102 results about "Model network" patented technology

The invention discloses an OCT (Optical Coherence Tomography) retina image denoising method based on a high-frequency enhanced diffusion model, which is characterized in that a double-branch diffusion model network constructed based on frequency perception Fourier transform attention (FFTA) is used for separating different frequency domains, enhancing specific frequency domain features and fusing the specific frequency domain features into a spatial domain feature map to realize stronger frequency domain perception and processing capability; the invention relates to a time step T coding method for a module of frequency domain attention. The time step T of the diffusion model is used as a parameter to be coded into a frequency domain attention module to participate in self-attention matrix calculation, so that frequency domain feature processing is aligned with the iteration step number of the diffusion model, and different frequency domain information is pointedly processed at different time steps T; the frequency selective hopping mechanism uses pooling operation to obtain a high / low frequency characteristic pattern, so that the network has adaptive frequency domain retention capability for different local parts.

The invention relates to the technical field of oil and gas reservoir development and geological modeling, and discloses a multi-scale factor constraint-based tectonic network fracture modeling method, which comprises the following steps of: synthesizing a Riemannian metric tensor field on the basis of earthquake, logging and geomechanics data, and defining non-Euclidean distance cost of a fracture expanded in an anisotropic medium; initial seed points are screened according to the elastic strain energy density, and initial growth potential energy in a limited range is distributed; solving the eikonal equation by using an anisotropic fast marching algorithm to carry out wavefront competitive growth, and dividing a grid region into generalized Voronoi units; identifying a wavefront contact interface, and extracting gradient features to judge a fusion or truncation type so as to establish fracture topological connection; and finally, tracking a geodesic line path along an anti-gradient direction to generate a three-dimensional discrete fracture network. According to the method, macro and micro constraints are unified through Riemannian geometry, clear physical significance is given to the fracture by utilizing an energy mechanism, and automatic and accurate construction of the complex fracture network topology structure is realized.

A system includes one or more processors and memory. The memory stores instructions for execution by the one or more processors, including instructions for: obtaining a configuration request for a communications network; configuring a network of models (e.g., oscillators or oscillators'settings) into an initial configuration representing the configuration request for the communications network; reading out a final configuration of the network of models, the final configuration representing a solution to the configuration request for the communications network; and providing information over the communications network according to the configuration request.

The invention relates to the technical field of artificial intelligence, in particular to a production scene-oriented multi-modal large model system, terminal equipment and a storage medium. The system comprises a data acquisition and analysis module used for acquiring multi-modal data; the teacher model module is used for configuring and constructing a heterogeneous multi-teacher model network and generating a lightweight multi-modal sub-network according to resource constraints of target terminal equipment, and the student model comprises a shared perception coding layer, a plurality of modal specific channels and a plurality of task decision heads; a distillation training module for training a student model according to a mixed distillation loss function, and a deployment optimization module for obtaining hardware configuration information of a target terminal, generating an optimal student model adapted to the target terminal according to the hardware configuration information and completing a deployment test; the terminal equipment has the comprehensive judgment capability close to that of an artificial expert, can quickly respond to problems and tasks, and solves the problem of timeliness.

Techniques for building and maintaining model webs are described. In some examples, a model web is built by selecting models for the model web from one or more available model types based on at least one or more of availability, tensor information, and compute type, instantiating synapses between the selected models to form the model web and updating information regarding availability of the selected models of the model web to indicate being in use.

The invention relates to a performance adjustment method and device of a large language model, equipment and a storage medium. The method comprises the steps that a first service sample is processed based on a large language model to obtain a first reference result, the large language model comprises N network layers with the same function, and M1 network layers are pruned from the large language model based on weight factors to obtain a first intermediate model; the network layer has a weight factor which is used for representing the contribution degree of the network layer in executing the text processing task; processing the first business sample based on the first intermediate model to obtain a first prediction result; performing quantitative comparison based on the first reference result and the first prediction result to obtain performance loss of the first intermediate model; and when the performance loss of the first intermediate model is within the threshold range corresponding to the loss threshold, the first intermediate model is used as the business model, so that the text reasoning performance of the business model can be improved.

This invention discloses a method for constructing a digital twin model for smart water conservancy based on artificial intelligence, relating to the field of smart water conservancy technology. The method includes: establishing a project twin model based on smart water conservancy, performing related project analysis, and obtaining associatable equipment; establishing a twin model network and obtaining the response conditions and content of each node; and constructing a response analysis model based on the response content of real-time nodes. This invention addresses the problem in existing methods for constructing digital twin models for smart water conservancy that, when multiple types of water conservancy data exist and a certain type of water conservancy data exhibits abnormal changes, it is impossible to construct a targeted, real-time digital twin model for water conservancy equipment with significant data fluctuations. This results in an inability to effectively monitor abnormal equipment in water conservancy projects.

