image feature information while reducing the amount of model parameters.It also integrates migration learning to pre-train the model parameters, which reduces the network's need for input data and improves the recognition accuracy of the model.The results show that: the parameter quantity of the model in this paper can be reduced to 2.24×106, and the average accuracy of interference and partial discharge pattern recognition in GIS reaches 95.8% and 92.1%, respectively.Compared with the traditional deep learning model, this model can significantly reduce the computational complexity and improve the accuracy of pattern recognition, which has certain value and significance for effective, intelligent and lightweight operation and maintenance of actual GIS equipment.

Abstract:The verification platform plays an important role in the design of the video acquisition and compression card chip. Aiming at the shortcomings of traditional verification platforms in terms of code coverage and test efficiency, this paper designed a SystemVerilog-based verification platform. The verification platform is designed with an object-oriented programming language, where the PCIe host (RP) adopts the Xilinx IP modeling link layer and the physical layer, ensuring that the PCIe bus environment is the same as the real host board card environment. The external verification environment adopts the SystemVerilog hierarchical design method, and adopts the class idea to design the upper-layer verification environment, so that more verification components can be transplanted to different types of SoC of the same interface protocol. Moreover, in the automated verification stage, the case state is automatically judged by the simulation report, adjusting the random benchmark and tracking the uncovered module paths in the coverage report, which greatly improves the corner coverage of the code and accelerates the regression convergence. The verification platform is analyzed from three aspects: acquisition and compression simulation process, verification automation, and coverage. The results show that the verification platform can quickly complete the horizontal verification module transplantation of similar design, improve the verification reliability of similar function chips, save manpower, accelerate the simulation progress, accelerate the coverage convergence, shorten the verification cycle and increase the success rate of flow chip.

Abstract:Aiming at the current manual operation of the kilogram weight verification process, there are problems of low verification efficiency, high labor intensity, and low degree of automation. This paper proposes a real-time recognition and spatial positioning technology for the weight handle of stacked kilogram weights. First, the Faster R-CNN target detection method is used to realize the identification and positioning of the weight and the weight handle; according to the identified pixel point coordinates of the weight handle, the depth value information of the weight handle is obtained correspondingly. Preliminary experiments show that the method in this paper can realize the real-time recognition and positioning of the weight handle of the stacked kilogram group. The average accuracy of the boundary box recognition of the weight handle is 98.5%, and the time to acquire the image and identify the spatial coordinates of the weight in a single time does not exceed 0.473 s, the accuracy of weight handle recognition in multiple experiments is 100%, which meets the real-time identification and positioning requirements of the weight handle of the stacked kilogram group.

Abstract:To improve the power generation ability of PVDF piezoelectric film, a PVDF piezoelectric energy collection structure based on vortex-induced vibration mechanism was explored, which converts wind energy into electric energy for micro power electronic products. A signal conditioning circuit was designed to convert the charge output of piezoelectric film into voltage output. The power generation performance of single PVDF piezoelectric film, double and three PVDF piezoelectric films in series and in parallel under different wind speeds was studied by micro wind tunnel experiment. Experimental results show that the resonant frequency of the selected PVDF piezoelectric film is stable at about 50 Hz. The signal conditioning circuit can filter out the high-frequency noise interference effectively, and the output voltage of the signal conditioning circuit increases with the increase of wind speed. The parallel connection of PVDF piezoelectric films and matching load at the end of the circuit can effectively improve the load power. The load power of 36 mW can be achieved when three piezoelectric films are connected in parallel and a 1.25 Ω load is connected at the end of the signal conditioning circuit.

Abstract:In-situ detection is an important means to ensure the readiness of airborne EW equipment, the current in-situ detection mode mainly carries out rough detection of functions, with incomplete detection coverage and incomplete troubleshooting of potential failure, which makes it difficult to find the degradation of equipment operational performance in time. Aiming at these shortcomings, the concept of in-situ depth detection of airborne EW equipment is proposed, and its connotation, mode and key technology are studied. Engineering practice is carried out in the in-situ fault depth diagnosis, equipment performance change trend monitoring and rapid positioning calibration of airborne EW equipment, which can keep the readiness rate of pilot equipment above 95%, The results show that the in-situ depth detection technology of airborne EW equipment plays an active role in improving the readiness rate of equipment.

