1403 results about "Similarity computation" patented technology

The invention relates to an experimental data processing method and device, an AI analysis module and computer equipment, and belongs to the field of data processing.The method comprises the steps that multi-dimensional original data are partitioned according to types, and formats are unified; generating a similarity matrix based on time and space neighborhood information, and marking abnormal fluctuation points; effective signals are separated through a time-frequency feature matching noise library; extracting multi-layer features of basic statistics, time sequence correlation and trend change; and dynamically screening core features to update the tracking type experimental model. The matched AI analysis module integrates hardware circuits of data partitioning, similarity calculation, anomaly marking, noise matching, feature extraction and model updating, and whole-process acceleration is achieved. According to the method, through multi-dimensional data compatibility processing, accurate anomaly detection, multilayer feature fusion and model adaptive optimization, the experimental data processing efficiency and conclusion reliability are remarkably improved, and the method is suitable for real-time analysis of multiple scenes such as scientific research and industry.

The invention discloses a multi-modal metadata alignment fusion method and device, equipment and a storage medium, and the method comprises the steps: carrying out the metadata extraction of structured data, unstructured text data and image data, and generating multi-modal metadata with semantic annotations; establishing a shared semantic embedding space, and mapping the multi-modal metadata to the shared semantic embedding space for coding to obtain a unified spatial vector; determining an alignment candidate pair from the unified spatial vector through similarity calculation, and identifying a semantic relationship of the alignment candidate pair; and performing conflict detection on the aligned candidate pairs and the corresponding semantic relationships, and resolving conflicts based on weight weighting fusion to obtain a unified metadata system. According to the method, improvement and optimization are carried out from multiple aspects of multi-modal data processing, semantic understanding, alignment accuracy, conflict resolution and the like, the defects in the prior art are overcome, and a more accurate and comprehensive multi-modal metadata alignment fusion result can be provided.

The invention provides a natural language text data intelligent classification method and system based on deep learning, and relates to the technical field of natural language processing, and the method comprises the steps: 1, employing a context awareness mechanism to analyze the real semantics of a target vocabulary according to an antagonistic variant existing in a text, and obtaining a target vocabulary; in combination with a word meaning library and a pre-training process of a dynamic learning rate adjustment strategy, generating a candidate replacement vocabulary set with consistent semantics; and step 2, based on the candidate replacement vocabulary set, performing multi-dimensional semantic similarity calculation and emotional tendency discrimination, determining applicable vocabularies conforming to an original culture background through a context adaptation strategy, and generating a standardized text sequence. According to the method, through multi-dimensional semantic analysis, cultural context fusion, cross-granularity feature construction and dynamic parameter correction, the accuracy and adaptability of natural language text classification are realized.

The invention relates to the technical field of artificial intelligence, and discloses an education data report content interaction method and system generated based on retrieval enhancement, and the method comprises the steps: judging whether a natural language problem is an education field problem or not through a large language model, and if yes, carrying out semantic analysis to generate a structured query instruction; when the problem relates to cross-document association analysis, retrieving the structured semantic index database to generate a retrieval result set; if policy association analysis is involved, matching a policy knowledge graph by combining semantic similarity calculation and an entity linking technology, and then performing cross-modal fusion processing to obtain a retrieval result set; and inputting the retrieval result set into the retrieval enhancement generation model, and calling an education field language model to generate an analysis report. According to the method, the industrial pain points of inaccurate intention recognition, low cross-document analysis efficiency, incapability of dynamically combining with latest policies and the like in a traditional interaction mode can be solved, the efficiency and quality of data report interaction in the education field are remarkably improved, and the user interaction experience is optimized.

The invention relates to the technical field of fault detection, and discloses a diesel generating set fault detection method and system based on deep learning, and the method comprises the steps: obtaining first vibration signal data, and carrying out the time-frequency decomposition, and obtaining a dynamic change feature; de-noising processing is carried out on the dynamic change features to obtain a time-frequency feature sequence; extracting a peak energy distribution data set, and calculating each frequency band entropy value to obtain a frequency band entropy value sequence; classifying the frequency band entropy sequence, determining a random fluctuation reference mode, and separating to obtain an abnormal frequency component; calculating a spectral line spacing and amplitude ratio, obtaining a spectral line feature data set, classifying the spectral line feature data set, and determining a fault classification result; obtaining current second vibration signal data, performing similarity calculation on the current second vibration signal data and a pre-established normal mode library, and outputting a fault feature vector; and verifying the fault feature vector to obtain a final fault detection result. According to the method, closed-loop diagnosis from signal acquisition to fault classification can be realized, and the fault detection precision of the diesel generating set is improved.

