35 results about "Consistency model" patented technology

The invention discloses a progressive fine tuning method and system for a multi-modal pre-training model, and belongs to the field of deep learning, and the method comprises the steps: obtaining a high-dimensional visual feature vector from a visual encoder of a pre-trained multi-modal large model, obtaining a text feature vector from a text encoder, and carrying out the construction to obtain a heterogeneous modal feature; the contribution degrees of different modes are analyzed through a resource allocation strategy, the whole fine adjustment process is dynamically guided, and limited computing resources are allocated to a multi-mode large model component which contributes to the current task most; and processing the heterogeneous modal features through a cross-modal comparison consistency model to obtain a final optimization target. According to the invention, a larger batch can be trained or a larger batch size can be used for fine tuning training under a limited hardware condition.

The invention provides a probabilistic multi-scale-based missile loading vehicle tail end relative attitude estimation method and system, and belongs to the technical field of aviation equipment automation. Comprising the following steps: firstly, performing coordinate system conversion and noise suppression processing on acquired point cloud data, and constructing a probability multi-scale consistency model; then, on the basis of weighted features obtained by the probability multi-scale consistency model, density self-adaptive multi-scale feature fusion is executed; and finally, according to the multi-scale features after feature fusion, performing dual-stage pose estimation. Through dynamic noise suppression, component-level multi-scale feature modeling and a real-time optimization strategy, millimeter-level and high-reliability alignment between the end effector of the missile loading vehicle and the missile loading frame is achieved.

The application discloses a kind of few sample X-ray defect intelligent detection methods based on diffusion generative model, belong to industrial nondestructive testing and artificial intelligence technical field;The method first acquires preprocessed X-ray image dataset, constructs defect condition embedding that retains physical characteristics;Build the condition diffusion generative model that introduces rectified flow reparameterization and consistency model distillation mechanism optimization, trains high-fidelity defect sample in stages generation;Build multiscale Transform detection network, through mixed sample training, dynamic loss optimization, course learning and feature consistency constraint, by the intermediate feature of detection network is fed back to generative model to adjust defect condition weight, realize the joint training of generative model and detection network;Finally input image to be detected completes defect positioning classification.The application can generate multiple types of real defect samples, eliminate sample imbalance and domain bias, with high detection accuracy and real-time performance under few sample conditions.

This application provides an image generation method, apparatus, device, medium, and program product. The method includes: acquiring text data; determining an initial latent code; iteratively reasoning on the text data and the initial latent code using a target model to obtain a target latent code; wherein the target model includes a latent consistency model and a text-to-image model; the text-to-image model includes a parameter-adjusted Unet module; the parameter-adjusted Unet module is trained using target sample data; the target sample data includes a set of images carrying text labels; the image set includes at least a first set; the first set includes a set of scene images corresponding to a target scene; the target scene includes the application scene of the target model; and generating a target image based on the target latent code. This application improves the efficiency of target image generation.

The application provides a surface feature-based anti-counterfeiting information generation method, device, equipment and medium, which comprises the following steps: firstly, a microstructure intensity map is obtained by extracting surface micro-relief features through a phase consistency model; a first height map is obtained by reconstructing an initial surface height field by using a MiDaS depth estimation network; a local displacement field caused by embedding a preset mark is quantified by using a RAFT optical flow model; an initial stress disturbance distribution is constructed by fusing the height field and the displacement field as physical input of a coupling prediction model, and then the surface morphology of a concrete sample after solidification is predicted; finally, the microstructure intensity map, the local displacement field and the first height map are taken as initial surface features at the sampling time, and the initial surface features and target surface features are encoded as anti-counterfeiting information, so that the reliability of sample identification is effectively improved.

The DoS attack strategy making method based on the multilayer cut points is provided for consistency control of the multi-agent system, and the state consistency of the multi-agent system is effectively influenced under the condition that it is guaranteed that the attacker consumes relatively little energy. The method comprises the steps that a multi-agent system consistency model and a DoS attack model are established, an objective function is obtained by taking energy consumption of an attacker and a system consistency error as two indexes, multi-layer cut points are used for replacing an attack action space to improve the optimization efficiency of the objective function, and an attack strategy is obtained by optimizing the objective function. The attack research essence is to guarantee the improvement of the system security performance, and the method provided by the invention can be used as a link for testing the system security to help discover the defects in the aspect of system defense.