Abstract:This paper presents an off-chip interleave calibration scheme for the Offset error, Gain error and Skew error of the high sampling rate time-interleaved analog-to-digital converter (TIADC). A time-interleave calibration algorithm is based on statistical approximation. Through Ethernet communication, the quantization information of the channel to be calibrated is transmitted to the PC to extract the mismatch parameters, and the error is compensated in the form of negative feedback. The scheme is not affected by high speed data transmission and synchronization, does not consume logical resources in error extraction stage, supports various large-scale and high-consumption calibration algorithms, and has a short development cycle. When applied to the self-developed 3GS/s -12bit four-way interleaving TIADC, the test results show that the ADC effective bit (ENOB) is increased by 2.69 bits on average in the 2.5G input signal bandwidth under the condition of other non-ideal factors. The calibrated SFDR improved by an average of 29.73 dBc. It is proved that the algorithm and the calibration scheme are effective.

Abstract:Aiming at the problems in environmental monitoring, such as the security of environmental information can not be guaranteed, remote areas are not easy to monitor, and there are blind areas in the use of communication network, this paper uses Beidou’s high-precision spatiotemporal technology and short message communication function, and the powerful graphic analysis function of geographic information system to build a space-sky-terrestrial integration environmental monitoring platform based on Beidou. Combined with the characteristics of the Yangtze River Basin and Dongting Lake, we combined the Beidou system with the relevant professional environmental monitoring sensors to form the space-sky-terrestrial integration environmental monitoring application scenario based on Beidou, then monitor various environmental indicators and send the monitoring data to database for analysis and visualization. The results show that the environmental monitoring platform based on Beidou can achieve 24-hour uninterrupted data acquisition and transmission, and provide data support for scientific management and decision-making.

Abstract:The development trend of ' reducing people and improving efficiency ' in coal mine production makes it more and more important to ensure the safety of workers. Aiming at the problems of large amount of data and weak robustness of current miner abnormal behavior detection methods, a miner abnormal behavior recognition method with structured discrete attitude perception is proposed. The Kalman filter technology is used to optimize the behavior perception information obtained based on the nine-axis attitude sensor. After the behavior information is intercepted by the sampling window, the three-channel RGB behavior image is structured according to the axial direction of the attitude perception. Combined with the CTFRN model designed to extract the temporal and spatial characteristics, the temporal and spatial characteristics of the five kinds of miners’ behaviors are accurately extracted, and the miners’ abnormal behaviors are monitored with low computational complexity and high robustness. Compared with other models, the results show that the proposed method has higher accuracy, up to 99.3%. The designed system and recognition method can be used for real-time monitoring of abnormal behavior of miners in actual environment to ensure the safety of miners.

Abstract:In the process of rotor position estimation of PMSM based on single-frequency pulse voltage injection method, there are some problems, such as incomplete removal of dc component, easy influence of inductance, when using filter method and direct calculation method to eliminate dc component. A synchronous modulation technology is proposed to remove the dc component. This rotating voltage vector is injected into stationary coordinate system, and using band-pass filter response to high frequency current is extracted, and then use the synchronization modulation technology to achieve four independent rotor position information, and cancel by poor response to the current of dc component, and finally using two-phase type phase-locked loop to get the final estimate rotor position. The simulation results show that compared with the filter method and the direct calculation method, the proposed method reduces the maximum error of rotor position estimation by 0.1rad and 0.14rad respectively when the motor starts. The position estimation error of the proposed method is minimum under the condition of inductance variation. It is proved that the proposed method has the advantages of high precision and insensitive to inductance change.

Abstract:A genetic algorithm array structure optimization scheme for Hanbury and Brown-Twiss (HBT) interference localization is proposed for the problem that the acoustic localization performance is easily affected by the microphone array structure. The method constructs intermediate populations with the distance difference between two adjacent array elements as individuals, and converts them to distance-spaced populations by gene sorted by ascending order. Then, the objective function based on the maximum parametric level is constructed with the directional map function as the fitness function, and the microphone spacing is taken as the optimization object to transform the objective function into an unconstrained optimization problem. The problem is solved by genetic algorithm to obtain the array structure with the highest localization performance. The simulation results show that the optimization effect of the seven-element array is the most significant, and the peak value of the near-point positioning partials is reduced from 0.4374 to 0, and the number of distant-point positioning interference partials is reduced to 0. This method can effectively improve the positioning accuracy.

Abstract:Aiming at the problems that the whale optimization algorithm is easy to fall into local optimum and low convergence accuracy when optimizing complex engineering, a whale optimization algorithm (ELWOA) based on elite backward learning and Lévy flight is proposed, which first optimizes the initialized population through elite backward learning to improve the diversity of the population; then increases the adaptive weight factor, which is beneficial to balance the global and local search ability of the algorithm; finally, the Lévy flight strategy is applied to the whale optimization algorithm to conduct a small search near the optimal position, which is beneficial to the algorithm to jump out of the local optimum later and improve the local search ability of the algorithm. Through the simulation and optimization analysis of several test functions, the results show that the ELWOA algorithm has faster convergence speed and better convergence accuracy than the WOA and MWOA algorithms.

