1066 results about "Feature aggregation" patented technology

The invention provides a dynamic routing selection method and system, electronic equipment and a storage medium, and aims to solve the problem that a routing strategy is difficult to adapt to a dynamically changing network, the method comprises the following steps: a terminal layer collects the state of a terminal and network data, and performs lightweight feature extraction; the edge node layer receives the data of the terminal layer, carries out space-time-semantic feature aggregation, and generates a region-level resource scheduling and routing decision strategy based on fragmented reinforcement learning; the central cloud service layer gathers whole network data, generates a global optimization strategy and issues the global optimization strategy; the edge node layer fuses global optimization and a region-level strategy, and executes dynamic routing selection; and security and privacy protection are provided through the trusted chain layer. According to the invention, adaptive path selection can be realized, the network resource utilization rate is improved, and the network stability is improved.

The invention relates to the field of power data management, and discloses a power big data adaptive management method and system fused with spatial-temporal feature mapping, and the method comprises the steps: obtaining a real-time operation data flow from a multi-source power terminal, and constructing an original power data set with a time sequence label and a device identifier; the method comprises the following steps: dividing an original power data set into parallel processing units based on a storage-while-computing architecture, performing dynamic index updating by adopting an event-driven index mapping rule, and constructing a multi-dimensional data index cache system with real-time responsiveness; identifying a key abnormal trajectory through a time-varying feature nesting mechanism, and performing hierarchical measurement and entropy disturbance analysis on a data fluctuation degree in the key abnormal trajectory by using a streaming feature aggregation network; identifying potential security risk nodes in combination with the structure matching degree between the historical abnormal event evolution graph and the key abnormal trajectory; and generating a multi-level response instruction chain based on the risk assessment result. The method has the advantage of improving the operation safety of the power grid.

The invention discloses a flowmeter-oriented embedded multi-parameter fusion calibration control method and system, relates to the technical field of industrial measurement and embedded control, and is used for solving the problem of unstable calibration caused by misalignment of multi-link flow signals, compensation lag and fusion oscillation. The method relates to synchronous acquisition, pretreatment and calibration of differential pressure type, ultrasonic wave and thermal type mass measurement links and environment parameters such as temperature, pressure and density. All-channel unified time scales are established through ternary binding, filtering and denoising, zero snapshot and alignment rules, a stability scoring mechanism is constructed in combination with offset behavior recognition, event classification and feature aggregation, and dynamic switching of a compensation path and a fusion strategy is driven. In a double-path fusion structure, table look-up compensation, model estimation, multi-algorithm parallelism and fusion track recording are adopted, version traceability and calibration result closed-loop control are achieved, and the stability and rechecking performance of mass flow output under complex disturbance are effectively improved.

The invention provides an ocean wind field prediction method based on a neural network, and belongs to the technical field of ocean wind field prediction.The method comprises the steps that sparse ocean observation data are collected, a spatial covariance matrix is established, the spatial covariance matrix is converted into a graph structure, and then multi-hop neighborhood feature aggregation is conducted through a graph convolutional network; a tensor decomposition algorithm is combined for modeling high-order feature interaction to generate a gridding wind field, a bidirectional long-short-term memory network encoder is used for extracting space-time invariant features, a multi-layer perceptron predictor is used for directly mapping a future multi-step wind field, and a course learning strategy and a Shenchang differential equation boundary layer are matched for correction. The technical problem that sparse ocean observation data are difficult to accurately reconstruct into a high-resolution gridding wind field is solved.

The invention discloses an indoor three-dimensional point cloud semantic segmentation method based on a super voxel Transform architecture, and belongs to the technical field of map making. The method comprises the following steps: acquiring point cloud data of different scenes, preprocessing the point cloud data, and constructing a training sample set; constructing a neural network for indoor three-dimensional point cloud semantic segmentation; training the neural network; and obtaining indoor three-dimensional point cloud data to be segmented, preprocessing the point cloud data, inputting the point cloud data into the trained neural network, outputting a super-voxel category probability and confidence, mapping a super-voxel label back to the original point cloud, and completing semantic segmentation of the indoor three-dimensional point cloud. According to the method, efficient dimension reduction and local feature aggregation of the point cloud are realized through a hierarchical structure of the super voxels, global semantic association is modeled in combination with a self-attention mechanism of Transform, semantic segmentation can be better performed on the indoor three-dimensional point cloud, and various indoor application requirements are met.

