360 results about "Graph structured data" patented technology

The invention discloses an intelligent management system and method for knowledge graph quality evaluation and self-repairing, belongs to the technical field of knowledge graphs, and aims to solve the problems that in traditional knowledge graph management, manual auditing efficiency is low, an effective automatic repairing means is lacked, and data complexity and real-time changes are difficult to deal with. The system firstly collects multi-source heterogeneous data in a target field, cleans the data through a deep learning noise recognition model, extracts entities and relationships by using a natural language processing technology, and adds metadata to convert the entities and relationships into graph structure data; then, a graph framework is defined based on the ontology, entity semantic alignment is achieved in combination with a graph neural network, and a knowledge graph is constructed by complementing implicit relations with the help of a pre-training language model. Then, the quality of the atlas is quantitatively evaluated through a four-layer quality evaluation system, meanwhile, a repair scheme is generated based on vulnerability feature extraction, knowledge base matching and decision fusion, and intelligent self-repair is achieved; the map can be monitored in real time and evaluated regularly, a repair strategy and a knowledge base are optimized through reinforcement learning, it is ensured that the map is kept accurate and time-efficient for a long time, and the practical value is improved.

The invention discloses an unmanned aerial vehicle charging path optimization method based on deep learning, and the method comprises the following steps: S1, collecting the flight state data of an unmanned aerial vehicle, and constructing a state input sequence; s2, normalizing the state input sequence to form modeling input data; s3, inputting the modeling input data into the improved closed continuous time state evolution model to generate a state trajectory prediction result; s4, generating a path adjustment instruction according to a prediction result, and constructing path diagram structure data; s5, inputting the path graph structure data into the improved graph pointer network to generate an optimal path instruction; s6, fusing the optimal path instruction and modeling input data, iteratively updating and re-predicting a state trajectory; s7, generating a charging path execution scheme; and S8, forming an unmanned aerial vehicle path optimization control instruction sequence. According to the method, the path planning and charging scheduling efficiency of the unmanned aerial vehicle in a complex environment is improved, and the method has relatively high intellectualization and adaptability.

The invention discloses a motor fault diagnosis method and system based on color image fusion symmetric point mode, and the method comprises the steps: converting a vibration signal and an electromagnetic signal of a motor into symmetric point mode images, and generating a color signal image fusing feature information; respectively abstracting the color signal images fused with the feature information into nodes and edges in a high-dimensional semantic space so as to construct graph structure data; and performing diagnosis classification on the graph structure data of the vibration signals and the electromagnetic signals by using respective capsule graph network models, and fusing diagnosis classification results of the vibration signals and the electromagnetic signals through a voting mechanism to obtain a final diagnosis classification result. According to the method, multi-channel time domain signals are converted into image expressions with dense information and consistent geometry, and unified feature modeling is carried out on the images based on a depth map structure network with topology perception capability, so that motor fault diagnosis with high diagnosis precision and strong robustness is realized.

The invention provides a traffic flow prediction method, system and device and a medium, and belongs to the technical field of intelligent traffic. The method comprises the steps of obtaining multi-source heterogeneous original traffic flow time sequence data, performing preprocessing to generate time sequence tensor data, and constructing graph structure data based on road network topology and traffic flow correlation; inputting the time sequence tensor data and the graph structure data into a pre-constructed time-space diagram neural network model, processing the graph structure data through a graph convolution module of the time-space diagram neural network model to extract spatial features, and processing the time sequence tensor data through a time sequence modeling module of the time-space diagram neural network model to extract time features, fusing the spatial features and the time features to form joint spatial-temporal features; inputting the joint spatial-temporal characteristics into a prediction layer of a spatial-temporal diagram neural network model for processing, and outputting predicted value data of the traffic flow in a future period as prediction result data; and feeding back prediction result data to the traffic control and management system to drive the traffic control and management system to execute control operation.

The invention relates to the technical field of graph neural network and causal representation learning, in particular to a graph neural network method for separating time-varying and time-invariant causal features, which comprises the following steps: acquiring graph structure data at a plurality of moments, and establishing a dynamic graph sequence; coding the dynamic graph sequence by adopting a time sequence graph neural network to obtain a node representation matrix; respectively adopting a time-varying attention head, a time-invariant attention head and a shortcut attention head to carry out attention score calculation, mask construction and sub-graph decomposition processing on the node representation matrix to obtain a time-varying cause sub-graph, a time-invariant cause sub-graph and a shortcut feature sub-graph; performing independent graph neural network coding and dynamic gating fusion to obtain fusion features; determining a current predicted value according to the fusion feature; according to the method, the causal discovery accuracy, robustness and interpretability can be improved.