Abstract:Bearing fault signal extraction is susceptible to strong background noise in the working environment,especially in the early fault signal detection,bearing fault signal is submerged by noise,resulting in limited detection.In view of the traditional adaptive stochastic resonance theory in bearing fault signal detection parameters optimization of a single defect,put forward a cuckoo algorithm to optimize stochastic resonance parameters based on the bearing fault detection algorithms,this method takes the output signal-to-noise ratio as fitness function,the theory of stochastic resonance in the two coordinate parameter optimization,get a set of optimal parameters,Adaptive stochastic resonance is best matched with input signal,noise and nonlinear system.Finally,through simulation comparison,the singal detection result of the proposed algorithm is better than that of the tradition stochastic resonance method.Experimental data of bearing fault diagnosis show that the detection error of bearing fault signals achieved by this algorithm is 0.15%.Experimental results show that the proposed method has the adwantages of high accuracy and good reliability,which is of great significance to the accurate detection of bearing faults and the stable operation of industrial equipment.

Abstract:Bearing fault signal extraction is susceptible to strong background noise in the working environment,especially in the early fault signal detection,bearing fault signal is submerged by noise,resulting in limited detection.In view of the traditional adaptive stochastic resonance theory in bearing fault signal detection parameters optimization of a single defect,put forward a cuckoo algorithm to optimize stochastic resonance parameters based on the bearing fault detection algorithms,this method takes the output signal-to-noise ratio as fitness function,the theory of stochastic resonance in the two coordinate parameter optimization,get a set of optimal parameters,Adaptive stochastic resonance is best matched with input signal,noise and nonlinear system.Finally,through simulation comparison,the singal detection result of the proposed algorithm is better than that of the tradition stochastic resonance method.Experimental data of bearing fault diagnosis show that the detection error of bearing fault signals achieved by this algorithm is 0.15%.Experimental results show that the proposed method has the adwantages of high accuracy and good reliability,which is of great significance to the accurate detection of bearing faults and the stable operation of industrial equipment.

Abstract:Dynamic obstacle avoidance is the key to the robot's autonomous movement and safe walking, in the face of complex and changeable indoor scenes, the robot needs to be able to detect obstacles in time and dynamically plan a safe walking route. In this paper， RGB-D depth camera and IMU unit was used to establish a robot environment perception system, multi-modal information such as three-dimensional vision and attitude angle were provided to the robot. At first, build an improved target detection model based on YOLOv4, The YOLOv4-M target detection algorithm was proposed to identify and locate obstacles in color images, and the depth map was aligned with the color map in order to calculate the size information of the obstacle and the distance information between the robot and the obstacle; The model of obstacle avoidance was built on modified artificial potential field method with the obstacle information in the environment and the posture and angle information of the robot movement, to solve the problem that the calculation of the total potential field falls into a local minimum solution. The model was designed with dynamic programming of the walking path, and the decision result was send to the robot chassis control unit to realize the autonomous movement of the robot in unfamiliar scenes. Simulation analysis and physical experiments show that this method can realize autonomous obstacle avoidance of robots. The research of this method provides a basis and reference for the robot to realize obstacle recognition and autonomous movement avoidance by relying only on vision and inertial navigation sensors.

Abstract:The distribution of the induced electric field generated intracranially by transcranial magneto-acoustic electrical stimulation is a key factor in determining the accuracy of neuromodulation. Phased arrays can generate highly focused acoustic fields, but the relationship between their structural parameters and intracranial induced electric fields is unclear. In this paper, a three-dimensional finite element model containing the phased array was developed, and the acoustic field distribution was calculated and evaluated by using the cross-sectional area of the focal zone, the depth of the focal zone, the acoustic intensity and the electric field strength. The actual electric field intensity at the focal area was measured by designing a mimic experiment. The effect of the phased array on the local field potential of rats at 0.8 MHz center frequency with different number of array elements coupled with 0.3 T static magnetic field was investigated using Wistar rats as the experimental object. The results showed that increasing the number of array elements could increase the electric field intensity in the focal area, and then decrease the oscillation frequency in the low frequency band (4-30Hz) and increase the energy in the high frequency band (30-80Hz), which revealed the mechanism of neuronal excitability regulation by phased array-based focused transcranial magnetoacoustic stimulation, and provided a reference for the design of transcranial magnetoacoustic stimulation system and the selection of phased array structure parameters.