The invention relates to the technical field of security and protection management, in particular to a campus security and protection management system based on deep learning, which improves the accuracy and robustness of identity recognition by acquiring access control card numbers, face images or fingerprint features and generating standardized identity authentication data. And on the basis of a comparison result of the identity authentication data and the campus database, a behavior chain initialization identifier is generated, and accurate identity binding of the school entering personnel is realized. Furthermore, by collecting multi-camera image stream data, pedestrian re-identification and similarity calculation are executed by using a deep feature matching network, and a cross-camera continuous trajectory data set is generated. And matching the behavior track data set with the conventional path template to generate a behavior offset feature vector. And carrying out joint modeling on the behavior offset characteristics and the identity information through a graph neural network model containing an attention mechanism, and outputting a behavior purpose label and a risk grade score. And a graded security response instruction is generated based on the risk score, so that the missing report rate and the false report rate are effectively reduced.

The invention discloses a missing mode gamma collaborative retrieval diffusion method for an open environment, and belongs to the field of multi-mode learning and missing mode processing. According to the method, a human brain multi-source context completion mechanism is simulated, and robust multi-modal learning is realized through three innovative modules: a context retrieval enhancement module: a multi-modal memory library is constructed, related instances are retrieved through similarity calculation under a gating mechanism, and context representation of a missing mode is enhanced; the prompt drive diffusion generation module is used for constructing a semantic prompt based on a retrieval result, fusing a de-noising diffusion probability model through an attention mechanism, and realizing context-aware knowledge migration and missing modal generation; and the inverse gamma noise optimization module is used for establishing a mixed normal-inverse gamma distribution model, dynamically sensing noise, realizing uncertainty estimation in multi-modal fusion and ensuring robustness and reliability of a regression result. According to the method, the dependence of the model on the available modal quality is effectively reduced, and the cross-modal knowledge migration effect and the multi-modal learning task performance are improved.

The invention provides a self-adaptive semantic-driven data set field matching method and system, and the system comprises a data preprocessing module which is used for carrying out the cleaning, standardization and preliminary analysis of an input data set, and extracting a field name, a data type, a field description and a data sample; the deep semantic representation modeling module is used for constructing a field-level semantic representation vector; the multi-level similarity calculation module is used for comprehensively calculating the grammatical similarity, the semantic similarity and the statistical similarity among the fields, dynamically adjusting the weight of the similarity of each level by adopting a weighted fusion algorithm, and generating a comprehensive similarity matrix; and the matching result management and application module is used for generating a field matching mapping table and a fusion suggestion according to the comprehensive similarity matrix. According to the method, high-precision automatic matching of data set fields is realized by fusing deep semantic understanding, multi-dimensional similarity calculation and incremental adaptive learning, and the efficiency and accuracy of data set fusion are remarkably improved.

The invention provides an industrial injury auxiliary identification method and system based on dual-channel retrieval enhancement generation, and relates to the technical field of artificial intelligence and industrial injury auxiliary identification, and the method comprises the steps: obtaining multi-modal data of a wounded movement video, a medical image and a case text; fusing the semantic vectors of the image, the text and the video into a unified semantic representation vector; inputting the semantic representation vector into a dual-channel retrieval enhancement generation model, and after inputting the semantic representation vector into the dual-channel retrieval enhancement generation model, respectively entering a law and regulation structured knowledge graph retrieval path and a historical case semantic retrieval path; extracting the final representation of each node in the regulation knowledge graph through the regulation structured knowledge graph retrieval path, and performing regulation node path extension to obtain a regulation graph matching basis chain; the historical case semantic retrieval path calculates historical related cases through semantic similarity to obtain case core summary information. According to the invention, the auxiliary evaluation efficiency is improved, and the interpretability of the result is enhanced.