The invention discloses a medical image classification method and system based on multi-scale spatial state modeling, and the method comprises the steps: firstly dividing an input medical image into a plurality of non-overlapping image blocks, and mapping the non-overlapping image blocks to a feature space through a learnable linear projection layer to obtain an initial feature map; then, multiple layers of stacked MS-SMamba blocks are used for carrying out layer-by-layer feature extraction, each MS-SMamba block comprises a main branch, an auxiliary branch, a dynamic gating fusion network, a residual error connection unit and a feedforward network, and long-range dependency relation capture and multi-scale feature fusion are achieved; and finally, processing the last-layer output feature map through a global feature aggregation and classification module, generating a global feature vector, and outputting a classification result. According to the method, the capturing capability of complex pathological features in the medical image is improved, the calculation efficiency and clinical applicability are improved, and the method is suitable for scenes such as disease screening and auxiliary decision making in medical image diagnosis.

The invention discloses a low-altitude wind field prediction method and system based on a space-time diagram convolutional network, and relates to the technical field of weather forecast and wind energy utilization, and the method comprises the steps: collecting wind field observation data and physical field data of all nodes of a target region, a dynamic space-time diagram is constructed based on a flow function-vorticity theory through a dynamic diagram construction module; extracting spatial information through a graph attention network to obtain a spatial feature tensor; the spatial feature tensor and the physical field data are processed by a PhysFusion-TransTCN encoder to obtain the deep spatial and temporal features of the wind field; performing hierarchical feature aggregation on the wind field deep spatial-temporal features through an output module to obtain a wind field prediction result of the target area; a wind field physical mechanism is deeply fused, multi-scale spatial-temporal feature fusion is realized, and prediction result precision and physical rationality are ensured.

The invention discloses a weak supervision video anomaly detection method based on prompt learning knowledge enhancement, and belongs to the technical field of video intelligent analysis. A video side gives a section of abnormal scene video, video sequence features and audio sequence features are obtained through a feature extraction network, then a trained and complete feature aggregation network is input to carry out multi-modal feature aggregation, an abnormal score is obtained through a score prediction network, and text representation is carried out based on prompt learning. A prompt template is constructed for abnormal video tags through a knowledge graph, semantic expansion is performed on normal tags through a plurality of learnable parameters, cross-modal alignment is performed on the normal tags and a video side, so that features of the video side are close to different normal semantics, knowledge enhancement is performed by introducing external information, positive abnormal boundaries of the video are learned, and the detection performance is improved. And finally, multi-task joint optimization is carried out through different loss functions, and abnormal video clip positioning is carried out.

The invention discloses a target multi-attribute identification method based on feature decoupling and cross-task collaboration, and belongs to the technical field of computers of specific calculation models, and the method comprises the following steps: firstly, extracting the initial features of each task through a lightweight backbone network, carrying out feature decoupling in a subspace, and according to the cross-task feature similarity, carrying out feature extraction; according to the target multi-attribute identification method based on feature decoupling and cross-task collaboration, an orthogonal constraint weight is dynamically adjusted, then mutual information confrontation minimization is adopted to further suppress statistical dependence between tasks, task residual errors are injected in a cross-task feature aggregation stage, differentiation enhancement is achieved, and finally unified joint feature representation is formed. According to the method, subspace statistical independence is realized, independence and necessary collaborative information are considered, a stable basis is provided for subsequent fusion, statistical dependence between tasks is further suppressed through mutual information confrontation minimization, complementation information is reserved while independence is ensured, and feature discrimination and robustness are improved.

