418 results about "Hierarchical clustering" patented technology

The invention belongs to the field of digital delivery, and particularly relates to a digital delivery topology mapping method and system for multi-source real-time data fusion, and the method comprises the steps: obtaining factory building distribution, equipment distribution and operation control logic and preset function block operation logic, and constructing a hierarchical clustering function mapping space through combining an association analysis and clustering algorithm; in response to a target function demand, obtaining a layered response mapping path in combination with a deep search algorithm; layered synchronous response and distributed node anomaly monitoring are realized based on the path, the three-dimensional simulation model and the display system equipment performance and the network state. Tracing abnormities based on a monitoring result in combination with a hidden Markov algorithm and a forward reasoning model, performing iterative verification after conflict resolution until the function is free of abnormities, and updating a mapping space; and adjusting the demand repeating steps to obtain a complete and updated mapping space, and realizing accurate function and picture collaboration under multi-source data fusion.

The invention relates to the field of 3D rendering, in particular to an intelligent scenic area three-dimensional image rendering method, which introduces a differentiable discrete decision into 3D fusion, supports end-to-end learning of a k value, performs discrete-continuous optimization based on an activation function, predicts an optimal k value of each voxel, introduces feature adaptive fusion based on dynamic neighborhood bidirectional retrieval, and realizes the 3D image rendering of the scenic area. The alignment of the color image and the point cloud is enhanced, the false detection rate of a small target is reduced, the detail reconstruction capability of a large target is improved, and the comprehensive rendering capability of a scenic spot is improved; according to the method, a lightweight grid is adopted to express a scenic spot subject, residual Gaussian is introduced to supplement high-frequency detail features, the number of Gaussian is reduced, rendering efficiency and capability are improved, textures are generated based on initial rendering back projection, fuzzy view angle dependence is avoided, hierarchical clustering and contour extraction from bottom to top are adopted on the basis, and the method is more efficient and efficient. The point cloud vertical structure change is dynamically detected, the point cloud is complemented, the accurate contour is extracted, the number of grid vertexes is reduced, and the rendering integrity is improved.

The invention relates to the technical field of artificial intelligence, in particular to a self-organizing knowledge base construction method, system and device based on a multi-agent system and a storage medium, and the method comprises the steps: S1, monitoring a dialogue information flow, calculating the self-reply confidence coefficient of an agent, and if the self-reply confidence coefficient exceeds a preset dynamic threshold value, judging that a potential knowledge value exists and triggering an extraction request; s2, in response to the request, performing semantic coding on a dialogue fragment, extracting knowledge elements, and generating a structured knowledge unit by utilizing graph neural network modeling and relation calibration; s3, through a pre-training semantic coding model, mapping the multi-dimensional semantic coding model into a multi-dimensional semantic vector; s4, carrying out hierarchical clustering comparison and combination, and if a threshold value is exceeded, establishing a new classification and adjusting a knowledge base index; and S5, knowledge fingerprints are generated and compared, and fusion updating is carried out based on the reputation scoring model when semantics conflict or redundancy occurs. According to the method, unstructured dialogue high-precision knowledge extraction is realized, so that the knowledge base structure is dynamically self-organized according to semantics, multi-source knowledge is coordinated, and the knowledge management automation level and quality are improved.

The invention discloses a high-dimensional time series data cluster structure prediction method based on dynamic hierarchical clustering and LSTM fusion. The method comprises the following steps: receiving a high-dimensional heterogeneous time series data set; performing clustering processing on the high-dimensional heterogeneous time series data set to obtain a cluster structure and a centroid matrix at the current moment; based on the cluster structure and the centroid matrix, time sequence features are extracted from three dimensions of a centroid track, a topological structure and scale dynamic, and a multi-channel time sequence feature tensor is generated through coding; and inputting the multi-channel feature tensor into the LSTM, and outputting a prediction result of the centroid offset and the topology change probability at the next moment by applying an attention gating mechanism. According to the method, the system and the equipment provided by the invention, efficient clustering and accurate prediction of high-dimensional heterogeneous time series data are realized through deep fusion of dynamic hierarchical clustering and LSTM, and the timeliness, the accuracy and the calculation efficiency of time series data processing are remarkably improved.