The invention discloses an intelligent class case retrieval method and system based on reinforcement learning self-feedback, and the method comprises the following steps: S1, constructing a multi-level semantic understanding framework, carrying out the semantic understanding of a query case through the multi-level semantic understanding framework, and generating a complete semantic representation; s2, carrying out multi-dimensional similarity calculation and sorting optimization on the query case and historical discriminants in the candidate case library; s3, constructing a dynamic user portrait and optimizing a recommendation strategy; and S4, establishing multiple rounds of dialogues and context awareness, and carrying out visual display and feedback mechanism optimization. According to the intelligent class case retrieval method and class case retrieval system based on reinforcement learning self-feedback, a more accurate, efficient and reliable law intelligent retrieval system is finally achieved, judicial practice requirements are met, and quantitative and measurable technical progress is achieved in the aspects of retrieval accuracy, sorting quality, personalized service, law adaptability and the like.

The invention discloses a short message template intelligent identification method and system based on semantic similarity calculation, relates to the technical field of information and communication, and realizes automatic matching of a request short message and a template library through semantic vectorization and similarity calculation; semantic extension, structural constraint and a dynamic feedback mechanism are combined, so that the matching accuracy and robustness are improved; an optimization process driven by a multi-level model and category features is introduced, and an identification-feedback-optimization closed-loop system is formed; the system structurally comprises a semantic modeling module, a similarity calculation module, a dynamic feedback module and a template optimization module, automatic and efficient short message template recognition and updating can be achieved, and the requirements for real-time performance and accuracy in a large-scale application scene are met.

The invention provides a tumor patient clinical test matching system and method based on a large language model and an OCR technology, and is applied to the field of medical data processing. The method comprises the following steps: analyzing clinical data and test information, processing an unstructured text, and generating structured clinical feature data through context association analysis; key data is extracted and subjected to double verification correction, and structured data supplementary information is generated; enhancing the structured clinical feature data and supplementary information based on a multi-modal processing assembly line module, extracting an image quantitative index, analyzing an immunohistochemical result, and generating a comprehensive matching score; through a rule engine and semantic similarity calculation, item-by-item comparison of patient features and entry and exhaust conditions is realized, and a preliminary matching result is generated; edge case misjudgment is corrected through context-aware multi-round reasoning, sorting is adjusted in combination with clinical test priority weights, and an optimized clinical test matching list is generated; and generating a clinical test matching report based on the data.

The invention provides a project cost big data management and analysis system, and relates to the technical field of data management, and the system comprises a data collection and preprocessing module which is used for collecting original cost data from a heterogeneous data source, and carrying out the preprocessing of the original cost data, and obtaining the preprocessed cost data; the semantic feature extraction module is used for converting the preprocessed cost data into a multi-dimensional feature vector based on a multi-level feature extraction system; the similarity calculation module is used for calculating similarities among different cost data based on the multi-dimensional feature vectors to obtain a similarity matrix; and the data storage and management module is used for storing the cost data, the multi-dimensional feature vector and the similarity matrix by adopting a mixed storage architecture, and providing retrieval and recommendation functions of cost projects based on a multi-level feature space index structure. According to the method, the limitation that a traditional method only depends on keyword matching is solved, and the system can recognize the deep incidence relation between the items.

The invention relates to the technical field of Internet, and discloses an Internet hotspot data mining system and method based on artificial intelligence, and the system comprises a data collection module, a semantic understanding module, a dynamic clustering module, a popularity evaluation module, a visual output module and a feedback optimization module. By setting a multi-source calibration unit, when cross-platform hotspot data acquisition is carried out, standardized processing of multi-source heterogeneous data acquisition is ensured by establishing a dynamic semantic feature library and configuring adaptive weight parameters for different data platforms; meanwhile, by monitoring the semantic offset of the collected data in real time, characterization deviation generated during cross-platform data collection can be detected and eliminated, the accuracy of hotspot clustering similarity calculation is guaranteed, hotspot recognition errors are reduced, and by deploying a semantic tracking engine, when hotspot event propagation analysis is carried out, the analysis accuracy is improved. The deviation degree of event core elements is calculated by constructing spatio-temporal feature vectors, and whether topic semantics are migrated or not is judged in real time.

The invention discloses a multi-agent interaction TOKEN compression strategy method and system based on LLMLINGUA, and belongs to the technical field of artificial intelligence and natural language processing. Aiming at the problems of high communication load and difficulty in semantic recovery caused by high-frequency interaction in a multi-agent system, the invention adopts the following technical scheme: inputting text data, and performing word segmentation and semantic analysis to form a low-dimensional semantic vector; dynamically selecting a compression strategy type according to semantic density and serial number density characteristics; the LLMlingua module is used for carrying out depth coding to generate a low-dimensional semantic identifier; storing the identifier in a dynamic reference library and distributing a unique reference code; reference codes and context information are transmitted among the intelligent agents; self-adaptive grouping is achieved through cosine similarity calculation, similar group debate is skipped, and debate between different groups is triggered; the modules are integrated to a multi-agent system. According to the method, the communication load and calculation overhead are remarkably reduced, semantic high-fidelity restoration is guaranteed, and the system response efficiency is improved.

