38 results about "Stochastic geometry" patented technology

The invention relates to the technical field of fault diagnosis, and discloses an intelligent fault diagnosis method and system for a multi-channel gas sensor, and the method comprises the steps: carrying out the random geometric analysis and response limitation region construction of a multi-channel gas sensor array, and obtaining an effective sensor set; performing correlation analysis and sensitivity precision dilution measurement function screening to obtain an optimal sensor combination; performing periodic sampling and feature extraction to obtain a feature data matrix; performing deep Q network analysis to generate a fault diagnosis trigger signal; performing sensor channel block mapping and position coding on the feature data matrix according to the fault diagnosis trigger signal to obtain a space-time separation feature vector; and inputting the space-time separation feature vector into a bidirectional long-short-term memory network to perform fault type determination, and outputting a fault type determination result, thereby realizing accurate identification and classification of different types of faults of the multi-channel gas sensor.

The invention discloses an unmanned aerial vehicle transformer substation inspection path planning method and device, equipment and a medium. The method comprises the following steps: firstly, generating a grid map; dividing the substation inspection space into a plurality of sub-regions; establishing an electromagnetic exposure intensity model, and setting an electromagnetic exposure intensity safety threshold; randomly sampling a plurality of nodes in each divided sub-region, constructing a random geometric graph, and generating an initial inspection path by adopting an RRT algorithm; optimizing the generated initial inspection path by using a path pruning optimization algorithm and a gradient descent convex optimization algorithm, and smoothing the path into a continuous curve by using a B-spline curve fitting technology to obtain a final optimized inspection path; according to the invention, through an overall strategy of first region division and then path planning, random geometric graph tree building, path pruning optimization, gradient curve smoothing and key region guiding sampling mechanisms are fused, and the efficient, stable and high-environment-adaptability unmanned aerial vehicle three-dimensional path planning method for the substation complex scene is realized.

The invention provides a brain tumor classification method based on deep learning, and the method comprises the steps: uniformly segmenting a multi-mode magnetic resonance imaging image into a plurality of small blocks, generating an enhanced puzzle through random geometric transformation, inputting the enhanced puzzle into a ResNet50 backbone network, extracting primary local features, and optimizing feature distribution; multi-scale feature maps of different stages are obtained, lightweight convolution blocks are constructed, and local pathological relevance is enhanced; performing full-connection layer splicing on the multi-scale feature maps in different stages, introducing a learnable weight matrix to dynamically allocate weights in each stage, and generating fusion features; a progressive parameter unfreezing mechanism is adopted to ensure orderly learning of the model from local to global; the sample weight is dynamically adjusted according to the category frequency, and the local discriminant force and the global consistency are balanced in combination with multi-stage classifier loss weighted fusion; and a final classification result is generated through multi-stage prediction probability weighted average and dynamic threshold adjustment, and clinical availability is improved in combination with a sigmoid calibration module.

The embodiment of the invention provides an intelligent reflecting surface coverage enhancement method and device for electric power inspection, and the method comprises the steps: building an intelligent reflecting surface and inspection equipment unified scheduling model, and building a unified scheduling model for a multi-intelligent reflecting surface networking network; deducing a signal interference distribution analytic expression, solving a space measure function of a point process and determining heterogeneous station density based on an interference level; constructing a multi-intelligent reflector networking correlation model, and analyzing a link coupling factor; based on modeling of a link coupling factor, the link performance under the collaborative networking is studied; a cross-domain reconfigurable capability boundary is explored, and a spatial statistical average analysis method for coverage, capacity and mobility performance is innovated; an adjustable intelligent reflecting surface distribution dependence model is provided, and a semi-closed theoretical analysis expression of the user coverage rate is given based on a stochastic geometric tool.

The invention discloses a two-dimensional angle estimation method for a random array in a non-uniform noise scene, and the method comprises the steps: obtaining response signals collected by an EMVS array in random geometric distribution, and obtaining a covariance matrix of the response signals; removing the covariance influence of non-uniform noise in the covariance matrix to obtain a noise-free signal covariance tensor estimated value; the noise-free covariance matrix is recovered based on tensor filling, and a third-order PARAFAC tensor is obtained; and decomposing the third-order PARAFAC tensor based on a parallel factorization algorithm to obtain an azimuth angle and a polarization angle of the source signal. According to the method, the two-dimensional angle of the random array in the non-uniform noise scene is estimated based on tensor filling and parallel factorization, the two-dimensional direction-of-arrival estimation precision in the non-uniform noise environment can be improved while the aperture of the array is kept, and the defects in the aspects of noise suppression, calculation complexity and estimation precision in the prior art are overcome.

