125 results about "Similarity distance" patented technology

The invention discloses a log template post-processing method and system based on semantic screening and similarity clustering. The method comprises the steps that log information is input into LLM to generate a log template; inputting the new log template and the cache log template into a semantic model, and respectively generating semantic embedding vectors; calculating a similarity distance between the new log template and the cache log template, and screening to obtain a candidate template; according to the generation confidence score, calculating the weighted structure similarity between the new log template and each candidate template, and clustering and combining the candidate templates exceeding a threshold value to generate a final template; and updating and replacing the final template into the cache structure. According to the technical scheme, the semantic similar templates can be recognized and combined more reasonably, efficiently and accurately, a semantic enhanced, controllable and stable template post-processing mechanism can be constructed on the premise that LLM calling burden is not increased, and automatic cache cleaning, template number control and structural consistency maintenance are achieved.

The invention discloses a power distribution network single-phase grounding reason diagnosis method based on confidence and ambiguity evaluation. The method comprises the following steps: acquiring original diagnostic data corresponding to a waveform to be detected; determining a confidence level corresponding to the original diagnosis data according to a preset grading rule based on the similarity distance corresponding to each of the plurality of neighbor sample waveforms in the original diagnosis data; determining a plurality of candidate reasons based on the single-phase grounding reasons corresponding to the plurality of neighbor sample waveforms; based on the number of the plurality of neighbor sample waveforms, determining occurrence frequencies corresponding to the plurality of candidate reasons; determining target ambiguity based on the occurrence frequencies corresponding to the candidate reasons and the reason label semantic association matrix; and based on the confidence level and the target ambiguity, adaptively generating a diagnosis result. According to the method, the technical problem that the information completeness of the conclusion output by the existing analysis method is insufficient, so that an objective judgment basis for the conclusion reliability is lacked when the operation and maintenance personnel make decisions is solved.

The invention discloses a remote sensing image aircraft fine-grained classification method based on multi-dimensional part feature modulation, and relates to the technical field of target detection and recognition. A self-supervised pre-trained ViT model is adopted as a backbone network, and remote sensing image set noise correction and dimension adjustment are input to generate a basic feature tensor containing part information; defining a key part prototype and a background prototype, calculating a similarity distance and adding noise, generating an attention map through convolution, calculating a local feature vector and a global feature vector, and splicing and reducing dimensions to obtain fusion features; and constructing a composite loss function, calculating a category score in combination with fusion features, converting the category score into probability distribution through Softmax, and selecting a category with the maximum probability as a final fine-grained classification result. Through multi-scale feature fusion and feature modulation of an aircraft target, the fine-grained classification precision can be effectively improved, and the limitation of poor performance under a complex background and a weak feature target is overcome.

The embodiment of the present invention provides a method and system for constructing a knowledge base in the field of power grid based on a vector database, which belongs to the field of emerging software and new information technology services. The method includes: obtaining a power grid data file to be stored; obtaining keywords in the power grid data file; using a preset vector database to process the keywords to obtain word vectors; determining similar word vectors of each word vector according to a preset similarity distance; word-processing the power grid data file; sequentially reading the power grid data file after word processing; processing the read word vectors or similar word vectors to obtain mapping values; counting according to the mapping values; using the counting result as the second feature of the power grid data file; clustering the power grid data file using the second feature and the keywords; index encoding the result of the clustering operation to obtain the power grid domain knowledge base.

The invention discloses a real-time multi-objective optimization decision-making method for shield tunneling parameters based on multi-mode NSGA-II (Non-dominated Sorting Genetic Algorithm-II) and LSTM (Long Short Term Memory). Comprising the following steps: acquiring a plurality of data sets, training a constructed shield tunneling parameter prediction model and a tool abrasion loss model, acquiring a tunneling parameter time sequence to be measured and a geological parameter time sequence, and inputting the tunneling parameter time sequence and the geological parameter time sequence into the shield tunneling parameter prediction model to obtain a tunneling parameter prediction value, inputting the tunneling parameter prediction value into a tool wear loss model to obtain a wear loss prediction value of each tool; constructing and solving a tunneling parameter multi-objective optimization model to obtain an optimal solution set of tunneling parameters; and performing minimum similar distance processing according to the optimal solution set, the tunneling parameter prediction value and the wear loss prediction value of each cutter, and taking the to-be-set tunneling parameter corresponding to the minimum similar distance value as the finally-set tunneling parameter at the current moment. According to the method, self-adaption to the geological condition of the tunneling stratum is achieved, the comprehensive tunneling performance of the shield is improved, and the total tunneling time and the construction cost in engineering are saved.

