283 results about "Pooling" patented technology

Disclosed in the present invention is a four-network integration architecture for an unmanned swarm system. The four-network integration architecture has the capabilities of heterogeneous platform resource pooling, intelligent dynamic computing power allocation, and timely decision planning, so as to maximize the overall benefit. The present invention focuses on abstracting and integrating independent submodules to form a mesh topology of a swarm. The present invention designs a four-network integration architecture for an unmanned swarm system, which comprises a computing power network, a perception network, a decision network and a communication network as core modules. The structure aims to achieve efficient cooperation of all parts in the swarm, thereby improving the overall performance and adaptability of the system. The system integrates environmental perception, a swarm network modeling component, a knowledge base and a resource pool, providing an intelligent environmental perception strategy and a network modeling strategy for the interior of the swarm. Therefore, the perception of environments, tasks and networks by nodes can be facilitated, thereby completing establishment of intelligent networks, so as to ensure the characteristics of the stability and flexibility of networks.

The invention provides an operational research course knowledge graph construction method based on multi-source data fusion. The method comprises the following steps: firstly, discussing logical association of courses, majors and students, collecting data such as textbooks, exercises and teaching programs by taking operational research knowledge points as a core and utilizing technologies such as OCR (Optical Character Recognition) and crawlers, and carrying out preprocessing and manual labeling; then, an improved deep learning model is adopted for entity recognition and relation extraction, BERT + BiLSTM + CRF is adopted for entity recognition, and a dynamic context pooling enhancement model is fused to improve the capture ability of a complex knowledge boundary; bERT + BiLSTM is adopted for relation extraction, a multi-head attention mechanism is combined, and hidden logical relation mining is enhanced. And finally, constructing a multi-level knowledge network which takes knowledge points as nodes and logic relations as edges, and embedding the multi-level knowledge network into a Neo4j graph database for visualization. The map can optimize a teaching path, provides personalized learning recommendation, and is widely applied to the fields of wisdom education, knowledge retrieval and the like.

The invention discloses a hybrid expert and KAN-based cyclic attention network time sequence prediction method. The method comprises the following steps: S100, inputting time sequence data needing to be predicted; s200, constructing a graph structure by using an attention mechanism, learning basic correlation characteristics among variables of the input time sequence data through an adaptive and learnable graph convolutional network, and then performing global averaging and maximum pooling on the basic correlation characteristics along a time dimension to obtain complementary time domain statistical information, so as to provide effective time-space correlation characteristics for the follow-up process; s300, after feature learning is completed, collaborative modeling of the KAN and an attention mechanism is brought into full play, rapid and efficient time sequence modeling is carried out on data by adopting a cyclic attention network embedded based on the KAN, and a foundation is laid for subsequent time sequence prediction; and S400, establishing a hybrid KAN expert-based time sequence prediction network, and adaptively fusing differentiation prediction results by a gating mechanism. The time sequence prediction method is designed from the three aspects of feature learning, time sequence modeling and time sequence prediction.

The invention discloses an intelligent computing power integration service management method and platform based on cloud side-end cooperation, and relates to the technical field of computing power integration management, and the method comprises the steps: sensing the computing power resource state of a side-end cooperation port and terminal equipment in real time at a cloud controller, and building a resource topological graph; a dynamic task demand is introduced, and a computing power scheduling vector is generated; carrying out lightweight segmentation on the cloud training model, and determining adjacent edge nodes under hierarchical limitation to form a regional elastic cluster; and deploying a digital twin simulation engine rehearsal computing power distribution scene, establishing a heterogeneous resource pooling mechanism, and carrying out computing power integration service management. The technical problems of low management efficiency and insufficient utilization rate of heterogeneous computing power resources in the prior art are solved, and the technical effects of realizing efficient management of intelligent computing power integration services and improving the utilization rate of the heterogeneous computing power resources are achieved.

