40results about How to "Efficient deployment" patented technology

The invention relates to a multi-agent cooperation-based language model reasoning resource scheduling method and system, and the method achieves the intelligent simulation of multi-view iterative thinking in a complex reasoning task through the construction of a multi-agent system which is clear in division of labor and is provided with a special knowledge base. View limitation and decision deviation of a single model in long-sequence and multi-step reasoning are effectively overcome; depending on the fusion of the general capability of the large language model and the task special knowledge base, the professionality and accuracy of the reasoning result are improved. Meanwhile, the problems of resource waste, redundant calculation and unstable convergence caused by cognitive overload in the cooperation process are solved by combining multi-round dynamic cooperation with a convergence mechanism of cognitive load perception and a dynamic token number limitation and low-confidence branch pruning strategy; therefore, on the premise that the reasoning quality is guaranteed, the utilization efficiency of computing resources is remarkably optimized, peak value occupation is reduced, the overall reasoning time delay is shortened, and reliable technical support is provided for efficient and stable deployment of a large language model in a complex task.

The application discloses a mixed expert network-based optical flow estimation method and system, belonging to the field of computer vision and artificial intelligence. First, the input two frames of images are preprocessed, low-resolution features are extracted through a mixed expert feature extractor (MoEE) containing a sparse activation mechanism to reduce redundant calculation; then, dot product operation is performed on the low-resolution feature maps to construct a 4D correlation volume to capture pixel motion matching relationship; subsequently, a mixed expert updater (MoEU) is used to iteratively update the hidden state and regress the residual flow increment through a dynamic expert selection mechanism to optimize the optical flow accuracy; finally, a multi-scale upsampling module is used to reconstruct the high-resolution optical flow field to output the high-precision optical flow result. The algorithm realizes dynamic resource allocation through the MoE architecture, significantly reduces the calculation cost while ensuring the accuracy of optical flow estimation, and can adapt to resource-constrained scenarios such as automatic driving and unmanned aerial vehicles, and can balance efficient inference and flexible deployment.

The application relates to a water quality change trend rapid prediction method based on multi-source data fusion and physical constraints, and specifically comprises the following steps: step 1: synchronously collecting data such as spectrum information, DO, COD, temperature and pH at key monitoring sites, constructing a water dynamics-water quality coupling equation, and simulating the space-time dynamic distribution of water quality parameters; step 2: outputting a water quality sensitive area through a water dynamics-water quality model, screening sensor layout points by combining information entropy evaluation and spatial clustering, realizing low-cost water quality sensor network deployment through a multi-objective optimization algorithm, and calculating a water body global water quality distribution map by adopting a spatial interpolation method; step 3: analyzing main pollution sources according to the synchronous collection of spectrum information at key monitoring sites, and adopting principal component analysis and an attention mechanism neural network; step 4: based on real-time optical characteristic value-DO data, combining a spatial topological network, a water quality gradient and a cross-region covariance to capture the spatial correlation of water quality parameters between different points, and constructing an optical characteristic value-DO-COD dynamic prediction model; combining pollution tracing and spectrum characteristic correction to correct the COD prediction value, adopting ensemble Kalman filtering to assimilate multi-source data, and improving the model accuracy.

This invention belongs to the field of industrial manufacturing technology, specifically providing a method and system for visual detection of industrial defects by decoupling defect features from imaging conditions. The method includes: collecting defect samples of industrial products under different operating conditions; wherein the defect samples include defect images and corresponding defect annotations; based on the defect samples, constructing a region-guided causal decoupling network model by combining a feature decoupling mechanism guided by defect regions and a causal invariance learning mechanism based on imaging simulation; performing multi-supervised loss joint optimization training on the causal decoupling network model to obtain a converged causal decoupling network model; and detecting defects in industrial products under target operating conditions based on the converged causal decoupling network model to obtain defect detection results. This invention can significantly improve the generalization and robustness of defect detection algorithms in complex industrial environments, enabling rapid and low-cost model transfer between different operating conditions.