The invention is suitable for the technical field of deep learning, and provides a data set generation method, a model training method and a computer, and the method comprises the steps: obtaining a video stream containing a target object, and enabling a camera to move around the target object, and carrying out the shooting of the video stream; receiving interactive annotation input of a user on the target object in a starting frame image of the video stream; calling an image segmentation model to generate an initial mask based on the interactive annotation input; based on the initial mask, automatically tracking and generating an object mask sequence of subsequent multiple frames of the video stream through a time sequence consistency model; and according to an object model of the target object, the initial mask and the object mask sequence, performing pose labeling on the multiple frames of images in the video stream to obtain real pose data. Through the method, the training data set can be automatically generated without manual annotation under the condition of not depending on equipment such as a mechanical arm and a rotary table.

The invention provides an anti-fake information generation method and device based on surface features, equipment and a medium. The method comprises the steps that firstly, surface micro-fluctuation features are extracted through a phase consistency model to obtain a microstructure strength diagram; reconstructing the initial surface height field by using a MiDaS depth estimation network to obtain a first height map, and quantifying a local displacement field caused by embedding of a preset identifier through an RAFT optical flow model; the height field and the displacement field are fused to construct initial stress disturbance distribution to serve as physical input of a coupling prediction model, and then the surface appearance of the concrete sample after solidification is predicted; and finally, the microstructure strength diagram, the local displacement field and the first height diagram are used as initial surface features during sampling, and the initial surface features and the target surface features are coded into anti-counterfeiting information, so that the reliability of sample identification is effectively improved.

The application provides a data center liquid cooling square tank cold source multi-parameter fusion monitoring and early warning method, relates to intelligent operation and maintenance, and comprises the following steps: collecting multi-source time sequence data representing fluid dynamics state and environment heat and humidity state through a multi-modal perception unit arranged in a liquid cooling square tank cooling loop and a cabin environment, and constructing a state observation sequence; based on the state observation sequence, using a physical consistency model embedded with fluid dynamics constraints, heat exchange constraints and loop response consistency constraints, the unobservable state in the cooling loop is inverted to obtain a hidden state vector; according to the hidden state vector and the state observation sequence, a time-varying coupling relationship reflecting parameter change amplitude and response time difference is established to generate a multi-dimensional correlation feature representation; taking the multi-dimensional correlation feature representation as a driving condition, a heat and humidity coupling digital twin model is constructed with the time-varying coupling relationship as a boundary driving condition, and a spatial temperature field, a humidity field and a condensation risk distribution are obtained, which can accurately locate the condensation risk area before the fault occurs and actively intervene, thereby improving the safety and reliability of the liquid cooling square tank operation.

Systems and methods, implemented as computer programs on one or more computers for training a consistency model for use in generating a frame of data, such as a frame of image data, and methods of using a trained consistency model to generate a frame of data. A consistency model is used to generate a frame of data by predicting a succession of de-noised frames starting with an initial, noisy frame at an initial time and ending with a final frame, without noise, at a final time. The consistency model is trained to generate self-consistent predictions over a trajectory of predicted frames corresponding to these times. Implementations of the described techniques divide the trajectory into segments and only require the model to generate self-consistent predictions over each segment. This can facilitate the rapid generation of high quality frames of data.

The invention belongs to the field of image consistency processing, and discloses an image consistency model adaptive discretization method for optimizing visual angle inspiration, which comprises the steps of diffusion model setting, adaptive discretization and the like. From the perspective of optimization, a joint target of local consistency and global consistency is constructed, a time step updating strategy in an analysis form is exported by using a Gaussian-Newton method and is seamlessly integrated into a training process, and self-adaptive adjustment and efficient convergence of time steps are realized. According to the method, various existing discrete strategies are unified, the training efficiency and final performance of image generation are remarkably improved, and a new theoretical tool and key technical support are provided for design and scheduling of a high-performance image generation model.

A consistency model is trained to mimic the output of a diffusion model at various points along the denoising trajectory. A trajectory of the diffusion model is determined by generating the data point with the diffusion model by sampling a noised data point and applying denoising steps of the diffusion model to obtain the denoised output. At each of the noise levels, the consistency model is applied to the corresponding data point to remove the remaining noise. The resulting data point from the consistency model is compared with the denoised output of the diffusion model. An error for the consistency model may then be determined based on the comparisons at the various points in the trajectory.

The application discloses a representation enhancement aspect sentiment triple extraction method based on contrast learning, and belongs to the technical field of artificial intelligence. In view of the problems of the existing method, such as too fine contrast granularity of word units, inaccurate boundaries, incomplete aspect word or sentiment word extraction, and lack of cross-sample semantic unified modeling, the method proposes a cross-sample segment-level contrast learning mechanism: obtaining word unit representation through a pre-training model, constructing a grid annotation matrix and a structure contrast loss, simultaneously extracting segment representation from multiple samples and calculating a segment-level contrast loss, and loss weighting back propagation optimization model. The method can enhance the consistency modeling capability of the model for similar semantic segments, and improve the extraction accuracy of semantic segments composed of multiple words and the generalization and accuracy of triple extraction.

