57results about How to "Reduce precision loss" patented technology

The invention discloses a quantitative training method and system for a deep neural network, and the method comprises the steps: obtaining an original neural network model and a verification data set, and generating a quantitative priority through hierarchical sensitivity evaluation and topology analysis; on the basis of the layer sensitivity matrix, differentiated quantization parameters are configured for each layer, and quantization constraints are established; constructing an error compensation mechanism, and generating an error compensation strategy; executing dual-scale adaptive quantization training, and respectively adopting an independently optimized quantization strategy for forward propagation and back propagation; dynamically adjusting quantization parameters according to the training state and the resource constraint; and evaluating model performance and resource efficiency. According to the method, through multi-dimensional sensitivity evaluation, a dynamic quantification strategy and system error compensation, the precision and efficiency of quantification training are improved, and meanwhile it is guaranteed that deployment resource constraints are met.

The invention belongs to the technical field of mixing precision quantification, and discloses a mixing precision quantification method of a deep neural network and a related device. The mixing precision quantification method comprises the following steps: carrying out diversity processing on a population by using a variation strategy in a differential evolution algorithm to obtain a variation population; recombining the sub-components of the variation individuals in the variation population with the sub-components of the target individuals in the current population by using binomial crossover operation to obtain a plurality of test individuals; performing hybrid precision quantification on the deep neural network based on the test individual and the target individual, and selecting a better individual from the test individual and the target individual; and based on the selected better individual, judging whether the differential evolution algorithm converges or not, if so, obtaining the deep neural network after final mixing precision quantification, and otherwise, selecting the obtained better individual as a population of the next generation. The method can realize mixing precision quantification of the deep neural network, and has the advantages of relatively low calculation complexity and relatively high adaptability.

The invention discloses a self-adaptive pruning method based on priority experience playback. The method comprises the steps of training a to-be-pruned image classification model, converting a pruning process of a neural network into a Markov decision process, and designing a corresponding state space, an action space and a reward function; constructing and initializing a value network, a strategy network, a target value network and a target strategy network of a depth deterministic strategy gradient algorithm; pruning the to-be-pruned image classification model; storing the pruning process of the to-be-pruned image classification model into an experience playback pool, and setting a priority for the pruning process; updating the sampling probability of each sample according to the time difference error; updating the value network and the strategy network, updating the target value network and the target strategy network, and carrying out loop training until a set maximum number of iterations is reached; and carrying out fine tuning on the image classification model to be pruned of image classification. According to the method, the pruning proportion of each layer of the network can be adaptively generated, and adaptive compression of the network to-be-pruned image classification model is realized.

The invention discloses a rapid marine ship target identification method based on YOLOv8, and is applied to the technical field of target identification. Comprising the following steps of collecting an image sample of a ship navigation water area in a digital vehicle, cleaning and marking the image sample, performing data enhancement on the image sample to obtain a data set which can be used for training and verification, constructing a target identification network model based on a YOLOv8 deep neural network, and obtaining a target identification result by setting hyper-parameters such as a learning rate and a batch processing size. And carrying out transfer learning by using the pre-training model, removing a repeated detection frame by using a non-maximum suppression algorithm and carrying out result filtering by using a confidence coefficient threshold to carry out post-processing on a reasoning result of the model so as to complete prediction of the ship target, and carrying out sparse pruning training on the trained model neural network so as to obtain the ship target. And carrying out calibration and quantification on the pruned model. According to the method, on the premise that the model detection precision is reserved to the maximum, memory consumption during equipment operation is reduced, and meanwhile the model detection speed is increased.

The invention discloses a hardware-friendly pruning method based on a Vision Transform neural network, and belongs to the technical field of calculation, reckoning or counting. The pruning method is a scheme combining layer-by-layer N: M weight pruning and non-parameterized multi-stage adaptive token pruning, and comprises the following steps: calculating an average Hessian trace of each layer of the network according to a neural network and a pre-training weight, realizing layer-by-layer N: M weight pruning according to the distribution of the average Hessian trace, carrying out weight pruning retraining and updating the weight, and carrying out multi-stage adaptive token pruning. And then deploying an adaptive token pruning module, carrying out token pruning retraining and weight updating, and finally deploying the neural network adopting the pruning method to a hardware platform. According to the method, the parameter quantity and the calculation quantity of the ViT network are effectively reduced by combining hardware-friendly weights and tokens with a pruning method, the reasoning delay and the hardware overhead of the ViT are reduced, high precision is kept, and an optimization method is provided for deployment of the ViT network at a resource-limited edge end.