The invention provides a multi-receptive-field network abnormal flow detection method and system based on endogenous security attributes, and relates to the technical field of network security and abnormal flow intelligent detection. The method comprises the following steps: firstly, performing multi-scale flow representation, preprocessing and data enhancement on network flow data to obtain enhanced input flow data; local features are extracted through basic convolution, and local and global fusion features are obtained based on a double-branch network comprising a multi-receptive field convolution branch and a Mama-self-attention branch; deep fusion representation is formed through multi-round feature extraction and tensor fusion, and finally binary classification and fine-grained classification results are output through global pooling and a linear classification layer. According to the method, high-precision, high-robustness and high-real-time detection of the abnormal traffic of the complex network is realized under low calculation overhead.

The invention provides a hit probability prediction method and device, computer equipment and a storage medium, and the method comprises the steps: replacing a random number sequence in a white shark optimization algorithm WSO with a chaotic mapping sequence determined by a chaotic mapping function, and carrying out the population optimization of the white shark optimization algorithm WSO, and obtaining an improved white shark optimization algorithm CWSO; a KAN layer and a fuzzy pooling layer are introduced into a convolutional neural network CNN, and a hit probability prediction model of the KCNN underwater launching device is constructed; obtaining influence factor training data, and performing parameter optimization on the KCNN model through an improved white shark optimization algorithm CWSO by adopting the influence factor training data to obtain a CWSO-KCNN model; and obtaining influence factor target data, and inputting the influence factor target data into the CWSO-KCNN model to obtain a hit probability prediction value of the underwater launching device. Therefore, the CWSO is obtained by improving the optimization algorithm, and the CWSO is adopted to perform parameter optimization on the KCNN model to obtain the CWSO-KCNN model, so that the prediction precision of the hit probability is effectively improved.

The invention discloses a punching machine bearing fault diagnosis method and equipment. The method comprises the following steps: establishing a diagnosis mechanism fusing time-frequency domain feature extraction and multi-branch depth feature modeling; original vibration signals are respectively converted into GADF images to capture time sequence dependence, CWT is adopted to extract transient frequency change, STFT is utilized to obtain frequency spectrum characteristics under a stable working condition, and joint input of a multi-view time-frequency graph is realized. On a model structure, a Swin Transform branch is designed for local space detail modeling, a global perception branch is constructed by combining a CNN module and a GAM module, and multi-category fault classification is completed through self-adaptive pooling and a full connection layer on fusion output. User-defined data loading and a standardized process guarantee the consistency of training data, multi-core maximum mean difference is introduced for feature distribution alignment, and the migration ability of the model under different working conditions and the decoupling recognition ability of composite faults are improved.

The invention discloses a multi-source heterogeneous time series data compression and motor fault prediction method and system, and the compression method achieves the intelligent compression and feature enhancement of time series data through the cooperation of multi-scale feature extraction, SE weighted fusion, mixed pooling and aggregator collapse. According to the architecture, reasonable reduction of the sequence length and effective improvement of the feature depth can be completed at the same time, on the premise that key fault features are reserved, the technical problem of multi-source heterogeneous sensor data fusion in an industrial scene is solved, and the limitation of serious information loss of a traditional dimension reduction method is overcome.

The invention discloses a dynamic space-time diagram flow prediction method and system based on course learning, and relates to the technical field of supply chain logistics data analysis, and the method comprises the steps: building a space-time matrix based on historical multi-source data, generating a dynamic adjacent matrix through learning, and carrying out the smooth fusion through combining a static diagram, and forming a dynamic diagram structure. And then space and time features are respectively extracted by using a graph convolutional network and a gating loop unit, and deep interaction and fusion are realized through a bidirectional cross attention mechanism. A multi-dimensional difficulty estimator is innovatively introduced, the prediction difficulty of each training sample is quantified from three dimensions of space, time and time-space coupling, the selection sequence of the training samples is dynamically adjusted based on an adaptive course scheduler, and progressive learning is realized. And finally, feature representation is obtained through global pooling, and multi-step traffic prediction is realized by adopting a parallel independent decoder, so that error accumulation is avoided. According to the invention, prediction precision and model training efficiency in a complex supply chain logistics scene are effectively improved.