This invention relates to a method and system for automated generation of TSN network configurations based on a reflective agent. The system comprises a network state perception and intent reflection module, a YANG model configuration parameter and structure decision module, and a YANG model intelligent generation module. The network state perception and intent reflection module is responsible for structurally perceiving network state and resource information. The YANG model configuration parameter and structure decision module integrates multi-source knowledge and contextual information, intelligently generating deployable TSN scheduling parameters and YANG configuration structures based on business needs. The YANG model intelligent generation module automatically generates the corresponding YANG configuration model based on the optimized results, performs syntax and semantic verification, and then distributes it to the target device, achieving automated deployment of network configurations. This invention significantly improves the intelligence level and deployment automation capability of TSN network configuration by introducing reflective reasoning and feedback optimization techniques.

The present application relates to the technical field of GPU computing optimization, and in particular to a sparse attention computing method, device and medium for GPU, wherein the method realizes the high efficiency of long context reasoning through the geometric perception sparse attention framework of ball hashing, and combines a large-scale parallel hashing optimization algorithm and a load adaptive computing kernel. Compared with the existing sparse attention methods based on heuristics or gradient learning, the present application realizes higher retrieval recall rate, lower preprocessing overhead and efficient hardware adaptation to irregular sparse patterns.

The application relates to the technical field of knowledge question answering, in particular to a knowledge question answering method based on a multi-modal document understanding large model of a graph structure, which comprises the following steps: obtaining a target task and a target document graph corresponding to the target task; generating a target heterogeneous document graph corresponding to the target document graph according to the target document graph; inputting the target task and the target heterogeneous document graph into a preset large model to obtain a target question and answer text corresponding to the target task; the method can achieve higher accuracy and robustness on multiple public benchmark tasks (such as document question answering, graph table question answering and table understanding); the method has excellent generalization ability for novel and complex document layouts; and the method realizes more simple and efficient system deployment and application through an end-to-end unified architecture.

The invention discloses a park weed accurate identification method and system based on an improved YOLOv11 deep convolutional network, and the method comprises the steps: 1, constructing a multi-scene park weed image data set, 2, carrying out the data enhancement and preprocessing, 3, constructing an improved YOLOv11 lightweight network model, 4, designing a self-adaptive loss function, and 5, carrying out the recognition of a multi-scene park weed image data set. Multi-stage model training and optimization are executed; and step 6, weed real-time identification and result output are realized. According to the method, a lightweight attention mechanism and an improved multi-scale feature fusion structure are introduced, the extraction and distinguishing capability of the network on small-scale weed features under a complex background is remarkably enhanced, the false detection and omission ratio is effectively reduced, and the dynamic label distribution strategy and a loss function fusing global context information are adopted, so that the robustness of the network is improved. And the detection robustness and the positioning precision of the model in dense, shielded and form-variable weed scenes are improved.

The application provides an image classification method, system, device and medium with explainability, obtains sample sub-blocks of a sample image and distinguishes flat regions and non-flat regions according to pixel brightness differences; color clustering is performed on the sample sub-blocks of the flat regions, shape clustering is performed on the sample sub-blocks of the non-flat regions, and positive and negative samples are independently processed to obtain clustering centers; a target sub-block of a target image is obtained and regions are distinguished; the responsiveness of the target sub-block is calculated based on the clustering centers; the responsiveness is quantified into an integer, and the distance between feature vectors is calculated by using city distance to generate a feature vector of the target image; and the target image is classified based on the feature vector. The application realizes internal mechanism transparency of the model, strong feature extraction pertinence and greatly reduced reasoning calculation complexity.

This invention discloses an adjustable activation neuron circuit and a multilayer inference acceleration device based on phase-change memory (PCM), relating to the field of micro-nano electronics technology. It includes: a neuron input terminal for receiving data to be activated; a neuron output terminal for outputting data after nonlinear activation processing; an activation parameter adjustment circuit connected to an activation function circuit, which includes: a first transmission gate, a second transmission gate, a third transmission gate, a fourth transmission gate, a PCM, and an operational amplifier. The inverting input terminal of the operational amplifier receives the data signal to be activated, the non-inverting input terminal is grounded, and the output terminal serves as the neuron output terminal. The activation parameter adjustment circuit is used to adjust the resistance state of the PCM through pulse signals, changing the nonlinear relationship between the output voltage and the input current. Based on the physical characteristics of PCM, this invention develops a neuron circuit with continuously adjustable activation functions, improving the data processing capability of neural networks.