The invention relates to the technical field of communication, and discloses a reinforcement learning-based cross-space resource autonomous configuration method, which comprises the following steps of: performing joint modeling on network resources and physical resources by using a unified heterogeneous graph, and expressing a cross-space resource comprehensive configuration problem as a random and time-varying mixed integer nonlinear programming problem in a rolling time domain; introducing reinforcement learning to convert the nonlinear programming problem into a constrained Markov decision process; a resource configuration language instruction output by the policy network is mapped into a specific network domain and physical domain control instruction through a domain specific compiler, the specific network domain and physical domain control instruction is transactionally issued and executed, and an execution receipt is collected and fed back to the policy network to form closed-loop learning; according to the method, the engineering feasibility and the theoretical interpretability are unified. By means of a resource configuration language and triple masks, strategy output naturally falls in a compilable and rollback instruction subspace, and invalid exploration and configuration conflicts are remarkably reduced.

The application provides a digital twin architecture of a network, a network session processing method and device, the method comprising: a data acquisition and control entity configured to acquire characteristic data of a physical network corresponding to a virtual network; a core entity connected to the data acquisition and control entity, configured to construct a network-related model corresponding to the physical network according to the characteristic data acquired by the data acquisition and control entity, the network-related model being configured to perform convergent processing on a network session event; and a user entity connected to the core entity, configured to visually display the network based on the network-related model. In the embodiment of the application, the digital twin architecture comprehensively uses perception, calculation, modeling, simulation and other technologies, realizes virtual-real mapping and interaction, can fully train and efficiently simulate more intelligent applications with high error cost, reduces the risk generated when a new technology is verified in a physical space, and reduces the possibility of errors occurring when the new technology is deployed in a real environment.

This application provides a method and electronic device for network topology repair of a 3D model. The method includes: constructing a first mapping information containing mesh faces and corresponding halves based on 3D model mesh data; determining a second mapping information containing each vertex and its corresponding half based on this; then combining the first two to determine the twin relationship information of each half; next, identifying non-manifold vertices based on the three types of information and counting their sector counts; and finally, completing the repair of non-manifold vertices based on the sector count and the second mapping information. By constructing multi-layer mapping information, a complete topology data system is built. Each step supports parallel processing. It accurately identifies non-manifold vertices based on topological features and achieves targeted repair based on the sector count. While ensuring repair accuracy, it significantly improves the efficiency of mesh topology repair and provides stable topology data support for subsequent mesh processing.

The embodiment of the invention provides a photo volume generation method and device based on a large model, and the method comprises the steps: describing the contents of all reconnaissance images in a preset range based on a multi-modal large model according to a first cue word, and generating a detail title of each reconnaissance image; carrying out character analysis on the detail title of each survey image according to a second cue word based on a large language model to obtain a spatial hierarchy corresponding to the detail title; according to the investigation images of different spatial levels, feature extraction is carried out on the investigation images based on a diffusion model network, and duplicate removal is carried out on the same view angle images of the same spatial level; on the basis of a diffusion model network, sequentially segmenting each survey image from the survey image set of the uppermost spatial hierarchy, searching for the most similar survey image in the next spatial hierarchy, and performing spatial matching; based on a result of the spatial matching, a photo volume from a complete spatial hierarchy is generated.

The present application provides a kind of based on multi-source data fusion's ultra-high voltage line mountain fire prediction method, this method first constructs multiple parallel training systems, based on consistent data and model architecture, different training parameters complete training, screen five network parameters of optimal precision for actual prediction system;Again satellite original drawing, ultra-high voltage camera original drawing and two kinds of image cross-time difference graph are collected four kinds of pixel data, respectively input network model 1, network model 2, network model 3 and network model 4 extract features and output probability;Finally, the probability value is combined as feature vector input fusion network, and the final mountain fire classification result is output.The proposed ultra-high voltage line mountain fire prediction method solves the problem of traditional prediction relying on single data, incomplete feature extraction, low precision in complex scene, realizes multi-source feature deep fusion and multi-network collaborative prediction, optimizes feature capture and model generalization ability, and improves the accuracy and reliability of ultra-high voltage line mountain fire prediction.