The invention belongs to the technical field of engineering simulation data processing, and particularly discloses a finite element grid graph structure construction method and system, a terminal and a medium. Comprising the following steps: analyzing original full-amount grid data exported by finite element simulation, segmenting unstructured grid data into a grid vertex coordinate set and a grid unit mark number set, and constructing a graph edge topological structure corresponding to a grid based on a unit mark number relationship; on the basis, loading physical field data and adopting a tolerance-based coordinate matching algorithm to realize accurate mapping of physical field labels and material attributes with grid nodes; and generating standardized finite element grid graph data which can be directly used for graph neural network processing. By means of the method, high-consistency and high-physical-reliability conversion from the finite element simulation data to the graph structure data is achieved, and the physical field modeling and simulation acceleration capacity based on graph learning is improved.

The invention discloses a building electrical safety protection method, which comprises the following steps of performing real-time data acquisition on a building electrical system to obtain multi-modal time sequence data; constructing the multi-modal time sequence data into graph structure data with a node-edge topological relation, and encoding the multi-modal original observation data corresponding to each node into a multi-modal initial feature vector of the node; graph neural network feature extraction is carried out on the graph structure data, and a global feature vector sequence used for representing the operation state of the electrical system is obtained in combination with an attention mechanism; inputting the global feature vector sequence into a pre-constructed time sequence prediction model to perform operation state prediction, and judging whether potential abnormality exists or not; and when the residual error exceeds a preset threshold value, fault backtracking positioning is carried out on the key node according to the attention weight, and an electrical fault point is generated in combination with node characteristics. According to the invention, the fault response speed and the emergency disposal efficiency can be effectively improved, and the safety accident rate is reduced.

The invention relates to a wind power cluster short-term power prediction method and device of a space-time diagram neural network fused with physical information and computer equipment, and the method comprises the steps: obtaining related information data of each wind power plant in a wind power cluster, and carrying out the preprocessing; forming a physical prior data set through an engineering analysis model fusing the wake flow analysis model and the blocking effect model; taking each wind power plant as a node of the graph, constructing graph structure data for predicting the power of the wind power plant, and forming a dynamic adjacent matrix; constructing a space-time diagram neural network WB-STGNN model architecture comprising a diagram convolutional neural network module, a gating time convolutional network and a multi-layer perceptron; the method comprises the following steps: pre-training by using a physical prior data set, and then performing formal training based on historical power data and a dynamic adjacency matrix to obtain a space-time diagram neural network WB-STGNN model; inputting the wind speed of the prediction day, and predicting the active power of the whole wind power cluster in 24 hours of the prediction day. By adopting the method, the precision and efficiency of wind power cluster power prediction can be effectively improved.

The invention discloses a drainage basin real-time flood control scheduling risk dynamic early warning and emergency scheme making method, which comprises the steps of obtaining a real-time total element information tensor and an engineering flood control design parameter set, and calculating to obtain a real-time hierarchical risk entropy time sequence front section based on safety margin distribution, and inputting real-time data and pre-constructed entropy flux Markov graph structure data into the trained entropy perception double-flow prediction model for reasoning to obtain a future risk prediction set. And a dynamic risk early warning result set is generated in combination with a preset entropy level threshold and a future failure chain probability, and an emergency scheduling scheme set is generated on the basis of meeting safety margin constraints and entropy reduction targets. According to the invention, dynamic early warning and decision making of strong physical-risk coupling are realized.

The invention discloses a real-time monitoring and closed-loop regulation and control method and device for construction quality of an asphalt pavement, and relates to the technical field of intelligent monitoring of road construction. The method comprises the steps of collecting multi-modal data in a construction process and performing space-time alignment to generate a space-time aligned multi-modal data set; constructing graph structure data of the construction area based on the set; extracting a node embedding vector by using a graph neural network; synchronously outputting the compactness, the temperature uniformity index and the flatness index through a multi-task prediction model based on the vector; identifying a construction track abnormal area based on the road roller track data; and generating a construction quality digital twinborn model by integrating the information, and performing feedback regulation on an automatic control system of the road roller based on the model. Through multi-modal data fusion, graph structure modeling and multi-task collaborative prediction, real-time monitoring and dynamic regulation and control of the construction quality of the asphalt pavement are realized, the prediction precision can be effectively improved, and the monitoring efficiency and accuracy are remarkably improved.