The invention discloses a high-precision distributed photovoltaic output prediction method and system based on micrometeorology and feature hierarchical clustering, and relates to the technical field of distributed photovoltaic output prediction.The method comprises the steps that numerical weather forecast data are collected, an initial micrometeorological field is generated through space-time alignment and self-adaptive KNN interpolation, and the initial micrometeorological field is subjected to feature clustering; a WRF-LES system and a bidirectional LSTM are combined to establish cross-scale mapping, a dynamic residual correction field is fused to generate hectometer-level high-resolution micrometeorological data, and the problem of insufficient resolution of traditional numerical forecasting is solved. MIC and PA-DTW are used for jointly analyzing the characteristics of the power station, and dynamic clustering is achieved through a sliding time window and incremental spectral clustering. According to the method, a physical information graph network and causal expansion convolution are coupled to extract features, federal learning cross-power-station cooperative training is combined, the distributed photovoltaic output prediction precision and robustness are improved, and privacy security is considered.

The invention relates to the technical field of health management of old people, and discloses a method and a system for predicting multiple diseases of old people. The method comprises the following steps: acquiring multi-modal health monitoring data of a target old person in a preset time period, wherein the multi-modal health monitoring data comprises physiological index time sequence data, a medication record sequence and a daily activity ability evaluation result; and performing cross-modal correlation analysis on the multi-modal health monitoring data to generate a disease interaction characteristic matrix containing metabolic disease correlation degree, circulatory system disease coordination index and neurodegenerative disease progress rate. And performing hierarchical clustering processing on the feature matrix by adopting a dynamic weight distribution algorithm, and outputting potential common disease combinations of the target old people and a priority score of each common disease combination. And generating a personalized intervention strategy set including a drug interaction avoidance scheme, a rehabilitation training intensity adjustment scheme and a nutrition intake ratio scheme according to the priority score.

The invention relates to the technical field of underground coal mine safety, and discloses an underground coal mine early warning system and method. The method comprises the following steps: collecting multidimensional monitoring data of an underground coal mine environment; performing dynamic window division on the multi-dimensional monitoring data to generate a data fragment set; performing abnormal mode recognition based on the data fragment set, and extracting abnormal feature points exceeding a dynamic threshold in each data fragment; inputting the abnormal feature points into a pre-constructed space-time correlation model, and analyzing diffusion paths of the abnormal feature points in a time dimension and a space dimension; according to the diffusion path, generating an abnormal propagation map containing the propagation direction and the intensity attenuation gradient of the abnormal feature points; hierarchical clustering is carried out on the abnormal propagation atlas, and abnormal clusters with similar propagation characteristics are divided; generating an early warning signal containing a priority label by using the spatial and temporal distribution characteristics of the abnormal cluster; performing synchronous verification on the early warning signal and underground equipment state data, and screening out a target early warning signal needing to be responded; and triggering a corresponding emergency control instruction according to the target early warning signal.

Disclosed in the present invention is an unmanned aerial vehicle flight trajectory anomaly tracing method based on a fusion clustering algorithm. A hierarchical clustering algorithm is applied to classify and organize complex and sensitive spatio-temporal data features to identify a potential anomaly point; then, weights of various data features obtained by using a Lasso regression algorithm are used to calculate the sum of products of the potential anomaly point and the weights of the various data features; a system detects an anomaly point in a flight trajectory of an unmanned aerial vehicle to ensure flight safety and stability; the anomaly point is confirmed; and the system instantly calls an improved KD tree algorithm to quickly notify nearest on-duty personnel to go to the site, thereby comprehensively inspecting and thoroughly verifying factors interfering with the normal flight of the unmanned aerial vehicle, quickly and accurately processing anomaly conditions of the flight of the unmanned aerial vehicle, ensuring the flight safety and data integrity of the unmanned aerial vehicle, and providing reliable guarantee and support for a flight system of the unmanned aerial vehicle.