The invention discloses a large model reasoning acceleration method and device and a medium, and the method comprises the steps: constructing a training set and a test set for training a large model according to scene data; expanding a prediction head in the large model to obtain an expanded target large model, and freezing original prediction head parameters in the large model so as to update the prediction head parameters corresponding to the expanded prediction head in the training process of the target large model; training and evaluating the target large model through the training set and the test set to obtain a trained target large model; constructing a calibration data set formed by partial scene data, and inputting the calibration data set into the target large model for forward reasoning to obtain a calibration matrix corresponding to each linear layer; and determining an input matrix and an output matrix corresponding to the linear layer according to the calibration matrix, and quantizing a weight corresponding to the linear layer based on the input matrix to obtain a quantized model weight.

The invention discloses a hierarchical quantization and dynamic priority management-based model cache optimization method and device. The method comprises the following steps of: acquiring an importance score of each Token of a word sequence corresponding to a reasoning long text by utilizing a mask language model; constructing a Token importance scoring model, and dividing the cloud platform GPU into a head cache region, a middle cache region and a tail cache region; the service reasoning model stores the key value pairs of the first # imgabs0 # Token in a Token reasoning sequence corresponding to the reasoning long text in a head cache region in an FP16 format, stores the key value pairs of the first # imgabs1 # Token before the current processing Token in a tail cache region in the FP16 format according to an FIFO principle, and stores the key value pairs shifted out of the tail cache region in an intermediate cache region in an FP8 format; and after the intermediate cache region is full, updating the intermediate cache region by using the Token importance scoring model and the importance score. According to the method, the reasoning efficiency of the large language model in long text generation can be improved.

The invention relates to the technical field of computers, and discloses a model quantification method and device, a storage medium and a program product. The method comprises the following steps: in an off-line quantization stage, calculating to obtain a leveling coefficient according to an input matrix and a weight matrix of each linear layer operator, and calculating to obtain an input quantization parameter and a weight quantization parameter according to the leveling coefficient; performing 8-bit floating point quantization, grouping and 4-bit unsigned shaping asymmetric quantization on the weight matrix according to the weight quantization parameter to obtain a scaling coefficient, a zero reference value and a 4-bit unsigned shaping quantization weight; and in the online reasoning stage, performing online reasoning on the reasoning task based on the leveling coefficient, the input quantization parameter, the weight quantization parameter, the scaling coefficient, the zero reference value and the 4-bit unsigned shaping quantization weight to obtain a reasoning result. According to the scheme of the embodiment, the hardware memory occupied by the large language model can be reduced, the model precision loss can be reduced, and the model reasoning efficiency can be improved.

The invention discloses a three-degree-of-freedom macro-micro composite piezoelectric nano probe station, and relates to the technical field of semiconductor manufacturing and biotechnology. In order to solve the technical problems that in the prior art, an existing motor-driven probe station is large in cooperative control difficulty, the positioning precision and the stability are difficult to consider, the structure is complex, and the energy consumption is high, the technical scheme provided by the invention is that the three-degree-of-freedom macro-micro composite piezoelectric nano probe station is provided. Comprising a probe bench surface, an XY macro-motion mechanism, an XY micro-motion mechanism and a Z-direction mechanism. The XY macro-motion mechanism comprises an X piezoelectric motor and a Y piezoelectric motor which are used for achieving dozens of millimeter level horizontal movement in the X-axis direction and the Y-axis direction respectively. The XY micro-motion mechanism is used for realizing accurate movement of several nanometers to dozens of nanometers in the X-axis direction and the Y-axis direction; the Z-direction mechanism is used for realizing dozens of microns of lifting in the Z-axis direction; the probe bench surface is used for bearing a bearing object and is also used for driving the bearing object to move along the X-axis and Y-axis directions. The device is suitable for semiconductor manufacturing and biotechnology testing work.