The invention discloses a low-altitude target recognition method based on an attention mechanism, and the method comprises the steps: collecting low-altitude image data, and carrying out the preprocessing, and obtaining standardized low-altitude image data; executing a maximum flow minimum cut algorithm to generate a foreground mask; constructing a space attention mechanism, and generating a space attention weight and a space weighted feature vector; performing channel statistics on the spatial weighted feature vector to generate a channel weight and a target feature vector; constructing a target association matrix and graph structure data to obtain an association enhanced target feature vector; and inputting the associated enhanced target feature vector into an improved Kohonen neural network model, determining a low-altitude target category, and outputting an identification result. According to the invention, by introducing the maximum flow minimum cut algorithm and improving the Kohonen neural network model, stable and accurate identification and classification of the flight target in the complex low-altitude environment are realized.

The invention provides a new energy station high-voltage voltage transformer error evaluation method and system, and the method comprises the steps: collecting the voltage phasors of all measurement nodes in a preset time period at a transformer substation and a new energy station at the same time, obtaining the equipment parameters of a voltage transformer and the line parameters between the measurement nodes, and constructing a measurement data set; the method comprises the following steps: constructing a corresponding adjacency matrix and a node characteristic matrix by extracting line impedance and node working condition characteristics based on a power system line topological structure and data in a measurement data set, and obtaining a time sequence diagram structure data set according to the adjacency matrix and the node characteristic matrix; constructing a graph neural network, and training the graph neural network based on the time sequence graph structure data set; and inputting the voltage phasor of each measurement node, and based on the output result of the trained graph neural network, judging the state of the voltage transformer through a preset threshold value. According to the scheme, online evaluation of the error state of the voltage transformer can be realized, and the accuracy and interpretability of error evaluation are improved.

The invention discloses an unmanned aerial vehicle wind power blade defect dynamic detection method based on AI vision, and relates to the technical field of JSLYMC, and the method comprises the following steps: S1, collecting an original dynamic image sequence; s2, generating an aligned image sequence; s3, generating a standardized blade surface image sequence; s4, constructing a blade topological structure diagram; s5, inputting the graph structure data into the improved TransGAT model, and outputting a defect candidate set; s6, inputting the defect candidate set into the FairMOT model, and outputting a defect space-time trajectory; and S7, generating a structured detection report. The method overcomes the limitations of dependence on manual inspection, poor dynamic adaptability and insufficient identification precision in a traditional wind power blade defect detection method, and provides an efficient and accurate solution for unmanned aerial vehicle wind power blade automatic inspection and intelligent maintenance decision.

The invention discloses a distributed photovoltaic power prediction method and device based on dynamic spatial correlation, and relates to the field of distributed photovoltaic technology, and the method can reflect the real influence relation between photovoltaic power stations at different moments in real time through constructing dynamic graph structure data which can reflect the spatial relation between the photovoltaic power stations more comprehensively and accurately. And the prediction result of the distributed photovoltaic power is improved. The method comprises the following steps: acquiring an influence factor feature set of distributed photovoltaic power prediction; by taking the distributed photovoltaic power station as a node, constructing dynamic graph structure data of the photovoltaic power station according to the influence factor feature set and the spatial weight matrix fused with the meteorological condition information; performing spatial correlation extraction on the dynamic graph structure data by using a pre-trained first network to obtain spatial correlation characteristics of the distributed photovoltaic power station; and performing time domain correlation extraction on the spatial correlation characteristics by using a pre-trained second network to obtain a point prediction result of the distributed photovoltaic power.

The invention provides an abnormal user detection method and system based on topology awareness and hub node guidance. The method comprises the steps of obtaining graph structure data; fusing the topological relation in the graph structure data and the initial features of the nodes by using a topological adaptive label evolution mechanism, and predicting and optimizing labels of unlabeled nodes to generate an optimized label distribution matrix; based on the node initial features, taking the optimized label distribution matrix as a supervision signal, and generating semantic hub node features through a double-layer multi-scale contrast learning mechanism; dynamically screening neighbors of each target node by adopting a strategy gradient driven neighbor aggregation mechanism based on semantic hub node features; performing message passing through a multilayer graph neural network based on the screened neighbor set, and fusing node representations of different layers by using an attention mechanism to obtain a final node representation; and inputting the final node representation into a classifier, and outputting an abnormal user detection result.

