533 results about "Knowledge transfer" patented technology

The invention relates to the technical field of industrial Internet of Things, in particular to a machine equipment online state monitoring and fault diagnosis system, which comprises the following steps of: acquiring multi-source heterogeneous sensing data through an edge computing node deployed on an equipment body, performing adaptive noise filtering and feature dimension reduction processing on original data, and acquiring multi-source heterogeneous sensing data; outputting a standardized equipment state vector set; inputting the equipment state vector set into a dynamic knowledge graph engine, constructing a fault evolution network comprising space-time correlation characteristics based on an equipment operation entropy change quantification model, and generating a graph node connection relationship with a weight coefficient; and inputting the fault evolution network into a migration reinforcement learning module, and outputting a diagnosis decision set comprising a fault type, a severity degree and an evolution path through knowledge migration of a cross-device fault mode. According to the method, the problems of edge redundancy and single feature expression in traditional rule-based atlas construction are effectively avoided, and the structuring ability and physical traceability of fault recognition are improved.

The invention provides an automatic heuristic algorithm planning method based on a large language model, and the method comprises the following steps: carrying out the initialization and problem modeling, starting from a basic heuristic mode through guiding the large language model, generating a candidate algorithm set in combination with a plurality of cognitive perspectives, and providing diversified starting points for a search space; configuring core parameters of Monte Carlo tree search; in each iteration process, planning is started in a heuristic space by utilizing Monte Carlo tree search, and the process is composed of five core stages of selection, reflection, expansion, simulation and back propagation; and after all iterations are completed, the path with the highest average reward and the corresponding optimal heuristic algorithm are returned, and the global optimality of the final solution is ensured. According to the method, the effective experience can be automatically extracted from the heuristic strategy generated historically, and real-time feedback adjustment and strategy induction optimization of the heuristic structure are realized, so that the knowledge migration and generalization ability in the search process is remarkably enhanced.

The invention relates to a large-scale road network traffic control method based on a deep reinforcement learning large model, and belongs to the technical field of intelligent traffic control. The method comprises the following steps: sensing real-time multi-modal road network information including urban road intersections, highway entrance ramps and emergency lanes, and generating a space-time fusion representation vector representing a current traffic network state by fusing a space diagram construction method and a time sequence embedding method; the space-time fusion representation vector and historical state memory are spliced to serve as input, a backbone network of a pre-training large language model is used for state feature distillation so as to enhance state representation, and a traffic control decision is output through a strategy network with a layered action space; through cross-modal knowledge migration and a progressive course learning strategy, a training process of a deep reinforcement learning algorithm is guided and optimized so as to improve model training efficiency and generalization ability. According to the method, the generalization performance and the accuracy of the control strategy are improved while the real-time response speed is ensured.

The invention discloses a multi-modal traffic large model real-time regulation and control method and system for vehicle-road cooperation, and relates to the technical field of artificial intelligence, and the method comprises the steps: combining multi-modal data fusion, edge intelligent reasoning, federated learning, cross-regional knowledge migration, reinforcement learning optimization and adaptive closed loop iteration; and efficient and accurate vehicle-road cooperative regulation and control are realized. The model is adopted to perform space-time alignment and high-dimensional feature extraction on vehicle-mounted, roadside and cloud data, so that the environmental perception precision is improved; cross-regional traffic knowledge sharing is realized through gradient aggregation and decentralized training, and data privacy leakage is avoided; a transfer learning and self-supervision mechanism is adopted, so that the model can quickly adapt to different cities and different road environments, and the generalization ability is improved; by adopting cloud multi-agent reinforcement learning, the optimal decision of signal lamp timing and path recommendation is realized, the traffic flow change is dynamically adapted, and the problem that efficient and real-time model adjustment cannot be realized in consideration of privacy and global optimization in the prior art is solved.

The invention relates to the field of data detection, in particular to a network traffic anomaly detection method based on aggregated mimicry distillation. According to the method, protocol level analysis is carried out on traffic through a multi-branch feature extraction network, and a unified representation vector is generated through a cross-layer fusion mechanism; calculating a first abnormal score based on a protocol perception weighted confrontation soft contrast mechanism; a heterogeneous teacher model is constructed and dynamically weighted and aggregated, and the student model generates a second abnormal score through knowledge distillation learning; forming a heterogeneous redundant detection pool by the student model, the teacher model and the rule detector, dynamically selecting the detector and applying adaptive disturbance to obtain a third abnormal score; and dynamically fusing the three types of abnormal scores to output a detection result. According to the method, the problems of protocol semantic segmentation, weak boundary sample discrimination, knowledge migration simplification and defense path predictability are effectively solved, and the detection accuracy, robustness and dynamic defense capability are remarkably improved.