The invention discloses a collaborative recommendation method and device fusing lightweight relation path completion and dynamic negative sampling, equipment and a medium, and relates to the technical field of recommendation. The collaborative recommendation method comprises the following steps: acquiring historical interaction data of a user, and constructing an interaction relationship knowledge graph; vectorizing the interactive relationship knowledge graph, and then dynamically constructing a multi-hop relationship path based on a lightweight relationship path completion mechanism to obtain a path completion representation. According to the lightweight relation path completion mechanism, a path order L belongs to {1, 2, 3} is set according to the sparse degree of a user-article pair, and a first-order, second-order or third-order reasoning path is constructed. And then performing intelligent feature fusion on the constructed path through a triple pooling strategy. And projecting the embedded representation of the entity to the corresponding relation subspace through orthogonal projection based on the path completion representation. And calculating a recommendation score according to the embedded representation after projection, and obtaining an article recommendation list.

The invention discloses a Tor network exit traffic identification system and method fusing a multi-scale LSTM and a Transform network, and belongs to the technical field of anonymous network traffic analysis and network security. The system comprises five core components, namely a multi-scale feature extraction module, a feature fusion module, a global dependency modeling module, a dynamic weighted aggregation module and a classification module. The multi-scale feature extraction module adopts parallel bidirectional LSTM branches with different time resolutions to capture a microcosmic burst mode and a macroscopic session behavior at the same time; the feature fusion module unifies the scale features to the same time sequence length and splices the scale features; the global dependence modeling module utilizes a multi-head self-attention mechanism to learn long-distance time sequence dependence; the dynamic weighted aggregation module highlights a key time slice through adaptive weight pooling; and the classification module outputs website category labels. According to the system, the recognition accuracy on a GTT23 data set is remarkably improved compared with that of an existing method, and good recognition capability and robustness are shown for various flow defense mechanisms.

The invention discloses a mapping knowledge domain embedding method fusing meta-information and logic rules, and relates to the technical field of computers. The method comprises the following steps: inputting a to-be-predicted knowledge graph into a knowledge graph embedding module; in the multi-granularity meta-information embedding module, for each entity to be predicted in the knowledge graph to be predicted, obtaining ontology information embedding and multi-hop neighbor relation embedding of the entity to be predicted; carrying out average pooling processing on each of the multi-hop neighbor relation embedding, and processing the neighbor relation embedding after the average pooling processing through a multi-layer perceptron and learnable parameters to obtain a hierarchical attention weight of each neighbor relation embedding; each multi-hop neighbor relation embedding comprises a plurality of neighbors; and performing weighted summation on the ontology information embedding and the multi-hop neighbor relation embedding based on the hierarchical attention weight to obtain meta-information embedding of a to-be-predicted entity in the to-be-predicted knowledge graph. According to the method, the determination precision of the meta-information embedding of the entity can be improved.

The invention requests to protect a network intrusion detection method based on a CKAN-BiLSTM (Content Kernel Area Network-BiLSTM). The method comprises the following steps: firstly, selecting an NSL-KDD data set widely applied in the field of network security, and performing preprocessing operation on the NSL-KDD data set to improve data quality and a model training effect; then, feature extraction is carried out on input data through a convolutional layer, size adjustment is carried out on features in combination with an adaptive pooling mechanism, and the perception ability of the model to a local mode is enhanced; secondly, inputting the extracted features into a BiLSTM network to fully capture a time sequence dependency relationship in the BiLSTM network; and finally, Kolmogorov-Arnold Network (KAN) is introduced to carry out weighted fusion on the key features, final attack type classification is completed, and efficient and accurate intrusion detection is realized.