The invention discloses a carbon fiber prepreg appearance defect detection method and system based on deep learning, and relates to the technical field of image analysis, and the method comprises the steps: collecting and preprocessing a sample image of the surface of a carbon fiber prepreg; inputting the image into a pre-trained convolutional neural network feature extraction model, extracting middle / high-level features, performing up-sampling alignment, and performing fusion to form a multi-scale fusion feature map; constructing a feature memory library by using a fusion feature vector extracted from a defect-free sample through a convolutional neural network feature extraction model; obtaining a feature vector of a to-be-detected image according to the same process, performing similarity comparison with the memory bank, and calculating an abnormal score graph based on a similarity distance result; and generating an abnormal region mask image according to the feature distribution of the abnormal score image based on the abnormal judgment threshold value, and visually outputting a defect detection result. Through the technical scheme of the invention, the missed detection risk is reduced, defect large sample labeling is not needed, the cross-scene stability is improved, and the result expression is more intuitive.

The invention provides an interest point recommendation method and device for heterogeneous multi-graph representation learning, equipment and a medium, and the method comprises the steps: obtaining a set of users and POIs containing multi-dimensional situation information in a location social network, building a user-POI interaction graph, and extracting feature vector representation of the users and the POIs in an interaction space by adopting a neighborhood similarity measurement method; based on user sign-in time information and text description information, calculating a semantic information similarity distance and a time pattern similarity distance, constructing a user-user perception graph, and learning feature vector representation of the user in a social space; constructing a distance hypergraph by utilizing POI position information, and obtaining feature vector representation of the POI in a position space after dimension reduction clustering processing; and fusing each feature vector representation in the social space and the interaction space as well as the position space and the interaction space, extracting a user representation vector and a POI representation vector, and predicting user preferences by using a neural network model to realize POI recommendation.

The invention discloses a short-term power prediction method and system based on a wind power plant cluster. The method comprises the following steps: constructing a similarity distance matrix of a key meteorological element time-varying rule between different stations; calculating a Copula entropy between the historical power samples of each station, and constructing a spatial similarity matrix of station power according to each Copula entropy; performing normalization processing on the similarity distance matrix and the spatial similarity matrix, and summing the similarity distance matrix and the spatial similarity matrix after normalization processing to obtain a space-time similarity distance matrix; clustering the similarity distance matrix by adopting a hierarchical clustering algorithm, determining a partitioning scheme in combination with the geographic distribution of the new energy stations, and dividing each wind power plant in the target region into each region; and outputting a total power prediction result of the obtained region according to the LSTM time sequence prediction model. The overall precision and robustness of power prediction are improved, and a more reliable scheduling basis is provided for large-scale access of a new energy station to a power grid.

The invention relates to a spatial-temporal trajectory similarity query method based on an improved LCSS, and belongs to the field of channel traffic control. The invention aims to solve the defects of the traditional LCSS algorithm in the process of processing the inland ship trajectory data. According to the method, the spatial position and the course of the trajectory are comprehensively evaluated by introducing course information, so that the similarity between the trajectories is more comprehensively reflected, and the query precision is improved. Meanwhile, a distributed architecture is adopted, similarity distance calculation tasks are processed in parallel, and the query efficiency is effectively improved. In addition, the method supports the query of a short spatio-temporal trajectory segment and a specific short time period, and meets the requirements under different application scenarios. The method can be applied to multiple fields of route optimization, abnormal behavior monitoring, false target discrimination, port throughput statistical analysis and the like, and provides effective technical support for inland waterway traffic control.

