82 results about "Graph mapping" patented technology

The invention relates to a data query analysis method based on a natural language, which comprises the following steps of: analyzing a natural language query request input by a user into a dynamic semantic intention tensor which comprises an operation type, a comparison relationship, a time constraint and a data role semantic dimension, and dynamically combining according to task semantics to form a complete data query and analysis operation expression; mapping the natural language intention to a controllable data operation space of a database, and generating a preliminary data operation chain; performing side effect perception processing on the operation tensor, predicting a potential execution consequence of the operation, generating a side effect tensor for simulating a side effect of the database operation, and if any index in the side effect tensor exceeds a preset threshold value, performing local correction on the operation tensor; according to the method, a complex natural language query task is disassembled into a plurality of minimum semantic units by adopting a step-by-step semantic resolution mechanism, each unit is mapped into a specific tensor operation, and a dependency relationship between operation units is dynamically reasoned to generate an operation execution chain.

The invention relates to the technical field of cognitive inference, and discloses a cognitive inference system fusing a dynamic knowledge graph and deep learning. The system comprises a cognitive event stream processing module which is used for acquiring a real-time event stream and a cognitive target stream, identifying entity state change and extracting a reasoning intention; the graph mapping generation module is used for receiving the entity state change and the reasoning intention, constructing a mapping relation with knowledge graph topology and generating a multi-dimensional graph mapping library; the cognitive path decision-making module is used for extracting key topological characteristics from the multi-dimensional map mapping library, calculating a reasoning path migration probability and generating an initial cognitive path set; the time sequence evolution prediction module is used for carrying out time sequence evolution modeling on the knowledge graph topology based on the initial cognitive path set, identifying node state offset and updating a multi-dimensional graph mapping library; and the reasoning optimization module is used for analyzing a reasoning stage and a reasoning dependency relationship of the cognitive target flow according to the updated multi-dimensional map mapping library, and generating a dynamic cognitive reasoning library.

The invention discloses a graph mapping method based on precise synthesis, and the method comprises the steps: constructing a precise synthesis structure library based on NPN equivalence classes, and inputting the precise synthesis structure library into a Boolean network; k-Cut cutting enumeration is carried out on the input Boolean network, a truth table is calculated, cutting is matched with a structure in a precise comprehensive structure library through Boolean matching, and rapid search is realized in combination with NPN classification; the overlay network is generated through multi-round mapping optimization, logic sharing nodes are mined in combination with structural hash, and the multi-round mapping optimization comprises delay-oriented mapping, global area topological optimization and local accurate area optimization; and a new target network is generated based on the optimized overlay network, redundant nodes are removed, and final network construction is completed. According to the method, manual intervention is not needed through full-process automatic mapping and a redundancy removal mechanism, the labor cost of a technical mapping link in chip design is remarkably reduced, and the method is particularly suitable for efficient design of complex circuits and emerging majority of logic base technologies.

The invention provides a user behavior intention recognition method and device based on multi-modal data fusion, and relates to the technical field of data processing.The method comprises the steps that a multi-modal observation sequence is obtained, initial features are formed through time synchronization and mutual information screening, a cross-space initial offset vector is generated in combination with virtual space offset parameters, and the cross-space initial offset vector is obtained; establishing an association relationship between a real action and a virtual action through nonlinear mapping to obtain a cross-space registration feature; generating a correction coefficient according to the proportion of the real action amplitude and the virtual action amplitude, and performing amplitude correction to obtain a cross-space correction feature sequence; compensating rendering delay and synchronous drift by using weighted fusion of the current feature and the lag feature to obtain a time sequence correction feature; outputting cross-space fusion features through a unified embedding space and a cross-modal attention mechanism; and non-linear intention mapping is constructed based on the fusion features, and a user behavior intention result is inferred and output in combination with time smoothing. According to the invention, the accuracy of user behavior intention analysis in a virtual field scene can be improved.

The invention discloses an interactive document editing canvas method, which comprises the following steps of: establishing a data-view mapping relation directly driven by a Vue2 responsive system between a canvas container and an element array, any addition, deletion and modification operation aiming at the node description data set can be instantly and accurately reflected to the style and attribute of the corresponding DOM child node without manual intervention, so that element positioning, rendering and follow-up editing are completed in the same technical action, the delay and inconsistency risk caused by frequent and direct operation of a real DOM in a traditional scheme is eliminated, and the operation efficiency is improved. The interaction fluency, the layout intuition and the system maintainability are obviously improved; by means of a virtual DOM mechanism of Vue2, batch diff is carried out and then submitted to a real DOM in a unified mode, the number of rearrangement and redrawing times of a browser is reduced, CPU and memory occupation is further reduced, and the canvas still keeps smooth response even in the scene with dense elements and frequent updating.

