541 results about "Learning architecture" patented technology

The invention discloses a multi-modal medical image data intelligent processing system, relates to the field of medical image analysis, and is applied to multi-modal medical image whole-process analysis of CT, MRI, PET, ultrasound and the like. According to the system, different modal image features are extracted and fused through a cross-modal manifold fusion network; a semantic guidance dynamic registration engine optimizes registration parameters to ensure that the registration error is less than or equal to 1.5 mm; the multi-task collaborative diagnosis network realizes multiple tasks such as disease classification; the clinical knowledge embedding and interpretable module generates a structured report and is in butt joint with an HIS system. Meanwhile, the model is optimized through a federated learning architecture, the adaptability of newly added data is improved by more than or equal to 20%, and intelligent processing and analysis of multi-modal medical images are realized.

The invention discloses a metal cutting process parameter optimization analysis method based on machine learning, and particularly relates to the field of machine learning. Comprising multi-dimensional process parameter feature extraction and preprocessing, cutting state intelligent identification based on integrated learning, dynamic process parameter sensitivity analysis and weight calculation, process parameter intelligent optimization under a multi-target constraint condition, and adaptive parameter adjustment and real-time control strategy. According to the method, the interaction relationship between complex nonlinear features and process parameters in the cutting process is comprehensively captured, and accurate and intelligent recognition of different cutting states such as normal cutting, tool abrasion and abnormal flutter is achieved through a three-layer integrated learning architecture; the technical bottlenecks that an existing system lacks real-time self-adaptive adjustment capacity and is low in process optimization efficiency are overcome, pertinence and effectiveness of parameter adjustment are ensured, and the technical current situation that machining quality fluctuates and repeatability is poor due to traditional fixed parameters is changed.

Deep learning tools such as convolutional neural networks (CNNs) and transformers have spurred great advancements in computational biology. However, existing methods are constrained architecturally in context length, computational complexity, and model size. This application introduces a sub-quadratic architecture for modeling, which combines projected gated convolutions and structured state spaces to achieve local and global context with, for example, single-nucleotide resolution. These models outperform CNN-, GPT-, BERT-, and long convolution-based models in many tested genomics tasks without pre-training and with 4×-781× fewer parameters. In the proteomics domain, these models similarly outperform pretrained attention-based models, including ESM-1B and TAPE-BERT, on remote homology prediction without pre-training and while using 3,308×-23,636× fewer parameters.

The invention discloses an intelligent load balancing method and system based on multipath fusion, and relates to the field of network load balancing. A distributed monitoring system is constructed to collect network performance indexes, and a Transform model is used to predict traffic; a hierarchical reinforcement learning architecture is adopted, a global strategy is generated according to a macroscopic network state, and flow distribution is optimized for a single path; kalman filtering and particle filtering are automatically switched according to the path stability; the flow is flexibly migrated based on a genetic algorithm; an index weight is calculated by using a Shapley value method and an entropy weight method, and path quality is evaluated; the distribution strategy is executed through the SDN controller, and the overall performance of the network is improved. According to the invention, network abnormity is quickly responded, the traffic migration efficiency is improved, and bandwidth waste is reduced; service differentiation scheduling is supported, the service quality is guaranteed, and the attack defense capability is enhanced; operation and maintenance efficiency is improved, fault positioning time is shortened, and efficient utilization of network resources and guarantee of service quality are realized.

The invention discloses a wind turbine generator performance dynamic evaluation method and system based on multi-source data fusion, and the method comprises the steps: synchronizing multi-modal heterogeneous data through a quantum encryption algorithm and an edge gateway; constructing a space-time semantic graph network through a deep semantic analysis technology, and generating a precise space-time feature matrix; based on an attention mechanism, generating a physical enhancement feature vector; constructing cross-working-condition health index mapping by applying a machine learning algorithm, and quantifying a cross-working-condition comparable health index; a hierarchical incremental learning architecture is adopted, and a multi-objective optimization algorithm is utilized to generate a dynamic maintenance priority sequence; and building a digital twin platform, performing closed-loop verification on the maintenance priority sequence, and generating a unit maintenance scheme. The problem that multi-source data fusion of the wind turbine generator is difficult is solved, the health index quantification accuracy under the complex working condition is improved, the dynamic maintenance strategy is optimized, and closed-loop optimization of the maintenance scheme is achieved.