The invention discloses a daily load curve clustering method based on a mixed distance improved FCM algorithm, and belongs to the technical field of data processing, and the method comprises the steps: collecting daily load data of a user at a sampling frequency of 5 minutes, carrying out the corresponding data preprocessing, analyzing the fluctuation characteristics of a daily load curve, and carrying out the clustering of the daily load curve; euclidean distance representing load curve power amplitude information and cosine distance representing load curve power fluctuation direction information are constructed respectively, and then an entropy weight method is adopted to construct a hybrid similarity distance (HSD) fusing the Euclidean distance and the cosine distance as a similarity criterion between daily power load curves; replacing the Euclidean distance in the FCM algorithm with the constructed HSD, then calculating a corresponding membership degree and a clustering center, iteratively updating the membership degree and the clustering center by using a target function based on an intra-class error weighted quadratic sum, and completing daily load curve clustering according to a preset maximum iteration number; and finally, determining the optimal clustering number and the corresponding optimal daily power load curve clustering result by using the sum of intra-class errors (SAE), thereby effectively solving the prominent problems of poor effect, low efficiency and the like of the traditional power load curve clustering method.

The application relates to the technical field of data management, in particular to an intelligent sports teaching data management method and system. The method comprises the following steps: collecting motion video stream data and inertial motion data of students; mapping a camera coordinate system of the video stream data and a body coordinate system of the inertial motion data to a world coordinate system; constructing a state space model based on an extended Kalman filter, and iteratively calculating to obtain a corrected student three-dimensional skeleton posture sequence; calculating an optimal matching path and a similarity distance between the student three-dimensional skeleton posture sequence and a standard action template sequence; and generating a quantitative action score according to the similarity distance. The application eliminates integral drift while ensuring the anti-occlusion capability by using an extended Kalman filter algorithm; meanwhile, an improved dynamic time warping algorithm is introduced, the standardization and objectivity of action quality scoring are realized, and the scientificity, safety and individualization level of sports teaching are effectively improved.

The invention discloses a spatial-temporal feature fusion concrete arch dam deformation prediction method based on multi-scale hierarchical clustering, and the method comprises the following steps: decomposing a dam deformation sequence into trend, period and fluctuation items, and constructing a new multi-scale similarity distance index; adopting a hierarchical clustering method based on a Ward criterion, and utilizing the constructed multi-scale similarity distance index to perform clustering partition on the deformation measuring points with the similar deformation rule; a deep learning model TCT-CANet fused with the spatial and temporal characteristics is constructed; the deformation of the concrete arch dam is predicted by utilizing the constructed TCT-CANet model, the prediction performance of the model is evaluated through various evaluation indexes, and meanwhile, the effect of each module in the aspects of spatial-temporal feature capture and prediction precision improvement is analyzed through an ablation experiment; the TCT-CANet model is compared with various time sequence prediction methods on all measurement points, and the superiority and generalization ability of the model are verified. According to the method, the prediction model TCT-CANet based on spatial-temporal feature fusion is constructed, and accurate prediction of deformation of the concrete arch dam is realized.

The invention discloses a small sample action recognition method based on multiple tasks and local parallel attention, which comprises the following steps: S1, initializing an SSL-LPA network model, modeling a human skeleton sequence into a space-time diagram, dividing the space-time diagram into a training set and a test set, and constructing a support set and a query set; s2, inputting a training sample, and executing a feature extraction task to extract a support sample and a query sample; s3, performing data enhancement on the support sample, and then executing a self-supervision auxiliary task; s4, executing a supervised classification task, expanding a sample space, and enhancing action uniqueness between a prototype vector and a query sample; s5, calculating the similarity distance between the prototype vector and the query sample to determine the category of the query sample; s6, the supervised classification task and the self-supervised auxiliary task have a synergistic effect, and the SSL-LPA network performance is optimized through mutual feedback; the robustness and generalization ability of the model are improved, and the ability of accurately identifying the action can be learned under the condition of limited label samples.

The invention relates to the field of power quality monitoring of a power system, in particular to a comprehensive power quality monitoring method and system based on an intelligent fusion terminal, and the method comprises the steps: obtaining and preprocessing power quality data of the intelligent fusion terminal, dividing electrical partitions, determining dominant harmonic components, and calculating the electrical distance between terminals in the same partition. And based on density peak clustering, combining feature difference and electrical distance weight fusion to obtain a similarity distance, determining an adaptive cutoff distance by using a low-density region adjustment factor, and drawing a decision diagram to identify a clustering center. Screening clustering results, analyzing electric energy quality characteristic dominance, identifying a mixed cluster and re-clustering, and positioning a disturbance source by using an attenuation law in combination with a significant clustering result and an electrical distance. According to the method, the accuracy of disturbance source positioning is improved through adaptive clustering and mixed cluster analysis, and the operation condition of the power grid and multiple disturbance sources can be accurately identified through the adaptive truncation distance and the adjustment factor of the low-density region.