The invention relates to a power distribution network safety analysis method, device and equipment based on a graph convolutional neural network, a storage medium and a program product, and relates to the technical field of artificial intelligence. The accuracy and reliability of the analysis result can be improved. The method comprises the following steps: constructing an adjacent matrix of the power distribution network according to structural feature information of the power distribution network, and constructing a node feature matrix of the power distribution network according to multi-source state information of the power distribution network; based on the adjacency matrix and the node feature matrix, generating a graph structured data set of the power distribution network in different operation states, and obtaining a training set, a verification set and a test set according to the graph structured data set; constructing a to-be-trained initial network model according to the target loss function and the target hyper-parameter; and training the initial network model by using the training set and the verification set through a deep learning algorithm and updating the model parameters of the initial network model until the performance index of the updated network model on the test set meets a threshold condition, thereby obtaining a trained graph convolutional neural network model.

The disclosure provides a graph neural network training method, a training system and an abnormal account identification method. The graph neural network training method comprises: obtaining initial graph structure data corresponding to a terminal device; the initial graph structure data obtained by a plurality of distributed training terminals respectively is derived from the same sample graph structure data; the following graph structure data processing stage and graph neural network training stage are executed cyclically until a target neural network meeting the training requirement is obtained: determining a processing time of the current execution graph structure data processing stage according to historical execution data of the historical execution graph structure data processing stage and the historical execution graph neural network training stage; performing graph structure data processing on the initial graph structure data in the graph structure data processing stage according to the processing time to generate target graph structure data; the graph structure data processing comprises data sampling processing and feature extraction processing; and training the target neural network based on the target graph structure data in the graph neural network training stage.

The invention discloses an electric power knowledge graph construction method and system based on multi-head attention and a medium, and belongs to the technical field of electric power system automation. The method comprises the following steps: constructing graph structure data including power equipment, parts, defects and environment entities, and encoding node features; secondly, inputting graph data into a graph attention module, reducing storage overhead through a low-rank compression key value vector, independently performing decoupled rotation position coding on a query vector to inject hierarchy and time position information, respectively modeling an equipment hierarchy relationship and a dynamic time sequence relationship through a division-of-labor multi-head attention mechanism, and performing multi-head attention mechanism modeling on the equipment hierarchy relationship and the dynamic time sequence relationship; obtaining a node vector representation; and finally, performing hierarchical processing and dynamic updating on the node representation, and constructing and continuously updating the electric power knowledge graph through knowledge reasoning after fusion. The problems that a traditional method is low in storage efficiency, insufficient in space-time relation modeling and difficult in dynamic updating are solved, and the efficiency, precision and real-time performance of graph construction are remarkably improved.

The invention provides an AI safety early warning method based on a temporary power system. The AI safety early warning method comprises the following steps: collecting three-phase voltage time sequence data of each bus and three-phase current time sequence data of each branch; taking the voltage data as node features, taking the current data as edge features, and constructing graph structure data in combination with power grid topology; a 1-D GAT model is constructed, time sequence features are extracted through a one-dimensional convolutional layer, spatial features are extracted through a graph attention layer, and multi-task output is completed through a full connection layer; dividing a training set and a test set, and training the model by adopting a multi-task loss function until convergence; and inputting graph structure data of a to-be-diagnosed system into the trained model, and outputting a complete diagnosis result including fault judgment, type identification and position positioning. According to the method, the problems of poor real-time performance, incomplete feature extraction, low diagnosis precision, single function and the like of a traditional method are solved, and the accuracy and the reliability of temporary power system safety early warning are remarkably improved.

The invention discloses a control method and device of a customer relationship management system and a storage medium, and belongs to the technical field of management systems. The method comprises the steps of generating an association problem set according to service scene information of at least one predefined service scene, constructing graph structure data based on the association problem set, forming a domain knowledge graph, distributing priority weights for conflict problems in the domain knowledge graph, and integrating the domain knowledge graph into an intention recognition module, and in response to a target service scene triggered by a user, outputting at least one guide problem matched with the target service scene through an intention recognition module on the basis of the domain knowledge graph and the priority weight. According to weight distribution and a context association mechanism of the knowledge graph, the accuracy of a recommendation system in a complex service scene is improved.