The invention discloses a method, a system and equipment for dynamic compensation of PCB (Printed Circuit Board) processing equipment and a medium. The method comprises the following steps: monitoring and collecting actual temperature data of parts of the PCB processing equipment in work; determining a temperature interval to which the actual temperature data belongs according to the actual temperature data; acquiring an expansion coefficient of the component according to the temperature interval; calculating the actual displacement required to be compensated through the expansion coefficient; and performing motion compensation on the corresponding part according to the displacement. According to the invention, the position feedback data is compensated under the most suitable actual expansion and contraction coefficient matched with the temperature and the compensation mode, and the compensation precision is effectively improved. Through the temperature monitoring and compensation technology, the precision loss of the PCB processing equipment caused by temperature change is remarkably reduced, the overall performance of the equipment is improved, the service life is prolonged, the PCB processing equipment is particularly suitable for precision processing occasions, and the product quality and the production efficiency are improved.

The invention provides a multidisciplinary airfoil profile inverse design method based on a deep neural network, and the method comprises the following steps: 1, selecting an airfoil profile as a design target, and selecting the aerodynamic low detection performance of an airfoil profile as a design variable; the profile of the airfoil is the vertical coordinate of the airfoil; determining a pneumatic and low-detectivity design state; 2, selecting a set number of airfoils from an existing airfoil library, and performing horizontal coordinate unified interpolation processing on the airfoils to obtain vertical coordinate values of the airfoils with unified horizontal coordinates; 3, calculating the airfoil profile after interpolation to obtain pneumatic and low detection performance of the selected airfoil profile in a design state; 4, establishing a plurality of deep neural network agent models according to the airfoil profile ordinate value obtained in the step 2 and the airfoil profile pneumatic and low detection performance obtained in the step 3; and step 5, predicting the airfoil profile with given pneumatic and low detection performance according to the deep neural network agent model obtained in the step 4, and obtaining the airfoil profile appearance meeting the pneumatic and low detection target performance.

A fully automatic press quenching tool for special-shaped driven bevel gears, used in the field of mechanical manufacturing technology. The centering aid is installed in the center bottom hole of the quenching press worktable, and its positioning rod is pressed against the inner wall of the center bottom hole of the quenching press worktable; the centering fixture is threadedly connected to the center rod of the quenching press, and the centering fixture and the centering aid body are hole-shaft matched, and the centering fixture is in contact with the inner hole of the part to center it; the outer die is pressed into contact with the tooth surface of the part. The present invention uses a centering aid to guide the part, which can ensure the concentricity of the center of the part and the center of the quenching press worktable, realizes a fully automatic press quenching process, effectively solves the quality problem of unstable gear heat treatment deformation, and can ensure the consistency of gear heat treatment quenching deformation. The gear heat treatment deformation is significantly improved, and the flatness deformation is 0.06mm-0.1mm. The tooth profile accuracy after heat treatment is good and can reach level 7-9, which effectively reduces the attenuation of tooth profile accuracy, has low labor intensity, and high operating efficiency.

The invention provides a method and equipment for realizing smooth switching between a ground-based enhanced positioning resolving mode and a satellite-based enhanced positioning resolving mode, a user side takes the ground-based enhanced positioning resolving mode as a main mode of enhanced positioning, judges whether ground-based enhancement is available or not according to a preset abnormal condition detection index, and smoothly switches to satellite-based enhancement if the ground-based enhancement is not available; when a main mode is switched from ground-based enhancement to satellite-based enhancement, an asynchronous positioning mode with additional time accumulation error compensation is adopted, the stability of the positioning precision of a user side is maintained in combination with deviation compensation, and meanwhile, the satellite-based enhancement is restrained and the convergence speed of the satellite-based enhancement is increased in combination with ground-based enhancement solution information; and when the main mode is switched from the satellite-based enhancement to the ground-based enhancement, the ground-based enhancement is constrained by using prior information provided by the satellite-based enhancement, and the restarting speed of the ground-based enhancement positioning mode is improved. According to the method, the continuous smoothness and positioning consistency of ground-based and satellite-based enhanced positioning resolving mode switching are improved by optimizing switching, and the reliability of the Beidou enhanced positioning technology can be improved.

The invention discloses a Sigma Delta modulator introducing element mismatch information. The modulator is composed of a three-stage discrete time integrator, a three-bit quantizer, a DWA and feedback DAC, a two-stage digital integrator, a one-bit quantizer, a counter, a Pointer Generation algorithm and a memory. According to the invention, the element mismatch information of the feedback DAC is introduced into the digital output of the Sigma Delta modulator, so that the signal-to-noise ratio is 116.90 dB, the signal-to-noise harmonic distortion ratio is 115.69 dB, the significant digit is 18.93 bits, and the spurious-free dynamic range is 124.71 dB. Compared with an ideal model, after feedback DAC element mismatch is added, the signal-to-noise ratio of the Sigma Delta modulator output scheme provided by the invention is far greater than that of a conventional output scheme. Compared with a traditional structure, the Sigma Delta modulator has the advantages that the signal-to-noise ratio advantage is obvious, and the precision of the Sigma Delta modulator can be effectively improved.