A method, computer program product, and computing system for transforming a plurality of content portions into a plurality of embeddings using a language model. A graph is generated with nodes representing respective embeddings and an edge between a pair of nodes representing a similarity distance between the respective embeddings that is less than or equal to a predefined threshold. A category prediction is generated for each content portion by processing the graph using a graph neural network. A loss function is determined using a plurality of predefined categories and the category predicted for each content portion. The language model and the graph neural network are finetuned for automatically tagging content portions with a category by maximizing the loss function.

Disclosed in the present application are a result generation optimization method, apparatus, device and system for a large language model. The method comprises: concurrently sending a preset prompt word to a plurality of preset large language models, so as to obtain a plurality of generation results; and performing comprehensive optimization processing on the plurality of generation results on the basis of a preset comparison selection algorithm or a preset merging algorithm, so as to obtain an optimization result, wherein the preset comparison selection algorithm is constructed on the basis of similar-distance analysis, and the preset merging algorithm is realized on the basis of merging the generation results. A plurality of preset large language models are used to simultaneously perform result generation on the basis of the same preset prompt word, and these generation results are then comprehensively optimized to obtain an optimized result; and during the process, optimization can be performed by means of selecting a preset comparison selection algorithm or a preset merging algorithm on the basis of the actual situations, so that it can be ensured that the optimized result is more accurate and reliable. In this way, the present application can solve the technical problem in the prior art of it not being possible to meet actual application requirements due to a result generated on the basis of a single large language model lacking accuracy and reliability.

The invention discloses a text matching method and device based on probability distribution, equipment and a storage medium, and relates to the technical field of natural language processing, and the method comprises the steps: obtaining semantic feature distribution of each professional knowledge text from a professional knowledge base, obtaining a knowledge probability distribution set, abstracting the intrinsic features of the text through probability distribution, and obtaining a knowledge probability distribution set; and standardized representation of the knowledge base is realized. The text input by the user is obtained, the corresponding user text probability distribution is calculated, and the fault tolerance of text matching is remarkably improved. The similarity distance between the user text probability distribution and each distribution in the knowledge probability distribution set is calculated, and the overall semantic similarity instead of local matching is measured through the distribution distance, so that the comparison process can tolerate the distribution offset. Finally, a text matching result is determined according to the minimum value of the similarity distance, the effect of stably retrieving semantic related professional knowledge under high noise is achieved, and fault tolerance and matching precision are remarkably improved.

A collective data leakage attacks on a Retrieval-Augmented Generation (RAG) application for a generative artificial intelligence (GenAI) application is detected and prevented based on an analysis of incoming queries to distinguish normal and potentially malicious querying behavior. Similarity distances associated with each query are determined, such as the distances between the query vector embeddings and the nearest neighbors in the vector embedding space, distances between each query vector embedding and other query vector embeddings for other queries from the same user, or distances between each query vector embedding and the nearest neighbor cluster of vector embeddings in the vector space. A data leakage attack on the RAG application may be determined based on the similarity distances associated with each query. In response to the identification of a potential data leakage attack, the release of data from the RAG application may be halted.

The application provides a method, device and equipment for identifying a relationship between trajectories. The method of the application connects the starting points of two target trajectories and the ending points of the two target trajectories to form a closed trajectory line containing the two target trajectories. The total area and total perimeter of the closed area surrounded by the closed trajectory line are determined, the similarity distance between the two target trajectories is determined according to the ratio of the total area to the total perimeter, the average time difference between the two target trajectories is determined according to the time stamps of the trajectory points in the trajectory point sequences of the two target trajectories, and whether the two target trajectories have a relationship is determined according to the similarity distance and the average time difference, the spatial similarity and the time similarity. The method can accurately identify whether the two trajectories have a relationship, has low computational complexity, and improves the efficiency of the relationship analysis.