The invention provides a bidirectional visual semantic interaction enhancement method and system for generalized zero sample learning, and the method comprises the steps: collecting visual samples containing visible classes and unvisible classes and corresponding cross-class semantic descriptions in a generalized zero sample learning scene; respectively generating a global visual feature vector and a semantic word vector through a visual feature extractor and a semantic encoder; fusing the two by using a multi-head attention mechanism to generate a semantic enhanced visual embedding representation; then visual-to-semantic and semantic-to-visual two-way comparative learning with expert attribute vectors as alignment targets and starting points is carried out, and visual semantic embedding and attribute vector alignment are promoted in combination with attribute regression loss; then training a bidirectional visual semantic interaction enhancement model based on multi-loss joint optimization; and finally, calculating visual semantic embedding for a test image by utilizing the trained model, and outputting a classification result of a visible class or a non-visible class by combining similarity calculation with a candidate class attribute vector with a self-calibration item.

The invention provides a traffic scene multi-target detection method and system based on deep learning, and relates to the technical field of traffic, and the method comprises the steps: carrying out the semantic prior driven multi-scale feature extraction of a multi-frame image, and carrying out the point-by-point fusion; obtaining a motion field through optical flow estimation and feature similarity calculation, and executing motion compensation to obtain a moving target mask; obtaining a static target boundary by using boundary regression decoupling and geometric consistency constraint; and finally, moving and static target results are combined, and quadratic regression is executed based on consistency evaluation. The dynamic and static targets in the traffic scene can be effectively detected, the boundary regression precision is improved, and the false detection rate caused by shielding is reduced.

The invention discloses a zero-sample liquid crystal display screen defect detection method. According to the scheme, the method comprises the following steps: 1) collecting and processing image data of a display screen; 2) constructing an image-text comparison pre-training model (CLIP) to carry out text-image similarity calculation; 3) designing a self-adaptive prompt network, combining static and dynamic prompts and designing an optimal fusion weight, realizing self-adaptive combination of prompt semantics, and improving the adaptability of the model; (4) semantic separation loss is added into the global features of the text, and the semantic separability of the normal text and the defect text is improved; and 5) before image classification, a feature improvement module for abnormal guidance is inserted into classified visual features to enrich the visual features, and the ability of alignment with the text is further improved. The method is suitable for a cold start stage of display screen defect detection, can solve the problems that in display screen zero sample detection, 'defect 'semantics are difficult to understand, and normal defect attributes are difficult to separate, and remarkably improves the classification and positioning capability of display screen defects.

The invention relates to the technical field of load prediction, and provides a charging load prediction method and system based on comprehensive similarity similar day screening, and the method comprises the steps: carrying out the similarity calculation and normalization of obtained historical load features, meteorological features and context features; by taking a mean value and a standard deviation of minimum fusion similarity scores as a target, solving to obtain a fusion weight of each similarity and then constructing a standard similar day set; calculating day pair features of similarity between meteorological features and context features of a to-be-predicted target day and a candidate day as input, and screening out a most matched similar day identification set through the trained model; and obtaining a load prediction result based on the meteorological features and context features of the target day to be predicted and the screened similar day identification set. Through multi-feature comprehensive similarity calculation and optimized similar day screening, similar day set construction based on multi-dimensional features is realized, and a high-precision input feature selection framework is provided for charging load prediction.

The invention discloses a radar communication radiation source identification method and system based on multi-modal feature alignment, and belongs to the technical field of electronic reconnaissance and signal processing. The method comprises the following steps: preprocessing a received radar signal to generate a standardized time-frequency graph; extracting a signal feature vector through a specially designed convolutional neural network encoder; extracting a text feature vector by using a Transform encoder; through joint optimization of cosine similarity loss and physical parameter constraint loss, alignment of signal-text features in a unified vector space is realized; and finally, zero sample identification of unknown radar signals is realized through vector similarity calculation. According to the method, the problems of low recognition rate and insufficient cross-modal information fusion in a low signal-to-noise ratio environment of a traditional method are effectively solved, and the recognition precision and the system robustness are remarkably improved.