The mechanical arm is provided with a shoulder joint center, an elbow joint center and a wrist joint center, and comprises at least four power sources arranged in parallel in the horizontal direction; the arm body comprises a large arm, a small arm, a mechanical clamping jaw and a power shaft, the power shaft comprises a power input part and a power output part which are located at the two ends of a shaft body, and the power input part is in transmission connection with the at least four output ends of the kinetic energy assembly; the transmission assembly is in transmission connection with the kinetic energy output part; the transmitted kinetic energy is used for driving the large arm to rotate around the center of the shoulder joint in the horizontal direction, driving the small arm to rotate around the center of the elbow joint in the horizontal direction, driving the mechanical clamping jaw to rotate around the center of the wrist joint in the horizontal direction and driving the mechanical clamping jaw to clamp a wafer conveying box. By the adoption of the scheme, the requirements for multi-angle large-torque movement in a conveying window and overall installation space limitation are met.

This utility model discloses a single-station hot plate welding machine, belonging to the field of hot plate welding technology. The hot plate welding machine includes a cabinet, an upper mounting part, a lower mounting part, and a heating component. The movement of the upper and lower mounting parts is driven by first lead screws driven by first servo motors on both sides. The synchronous control of the dual motors can achieve precise synchronization of movement in opposite directions, and the lead screw drive can provide continuous and stable mold closing pressure. Compared with the traditional slide bar drive, it avoids welding gaps caused by pressure fluctuations, ensures tight contact of the workpiece welding surface, and improves welding strength and consistency. The linear slide rail with built-in ball bearings, cage, and return channel is used in conjunction with the sliding block to replace the sliding friction of the traditional slide bar. The low friction coefficient of rolling friction can significantly reduce the wear of the slide rail and the slider, avoiding the problem of precision decay caused by the thinning of the slide bar and the enlargement of the sliding hole. At the same time, the multi-directional contact design of the slide rail can limit the movement of the upper and lower mounting parts, ensuring accurate workpiece alignment during mold closing.

The embodiment of the present invention provides a high-speed and high-precision single-drive gantry structure, including an equipment frame, a first crossbeam assembly and a second crossbeam assembly, wherein the first crossbeam assembly and the second crossbeam assembly are slidably arranged on the equipment frame, and the equipment frame includes a first active side guide rail and a first driven side guide rail, and the first crossbeam assembly is slidably connected to the first active side guide rail and the first driven side guide rail at both ends along the X direction. The high-speed and high-precision single-drive gantry structure provided by the present invention can carry the first crossbeam assembly and the second crossbeam assembly at the same time. Here, the two-layer structure has good casting stability and meets the processing technology requirements for high precision requirements, and greatly reduces the precision loss caused by the impact of the acceleration and deceleration process during high-speed movement.

The invention provides an intelligent super surface wave angle of arrival estimation method and device, and the method comprises the steps: dividing a search space into a plurality of grids, and enabling each grid to correspond to a codebook; traversing all codebooks by using an intelligent metasurface array, and counting power information reported by a receiving end; screening maximum power information in the power information, and determining a preliminary direction of arrival according to the maximum power information; dividing the intelligent metasurface array into a plurality of sub-arrays, and performing phase assignment on each array of the sub-arrays based on the initial direction of arrival; and calculating the direction of arrival according to the assigned phase value of each oscillator. According to the invention, the problem of inaccurate array phase acquisition of the RIS under the condition of weak coverage is effectively solved, and high-precision DOA estimation of the RIS is realized.

The application provides a hardware perception-based model quantization method, system, medium, product and terminal, a hybrid execution mechanism is constructed between a high-level software framework and a bottom layer hardware behavior model, only core calculation operators in a quantization calculation graph of an initial quantization model which need to be executed on target hardware are handed over to a hardware behavior model for bit precision simulation, real calculation characteristics of the target hardware are simulated in a quantization process, and other auxiliary, non-computation-intensive operators are reserved in the high-level software framework for execution, so that the simulation verification efficiency is improved, the simulation precision and the execution efficiency are considered, and the goal that the quantized model runs more efficiently and has lower precision loss on the target hardware is achieved.