The invention relates to the technical field of constructional engineering information retrieval, and discloses a constructional engineering business opportunity information retrieval agent system and method based on a large model. The system comprises an information acquisition module, an intention mapping module, a mapping relation construction module, a business opportunity matching module, a dynamic updating module and a retrieval management module. The information acquisition module acquires multi-modal business opportunity information from a multi-heterogeneous data source, and generates a business opportunity feature warehouse through integration and feature extraction; the intention mapping module receives a user retrieval request, analyzes a semantic intention by means of a big language model to generate an intention feature vector, and performs preliminary matching in a business opportunity feature warehouse; the mapping relation construction module constructs a dynamic mapping relation between intention and business opportunity feature space according to the association strength; the business opportunity matching module combines the relationship with the time weight to calculate semantic relevancy, and generates a business opportunity matching sequence; the dynamic updating module monitors data change, updates the warehouse and adjusts the mapping relation; the retrieval management module generates a hierarchical retrieval strategy, constructs a business opportunity retrieval knowledge base, and meets the retrieval requirements of building enterprises.

The invention provides a three-dimensional part retrieval method based on graph similarity search, relates to the field of computer graphics, and solves the technical problems that in the prior art, geometric and design semantic information of CAD cannot be fully utilized, and a topological relation is difficult to capture and display, so that the retrieval efficiency is low. The method comprises the following steps: acquiring part three-dimensional data of a computer-aided design (CAD) model, and constructing a training data set; constructing an edge-surface connection diagram based on the three-dimensional data of the part; calculating a graph editing distance (GED) matrix of all edge surface connection graphs in the training data set as a supervision signal; based on the GED matrix, training a sorting model, mapping an edge-surface connection graph to a hidden space, and constructing a part vector database according to an output graph-level embedding vector; the sorting model is constructed based on a graph attention network; and inputting a to-be-queried CAD part into the trained sorting model to obtain a feature vector, carrying out nearest neighbor search in the part vector database, and returning a similar part result.

There is provided a computer implemented method of managing access for user account, comprising: accessing a graph mapping a plurality of nodes denoting real users to a plurality of user accounts and a plurality of resources hosted by a plurality of service computing environments external to a target computing environment accessed by the user accounts, and defining different permissions for user accounts for accessing the different resources, receiving a query for identifying user accounts for a target user, executing the query on the graph, and providing details regarding user accounts of the target users from the execution of the query on the graph, the details including at least one of: identifiers of the user accounts, resources accessible to the user accounts, access privileges for accessing the resources.

The invention relates to the technical field of biological detection signal analysis, and discloses a method for rapidly detecting the content of active collagen. The method comprises the following steps: collecting an original optical response signal of a reaction between a to-be-detected solution and a specific reagent, and constructing a time-space continuous primary reaction track through multi-level slicing and cross-domain splicing of a time domain, a frequency domain and a time-frequency domain; a signal mode set related to activity is separated through convolution kernel group depth scanning. A graph association network between signal modes is constructed based on the set, and a steady-state feature graph representing reaction kinetics is generated through iterative calculation. The feature map is mapped to a virtual reaction space defined by standard data, and quantitative judgment is achieved by calculating the distance between a projection vector and a preset concentration anchor point. According to the method, the sensitivity and the anti-interference performance of detection are enhanced through deep analysis reaction dynamic signals and correlation network modeling.

The invention discloses a task processing method and device based on quantum computing, equipment and a storage medium. The method comprises the steps of determining a maximum cut problem corresponding to a target task and a weighted undirected graph of the maximum cut problem; performing community detection division on the weighted undirected graph to obtain a plurality of community sub-graphs; mapping the community sub-graph into a sub Hamiltonian and constructing a parameterized quantum circuit; updating parameters of the parameterized quantum circuit by adopting a gradient descent algorithm to minimize a global Hamiltonian expected value, and outputting a binary string of a quantum state corresponding to the global Hamiltonian expected value as an initial solution; generating a plurality of candidate solutions for the initial solution based on a neighborhood search algorithm, calculating cut values of the initial solution and the candidate solutions, and selecting a better feasible solution based on the optimal cut value; and applying a preset disturbance operator in the parameterized quantum circuit to construct a disturbance quantum circuit, updating parameters of the disturbance quantum circuit by adopting a gradient descent algorithm, and outputting a binary string of a quantum state corresponding to the global Hamiltonian expected value as a final solution of the target task by minimizing the global Hamiltonian expected value after disturbance.

