Intelligent analysis method for target activity law based on sky net big data

Abstract

The present application relates to the technical field of electric digital data processing, and especially relates to a target activity rule intelligent analysis method based on sky net big data, comprising the following steps: constructing a knowledge graph and knowledge organization of a target based on sky net big data; acquiring real-time or non-real-time situation data of the target, and tracking and analyzing the activity situation of the target based on the knowledge graph and knowledge organization; analyzing the activity rule of the target based on the tracking analysis result of the activity situation, the knowledge graph and the knowledge organization, and generating and visually displaying a target portrait. Through the research on related technologies such as target knowledge modeling, knowledge organization, intelligent identification, situation tracking research and judgment, activity track information extraction, activity rule reasoning, target portrait generation and final target portrait visual display, the present application realizes the establishment of the correlation between target activities and space-time events from a large number of fragmented and static data, and the mining of certain activity rules.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an intelligent analysis method for target activity patterns based on Skynet big data. Background Technology

[0002] The analysis of target activity patterns plays a crucial role in target information acquisition and target tracking. Extracting and predicting target activity trajectory information can effectively grasp the target's direction of movement and intention. Traditional target activity analysis usually relies on calculations and inferences based on images collected at different times, resulting in a relatively limited data source and low accuracy in predicting target activity patterns.

[0003] "Sky" refers to Earth observation data such as images and locations acquired through aerospace / aviation / UAV sensors; "Network" refers to various publicly released multimedia data circulating and stored on the Internet. SkyNet Big Data possesses virtually unrestricted data collection capabilities and contains abundant, timely, and valuable data information. With the large-scale application of satellite remote sensing data and open-source network data, using them as data sources for target activity analysis can improve the accuracy of predicting target activity patterns. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent analysis method for target activity patterns based on SkyNet big data, so as to improve the accuracy of analysis and prediction.

[0005] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0006] A method for intelligent analysis of target activity patterns based on Skynet big data includes the following steps:

[0007] Based on the Skynet big data, construct a knowledge graph and knowledge organization for the target;

[0008] Acquire real-time or non-real-time situational data of the target, and track and analyze the target's activity status based on the knowledge graph and knowledge organization;

[0009] Based on the tracking and analysis results of the target's activity status, knowledge graph, and knowledge organization, the activity patterns of the target are analyzed.

[0010] Based on the results of target pattern analysis, activity status tracking analysis, knowledge graph, and knowledge organization, a target profile is generated and visualized.

[0011] Based on an example of an implementable method, the steps for constructing a knowledge graph and knowledge organization of the target based on Skynet big data include:

[0012] S101, collects remote sensing image data and open-source data from the network of the target;

[0013] S102, which maps and merges remote sensing image data and open-source network data with the target as the main body;

[0014] S103 encapsulates and expands the text description of the target and related satellite imagery data based on the OWL language;

[0015] S104. Based on the data obtained from steps S102 and S103, construct the target's knowledge graph and knowledge organization respectively. The knowledge graph is centered on the specified target and contains all the elements of the specified target. The knowledge organization is composed of various targets and contains the element characteristics, time characteristics, spatial characteristics, morphological characteristics, and source characteristics of each target.

[0016] As an example of an implementable method, the tracking and analysis of the target's activity status includes target identification and situation assessment. The target identification uses a densely connected network fused with a YOLOv3 model, which connects feature maps of the same size and only transmits them to the input of the last convolutional layer of the same scale for addition.

[0017] In the above scheme, only feature maps of the same scale are connected and transmitted only to the input of the last convolutional layer of the same scale. This can reduce parameters, improve computational efficiency, and increase the accuracy of target recognition.

[0018] The analysis of the target's activity patterns includes target activity trajectory information extraction and target activity pattern inference; the target activity trajectory information extraction adopts a high-speed multi-kernel convolutional neural network extraction model based on small samples, and the high-speed multi-kernel convolutional neural network extraction model based on small samples includes a sentence embedding module, a package embedding module, and a group embedding module;

[0019] First, the text data in the target's historical activity dataset is converted into vector form through word embedding to obtain the sentence representation vector, which is then used as the input to the sentence embedding module. Next, in the bag embedding module, the representation vectors of all sentences containing the same entity pairs are put into a bag, and the bag representation vector is obtained through an attention mechanism. Subsequently, in the group embedding module, the representation vectors of all bags classified as the same relation type are put into a group, and the group representation vector is obtained through a self-attention mechanism.