The invention provides a malicious user identification method and system giving consideration to privacy protection in a social network, and relates to the technical field of network privacy security, and the method comprises the steps: carrying out the structure perception sub-graph segmentation of a social network graph through an METIS algorithm, dividing an original graph into a plurality of sub-graphs, minimizing the number of edges crossing the sub-graphs, and keeping the scale balance of the sub-graphs; constructing a privacy perception GNN of an integrated gating residual attention module, wherein the privacy perception GNN comprises a privacy perception linear layer and a gating residual mechanism; based on differential privacy stochastic gradient descent framework training, combining an adaptive noise scheduling strategy, dynamically adjusting the noise scale according to privacy consumption deviation, and performing closed-loop control budget to obtain a trained model; and malicious users are identified through the trained model. According to the method, the problem of performance reduction caused by fixed noise injection and noise amplification is solved, and efficient and robust identification of malicious users is realized while differential privacy constraints are met.

The application discloses a kind of equipment and control method for automatically fish slice defect finishing.The equipment includes fish slice single piece loading unit, image acquisition and processing unit and finishing unit.Fish slice single piece loading unit utilizes the action of gravity to disperse and load the fish slice overlapping to image acquisition area;Image acquisition and processing unit, image acquisition and processing unit are used to collect fish slice defect image, and real-time picture information is transmitted to industrial computer, and the fish slice defect detection network based on YOLOv8 model improved in the middle part of industrial computer is used to identify defect area and plan defect finishing track;Finishing unit receives track and posture information from industrial computer, and accurately finishes fish slice, so as to complete the process of fish slice loading, detection and finishing.The application replaces traditional manual finishing mode by automatic processing, realizes real-time finishing of fish slice defect, effectively improves the operation efficiency and product quality of fish slice processing.

This invention discloses a graph search-based farmland navigation path planning system for autonomous vehicles. The system includes: a farmland data collection component for acquiring farmland latitude and longitude information, starting ridge width, vehicle travel direction angle, vehicle turning anchor point position, and farmland size; a farmland structure partitioning component for partitioning the farmland structure, including starting ridge, ending ridge, far end, near end, number of ridges, turning anchor point position, and obstacle positions within the turning radius of the far and near ends; a node graph construction component for constructing a cost matrix for transitions between any two nodes; and a path search and selection component for performing a depth-first path search on the node graph based on the cost matrix, selecting the path with the minimum total path cost from the searched paths as the vehicle's farmland navigation path for the current farmland.

The application discloses a farmland navigation path planning system for an automatic driving vehicle and a method thereof. The system comprises a farmland data collection component, a farmland structure division component, a basic path planning component, a land division component and a farmland navigation path merging component. The farmland data collection component collects farmland information by driving a vehicle around the outermost ridge of the farmland. The farmland structure division component divides the starting ridge, the ending ridge, the far end, the near end, the number of ridges and the turning anchor point position of the farmland based on the average ridge width, the turning anchor point position and the turning radius of the vehicle. The basic path planning component plans the continuous driving route of the vehicle in different basic land blocks in a manner of crossing at least one ridge and excluding repeated driving in the same ridge. The land division component divides the farmland into a plurality of basic land blocks and a last land block in sequence from the starting ridge with a basic ridge number of at least 5. The farmland navigation path merging component connects the navigation paths of all the basic land blocks and the last land block in sequence to obtain an automatic driving navigation map of the vehicle in the farmland.

The invention discloses a lightning arrester instrument data image recognition method, system and device and a medium, and the method comprises the steps: obtaining an original image of a lightning arrester instrument, adaptively adjusting collection parameters according to environment illumination, carrying out the geometric correction and quality evaluation, carrying out the dynamic contrast enhancement and multi-scale noise suppression through an illumination decoupling model, and generating a standardized image; inputting a lightweight mobile terminal target detection model, and positioning and cutting an instrument display area; enhancing the sub-images, extracting edge features, and judging the type of the instrument: if the sub-images are digital, segmenting character areas and identifying the character areas by using a private network to obtain initial readings; verifying validity in combination with a preset physical quantity reasonable range, and generating a final reading; and automatically associating a current inspection task, adding time, a position and an image index, storing into a local database through transaction operation, and supporting generation of a structured inspection report. According to the invention, the accuracy and efficiency of inspection work are improved, and the reliability and traceability of data are ensured.