This invention relates to the field of electronic digital data processing and digital information transmission technology, specifically to a cloud-based intelligent data processing system for multiple nodes. The system includes a data preprocessing unit, a topology-aware scheduling engine, a graph neural network device, and a task execution unit. The data preprocessing unit extracts the temporal dependencies and logical associations of the input data stream to generate a directed acyclic graph (DAG). The topology-aware scheduling engine maps the graph to the physical network topology and calculates link bandwidth margin and congestion status. The graph neural network device jointly encodes node computational load and edge link status to output a task allocation strategy. The task execution unit schedules strongly dependent tasks to adjacent nodes with sufficient link margin and idle computing power according to the strategy, and dynamically migrates tasks to alternative path nodes based on link packet loss rate during execution. This invention reduces data transmission latency across switches in multi-node cascading and avoids node computing power idleness caused by local link congestion.

The application belongs to the technical field of image segmentation, and particularly relates to a complex scene target segmentation method and system based on image recognition, which comprises the following steps: performing superpixel adaptive division on an input image, fusing gray scale and texture features to determine a superpixel boundary, mapping the superpixel boundary into a graph node and calculating an edge weight, and constructing an undirected weighted graph; adaptively encoding a graph signal, utilizing a hybrid graph wavelet-Fourier joint transform to optimize and separate features, and obtaining purified graph frequency domain features; sparsely reconstructing features through adaptive super-complete dictionary learning, and obtaining target enhanced features; extracting topological parameters based on an improved persistent homology, constructing a topological constraint feature graph, mapping the topological constraint feature graph into an initial contour field, iteratively optimizing a level set and a contour through an adaptive partial differential equation, and obtaining a high-fidelity coarse segmentation result; extracting geometric features to construct a joint constraint model to repair an occluded area, and outputting a precise segmentation result. In the application, sparse topological modeling is adopted, weak features are strengthened, and target discrimination accuracy is improved.

Various embodiments include a method for finding and erasing errors within a mapping process transforming an input file into a graph using a transformation instruction written in a graph mapping language and saving the graph in an output file. An example includes: acquiring the input file to be transformed into the graph by the mapping process; acquiring a mapping file written in the graph mapping language and comprising a plurality of transformation instructions for transformations of the input file by the mapping process; executing the transformation instruction in an order specified by the mapping file; checking whether a predetermined condition is satisfied for the particular transformation instruction to be executed and / or the result of the execution; executing a predetermined operation and / or outputting the transformation instruction executed and / or the result and / or the output file in dependence on the condition; and debugging the mapping file in correspondence to the input file.

The application provides a three-dimensional part retrieval method based on graph similarity search, relates to the field of computer graphics, and solves the technical problem that the prior art cannot fully utilize the geometry and design semantic information of CAD, and it is difficult to capture and display the topological relationship, resulting in low retrieval efficiency. The method comprises the following steps: obtaining part three-dimensional data of a computer-aided design (CAD) model, and constructing a training data set; constructing an edge-face connection graph based on the part three-dimensional data; calculating a graph edit distance (GED) matrix of all edge-face connection graphs in the training data set as a supervision signal; training a ranking model based on the GED matrix, mapping the edge-face connection graph to a hidden space, and constructing a part vector database according to the output graph-level embedding vector; the ranking model is constructed based on a graph attention network; inputting a CAD part to be queried into the trained ranking model to obtain a feature vector, performing nearest neighbor search in the part vector database, and returning a similar part result.

The invention provides an agent anthropomorphic interaction method and system based on cognitive graph tensor decomposition, and the method comprises the steps: receiving multi-modal input data issued by a user, converting the multi-modal input data into structured cognitive elements, and quantifying the cognitive elements into mathematical expressions, so as to obtain a three-dimensional cognitive graph tensor; decomposing the three-dimensional cognitive schema tensor to obtain a core tensor and a factor matrix; performing double-graph mapping based on the core tensor and the factor matrix to obtain a deep intention recognition result and a topic strategy; the anthropomorphic natural language response is generated and output based on the deep intention recognition result and the topic strategy, smooth transition of topics can be achieved, the embarrassment of'rigid turning 'of a traditional dialogue robot is avoided, and anthropomorphic experience is improved.