The invention relates to the technical field of robot motion planning, in particular to a robot motion planning method and device inspired by spatial relationship memory. The method comprises the steps that according to motion starting position information, motion target position information and map information, a first random geometric graph is constructed through a K-NN algorithm by using a selective sampling strategy in a configuration space of a robot; based on a preset confidence threshold, inputting the first random geometric graph and the obstacle position information into an intelligent edge selector for edge screening to obtain a second random geometric graph; according to the second random geometric graph, heuristic value calculation is carried out through a memory perception predictor, and a node heuristic value is obtained; and according to the node heuristic value and the second random geometric diagram, performing optimal path search by using a micro # imgabs0 # module to obtain a second optimal path. The method is an efficient and accurate robot motion planning method based on human brain space relation memory inspiration.

The invention discloses a method for analyzing the performance of an MIMO (Multiple Input Multiple Output) heterogeneous network under the influence of a strong electromagnetic pulse, and relates to the technical field of wireless communication network performance analysis. Considering the base station degradation condition of damaged partial antennas in partial base stations under the influence of strong electromagnetic pulses; based on stochastic geometry, constructing a spatial distribution model of MIMO heterogeneous network surviving base stations and degraded base stations under the strong electromagnetic pulse influence; deriving a coverage probability and lowest area spectrum efficiency estimation formula of the network, and calculating the coverage probability and the lowest area spectrum efficiency of the MIMO heterogeneous network under different base station degradation probabilities and antenna damage probabilities; the specific influence of the base station degradation probability and the antenna damage probability on the coverage performance of the MIMO heterogeneous network and the lowest area spectrum efficiency can be determined, and a thought is provided for researching the comprehensive action effect of the strong electromagnetic pulse on the information network and pre-disaster deployment or post-disaster recovery.

The invention discloses a random geometry-based cellular network signal-to-interference ratio distribution analysis method, which comprises the following steps of: constructing a cellular space node distribution, service time arrival and equipment energy state three-dimensional model; therefore, a receiving end signal to interference ratio expression, a conditional transmission success probability of the equipment and an overall active probability of the equipment in a stable state are obtained, a moment generation function of a final transmission success probability is obtained by utilizing random geometry, and finally an expression of cellular network signal to interference ratio element distribution is obtained by utilizing a Gil-Pelaez theorem. And the proportion of the successful transmission probability of the equipment in the cellular network can be analyzed by using the expression. The method disclosed by the invention is applied to a large-scale cellular network with randomly distributed equipment positions, the characteristics of equipment energy, storage and the like are considered, the system performance can be effectively evaluated, and the influence of parameters such as a signal-to-interference ratio threshold value and an energy arrival probability on signal-to-interference ratio element distribution is analyzed.

A double-lane millimeter wave vehicle networking safety transmission method, system, device and medium based on STAR-RIS, the method comprising: constructing a STAR-RIS assisted double-lane millimeter wave vehicle networking communication scene; based on the propagation characteristics of millimeter wave signals, combining the channel fading characteristics under the conditions of line-of-sight and non-line-of-sight transmission and the transmission and reflection functions of STAR-RIS, calculating the channel gain of the cascaded channel of the STAR-RIS assisted link; based on the random geometry theory, respectively calculating the probability density functions of the associated distances of the direct link, the STAR-RIS reflection link and the STAR-RIS transmission link; constructing a dynamic link association strategy supporting ternary selection; establishing a signal-to-interference-and-noise ratio model under three association modes; combining the physical layer security performance index definition, calculating the closed-form analytical expressions of the connection interruption probability and the secrecy interruption probability, and verifying the security gain of STAR-RIS; the present application has the advantages of effectively overcoming the half-space coverage blind area of the traditional reflective RIS, and improving the link connection reliability and physical layer safety transmission performance under the double-lane complex shielding environment.