The invention relates to a learning state monitoring method, system and device based on artificial intelligence and a medium. The method comprises the following steps: collecting signals such as physiological and behavior images of a target object through a multi-source heterogeneous sensor, and performing normalization processing on aligned features to generate a standardized feature vector; calculating a comprehensive cognitive state score of the feature vector by adopting a dynamic weighted fusion algorithm, and judging and outputting a soft label according to a preset threshold value; applying constraint random disturbance based on statistical distribution characteristics of the feature vectors to generate synthetic data, and associating the synthetic data with the soft labels to construct a soft label data set; by taking the feature vector of the data set as input and the soft label as a supervision signal, iteratively calculating loss through a preset loss formula, updating model network parameters to convergence, and obtaining an optimized neural network model; during real-time monitoring, processed sensor signals are input into the model, and quantitative learning state scores are output through end-to-end reasoning. The method can break through the limitation that traditional monitoring is high in subjectivity and insufficient in real-time performance.

The application discloses a kind of based on source domain and target domain joint training of self-supervision's covering manifold estimation method, comprising: the base station position sample corresponding to each region, topographic information sample and covering manifold sample are obtained to the specified region division;Source domain and target domain sample are sequentially subjected to data enhancement processing and dynamic mask processing;The neural network prediction model framework consisting of feature coding fusion module and covering manifold generation module is constructed, for sample input is executed progressive mask generation strategy after, until generating full mask, finally obtain covering manifold prediction result;Based on prediction result, calculate its reconstruction error with real covering manifold in mask area, and update model network parameters by back propagation.The application has reconstruction ability to target domain feature and generation ability based on source domain information simultaneously in training process, and finally model only needs to input source domain information to generate covering manifold prediction result.

The application belongs to the technical field of polarization imaging, and particularly relates to a water image despeckling model construction method based on polarization imaging, which comprises the following steps: using a feature extraction encoder to map four polarization angle images and calculated intensity images to a low-dimensional latent feature space; introducing a diffusion model network, a forward process of which is used to gradually add Gaussian noise until the noise data conform to a standard Gaussian distribution, starting from clear latent features extracted based on a clear water target intensity image and a turbid water target polarization image; a reverse process of the diffusion model network is used to denoise the noise data under the guidance of polarization features of the turbid water target, generate clear latent features of the underwater target, and obtain clear latent feature estimation results; using a reconstruction module network, based on the estimation results and the intensity image of the underwater target in the corresponding sample, a corresponding clear underwater target image is generated. The application can adapt to scattering environments of water bodies with different turbidity, and has high generalization and high robustness.

The invention provides a tidal current sediment motion prediction method and system based on deep learning, and relates to the technical field of deep learning, and the method comprises the steps: calling a pre-trained tidal current sediment prediction model to carry out the preliminary prediction of tidal current sediment associated data containing tidal current power associated information and sediment characteristic associated information, and generating an initial prediction result; acquiring actual observation data and an initial prediction result, performing deviation analysis to obtain deviation characteristics, inputting the deviation characteristics into a parameter adjustment module to dynamically adjust model network parameters to obtain an optimization model, calling the optimization model to perform prediction again to generate an intermediate prediction result, and finally performing iterative optimization in combination with the deviation characteristics to obtain an optimal prediction result. According to the method, the prediction deviation is continuously corrected through an iteration updating mechanism, a final accurate tidal current sediment movement prediction result is obtained, dynamic and accurate prediction of tidal current sediment movement is achieved, interaction and real-time changes among complex factors are fully considered, and the prediction reliability is greatly improved.

The invention discloses a self-supervised source domain and target domain combined training-based coverage manifold estimation method. The method comprises the following steps of: dividing a specified region to obtain a base station position sample, a topographic information sample and a coverage manifold sample corresponding to each region; sequentially performing data enhancement processing and dynamic mask processing on the source domain sample and the target domain sample; a neural network prediction model framework composed of a feature code fusion module and a coverage manifold generation module is constructed and used for executing a progressive mask generation strategy after sample input until a full mask is generated, and finally a coverage manifold prediction result is obtained; and based on a prediction result, calculating a reconstruction error between the prediction result and a real coverage manifold in a mask area, and updating model network parameters through back propagation. In the training process, the method has the target domain feature reconstruction capability and the source domain information-based generation capability at the same time, and the final model can generate a coverage manifold prediction result only by inputting the source domain information.

A system and method for automatically generating deep neural network architectures for time series prediction. The system includes a processor for: receiving a prediction context associated with a current use case; based on the associated prediction context, selecting a prediction model network configured for a current use case time series prediction task; replicating the selected prediction model network to create a plurality of candidate prediction model networks; inputting a time series data to each of the plurality of the candidate prediction model network; train, in parallel, each respective candidate prediction model network of the plurality with the input time series data; modifying each of the plurality of the candidate prediction model network by applying a respective different set of one or more model parameters while being trained in parallel; and determine a fittest modified prediction model network for solving the current use case time series prediction task.