The invention discloses an air-space-ground caching and resource optimization method and system based on a hierarchical intelligent agent, and aims to construct a deep reinforcement learning architecture in which a high-layer DQN and a low-layer DDPG are coordinated for high dynamics and information uncertainty of an air-space-ground integrated network. The high-level intelligent agent generates a period-level content cache and access control strategy based on global states including a cache state, a task request, a node resource and the like, and the low-level intelligent agent executes time slot-level resource allocation, task unloading rate control and UAV deployment optimization under the constraint of the high-level strategy. The system triggers low-level optimization through a double-stage reward mechanism and constraint verification, evaluates a strategy effect based on time slot income and full-period income, and combines state perception and experience playback technologies to realize collaborative optimization of high and low-level decisions. A simulation result shows that compared with a traditional method, the method has remarkable advantages in the aspects of reducing content acquisition delay, reducing return communication overhead, improving task processing success rate and the like, and the space-air-ground MEC network resource utilization rate and user experience are effectively improved.

The invention discloses an intelligent monitoring decision-making method based on a knowledge graph and federal learning, and the method comprises the following steps: S1, collecting and preprocessing multi-source health data of a user, and generating a health data set; s2, constructing a local medical knowledge graph and performing knowledge embedding modeling to generate a knowledge representation vector; s3, constructing a health risk assessment model, and performing modeling in combination with knowledge representation and health data; s4, initializing a federated learning architecture, setting a client and an aggregation end, and distributing a model structure and parameters; s5, locally training the model by each federated client, and uploading parameters to an aggregation end to complete parameter aggregation; s6, combining the updated model with the real-time health data and a knowledge graph reasoning result to generate a personalized monitoring decision; and S7, collecting user feedback and newly added data, updating the knowledge graph and the model, and entering a new round of optimization. The method is used for realizing personalized health risk assessment and intelligent monitoring fusing the knowledge graph and federal learning while ensuring privacy.

The invention discloses a marine environment real-time monitoring and early warning system based on machine learning, and relates to the technical field of machine learning. Comprising the steps that an ocean multi-source sensing module collects ocean environment data in real time through a sensor and a combined collection scheme; the multi-source feature extraction module performs time domain, change rate and frequency domain feature analysis on the data to construct a unified multi-dimensional feature vector; the multi-model fusion prediction module outputs a marine environment state vector through dynamic weighting and deviation correction based on a parallel learning architecture of a deep neural network, a long-short-term memory network and a one-dimensional convolutional neural network; and the ocean risk identification and early warning module generates graded and classified early warning information through double study and judgment of a sea condition classifier and an abnormal event detector. According to the method, comprehensive acquisition, deep feature mining, high-precision prediction and accurate early warning of marine environment data are realized, the problems of low prediction precision, risk identification lag and the like in the prior art are effectively solved, and reliable guarantee is provided for marine operation safety.

The invention relates to the field of artificial intelligence medical application, and discloses a lightweight intelligent traditional Chinese medicine inquiry system and a construction method thereof, and the system comprises a multi-dialect adaptive speech recognition module, a traditional Chinese medicine intelligent dialogue large language model module, a natural speech synthesis module, and a continuous learning mechanism module. The multi-dialect adaptive speech recognition module is used for converting dialect speech input of a patient into a standard text; the traditional Chinese medicine intelligent dialogue big language model module is the core of the system and is used for carrying out natural language understanding, dialectical reasoning and inquiry dialogue generation, and the natural speech synthesis module is used for converting a text response generated by the system into speech output; and the continuous learning mechanism module realizes continuous optimization of the large language model through incremental learning architecture and clinical feedback integration. According to the method, while the professional traditional Chinese medicine diagnosis capability is maintained, the calculation complexity is remarkably reduced, and the universality and sustainable development capability of system application are improved.

Artificially intelligent systems and methods for financial coaching provide personalized, fiduciary-compliant financial guidance through advanced machine learning architectures with measurable performance criteria. The systems implement privacy-preserving processing pipelines that detect personally identifiable information using multi-layered pattern recognition including regular expressions for formatted data sequences, named entity recognition with confidence thresholds above 0.85, and contextual analysis algorithms. A multi-step artificial intelligence processing workflow includes automated language detection, emotional tone classification with confidence scoring, financial profile transformation using predefined templates, context-aware question rephrasing, and semantic similarity matching employing vector embeddings with financial domain vocabulary weighting applying multiplier values between 1.3-2.0. Specialized training methodologies expand datasets through mathematical transformation functions utilizing statistical standard deviations with incremental variations between 0.5-2.0. Mood-based escalation logic automatically transfers users to human advisors when emotional indicators exceed confidence thresholds above 0.8. The systems maintain response times below 5 seconds while providing regulatory compliance through curated content sources and predefined fiduciary instruction parameters.

