67 results about "Historical model" patented technology

The invention relates to a heterogeneous federated learning method and system based on historical model double knowledge distillation, and belongs to the technical field of federated learning and knowledge distillation. The method comprises the steps that a server initializes a global model and constructs a historical global model storage list, and then the server generates a historical fusion model through weighted aggregation of multiple rounds of historical global models and sends related models to a client; after the client receives the global model and the historical fusion model of the round, local model training is carried out through a double knowledge distillation strategy based on a local private data set, the updated local model of the round is stored, global knowledge distillation takes the historical fusion model as a teacher model, and local knowledge distillation takes the historical local model of the client as a teacher model; global and local knowledge migration is balanced through hyper-parameters; the server averagely aggregates the model parameters updated by the client to generate and store a new round of global model; and iteratively optimizing to a convergent or preset training round number. According to the method, the performance of the model in a heterogeneous data environment is remarkably improved.

The invention provides an undercarriage system modeling method based on a natural language demand text, and relates to the technical field of model-based computer aided design, and the method comprises the steps: S1, constructing a SysML meta-model knowledge base and an undercarriage system model knowledge base; s2, analyzing the natural language demand text, and generating a structured demand text of the undercarriage system; s3, mapping a structured demand text entity of the undercarriage system into a node type of a semantic graph; and S4, in combination with a historical model template and real-time retrieval, generating an undercarriage system SysML model according to the semantic map obtained in the step S3, and performing undercarriage system engineering analysis through the undercarriage system SysML model, thereby improving the efficiency and quality of aviation equipment development. On the basis of a retrieval enhancement generation technology and an intention routing mechanism, the modeling efficiency and accuracy of the undercarriage system are remarkably improved; a graph neural network and a semantic vector space model are adopted to realize deep fusion and conflict resolution of cross-domain knowledge; and through an incremental learning mechanism, model continuous evolution and design experience closed-loop iteration are supported.

The invention discloses a computing node selection method and device and a storage medium, relates to the technical field of communication, is used for improving the computing node selection efficiency, and comprises the following steps: obtaining historical information of each computing node in a plurality of computing nodes; the historical information comprises a historical contribution degree, a historical model verification passing rate and a penalty factor; the computing node selection device determines the reputation value of each computing node in the plurality of computing nodes according to the historical information of each computing node in the plurality of computing nodes; the computing node selection device selects a target computing node set meeting the reputation value requirement from the plurality of computing nodes; the target computing nodes in the target computing node set are used for participating in model training. The method and the device are applied to the process of computing node selection.

The invention provides an underground traffic construction pipeline relocation optimization method and system. According to the method, the spatial position, the cross section size, the joint position and the maintenance record of the existing cable are acquired, and a channel structure is combined to construct a spatial layout model; scanning the side wall and the bottom plate of the channel by using a non-contact detector, and generating a channel distribution diagram based on the identified obstacle position, size parameter and structure boundary dynamic updating model; dividing the distribution map into grids, calculating a safety distance by taking a section size as a constraint, traversing a grid connection relation, and determining an initial installation path by avoiding a high-risk area marked at a joint position; then extracting a fault position and a maintenance frequency, binding an initial path with a spatial position by means of a historical model, calculating a path risk, and outputting a maintenance parameter; and adjusting the path trend, generating a target scheme and outputting a cable arrangement diagram. According to the invention, accurate optimization of the cable path is realized, the construction collision risk is avoided, and the long-term operation and maintenance cost is reduced.

The invention discloses a denture three-dimensional model feature construction and matching method based on curvature adaptive sampling, which comprises the following steps: S1, acquiring denture surface point cloud data through three-dimensional scanning, converting the data into an STL model, and comparing a historical STL model pre-stored in a folder; s2, constructing a self-adaptive sampling weight, selecting a triangular patch through replacement random sampling, and generating a sampling point; s3, carrying out normalization processing on the sampling points; s4, calculating the distribution of D2 and A3 by using the normalized sampling points, and obtaining corresponding shape features; s5, splicing the features to form a joint geometric feature vector; and S6, performing S1-S5 on the to-be-matched false tooth to obtain a matching vector, performing S2-S5 on the historical models one by one to obtain comparison vectors, calculating cosine similarity between the matching vector and each comparison vector, and outputting a Top-K similarity model. According to the method, key geometric structure feature expression is enhanced, robustness to point cloud density change, local defects and the like is high, and matching is ensured to be reliable and consistent.