The invention discloses a photovoltaic module subfissure fault diagnosis method fusing IV characteristic spectrum and infrared thermal imaging. Comprising the following steps: firstly, acquiring infrared thermal imaging data and IV scanning data of a photovoltaic module to be diagnosed; then, dynamic thermal characteristics such as phase delay and amplitude attenuation are extracted from the infrared thermal imaging sequence, and electrical characteristics including basic electrical parameters and differential characteristic spectrums are extracted from the IV scanning data; secondly, abstracting the photovoltaic module into a graph structure; then, processing graph structure data by adopting an asynchronous time-space diagram attention network, and realizing efficient multi-modal feature fusion by learning dynamic association weights among nodes; and finally, carrying out super-resolution positioning by utilizing a learnable iterative shrinkage-threshold network, and carrying out quantitative evaluation on the hidden crack fault. Compared with the prior art, the method has the advantages that the diagnosis accuracy is remarkably improved, the false alarm rate is greatly reduced, super-resolution accurate positioning is achieved, quantitative evaluation indexes are provided, and the model generalization ability is enhanced.

The invention discloses a salinized frozen soil roadbed settlement prediction method and system, and relates to the technical field of roadbed settlement prediction, and the method comprises the steps: obtaining multi-source monitoring data of a target salinized frozen soil roadbed, the multi-source monitoring data comprises temperature, moisture content, conductivity, stress and ground surface settlement time sequence observation values distributed along the depth, and a ground surface remote sensing image; establishing a physical mechanism model reflecting the heat-water-salt-force coupling effect of the salted frozen soil based on the multi-source monitoring data; integrating the physical mechanism model as a constraint condition into a neural network training process, and constructing a settlement response prediction sub-model with physical consistency in combination with multi-source monitoring data; performing spatial reconstruction on the multi-source monitoring data to form a two-dimensional profile image containing a temperature field, a humidity field and a salt field, and inputting the two-dimensional profile image into a convolutional neural network to extract spatial structure features; and constructing graph structure data by utilizing spatial structure characteristics, modeling a mutual influence relationship among different salinized areas through a graph neural network, and outputting settlement tendency distribution.

The invention provides a sewage plant water-energy-drug collaborative optimization method and system, and the method comprises the steps: obtaining the data of a technological process and a material transfer relationship of a sewage plant, constructing a graph network structure model with a technological unit as a node and material flow as an edge, and carrying out the preprocessing, thereby obtaining dynamic coupling graph structure data; a dynamic coupling graph neural network model containing a node feature coding layer, a time sequence coding module, a space message passing layer and an attention mechanism layer is constructed based on the data, and a prediction model capable of representing the dynamic coupling relation of the process unit is obtained through historical data training; then, a multi-objective optimization function which takes the lowest ton water treatment cost as an objective and covers water quality standard reaching, energy consumption and medicament dosage constraints is constructed; and in combination with the prediction model and the optimization function, the optimal operation parameters are solved through an optimization algorithm, and a whole-plant collaborative optimization decision scheme is generated. According to the invention, the dynamic coupling GNN prediction model and the virtual element multi-domain parallel optimization technology are integrated, and intelligent operation management of the sewage treatment plant is realized.

The invention relates to an example-based clothing template automatic generation method and device. The method comprises the following steps: firstly, constructing a standard clothing model database containing semantic tags based on a graph structure; constructing a graph neural network segmentation framework based on a surface adjacent graph, converting the triangular mesh model into graph structure data, and constructing a surface patch classification probability prediction model; then training the framework and semantic tag prediction, iteratively optimizing semantic tags by adopting a graph cut optimization algorithm, and clustering to form an independent cut piece set through a region growing algorithm based on the optimized tags; performing least square conformal parameterization on each cutting piece to obtain an initial two-dimensional sample plate, and constructing a geometric constraint energy function containing fabric characteristics and sample plate process requirements; and finally, solving a four-dimensional comprehensive optimization energy function containing physical characteristics and geometric constraints, and obtaining a final sample plate by minimizing the function. According to the method, highly-automatic end-to-end model generation is achieved, and the industrial platemaking standard is met more accurately.