The invention relates to a network task model construction method based on cross-architecture knowledge distillation, and the method comprises the steps: carrying out the encoding and projection of a multi-modal network operation data set through a multi-modal encoder, and obtaining a unified token; inputting the unified token into a teacher model comprising a Transform module and a first task specific network header, and outputting through the Transform module and the first task specific network header in the teacher model to obtain a teacher model network state prediction result; inputting the unified token into a student model comprising a Mama module and a second task specific network header, and outputting through the Mama module and the second task specific network header in the student model to obtain a student model network state prediction result; and constructing a total loss function based on the teacher model network state prediction result and the student model network state prediction result, and training the student model by using the total loss function to obtain a trained network task model. According to the method, the technical problems of calculation complexity and storage redundancy can be solved at the same time.

An image detection method based on embedded manifold representation and information bottleneck constraints is proposed. In the offline stage, a discriminator network and a generative flow model network (acting as an encoder or decoder) are initialized, and reconstruction loss function, information bottleneck loss function, and Jacobian matrix isometry loss function are calculated sequentially to obtain the total loss function. This allows for backpropagation training of both the generative flow model network and the discriminator network. In the online stage, the trained generative flow model network obtains discriminative and common features of the image to be tested. Image reconstruction and submanifold image reconstruction are performed only on images whose common features belong to the common feature space. Image classification is achieved by calculating the distance between the reconstructed image and the reconstructed images of each category's submanifold. This invention avoids meaningless calculations on invalid samples by pre-analyzing the images to be classified, saving computational resources. Furthermore, by utilizing the guidance of embedded manifold representation algorithm in image sample projection and information bottleneck constraints in feature extraction, it demonstrates excellent classification robustness.

The invention discloses a remote sensing crop identification domain adaptive model based on feature alignment, and the model comprises the following steps: 1, making a sample set; step 2, constructing a whole network architecture of the model; step 3, a detailed construction scheme of the model framework; step 4, a loss function fusion strategy; according to the feature alignment-based remote sensing crop recognition domain adaptive model, a cross-domain feature alignment strategy is introduced, so that effective fusion of a source domain and a target domain in a feature space is realized, and the adaptability and generalization ability of the model to different regions and time phase data are improved; the method not only improves the robustness and precision of crop identification, but also provides a technical path with higher generalization for remote sensing agricultural intelligentization.

The application discloses a paper medicine box steel seal character recognition method based on an improved YOLOV5 model, and the method comprises the following steps: collecting the steel seal character image of the medicine box by using an image acquisition device; performing image enhancement and other pretreatments on the image and labeling the recognition area; inputting into the improved YOLOV5 model for training to obtain the trained YOLOV5 model; the improved YOLOV5 model comprises the following steps: adding an efficient position attention mechanism CA (Coordinate attention) in the backbone network of YOLOV5, so that the model network can pay attention to a wide range of position information without bringing too much calculation amount, which is helpful to improve the model performance and better locate and identify the target; using SimSPPF to replace SPPF fast pooling layer to improve the training speed; for the dense small target, an additional smaller prior box is used in the model, and a smaller detection head is additionally added; inputting the data set to be recognized into processing to obtain the final prediction result.

The present application belongs to the technical field of speech conversion, and in particular relates to a singing speech conversion method based on CBAM and dynamic convolution decomposition, which comprises a training stage and a conversion stage, and a model network comprises a generator, a discriminator and a style encoder. First, dynamic convolution decomposition is introduced in the generator, dynamic channel fusion is used to replace the dynamic attention of the channel group, the problem that dynamic convolution in the generator causes the number of convolution weights to increase by K times is solved, the difficulty of joint optimization is reduced, and the model requires fewer parameters without sacrificing accuracy, thereby improving the running performance of the algorithm. Further, CBAM attention modules are introduced in the encoding network and the decoding network of the generator, attention is applied to the channels and the space, the attention and capture of the detailed information in the spectrum are improved, and the quality of the converted singing speech is significantly improved.