A system, method, and a computer program product for determining anomalies in a dataset are provided. The dataset, which stores data structures of parameters, is received at an anomaly detection framework. The anomaly detection framework selects multiple models, including time series models, historical models, artificial intelligence models and isolation forest models to analyze the parameters in the dataset and determine anomalies in the data structures.

The application provides a data lake-based data processing method, device, equipment and medium. The method comprises the following steps: determining a preset data lake table; obtaining time series data and model data of a target device based on a flink framework and the data lake table; comparing the time series data with historical time series data stored in a target state backend of the flink framework to obtain a first comparison result; comparing the model data with historical model data stored in the target state backend to obtain a second comparison result; updating the historical time series data and the historical model data in the target state backend based on the first comparison result and the second comparison result to obtain an update result; and storing the time series data and the model data in a target database according to the update result. Through the method, unified real-time access to a data lake can be realized based on a data lake storage technology, stream batch integration at a storage level is realized, and stream batch integration at a computing level is realized based on a flink computing engine.

The invention provides a CO emission prediction method based on robust kernel feature space semi-supervised drift detection. The method comprises the following steps: calculating a first-order difference component of a historical sample through a pre-constructed historical model to construct a historical data set; calculating feature space statistics of the real-time data, and performing drift detection through the feature space statistics to obtain a drift sample; union set taking operation is carried out on the historical samples and the drift samples, and a new training set is obtained; and performing data updating on the historical model through the new training set, and predicting the CO emission through the updated prediction model. According to the method, through semi-supervised drift detection, the problems of concept drift identification and model dynamic updating under truth value deficiency during CEMS fault are solved, and the CO emission prediction precision and the robustness in a complex industrial scene are improved.

The application discloses a method for defending against poisoning attacks in blockchain federated learning supporting privacy protection, and relates to five links of initialization, model training and blinding, malicious detection, model aggregation and model updating. In the process of malicious detection, the method accumulates abnormal fluctuations in the form of historical model updates to highlight the characteristics of malicious behavior and realize accurate identification of malicious behavior, thereby effectively defending against poisoning attacks and improving the robustness of the global model. In addition, the poisoning attack defense effect of the method will not be affected by the protection of model updates, which can realize malicious detection using original features and avoid direct exposure of participant data privacy, to a certain extent, alleviate the contradiction that the original features of model updates are hidden after the implementation of privacy protection, which is not conducive to the accurate identification of malicious detection, thereby further enhancing the overall security of blockchain federated learning.

The application discloses a model training method and device, a computer readable storage medium and an electronic device. The method comprises the following steps: obtaining historical model training information of a plurality of nodes in a preset time period before the current time of each node in a joint learning task; when there is an interrupt node in the plurality of nodes, determining a current model training scheme of the plurality of nodes according to the historical model training information of each node. Through the technical scheme of the application, the current model training scheme can be determined in real time when an interrupt node appears in the process of executing the joint learning task, and the normal progress of the joint learning task is ensured.

Media plan generation systems and methods for TV advertising are disclosed. Embodiments of these systems and methods are adapted to generate a media plan comprising a core plan and a test plan. The core plan may be produced based on historical models regarding historical performance of an entity's media while the test plan may be produced based on a predictive model of media performance.

The invention relates to a method for accelerating model training recovery, electronic equipment and a storage medium. The method comprises the steps of executing a first task for constructing a training component based on an instruction of recovery model training, and independently executing a second task through a concurrent execution unit corresponding to a main execution unit; wherein the second task is used for loading the check point file corresponding to the training component to the memory from the external storage system; wherein the training component comprises a model object and an optimizer object, the check point file comprises state data, stored in historical model training, of the training component, and the first task and the second task are at least partially overlapped in execution time; and in response to an event of completing the first task and the second task, loading the state data in the memory into the corresponding training component to complete state recovery of the training component. According to the method, the parallel support of each task in the model training recovery process can be realized, and the time consumption of model training recovery is effectively reduced.

The invention relates to the technical field of operation efficiency evaluation, in particular to a civil aviation autonomous operation efficiency evaluation method, which comprises the following steps: acquiring operation information historical data which comprises operation information accumulated in the civil aviation operation process; multiple historical models are established based on the operation information historical data, each historical model is used for representing different autonomous operation scenes of the civil aviation, and each autonomous operation scene comprises a corresponding reference efficiency index; obtaining to-be-evaluated operation information, matching the to-be-evaluated operation information with a plurality of historical models, and obtaining a current autonomous operation scene; calculating an actual efficiency index corresponding to the current autonomous operation scene according to the to-be-evaluated operation information; performing visual comparison display on the actual efficiency index corresponding to the current autonomous operation scene and the reference efficiency index; the civil aviation operation efficiency can be comprehensively evaluated.