The invention discloses an AI-based automatic production line scheduling system in an industrial internet, which relates to the technical field of production scheduling and comprises a production plan management module S1, a dynamic scheduling engine module S2, a resource scheduling module S3, a real-time monitoring system module S4, an exception handling center module S5 and a data optimization platform module S6. In the industrial internet, an AI-based automatic production line scheduling system, an X dynamic scheduling engine millisecond response and a multi-agent reinforcement learning engine based on a federated learning architecture realize millisecond response scheduling, each device is used as an autonomous decision-making unit, and dynamic coordination is performed through a distributed Q learning algorithm, so that the vacancy rate of the devices is greatly reduced, and the scheduling efficiency is improved. According to a long-short-term memory network deep analysis model of order delivery cycle compression, emergency order insertion response speed improvement, multi-modal AI quality monitoring, fusion of vibration, thermal imaging and current spectrum, the detection rate is greatly improved compared with a unified sensor, causal reasoning and root cause analysis are performed, a fault causal graph is constructed to position a deep problem, and the average repair time is shortened.

The invention relates to an unmanned aerial vehicle group-oriented sensing communication integrated network multi-resource joint scheduling method, and belongs to the technical field of wireless communication, and the method comprises the steps: building a system model of multiple UAV-ISAC tasks, and defining a joint optimization problem; extracting spatio-temporal features from the dynamic heterogeneous graph in which the unmanned aerial vehicle, the user and the sensing target are abstracted as nodes and the relationship is abstracted as edges; taking the features as input, and adopting a layered multi-agent reinforcement learning architecture to solve the joint optimization problem on line; in the architecture, resource allocation and trajectory planning actions are generated through cooperation of a central Actor and all unmanned aerial vehicle Actors, and system performance is evaluated by a central Critic; constructing a multi-target weighted reward function, stabilizing a training process by combining experience playback and a Mini-batch sampling mechanism, and updating network parameters in parallel; and obtaining an optimal resource allocation and unmanned aerial vehicle trajectory strategy through training. The sensing performance is improved, the communication quality is guaranteed, and the defects in the aspect of dynamic resource scheduling in the prior art are overcome.

The invention provides an arch dam shape efficient intelligent optimization method and system based on an agent model. The method comprises the steps that an arch dam physical-numerical model is constructed, a double optimization target with structural safety and economical efficiency as the core is determined, design parameters are selected, a constraint function is set, and an evaluation index system is established; samples are generated through Latin hypercube sampling, a data set is constructed in combination with finite element calculation, and a multi-task learning architecture is adopted to train a high-precision agent model. And then, coupling the proxy model with a multi-objective optimization algorithm, quickly searching a Pareto optimal solution set, and screening out a comprehensive optimal figure by using a multi-attribute decision-making method. And finally, through a finite element simulation verification result, prediction precision and performance improvement are ensured. The method has the advantages of lightweight modeling, efficient prediction and accurate search, and provides an effective tool for intelligent optimization and rapid decision making of hydraulic structures such as arch dams and the like.

The invention discloses a protocol conversion and protocol self-learning method and system for transformer area optical storage and charging cooperation, and the method comprises the steps: constructing a digital twinborn body of a transformer area optical storage and charging system, carrying out the parallel operation of a protocol agent and a protocol agent in a virtual environment, and achieving the cooperative training of the protocol agent and the protocol agent through a hierarchical reinforcement learning architecture, the protocol agent is responsible for learning a dynamic priority scheduling and compression strategy for heterogeneous protocol messages such as Modbus, CAN and IEC 104, the protocol agent is responsible for learning a power balance and voltage stability control strategy based on photovoltaic output, energy storage SOC and charging load, and the two agents realize cross-domain collaborative optimization through a reward function mutual coupling mechanism. And finally, safely deploying the collaborative strategy obtained by training to the edge control equipment of the physical transformer area. According to the method, the problems of disjunction of protocol conversion and cooperative control, protocol strategy solidification, insufficient cross-domain cooperation and the like in the prior art are solved, and the operation efficiency and the self-adaptive capability of the transformer area optical storage and charging system are improved.