The application discloses an indocyanine green fluorescence image classification and recognition method based on a diffusion model and contrast learning, and comprises the following steps: step one, collecting indocyanine green fluorescence images as a data set; step two, inputting the images obtained in step one into a diffusion model network after cutting and scaling processing, and training a diffusion model; step three, training a contrast learning model based on the diffusion model based on the pre-trained diffusion model in step two; and step four, fine-tuning the model in step three, and detecting and classifying indocyanine green fluorescence images. The application learns the distribution between real samples by using the diffusion model, and fits the real distribution by using the learned distribution, thereby solving the problems of low data set quality and insufficient samples, and improving the classification ability of the model for indocyanine green fluorescence images.

The invention belongs to the technical field of rotor dynamics, and discloses a gas turbine multi-disc rotor imbalance fault diagnosis method and a related device. The method comprises the following steps: extracting a feature vector based on acquired abnormal vibration data; on the basis of the feature vectors, a rotor unbalance fault diagnosis model is used for identifying and positioning a fault source, and a multi-disc rotor unbalance fault online identification and tracing result is obtained; the rotor imbalance fault diagnosis model is composed of a low-fidelity model network and a high-fidelity model network, and the low-fidelity model network is used for predicting a low-fidelity result under the condition of given input data after training. And the high-fidelity model network is used for acquiring a high-fidelity prediction result by learning a mapping relationship between the low-fidelity output result and the high-fidelity output result. According to the invention, accurate identification and online traceability of the unbalance fault parameters of the multi-disc rotor are realized.

The invention discloses an unmanned aerial vehicle visual positioning method and system based on multi-model intelligent extraction, and relates to the technical field of unmanned aerial vehicle visual positioning, and the method comprises the steps: constructing a multi-model network; training the multi-model network; inputting the satellite image into a multi-model network, and establishing a satellite feature library; inputting the unmanned aerial vehicle image into a multi-model network to obtain unmanned aerial vehicle features; similarity matching is carried out on the unmanned aerial vehicle features and each satellite feature in a satellite feature library, and the satellite feature with the highest similarity is determined; and determining a geographic position of a satellite image corresponding to the satellite feature as a visual positioning result of the unmanned aerial vehicle. According to the method, the consistent robust feature representation of the multi-source image can be extracted, the separated training strategy can fully optimize the multi-model network and improve the network performance, the intelligent extraction technology based on the intelligent transmission network thought can realize variable scene adaptation, and finally the unmanned aerial vehicle visual positioning task under the complex condition is completed.

The invention discloses a multi-energy complementary operation control system, and relates to the technical field of electric power systems and automation thereof. A multi-micro-grid parallel system structure comprising various renewable energy sources such as wind power, photovoltaic energy and energy storage is introduced, and a distributed controller is configured at each micro-grid end; the graph neural network is utilized to perform unified modeling on the network topology and the real-time power operation state, and the power oscillation risk index and the target output power share correction can be output at the same time, so that the power distribution strategy is upgraded from the traditional static rule to the risk-oriented dynamic distribution. The power swing of a high-risk channel is actively weakened on the premise that the total power requirement and multi-energy complementary utilization are met; in combination with energy storage charging and discharging coordination control based on oscillation risk and power share, an energy storage unit provides effective damping support for power oscillation while ensuring a power distribution target.

The embodiment of the present disclosure provides a model parameter determination method, a model inference system, a decoding device and electronic equipment, which relates to the technical field of model inference, and comprises the following steps: determining a to-be-decoded code stream in a first weight code stream based on a weight parameter, the first weight code stream being obtained by performing prefix encoding processing on an initial weight matrix of a current network layer; and performing prefix decoding processing on the to-be-decoded code stream to obtain target weight data, the target weight data being used for sparse matrix multiplication calculation of the current network layer. The method provided by the present disclosure compresses the data transmission amount through prefix encoding processing, reduces the bandwidth requirement, and simultaneously reduces the bandwidth overhead during model network inference through the cooperative processing of prefix decoding of the to-be-decoded code stream determined by the weight parameter part, thereby improving the model processing efficiency.

The application discloses a kind of unmanned plane small target identification method and system, comprising: S1.unmanned plane collects image data, and image data is preprocessed;S2.YOLOv11 model network framework is constructed, SCSA module and SAFM module are introduced in main network and loss function is optimized, combined with Mosai c data enhancement technology, to enhance the recognition and extraction ability of small target features of model;S3.YOLOv11 model training learning optimization is carried out;S4.the model trained is deployed to embedded device, real-time processing unmanned plane collected image data is used to output target identification result;S5.lastly, the result is output and fed back to unmanned plane.The network structure of the application optimizes YOLOv11 algorithm, significantly improves the recognition accuracy of small target of unmanned plane, significantly improves the recognition accuracy and real-time performance of small target of unmanned plane, and adapts to the calculation resource limitation of embedded device of unmanned plane.