This invention relates to the field of speech synthesis technology, specifically to a speech synthesis method based on an implicit continuous consistency model. The method involves constructing a dataset containing audio and its text annotations; building a residual vector quantization variational autoencoder, training it using a joint loss until convergence, and extracting the mean and variance of all audio data mapped to a latent vector distribution; sampling Gaussian noise latent variables based on the mean and variance, sampling Gaussian noise and time steps, and adding noise to obtain speaker features; calculating the continuous consistency loss using the time steps, speaker features, text, audio data, and the consistency model; and optimizing the continuous consistency loss until the consistency model converges. This method utilizes residual vector quantization technology to achieve high-rate audio feature compression and decoupling, and combines the single-step sampling characteristics of the consistency model in the latent space to improve the inference efficiency and training stability of speech synthesis, enabling the model to generate high-fidelity speech in a very small number of iterations.

The invention relates to a reinforcement learning-based text-to-image consistency acceleration method and system, which can model the reasoning process of a consistency model into a Markov decision process (MDP), and directly optimize a task-specific reward function by using reinforcement learning, thereby realizing rapid training and efficient reasoning text-to-image generation.

The invention discloses an end-to-end reservoir parameter inversion method based on a consistency model and phase control constraint, and relates to the technical field of reservoir engineering. The method comprises the following steps: firstly, acquiring geological parameters and production dynamic data of a target area, constructing a sample database and carrying out data preprocessing, then constructing a consistency model and an improved U-shaped network, and integrating the consistency model and the improved U-shaped network to construct an end-to-end oil reservoir parameter inversion model considering phase control constraint; and training and verifying the end-to-end oil reservoir parameter inversion model by using the database, and after the end-to-end oil reservoir parameter inversion model is optimized, performing oil reservoir end-to-end history fitting on a specified oil reservoir by using the end-to-end oil reservoir parameter inversion model, and inverting an oil reservoir geological parameter field. According to the method, the physical mechanism constraint is combined with a model efficient generation method, the consistency model and the sedimentary facies distribution constraint are introduced into the end-to-end oil reservoir parameter inversion model, the inversion robustness of the end-to-end oil reservoir parameter inversion model is improved, and technical support is provided for accurate inversion of an oil reservoir geological parameter field.

The application discloses a kind of under DoS attack based on dynamic event triggering's spacecraft formation system's self-adapting consistency control method.The method is first based on multi-agent consistency model theory, establishes the multi-agent system model of multi-agent spacecraft formation system, then consider based on time series DoS attack, by the frequency and duration of DoS attack are analyzed and researched, to solve the consistency problem of multi-agent spacecraft formation under DoS attack, then design a dynamic event triggering mechanism to reduce communication transmission frequency, and it is proved that there is no Zeno behavior existence, finally based on Lyapunov stability theory analysis, design a kind of event triggered distributed adaptive controller, solve the consistency problem of multi-spacecraft formation.The method is applied to multi-spacecraft formation system, guarantee the normal operation of system under DoS attack.

Systems and methods are disclosed that perform a truncated consistency model training framework that includes two stages. For example, in the first stage, embodiments of the present disclosure may train a consistency model using first and second time step samples. The first time step samples may be obtained based on sampling from a plurality of time steps. Following, a truncated time range that does not include all of the time steps from the plurality of time steps is obtained. Then, third time step samples are obtained based on sampling from the truncated time range and fourth time step samples are determined based on the third time step samples and a time step difference. Afterwards, in a second stage, the consistency model is further trained using the third time step samples and the fourth time step samples.

The application discloses a big data-based information technology detection system with security protection, comprising the following modules: a log access and preprocessing module for collecting multi-source heterogeneous log big data and generating standardized feature data stream; a label alignment and embedding module for performing local label standardization and embedding mapping on the standardized feature data stream at each participating node and generating a fusion feature vector; a model training and updating module for generating an encrypted gradient; a federal coordination module for obtaining a global model; a distillation guidance module for generating a distillation consistency model; a blockchain storage module for writing parameters in the global model of the current round into a side chain storage; and an inference deployment module for deploying the fused global model and the distillation consistency model to each node. The application provides basic support for security and governance in a multi-agency collaborative environment.