[0020] The above scheme employs a high-speed multi-kernel convolutional neural network (HMKN) extraction model based on small samples to address the problem of large corpus size.

[0021] In the feasible implementation, the target activity pattern reasoning adopts a knowledge completion model based on convolutional neural networks. First, the vector representation of the triples is initialized, and all vectors are concatenated to form matrix A. This matrix A is then input into the convolutional layer. Next, in the convolutional layer, row-wise convolution is performed on the triple matrix using filters of 1x3, 3x3, and 5x3 shapes to generate feature extraction of the triples. All feature extractions are then expanded and input into two fully connected layers. Finally, a score representing the rationality of the triple combination is calculated in the fully connected layer. The score with the maximum value is selected as the correct entity of the target activity pattern. The score is calculated as follows: f(h,r,t)=g(vec(g(A*Ω))·w1)·w2, where A represents the matrix representation of the head entity h, relation r, and tail entity t in the triple, * represents convolution, Ω represents a set of filters, g represents the ReLU function, vec represents vectorization, · represents dot product, and w1 and w2 both represent matrices.

[0022] The method described above obtains a reasonable representation of triples through backpropagation training. Based on this representation, it predicts the tail entity given the head entity and relation, or predicts the head entity given the relation and tail entity. This method not only predicts the acquisition pattern of the target but also achieves high accuracy.

[0023] As an example of an implementable method, the generation and visualization of a target profile includes the following steps:

[0024] S401, Select target;

[0025] S402 obtains basic information about the selected target, such as radar equipment, target threat, target location, historical data, real-time radar status, and target image slices from knowledge graphs and knowledge organizations, and constructs a target profile based on this information.

[0026] S403, acquire target data in real time, the data including: the latest detected latitude and longitude of the target and alarm information;

[0027] S404: Set the start and end times of the target trajectory data, query the trajectory data within the specified time period, display the target's array map and heat map in the form of a graph, and display the target trajectory point details in the form of a list;

[0028] S405: Set the start and end times for target real-time radar data, query the real-time radar data within that time period, and display it in a list;

[0029] S406: Set the start and end times for the target image slice data, query the database for the target image slice data within that time period, and display it in the form of a graph.

[0030] Compared with existing technologies, this invention, through research on related technologies such as target knowledge modeling, knowledge organization, intelligent recognition, situational tracking and judgment, activity trajectory information extraction, activity pattern reasoning, target profile generation, and final target profile visualization, enables the establishment of correlations between target activities and spatiotemporal events from a large amount of fragmented and static data, the discovery of certain activity patterns, and the creation of multi-dimensional profiles of target activity information, providing auxiliary support for command and decision-making. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart of the intelligent analysis method for target activity patterns based on SkyNet big data in the embodiment.

[0033] Figure 2 This is a flowchart illustrating the process of constructing a knowledge graph and organizing knowledge based on Skynet big data in this embodiment.

[0034] Figure 3 This is a flowchart illustrating the process of encapsulating and expanding the text description of the target and related satellite imagery data based on the OWL (Web Ontology Language) language in this embodiment.

[0035] Figure 4 This is a schematic diagram of the knowledge graph of the target in the embodiment.

[0036] Figure 5 This is a schematic diagram illustrating the knowledge organization of the target in the embodiment.

[0037] Figure 6 This is a schematic diagram of residual blocks in a densely connected network in the embodiment.

[0038] Figure 7 This is a flowchart of the situation assessment process in the embodiment.

[0039] Figure 8 This is a flowchart illustrating the process of extracting target activity trajectory information in this embodiment.

[0040] Figure 9 This is a schematic diagram of a high-speed multi-kernel convolutional neural network for the sentence feature extraction module in the embodiment.

[0041] Figure 10 This is a flowchart illustrating the process of reasoning about the target activity patterns in this embodiment.

[0042] Figure 11 This is a flowchart illustrating the process of generating and visualizing the target image in this embodiment. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0044] Please see Figure 1 This embodiment provides an intelligent analysis method for target activity patterns based on Skynet big data, which includes the following steps:

[0045] S10, based on Skynet big data, constructs a knowledge graph and knowledge organization for the target.