The application discloses an RIS-assisted zero-forcing beamforming method based on a physical layer security technology, and the method comprises the following steps: constructing an RIS-assisted downlink MISO security transmission system model, constructing a passive eavesdropping user mathematical model based on random geometry, and constructing a legal user and eavesdropping user received signal-to-noise ratio expression; constructing a legal user received signal-to-noise ratio maximization optimization problem model, including determining optimization variables, an objective function and constraint conditions, and pre-designing a transmission beamforming vector based on a zero-forcing algorithm; constructing a joint optimization transmission beamforming and reflection beamforming vector algorithm, dividing the legal user received signal-to-noise ratio maximization optimization problem into two sub-problems for solving, and optimizing the reflection beamforming vector at the RIS and the transmission beamforming vector at the base station. The application jointly optimizes the transmission beamforming vector and the reflection beamforming vector, and realizes the optimization target of legal user received signal-to-noise ratio maximization.

According to the intelligent reflecting surface auxiliary uplink transmission method and equipment based on power control, the signal is enhanced by adjusting the intelligent reflecting surface unit, so that the coverage performance of uplink transmission is effectively improved. Specifically, the invention provides a power control model for the uplink transmission of the IRS auxiliary network, the constraint relation between power control and IRS deployment is represented, and uplink signal power distribution is deduced again through stochastic geometry. According to the invention, the power control uplink transmission of the ground user in the IRS auxiliary network under the influence of the IRS cascade channel is analyzed. In addition, an analysis framework is established by using a Poisson point process, and a theoretical expression of a network performance index in an uplink of the IRS-assisted network is calculated. The network performance of an IRS auxiliary uplink is reflected by measuring coverage rates under different system parameters, the deployment density and the number of IRS service units when the joint coverage performance is optimal are obtained, and the tradeoff relation between IRS auxiliary network system deployment parameters and uplink power compensation is represented.

The invention discloses a region-centered unmanned aerial vehicle coverage rate analysis method. The method comprises the following steps: constructing a region-centered unmanned aerial vehicle cooperative transmission model; fractional power control is carried out on the service unmanned aerial vehicle and the interference unmanned aerial vehicle; calculating the signal-to-interference ratio of the typical user based on the transmitting power of the service unmanned aerial vehicle and the interference unmanned aerial vehicle; based on the stochastic geometry theory, the coverage probability of typical users is obtained through service distance distribution, signal power distribution and interference distribution in combination with a Cauchy-Schwarz inequality and Gamma distribution approximation; the unmanned aerial vehicle density, the power control coefficient, the path loss index and the Nakagami fading parameter are set, the influence of the number of different service unmanned aerial vehicles, the power control coefficient and the channel condition on the coverage probability is analyzed, and simulation verification is carried out. According to the method, the technical problem of how to suppress the network interference of the unmanned aerial vehicle through power control in a complex interference scene is solved, and the network coverage rate of the unmanned aerial vehicle is calculated.

The invention provides an ultra-dense network multi-beam directional transmission performance analysis method and system based on stochastic geometry, and the method comprises the steps: carrying out the network performance index analysis of the coverage probability, the unit area spectrum efficiency and the energy efficiency for the ultra-dense network multi-beam directional transmission when a preset condition is satisfied. Modeling is carried out according to network deployment, channel characteristics, directional transmission multi-beam characteristics and the like in the ultra-dense network, expressions of the coverage probability, the unit area spectrum efficiency and the energy efficiency of the network are given through a random geometric method, and the network performance can be effectively evaluated.

The application discloses a random geometry modeling three-dimensional data set generation method, device, medium and product, relates to the technical field of data set generation, and comprises the following steps: for any three-dimensional geometry under the current cycle number, randomly generating a depth, a space parameter and a size in a set numerical range, so as to generate an initial three-dimensional geometry in a three-dimensional scene; randomly generating a noise texture and a target energy value; applying the noise texture to the corresponding initial three-dimensional geometry respectively to obtain an initial three-dimensional scene; in the initial three-dimensional scene, setting a sunlight light source with the target energy value to obtain a target three-dimensional scene; rendering the target three-dimensional scene to obtain an image data group, adding the image data group into an initial data set to obtain a target data set, and performing the next cycle until a maximum cycle number is reached, so that a random geometry modeling three-dimensional data set is obtained. The application realizes the rapid generation of the random geometry modeling three-dimensional data set.