Abstract:Using python language, the fuel consumption prediction model is built based on the vehicle operation data collected by OBD.Taking vehicle running state parameters such as vehicle speed v, engine speed n, intake pipe absolute pressure P, throttle position TP, coolant temperature CT, load rate L, idle time IT, acceleration an as independent variables and 100 km fuel consumption as dependent variables, the correlation intensity between parameters and dependent variables is sorted by SelectKbest function and briefly analyzed.The (MLP) neural network model of multilayer perceptron based on tensorflow and the multiple linear regression model of support vector machine (SVM) are used to predict the fuel consumption at the same time.Support vector machine model RMSE is 0.088 MAE is 0.56 tensorflow neural network model RMSE is 0.132 MAE is 0.70.The results show that the two models are accurate in the prediction of fuel consumption, which can provide a theoretical basis for further elucidating the relationship between vehicle fuel consumption and vehicle running state parameters.

Abstract:Reinforced concrete structures are widely used in the power infrastructure, and the detection of rebar can effectively judge the durability of the structure, which is of great significance to ensure the stable operation of the power system. Aiming at the problem that the traditional reinforced concrete detection method cannot determine the direction of the rebar and the measurement of the buried depth is not accurate enough, a method of positioning and buried depth detection based on electromagnetic induction is proposed, through the analysis of the measurement principle of the rebar and the layout design of the Hall sensor. This method is used to judge the center position of the rebar and measure its deflection angle, and the relationship between the detection value and the buried depth is reflected according to the fitting function. Experiments show that the buried depth detection value is affected by the adjacent rebar, the smaller the spacing, the greater the error. Therefore, the detection value under different spacing is corrected by the back propagation neural network, which effectively improves the detection accuracy of the buried depth of the rebar.

Abstract:Gas-solid two-phase flow parameter detection is widely used in industrial development， and the real-time monitoring of its parameters plays a decisive role in production efficiency and environmental protection. Based on the gas-solid two-phase flow experiment platform，this paper analyzes the sensitivity characteristics of the electrostatic sensor under different wind speeds by finite element; extracts the characteristic value of the collected electrostatic signal，combines the linear fitting analysis with the least square method，and proposes an electrostatic signal The mathematical model of characteristic quantity and dust particle concentration; finally verified by relative error analysis. The experimental results show that the root mean square value of the electrostatic signal is more related to the particle concentration，and the relative error is controlled within a very small range of 0.5% to 1.24%，which confirms the reliability of the mathematical model and is a gas-solid two-phase flow parameter detection technology Provides new ideas.

Abstract:In order to solve the problem of how to quickly and accurately measure sheep weight dynamically and improve the intelligence of smart farms, a dynamic processing algorithm based on BP neural network is proposed. A sheep dynamic weighing system is built, and the LabVIEW host computer is used to collect data. Four pressure load cell signals are selected as network inputs, and real sheep weight data are used as network outputs, and the input and output of BP neural network are trained and tested, because there are problems such as local minimum in BP neural network, the average relative error of test samples is large. The weights and thresholds of the neural network are optimized using the particle swarm algorithm. The results show that the average relative error of the test samples of BP neural network algorithm is 7.9%, and the average relative error of the test samples of PSO-BP algorithm is 5.3%, which indicates that PSO-BP neural network is more effective in reducing the dynamic weighing error of flock and has potential application value.

Abstract:With the development and application of airborne radar technology, it is faster and more convenient to obtain the spatial distribution of ground objects. To realize the terrain classification for airborne radar point cloud data, we first adopt the elevation histogram to remove the outliers of the original point cloud. Then, the voxelgrid filter is used to sample the point cloud, which can greatly reduce the number of points while maintaining the shape characteristics of the point cloud. Next, the K-means clustering method is improved. The K value and initial clustering center are determined through the height information of three-dimensional point cloud data, and the point cloud surface roughness is used for cluster analysis to realize the classification of different ground objects. The experimental results on the collected data and the open data show that the proposed method has high accuracy for point cloud data classification, with the overall classification accuracy reaching 88.17% and 90.47%, respectively, and the Kappa coefficient being 0.81 and 0.85, respectively.