The application discloses a kind of railway along the line environment cross-modal remote sensing data fast matching method, comprising: S1, data acquisition;S2, generate anisotropic weighted moment diagram by phase consistency model diffusion;S3, phase directional description matching is carried out, and initial matching point pair is obtained;S4, initial inner point pair set and initial transformation matrix are obtained by initial solution estimation of FSC;S5, initial transformation matrix is optimized using weighted average optimization algorithm, and the best transformation matrix after expansion is obtained;S6, according to the best transformation matrix after expansion, calculate matching error, dynamic adjustment error threshold;S7, check convergence: iteration is carried out until convergence is stopped.The method considers global optimization, avoids convergence solution to fall into local optimum, solves the problem that the high mismatching rate of heterogeneous image is caused by spectral difference, and the problem that inner point screening is not complete, and the registration precision is significantly optimized.

The invention discloses a big data-based information technology detection system with security protection, and the system comprises the following modules: a log access and preprocessing module which is used for collecting multi-source heterogeneous log big data and generating a standardized feature data stream; the label alignment and embedding module is used for executing local label standardization and embedding mapping on the standardized feature data flow at each participating node and generating a fusion feature vector; the model training and updating module is used for generating an encryption gradient; the federation coordination module is used for obtaining a global model; the distillation guidance module is used for generating a distillation consistency model; the block chain evidence storage module is used for writing parameters in the global model of the current round into a side chain evidence storage; and the reasoning deployment module is used for deploying the fused global model and the distillation consistency model to each node. According to the invention, basic support is provided for safety and treatment in a multi-mechanism collaborative environment.

The application discloses a kind of space intelligence depth consistent and geometric correction high-precision point cloud reconstruction system, it is related to space intelligence field, solve the problem of insufficient accuracy of present point cloud model reconstruction, including: multi-modal perception module: data acquisition is carried out to space environment;Consistency modeling module: space environment is segmented, and unit model is obtained;Consistency optimization module: unit model is analyzed to overlap, and unit model is optimized;Self-adapting correction module: construct space residual graph, generate compensation field, and carry out self-adapting adjustment based on compensation field;Structure fusion and balance module: construct overall point cloud model, and optimize fusion is carried out to model;Feedback optimization module: point cloud reconstruction is carried out to overall point cloud model, compare the residual change of space residual graph, and feedback optimization is carried out to point cloud reconstruction;Check output module: consistency check is carried out to overall point cloud model;The application can effectively improve the accuracy of model reconstruction.

The invention discloses a high-precision point cloud reconstruction system for space intelligence depth consistency and geometric correction, relates to the field of space intelligence, and solves the problem of insufficient reconstruction accuracy of an existing point cloud model, and the system comprises a multi-modal sensing module which carries out the data collection of a space environment; the consistency modeling module is used for segmenting a space environment to obtain unit models; the consistency optimization module is used for performing overlap analysis on the unit models and optimizing the unit models; the self-adaptive correction module is used for constructing a spatial residual image, generating a compensation field and carrying out self-adaptive adjustment based on the compensation field; the structure fusion and balance module is used for constructing an overall point cloud model and carrying out optimization fusion on the model; the feedback optimization module is used for performing point cloud reconstruction on the overall point cloud model, comparing residual changes of the spatial residual images and performing feedback optimization on the point cloud reconstruction; the verification output module is used for carrying out consistency verification on the whole point cloud model; the method can effectively improve the accuracy of model reconstruction.

The invention discloses a method, a system, equipment and a medium for transforming an in-situ control unit, and belongs to the technical field of monitoring, and the method comprises the following steps: constructing a homologous bijection data consistency model, configuring new and old master control sink nodes to map the same database table, and determining a coexistence basis; establishing a time-space replacement type relay topology, moving a source node to a temporary area reconstruction link, and maintaining monitoring on a lower-level edge node while releasing an in-situ space; deploying a target node in situ to form a new and old node parallel double-track complementary monitoring architecture; based on the architecture, incremental iteration replacement is carried out on edge expansion nodes, and smooth migration of the system is realized by dynamically switching links. The system comprises a consistency model construction module, a topology reconstruction module, a double-track architecture deployment module and a smooth migration execution module. The space contradiction is solved through data mapping and shifting, the out-of-control range is narrowed through double-track parallel and incremental replacement, and the transformation flexibility and safety are improved.