The present invention proposes a power control method, system and storage medium for dynamic spectrum allocation, which includes building a distributed SU power control model of several unit areas based on existing users IU, secondary users SU, wireless spectrum sharing coordination center SAS, and wireless spectrum sharing perception engine ESC; using ESC sensors, combining random process derivation and random geometric modeling to calculate the SU interference aggregated in the unit area and the ESC local maximum allowable interference threshold, and calculate the ESC update threshold; based on the ESC update threshold, calculate the global parameters broadcast by the SAS coordination center to the secondary user transmitter SU‑TX, SU‑TX adjusts the transmission power of SU‑TX based on the global parameters, the signal attenuation from SU‑TX to ESC, and the signal-to-interference-to-noise ratio SINR of the secondary user receiver SU‑RX, and repeats the above steps until the system converges to a stable state. The present application successfully achieves multiple optimization goals of SU local power control with minimized energy consumption, IU location privacy protection, and system scalability assurance.

The present invention discloses a calculation method for an antenna pattern adapted to a wireless channel. The method is as follows: Discretely model the antenna pattern: Divide the spatial sphere into N equal parts, and within each part of the spatial angle, the radiation characteristics of the antenna are the same; The communication environment is the downlink of a base station and a mobile phone user. The number of antennas of the base station as the transmitting end is Tx, and the number of antennas of the mobile phone user as the receiving end is Rx. Under the random geometric channel model, establish the antenna pattern and the channel matrix, obtain the relationship between the channel capacity and the antenna pattern, and use the channel capacity as the performance index of the system; The antenna pattern that maximizes the channel capacity is defined as the best-adapted antenna pattern for the wireless channel, with the constraint that the total power of the transmitting antennas is P; Use SVD decomposition and the water-filling algorithm to calculate the optimal antenna pattern and determine the power allocation of each transmitting antenna; The optimal pattern adapted to the wireless channel is given, which can be used as the target pattern for antenna synthesis.

The invention discloses a satellite-ground fusion network cross-layer secure transmission method based on an optimal detection contact angle, and the method comprises the following steps: 1, collecting a plurality of parameters of a ground terminal, a service satellite and an eavesdropping satellite, and constructing a double-layer satellite satellite-ground fusion network model based on stochastic geometry; 2, constructing a signal model, and determining a reachable secrecy rate according to the signal model; 3, determining the security coverage probability and the average secrecy rate under different network layer information exchange degrees; and step 4, according to the optimal detection contact angle, dynamically adjusting an information exchange strategy of an inter-satellite link, informing related service satellites of an eavesdropping satellite range in which information needs to be acquired, implementing eavesdropping avoidance or resource scheduling by the service satellites in a known area, taking conservative security measures for an unknown area, and transmitting data to be transmitted. The method has the characteristics that the security performance of the physical layer is improved on the premise of not depending on the CSI of the eavesdropper, and the practicability and robustness of the security scheme are greatly improved.

The invention provides a network optimization method, device and equipment based on geometry and deep learning fusion, and belongs to the field of network resource allocation. The method comprises the following steps: generating random distribution of a macro base station, a road side unit and a vehicle by adopting a Poisson point process so as to construct a topological structure of a communication network; constructing a channel analytic function of the communication network according to the topological structure; deriving an analytic function to obtain a performance analysis result of the communication network; constructing a deep deterministic policy network by taking an analysis result as a precondition; in combination with a deep deterministic policy network and a near-end policy optimization algorithm, resource allocation and power control of the communication network are jointly optimized to obtain a preliminary communication network; and performing local optimization on the preliminary communication network according to the optimization limiting function to obtain a target communication network. By combining stochastic geometry, a deep learning algorithm and local optimization, the wireless network performance is effectively improved, the link outage probability is reduced, and the resource utilization efficiency is improved.

The invention discloses an SPMA protocol priority threshold design method based on stochastic geometry and a queuing theory, and belongs to the technical field of SPMA protocol performance optimization. The method solves the problems that an existing method cannot obtain an optimal scheme capable of maximizing the channel utilization rate and guaranteeing priority scheduling, and is poor in flexibility and adaptability and high in test cost. The random geometry and the queuing theory are fused, the position distribution of a transmitter and a receiver and the arrival process of data packets with various priorities are obtained based on the random geometry, the distribution obeyed by the service time of the data packets is obtained through the queuing theory, and then the constraint relation between the priority threshold value and the data packet sending success rate is determined through mathematical derivation. Theoretical guidance is provided for priority threshold design, more channel access opportunities are provided for low priority on the premise that high-priority transmission performance is guaranteed, and dynamic balance between maximization of channel utilization rate and multi-priority scheduling is guaranteed. The method can be applied to the priority threshold design of the SPMA protocol.