The invention discloses an NVMe storage array virtualization device based on an FPGA (Field Programmable Gate Array), which is characterized in that single I / O (Input / Output) virtualization (SR-IOV) is realized in the FPGA, and a plurality of NVMe virtual functions which can be directly accessed by a virtual machine are externally provided; mapping from a virtual volume logic address to a physical storage address is completed in hardware, striping management and RAID verification calculation are carried out on a plurality of NVMe solid state disks at the rear end, and storage resource pooling is achieved; the partial reconfigurable characteristic of the FPGA is utilized, I / O scheduling, data compression and prefetching strategies are dynamically adjusted, and the adaptability and performance efficiency of the system are improved. Hardware unloading of the whole I / O path is completed in the FPGA, intervention of a host CPU is not needed, delay is remarkably reduced, the throughput capacity is improved, service quality isolation and elastic expansion in a multi-tenant environment are supported, and an effective technical path is provided for constructing a next-generation high-performance and low-delay storage infrastructure.

The invention provides a multi-source heterogeneous data integrated management method based on a digital main line, a medium and a system, and belongs to the technical field of industrial digital main lines. A digital main line platform is used for collecting data of various formats and conducting standardization processing, a deep ontology fusion model is adopted for achieving cross-domain semantic understanding, and the data of various formats is obtained. Semantic understanding efficiency is improved through adaptive pooling feature dimensionality reduction and separable attention calculation, a storage space is optimized by adopting similarity aggregation de-duplication or a distributed independent storage strategy according to a data overlap ratio, and a digital principal line data consanguinity tracing relation is established to realize full-life-cycle data link construction. Based on the semantic mapping deviation value, a fine fine tuning or coarse tuning mode is adopted to optimize semantic understanding parameters, longitudinal integrated mapping from a function model to a performance model and a physical model and a transverse integrated framework of multi-domain comprehensive simulation are constructed, and the technical problem that cross-domain model semantic understanding accuracy is insufficient is solved.

The invention provides a dynamic network security adaptive defense method and system, and the method comprises the steps: collecting multi-source security event data, carrying out the time synchronization and context information addition of the multi-source security event data, and constructing a weighted attack behavior graph model; based on the weighted attack behavior graph model, graph-level feature vectors are extracted in a weighted pooling mode, and a strategy index is generated in combination with predefined strategy template vectors in a strategy library; retrieving a corresponding parameterized strategy template from a strategy knowledge base according to the strategy index, performing parameter filling in combination with real-time network state information, dynamically adjusting parameter values, and generating an executable strategy instruction set; and performing priority ranking and conflict detection on actions in the strategy instruction set, translating the actions into instructions which can be executed by specific equipment in batches, issuing the instructions to target safety equipment in sequence for execution, and generating a strategy execution state record for auditing.

The invention discloses a reverse knowledge distillation-based cross-stage feature fusion method for industrial anomaly detection and positioning, and relates to the technical field of knowledge distillation and transfer learning. The method comprises the following steps: step 1, constructing a teacher network encoder, a student network encoder and a decoder; 2, extracting input image features through a student network encoder, and reconstructing the features through a student decoder; step 3, student network cross-stage feature fusion design; 4, locally sensing a dynamic attention module LDA, and adding a sliding window and convolution to replace original global pooling; and step 5, a self-adaptive multi-scale feature fusion module CAAMS-FF establishes a dependency relationship of a context so as to realize a fine-grained high-quality feature reconstruction effect. And step 6, inputting an anomaly graph obtained by the teacher-student network into the segmentation sub-network to obtain a final anomaly detection and positioning result. While the calculation efficiency is maintained, the spatial positioning precision and the multi-scale adaptability are significantly improved.

