9 results about "Local consistency" patented technology

The invention discloses a medical text condition generation method under rule constraints, and belongs to the technical field of natural language processing and medical information. According to the method, a medical rule knowledge graph is constructed, graph neural network coding is adopted, and medical business rules are converted into constraint representation; designing a rule-perceived attention mechanism, and fusing medical rule constraints in a generation process; constructing a constraint-aware Transform decoder, and ensuring the generation normalization by adopting an improved beam search algorithm; and establishing a multi-level quality control system, and performing three-dimensional evaluation from local consistency, global consistency and rule conformity to generate a result. According to the method, the technical problem that existing medical text generation cannot meet complex rule constraints while guaranteeing fluency is solved, the rule coincidence rate reaches 95% or above, the quality score is improved by 18-25% compared with an existing method, and the method is excellent in performance in application of multiple medical departments.

The invention relates to the technical field of internet security, and particularly provides a black box graph fraud detection collusion model editing method based on a conditional diffusion model, which comprises the following steps: S1, constructing a local subgraph context coding mechanism; s2, establishing a condition diffusion generation model of the potential space; s3, designing a joint generation strategy of features and edges; and S4, collaborative optimization based on adversarial loss. Under the combined action of related technical schemes, the problem that in the prior art, a method for cooperatively generating nodes with collusion characteristics and edge attacks based on a conditional diffusion model under the black box and only local view constraint is lacked is solved; according to the method, real collusion attack simulation is realized, local consistency and high concealment are ensured, the problem of segmentation of feature-topology generation is solved, strict black box and local access constraints are met, potential security vulnerabilities in an industrial financial risk control system can be revealed, and a key technical support is provided for developing a next-generation robust defense mechanism.

A residual life prediction method based on domain invariance and consistent ordinal number representation learning comprises the following steps: firstly, calculating an MMD distance between different bearing domains in a training set, and minimizing the distance through gradient descent to enable feature distribution of each source domain to be globally aligned, extracting domain-invariant and ordered degeneration features, introducing a Gaussian mixture prior coding-decoding structure, and finally obtaining a residual life prediction result; the features are mapped to a potential space constrained by Gaussian mixture distribution through a variational auto-encoder, and the inadaptability of traditional single Gaussian distribution to nonlinear degradation is reduced; and in combination with MMD alignment, ordinal number loss and local consistency constraints, the dominant situation of a single constraint is avoided, the feature space has domain invariance and continuous orderliness at the same time through joint optimization, and the generalization ability of the model in a cross-domain task is remarkably improved.

The invention provides a local consistency guided sparse label enhancement method, which is suitable for detecting a drivable area on a road and belongs to the technical field of images. The method aims at solving the problem that an existing deep learning model depends on a large amount of pixel-level annotation data, firstly, sparse annotation is conducted on an input image, context enhancement features are constructed according to local and global image representation, and the similarity relation between super-pixel nodes is established; and then constructing a label propagation model based on the graph convolutional network, and propagating the sparse labels to the unlabeled areas to generate pseudo labels. According to the method, a weak supervision training strategy guided by local consistency is adopted, a joint loss function is designed, and collaborative supervision is carried out on a labeled region and an unlabeled region, so that the reliability of a pseudo label and the overall segmentation precision are improved. Experimental results show that the method can be suitable for various road drivable area detection tasks, and the obtained high-quality pixel-level pseudo label can be used for subsequent fully supervised model training.

The invention relates to the technical field of distributed service registration, in particular to an anti-partitioning method for a distributed service registration center, which comprises the following steps of: S1, deploying a Raft algorithm-based cluster in each machine room of a system, and realizing strong consistency of service registration data in a region by the cluster through a free time number and majority log copying mechanism, wherein the Raft cluster is composed of 2f + 1 nodes, and f is the tolerable maximum fault node number; s2, a Gossip ring network is formed among the machine rooms, service registration increments are propagated through asynchronous event distribution and an adaptive probability adjustment formula, in the method, Raft clusters are deployed in the machine rooms to guarantee local area strong consistency, and a Gossip mechanism with probability adjustment is adopted among the machine rooms to achieve final consistency; an independent arbitration subsystem is introduced to prevent split brain, and the availability of a copy is improved in combination with a fault domain isolation strategy; and meanwhile, GSLB / DNS and SDN intelligent scheduling are fused, so that a client request is always guided to an optimal service node.