The invention relates to a data fusion method based on a multi-channel image acquisition card and related equipment, and the method comprises the following steps: synchronously receiving a multi-view target scene shot by a camera through the multi-channel image acquisition card, and obtaining original multi-channel image data; performing geometric correction on the original multi-channel image data to obtain a geometric correction image group, and performing space-time coordinate conversion based on the geometric correction image group to obtain a space-time registration image group; performing cross-scale feature aggregation on the space-time registration image group to obtain a multi-scale feature pyramid layer, and constructing a multi-scale feature pyramid based on the multi-scale feature pyramid layer; and carrying out adaptive weight convolution fusion on the multi-scale feature pyramid to generate a target fusion image, thereby solving the technical problems that most methods depend on a complex preprocessing process or have relatively strong hypothesis for a specific scene and are difficult to adapt to a dynamically changing real environment.

The invention discloses a re-parameterization unmanned aerial vehicle target detection method based on multi-core fusion and omnidirectional connection, and the method comprises the following steps: obtaining an unmanned aerial vehicle aerial image data set, and constructing a detection model comprising a backbone network, a feature fusion network and a deformable task decoupling detection head; the backbone network extracts multi-scale and multi-direction features by six parallel paths in a training stage through a wide-branch re-parameterization convolution module, and re-parameterization is carried out in a reasoning stage to obtain single-path convolution; the feature fusion network performs down-sampling through a spatial deep convolution module and reduces spatial information loss, and realizes cross-stage omnidirectional feature interaction and double attention enhancement in combination with an omnidirectional kernel cross-stage partial connection module; and the deformable task decoupling detection head decouples the classification and regression features, optimizes feature expression, weights the classification features and then performs aggregation decoding. The model is trained to be used for a test set to output a detection result, detection precision and reasoning efficiency can be balanced, and the model adapts to a complex aerial photography scene of an unmanned aerial vehicle.

A computer-implemented system for multidimensional generative value assessment and applicant selection, consisting of: a data collection unit configured to electronically receive applicant data consisting of structured academic records, work experience records, digital documentation, and unstructured narrative responses generated from generative self-assessment instruments and contextual interviews; a feature extraction unit coupled to the data acquisition unit, configured to apply computer-assisted text processing, semantic analysis, and token-level attribute identification to transform narrative responses and structured data into multidimensional feature vectors that represent generative indicators of innovation, mentoring, collaborative performance, resilience, social contribution, ethical consistency, and predicted institutional impact; a weighting calculation unit configured to assign weight values ​​to the extracted feature vectors based on a digital generative profile definition matrix that includes dimensions, sub-criteria, indicators, documentation requirements and importance coefficients, with the weighting being distributed across the generative dimensions defined in the digital matrix and configurable according to the institutional context; a quantitative rating unit configured to calculate a generative rating score by aggregating weighted feature vectors derived from self-assessment inputs, interview-based ratings, document analyses, and authenticity predictions, with the aggregation including normalization, nonlinearity correction, conflict handling, and artifact frequency balancing to obtain a consolidated score; a proof verification unit configured to electronically validate referenced digital evidence by performing content extraction, metadata verification, pattern matching, and cross-document correlation to determine authenticity, credibility, and contextual relevance with respect to the calculated feature vectors; a classification determination unit configured to assign a classification level to an applicant by comparing the generative assessment score with a set of system-defined calculation thresholds, including at least a lower threshold, a middle threshold and an upper threshold, the classification levels representing different generative maturity states and determining subsequent eligibility for selection; a decision generation unit configured to produce a digital output data set that includes classification level, feature aggregation summaries, evidence validation results, and recommended organizational actions, wherein the decision generation unit encodes the data set in a digitally signed, tamper-proof format and stores it on a non-volatile storage medium; and A system control unit acts as an operational interface to all other units and is configured to orchestrate data flow, scheduling, process state transitions, and event logging to ensure verifiable traceability, consistency, and auditability of the evaluation and selection processes.