This invention discloses a high-efficiency video analysis method based on cloud-edge collaboration and adaptive MoE. The method includes: when an edge device performs a video analysis task, it first constructs a meta-request containing current task data and device resource constraints, and sends it to the cloud; the cloud uses a pre-trained basic model to extract features from the video data and clusters the data based on the feature space, dividing complex video scenes into multiple semantic subdomains; a lightweight expert model is trained in the cloud for each semantic subdomain; a routing decision model is further trained in the cloud; after deploying an expert model pool and routers on the edge device, for input data, the router first calculates the matching degree of each expert model and performs a comprehensive evaluation based on its computational cost. Through knowledge distillation, the large-scale basic model is decomposed into lightweight expert models, achieving a significant compression of the model size, enabling efficient deployment on resource-constrained edge devices.

The embodiment of the invention provides an agent deployment method and device, and the method is applied to a configuration center, and comprises the steps: receiving registration information sent when an atomic power service and an application service are started; receiving service dependency information sent by the application service; constructing a service dependency relationship graph according to the service dependency information, and determining configuration dependency coverage according to the service dependency relationship graph, the service access address and the service configuration information; and when it is determined that the configuration dependency coverage meets a preset coverage threshold, determining that deployment of the target agent is completed. According to the method, full-process automation from parameter configuration to state verification is achieved, the operation and maintenance threshold is remarkably reduced while efficient and accurate agent deployment is guaranteed, and technical support is provided for intelligent transformation of enterprises.

The invention relates to the technical field of data analysis, in particular to a palmprint recognition anti-fraud method based on double-branch self-supervised learning, and the method comprises the steps: constructing a double-branch self-supervised reconstruction architecture, determining a first to-be-processed image based on a high-frequency self-supervised branch, and determining a second to-be-processed image based on a chromaticity regularization self-supervised branch. Dividing the two images into patch blocks which are not overlapped with each other, independently generating random mask matrixes with different spatial distributions, determining visible areas, inputting the patch blocks of the visible areas into a shared encoder for feature extraction, outputting latent variable features, and splicing the latent variable features with learnable mask marks to obtain a patch matrix; and respectively sending to a high-frequency decoder and a chroma decoder, determining total loss, carrying out physical decoupling on illumination and material attributes, and carrying out fine adjustment on the architecture to finish convergence of the architecture. According to the method, the palm print features with discrimination are extracted through the shared encoder, and chromaticity distribution consistency constraints are introduced, so that effective decoupling of ambient light and real material attributes is realized.

The invention discloses a lightweight security detection method and system for energy efficiency management of industrial equipment, a terminal and a storage medium, and the method comprises the steps: converting multi-protocol industrial data into a unified feature matrix through zero-copy analysis on an edge gateway, generating an anti-interference robust feature matrix through composite normalization processing, and carrying out the detection of the robust feature matrix; the matrix is input in parallel into a lightweight automatic encoder and an optimization isolation forest for bimodal analysis, a comprehensive abnormal score is output through a dynamic fusion layer, the energy efficiency deviation degree is calculated in combination with an equipment energy efficiency baseline, a safety and energy efficiency joint decision vector is formed, a progressive response strategy is matched according to the decision vector, and a control instruction is executed in real time; meanwhile, the carbon footprint in the detection process and the energy-saving strategy emission reduction amount are quantified, and the model is optimized through feedback. According to the method, efficient deployment on the resource-constrained edge side is realized, the method has strong interference resistance and high real-time performance, a linkage closed loop of safety protection and energy efficiency optimization is opened, and verifiable carbon emission reduction benefits are provided.

The application discloses a self-adaptive wordpiece table level distillation method for a large language model, comprising the following steps: obtaining wordpiece probability distribution of teacher-student models at each generation step; extracting the union of the first K wordpiece indexes with the highest probability of each step respectively, and constructing an aligned candidate wordpiece set; calculating the absolute value of the teacher-student probability difference as the learning difficulty score of each candidate wordpiece, and collecting and constructing a global difficulty matrix; selecting high difficulty wordpieces from the difficulty matrix according to the current focus ratio to form a target distillation wordpiece set, and re-normalizing the student model probability; only calculating the divergence loss of the target wordpiece set and updating the student model parameters; and dynamically adjusting the focus ratio according to the real-time loss of the student during the training process. The application concentrates computing resources on difficult and high information wordpieces through dynamic self-adaptive screening, reduces the computing power consumption, improves the knowledge transfer accuracy, and is compatible with various distillation targets, and is suitable for deployment of large language models on resource-limited devices.