This invention relates to the field of system-on-chip (SoC) testing technology, specifically to a dynamic verification method for SoC chips based on a memory consistency model. The method includes collecting memory access data entering the cache controller and generating feature vectors; classifying the feature vectors using a decision tree to filter out risky memory data; expanding test cases based on the risky memory data to obtain test cases; testing the SoC chip based on the test cases; and iterating the decision tree model based on the test results. Addressing the problem that existing SoC verification test cases are ineffective against dynamically changing memory read / write patterns, this invention introduces a hardware monitoring unit into the cache controller to directly capture memory read / write data appearing in the cache. A decision tree model is then used to filter out risky memory data that may cause failures. Based on this, test cases are constructed and expanded to facilitate testing of the SoC's memory read / write mechanism, and the decision tree model is improved to obtain more test cases.

The application discloses a kind of progressive fine-tuning method and system of multimodal pre-training model, belong to the field of depth learning, the method includes from the visual encoder of multimodal large model that has completed pre-training obtains high-dimensional visual feature vector, and obtains text feature vector from text encoder, and heteromodal feature is constructed;The contribution degree of different modalities is analyzed by resource allocation strategy, and the whole fine-tuning process is dynamically guided, and the limited computing resources are allocated to the multimodal large model component that contributes most to the current task;The final optimization target is obtained by processing the heteromodal feature by cross-modal contrast consistency model.The application can train larger or use larger batch size for fine-tuning training under limited hardware conditions.

The invention discloses a quick matching method for cross-modal remote sensing data of an environment along a railway. The method comprises the following steps: S1, data acquisition; s2, generating an anisotropic weighted torque graph through phase consistency model diffusion; s3, performing phase orientation description matching to obtain initial matching point pairs; s4, estimating through an initial solution of the FSC to obtain an initial inner point pair set and an initial transformation matrix; s5, optimizing the initial transformation matrix by using a weighted average optimization algorithm to obtain an expanded optimal transformation matrix; s6, calculating a matching error according to the expanded optimal transformation matrix, and dynamically adjusting an error threshold value; and S7, checking convergence: carrying out iteration until convergence. According to the method, global optimization is considered, the convergence solution is prevented from falling into local optimum, the problems of high mismatching rate and incomplete interior point screening caused by spectral difference of different-source images are solved, and the registration precision is remarkably optimized.

This invention discloses a diffusion model processing method, apparatus, electronic device, and storage medium for unified image restoration. Based on a virtual consistency function, this invention establishes a virtual consistency model and introduces a lightweight noise correction and cue word refinement network, which can be used for noise and cue embedding modification, enhancing image restoration performance without retraining the pre-trained diffusion model. By using the virtual consistency function, the virtual consistency model VCMUIR leverages the uniform degradation characteristics of the high-noise space of the diffusion model, effectively solving general image restoration problems while significantly improving fidelity. This invention can adaptively handle different degradation types to enhance robustness and can simultaneously handle different degradation tasks, thus providing a general image restoration model that eliminates the need for task-specific models, thereby improving image restoration efficiency. This invention has wide applications in the field of image processing technology.

This invention discloses a dynamic laser SLAM method based on adaptive point cloud clustering and model residual verification. First, the original point cloud is downsampled and a KD-tree is constructed. Then, ground points are removed using an adaptive RANSAC algorithm based on height priors and normal vector constraints. Subsequently, an improved DBSCAN algorithm is employed, combining local density and FPFH feature similarity for adaptive clustering. Finally, a motion consistency model is constructed, and dynamic targets are removed through centroid residual calculation and multi-frame verification using a sliding window, resulting in a static point cloud for localization and mapping. This invention effectively reduces interference from dynamic objects and improves the robot's localization accuracy and map robustness in indoor dynamic environments.

The invention discloses a transformer substation switching operation intelligent monitoring method and system based on multiple data, and the method comprises the steps: constructing a state node which comprises a topological matrix, an electrical quantity vector, a step mark and a timestamp according to a state change event in a switching operation process, and forming a time sequence topological evolution diagram; extracting evolution embedding features through a graph neural network, and matching the evolution embedding features with a legal pattern library to obtain a structural deviation score; meanwhile, an equivalent impedance network is constructed based on a current topology and a protection section boundary, virtual fault calculation is performed in combination with a key fault point, a protection expected action matrix is generated, and a semantic deviation score is obtained according to the protection expected action matrix. And on the basis, the evolution embedded features and the protection semantic features are fused to construct a joint feature vector, a joint deviation score is obtained through a joint consistency model, and finally, the deviation scores are synthesized to form a risk assessment result. According to the method, the structural and semantic consistency analysis of the whole switching operation process is realized, and the accuracy and real-time performance of anomaly recognition are improved.