The invention discloses a short-term wind speed prediction method for multiple offshore wind power plants, relates to the technical field of power system intellectualization, and constructs a dynamic graph structure fusing the correlation between geographic distance and wind speed according to the correlation between the geographic position of a target area and the wind speed, namely a multi-wind-plant connected graph, and represents the spatial topological relation of a wind power plant group. A wind power plant group is mapped into a node network by constructing a dynamic graph structure fusing geographic distance and wind speed correlation, spatial dependence intensity between nodes is quantized by using a weighted adjacent matrix, spectral domain convolution operation is carried out by a graph convolution network based on a normalized Laplacian matrix, and the spatial dependence intensity between nodes is quantized by using a normalized Laplacian matrix. Efficient neighborhood feature aggregation is achieved through Chebyshev polynomial approximation, complex spatial association caused by geographic position difference and meteorological condition interaction can be accurately captured, the defect of non-Euclidean spatial relationship modeling in a traditional method is overcome, and the representation capacity of the spatial dependency relationship in the multi-wind-power-plant environment is remarkably improved.

The invention discloses a mixed Mamb-Attention air quality prediction model based on multi-scale decomposition and a construction method of the mixed Mamb-Attention air quality prediction model. The method comprises the following steps: firstly, constructing a multi-scale decomposition module (MSD), decomposing an input time sequence into a trend term, a season term and a residual term through a parallel sliding window group, and realizing cross-scale feature fusion by utilizing group normalization and convolution; then designing a periodic pyramid module, extracting multi-level periodic features based on fast Fourier transform (FFT), and enhancing the perception ability of the model to different time scale periodic laws; the Mama branch is used for capturing long-range dependence, the self-attention branch is used for extracting a local dynamic mode, and the output of the Mama branch and the output of the self-attention branch are fused through residual connection and layer normalization; and finally, the prediction head module completes feature aggregation and result output. The model gives consideration to long sequence modeling capability and calculation efficiency, can accurately capture multi-scale dynamic change and non-stationary features in air quality data, improves the precision and stability of air quality prediction, and has good practical value and popularization prospect.

The invention discloses a crop disease and insect pest image recognition algorithm based on dynamic adaptive multispectral fusion Transform, and belongs to the crossing field of agricultural information technology and computer vision. The objective of the invention is to solve the problems of insufficient multi-spectral feature fusion, poor complex background adaptability and insufficient precision in traditional recognition. Acquiring pest and disease damage images of crops in different wave bands (visible light, near-infrared light and the like) to construct a data set; through a dynamic adaptive fusion module, spectral weight distribution is learned in real time based on an attention mechanism, weights are adjusted according to spectral response differences of disease and insect pest areas, and accurate feature aggregation is achieved; the fusion features are input into an improved Transform model, a self-attention mechanism of crop semantic priori knowledge is introduced, focusing of key features of diseases and insect pests is enhanced, and background interference is inhibited; and finally outputting the disease and pest category and confidence. According to the method, through dynamic fusion and Transform cooperation, the recognition accuracy and robustness in a complex scene are improved, support is provided for early warning and prevention of diseases and insect pests, and the application value is remarkable.

The invention discloses a multi-view construction personnel tracking method and system based on attention perception. The method comprises the following steps: giving synchronous images from S cameras, and inputting the synchronous images into an encoder for feature extraction to obtain a multi-view feature map; transforming the multi-view feature map into a unified aerial view space by using perspective projection, and aggregating features after projection transformation of all views by using a convolutional layer; inputting the aggregated aerial view features into a decoder for decoding; a cross attention module is introduced, bird's-eye view features of a current frame and an adjacent frame are processed through 3D position coding, instance tokens are extracted as query, keys and values, an affinity matrix is generated through CNN coding similarity, features are propagated through matrix multiplication, and the bird's-eye view features of the current frame are updated. According to the method, the multi-view feature map is projected to the aerial view to realize early fusion, and a cross-frame attention mechanism is introduced, so that the problem of appearance feature distortion caused by perspective transformation is solved.

The invention discloses an intelligent water quality monitoring and pollution source identification method based on full spectrum analysis, and relates to the technical field of environmental monitoring optical sensing, and the method comprises the following steps: carrying out dynamic background stripping on a water body pollution metadata set to obtain a dynamic background spectrum, and building a spectrum purification model to carry out characteristic purification on the dynamic background spectrum to obtain a water body pollution data set; generating a pollution characteristic spectrum; performing multi-dimensional feature aggregation and hydraulic drive topology reconstruction on the pollution feature spectrum to generate a watershed pollution dynamic propagation relationship network diagram; performing physically constrained space-time convolution and propagation inversion on the watershed pollution dynamic propagation relationship network diagram to generate a pollution source probability distribution thermodynamic diagram; and carrying out responsibility association mapping on space anchor point coordinates of the pollution characteristic spectrum and pollution source center positioning of the pollution source probability distribution thermodynamic diagram, and generating a visual traceability report. Through dynamic background stripping and feature purification, high-fidelity extraction of pollution features under a complex optical background is realized, and the detectability of trace pollutants is remarkably improved.