Various embodiments of the present disclosure provide machine learning architectures and data processing techniques for improving computer-based text comprehension. The techniques may include identifying a plurality of data entity tokens from a target section of a multi-section natural language document and generating, using an embedding layer of a semantic chunking model, a text span embedding for a text span of the target section. The techniques may include leveraging the semantic chunking model to generate an attended span representation for the text span based on the text span embedding and the plurality of data entity tokens. The techniques may include identifying an entity topic that corresponds to the text span based on the attended span representation and, responsive to an identification of the entity topic, generating a subgraph data object for a knowledge graph using the text span.

The invention discloses a flow field video generation method based on a policy value architecture and online physical exploration, and belongs to the technical field of crossing of artificial intelligence and computational fluid dynamics (CFD), and the method comprises the following steps: step 1, constructing an unsteady flow field multi-modal training data set, step 2, constructing a generative network system based on an Actor-Critic architecture, step 3, constructing an unsteady flow field multi-modal training data set, and step 4, constructing an unsteady flow field multi-modal training data set. Step 4, supervised fine tuning training is carried out in the first stage; step 5, online physical exploration of a generator is carried out in the second stage; step 6, feedback co-evolution of a physical encoder is carried out in the third stage; and step 7, reasoning generation of an unsteady flow field video is carried out. According to the method, a reinforcement learning architecture containing an Actor and a Critic is constructed, a physical equation is packaged into a digital environment, and a training strategy of basic supervision fine tuning, online physical exploration of a generator and coevolution feedback of an encoder is adopted.

A marine zooplankter identification method based on in-situ images and deep learning is used for processing a multi-stage series deep learning architecture constructed for all in-situ images, and comprises the following steps: quickly positioning and framing zooplankter individuals under a complex background based on a YOLOv8 skeleton construction model; performing pixel-level fine segmentation on individuals in the in-situ image by using a U-Net model, and extracting an accurate contour to obtain a high-quality individual image; and finally, inputting the image into a PlanktonNet network innovatively designed based on a ViT architecture, and through adaptive small-size slice embedding and introduction of a convolution word embedding layer, improving small-scale target feature extraction capability, and realizing high-precision and fine-grained classification of genera and species. The method has the advantages that target detection, semantic segmentation and recognition tasks are organically fused, the problems of low recognition efficiency and low automation degree in the prior art are effectively solved, and marine zooplankton can be quickly and accurately recognized from in-situ images on a large scale.

The invention provides a control system and method applied to a smart home terminal and a terminal, is applied to the technical field of smart home Internet of Things, and has the advantages that a user behavior prediction model driven by machine learning and a dynamic rule generation engine are introduced, so that the system can break through a traditional three-level linkage mechanical execution framework; an intelligent control closed loop with environment perception, intention reasoning and autonomous decision-making capabilities is constructed, multi-dimensional data fusion analysis from equipment state parameters and user historical operation tracks to real-time environment variables is realized, and a composite control strategy including equipment linkage sequence optimization, scene parameter dynamic calibration and abnormal operation self-correction is automatically generated. And in combination with a federated learning architecture, cross-user knowledge migration and group behavior pattern mining are realized, and the generalization ability and scene adaptability of a control strategy are continuously improved.

The invention discloses a data deep learning and intelligent analysis method based on an AI artificial intelligence technology, and relates to the technical field of basic AI models, and the method comprises the steps: employing a multi-modal data preprocessing module to carry out the expansion of small sample data through a generative model, and combining with meta-learning to extract prototype features, meanwhile, an epsilon-differential privacy budget is dynamically allocated based on the data sensitivity level so as to inject dynamic noise; establishing a layered federated learning architecture, training a model by local training nodes through a loss function containing a self-adaptive regularization item, and performing sparse processing and gradient disturbance before uploading parameters; the global aggregation node adopts a weighted federated average algorithm to aggregate parameters, and dynamically adjusts the communication frequency according to the loss convergence speed; and a target model is obtained through iterative training, and a decision interpretation report containing the attention thermodynamic diagram and the desensitization identifier is generated when a result is output. According to the method, the problems of small sample overfitting, data islands and privacy disclosure are effectively solved, and the accuracy and practicability of the model are improved.