A system and method for sampling a flight envelope of an air vehicle under test. The method includes obtaining engineering models of the vehicle to identify how the vehicle behaves, obtaining empirical models through test data, and obtaining historical models of similar legacy aircraft. The method further includes ensembling the models into an integrated model and assessing the ensemble uncertainty at any particular test point, for a variety of aircraft conditions, to determine model reliability. The method includes determining the optimal order of test points based on local model uncertainty, testing the air vehicle at a point, revising the integrated model with that data and reassessing the uncertainty, and determining if a stop testing condition has been met, in which case the testing series is closed. The last calibrated integrated model could then be used to generate trusted data for any remaining required test using a certification by analysis method.

The invention belongs to the field of power system risk early warning, and particularly relates to a power grid node connectivity modeling and operation risk accurate early warning method and device based on meteorological disturbance driving, computer equipment and a storage medium. According to the method, real meteorological data of a plurality of preset time points are acquired and predicted meteorological data of a plurality of reference time points are obtained through prediction, so that continuous tracking of a meteorological change trend is realized; meanwhile, reference meteorological data are determined by using historical meteorological data, association between predicted meteorology and historical modes is established, risk level assessment is more fit with an actual meteorological disturbance rule, and the risk level assessment accuracy is improved by identifying a change time point and a corresponding level change amount in a basic risk level sequence, calculating a level adjustment coefficient and adjusting the sequence. The risk level is adjusted by effectively utilizing the time domain information, so that the influence of time domain accumulation can be introduced during risk early warning, the accuracy of risk early warning is improved, and the reliability of risk early warning of the power system is further improved.

The embodiment of the invention provides a model configuration parameter generation method and device, equipment and a computer storage medium. The method comprises the steps of obtaining preset target model task data; and based on the target model task data, according to a corresponding relationship between preset historical model task data and a historical model predicted calculation power value, determining a target model predicted calculation power value corresponding to the preset target model task data. And under the condition that the target model predicted computing power value is equal to the historical model predicted computing power value, historical model configuration parameters corresponding to the historical model predicted computing power value are used as target model configuration parameters, and the target model configuration parameters are used for constructing a target model. According to the invention, under the condition that the predicted calculation power value of the target model is equal to the predicted calculation power value of the historical model, the historical model configuration parameter corresponding to the predicted calculation power value of the historical model is used as the target model configuration parameter, so that the time for determining the target model configuration parameter can be shortened; therefore, the time for constructing the target model based on the target model configuration parameters is shortened.

The invention discloses a federated learning-oriented data-free hidden model poisoning attack method. The method does not need to access local training data of any client, only depends on historical global model parameters issued by a server, constructs simulated benign parameters based on an improved Shampoo optimization algorithm by estimating a global gradient, and generates reference attack parameters on the basis; then fusing historical attack directions and superposing multi-dimensional composite disturbances such as projection disturbance, low-rank disturbance, adversarial disturbance and statistical disturbance; and finally, dynamically limiting the disturbance amplitude by adopting a self-adaptive norm cutting mechanism based on historical model statistics and direction information to ensure that malicious parameters are difficult to distinguish from benign updating. Experiments prove that the method can reduce the model accuracy from 70% to about 10% under various mainstream aggregation strategies, and has high concealment and universality. The method reveals that the federal learning system still faces serious model security threats under the condition of no local data.

The present application relates to an incremental detection method based on elastic weight interference isolation and prototype drift calibration, comprising: in the elastic weight interference isolation stage, detecting the interference area of the current data through the historical model and generating an interference area set (new class target mistaken as background), combining the historical and current interference information importance parameters to construct a perception score, constructing a loss function to train the model and updating the old knowledge importance parameter; in the prototype drift calibration stage, establishing the old stage class prototype based on the historical features, performing drift compensation through the projection layer and generating drift compensation features, splicing the drift compensation features with the model output features to retrain the classification head, so as to calibrate the feature distribution and complete the two-stage technical closed loop. The incremental detection method can effectively solve the problem of catastrophic forgetting.