The invention relates to the technical field of deep learning, in particular to a photovoltaic cell defect detection method fusing multi-scale features and a re-parameterization strategy, and the method comprises the steps: obtaining a to-be-detected image; the image is input into a defect detection model, in the defect detection model, the backbone network is used for extracting a feature map of the image, and a first C3K2MSDA module is used for carrying out multi-scale cavity sliding window self-attention aggregation on the feature map to output a multi-scale feature map; the neck network is used for performing up-sampling on the multi-scale feature map and splicing the up-sampled feature map with the feature map from the backbone network to generate a fusion feature map, and performing multi-scale feature aggregation again by using a second C3K2MSDA module and an EMA-AFF module and generating a cross-scale fusion feature map; a DEC-Head module in the head network performs detection processing on the cross-scale fusion feature map, and outputs a defect category probability of each candidate region and a corresponding bounding box position as a detection result; and designing a loss function and optimizing the defect detection model by using the loss function. According to the method, the model defect detection performance can be improved.

The invention provides a high-altitude operation emergency scheduling method and system, and the method comprises the steps: firstly obtaining a real-time video data set of a high-altitude operation region, carrying out the preprocessing of the real-time video data set, and generating a standardized video data flow, and then calling a video analysis model to extract an operation environment feature set and an operator behavior feature set; and then constructing a graph structure based on the working environment feature set, generating an environment state feature vector through graph convolutional network multi-level feature aggregation, and generating a dynamic emergency scheduling strategy in combination with the environment state feature vector and the working personnel behavior feature set. And finally, generating an emergency scheduling instruction set after strategy verification, and transmitting the emergency scheduling instruction set to the operation terminal equipment to trigger emergency response operation, so that a targeted emergency scheduling strategy can be dynamically generated according to actual conditions, automation and intelligence of emergency scheduling are realized, and the emergency response efficiency is improved.

The invention discloses a bridge inclination and settlement monitoring method and system based on multi-sensor fusion, and belongs to the field of bridge structure monitoring, and the method comprises the steps: obtaining the data of a multi-source heterogeneous sensor; the sensor data is converted into space-time diagram data, and the space-time diagram data comprises the steps that each sensor is mapped into nodes in a diagram, edges between the nodes are defined according to the physical connection relation of the bridge structure, and multi-source heterogeneous sensor data are unified into dynamic feature vectors with the same dimension on the nodes through learnable feature mapping; inputting the time-space diagram data into a preset neural network model, performing spatial feature aggregation on the dynamic feature vector through a diagram attention mechanism, performing time feature extraction through a time convolutional network, and generating a hidden state vector fused with time-space information; and generating a monitoring state value of the bridge based on the hidden state vector. According to the invention, depth feature fusion of spatial perception is realized, and the sensitivity of anomaly recognition is improved.

The invention discloses an unmanned vehicle inspection small target detection method based on an efficient attention mechanism. The method comprises four stages of image feature extraction, vocabulary embedding extraction, efficient attention coding and cross attention decoding. And image feature extraction: performing feature extraction on the input image by using the backbone network to generate a multi-scale feature map. And vocabulary embedding extraction: generating vocabulary embedding in the defined category vocabulary through a CLIP text encoder. And efficient attention coding: performing deep feature interaction, space attention guidance and multi-scale feature aggregation processing on the multi-scale feature map of the picture to obtain an image feature map fusing the visual context and the multi-scale information. And cross attention decoding: embedding the aggregated feature map and vocabulary, and outputting a final detection result through cross attention fusion, IoU perception query and regional text comparison processing. Compared with the prior art, the method has the advantages of good prediction effect, good practicability and the like.

