A Situation Collaborative Cognition Method Based on Distributed Large Model
By calculating the deep feature parameters of the situation data on multiple end nodes, and fusion on the cloud server, the problem of low efficiency and poor timeliness of the existing situation cognitive method is solved, and the accuracy and timeliness of multiple goals are realized. Situation cognition.
Patent Information
- Application Number
- CN202311854747.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-29
- Publication Date
- 2025-01-24
- Estimated Expiration
- 2043-12-29
AI Technical Summary
The existing situation cognitive methods are inefficient in computing efficiency and poor timeliness due to centralized processing, which cannot meet the needs of some application scenarios.
Adopt a combined cognitive method based on a distributed -based model, to obtain the timing data of the target on multiple end nodes, calculate deep feature parameters, and integrate and recognize the situation on the cloud server.
The fusion of multi -end timing information is realized, and multi -tasking disassembly is introduced in the introduction of ultra -networks for large models, which can accurately and timeliness of multiple targets.
Smart Images
Figure CN117786462B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data identification and data representation, and in particular to a situation collaborative recognition method based on a distributed large model. Background Art
[0002] Situational awareness not only includes the positioning and identification of the target, but also the prediction of the motion trajectory of the dynamic target. However, if accurate situational awareness of the target is to be achieved, the model needs to have extremely strong information processing capabilities and requires all-round and multi-angle time series observation data.
[0003] Existing situational awareness methods generally adopt centralized processing, using large models deployed on cloud servers to process time-series observation data gathered from different observation platforms. Although this model can obtain relatively accurate situational awareness of the target, the amount of time-series observation data required for aggregation is large, the transmission speed is slow, and the calculation efficiency is low. Therefore, the situational awareness of the target lacks timeliness and cannot meet the needs in some application scenarios. Therefore, designing an accurate and timely target situational awareness method is an urgent problem to be solved. Summary of the invention
[0004] 1. Technical solution
[0005] An embodiment of the present invention provides a situation collaborative recognition method based on a distributed large model, including: obtaining N time series data of a target based on N end nodes, wherein each end node obtains one time series data of the target respectively; calculating at least one deep feature parameter of the target based on each time series data respectively; and obtaining all deep feature parameters of the target on a cloud server, and performing situation recognition on the target based on all deep feature parameters of the target, wherein N is an integer greater than or equal to 1.
[0006] Optionally, the scenes of the target corresponding to different time series data are different, and calculating at least one deep feature parameter of the target based on each time series data includes: extracting a shallow feature parameter of the target based on each time series data; extracting a corresponding deep feature parameter from a shallow feature parameter of the target based on at least one deep feature extraction layer, wherein different deep feature extraction layers use different extraction parameters to extract deep feature parameters.
[0007] Optionally, before extracting a corresponding deep feature parameter from a shallow feature parameter of the target based on at least one deep feature extraction layer, the method further includes: allocating a shallow feature parameter of the target to at least one corresponding deep feature extraction layer based on a hypernetwork.
[0008] Optionally, extracting a shallow feature parameter of the target based on each time series data includes: extracting a shallow feature parameter of the target using a CNN convolutional neural network.
[0009] Optionally, allocating a shallow feature parameter of the target to at least one corresponding deep feature extraction layer based on the hypernetwork includes: allocating a shallow feature parameter of the target to at least one deep feature extraction layer based on a gating function.
[0010] Optionally, assigning a shallow feature parameter of the target to at least one corresponding deep feature extraction layer based on the gating function includes: calculating a weight value of the correlation between a shallow feature parameter of the target and the deep feature extraction layer; and assigning a shallow feature parameter of the target to at least one deep feature extraction layer according to the weight value.
[0011] Optionally, extracting corresponding deep feature parameters from a shallow feature parameter of the target based on at least one deep feature extraction layer includes: at least one deep feature extraction layer extracting corresponding deep feature parameters from a shallow feature parameter of the target based on a standard feedforward neural network.
[0012] Optionally, performing situation awareness on the target based on all the deep feature parameters of the target includes: performing feature fusion on all the deep feature parameters of the target to obtain target feature information; and performing situation awareness on the target based on the target feature information.
[0013] Optionally, performing feature fusing on all deep feature parameters of the target includes: performing feature fusing on all deep feature parameters of the target based on a self-attention mechanism.