The application discloses a kind of house building engineering intelligent surveying and mapping method, it is related to the technical field of image processing.The original surveying and mapping chart of target building is acquired, and the gray texture matrix and edge feature matrix are obtained by extracting original surveying and mapping chart;Gray texture matrix and edge feature matrix are enhanced to obtain enhanced texture matrix and enhanced edge matrix respectively;The R channel feature map is obtained by mapping enhanced feature matrix to R channel, the G channel feature map is obtained by mapping original surveying and mapping chart to G channel, and the B channel feature map is obtained by mapping enhanced edge matrix to B channel;R channel feature map, G channel feature map and B channel feature map are fused to obtain enhanced surveying and mapping chart;The surveying and mapping data are obtained by identifying enhanced surveying and mapping chart through target model, and the surveying and mapping data are uploaded to cloud.The building texture and edge feature are extracted and enhanced, the details are highlighted using multi-channel fusion and the original information is retained, and the identification accuracy and efficiency of model surveying and mapping data are improved.

This invention proposes a collaborative evolutionary classification method based on cross-view structural attribute dual-domain semantic mining, belonging to the field of multi-view semi-supervised classification. This invention acquires multi-view features from training samples, constructs K-nearest neighbor graphs for each view, and adaptively prunes based on the shared support relationships of the graph structure to obtain a purified adjacency matrix. Subsequently, structural and attribute representations are extracted through a shared graph convolutional network, and the consistency between the two in a unified space is constrained using a structure-attribute consistency loss. Simultaneously, learnable view position encoding and masked self-attention mechanisms are introduced for progressive fusion, and cross-view evolutionary patterns are learned through adjacent view mapping to form an evolutionary loss. Finally, the evolutionary representations of each view are concatenated and mapped to a sample-level fused representation, which is then input into the classifier for prediction. The model is jointly optimized by the classification loss, consistency loss, and evolutionary loss. Compared with other methods, this invention significantly improves the accuracy of classification results with a small amount of labeled data.

The application discloses a kind of based on BIM technology's building feature extraction and mapping method, comprising the following steps: constructing three-level building feature system including microcosmic, mesoscopic and macroscopic features, and being associated with BIM model component by unique ID;BIM model is exported and feature data is extracted, and original sample set is formed;By data enhancement, construct heterogeneous sample library;Three-way scanning is carried out along XYZ axis, and series of profile, plan and three-dimensional point cloud model are generated, and standardized pretreatment is carried out;BIM parameters and multi-view features are input into dual-channel fusion network, and high-dimensional feature tensor is output after tensor fusion and decomposition;Feature tensor is converted into vector graphics by mapping module;Mixed loss function is used to pretrain and fine-tune the network, and finally realize lightweight deployment.The method realizes efficient and automated generation from BIM model to standard engineering drawings, and improves the precision and efficiency of building information extraction and mapping.

This invention discloses a deep learning-based 3D reconstruction method for ship engines, belonging to the field of computer vision technology. The method includes: obtaining 10 input images containing a reference image and neighboring images from 10 images; inputting each of the 10 input images into a pre-trained DAS-MVSNet network, which performs the following steps: sequentially inputting each input image into a feature extraction module, a Deep-ASPP module, and a Conv1 convolutional layer to obtain 10 first feature maps; establishing a reference camera viewpoint and uniformly sampling along the optical center direction within its view frustum to form multiple planes of different depths; mapping each first feature map to multiple planes of different depths to generate 1 feature volume; aggregating the 10 feature volumes into a single cost volume and inputting it into a four-scale 3D convolutional neural network to obtain an optimized depth map; and finally, after traversing all images, generating a point cloud model based on the optimized depth map. This invention improves the accuracy and completeness of 3D reconstruction.