The invention relates to an out-of-distribution detection method based on feature consistency enhancement. The method is especially suitable for identification and processing of near-out-of-distribution (Near-OOD) samples in a natural language processing system. According to the method, through a cross-layer complementary consistency alignment mechanism, differences among multi-scale feature representations are disclosed; further adopting an edge self-adaptive adversarial regularization mode to guide the model to learn a decision boundary with higher discrimination; meanwhile, a local consistency fusion detector is introduced, fusion judgment is carried out on the similarity of the input samples in the global level and the local level, and therefore the recognition capacity of the model for the Nearar-OOD samples in the feature space is improved. According to the method, the detection robustness and the discrimination precision of the model in a feature height overlapping region can be remarkably enhanced, and the method is suitable for constructing a natural language processing model with higher safety and generalization ability.

The invention provides a combined and deconstructed block matching learning anatomical consistency self-supervision method and system, and the method comprises the steps: generating a multi-scale cutting block from an original image through a unique learning branch, extracting block embedding and global embedding through a student encoder and a teacher encoder, obtaining output embedding through the projection of a unique head, and carrying out the self-supervision of the block matching learning anatomical consistency. Calculating cross entropy loss; dividing an image block into grids through a consistency learning branch, extracting two overlapped cutting blocks, calculating an embedded correlation matrix, and optimizing local consistency based on step loss; through combination and decomposition learning branches, cutting blocks of different scales are extracted, combination learning and decomposition learning are carried out respectively, a combination matching matrix and a decomposition matching matrix are calculated, and combination and decomposition consistency is optimized based on cross entropy loss; and alternately training the uniqueness learning branch, the consistency learning branch and the combination and decomposition learning branch according to a preset rule by adopting a cyclic pre-training strategy.

The invention belongs to the technical field of electronic design automation, and discloses a system block diagram automatic analysis and task question and answer method and system based on multi-source fusion, and the method comprises the steps: carrying out the preprocessing of a system block diagram, carrying out the component recognition of a preprocessed image through a target detection model, carrying out the line segment extraction through a multi-algorithm fusion strategy, and carrying out the task question and answer. Constructing a physical topology, and judging the signal flow direction by adopting a four-stage strategy comprising loop closing and local consistency; a circuit knowledge question and answer data set is created, Qwen2.5-VL-3B is finely tuned by adopting a secondary fine tuning method from overall reasoning to local enhancement, a question and answer text generated by a large model is analyzed, and structured verification and path correction are performed on a question and answer result based on connectivity of physical topology. According to the method, the problems of character interference, line segment breakage and illusion connection generated by a large model in the block diagram are effectively solved, and the accuracy of complex circuit diagram analysis and the interpretability of content are remarkably improved through closed-loop fusion of visual hard constraint and semantic soft reasoning.

The application discloses a sports event stream real-time analysis method, and particularly relates to the technical field of event data processing; the time stamp of a multi-source heterogeneous original event stream is standardized to obtain an event sequence under a unified time axis; subsequently, a sliding window is used to extract a candidate event set, and the time sequence offset feature and the semantic similarity feature between events are calculated to construct a feature set; on the basis, an event correlation graph is constructed, the connection relationship between nodes is generated by weighting the time sequence offset feature and the semantic similarity feature; further, topology reconstruction is performed under global constraints, implicit conflicts are eliminated by combining minimum conflict path search and local consistency backtracking, and the event sequence rearrangement basis is formed; finally, event state convergence judgment and semantic fusion are performed, and the final event stream with structured and continuous semantics is output; the application can effectively solve the multi-source event time sequence disorder and semantic conflict problem, and improve the real-time performance and accuracy of sports event analysis.