[0014] Optionally, performing situation awareness on the target based on the target feature information includes: decoding the target feature information based on a Transformer decoder to achieve situation awareness on the target.
[0015] (II) Beneficial effects
[0016] The present invention provides a situation collaborative recognition method based on a distributed large model. By acquiring characteristic parameters of a target in multiple terminal nodes and realizing situation recognition of the target in a cloud node according to the characteristic parameters, a "one cloud, multiple terminals" collaborative framework is constructed, which can realize the fusion of multi-terminal time series information, and introduce a super network to realize multi-task decomposition of the large model. Different terminal nodes are used for analysis based on different scenes where the target is located, so that the characteristic parameters of the target are obtained through collaboration between the end nodes. Finally, accurate and timely situation recognition of multiple targets can be performed. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 Schematically shows a flow chart of a situation collaborative recognition method based on a distributed large model provided in an embodiment of the present invention;
[0018] Figure 2The flowchart of calculating the target deep feature parameters provided by the embodiment of the present invention is schematically shown. DETAILED DESCRIPTION
[0019] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.
[0020] The description of the structural embodiments and methods of the present invention is disclosed herein. It should be understood that this is not intended to limit the present invention to the specific disclosed embodiments, and the present invention can be implemented by using other features, elements, methods and embodiments. Similar elements in different embodiments are usually marked with similar numbers.
[0021] Figure 1 A flow chart of a situation collaborative recognition method based on a distributed large model provided by an embodiment of the present invention is schematically shown.
[0022] Figure 1 Includes operations S110~S130.
[0023] In operation S110, N time series data of the target are acquired based on N end nodes, wherein each end node acquires one time series data of the target respectively.
[0024] In this embodiment, the N end nodes may be satellite nodes, and the target scenes included in the N time series data are different.
[0025] In operation S120 , deep feature parameters of the target are calculated based on the N time series data.
[0026] In this embodiment, each end node includes a deep feature extraction layer, and the extraction parameters of the N deep feature extraction layers of the N end nodes are different, which are used to calculate the deep feature parameters of the target in each scene, and a deep feature extraction layer is used as an expert scene model. For example, a building scene, an airplane scene, etc., so that corresponding feature extraction can be achieved for different scenes where the target is located.
[0027] Figure 2 The flowchart of calculating the target deep feature parameters provided by the embodiment of the present invention is schematically shown.
[0028] Figure 2 Includes operations S210~S230.
[0029] In operation S210, shallow feature parameters of the target are extracted based on the time series data.
[0030] In this embodiment, each end node uses a CNN convolutional neural network to extract shallow feature parameters of the target, and the neural network model used by each end node is the same and has the same neural network model parameters, ensuring that the extracted shallow feature parameters of the target are consistent.
[0031] In this embodiment, each end node uses a ResNet-18 network to extract shallow feature parameters of the target. The network model consists of four residual blocks with the same structure but different dimensions, and each residual block is repeated twice during the operation process.
[0032] In this embodiment, before performing the operation of the residual block, it is necessary to obtain the convolution parameters and pooling parameters of the residual block.
[0033] In this embodiment, each residual block includes a straight line branch and a shortcut branch. The feature matrix of the target time series data is obtained and input into the straight line branch and the shortcut branch respectively. In the straight line branch, the feature matrix is convolved multiple times and the convolution matrix is output. In the shortcut branch, the feature matrix is not processed. Finally, the result of the straight line branch is added to the result of the shortcut branch to obtain the shallow feature parameters of the target.
[0034] In operation S220, shallow feature parameters are allocated to corresponding deep feature extraction layers based on the hypernetwork.
[0035] In this embodiment, the hypernetwork uses a gating function to allocate shallow feature parameters, and the gating function is used to determine which expert scene models to assign the shallow feature parameters to.
[0036] In this embodiment, the gating function calculates the weight value of the correlation between the shallow feature parameters and the expert scene model, and assigns the shallow feature parameters to one or more expert scene models according to the weight value. At the same time, in order to ensure the sparsity of the output, the hypernetwork retains the top K values of the gating function (Top-K Sampling) and sets the non-Top-K values to zero. The output of the gating function designed in this way is sparse, and only some expert scene models have non-zero weights, while expert scene models with a weight of 0 do not participate in the calculation.
[0037] In operation S230, deep feature parameters of shallow feature parameters are extracted based on N different deep feature extraction layers.