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.

A system and method for cross-domain knowledge transfer in federated compression networks. The system enables efficient lossless data compression across diverse data types by intelligently sharing compression strategies between domains. A cross-domain knowledge transfer system identifies relationships between different data domains, adapts compression parameters accordingly, and optimizes learning processes to maximize knowledge reuse. The architecture may include a knowledge repository for storing domain features and compression patterns, domain mapping components that identify similarities, and transfer learning optimization that enables efficient adaptation with minimal examples. This approach significantly accelerates model training for new domains while improving compression performance. Applications include satellite telemetry systems where efficient compression is critical for transmitting large information sets between distant locations. The system may employ probability prediction driven arithmetic coding paired with long short-term memory networks, enhanced by cross-domain knowledge sharing that adapts successful compression strategies from one domain to another while preserving domain-specific optimization.

The invention discloses a distribution line unmanned aerial vehicle edge end real-time visual identification and target detection method and system based on lightweight AI. The method is based on a YOLOv8 architecture, and constructs a complete lightweight detection framework comprising a feature extractor, an enhancement module and a simplified detection head by introducing a space structure maintaining assembly, a separated large kernel convolution and a weighted reconstruction feature pyramid. A cross-dimension semantic relation model is constructed by utilizing a combined attention structure, multi-level feature fusion is realized by reconstructing a feature pyramid network, model training is performed by adopting a hybrid optimization function and a dynamic sample adjustment mechanism, and the model is deployed on a mobile computing platform after being optimized by a hierarchical knowledge transfer method. According to the method, the detection speed is remarkably increased while high precision is kept, and real-time identification and anomaly analysis of the power line element are effectively realized.

The invention provides a robot social adaptive navigation knowledge learning and migration method and system, and relates to the field of mobile robot navigation. Aiming at the problems that an existing path planner lacks time sequence memory and neglects pedestrian social intent, a man-machine co-fusion scene is constructed, a training set containing an expert demonstration path is made, and a recursive generation model is input; designing a recurrent neural network embedded RRT, generating an RNN-RRT planner, and fusing historical information and pedestrian convergence probability in training; new scene loading training parameters are finely adjusted to realize knowledge migration, loss convergence or output RNN final parameters after reaching a preset round number. According to the method, the path anthropomorphism and generalization ability are improved, and the method is suitable for complex human-computer interaction scenes.

The invention discloses an energy storage battery health feature extraction and state evaluation method based on transfer learning. The method comprises the following steps: S1, constructing a source domain health feature library and pre-training a model; according to the method, dependence on complete cyclic data is broken through, high precision and robustness are still achieved under the conditions of data sparsity and working condition difference, and the method is suitable for intelligent operation and maintenance and predictive maintenance of an energy storage power station. And meanwhile, common incomplete and partial charge and discharge data fragments under actual working conditions can be directly utilized for feature extraction and state evaluation, dependence on complete charge and discharge cycles is avoided, and the application scene of the data driving method is greatly widened.

The invention discloses a small sample self-learning accurate identification method based on distillation knowledge migration. The method comprises the following steps: S1, extracting deep semantic features of a source domain and shallow features of a small number of samples of a target domain, and calculating a mapping matrix; s2, calculating an entropy difference distillation excitation function based on the initial alignment features; s3, executing domain knowledge distillation and generating staged distillation representation; s4, constructing a composite fitness function and initializing a parameter population; s5, performing iterative optimization by adopting a variable step size dynamic feedback compression strategy; s6, loading the optimal parameters and performing coupling alignment with the historical distillation representation; and S7, performing combined fine adjustment on the distillation weight and the model parameters through self-learning feedback. According to the method, through adaptive knowledge distillation and dynamic optimization feedback closed loop, high-precision and adaptive identification under extremely few labeled samples is realized, the generalization ability of the model is remarkably improved, and overfitting is effectively inhibited.