Abstract:Current traffic classification methods often suffer from poor classification effects on minority classes when facing class imbalanced flows. To solve this problem, an end-to-end classification model for class-unbalanced encrypted traffic is proposed. The proposed model adds an Inception module to the traditional CNN model for feature fusion, so that the model can extract richer features, which makes up for the lack of feature learning caused by the small number of samples in a minority of categories; at the same time introduces a The channel-spatial domain attention module assigns corresponding weights to the features fused by the Inception module according to their importance, so that the model pays more attention to more important features and enhances the characterization ability of traffic features. At the same time, in order to reduce network parameters, a combination of convolutional layer and global average pooling layer is used to replace the fully connected layer in the model. The experimental results show that compared with the current typical traffic classification model, the proposed model has better classification performance on the minority classes of the data set, and the accuracy rate, recall rate and F1-Score have been significantly improved. And the comprehensive performance index F1-Score is on some minority classes. The improvement reached 15%~18%.

Abstract:Feature pyramid network (FPN) has become an effective framework for extracting multi-scale features in object detection. However, FPN has problems such as loss of semantic information due to channel reduction, high-level features only contain single-scale context information, and the direct fusion of different layer features with semantic differences resulting in aliasing effects. In response to the above problems, this paper proposes a feature pyramid network based on attention enhancement guidance, which is composed of channel feature enhancement module, context enhancement module and attention guidance fusion module. Specifically, the channel feature enhancement module reduces the information loss caused by channel reduction by modeling the dependency between the features, the context enhancement module uses different levels of features to extract context information to enhance high-level features，and the attention guidance feature fusion module uses the attention mechanism to guide the feature learning of adjacent layers to enhance the consistency of semantic information with each other. This paper replaces the FPN in the Faster R-CNN and Mask R-CNN object detectors with AEGFPN and performs experiments on different data sets, which experimental results show that the average accuracy of the improved Faster R-CNN detector on the PASCAL VOC and MS COCO datasets is increased by 1.5% and 1%, respectively, and the improved Mask R-CNN detector also improves the performance of Mask AP and Box AP by 0.8% and 1.1% on the MS COCO data set.

Abstract:In order to solve the issue that the existing ground-based cloud classification methods can not make full use of multi-modal information, we propose the Dense Fusion Convolutional Neural Network (DFCNN) for multi-modal ground-based cloud classification to effectively integrate the visual features and the multi-modal features of ground-based cloud samples. The DFCNN utilizes convolution neural network as the visual subnet to extract visual features and adopts the multi-modal subnet to extract multi-modal features of cloud samples. There are five Dense Fusion Modules (DFM) in the DFCNN and they are employed to fully fuse visual features and multi-modal features. The DFM could be injected into the subnet independently without changing the original network structure, and therefore it possesses great flexibility. The DFCNN achieves the classification accuracy of 89.14% on the public multi-modal ground-based cloud dataset MGCD, which verifies the effectiveness of the proposed DFCNN for the ground-based cloud classification task.

Abstract:In order to improve the recognition accuracy and positioning accuracy of fruit picking robot, a target recognition and positioning algorithm based on deep learning Faster-RCNN framework was proposed. Firstly, the convolutional neural network VGG16 model was used to extract the characteristics information of the input image, and the region proposal network RPN was used to generate the candidate box containing the target. The adaptive number of candidate boxes was introduced to improve the performance of the algorithm. Then, the multi task loss function was used to classify the target and correct the prediction box. Finally, the mapping relationship between the two coordinate systems of the hand and eye of the picking robot was solved by calibration, so as to realize the accurate recognition and positioning of the fruit. The experimental results of apple recognition and location show that the proposed algorithm has high recognition accuracy, the average accuracy is 97.5%, and the location error is lower, the maximum error is only 1.33cm, which can provide strong technical support for the development of smart agriculture.

Abstract:The traditional Faster r-cnn positioning algorithm uses RoiPooling, which is interpolated by the nearest neighbor interpolation algorithm. The recognition effect of small defects is not good. This article will improve it to RoiAlign using bilinear interpolation algorithm, which improves tire abnormality detection. Accuracy. Traditional tire defect sample detection faces the problem of difficult feature extraction. In this paper, the RSDC-Net (resnet and densenet converged network) network model built by fusing the two network models of ResNet and DenseNet has improved the generalization and perception capabilities of the network. , Enhance the feature extraction ability, and apply the network to the interpretability research of deep learning, and realize the visualization of deep learning. At present, there is still a big gap in the research field of neuron classification. Therefore, in order to carry out the research on the neuron classification of the latent layer according to the image results of the sensitive area, this paper designs a double convolution threshold recurrent neural network as a network model to complete the neuron classification , The network model performed best in the four model comparison experiments.