The invention relates to the technical field of user behavior data processing, and discloses a user behavior analysis and personalized recommendation method and system, and the method comprises the steps: collecting multi-dimensional behavior data of a user, and carrying out the feature extraction of the multi-dimensional behavior data; key features are defined based on behavior data, redundant features are reduced through a feature selection algorithm, the dimension of numerical features is unified through normalization processing, and an input feature vector used for a deep learning model is generated; processing the input feature vector by adopting a hybrid deep learning architecture, and generating a user-commodity interaction model in combination with the user behavior sequence and the commodity features; performing real-time prediction on the new user behavior data based on a deep learning model, and dynamically adjusting a recommendation strategy according to a prediction result; and displaying a user behavior analysis result through a visual interface, and continuously optimizing a recommendation strategy in combination with an A / B test framework. According to the method, the defects of a traditional recommendation system in the aspects of multi-source data processing, recommendation individuation and the like are effectively overcome.

The invention discloses a distributed photovoltaic short-term power prediction method and system, a medium and equipment, and belongs to the technical field of new energy power prediction, and the method comprises the steps: building a sample set based on distributed photovoltaic station power data, meteorological measured data and meteorological forecast data of all participants, and dividing the sample set into a training set and a test set; through a federated learning architecture based on edge equipment-master station server-cloud server and in combination with a client selection mechanism, each participant utilizes an artificial intelligence algorithm to cooperatively train a prediction model; after the performance of the model is evaluated, the model meeting the precision requirement is used for prediction. According to the method, the problems of low communication efficiency and poor training effect of small samples and traditional federal learning architecture can be solved, and the prediction precision is improved while secure sharing of data is ensured.

The invention provides a seed screening method and system based on a laser radar technology, and relates to the technical field of agricultural automation, and the method comprises the steps: extracting geometric features and defect features from optimized point cloud data; extracting mildew and insect pest spectral features from the optimized spectral data; a multi-dimensional feature set is obtained; calculating seed plumpness, roundness and surface roughness based on the optimized point cloud data; calculating a sag index and a crack feature based on point cloud edge detection; fusing the geometric features, the defect features and the pest spectral features to generate a comprehensive score; a machine learning model is utilized to execute final grading, and a first probability predictor based on laser radar features and a second probability predictor based on pest spectral features are trained respectively; and inputting the probability outputs of the two predictors into a second-layer grading model of the meta-learning architecture, and outputting a final quality grade. According to the method, efficient and accurate seed screening is realized through automatic process and algorithm design.

The invention discloses a garment supply chain risk analysis method and system based on artificial intelligence, belongs to the technical field of garment supply chain analysis, and effectively enhances crisis pre-judgment and response capability of a supply chain through a dual mechanism of simulation deduction and dynamic optimization. A built-in game learning framework of the system can autonomously construct a multi-factor coupled complex interruption scene, for example, an extreme condition that a port is closed in a typhoon season, cross-border tax policy mutation and alternative raw material transportation route congestion occur at the same time is simulated. The deduction not only reveals the problem of excessive dependence of a single node, which is difficult to discover by traditional auditing, but also can verify the practical feasibility of a standby scheme, and helps an enterprise to establish a multi-layer defense system. The continuous self-optimization characteristic of the system enables the system to be prominent in response to novel challenges, and when the international logistics network is suddenly adjusted, a supplier recombination scheme considering both cost and time efficiency can be quickly generated, and the potential chain breakage risk is resolved in the germination stage.

The invention relates to the field of power data processing, in particular to a power transformer fault prediction method and system based on digital twinning. Comprising the following steps: acquiring a focusing data set, and constructing a to-be-calibrated reference model comprising multi-physics field coupling; adjusting preset uncertainty parameters through a Bayesian calibration method, and outputting a digital twinborn model; selecting a fault anchor point by adopting a filling space experimental design method, simulating the fault anchor point to generate a sample for training an agent model, and generating a virtual fault sample set by utilizing the agent model; performing enhancement processing on the virtual fault sample set according to noise characteristics of the sensor to obtain an enhanced fault sample set; spatial features and time sequence features are extracted to form a feature vector data set, and the feature vector data set serves as training data to train a fault prediction model through a multi-task learning architecture. According to the method, a large number of fault samples are generated through the digital twin model and the proxy model, and comprehensive decision support is provided for state maintenance and health management of the transformer.

The present disclosure is directed toward systems, methods, and non-transitory computer readable media that provide a contextual content generation system that trains and implements a unique machine learning architecture to generate context-specific digital content items based on a digital guideline document. In particular, the disclosed systems select a content generation method from among prompt engineering and / or updating one or more machine learning models to generate digital content. For example, the disclosed systems utilize machine learning models to extract key elements from a digital guideline document comprising context-specific guidelines for digital content. Further, the disclosed systems generate an augmented prompt comprising indications of key elements from the digital guideline document. In addition, the disclosed systems select a content generation method from among prompt engineering and / or updating machine learning models to generate the digital content item which incorporates digital content corresponding to the context-specific guidelines based on the augmented prompt.