The invention discloses an internal medicine drug curative effect evaluation method based on big data analysis of an electronic health record, and the method comprises the steps: extracting basic health data, diagnosis and treatment time sequence data and drug intervention data from the electronic health record, and carrying out the time-space alignment to generate a dynamic feature set; subgroups are obtained based on disease typing standard hierarchical clustering, and historical data and real world data are fused through transfer learning to construct a subgroup curative effect reference matrix; collecting data after medication in real time, and generating an evaluation vector containing short-term physiological response, middle-term symptom improvement and long-term prognosis risk through deep learning; dynamically matching the evaluation vector with the reference matrix, and introducing an individual weight coefficient to correct deviation; taking the deviation correction value as input, constructing a self-adaptive evaluation model through reinforcement learning, and performing iterative optimization; and generating an individualized report containing the curative effect level, the medication suggestion and the risk early warning, and quantifying the curative effect level through a fuzzy comprehensive evaluation method. According to the method, individual differences are accurately captured, full-cycle dynamic evaluation is realized, and the curative effect evaluation accuracy and the clinical decision-making efficiency are improved.

The invention discloses a layered densification Gaussian sputtering method based on visibility, and relates to the technical field of artificial intelligence and computer vision. According to the scheme, an initial three-dimensional Gaussian primitive set is generated based on sparse multi-view observation data, and initial scene representation is established by extracting spatial distribution parameters, morphological parameters and radiation parameters; performing fusion analysis on the Gaussian primitives based on the multi-dimensional observability parameter set to obtain comprehensive observability index data; executing hierarchical clustering according to the index data, and constructing a multi-layer Gaussian structure of a significant layer, a transition layer and a background layer; performing density enhancement, geometric continuity constraint and parameter update processing on different levels of Gaussian structures, and generating a rendered image of a target view angle under a volume light traveling and transparency hybrid mechanism; according to the method, continuous reconstruction of a scene structure and accurate expression of radiation characteristics can be realized under the sparse view condition, and the geometric fidelity and rendering consistency of new view angle synthesis are improved.

The invention discloses a single-molecule conductance signal semantic segmentation method based on multi-domain feature fusion, and relates to the field of single-molecule electric transport data analysis. The core of the method is a deep learning framework, and the deep learning framework comprises a time domain-frequency domain double-flow encoder, an attention mechanism module, a projection layer, a hierarchical clustering module and a segmentation head. The method specifically comprises the following steps of: performing data enhancement on time domain and frequency domain information of an input sample, generating views which are related to semantics and have different forms, and inputting the views into corresponding encoders; aligning two-modal coding features by adopting an attention mechanism; inputting the projected original time domain features into a hierarchical clustering module to generate a high-quality pseudo-tag, and taking the high-quality pseudo-tag as a supervision signal training segmentation head; and performing joint optimization by taking the weighted sum of the comparison loss and the segmentation loss as a total training target. The evaluation indexes comprise the accuracy rate, the F1 score and the MIoU. The method is applied to single-molecule electrical transport signal analysis, and provides effective data method support for single-molecule electronics basic research and application research.

The invention discloses an intelligent agent system and a control method thereof, and relates to the technical field of natural language processing. The system comprises an execution layer, a planning layer and an auditing layer, the planning layer identifies and extracts innovation task meta-information, constructs a task tree to complete operations such as hierarchical clustering, and determines a task execution path; the execution layer calculates the semantic similarity between the content of the knowledge graph and the user innovation question, determines the optimal question and answer, extracts heuristic information, and decomposes the heuristic information into sub-questions to construct a dependency graph; generating an answer set based on the graph solving sub-problems, and integrating answers by using a genetic algorithm to obtain an overall solution; and the auditing layer performs multi-dimensional scoring on the scheme, and determines a structured scheme report and optimization suggestions for users to use according to a scoring result. The method has the capabilities of structured reasoning, problem recursive decomposition and scheme closed-loop optimization, can generate a multi-dimensional evaluation result, and efficiently realizes systematic modeling of a clear path for analogy heuristic information support problem solution.