The invention discloses a data processing method and device, equipment, a storage medium and a program product. In response to a first configuration operation input by a user to the target table, configuring an upstream data source associated with the target table, the upstream data source comprising a target client table and a target performance table, and the target client table and the target performance table comprising at least one index; in response to a second configuration operation of the user on each index in the target table, configuring a calculation rule of each index associated with the target table; analyzing a target client table, a target performance table and a calculation rule of each index associated with the target table, and generating a structured query language script; executing the structured query language script, generating a target table, and inserting the target table into a mapping relation sub-table of a physical table and a view of a preset global parameter table to obtain a target view; according to the target view, a database server is connected and accessed to obtain data of the target client table and the target performance table, processing is carried out according to the calculation rules of the indexes, and a target report corresponding to the target table is generated.

The invention discloses a heterogeneous multi-robot collaborative coverage method for a mine environment, and belongs to the technical field of robot path planning. According to the method, firstly, a mine roadway structure is abstracted into a directed graph with characteristic parameters, and a multi-layer virtual graph is constructed based on different robot capabilities; aiming at the characteristics of asymmetric traffic cost and empty driving demand in a mine, an improved enhancement-merging algorithm is designed, and an efficient service route meeting resource constraints is generated. And multi-graph polling and main task graph mapping are combined, and heterogeneous multi-robot collaborative coverage is realized through multi-round iterative planning. According to the method, the path planning efficiency and the task coverage quality are improved, and the method is suitable for heterogeneous multi-robot collaborative coverage under the catastrophe mine.

The embodiment of the invention provides an information extraction method and device, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining original data, including static attributes and dynamic attributes, of a target object, and constructing a global hierarchical attribute graph based on the original data by taking the target object as a root node and taking the dynamic attributes and the static attributes as non-root nodes, mapping the global hierarchical attribute graph to a unified embedding space, and performing information aggregation in the space to obtain basic representation information; a selective aggregation mechanism is adopted to perform secondary information aggregation to obtain enhanced representation information, and then target representation information of the target object is obtained based on the basic representation information and the enhanced representation information, so that a high-order interaction relationship which is implied between static and dynamic attributes and is beneficial to tasks is effectively utilized, a semantic gap is eliminated, and the target representation efficiency is improved. The whole feature extraction process can be automatically carried out, and the performance of the model on complex decision-making tasks is remarkably improved.

The invention discloses an Isin annealing SAT solver based on sparse sensing and hierarchical topology driving, relates to an Isin model technology, and provides the scheme for solving the problems of redundancy and the like in the prior art. The method comprises the following steps: firstly, mapping an SAT problem to an Isinus model by using a mixed topological graph comprising a sparse connection graph and a lattice graph, then solving by using a simulated annealing algorithm of a layered topological structure, and after the problem is successfully solved, reflecting a spin reflection to a variable assignment of an SAT problem expression; a sparse perceptual coding mechanism is adopted to represent a model coefficient; the sparse perception coding mechanism is composed of a sparse format storage mechanism and a weight perception mechanism. The spinning type comprises original spinning and auxiliary spinning; the original spins are emitted to a reconfigurable sparse connection graph, and the auxiliary spins are mapped to a lattice graph. The method has the advantages of good mapping effect, less resource occupation and high solving efficiency.

The invention relates to the technical field of image matching, can be applied to business scenes of financial science and technology, medical health and the like, and discloses an image background similarity determination and task processing method, device, equipment and medium. The method comprises the following steps: extracting image features of a to-be-detected image and a human body key point thermodynamic diagram, generating a foreground probability graph and deducing a background attention mask, modulating the image features by using the background attention mask to obtain a weighted feature graph, mapping the weighted feature graph to a metric space to determine background similarity, and executing a target task related to the image background according to the background similarity. According to the method, attitude estimation and background separation are jointly optimized, and the background attention mask is generated by using the key point thermodynamic diagram, so that the stability and accuracy of background similarity determination in a complex scene are improved.

The application discloses a graph mapping method based on exact synthesis, comprising: constructing an exact synthesis structure library based on NPN equivalence class, and inputting a Boolean network; performing k-Cut cutting enumeration on the input Boolean network and calculating a truth table, matching the cutting with a structure in the exact synthesis structure library through Boolean matching, and realizing fast searching in combination with NPN classification; generating a cover network through multi-round mapping optimization, and mining a logic shared node in combination with structure hashing, wherein the multi-round mapping optimization comprises delay-oriented mapping, global area topology optimization and local exact area optimization; generating a new target network based on the optimized cover network, removing redundant nodes and completing final network construction. Through the automatic mapping and redundant removal mechanism of the whole process, the application does not need manual intervention, significantly reduces the manual cost of the technical mapping link in chip design, and is especially suitable for efficient design of complex circuits and emerging majority logic base technologies.