The invention relates to the technical field of data management, in particular to an intelligent physical education data management method and system. The method comprises the following steps: collecting motion video stream data and inertial motion data of a student; uniformly mapping a camera coordinate system of the video stream data and a body coordinate system of the inertial motion data to a world coordinate system; a state space model based on extended Kalman filtering is constructed, and a corrected student three-dimensional skeleton attitude sequence is obtained through iterative calculation; calculating an optimal matching path and a similarity distance between the three-dimensional skeleton posture sequence of the student and the standard action template sequence; and according to the similarity distance, generating a quantized action score. According to the invention, by using the extended Kalman filtering algorithm, integral drift is eliminated while the anti-shielding capability is ensured; meanwhile, an improved dynamic time warping algorithm is introduced, standardized and objective scoring of the motion quality is achieved, and the scientificity, safety and individuation level of physical education are effectively improved.

The present disclosure relates to a recognition processing method and device and a non-transient computer readable storage medium. The recognition processing method comprises: obtaining first thermal distribution data of an input image, and, according to the first thermal distribution data, calculating first color feature data of the input image; obtaining second thermal distribution data of each candidate image among multiple candidate images, and, according to each item of second thermal distribution data, respectively calculating second color feature data of a corresponding candidate image; calculating the color similarity distance between the first color feature data and each item of second color feature data respectively, and, according to each color similarity distance, determining the candidate image, among the multiple candidate images, that matches the color of the input image as an output image.

The invention relates to a construction method of an increased-class sample classification model and related equipment thereof. The method comprises the following steps: acquiring a trained basic sample classification model; acquiring an increased-class sample comprising increased-class sample feature data and an increased-class category label, and processing the increased-class sample feature data by adopting the basic feature representation model to obtain feature representation of an increased-class category prototype corresponding to the increased-class category label; on the basis of the similarity distance between the feature representations of the added category prototype and the plurality of basic category prototypes, calibrating the feature representation of the added category prototype to obtain the calibrated feature representation of the added category prototype; and adding the calibration feature representation of the increased class prototype as the feature representation of the increased class prototype of the increased class label into a basic classifier model of the basic sample classification model to obtain an increased class sample classification model. Through the method and the device, the problems of disastrous forgetting and over-fitting of an increment small sample classification system in related technologies are solved.

To provide an EMC countermeasure presentation system capable of efficiently extracting and presenting an emission countermeasure case in another field and another model with high similarity (countermeasure effectiveness) in consideration of a device configuration of an electronic apparatus.SOLUTION: A similar configuration extraction unit 3 configured to extract, based on device configuration data 11 at the time of noise measurement, device configuration data at a short distance by configuration learning data 9 obtained by learning a graph model representing a device configuration when the noise countermeasure and reduction amount frequency data is acquired; and a countermeasure estimation unit 4 configured to estimate, from data obtained from the reduction countermeasure extraction unit and the similar configuration extraction unit, a recommended countermeasure content having a high degree of similarity of a device connection configuration and expected to have an excessive noise reduction effect.SELECTED DRAWING: Figure 1

The invention discloses a vector retrieval method and device based on a neighbor graph, and relates to the field of vector retrieve.The method comprises the steps that according to a vector data set, an integral function of non-centrality chi-square distribution is adopted, and the centrality degree of each vector in the whole vector data set is determined; according to the centrality degrees of all the vectors, a centrality degree set # imgabs0 # is obtained; according to the similarity distance between the vectors in the vector data set and # imgabs 1 #, adopting a gradual vector adding method to construct a neighbor graph # imgabs 2 #; performing vector central degree traversal calculation on the # imgabs3 #; based on the central degree set # imgabs4 #, performing tuning and edge cutting operation on the neighbor graph to obtain a final result neighbor graph # imgabs5 #; and searching on # imgabs6 # by taking the feature vector of the query object as a query vector to obtain a vector retrieval result. According to the method, the precision and efficiency of vector retrieval based on the neighbor graph can be improved.

The application provides a point of interest recommendation method and device based on heterogeneous multi-graph representation learning, equipment and medium, the method comprises the following steps: obtaining a user and a POI set containing multi-dimensional context information in a location social network, and establishing a user-POI interaction graph; using a neighborhood similarity measurement method to extract feature vector representations of the user and the POI in the interaction space; based on user check-in time information and text description information, calculating semantic information similarity distance and time pattern similarity distance, constructing a user-user perception graph, and learning feature vector representations of the user in the social space; using POI location information to construct a distance hypergraph, obtaining feature vector representations of the POI in the location space after dimension reduction clustering processing; fusing the feature vector representations in the social space and the interaction space, and the location space and the interaction space, extracting user representation vectors and POI representation vectors, and using a neural network model to predict user preferences to realize POI recommendation.