The invention discloses a highway tunnel surrounding rock deformation prediction method, and relates to the technical field of geotechnical engineering monitoring, and the method comprises the steps: collecting existing data, determining a monitoring arrangement scheme, and collecting on-site monitoring data through a monitoring data real-time feedback technology; processing multiple types of monitoring data and screening proper input features; clustering analysis is carried out on the monitoring data of the multiple sections, and classification of the excavated sections is completed; extracting the shape center of each section classification, and comparing the shape center with early-stage monitoring data of a predicted section to realize classification of the predicted section; and establishing a prediction model according to the classification of the prediction sections, and performing long-term prediction on the tunnel surrounding rock deformation response. According to the method, fine processing of the monitoring data is realized through feature screening and hierarchical clustering, and classification and classification of the known section and the predicted section are realized by fully utilizing multiple types of monitoring data, so that the long-term response of tunnel section surrounding rock deformation is predicted more accurately.

The invention discloses a double-platform point cloud-oriented bidirectional hierarchical clustering single tree segmentation method, which belongs to the technical field of forestry monitoring, and comprises the following steps of: firstly, preprocessing and filtering an original point cloud, and then vertically layering along a Z axis according to a set slice thickness; in each layer, a clustering radius is adaptively determined by using a local point density estimation result, point cloud clustering is carried out by using a region growing method based on a search radius, and an initial single wood structure unit is extracted; and then, space connection operation is executed between the upper layer and the lower layer, points with continuous structures are assigned to the same target based on triple constraints of adjacent distance, lateral offset and height continuity, and finally accurate segmentation of the complete single-tree point cloud is realized. The BLS method supports a top-down segmentation strategy and a bottom-up segmentation strategy, automatic adaptation can be carried out according to platform types, trunk point clouds are preferentially segmented in TLS data, and a canopy structure is preferentially constructed in ULS data.

The invention relates to the technical field of video behavior analysis, in particular to a post-examination illegal behavior detection method based on cross-mirror tracking and identity authentication, and the method comprises the steps: collecting monitoring videos of all cameras in an examination scene; carrying out identity authentication by adopting Reid identity matching, and forming and maintaining a personnel identity mapping relation; determining a target bounding box, dynamically evaluating and enhancing the quality score of the target bounding box, constructing an optimized trajectory feature, and tracking a trajectory through hierarchical clustering to obtain a corresponding identity ID; a historical feature library is established and dynamically updated, identity conflict detection is executed, and an identity ID is bound or reset; inputting a GAN generator to synthesize an enhanced spatiotemporal feature sequence, inputting the enhanced spatiotemporal feature sequence into a Bi-LSTM discriminator in combination with an image block sequence, executing adversarial discrimination and action classification, and identifying illegal behaviors; when the illegal behavior is identified, real-name system alarm information including an identity ID and an illegal behavior type is generated; therefore, the automatic discovery and real-name traceability of illegal behaviors after examination can be improved.

The invention discloses a topology identification method based on intelligent measurement data of a low-voltage transformer area. According to the technical scheme, the topology identification method comprises the steps that 1, a master station issues a timing command to an intelligent measurement terminal and an electric meter; step 2, the master station issues a transformer area electric meter file to the intelligent measurement terminal, configures an acquisition task and issues the acquisition task to the intelligent measurement terminal; step 3, establishing an electric quantity sum linear regression model based on a Lasso algorithm by using daily frozen electric quantity data, and converting the identification problem of the transformer area user-transformer relation into regression coefficient solving; step 4, on the basis of user change relation identification, performing dimension reduction on historical voltage data by using a t-SNE algorithm to improve the identification efficiency of the model; and step 5, clustering the voltage data after dimension reduction by using agglomerated hierarchical clustering, identifying a user phase and a meter box to which the user belongs, and completing identification of a topological structure. According to the method, the power grid topology information of the transformer area can be automatically identified, the efficiency and the accuracy of topology identification are improved, and the method has far-reaching significance for construction of an intelligent power grid.