The invention relates to the field of biomedical engineering, and discloses a breast cancer neoadjuvant chemotherapy curative effect prediction method based on a graph neural network, and the method comprises the steps: obtaining multi-modal data of a breast cancer patient, and extracting a time sequence feature vector; constructing a plurality of meta-tasks containing a support set and a query set, and constructing the time sequence feature vector of the patient sample in each meta-task into graph structure data; inputting the graph structure data into a graph neural network model for curative effect prediction and loss calculation; back propagation prediction loss update model parameters; and constructing the time sequence feature vector of a new patient to be predicted and the support set into new graph structure data, inputting the new graph structure data into the trained model, and outputting a prediction result. The system comprises a data processing module, a graph construction module, a model training module and a curative effect prediction module. According to the method, through the meta-learning strategy and the graph neural network, multi-modal time series data are effectively integrated, accurate prediction of the breast cancer neoadjuvant chemotherapy curative effect in a small sample scene is realized, and support is provided for clinical diagnosis and treatment.

The invention relates to the technical field of medical information processing, in particular to a gastric cancer / pancreatic cancer risk assessment method and system based on periodontal data, and the method comprises the steps: obtaining periodontal index data and clinical index data of a user, and extracting a periodontal image feature vector in the periodontal image data; fusing features reflecting similar information in the periodontal index data, the periodontal image feature vector and the clinical index data, and mapping the fused data into nodes in a heterogeneous graph; on the basis of statistical correlation, domain knowledge and biological association, defining connecting edges between different nodes, and constructing initial heterogeneous graph structure data; carrying out aggregation updating processing on the initial heterogeneous graph structure data by utilizing a heterogeneous graph neural network model to obtain feature representation of each node; and performing classification processing based on the updated feature representations of all the nodes, and outputting a risk assessment result that the user belongs to gastric cancer, pancreatic cancer or non-cancer people. And constructing a relational graph model by using the multi-modal periodontal data to realize quantitative evaluation of cancer risks.

The invention discloses a code processing method and device fusing grammar structure and graph structure learning, and the method comprises the steps: collecting high-performance project codes, extracting hot code segments, and analyzing the structural features of the hot code segments; converting the hot code segment into an abstract syntax tree AST file, performing structured analysis, converting the file into Python tree structure data, and expanding node attributes; converting the tree structure data into graph structure data, generating edge, graph and node index files, and constructing the graph structure data and corresponding code structure labels; constructing a graph attention network model for different code structures, and training a plurality of GAT models based on graph structure data and labels; and converting the user input code, inputting the converted code into each GAT model for prediction, and integrating and outputting a feature vector. According to the method, grammar logic is accurately captured through AST conversion and node expansion, multi-language AST is unified into a graph structure, the recognition accuracy of a complex code structure is improved through a GAT multi-model architecture, full-process automation is achieved, and the labor cost is reduced.

The invention discloses a digital regularization and graph conversion method for large model training, which belongs to the technical field of image processing, and comprises the following steps: acquiring a target document, carrying out preprocessing and layout analysis on the target document, and extracting title text block information, body text block information and document picture data; based on multi-scale density analysis, determining a column layout of each page in the target document and a column to which each text block belongs, and establishing a chapter structure; judging the chapter attribution of the text block through two-stage decision, and judging the reference attribution of the document picture based on the reference relationship and the document picture data; reconstructing a document structure, and sequentially generating graphic structured data; and constructing a multi-modal knowledge graph. According to the method, the logic structure of the complex typesetting document is identified and reconstructed, the structured data is extracted, the data utilization rate and the information depth of the complex document are improved, and a high-quality training basis is provided for a large model.

The invention provides a multi-modal large language model construction method for automatic accounting document auditing, and belongs to the technical field of large language models.The method includes the steps that table graph structure data are constructed, a graph convolutional network is input for node feature propagation, and a sequential dependency relationship is modeled in combination with a bidirectional long-short-term memory network; a cross-modal semantic alignment encoder is constructed to map a visual token sequence and a language token sequence to a unified semantic space, a conditional generative adversarial network is adopted to carry out data enhancement, a time sequence modeling framework of a dynamic graph neural network is constructed to process time-varying graph structure data, an active learning framework and a noise adaptation network are designed, and time-varying graph structure data processing is carried out. A knowledge distillation technology and a model optimization method are used, a multi-task uncertainty weighted loss balance algorithm is adopted for training to obtain a multi-modal large language model for accounting document automatic auditing, and the problem that in the prior art, it is difficult to accurately understand complex table structures and cross-modal semantic association is solved.