This invention provides an approximate 4-bit lookup table multiplier, relating to the field of integrated circuits, comprising: a conversion module, a first-stage multiplication module, and a second-stage multiplication module. During the multiplication process, for two compressed partial products with the same input, a 6-input, 2-output lookup table (LUT6_2) is used for simultaneous calculation. For three compressed partial products with the same input, the precision loss of each compressed partial product is calculated, and a 6-input, 2-output lookup table (LUT6_2) is used to simultaneously calculate the two compressed partial products with the smallest precision loss. The compressed partial product with the largest precision loss is calculated separately using a 6-input lookup table (LUT6). For compressed partial products with an input number exceeding the lookup table's input limit, and for unprocessed partial products, a partial carry-truncation method is used to calculate the product result. This invention can improve multiplier performance while reducing multiplier precision loss.

The invention provides a key value caching method and device, electronic equipment and a storage medium. The method comprises the steps of obtaining a plurality of text segments corresponding to a current text and a segment attention score corresponding to each text segment; based on the fragment attention score, screening at least one target text fragment from the plurality of text fragments; performing attention analysis on the plurality of candidate lexical elements in the at least one target text fragment to obtain first attention scores corresponding to the plurality of candidate lexical elements; based on the first attention score, screening out at least one target lexical element from the plurality of candidate lexical elements; and updating the original cache based on the key value cache corresponding to the at least one target lexical element. By utilizing the embodiment of the invention, the recall of the key value cache of the target lexical element can be realized, the information loss caused by relatively far lexical element information is avoided, the reasoning of relatively long context can be supported, and the problem of precision loss caused by information loss in a rarefaction process is alleviated.

The invention discloses a virtual scene real-time color correction method and system, and relates to the technical field of color correction, and the method comprises the steps: presetting color space projection and capture, and carrying out the generation of an RGB channel mapping relation; performing parallel processing on the color area of the LED screen; the bit depth reduction and the color grading mapping are adaptively matched according to the regional color type, and the color index is quickly calculated through displacement operation to carry out real-time correction of the color deviation. According to the virtual scene real-time color correction method provided by the invention, the RGB color space is divided into the non-overlapped triangular grids through the Delou triangulation algorithm, the color distribution of the three-dimensional space is effectively processed, the smooth transition between the RGB colors is ensured, the mapping relation of each color point is ensured to have high continuity and consistency, and the real-time color correction of the virtual scene is realized. Abrupt change and fault in color mapping are prevented, and the smoothness and precision of color correction are improved.

The utility model relates to a kind of mobile press pin machine devices, including product placement unit, cylindrical pin positioning guide unit and mobile press pin machine unit;Wherein product placement unit includes the mobile trolley of placing camshaft product;Cylindrical pin positioning guide unit includes cylindrical pin positioning guide sleeve, locking screw and cylindrical positioning bolt;Mobile press pin machine unit includes product positioning seat, press pin head, gas-liquid pressure cylinder front seat, gas-liquid pressure cylinder, gas-liquid pressure cylinder rear seat, cylindrical guide rail, flange linear ball bearing, gas-liquid pressure cylinder bottom plate, round flange seat, lifting mechanism bottom plate frame, directional wheel, universal wheel, cylinder, bevel gear elevator, hand wheel, handle, pneumatic element bottom plate, gas source two joint parts, solenoid valve and controller.The utility model can improve the precision and efficiency of camshaft workpiece press-fit positioning pin, and the utility model can reduce workpiece collision risk and personnel safety hazard.

The invention provides a vehicle-mounted wheelset measurement sensor, which is installed on the inner side of a wheel and comprises a camera, a laser and a controller; the controller comprises a laser control module, an acquisition module, a tread defect detection module, a tread information storage module and a communication module; the camera acquires a laser bar image; the tread defect detection module acquires point cloud information, searches for points whose coordinate values ​​exceed a normal coordinate interval in a tread area of ​​the laser bar, marks them as defective points, and transmits the coordinates of the defective points to the tread information storage module; the tread information storage module counts the number of defective points and issues a prompt when an abnormality occurs; the invention also comprises a rim size measurement module connected to the acquisition module and the rim information storage module; the rim size measurement module acquires point cloud information, calculates rim geometric dimensions using the point cloud information, and monitors rim status; the sensor of the invention adopts a vehicle-mounted manner to dynamically and real-time monitor the conditions of the wheelset tread and rim, and timely discovers wheelset abnormalities, with high safety.