A method for enhancing summary reply capability of an intelligent question-answering system based on graph clustering comprises the following steps: step 1, acquiring text data of a document, and converting unstructured text data into a structured text information graph; step 2, performing hierarchical clustering on the text information atlas, and dividing the text information atlas into a plurality of different information communities; step 3, generating an information abstract for each information community, selecting an information abstract similar to a user question, and generating a candidate document set; and step 4, integrating the candidate document set to obtain a final summary reply. According to the method, firstly, the text information atlas and the graph clustering technology are creatively combined, and LLM optimization is supplemented, so that the summarization reply capability of the intelligent question-answering system is remarkably improved. Secondly, the deep semantic structure and the document content are organically combined, so that the recall quality, the information comprehensive efficiency and the reply continuity of the intelligent question-answering system are improved, and efficient and reliable technical support is provided for application of the intelligent question-answering system in a complex task scene.

The invention relates to the technical field of document intelligent analysis, and discloses a document content self-adaptive analysis method and system based on a large model. The method comprises the following steps: acquiring original data flow of a to-be-analyzed document, wherein the original data flow comprises a text coding sequence, a layout structure mark and a multimedia embedding feature; the data stream is input into a pre-trained multi-modal large model, and a document semantic graph structure, a concept-containing node set, a relation edge weight matrix and a cross-modal alignment index are generated through context sensing analysis; performing dynamic hierarchical clustering on the semantic graph structure to obtain a hierarchical topic tree containing core topic branches, secondary topic branches and leaf node association strength; extracting a document logic framework containing chapter division suggestions, key information positioning coordinates and a cross reference mapping table according to the topic tree; a result is generated based on an adaptive analysis strategy optimization framework, and the strategy adjusts clustering granularity and relation mining depth according to document type features.

The invention discloses a multi-target line spectrum feature distinguishing method based on Single-link hierarchical clustering. The multi-target line spectrum feature distinguishing method comprises the following steps: step 1, performing short-time frequency modulation Fourier transform; step 2, extracting line spectrum frequency modulation slope identifiable features; step 3, outliers are removed; and step 4, Single-link clustering analysis and result output are carried out. In order to solve the problem that multi-target line spectrum features are difficult to identify in an underwater acoustic complex environment, frequency modulation slope matrixes of different frequencies and different time are extracted through short-time frequency modulation Fourier transform and serve as line spectrum identifiable features, and a Single-link hierarchical clustering method is utilized to process different-frequency identifiable feature sets. Finally, a line spectrum set of different targets is obtained, and multi-target line spectrum feature resolution is realized. A data analysis result shows that the method can effectively distinguish line spectrum features from different targets, lays a foundation for subsequent extraction of target motion features and realization of multi-target positioning, and has good engineering practical value.

The invention relates to a three-dimensional point cloud registration method and device, electronic equipment and a storage medium. The method comprises the following steps: acquiring source point cloud data and target point cloud data; identifying the source point cloud data and the target point cloud data, and determining a target reference point; performing hierarchical clustering on the target reference points to obtain a reference point cluster; determining a global transformation matrix according to the first reference point cluster of the source point cloud data and the second reference point cluster of the target point cloud data; wherein the global transformation matrix represents a pose transformation relationship between the source point cloud data and the target point cloud data; and iterating the source point cloud data and the target point cloud data according to the global transformation matrix in combination with ICP fine registration until point cloud registration is completed. Thus, by extracting the target reference points and performing hierarchical clustering to obtain the reference point clusters, key local units with global structure significance can be extracted from the disordered point cloud, and compared with a traditional point-by-point matching point cloud registration method, redundant calculation and local structure misjudgment can be reduced.

A three-stage pipeline is used to create a data structure for efficiently producing grounded recommendations that guarantee that the recommended items are part of a set, D. In the first stage, each item in D is converted into a vector representation. In the second stage, a hierarchical clustering method is used to build a tree based on the vector representations. Each item is a leaf node of the tree. Each non-leaf node represents a group of items or a group of groups of items, and so on. In the third stage, an LLM is used to generate text that encapsulates the information of the group (or groups) of items represented by each node. The generated tree is recursively traversed to generate recommendations.