The invention relates to the technical field of semantic understanding, in particular to an environmental impact report auxiliary auditing system based on semantic understanding, which comprises a text cleaning module, a chapter segmentation module, a rule generation module, a content auditing module and a result output module. According to the method, the paragraph structure and the layout identification are analyzed in a unified mode, key paragraph types in environmental influence chapters are efficiently recognized, the text integration accuracy is improved by combining inter-paragraph similarity calculation and repetition rate screening, chapter boundary accurate positioning and affiliation adjustment are achieved on the basis of format feature comparison of serial numbers and title styles, and the text integration efficiency is improved. A multi-layer parameter alignment rule set is constructed to support comprehensive verification of standard numbers, monitoring frequencies and periodic elements, an exception labeling mechanism is combined to complete difference item logic judgment and consistency evaluation, an auditing basis with rule driving and data comparison capabilities is formed, chapter positioning precision, element extraction comprehensiveness and logic exception recognition efficiency are improved, and the method is suitable for popularization and application. And the pertinence and the automatic processing depth in the auditing process are enhanced.

The invention discloses a private network content copyright monitoring and evidence obtaining system based on AI and a large model, which utilizes AI and large model technologies to carry out copyright monitoring and evidence obtaining on multimedia content in a private network environment, and carries out deep semantic understanding and cross-modal feature extraction through a large model processor to generate unified semantic representation. The method comprises the steps that a copyright content database is established and stored in a copyright content knowledge base, the copyright content knowledge base is used for storing metadata of original content protected by copyright, unified semantic representation and copyright declarations in a natural language form provided by a copyright party, and the copyright declarations are converted into query vectors through a semantic understanding technology; the method comprises the following steps: acquiring a semantic representation of a multimedia content, performing similarity calculation with the semantic representation of the multimedia content, identifying infringement content, when the infringement content is identified, recording an original source, publishing time and publisher information of the infringement content, performing differentiation analysis, generating an evidence chain, displaying a copyright monitoring result through a user interface, generating infringement alarm information, and presenting details of the evidence chain.

The present invention relates to the technical field of intelligent out-of-distribution fault detection for construction machinery, and discloses an out-of-distribution fault detection method and system based on energy propagation and graph learning, and the method includes: acquiring vibration acceleration signals in typical fault states, carrying out similarity calculation to obtain an adjacency matrix composed of the maximum mutual information coefficients, and taking the adjacency matrix as input in a graph neural network; carrying out feature extraction on the adjacency matrix through adopting a GraphSage graph convolution method, and generating each node representation; calculating an energy score of each node, and distinguishing between in-distribution data and out-of-distribution data; and enhancing out-of-distribution data confidence estimation for each node, and carrying out out-of-distribution data identification and out-of-distribution data detection under different working conditions of a rolling bearing.

The invention discloses a large model calling system and method based on an MCP tool chain, and the method comprises the steps: receiving user input, and carrying out the preliminary matching of tools: querying whether the user input contains an entity in an entity database or not through a fuzzy matching technology, and carrying out the preliminary screening of a plurality of candidate tools based on a matching result, so as to obtain limited to-be-selected tools; user input is converted into semantic vectors, semantic matching is conducted in a vector database, and semantic similarities of all tools to be selected are ranked through a similarity calculation method; extracting key information in user input; abstracting the user input into a specified format; and matching the specified format with the historical workflow template, and if the historical workflow template is matched, replacing the key information to a corresponding position of the historical workflow template by adopting a structure-preserving parameter replacement algorithm to realize workflow reuse. According to the method and the device, a complete technical link from accurate tool screening to intelligent workflow multiplexing can be realized.

The invention relates to an intelligent procurement collaboration method and system, and the method comprises the following steps: S1, building a supplier knowledge graph based on multi-source heterogeneous data fusion through employing a graph embedding algorithm, achieving the associated storage of industrial and commercial information, performance records and quality reports through a Neo4j graph database, calculating the weight of a supplier node through employing a PageRank improved algorithm, and carrying out the calculation of the weight of the supplier node; s2, based on the three-dimensional supplier portrait matrix, using an improved collaborative filtering algorithm to carry out demand matching, analyzing a purchase demand document through an ElasticSearch semantic analysis engine, combining TF-IDF weighted cosine similarity calculation to realize intelligent recommendation, and generating a purchase demand scheme with a weight score. The supplier knowledge graph is constructed through the multi-source heterogeneous data fusion and graph embedding algorithm, the scattered industrial and commercial information, performance records and quality reports are stored in an associated mode, traditional data island limitation is broken through, and effective integration of supplier multi-dimensional features is achieved.