The invention discloses a PCB bare board defect detection method and system based on an improved YOLOv11 neural network, and the method comprises the steps: obtaining a PCB defect data set, carrying out the preprocessing of an image of the data set, and constructing a PCB surface defect data set; an improved YOLOv11 detection model is constructed, in the Backbone stage, the depth separable convolution of improved PPLCNet and an SE module are utilized to reduce the calculated amount and enhance feature expression, in the Neck stage, multi-scale information fusion is optimized through cross-scale feature splicing and SPPF pooling, in the Head stage, an improved ShapeIoU loss function is introduced, and an improved edge enhanced EE-SimAM attention mechanism module is inserted in front of the object; training the model by using the preprocessed data set, and optimizing the initial weight by adopting a migration analysis strategy; and deploying the trained model to an edge detection device. The method has the beneficial effects that by constructing and improving the YOLOv11 network model, the size of the model is reduced, the model can be conveniently deployed to an edge equipment end, the PCB defect detection precision is improved, the omission ratio is reduced, and the method is suitable for industrial defect detection scenes.

A hydraulic loading system fault diagnosis method based on sequence learning comprises the steps that firstly, sample data and corresponding labels are imported, and the imported sample data are flattened and then subjected to normalization processing; then, a three-layer one-dimensional CNN convolution structure is adopted, each layer comprises convolution, batch normalization, ReLU activation function and maximum pooling operation, and a feature matrix is obtained; then, inputting the feature matrix into a single-layer one-way GRU network to capture a dynamic time sequence dependency relationship, and taking a final hidden state of the dynamic time sequence dependency relationship as a global feature representation; and finally, the global features are mapped to a fault category space by a full connection layer, and a Softmax activation function is used to complete fault classification. Cross entropy is adopted as a loss function, the normalized data sample is used for training, and when the accuracy of the verification set reaches a design value or reaches the maximum training round number, the training is ended; according to the method, the automation degree and reliability of fault diagnosis of the hydraulic loading device are remarkably improved by fusing the advantages of the CNN in feature extraction and the GRU in time sequence modeling.

The invention provides a small target detection method for unmanned aerial vehicle data, and the method can effectively improve the detection precision, achieves better balance in three aspects of light weight, high precision and real-time performance, and can be suitable for more small target detection scenes. The method comprises the following steps: constructing a PLDP-SPP module, calculating a regional significance score S and a texture complexity score T for each pixel point in an input feature map in a feature content analysis module, and calculating to obtain a comprehensive score V by taking S and T as quantitative indexes; different pooling scales are preset according to the comprehensive scores V in different ranges, and it is ensured that more detailed context information extraction can be carried out on an area with large information density; in a size decision mechanism module, feature densities of different areas are matched through dynamic pooling, the scale adaptability is higher, and particularly, the detection precision of small targets and medium targets is remarkably improved.

The invention discloses a switching communication device, a switching communication method and a server, relates to the technical field of computers, and realizes equipment-level decoupling and resource pooling basis by actively managing a GPU (Graphics Processing Unit) through a second interface configured to be in a root complex mode. The on-chip management system and the switching logic module form a control core, a data switching strategy is dynamically executed, and conversion from passive forwarding to autonomous management is achieved. And the virtual port module performs software definition and dynamic allocation on network resources, so that the flexibility of resource utilization is improved. Meanwhile, MAC ports in two modes of direct connection and exchange are supported, so that the device can flexibly construct a high-performance direct connection network or a large-scale exchange network, and the limitation of fixed topology is broken through. And the data processing module ensures high efficiency of data conversion between different interface protocols. According to the device, the communication efficiency and schedulability between the GPUs are improved by integrating high-speed exchange, intelligent management, resource virtualization and flexible networking capabilities.

The invention discloses a main and distribution network collaborative analysis method, system and device and a storage medium, and belongs to the technical field of power system analysis and calculation. The method comprises the following steps: abstracting a main power distribution network integrated model into a heterogeneous graph and generating a fusion node feature matrix; a time-space diagram neural network model is constructed and trained, the model ensures that a prediction result conforms to a power system physical law by introducing physical information neural network constraints, and multi-task parallel prediction is completed by adopting a diagram self-attention pooling and multi-task learning architecture; performing cross-domain interaction verification on a collaborative analysis result output by the model, wherein the cross-domain interaction verification comprises boundary consistency verification and energy / power balance verification; and dynamically adjusting calculation granularity or model parameters according to a verification result, and accelerating a calculation process based on a linear self-attention mechanism. According to the method, the problems of model splitting and inconsistent calculation results caused by independent analysis of the traditional main and distribution networks are solved, and global consistent, physically credible, efficient and collaborative power grid analysis is realized.