The invention discloses a remote sensing rotating target detection method and system based on comparative heterogeneous knowledge distillation. The method comprises the following steps: constructing a comparative heterogeneous knowledge distillation network detection framework; the Token sequence of the teacher network is converted into a two-dimensional feature map, and the two-dimensional feature map is matched with the inductive bias of the CNN of the student network to obtain transformed teacher features; converting the two-dimensional feature map of the student network into a Token sequence, and inputting the Token sequence into a teacher network for global interaction to obtain converted student features; constructing an assistant transformation model by sharing teacher network weight fusion transformation teacher features and student features; dividing positive and negative sample sets by cooperatively enhancing feature expression through space attention and channel attention and combining cosine similarity matching and a dynamic threshold division strategy; and combining a teacher network, a student network and an assistant transformation model, constructing comparative learning knowledge distillation in combination with positive and negative sample sets, realizing knowledge migration through feature distillation loss and logic distillation loss, and outputting a detection result image.

The invention provides a risk control cross-domain risk prediction method and system in combination with transfer learning, and the method comprises the steps: carrying out the domain difference measurement of a source domain and target domain risk control scene, generating a domain difference measurement matrix, building a hierarchical knowledge transfer channel based on the domain difference measurement matrix, and carrying out the prediction of the risk control cross-domain risk. And selectively migrating the source domain risk prediction model parameters, generating a target domain initialization risk prediction model, and carrying out dynamic adaptive training on the target domain initialization risk prediction model by adopting a target domain risk sample to obtain a cross-domain risk prediction model. And inputting the to-be-evaluated sample of the target domain into the cross-domain risk prediction model to generate a preliminary risk prediction score, and finally performing cross-domain calibration on the preliminary score based on the domain difference metric matrix to generate a final risk prediction result, thereby improving the accuracy of risk control cross-domain risk prediction.

The invention relates to the field of robot operation and three-dimensional perception, in particular to a monocular image-based three-dimensional perception robot operation knowledge distillation method, which comprises the following steps of: establishing a strategy learning framework comprising a student model and a teacher model; constructing a strategy prediction model in the strategy learning framework; training the strategy prediction model, wherein a three-level knowledge distillation mechanism is adopted in the training process to complete knowledge migration between the teacher model and the student model; combining the three distillation losses with strategy optimization losses to form a total loss function, and performing end-to-end training on the student model until convergence to obtain a deployable monocular strategy model; the deployable monocular strategy model only retains a student model, and generates a robot operation instruction. The method has the beneficial effects that the robot under monocular RGB input has three-dimensional perception and high-precision operation capabilities while the reasoning efficiency is kept, the task success rate and generalization performance are remarkably improved, and the effectiveness and robustness of the method are verified.

The invention relates to a cross-software operation agent training method based on structure perception and few-sample learning, which comprises the following steps: receiving an operation intention of a user on source software, preprocessing, and encoding the intention to obtain a high-dimensional intention semantic vector; the method comprises the following steps: extracting control attribute information of a software GUI on source software, constructing a control tree, and converting the control tree into a high-dimensional structure semantic vector; capturing a GUI (Graphical User Interface) screen on source software and carrying out image coding to obtain a visual feature vector; performing attention semantic consistency alignment and fusion on the high-dimensional intention semantic vector, the high-dimensional structure semantic vector and the visual feature vector by adopting a cross attention mechanism to obtain fused feature representation; and based on the fused feature representation, constructing a prototype library, obtaining a small number of user operation samples on the target software, and updating the prototype library by using the small number of user operation samples to realize knowledge migration and complete a training process. Compared with the prior art, the method has the advantages of high generalization ability, high efficiency and the like.

The invention discloses a wind power prediction method based on multi-source domain deep transfer learning, and relates to the field of new energy power prediction.The method comprises the steps that the data distribution difference between a multi-source domain and a target domain is reduced through an Euclidean alignment method, and the maximum mean value difference between the domains after optimization is obtained; setting a migration weight factor for each source domain based on the maximum mean value difference, and constructing a weighted migration training set; constructing a multi-source domain deep migration learning model, and carrying out migration training on the model based on a weighted migration training set; and finely tuning the pre-training model according to a small amount of data of the target domain to obtain a target domain wind power prediction model. The wind power data distribution difference between the multi-source domain and the target domain is reduced through the Euclidean alignment method, the knowledge migration efficiency is improved, and the negative migration risk is reduced; setting a migration weight factor for each source domain based on inter-domain MMD, so that the model preferentially learns high-correlation source domain wind power characteristics; and the multi-source domain deep migration learning model is fused into a dynamic weight module, so that low-efficiency migration is avoided.