The invention relates to the technical field of writing intelligent tutoring, and discloses a deep learning personalized tutoring method based on a real-time writing track. The method comprises the following steps: capturing writing motion data of a user in real time through a track sensing device, wherein the writing motion data comprises a pen point coordinate sequence, a timestamp sequence and a pressure intensity sequence; performing initial processing such as noise suppression and data alignment on the captured data; constructing a deep learning architecture based on the processed data to learn the writing feature representation; analyzing a real-time writing track sequence by applying the framework, and identifying writing deviation and a user habit mode; generating a customized tutoring instruction set according to the analysis result; and the deep learning architecture and the tutoring instruction set are dynamically updated by using the user interaction data and the continuously collected writing traces. According to the method, writing details can be comprehensively captured, writing problems and habits can be accurately identified through deep learning, personalized tutoring is provided, tutoring strategies can be dynamically adjusted, and the pertinence and effectiveness of writing tutoring are improved.

The invention relates to the field of noninvasive prenatal diagnosis, and particularly discloses a preeclampsia noninvasive screening method based on deep sequencing 8bp oligonucleotide double-fragment characteristics, which comprises the following steps: collecting preeclampsia and healthy pregnant woman peripheral blood samples, and extracting free DNA for high-throughput sequencing; the method comprises the following steps: extracting core 8-mer sequences' GTGCGCCC 'and' GATGGGGT 'in a long fragment of 150-200bp through bioinformatics analysis; an integrated support vector machine, K-nearest neighbor, extreme gradient lifting, a random forest and a multi-layer perceptron are combined with a logistic regression element classifier to construct a stacking model, the frequency of a core sequence is normalized, machine learning analysis is carried out, and the preeclampsia risk is predicted. According to the invention, two 8bp oligonucleotide characteristic fragments are specifically screened, and a deep learning architecture of multi-model fusion is combined, so that the limitations of low specificity and invasive detection of a traditional screening method are effectively broken through.

A method includes receiving a video comprising one or more frames; executing a first machine learning model using the one or more frames of the video to generate a dynamic mask configured to track a predicted magnitude of attention that individuals will give to different portions of each of the one or more frames of the video during playback of the video, the dynamic mask comprising attention scores for individual portions of each of the one or more frames of the video; generating one or more attention metrics for the video based on an aggregation of attention scores for the individual portions of each of the one or more frames of the video; and generating a record identifying the one or more attention metrics for the video.

Embodiments described herein provide for a voice biometrics system execute machine-learning architectures capable of passive, active, continuous, or static operations, or a combination thereof. Systems passively and / or continuously, in some cases in addition to actively and / or statically, enrolling speakers as the speakers speak into or around an edge device (e.g., car, television, radio, phone). The system identifies users on the fly without requiring a new speaker to mirror prompted utterances for reconfiguring operations. The system manages speaker profiles as speakers provide utterances to the system. Machine-learning architectures implement a passive and continuous voice biometrics system, possibly without knowledge of speaker identities. The system creates identities in an unsupervised manner, sometimes passively enrolling and recognizing known or unknown speakers. The system offers personalization and security across a wide range of applications, including media content for over-the-top services and IoT devices (e.g., personal assistants, vehicles), and call centers.

The invention relates to the technical field of computers, and discloses a fragmentation block chain federal learning method based on a large language model multi-agent, and the method comprises the steps: S1, initializing a client; s2, dynamic fragmentation scheduling and distribution; s3, generating and uploading local knowledge; s4, intelligent agent collaborative routing and knowledge acquisition; s5, knowledge fusion and model updating; and S6, repeatedly executing the steps S3 to S5 until the model converges or reaches a preset number of iterations. According to the invention, under a decentralized and fragmented federated learning architecture, the complex reasoning ability of a large language model and the autonomous cooperation mechanism of three multi-agent systems, namely a fragmented scheduling agent, a fragmented knowledge state agent and a global knowledge routing agent, are deeply fused; according to the mechanism, dynamic optimization of a bottom layer fragment structure and intelligent routing of high-value knowledge are achieved in an intelligent mode, and therefore the overall efficiency and model performance of a system under the condition of heterogeneous data and heterogeneous equipment are remarkably improved.