The invention discloses an open source environment software hidden vulnerability patch identification method, which is based on a multi-stage architecture and collaborative relationship modeling, and provides a new patch group identification scheme: firstly, obtaining candidate code submission from an open source warehouse, extracting correlation characteristics of vulnerabilities and candidate code submission, including rule-based characteristics and semantic characteristics, calculating a correlation score and screening high-correlation submission; pairwise pairing the high-correlation submissions, and fusing multi-dimensional features to predict a cooperative relationship between the submissions; and finally, constructing an undirected graph based on the correlation score, dividing a maximum connected sub-graph, fusing the features in the group through maximum pooling, calculating the correlation with the vulnerability, and outputting an optimal patch group. Meanwhile, an existing patch identification technology based on sorting learning is combined, an enhancement method based on a submission cooperative relation is provided, ranking logic submitted by candidate codes is updated through internal association between code submission and by means of correlation between group vectors and vulnerabilities, and the identification precision in a multi-patch scene is improved.

The invention discloses a satellite network topology discovery method and a path construction method and device, and the method comprises the steps: dividing a satellite network into autonomous domains through distributing an extended SRv6 SID which comprises a network prefix, an orbital plane number, a node number, a service type and a resource margin for each satellite, and building an SID matrix and a satellite node adjacency matrix; the satellite nodes sense own resource states, periodically detect link states by adopting a BFD (Bidirectional Forwarding Detection) protocol, and synchronize values of elements in a matrix through an incremental updating mechanism to realize dynamic topology sensing; meanwhile, a routing path construction method based on an SID matrix and an adjacent matrix is provided, and satellite nodes are supported to carry out path calculation locally. The method supports heterogeneous service identification and resource state awareness, can effectively reduce signaling overhead, improves accuracy and routing efficiency of satellite network topology discovery, and improves cross-constellation computing power pooling and cooperative computing capability.

The invention relates to the technical field of computer vision, in particular to a small target fine segmentation method fusing improved pyramid pooling and edge branch supervision, and the method comprises the steps: selecting a data set, carrying out the preprocessing of a data set image, carrying out the feature extraction of the preprocessed image through a backbone network MobilenetV3, and generating four sets of feature maps F1 to F4, the method comprises the steps of inputting F1 and F2 into an ABG module, generating feature maps F5 and F4 containing edge information, sending the feature maps F5 and F4 into an improved ASPP structure, obtaining a feature map F6 containing multi-scale context information, inputting F3 and F4 into an FCE module for feature fusion so as to supplement spatial details, finally splicing and fusing F5 and F6, enhancing through the FCE module, and gradually performing up-sampling back to the original size in combination with the supervision of the feature maps F3 and F4, so as to obtain the multi-scale contextual information. And a final segmented image is obtained. According to the invention, through integration of the segmentation network, full-link optimization of'feature extraction-edge enhancement-multi-scale modeling-feature calibration-precise decoding 'is realized.

The application discloses an edge computing processing system based on a regional comprehensive energy station, and the edge computing processing system comprises a data input module, a data processing module, a storage module, a communication module and a power conversion module; the data input module converts analog flow from the regional comprehensive energy station into data information; the data processing module performs task distribution on the data information in a time sequence pipeline mode; the storage module stores and calls the task distribution; the communication module transmits the task distribution to a server; and the power conversion module is used for supplying power to the edge computing processing system; wherein the neural network operation module comprises an input buffer unit, an intermediate calculation buffer unit, an average pooling unit, a control unit and a convolution operation unit; and the application solves the technical problems of complex hardware structure of an edge computing terminal and slow data processing task in the prior art.