The invention discloses a logging rock facies clustering method, system and device, a medium and a program product, and relates to the field of geological rock facies clustering, the method comprises the following steps: obtaining a sampling sequence formed by a plurality of sampling segments of each well, each sampling segment comprising a plurality of logging channel items; carrying out image conversion on the logging channel item of each sampling segment by utilizing a Grubm angle field function to obtain a three-dimensional tensor image of each sampling segment; extracting a first depth feature of the three-dimensional tensor image of each sampling segment, and performing linear transformation on the first depth feature according to a preset feature mapping vector to obtain a second depth feature of each sampling segment; calculating a cosine similarity between any two sampling segments according to the second depth feature, constructing a similarity distance matrix between the two sampling segments according to the cosine similarity, and aggregating the similarity distance matrix to obtain a clustering tree; and clustering the logging curve of the to-be-logged well based on the clustering tree to obtain a clustering result of the logging lithofacies.

This invention proposes a method for retrieving standard questions in a digital human system, comprising the following sub-steps: clustering all standard questions in the system to obtain the centroids of each category and the maximum Euclidean distance within each category; inputting a user question, calculating the distance between the word vector of the user question and the centroids of each category to obtain similarity distances, and constructing a similarity set; comparing each similarity distance in the similarity set with the maximum Euclidean distance within the corresponding category, constructing an attribution model, and calculating and determining the final category to which the user question belongs; calculating the similarity between the word vector of the user question and all standard questions in the final category, selecting the standard question with the highest similarity, and outputting the answer corresponding to that standard question as the result, thus providing the user with the answer. This method significantly reduces the workload of similarity calculation and improves the speed of standard question retrieval while maintaining the accuracy of the results.

The present invention discloses a method for anti-occlusion pedestrian re-identification with posture-guided attention, which respectively extracts the joint point heat map, joint point confidence and global feature map of the pedestrian image, multiplies the global feature map and the joint point heat map element by element, obtains the joint point local feature map, then passes the joint point local feature map and the global feature map through the posture-guided spatial attention module to obtain the spatial attention feature map; then passes the joint point local feature map and the spatial attention feature map through the posture-guided channel attention module to obtain the channel attention feature map, calculates the similarity distance between the pedestrian image to be retrieved and the gallery image, and selects the most similar gallery image as the final recognition result. The present invention overcomes the singularity problem of simply using global features and local features, as well as the problem of mismatch between features, and improves the accuracy of re-identification of occluded pedestrians.

The invention provides a method and a device for screening lesions by using a computer, and a medium. Obtaining a training sample set of the medical images with lesion category labels and a category similarity matrix representing the similarity between the categories; in the deep learning network training process, intra-class and inter-class similarity distances of lesion classes are obtained based on an output feature matrix of a convolutional layer, a first loss function is obtained by combining corresponding elements in a class similarity matrix, the intra-class similarity distance is in positive correlation with the first loss function, and the inter-class similarity distance is in negative correlation with the first loss function; obtaining a second loss function based on the probability matrix of the category to which the training sample output by the deep learning network belongs and the lesion category label; training the deep learning network based on the first loss function and the second loss function; and predicting the lesion category of the medical image by using the trained deep learning network. And the prediction accuracy is improved through lesion category similarity priori knowledge.

The present application belongs to the field of SA radar fault, and relates to a kind of extraction method of SA radar fault alarm code based on hierarchical clustering, comprising the following steps: S1 calculates the similarity distance between alarm codes;S2 calculates the distance between alarm code clusters;S3 alarm code hierarchical clustering process;S4 extracts key alarm code.The present application solves how to distinguish key alarm code and invalid alarm code for SA radar fault, takes key alarm code and invalid alarm code as two kinds of alarm codes, realizes the extraction of key alarm code by distinguishing the difference between the two kinds of alarm codes when fault occurs, and realizes the distinction between key alarm code and invalid alarm code by clustering alarm codes in the form of tree through hierarchical clustering method.