The invention discloses a spine image key point detection algorithm based on multi-task learning. The algorithm comprises the following steps: constructing a multi-task deep learning network model comprising a segmentation branch and a key point positioning branch; a feature aggregation module based on multi-scale cavity convolution is embedded in the segmented branches, and the multi-scale cone feature extraction capability of the model is enhanced through splicing fusion of multiple cavity rate convolution branches and global pooling branches; a cross-task attention fusion module is introduced between the two branches, and bidirectional dynamic interaction and complementation between segmentation features and key point features are realized by generating and fusing first-order and second-order context attention maps; semantic alignment loss is designed in a training stage, collaborative optimization of two tasks is promoted by constraining the consistency of segmentation masks and key point heat maps in a high-level feature space, global information of segmentation and local information of key point detection are fully utilized, and the accuracy and stability of spine centrum key point positioning and the accuracy of a segmentation result are improved.

The invention relates to the technical field of food quality detection, in particular to a tea oil authenticity identification method based on an in-situ mass spectrum-machine learning coupling technology. Comprising the following steps: S1, preparing a sample; s2, acquiring mass spectrum data; s3, feature extraction: performing dynamic binning processing on the mass spectrum data, extracting feature peak information, and performing time dimension feature aggregation to form an initial feature matrix; s4, data preprocessing; s5, training a machine learning model; and S6, intelligent identification and visual output: inputting the mass spectrum original data of the tea oil sample to be detected into the trained multi-model integrated system, outputting an identification result through weighted voting or an optimal model strategy, and giving a confidence interval to quantify the reliability of the result.

The invention discloses an echocardiogram video classification method, system and terminal, and the method comprises the steps: constructing a classification model network which comprises a feature extraction module, a feature enhancement module and a feature aggregation module; obtaining an echocardiogram video, obtaining a plurality of standard section views according to the echocardiogram video, performing interpolation processing and feature extraction on the plurality of standard section views through a feature extraction module, and outputting a plurality of video features; inputting the plurality of video features into a feature enhancement module for aggregation enhancement of spatial features and time sequence features, and outputting a plurality of enhanced features; and inputting the plurality of enhanced features into a feature aggregation module for frame-level feature weighted fusion to obtain a plurality of key frame features, selecting related features from the plurality of key frame features, obtaining fusion features according to the related features, and classifying the fusion features to obtain a classification result of the echocardiogram video. According to the method, the classification accuracy of the echocardiogram videos is effectively improved.

The invention provides a plant insect disease image identification method and system based on deep learning, and belongs to the technical field of deep learning, and the method comprises the steps: firstly obtaining a multi-source image set of a plurality of growth stages of a plant, including the whole, leaf and stem images of the plant, the growth stage time, and the shooting angle identification information; constructing a plant image time sequence association network to perform organ-level time sequence alignment processing to obtain an aligned image set and an organ association coefficient, performing cross-growth-stage focus feature association through the association coefficient to obtain a cross-stage focus association set and time sequence evolution information, and performing cross-growth-stage focus feature association; and inputting the association set into an exclusive deep learning model for feature aggregation and type matching to obtain a preliminary result, and finally performing organ adaptation verification on the preliminary result in combination with time sequence evolution information to generate a final identification result and a focus organ time sequence evolution atlas, thereby improving the accuracy and reliability of plant insect disease identification.

The invention discloses an image semantic segmentation method based on a graph convolutional network, and the method comprises the following steps: obtaining original image data, and carrying out the preprocessing; performing feature extraction to generate a multi-scale feature map; dividing regions based on the multi-scale feature graph, defining region units as graph nodes, and constructing a node feature set; constructing an initial adjacency matrix according to the spatial proximity relation and the feature similarity of the region units; carrying out self-adaptive updating on the initial adjacent matrix, and correcting an edge weight based on semantic similarity and spatial connectivity between nodes; inputting the adaptive adjacency matrix and the node feature set into a graph convolution network, and performing graph convolution operation and feature aggregation; fusing the node feature representation with the multi-scale feature map to obtain a fused feature; and performing up-sampling and pixel-level mapping on the fused features to generate a semantic segmentation result map. According to the method, the image convolutional network and a multi-layer attention mechanism are fused, adaptive modeling of image semantic segmentation is realized, and the method has the advantages of high precision, strong robustness and global perception.