The invention discloses a processing method and device for executing maximum pooling in a neural network, and the method sets kh * kw as the height and width of a pooling window of a target to be realized, tkh * tkw as the remaining height and width after the split pooling window, and tkh = kh and tkw = kw at the initial time, and comprises the steps: S1, calculating the size kh '* kw' of the pooling window split for the first time; s2, carrying out primary maximum pooling operation with the pooling window size of kh '* kw'; s3, tkh, tkw and the following pooling window parameter kh '* kw' are updated; s4, if at least one of the tkh and the tkw is not 1, performing maximum pooling operation with a pooling window of kh '* kw', and returning to the step S3; s5, judging whether to add the extraction point extraction operation or not according to the step length: if the step length height is shgt, judging whether to add the extraction point extraction operation or not; 1 or a wide swgt; if so, adding the extraction operation of sh * sw; if both the sh and the sw are 1, the operation is not added. According to the invention, hardware design logic and implementation complexity can be reduced, and hardware only needs to implement simple size (lt; = 3 * 3), and the size gt does not need to be achieved; and 3, maximum pooling.

The application discloses a media cloud desktop management method and system based on a bare metal server. The application comprises the following method: providing the basic resources of the cloud desktop in a bare metal pooling mode, adding instance matching specifications suitable for different vGPU use scenarios in the media cloud desktop system before the cloud desktop is created, creating cloud desktop template kits belonging to different virtual laboratories according to the corresponding specifications, and connecting various tool-type cloud desktops belonging to different laboratories through network access mode of the client. The application provides a media cloud desktop management method and system based on a virtual laboratory mode of an elastic bare metal server, realizes efficient management of the vGPU resources of the bare metal pooling, provides a convenient management mechanism for cloud desktop authorization and recycling, reduces the idle rate of the tool-type desktops, greatly improves the desktop use efficiency of the vGPU, realizes efficient operation and maintenance and elastic publishing and deployment capabilities of the exclusive desktop and a large number of multi-version desktop templates.

The invention belongs to the technical field of data mining, and discloses an aero-engine missing data filling method based on a neural network with an exogenous variable graph. Aiming at the detection parameter data and the operation parameter data, a sliding window is used for dividing according to the time dimension to obtain an intermediate matrix; the method comprises the following steps of: performing sliding window division on high-dimensional representation of input data to construct a time sequence-attribute graph; the attention is used for calculating the feature representation of the action intermediate matrix and the action time sequence-attribute graph, and a final weight matrix of the time sequence-attribute graph is obtained through combination. Convolution operation is carried out on the time sequence-attribute graph, so that the node can dynamically aggregate influences from multiple historical moments and multi-dimensional attributes, and meanwhile, the directional effect of exogenous variables on endogenous variables is effectively fused. And splicing all the convoluted feature representations, fusing time sequence information of different time points through a pooling layer, and finally calculating by adopting two linear layers and an activation function to obtain a final filling result.

The invention relates to the technical field of part identification, and discloses a part identification method and device based on an industrial large model, and the method comprises the steps: carrying out the image block division and embedded transformation of an original image of an industrial part, obtaining a first Token sequence, carrying out the two-dimensional discrete cosine transformation of the first Token sequence, and obtaining a second Token sequence; fusing the second Token sequence with a plurality of prototype feature vectors selected from a category prototype memory library to obtain a memory guide vector; performing gating modulation and residual connection processing on the second Token sequence based on the memory guide vector to obtain a third Token sequence; and performing iterative refinement on the third Token sequence to obtain a fourth Token sequence, and performing global pooling and classification prediction on the fourth Token sequence to obtain a category identification result of the industrial parts, thereby enhancing the discrimination of the part characteristics, and effectively solving the technical problem of difficult identification of small samples of rare parts.