The invention provides a data processing method and device based on a hybrid expert model and related equipment, and relates to the technical field of natural language processing, and the method comprises the steps: inputting to-be-processed data inputted by a user into the hybrid expert model for correlation matching, and determining an expert model set related to the to-be-processed data; performing knowledge distillation processing and knowledge migration processing on an expert model set in the mixed expert model and unselected expert models to obtain a target expert model set; and inputting the to-be-processed data into the target expert model set, and outputting a corresponding data processing result. According to the method, the performance of the hybrid expert model in multiple tasks can be effectively improved, and meanwhile, the utilization efficiency of computing resources is optimized.

The invention discloses an equipment agent processing program conversion method and device based on cooperation of a large model and a small model. An equipment agent is constructed by analyzing control systems, kinematics structures and process constraint information of a source machine tool and a target machine tool, and line-level instructions are abstracted into system-independent unified semantic representation based on a multi-system numerical control corpus. And performing grammar analysis and semantic mapping on the source program, and generating a conversion task in combination with grammar rules, kinematics accessibility and process security constraints of a target machine tool. A small model is finely adjusted through supervised learning to realize basic generation capability, a large model is introduced as a patch evaluator, candidate results are jointly scored from different dimensions, knowledge migration is realized based on improved near-end strategy optimization, and a small model generation strategy is continuously optimized. And finally, high-quality results are screened through confidence degree sorting, cross-system G / M instruction and multi-axis track generation and verification are completed, and high-precision automatic conversion of numerical control programs is achieved.

The invention discloses a hyperspectral image classification method for cross-domain small sample learning based on diffusion enhancement prototype knowledge distillation, and belongs to the technical field of hyperspectral image processing. The method comprises the following steps: extracting a neighborhood data cube, aligning spectrums, dividing a support set and a query set, and applying a mask and enhancing noise; executing domain adversarial denoising and reconstruction tasks, aligning feature distribution, and outputting a pre-training encoder; decoupling features, capturing spectrum-space global and local dependency relationships, and calculating similarity between a query set and a category prototype; constructing a distillation framework to realize knowledge migration; optimizing model parameters, and introducing a signal-to-noise ratio to enhance loss suppression noise; and performing feature extraction by using the optimized student model to generate a hyperspectral image classification result. According to the method, the problems of domain offset, intra-class feature dispersion, inter-class boundary fuzziness, noise interference and the like are solved, and the classification accuracy in a small sample scene is improved.

The invention discloses a 3D generative model training method and device based on a hybrid training framework, equipment and a storage medium, and belongs to the technical field of artificial intelligence. The method comprises the following steps: pre-training a 3D generation model by utilizing a preset 3D data set; quickly generating initial 3D assets according to the text cue words through a 3D generation model; a strategy based on two-dimensional knowledge distillation is adopted, a text-to-2D image generation model is used as a teacher model to carry out iterative distillation optimization on the initial 3D assets, and the 3D assets with improved visual fidelity are obtained; in the distillation process, dynamically adjusting the teacher model to adapt to the distribution of the initial 3D assets by using a self-adaptive teacher model guide strategy; and taking the optimized high-fidelity 3D assets as enhanced training samples, and training the 3D generation model again to improve the generation quality. According to the method, the 3D generation model is pre-trained, so that the 3D generation model is initially converged; according to the distillation method, strong text-to-2D image generation model knowledge is transferred to initial 3D assets and serves as an enhanced sample to train a 3D generation model again for knowledge internalization. The adaptive teacher model guiding strategy reduces the distribution difference between the teacher model and the student model, and ensures the generation quality of the 3D generation model.