The invention discloses a reconfigurable intelligent surface (RIS)-assisted multiple-input-multiple-output (MIMO) implicit channel estimation method based on a graph attention network, which is used for efficient downlink transmission in a multi-user scene. Firstly, a graph attention network is designed, user nodes and RIS nodes are modeled in a unified mode, received pilot signals serve as initial features, spatial feature expression is enhanced in combination with user three-dimensional position information, and therefore interference between users and a spatial correlation structure are accurately represented; secondly, end-to-end feature aggregation is achieved based on a message passing mechanism, a base station beam forming matrix and an RIS reflection coefficient are directly predicted under the condition that explicit channel estimation is not needed, and the total transmitting power constraint and the unit mode constraint are met through normalization processing so as to complete joint optimization; according to the method, the users and the speed of the system can be remarkably improved under limited pilot frequency overhead, and the method has excellent generalization performance and robustness in a multi-user complex propagation environment.

The invention discloses a voiceprint fusion and knowledge distillation insulator modeling evaluation method and system. The method comprises the steps that firstly, insulator voiceprint signals and synchronous working condition data are collected, and preprocessing is carried out to generate a multi-dimensional voiceprint feature map and a working condition scalar feature vector; secondly, constructing a teacher model, extracting voiceprint time-frequency features by using a double-flow spatial-temporal feature aggregation architecture, deeply embedding working condition scalar quantities into voiceprint feature tensors through a broadcast mechanism, and generating multi-modal fusion features; thirdly, constructing a lightweight student model, guiding the lightweight student model to perform knowledge distillation training by using multi-modal fusion features of the teacher model, and constructing a multi-task joint loss function optimization model parameter in combination with cross-view feature alignment loss, focus loss and center loss; and finally, deploying the student model subjected to training convergence to a monitoring terminal for real-time reasoning. According to the method, the problems of difficulty in insulator fault feature extraction and difficulty in model edge deployment under complex working conditions are effectively solved, and accurate evaluation of the health state is realized.

The invention relates to the field of distributed information retrieval, and particularly discloses a memory architecture vector approximate retrieval method based on a graph neural network, which comprises the following steps of: constructing and dynamically maintaining a historical query-hit vector association graph for modeling a deep semantic relationship between a historical query and a successful retrieval result; inputting the features of the current query vector, the information of the current query vector subjected to neighborhood sampling and feature aggregation in the graph and the service scene label into a lightweight graph neural network, and predicting the approximate retrieval tolerance level of the query; on the basis of the prediction result, an optimal retrieval strategy is generated in a self-adaptive mode; and in combination with asynchronous result return and a progressive refinement mechanism based on residual error reordering, a user is responded at the first time, and continuous optimization and pushing of a better result are realized. According to the method, a query-level personalized retrieval strategy is realized, the retrieval precision, the response delay and the system resource consumption are effectively balanced, and the method is suitable for a large-scale high-dimensional vector retrieval scene.

The invention discloses a light-weight small target detection method and system for images shot by an unmanned aerial vehicle, and the method specifically comprises the steps: constructing a neural network architecture which comprises a backbone network, a feature aggregation network and a detection head; improvement of light weight and attention enhancement is implemented in the backbone network, and a multi-scale initial feature map is extracted; constructing a feature aggregation network Neck, performing cross-level fusion and refining processing on the multi-scale initial feature map, and outputting a refined feature map; a lightweight target detection head Head is constructed in combination with a large-kernel depth separable convolution module and a special decoupling head structure of a YOLOv11 network, and decoupling prediction is performed on the refined feature map; training is carried out by adopting a mixed loss function based on a normalized Wasserstein distance and modulated IoU, a trained lightweight network is obtained, and detection of a lightweight small target is realized. According to the invention, the complexity of the model is reduced, the detection speed is improved, the high detection precision is maintained, and the method is suitable for real-time detection tasks on an unmanned aerial vehicle resource limited platform.