The invention relates to the technical field of three-dimensional modeling, in particular to an indoor space visual presentation system and method, and the system comprises a contour modeling module, an interference detection module, a path optimization module and a deformation feedback module. According to the method, a geometric structure is constructed through a non-uniform rational B-spline algorithm to realize automatic generation of a topological relation, a two-dimensional contour line is established in combination with a perspective projection converter to eliminate an artificial translation error, line segment intersection detection and hierarchical clustering analysis cooperatively improve interference identification precision, and collision prediction prejudges contour change based on a multi-directional displacement vector. A free deformation grid and a vertex shader are linked to realize surface dynamic updating, a closed-loop process from three-dimensional modeling to visual feedback is formed, the dynamic analysis capability of a spatial relationship is enhanced, automatic contour generation and real-time deformation compensation reduce manual intervention, a path optimization and deformation feedback iteration system is constructed, and the dynamic analysis capability of the spatial relationship is improved. And the visual interference response timeliness and the space adaptation accuracy are improved.

The invention discloses a Non-IID federated learning backdoor attack defense method and system and a medium, and relates to the field of federated learning and network security. In order to solve the problems that an existing method depends on model parameters and is easy to avoid by a malicious client, and adaptive Non-IID scenes are poor, truncated singular value decomposition is executed through the client to extract first p left singular vectors, and the first p left singular vectors are uploaded to a server; the server constructs a matrix based on left singular vector cosine similarity and performs hierarchical clustering, and calculates a client similarity score and an aggregation weight by using a zoom dot product attention mechanism in combination with a left singular vector of the clean reference data set; the server distributes a global model, the client uploads parameters after local training, and the server weights and aggregates the model parameters in the cluster according to the weight and iteratively optimizes the model parameters. According to the method, the malicious client is identified from the data essential features, the malicious proportion does not need to be preset, the good client contribution and the data privacy are guaranteed while the backdoor attack is inhibited, the method is suitable for a Non-IID scene, and the model robustness and the main task performance are improved.

The invention discloses a power plant node impedance matrix collaborative correction method based on distributed parallel computing, relates to the technical field of power plant node correction, and aims to solve the problem of poor correction effect caused by inaccurate node analysis in a power plant. Regions are divided in a hierarchical clustering mode based on topological relevance and other characteristics, dynamic optimization is carried out in combination with load balancing and other constraints, node resource accurate allocation tasks are matched, regional topological integrity is guaranteed, calculation loads are balanced, resource waste is reduced, key parameters are synchronized in time through a standardized boundary information exchange mechanism, and the reliability of the system is improved. Coupling errors caused by region segmentation are eliminated, local autonomy and boundary collaboration are achieved, local calculation pertinence is reserved, consistency of electrical characteristics among regions is ensured, global deviation is corrected by combining layered optimization, it is ensured that a matrix conforms to an electrical law, multi-reference verification and four-dimensional index evaluation are conducted, and it is ensured that results are accurate, stable and practical.

The invention provides a user consumption behavior multi-dimensional portrait analysis method and system based on a neural network, and relates to the technical field of data analysis, and the method comprises the steps: obtaining user historical consumption behavior data, and constructing a basic feature vector; key time sequence features are determined through a sub-sequence dynamic pruning algorithm and entropy value weighted mapping; extracting sequence features by adopting a bidirectional long-short-term memory network; constructing a multi-task adversarial feature extraction network to obtain scene invariant features; performing feature fusion to obtain multi-dimensional combined features; constructing a feature index tree to calculate user similarity; and hierarchical clustering is carried out to obtain a consumption behavior portrait. According to the invention, high-precision user portraits are realized, and scene adaptability and calculation efficiency are improved.

The invention belongs to the technical field of electronic search, and relates to a radio reconnaissance method and system based on multichannel parallel processing and intelligent clustering, and the method comprises the steps: carrying out the multi-phase filtering digital channelization processing of a broadband radio signal, so as to obtain adjacent channel sequences with overlapped frequency bands; performing square law detection and constant false alarm detection on each channel sequence to generate an initial pulse description word; establishing a candidate corresponding relationship in adjacent channels based on arrival time similarity and noise consistency, and determining a cross-channel merging relationship according to frequency range continuity and frequency-time relationship slope consistency to generate a complete pulse description word; performing hierarchical clustering sorting based on the arrival angle and the pulse width; and performing multi-stage differential analysis on a clustering result to extract a repetition frequency sequence and determine a repetition frequency type. According to the technical scheme, the stability and consistency of cross-channel pulse merging and sorting results can be kept under the broadband high-density condition, and the radio search processing precision under the complex electromagnetic environment is improved.