The invention belongs to the technical field of concrete models, and discloses a defect-containing concrete mesoscopic model construction method and related device.The method comprises the steps that characteristic parameters of a to-be-simulated concrete test piece are obtained; the characteristic parameters of the to-be-simulated concrete test piece comprise the geometric dimension of the concrete test piece, the characteristic parameters of the aggregate and the characteristic parameters of defects in the aggregate; generating a concrete matrix model based on the geometric dimension of the concrete test piece; based on the aggregate characteristic parameters, generating a random aggregate model by using a random geometric algorithm; based on the defect characteristic parameters in the aggregate, randomly generating an aggregate internal defect model; putting the random aggregate model and the aggregate internal defect model into a concrete matrix model to obtain a concrete mesoscopic model containing defects; according to the method, the concrete random aggregate model containing the defects is simply, conveniently and intuitively generated, the randomness of the aggregate and the defects in the model is improved, and the aggregate and the defects are closer to the form and distribution of real concrete aggregate.

The present invention is a method for generating random numbers using discrete randomized geometric graphs and cellular automaton mechanisms. Number generators often fail to provide unpredictable results which pass tests for statistical randomness. The present invention method increases unpredictability while meeting the tests for cryptographic random number generation.

This application provides a region-aware driven self-supervised image denoising method, system, apparatus, and medium, belonging to the field of image processing technology. The method includes: constructing an adaptive region-aware network framework, including an initial adaptive region-aware blind spot network, an adaptive hybrid scrambling branch, a non-scrambling branch, and a random geometric transformation pixel downsampling module; training the adaptive hybrid scrambling branch and the non-scrambling branch to generate a trained adaptive region-aware blind spot network; inputting a noisy image into the trained adaptive region-aware network to obtain a tested and verified adaptive region-aware blind spot network; and inputting the noisy image into the tested and verified adaptive region-aware blind spot network to obtain a denoised image. This achieves superior denoising performance.

The present invention discloses a method and system for calculating the optimal antenna spacing in a MIMO system. The method comprises the following steps: establishing a relationship between a channel matrix, an antenna vector matrix, and a multipath cluster environment matrix under a random geometric channel model; establishing a relationship between a transmitting antenna vector matrix, a departure angle, and a transmitting antenna spacing; establishing a relationship between a receiving antenna vector matrix, an arrival angle, and a receiving antenna spacing; in the MIMO system, comprehensively considering the condition number of the channel matrix and the sum of squares of the matrix's singular values as optimization targets, setting the case of omnidirectional antennas, calculating and solving the optimal antenna spacing to obtain the best MIMO channel; dividing the problem into three cases: the number of transmitting antennas is equal to the number of environmental clusters, the number of transmitting antennas is greater than the number of environmental clusters, and the number of transmitting antennas is less than the number of environmental clusters, and solving the three cases to obtain the optimal antenna spacing corresponding to the three cases; quantitatively analyzing the influence of the antenna vector matrix on the channel matrix condition number, and providing the optimal antenna spacing.

The invention discloses a random geometric modeling three-dimensional data set generation method and device, a medium and a product, and relates to the technical field of data set generation, and the method comprises the steps: for any three-dimensional geometry under the current cycle index, randomly generating depth, space parameters and size within a set value range, generating an initial three-dimensional geometry in the empty three-dimensional scene; randomly generating a noise texture and a target energy value; applying the noise textures to the corresponding initial three-dimensional geometries to obtain an initial three-dimensional scene; in the initial three-dimensional scene, setting a fluorescent lamp light source with a target energy value to obtain a target three-dimensional scene; and rendering the target three-dimensional scene to obtain an image data set, adding the image data set into the initial data set to obtain a target data set, and performing the next cycle until the maximum cycle index is reached to obtain a random geometric modeling three-dimensional data set. According to the invention, rapid generation of the stochastic geometric modeling three-dimensional data set is realized.