The present invention relates to the technical field of artificial intelligence, and provides a diversity-enhanced text retrieval-augmented generation method and system. The method comprises: using knowledge data to construct a local knowledge base; acquiring a user question and the desired number of retrieval results; in a retrieval stage, using a large language model to paraphrase the user question to obtain a paraphrased question, then expanding the retrieval range, expanding the number of retrieval results, and performing vector retrieval in the knowledge base on the user question and the paraphrased question to obtain retrieval results; calculating the similarity among the retrieval results, and screening for diverse retrieval results on the basis of the similarity calculation result; and integrating the screened retrieval results and the user question by means of prompt engineering, inputting the integrated result into the large language model, and taking a model generation result as an answer to be returned to the user, thereby providing the user with richer and more informative search results, and improving the richness and accuracy of content generated therefrom.

A long-term continuous learning method based on task core memory management and consolidation aims to enable a model to sequentially learn from a large number of task sequences, new knowledge is obtained, information of previous learning is reserved, and the method is similar to a human learning mode. The method comprises the following steps: 1) task input and instruction fine tuning; 2) performing difference analysis on the model parameters of the current task and the previous task, identifying a task core memory unit, calculating an adaptive weight based on task prototype similarity, and dynamically updating the memory unit; 3) constructing an experience playback buffer area through a difficult sample selection strategy and a difference sample selection strategy; and 4) utilizing the joint loss function training model to keep the memory of the historical tasks while learning the new tasks. According to the method, the problem of disastrous forgetting in long-term continuous learning is mainly solved, and the performance of the model in a long-term sequence task is remarkably improved.

The invention belongs to the computer vision technology, and particularly relates to a small sample multi-modal prototype learning-oriented remote sensing scene classification method, which comprises the following steps of: acquiring RGB (Red, Green and Blue) images with category labels and text prompts of the RGB images as a support set; establishing a text prototype, an RGB prototype and a hyperspectral prototype of each category according to the support set; and extracting to-be-classified query set image features by using a pre-trained CLIP image encoder, calculating cosine similarities between the query set image features and the text prototype, the RGB prototype and the hyperspectral prototype of each category of the support set, taking the cosine similarities as input of a multi-layer perceptron, and obtaining the category of the to-be-classified RGB image through classification of the multi-layer perceptron. High-precision and high-robustness remote sensing scene classification is realized under the small sample condition, only prototype and similarity calculation is needed in the reasoning stage, and deployment and expansion are easy.

The invention discloses a work order similarity calculation method and device based on multi-modal feature fusion, equipment and a storage medium, and relates to the technical field of computer data processing, and the method comprises the steps: extracting multi-modal features from work order data, the multi-modal features including text semantic features, structured features and time sequence behavior features; mapping the multi-modal features to a unified semantic space for multi-level fusion processing to obtain fusion features; generating a dynamic weight according to work order type information contained in the work order data and query intention information contained in the received query request, and performing weighted fusion processing on various similarity indexes obtained by fusion feature calculation to obtain a final similarity; and the contribution value of each modal feature to the similarity is calculated based on the final similarity, and a visual interpretation result is generated according to the contribution value, so that the accuracy, adaptability and interpretability of cross-platform work order matching are remarkably improved.

The embodiment of the invention provides a knowledge graph fusion method and system based on a large model, and the method comprises the steps: carrying out the data standardization, entity feature enhancement and relation semantic annotation of heterogeneous knowledge graphs from different sources; through semantic similarity calculation, context reasoning and alignment confidence evaluation, a multi-stage and multi-mode entity alignment mechanism is constructed in combination with a large language model, and cross-source entity matching is performed on entities in heterogeneous knowledge maps of different sources; based on conflict detection, dynamic weight distribution and a conflict resolution strategy, generating a data fusion result, and unifying multi-source data; and generating a high-quality unified knowledge graph through missing relationship prediction, logic consistency verification and an incremental updating mechanism. According to the knowledge graph fusion method and device, full-process automation, precision and dynamics of knowledge graph fusion are realized, the problems of high manual dependence, weak semantic processing capability, poor cross-domain adaptability and the like in the prior art are effectively solved, and the efficiency and quality of knowledge graph fusion are remarkably improved.