The invention provides an intention recognition method and device based on virtual adversarial training, and the method comprises the steps: obtaining an original text, and determining initial disturbance information based on the similarity between vocabularies in the original text and historical disturbance vocabularies; performing iterative updating on the initial disturbance information, determining an average value of a plurality of gradients obtained in an iteration process, and determining final disturbance information; and constructing a confrontation sample based on the final disturbance information and the original text, updating model parameters of the pre-trained intention recognition model based on the confrontation sample, and performing intention recognition based on the trained intention recognition model. The initial disturbance information of the original text is determined by introducing the disturbance vocabularies in the historical model training process, and is continuously updated and used in the adversarial training process, so that the noise of random initialization is reduced, a large amount of invalid disturbance is avoided, the adversarial training efficiency is improved, and the period of the model from development to deployment is shortened; and the service demand can be responded more quickly.

The invention discloses a geological exploration data three-dimensional modeling analysis G3DVelocity system based on IDL and an application method thereof, a storage medium and a program product. The geological exploration data three-dimensional modeling analysis G3DVelocity system comprises a database management subsystem, a model establishment analysis subsystem, a graph editing subsystem and a data management subsystem. The application method comprises the following steps: S1, performing format conversion on seismic exploration original data provided by a user in the database management subsystem, and storing the seismic exploration original data in a database; s2, according to user needs, the data can be edited and processed through the database management subsystem; s3, selecting data through a model establishment and analysis subsystem to perform spatial interpolation processing, establishing a three-dimensional model by using the data subjected to interpolation, and editing and analyzing the three-dimensional model according to the needs of a user; and S4, finally, the result model is stored in a graph editing subsystem, so that the historical model can be conveniently looked up and edited.

The invention discloses an atmospheric numerical forecasting method and system based on machine learning, and a storage medium, and relates to the technical field of meteorological information, and the method comprises the steps: generating a preliminary prediction field of a plurality of atmospheric physical variables in a specified time period in the future based on a deep learning model integrated with a multi-scale space-time attention mechanism; correcting the preliminary prediction field according to physical constraints; and comparing the corrected atmospheric state field data with atmospheric observation data in a historical model error library, adjusting a physical constraint weight of a corresponding region for an extreme weather scene according to a comparison result, and using a correction result for model training. Through deep fusion data driving prediction and a physical constraint correction mechanism, the reliability and precision of a prediction result are remarkably improved, a deep learning model integrated with a multi-scale space-time attention mechanism is utilized, complex space-time features in atmosphere state evolution are effectively captured, and the simulation capability of a multi-scale atmosphere process is enhanced.

This invention relates to the field of data processing technology and discloses a personalized federated learning method, system, and medium for edge scenarios. The method calculates a personalized model set composed of multiple client models, calculates the model combination weights using historical model combination weights and mutual information bias terms, and performs model interpolation training using the personalized model set and model combination weights. This can improve the personalization capability of client models and meet the client's expected training objectives.

The present application relates to the technical field of building curtain walls, and discloses an industrialized implementation method for spatial special-shaped unit curtain walls. The method includes: the design side uses a BIM modeling tool to generate a first model; the production side performs production feasibility verification based on historical models; determines a first number of historical models whose similarity to the first type of unit corresponding to the first model that has not been produced is greater than a first threshold; and then analyzes whether production is possible based on these historical models; if so, trial production is performed; otherwise, feedback is given to the design side for redesign. Furthermore, the degree of customization of the first type of unit is evaluated, and the first number is dynamically adjusted based on its level; the degree of customization is set based on the score of the first information of the unit. This method strengthens the coordination between design and production by making full use of historical production data and experience, improves the accuracy and efficiency of production feasibility judgment, and reduces production risks and costs.

The invention provides an intelligent knowledge question-answering method, device and equipment, and relates to the technical field of artificial intelligence. The method comprises the following steps: acquiring a plurality of sample sequences and a general knowledge vector set of power vertical domain knowledge, wherein the sample sequences comprise a plurality of word units and time steps; performing word unit prediction analysis on the global context and the local context corresponding to each time step to obtain an alignment loss item; the alignment loss item represents the prediction difference between the global context and the local context; based on the general knowledge vector set, the historical model parameters and the currently updated model parameters, performing calculation to obtain knowledge loss items; the knowledge loss item represents the disturbance of the model parameter change on the general knowledge; based on the sample sequence, constructing a loss function by using a prediction loss item, an alignment loss item and a knowledge loss item, and training to obtain a question and answer model; and predicting the question of the user through the question and answer model to obtain the answer of the question. The answer prediction accuracy can be improved.