The invention belongs to the technical field of data security and privacy protection, and particularly relates to an attitude data privacy processing system and method based on deep information encryption and identity desensitization. The method is executed by an end side device, and comprises the following steps: collecting depth image data of a human body; performing normalization processing on the depth image data; a lightweight model is used to carry out feature extraction and identity desensitization processing on the normalized depth image data, the identity desensitization processing comprises weakening identity information in features through an adversarial learning mechanism and mapping the features to a desensitization subspace through feature projection, and the lightweight model is trained by the cloud platform; performing encryption operation on the desensitized feature data to generate encrypted data; and transmitting the encrypted data to a cloud platform. The attitude data privacy processing method provided by the invention has the advantages of full-life-cycle privacy protection, thorough stripping of identity information and high-efficiency operation at the end side.

The application discloses a kind of force partition weighted railway bed dirt visual detection method and system, belong to railway engineering intelligent detection technical field.The application constructs the dynamic mapping model of interframe displacement and field of view overlap rate by real-time acquisition train motion parameter, and effectively frame is adaptively filtered;Adopt coarse cutting-fine identification hierarchical parallel vision architecture to identify dirty area;Based on the stress characteristics of bed, high stress area and low stress area are divided, and a mechanical weighting model is introduced to calculate the comprehensive risk index;The risk index is associated with the geographic location to generate a risk space distribution map, triggering hierarchical early warning and maintenance decision.The application solves the problems of data redundancy, heavy computing burden, and the disconnection between evaluation indicators and stress state in the prior art, achieving efficient and accurate detection of bed dirt and risk quantification, providing a scientific basis for railway maintenance decisions, and being suitable for heavy haul railways, coal lines and other scenarios.

The embodiment of the invention provides an audio enhancement method, electronic equipment, a storage medium and a program product. Relates to an audio processing technology. The method comprises the following steps: acquiring a to-be-enhanced initial audio signal; inputting the initial audio signal into a pre-trained audio enhancement model for enhancement processing to obtain an enhanced audio signal; wherein the audio enhancement model is obtained by performing generative adversarial training on a generative model constructed based on an initial audio enhancement model and a pre-constructed discrimination model by taking a loss function containing perception loss as an optimization target; wherein the perception loss is used for representing the feature difference between the output audio of the generative model and the clear audio in the training audio pair in the feature space. The method is used for achieving the technical effect of improving the practicability of audio enhancement in an actual scene.

The utility model belongs to the technical field of environmental protection equipment, and particularly relates to a first-aid device for an electric inclined bed, which comprises a bed frame, at least one wheel is rotatably arranged on the bed frame; one end of the inclined bed body is hinged to the bed frame, and the bed frame is provided with a driving assembly capable of enabling the angle of the inclined bed body to change; the storage mechanism is arranged on the bed frame, the storage mechanism comprises a linear guide component and a supporting rod which are connected to the bed frame, the linear guide component is slidably connected with a movable component detachably connected with the supporting rod, and arc-shaped guide pieces matched with the movable component are arranged at the two ends of the linear guide component; and a locking assembly matched with the movable component is arranged in the arc-shaped guide part and used for limiting movement of the movable part, and the device has the beneficial effects of efficient storage, rapid deployment, space optimization design and the like.

The invention discloses a method, a device, a medium and equipment for intelligently deploying a drilling machine. The method comprises the steps that S10, a well drilling operation tracking theme library is constructed to obtain data of all available drilling machines and well sites, the well drilling operation tracking theme library comprises a well drilling dynamic database, and well drilling daily reports are stored in the well drilling dynamic database; s20, main work content of the drilling daily report is intercepted, data preprocessing is conducted on the intercepted content, and the drilling machine without the drilling footage state is screened out; and S30, outputting a drilling machine operation arrangement plan based on bilateral matching and a GSA algorithm so as to intelligently deploy the drilling machine in the drilling footage-free state. The method comprises the following steps: constructing a drilling operation tracking theme library to obtain data of all available drilling machines and well sites; by effectively fusing bilateral matching and the GSA algorithm, the allocation condition of the drilling machine can be efficiently and accurately tracked, the operation efficiency of drilling construction and the management level of drilling operation are improved, and the utilization rate of drilling machine resources is improved.