This invention discloses a relay selection method and system for vehicle-to-vehicle multi-hop cooperative communication systems. It constructs a stochastic geometric model of the vehicle network and determines the location and size of candidate relay areas based on a no-relay probability threshold and the communication service quality of relay links. All candidate relay vehicles within the relay selection area are acquired to obtain a candidate relay set. A relay vehicle is selected from the candidate relay vehicles according to a relay selection strategy that guarantees the service quality between the source node and the relay vehicle. The source node sends messages to both the relay vehicle and the destination node. After decoding the message, the relay vehicle forwards it to the next relay vehicle or destination node based on its relative position to the destination node. When the distance between the relay vehicle and the destination node exceeds the communication range, the currently selected relay node is used as the source node after one hop, and this process is repeated until the message is forwarded to the destination node. This invention effectively reduces the number of relay hops, making communication more reliable and reducing latency.

The present application relates to a kind of based on edge cache adaptive downlink non-orthogonal multiple access transmission method and system, BS is adaptively selected transmission mode according to user local cache hit situation, in three kinds of mode between mute transmission, single-user transmission and NOMA transmission, flexible switching, reduce redundant data transmission from source and improve spectral efficiency;When two users all need to receive data, user can adaptively carry out interference cancellation according to whether the content required by the other party is cached, realize the optimization of cache-aided interference suppression and signal decoding.At the same time, the present application constructs large-scale cellular network model based on random geometry theory, system analyzes the probability of user successfully obtaining content under different cache states and transmission modes, reveals the influence mechanism of cache state, power allocation and decoding mode on system performance;And carry out performance comparison and parameter optimization under multi-reference scheme, realize the maximization of system downlink throughput by jointly optimizing power allocation factor and time allocation factor.

The invention provides a region perception driven self-supervised image noise reduction method, system and device, and a medium, and belongs to the technical field of image processing, and the method comprises the steps: constructing a self-adaptive region perception network framework, comprising an initial adaptive region sensing blind spot network, an adaptive hybrid scrambling branch, a non-scrambling branch and a random geometric transformation pixel downsampling module; training the self-adaptive hybrid scrambling branches and the non-scrambling branches to generate a trained self-adaptive region sensing blind spot network; inputting a noise-containing image into the trained adaptive region sensing network to obtain an adaptive region sensing blind spot network after test verification; and inputting a noise-containing image into the self-adaptive region perception blind spot network after test verification to obtain a de-noised image. And a better denoising effect is obtained.

The application discloses a kind of leather surface flaw detection grading method and system based on AI vision, specifically includes: step 1: collection leather surface image data and divide leather surface analysis area, generate spatial position mark and texture feature vector;Step 2: construct random geometric model including spatial intensity parameter set, spatial interaction parameter set and flaw grade state variable set;Step 3: construct grading condition adaptive proposal distribution and generate sample candidate parameter set;Step 4: introduce grade consistency potential function to construct target probability distribution;Step 5: execute Markov chain Monte Carlo acceptance determination to generate parameter update result;Step 6: generate flaw spatial distribution state;Step 7: generate flaw grading determination result and output leather overall grading result.The application introduces random geometric constraint, realizes accurate grading.

This invention relates to the field of geometric modeling technology for marine pile foundation engineering structures, and provides a method for modeling stochastic geometric defects in tubular structures based on a node-layered continuous construction. The method includes the following steps: establishing a finite element geometric model of the tubular structure to be analyzed; constructing nodes and element meshes; dividing the nodes involved in geometric updates into node layer sets; establishing a mapping relationship between node spatial coordinates and parameter spatial coordinates; based on externally input stochastic geometric defect field data, performing two-dimensional interpolation calculations on the node parameter coordinates in the parameter space to obtain the target defect values ​​on the outer and inner surfaces; calculating a continuous construction relationship; geometrically updating the spatial coordinate information of the nodes; and generating a finite element model of stochastic geometric defects. This invention, by no longer treating finite element nodes as independent discrete update objects, provides a geometric basis for the continuous construction of defect geometry in the thickness direction through this layered structure.

The present invention belongs to the field of adversarial sample defense, and particularly relates to a two-stage adversarial defense method and system for image classification. The method includes: training multiple deep neural networks with different structures using the same training set to obtain multiple trained classifiers as alternative models; constructing an optimal model pool from the multiple alternative models; performing random geometric transformation on the adversarial images to destroy the specific spatial structure of the perturbations in the adversarial images; randomly selecting a classifier from the optimal model pool to classify the images after the random geometric transformation. The adversarial defense method proposed by the present invention constructs defenses from both the data and model levels, can more comprehensively resist adversarial images of various types and different perturbation intensities, and hardly affects the classification performance of the model for clean samples.