The invention discloses a machine tool fault prediction system based on digital twinning, and the system comprises a data collection module which is used for collecting multi-source heterogeneous operation data of a machine tool; the edge calculation module is used for data preprocessing and feature extraction; the dynamic twinborn model module is used for establishing and updating a digital twinborn model and realizing incremental learning and multi-time scale fusion; the working condition self-adaptive module is used for identifying working conditions and carrying out cross-working-condition knowledge migration; the fault prediction module is used for predicting faults and evaluating cascade influences; the maintenance optimization module is used for generating an optimized maintenance strategy; the visual monitoring module is used for displaying a prediction result and pushing early warning; the data storage module is used for managing historical data and a knowledge base; and the model verification module is used for evaluating and calibrating the model precision. Through the dynamically updated digital twinborn model and the intelligent fault prediction algorithm, accurate prediction of the machine tool fault and optimization of the maintenance decision are realized, and the equipment reliability and the production efficiency are effectively improved.

The invention discloses a deep reinforcement learning optimization method for injection molding process parameters, and belongs to the technical field of intelligent manufacturing. The method comprises the following steps: constructing a dynamic causal graph network through information entropy flow analysis and transfer entropy calculation, and revealing a causal relationship and time delay characteristics among process parameters; manifold learning is adopted to map a high-dimensional parameter space to a low-dimensional manifold, and Riemannian metric guide optimization search is constructed based on the quality gradient; generating enhanced state representation fusing causal association and manifold geometric information; identifying a production element state and selecting a corresponding optimization strategy; a geodesic line is planned in a manifold space to obtain an optimal parameter adjustment path; historical experience is utilized through memory retrieval and case adaptation; cross-task knowledge migration is realized; adopting a depth deterministic strategy gradient algorithm to optimize the decision; and online learning is realized through elastic weight consolidation. According to the method, the problems of black box decision, slow convergence, difficulty in knowledge reuse and the like in the prior art are solved, the optimization efficiency and the interpretability are improved, and the method has the capability of quickly adapting to new tasks.

The invention relates to the field of artificial intelligence and medical image processing, in particular to a mammary gland pathology visual model establishing method based on multi-model fusion and combined distillation, which comprises the following steps: performing tissue segmentation and dyeing normalization on a full-slice image; inputting the image blocks into a pre-training teacher model of a plurality of freezing parameters in parallel to extract high-dimensional features, and generating unified enhanced features through a learnable feature fusion network; constructing a student model, and performing end-to-end training by using a joint loss function including feature simulation, logic output distillation and multi-task supervision; and connecting a plurality of task specific prediction heads to the student model, and realizing full-slice-level multi-task diagnosis and treatment prediction through an aggregation strategy. According to the technical scheme, efficient knowledge migration and multi-task cooperation can be achieved, and the accuracy, generalization ability and reasoning efficiency of mammary gland pathology analysis are remarkably improved.

The invention provides an online rail fault diagnosis method and system based on knowledge transfer learning, and belongs to the technical field of crossing of intelligent monitoring and artificial intelligence of railway infrastructures. The method comprises the steps that S1, a server trains a model framework through a source domain data set to obtain a teacher diagnosis model; based on the target domain data set, training the teacher diagnosis model by adopting a mixed training strategy to obtain a student diagnosis model; s2, acquiring real-time multi-modal monitoring data from a target domain line by the edge computing equipment, compensating to obtain corrected data, and dynamically selecting the most important feature subset from the corrected data to form a simplified feature set; and S3, inputting the simplified feature set into a student diagnosis model to obtain a student fault diagnosis result and confidence thereof, and introducing a D-S evidence theory to generate a student fault diagnosis report. The method has the advantages that the accuracy, the real-time performance, the cross-domain adaptability and the overall system safety of railway track fault diagnosis are greatly improved.

The invention belongs to the technical field of artificial intelligence and multi-modal emotion calculation, and discloses a multi-modal emotion recognition method based on a Mama state space model and cross-modal self-distillation. Through the organic combination of the efficient sequence modeling capability of the Mamba state space model and the knowledge sharing mechanism of cross-modal self-distillation, the advantages of the state space model in the aspects of time sequence modeling and calculation efficiency are fully played, and meanwhile, the limitation of a single model architecture is made up through a cross-modal attention mechanism; the technical bottlenecks of an existing multi-modal emotion recognition method in the aspects of long sequence processing efficiency, cross-modal information fusion and knowledge transfer sufficiency are effectively solved, and an efficient and reliable technical solution is provided for further development and practical application of the multi-modal emotion recognition technology.