This invention discloses an intelligent layout system for the installation of exterior wall composite panels, relating to the field of intelligent construction technology. It collects building facade information to generate a structured digital twin model of the building site; matches the site model with historical construction digital twin models based on feature fingerprint similarity, extracting an initial intelligent layout model from the graph neural network architecture of the matching cases; analyzes the differences between the site model and the historical models and encodes them as adjustment instructions for the model input graph structure; performs transfer learning on the initial model, focusing on optimizing the difference areas, to obtain an optimized intelligent layout model; acquires site environmental data, combines it with a built-in physical and process knowledge base, drives the optimized model to generate multiple candidate schemes and perform construction feasibility simulations, outputting multi-dimensional construction scores for each scheme; compares the multi-dimensional construction scores and outputs the optimal layout scheme, solving the problems of traditional layout methods being unable to handle complex structures, ignoring environmental factors, and lacking reusable experience.

The invention discloses a federated learning-based risk assessment method, device and equipment and a storage medium, and is applied to the technical field of artificial intelligence, and the method comprises the steps: obtaining current training sample data, and carrying out the model updating of a local risk assessment model deployed under the own equipment based on the current training sample data, obtaining the current model parameter update quantity in the current model update time period; obtaining a historical model parameter update quantity, and determining a data quality score according to the current model parameter update quantity, the historical model parameter update quantity and the current training sample data; encrypting and transmitting the current model parameter update quantity, the data quality score and the node performance data of the equipment to a central server, performing model updating on a global risk assessment model deployed in the server, and encrypting and transmitting the updated global model parameter update quantity to each client; and updating the local risk assessment model under the own equipment based on the global model parameter update quantity.

A system and method for sampling a flight envelope of an air vehicle under test. The method includes obtaining engineering models of the vehicle to identify how the vehicle behaves, obtaining empirical models through test data, and obtaining historical models of similar legacy aircraft. The method further includes ensembling the models into an integrated model and assessing the ensemble uncertainty at any particular test point, for a variety of aircraft conditions, to determine model reliability. The method includes determining the optimal order of test points based on local model uncertainty, testing the air vehicle at a point, revising the integrated model with that data and reassessing the uncertainty, and determining if a stop testing condition has been met, in which case the testing series is closed. The last calibrated integrated model could then be used to generate trusted data for any remaining required test using a certification by analysis method.

Technologies are described herein for classifying personalized interactions at an application. A method can include receiving and storing user inputs associated with interactions between users of the application, building historical models based on the user inputs to classify the interactions, wherein building a historical model for a particular user comprises aggregating particular user inputs associated with a set of previous interactions between a particular user and other users, in association with a pending interaction for the particular user and during the pending transaction associating a current user input with the historical model for the particular user and based on analyzing the historical model for the particular user, determining that the pending interaction satisfies a condition, and interrupting the pending interaction based on determining that the pending interaction satisfies the condition.

The invention provides a method and system for determining machining time and a computer program product, and relates to the technical field of intelligent manufacturing. The method for determining the machining time comprises the steps that a three-dimensional model of a to-be-machined part is obtained, and the to-be-machined part is a machined part; based on the three-dimensional model, matching a first historical model from a historical database, and taking related data of the first historical model as template data; identifying to-be-processed features of the three-dimensional model through a part identification model; based on the template data, processing requirements of the to-be-processed features are determined; identifying feature parameters of the to-be-processed features by using a CAD module; and determining the machining time of the to-be-machined part based on the machining requirements, the feature parameters and the machining process knowledge base of the to-be-machined features. According to the method, the defects of inconsistent man-hour estimation and the like caused by disjunction of the process and calculation, lack of intelligent decision and incapability of self-optimization of the system in the traditional method are overcome.

The present invention relates to the field of power grid dispatching technology, specifically a method for predicting the aggregated flexibility of a photovoltaic, storage and charging integrated power station with mechanism-fused data; specifically: by constructing a unified linearized model of the operating constraints of photovoltaic, storage and charging equipment and the power distribution network current, combining Kirchhoff's laws to construct the multi-phase distribution network current equation and adopting the fixed point method for linearization to form a unified aggregation model; based on a closed-loop learning algorithm, utilizing network physical information and data-driven collaborative mechanisms, through dynamic screening of high-uncertainty samples, derivation of convex set properties and historical model transfer learning, under the premise of ensuring the convexity of the feasible solution space, the classifier is iteratively trained for multiple rounds to expand the coverage of the flexibility space. The present invention embeds the power grid topology information into the classifier training process through matrix operations, realizes the efficient fusion of mechanism constraints and machine learning, and effectively solves the problem of excessive conservatism caused by traditional random sampling or fixed geometric assumptions.