[0038] In this embodiment, the expert scene model uses a standard feedforward neural network (FNN) to extract deep feature parameters, wherein different expert scene models have the same network model, but the network parameters corresponding to the network model are different, thereby enabling the extraction of deep feature parameters for targets in different scenarios.
[0039] In this embodiment, the FNN includes two fully connected layers, the first layer uses Relu as the activation function, and the second layer does not use an activation function.
[0040] Return to reference Figure 1 In operation S130, deep feature parameters of the target are obtained on the cloud server, and situation awareness of the target is performed based on the deep feature parameters.
[0041] In this embodiment, all deep feature parameters of the target obtained by each end node are first obtained on the cloud server.
[0042] In this embodiment, performing situation recognition on the target includes: performing feature fusion on all deep feature parameters to obtain target feature information; and performing situation recognition on the target based on the target feature information.
[0043] In this embodiment, the cloud server uses the self-attention mechanism to achieve temporal feature fusion in multiple scenarios. The attention mechanism measures the degree of association between different positions by calculating the correlation score between each position of the target and all other positions, and then multiplies the input deep feature parameters by the attention parameter matrix to obtain the query, key, and value. Then, using the correlation scores of the query and key, the attention weight is calculated by scaling and normalizing the correlation scores. Finally, the final self-attention representation is obtained by weighted summing of the values using the attention weights, thereby achieving feature fusion of the target.
[0044] In this embodiment, based on the Transformer decoder as the task head of situation cognition, situation cognition recognition can be realized without generating anchor frames, and trajectory prediction can be realized based on the historical trajectory and historical features of the target. The Transformer encoder is used to embed historical trajectory and environmental information into context information and input it into the Transformer. The specific input is target information, historical trajectory, target interaction information, etc., and the output is target situation cognition information.
[0045] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A situation collaborative recognition method based on a distributed large model, characterized in that: include: Acquire N time series data of the target based on N end nodes, wherein each end node acquires one time series data of the target respectively, and different time series data correspond to different scenes of the target; Calculating at least one deep feature parameter of the target based on each of the time series data respectively; specifically comprising: extracting a shallow feature parameter of the target based on each of the time series data respectively; extracting a corresponding deep feature parameter from a shallow feature parameter of the target based on at least one deep feature extraction layer, wherein different deep feature extraction layers use different extraction parameters to extract the deep feature parameters; Before extracting the corresponding deep feature parameter from a shallow feature parameter of the target based on at least one deep feature extraction layer, the method further includes: allocating a shallow feature parameter of the target to at least one corresponding deep feature extraction layer based on a hypernetwork; Allocating a shallow feature parameter of the target to at least one corresponding deep feature extraction layer based on the hypernetwork comprises: allocating a shallow feature parameter of the target to at least one of the deep feature extraction layers based on a gating function; The allocating a shallow feature parameter of the target to at least one corresponding deep feature extraction layer based on the gating function comprises: calculating a weight value of the correlation between a shallow feature parameter of the target and the deep feature extraction layer; allocating a shallow feature parameter of the target to at least one of the deep feature extraction layers according to the weight value; and All deep feature parameters of the target are obtained on the cloud server, and situation awareness of the target is performed based on all the deep feature parameters of the target, where N is an integer greater than or equal to 1.
2. The method according to claim 1, characterized in that The extracting a shallow feature parameter of the target based on each of the time series data comprises: A CNN convolutional neural network is used to extract a shallow feature parameter of the target.
3. The method according to claim 1, characterized in that The extracting corresponding deep feature parameters from a shallow feature parameter of the target based on at least one deep feature extraction layer includes: The at least one deep feature extraction layer extracts corresponding deep feature parameters from a shallow feature parameter of the target based on a standard feed-forward neural network.
4. The method according to claim 1, characterized in that: The performing situation recognition on the target based on all the deep feature parameters of the target includes: Performing feature fusion on all the deep feature parameters of the target to obtain target feature information; and Perform situation awareness on the target based on the target feature information.
5. The method according to claim 4, characterized in that The feature fusion of all the depth feature parameters of the target includes: Based on the self-attention mechanism, all the deep feature parameters of the target are fused.
6. The method according to claim 4, characterized in that The performing situation awareness on the target based on the target feature information comprises: The target feature information is decoded based on the Transformer decoder to achieve situational awareness of the target.