The application discloses a method for equipment health diagnosis and training support, and relates to the technical field of intelligent equipment support. The specific steps of the method are as follows: firstly, task priority calculation is performed, priority coefficients and grades are calculated according to collected parameters, and then the parameters are transmitted; secondly, fault atlas reasoning is performed, and a fault is located according to the atlas; thirdly, health state evaluation is performed, a health index is calculated by fusing data, and an early warning is generated; fourthly, resource scheduling decision is carried out, a scheduling scheme is generated by receiving the early warning; and finally, closed-loop parameter correction is performed, algorithm parameters are corrected by collecting execution data, and the scheme is reconstructed. In one aspect, the method fuses multiple features to complete priority calculation, builds a fault space-time knowledge atlas, and realizes accurate equipment health determination and trend prediction. In another aspect, the method models support resources in a digital manner, generates a support scheduling scheme by using a multi-element coupling algorithm, collects data for closed-loop correction, realizes efficient resource allocation, and promotes intelligent upgrading of the support mode.

The invention discloses a semantic communication privacy enhanced federal multitask learning method for mobile edge computing. The method comprises the following steps: initializing a global model comprising a shared encoder and a plurality of personalized decoders by an edge server, and issuing the global model; after each mobile edge client performs multi-task training locally, performing semantic compression processing of integrated error compensation, sparsification, quantification and coding on model updating, performing discretization mapping on features by using a local private codebook to enhance privacy, and uploading compressed data; and after the server side receives the data, inferring a nonlinear relation graph among global tasks through adaptive kernel learning, and according to the nonlinear relation graph, performing differential weighted aggregation of structure perception on updating execution of the shared encoder and each personalized decoder, updating a global model, and issuing the updated global model to the client side for next iteration. According to the method, the endogenous privacy protection is realized while the communication overhead is reduced, and the efficiency, safety and performance of federal multitask learning in a resource-limited mobile edge environment are effectively improved.

This application discloses an adaptive multi-granularity text embedding representation learning method based on gradient optimization. The method includes: text encoding, multi-granularity embedding generation, difficulty-aware dynamic sampling, sliding window negative sample management, adaptive loss optimization, and model training and validation. It aims to address the performance degradation caused by existing embedding vector dimensionality compression techniques. Through a difficulty-aware dynamic sampling mechanism, a Beta distribution is introduced to achieve progressive learning, significantly improving low-dimensional performance; sliding window negative sample management improves sample quality and training efficiency; and a gradient-oriented update strategy avoids gradient conflicts and optimizes training stability. Furthermore, this invention supports incremental dimensionality expansion, reducing training costs while maintaining consistency across multi-granularity embeddings. This method significantly improves performance in low-dimensional environments, maintains lossless performance in high-dimensional environments, significantly improves training efficiency, greatly reduces storage costs, and exhibits strong generalization ability, making it suitable for data from various domains.

This invention provides an integrated airflow-type high-intensity sound source, relating to the field of sound wave transducer equipment. It includes a motor, with one end of the motor shaft connected to an air pump and the other end connected to a sound source rotor via a reducer. The sound source rotor is rotatably mounted within a sound source stator, which has a sound source port. The sound source rotor has a hollow cavity, and the air pump outlet is connected to the cavity of the sound source rotor via an air guide pipe. Multiple frequency modulation ports are located on the outer wall of the sound source rotor, intermittently aligned with the sound source port. The sound source port is connected to an emitter with a horn-shaped nozzle. When the frequency modulation port is aligned with the sound source port, sound is emitted. Through the deep coupling design of the air compression mechanism and the acoustic modulation structure, the performance bottleneck of traditional high-intensity sound wave generators is overcome, demonstrating irreplaceable technical advantages in the field of acoustic applications. In the field of meteorological intervention, it can successfully output low-frequency sound waves of 30-100Hz, enabling meteorological intervention applications such as artificial rain enhancement and fog dispersal.

This invention discloses a method, apparatus, device, and medium for recognizing CAD vector graphics based on structured path encoding. The method includes: acquiring a command sequence and parameter sequence corresponding to each SVG path; obtaining a fixed-length vector representation of the path through an encoding network based on the command sequence and parameter sequence; arranging the fixed-length vector representation of the path sequentially according to path number to obtain a path feature matrix; obtaining a graph-level feature vector of the vector graphic based on the path feature matrix; and obtaining the recognition result of the vector graphic through a neural network based on the graph-level feature vector. This invention enables structured path encoding of SVG format vector graphics without requiring rasterization or conversion to point sets or graph structures, thus uniformly completing the recognition of symbols in CAD drawings.