The invention provides a real-time image stabilization method and system for intelligent image processing, and relates to the technical field of image stabilization. The method comprises the following steps: synchronously obtaining continuous image frames under strong glare interference, and identifying a texture blurred region; for the region, calculating a texture complexity distribution diagram by combining the local structure variation of the adjacent frames, and screening candidate regions according to the texture complexity distribution diagram; polarization filtering processing is applied to the candidate area, the filtering angle is dynamically adjusted to be matched with the polarization direction of incident light, and a polarization filtering image is generated; extracting a candidate region motion vector field based on the gradient continuity feature of the image, and performing spatial superposition on the candidate region motion vector field and the texture complexity distribution map to form a motion vector set carrying texture weight factors; and according to the texture weight factor, carrying out hierarchical clustering to separate a background vector group and a foreground abnormal group, only selecting a geometric center value of the background group to execute inter-frame motion compensation, and outputting a stable image sequence. According to the invention, the stability and quality of images collected by the unmanned intelligent turntable in a strong glare dynamic scene are improved.

The invention belongs to the field of computer system structures and artificial intelligence accelerators, and discloses a software and hardware collaborative heterogeneous storage computing accelerator oriented to full-chip DLRM reasoning. A control domain composed of a static random access memory and an embedded dynamic random access memory, an on-chip storage domain, an in-memory computing domain and an on-chip computing domain are integrated in the same chip, an embedded table is compressed by introducing a hierarchical clustering quantization frame, and a dynamic storage management strategy of access frequency perception is combined. The embedded access path is shortened, and the bandwidth requirement is reduced. According to the method, the reasoning process of the whole set of DLRM can be independently completed, and the memory access overhead in a traditional terminal SoC is remarkably reduced by reducing cross-domain data migration between the calculation unit and the storage unit, so that the model reasoning performance and energy efficiency are improved.

The invention relates to the technical field of federated learning, in particular to a hierarchical clustering federated learning method based on SMC, and aims to solve the contradiction between privacy protection and model accuracy in federated learning. According to the method, an uncertainty evaluation mechanism is provided, the uncertainty of data classification is reduced by calculating the average classification score of multiple enhancement instances of each sample, the consistency of data distribution is judged by using stationary points on the aspect of a grouping strategy, and when the data distribution of clients is consistent, the data classification is carried out. The stationary point obtained by minimizing the local empirical risk function is both a global minimum point and a stationary point of a single client, and the gradient norm of the client approaches zero at the moment; on the contrary, when data distribution is inconsistent and the gradient norm is larger than zero, local model parameters are encrypted through a Paillier homomorphic encryption system, parameter aggregation under the non-decryption condition is achieved through a federated training security aggregation protocol (FTSA), and a multi-stage aggregation method is further provided for aggregating the model parameters step by step in a staged mode.

The invention discloses a defense method and device for defending federated learning backdoor attack and a medium, and the method comprises the steps: a client global model is randomly distributed to a client set, and is trained and updated based on client local data; after model update quantity is collected, hierarchical clustering is carried out based on update direction similarity; for clusters obtained by clustering in each direction in hierarchical clustering, extracting a corresponding L2 norm as an amplitude feature, determining the optimal number of sub-clusters by adopting a contour coefficient, and performing secondary clustering; for each clustered sub-cluster, taking the median of the updating amplitude of the updating quantity of a plurality of groups of models as a clipping threshold, calculating the scaling of each node, and scaling the sub-cluster with the amplitude exceeding the threshold; aggregating the clipped model update quantity, carrying out weighted average to generate a new generation of global model, and distributing the new generation of global model to a client group for iteration; and repeating the steps until the model converges or reaches a preset training round. According to the invention, high-precision identification and isolation of malicious clients are realized.