The invention discloses a structure cross-domain damage identification method based on a pulse graph neural network. The method is suitable for structure health monitoring in the fields of civil engineering, mechanical equipment and the like. According to the method, structure dynamic response signals are collected through a sensor array, a graph structure is constructed, the graph structure is converted into space-time pulse graph data through pulse coding, and the space-time pulse graph data are input into a pre-trained pulse graph neural network for feature extraction and recognition. According to the method, the low power consumption and event-driven characteristics of the spiking neural network are creatively combined with the spatial topology modeling capability of the graph neural network, and a domain adversarial transfer learning mechanism is introduced, so that the problem of generalization of the model in cross-domain scenes of different structures, different environments and the like is effectively solved. According to the method, accurate positioning and quantitative evaluation of damage can be realized, and the method is particularly suitable for knowledge migration and long-term online monitoring from a laboratory model to a real structure.

The embodiment of the invention discloses an unannotated code vulnerability detection method and device based on distillation learning, electronic equipment and a computer readable medium. A specific embodiment of the method comprises the following steps: constructing a teacher model and an initial student model; inputting the target code related information into an embedded layer, a multi-layer encoder, a selector and an interaction module included in a teacher model to obtain teacher feature representation information; determining joint loss information by using the teacher feature representation information, the initial student model and the real label; training the initial student model to obtain a student model after distillation; inputting the target code information into the student model after distillation to obtain a detection result; and generating early warning information, and sending the early warning information to the target terminal for display. According to the implementation mode, the vulnerability detection capability of the model on the unannotated code is improved, the dependence on the annotation information is reduced, and the knowledge migration and generalization capability of the model is enhanced.

The invention relates to the technical field of geotechnical material local modeling, in particular to a geotechnical material data driving local modeling method and device based on a knowledge migration strategy, and the method comprises the steps: building a first mapping data pair based on the stress-strain physical quantity of a test curve sampling point, and building a high-fidelity data set of a geotechnical material; constructing a second mapping data pair based on the stress-strain state of the simulation curve sampling point and the intermediate physical quantity label, and establishing a low-fidelity data set of the geotechnical material; pre-training the target data-driven constitutive model by using the low-fidelity data set to generate an initial geotechnical material data-driven constitutive model; and driving the model to be fine-tuned by using the high-fidelity data set for the initial geotechnical material data, and generating final geotechnical material data to drive the model. Therefore, the problem of insufficient generalization ability and prediction precision of a data-driven model due to limited training data and unpredictability of a stress path in the prior art is solved.

The invention discloses a hybrid compression method combining knowledge distillation and meta-gradient initialization pruning, and relates to the field of neural network model compression. According to the mixed compression method combining knowledge distillation and meta gradient initialization pruning, firstly, in an HSD stage, a parameter mirror image initialization strategy is designed, pre-trained teacher network parameters are frozen, a topological symmetric student network is adopted, parameter smoothness and model performance are enhanced through self-distillation optimization, and a teacher model with better performance is obtained; and then, carrying out initialization pruning on the untrained network by using a ProsPr (Prospect Pruning, ProsPr) method to obtain a sparse subnet, and taking the sparse subnet as a student model of the next stage. In a knowledge distillation stage after pruning is completed, a sparse-sensing temperature coupling (SATC) mechanism is provided, adaptive matching of knowledge migration intensity and pruning rate is realized by dynamically adjusting distillation temperatures of a correct class and an error class of a teacher model, and the problem that the capacity difference between the teacher model and a student model is too large due to too high pruning rate is solved, so that the knowledge migration intensity and the pruning rate are greatly improved. And the knowledge distillation effect is reduced.

The invention provides a heterogeneous model knowledge migration method based on federated learning. The method comprises the following steps: a data collaborative preparation stage; screening local data, performing multi-dimensional labeling, and performing cooperative training and federal modeling preparation on a plurality of participants; a privacy initialization stage; privacy protection processing is carried out on local data, and local model parameters are initialized; a self-learning stage; each participant completes localized updating of the model parameters by using the local data set; a mutual learning stage; the server extracts valuable knowledge fragments from the knowledge representations uploaded by the participants through the super network for mixing, and updates model parameters of the participants; a server-side super network module is updated and iteratively cycled; and the server side aggregates the valuable knowledge fragments, updates the super network parameters and then returns to the step S3, and loop iteration is carried out until the model converges. The problems that computing resources of participants are limited, generalization is insufficient due to model heterogeneity, and a knowledge migration technology cannot adapt to a heterogeneous model architecture are solved.