Embodiments of the present disclosure relate to methods, devices, media and program products for layout mapping. Some embodiments of the present disclosure provide a method for layout mapping. In the method according to the present disclosure, constraint information for layout and routing is extracted from the original layout and original circuit to be mapped, and a new circuit under a new process is automatically laid out and routed based on the extracted constraint information to obtain a new layout for implementing the new circuit. In this way, the scheme according to the present disclosure avoids the misalignment, overlap and design rule checking problems caused by direct replacement of hierarchy or parameterized cells in the conventional scheme to the greatest extent, so that the workload of the designer can be greatly reduced.

The application relates to a social network repeated event detection method based on repeated subgraph pattern mining. The method comprises the following steps: obtaining a social network snapshot graph of a current timestamp, mapping an m-edge subgraph related to a newly inserted edge e in the social network snapshot graph into a corresponding bucket of a first auxiliary data structure by using a first hash function, and updating the frequency and frequency counter state of a subgraph pattern in the corresponding bucket; obtaining the frequency of the updated subgraph pattern, and if the frequency is equal to a support threshold value sigma, mapping the corresponding four-element information of the subgraph pattern into a corresponding unit of a second auxiliary data structure by using a second hash function, and updating the four-element information in the corresponding unit; and obtaining repeated events in the current social network according to the events corresponding to the subgraph pattern in the updated second auxiliary data structure. The method can efficiently detect repeated events on a social network stream graph.

The application provides a subway small obstacle recognition method, system, device and medium, the method comprises: taking a frame subway track area image as input, generating a feature map sequence through an image feature extraction model; for each shallow feature map used for detection, a shallow embedding mapping method is used to map the current shallow feature map and the subsequent shallow feature map to an embedding space feature map with a preset embedding dimension respectively and construct a matching score matrix; each deep feature map is mapped to a deep embedding space feature map with a preset embedding dimension, and a deep feature map matrix is generated by splicing, the deep feature map matrix is weighted and fused to generate a borrowed feature matrix; for each shallow feature map used for detection, the corresponding output feature map is generated based on the current shallow feature map, the corresponding borrowed feature matrix and the context feature information; for each output feature map, the boundary box parameters and the class confidence of a plurality of candidate prior boxes are generated, which are used for outputting the detection result of the target obstacle.

The application discloses a kind of intelligent behavior detection method and system based on motion feature enhancement, construct based on adaptive mixture Gaussian model, the Mahalanobis distance of pixel point and background distribution is calculated and model weight is dynamically updated, realize the accurate decoupling and separation of foreground motion region in video frame;Then, the foreground mask obtained initially is based on the connected domain analysis of area threshold to filter out discrete noise, and the morphological closing operation of structure element adaptation is used to fill target internal cavity and repair edge fracture, to generate the high signal-to-noise ratio of motion region binary graph;Again, the purified binary graph is mapped as high visual saliency color heat map, it is weighted linearly fused with original RGB video frame, the fusion feature map is used to guide deep learning model to pay attention to motion region, and through double threshold detection strategy, the perception and identification ability of system to small target and hidden uncivilized behavior under complex dynamic scene is greatly improved.

The invention discloses a high-capacity low-power-consumption internet of things communication de-noising and de-interference method, which relates to the field of wireless communication physical layer signal processing, and comprises the following steps: carrying out short-time Fourier transform on a received Chirp signal to generate a time-frequency diagram, and decomposing the time-frequency diagram into independent time-frequency units through a sliding window; learning noise distribution through forward diffusion and reverse de-noising processes by using a diffusion generative model, and performing iterative de-noising on the noisy time-frequency matrix to obtain a pure time-frequency graph; recording multi-scale hidden layer features in the denoising process, and obtaining a fusion feature map through spectrum normalization and attention weighted fusion; calculating normalized mutual information based on the pure time-frequency graph and the fused feature graph, and if the normalized mutual information is lower than a threshold value, de-noising again; and finally, mapping the reconstructed time-frequency graph into symbol representation through a decoder to complete adaptive demodulation. According to the method, the communication reliability and demodulation accuracy in a complex electromagnetic environment are improved by effectively solving the problems of low signal-to-noise ratio and multi-packet conflict in a high-density Internet of Things scene.