The present invention relates to a pixel circuit system, a driving method of the pixel circuit system and a display device. The pixel circuit system comprises: a video processing module, which is used to generate a pixel circuit driving signal according to video source data; a pixel circuit module, which is connected to the video processing module, and is used to separate the high-order signal and the low-order signal in the pixel circuit driving signal, and respectively perform digital-to-analog conversion on the high-order signal and the low-order signal to obtain a high-order voltage value and a low-order voltage value; according to the high-order clock frequency and the low-order clock frequency, the high-order voltage value and the low-order voltage value are respectively compared with the corresponding ramp voltage data, and a high-order pulse width modulation signal corresponding to the high-order voltage value and a low-order pulse width modulation signal corresponding to the low-order voltage value are output. The pixel circuit system can reduce the loss of display grayscale accuracy without increasing the design area and power consumption of the original pixel circuit system.

The present invention proposes a model generation and real-time construction method for unsupervised learning models, including: step 1, obtaining the original model; step 2, generating a model block, obtaining a derived model block according to the model block, and constructing a training environment for each derived model block; step 3, training the derived model block and updating the parameters of the derived model block; step 4, evaluating all model blocks to quantify the performance indicators and use value of the model block; step 5, modeling the edge reasoning scenario, constrained optimization problem; step 6, solving the constructed optimization problem; step 7, generating a new model according to the solved model block selection scheme, and scaling the model. The method avoids the influence of instability in unsupervised learning on the generation of derived models, ensures the high precision of derived models; reduces the offline generation cost of derived models in unsupervised learning and the online cost of real-time scaling models in edge reasoning.

The utility model discloses a photovoltaic power station support rotating part of automatic lubrication function belongs to photovoltaic power station support technical field, it includes base, the base upper end surface center place is equipped with stand, the stand upper end surface center place is equipped with rotating structure, and the rotating structure includes rotating frame, and the both sides wall center upper of rotating frame is equipped with bearing evenly, and the rotating frame inside center upper is equipped with rotating rod, in addition, the utility model discloses, realize through the linear motion of hydraulic telescopic link efficient stably conversion is driven support main body rotation's circular motion to this adjustment photovoltaic board angle better receive sunlight, has reduced equipment cost, has reduced the precision loss of the intermediate transmission link, just control the telescopic amount of hydraulic telescopic link, can accurate control rotating frame and support main body's rotation angle, guarantee photovoltaic board can continuously track sunlight, effectively improve power generation efficiency.

The present invention provides a lightweight pulse neural network single photon radar back-end signal processing method, including: simulating the optical and electrical processes of a SPAD sensor to collect images, calculating the photon trigger probability of the SPAD sensor; using the photon trigger probability and the grayscale image data set to generate pulse sequence data that obeys the Poisson distribution, and then training the constructed pulse neural network to obtain a trained pulse neural network; quantizing the network parameters of the trained pulse neural network to obtain a lightweight pulse neural network, and then deploying it to a single photon radar signal processing module, and predicting the pulse data collected by the SPAD sensor to obtain classification or imaging results. The present invention can minimize the precision loss of quantization by quantizing the network parameters of the pulse neural network, and converting the network parameters into integers, which can reduce the overhead of computing and hardware resources, and realize high-accuracy, high-efficiency, and high-dynamic range processing of on-chip pulse signals.

This invention discloses a convolutional neural network pruning method and apparatus based on feature similarity and feature compensation. The method includes: adjusting the size of the input images in the dataset to a fixed size, standardizing the pixel values ​​of the images, and increasing the training data using image enhancement techniques; initializing the network structure and setting model parameters, and training the model using the training data; for the trained model, obtaining the similarity between convolutional kernels, performing cluster analysis based on the similarity, grouping each layer of convolutional kernels into similar groups, selecting the convolutional kernels to be retained in each similar group, and generating a new network structure; copying the parameters of the retained convolutional kernels into the new network structure, and compensating the weight parameters of the pruned convolutional kernels in each similar group into the retained convolutional kernels through parameter superposition; performing model accuracy recovery training using the original dataset, and saving the model parameters and network structure. This pruning method can maintain model accuracy.