[0046] S20: Obtain real-time or non-real-time situational data of the target, and track and analyze the target's activity status based on the knowledge graph and knowledge organization.

[0047] S30, based on the tracking and analysis results of the target's activity status, knowledge graph, and knowledge organization, analyzes the activity patterns of the target and mines information such as the activity area and positional distribution patterns of the target or target group. A target group refers to a group composed of multiple targets.

[0048] S40 generates a target profile and displays it visually based on the results of target pattern analysis, activity status tracking analysis, knowledge graph, and knowledge organization.

[0049] The above method utilizes remote sensing image data (Sky Data) and open-source network data (Network Data) to construct a knowledge graph and knowledge organization for the target. Based on the knowledge graph and knowledge organization and real-time SkyNet big data, it generates a target profile. The data sources are rich and unrestricted, so the target profile is more accurate and the analysis of target activity patterns is more comprehensive.

[0050] Please see Figure 2 As an example of an implementation method, step S10 above may include the following steps:

[0051] S101 collects remote sensing imagery data and open-source data from the network for the target.

[0052] S102 maps and merges remote sensing image data and open-source network data with the target as the main body.

[0053] S103 encapsulates and expands the textual description of the target and related satellite imagery data based on the OWL (Web Ontology Language) language.

[0054] S104. Based on the data obtained from steps S102 and S103, construct the target's knowledge graph and knowledge organization respectively.

[0055] In step S102 above, the remote sensing image data and open-source network data are mapped and merged. The specific processing procedure can be found in Chinese Patent Publication No. CN114996488B, entitled "A Decision-Level Fusion Method for Skynet Big Data," or other existing technologies. This embodiment will not elaborate on the details here. For example, for the same aircraft target, its satellite remote sensing image data and open-source network data are mapped and merged to mutually verify each other, thereby improving data accuracy.

[0056] When mapping data, spatiotemporal relationships can be defined, as shown in the table below.

[0057]

[0058] In step S103 above, as an example of an implementable method, based on Figure 3 The encapsulation and expansion are performed as shown. Figure 3 In this context, a domain is a division of the scope of knowledge. The target knowledge graph consists of several subgraphs, each of which can be considered a domain, and domains can be directly referenced. An attribute block is a collection of attributes, forming a multi-layered structure. The target knowledge graph includes an appearance attribute block, which consists of length, width, and height. Association classes are classes that handle complex structures, but these classes do not need explicit naming in business logic and often depend on ordinary classes. For example, a mounting scheme class depends on an aircraft class and consists of mounting points, mounts, and mount bodies, which is a typical association class. Views are virtual graph structures that selectively reorganize the original knowledge graph to form a more usable graph form. Multiple inheritance refers to a class inheriting from multiple classes simultaneously.

[0059] It should be noted that step S103 is an optional step to improve the completeness of the knowledge graph and knowledge organization construction. Essentially, it involves first building a relatively complete knowledge structure based on the current data. Subsequent data updates can continuously refine the specific content within the structure or add more element data to improve the structure, making the knowledge graph and knowledge organization more complete and the information content richer, thereby improving the accuracy of subsequent activity pattern analysis.

[0060] For example, the target could be an aircraft (airplane), a ship, etc., with separate knowledge graphs built for each. For example, the knowledge graph for an aircraft would look like this: Figure 4 As shown. However, it should be noted that the target's knowledge graph includes not only categories, but also all elements of the target such as type, country of origin, dimensions, weight, structure, and permanent location. It should also be noted that the knowledge graph is constructed centered around a specific target (such as a drone). Figure 4 This represents a broad category; drones are... Figure 4 The fourth level of the organizational structure shown.

[0061] The target ontology gradually develops a series of characteristics during its lifecycle, which can be summarized into four main characteristics: essential characteristics, morphological characteristics, source characteristics, and other characteristics. Among them, essential characteristics include element characteristics, temporal characteristics, and spatial characteristics; essential characteristics are the attributes of the target's essential connotation, and different data resources have different essential characteristics.

[0062] Element characteristics: The result of abstracting and simplifying the spatial location of the environment, phenomena and events distributed within a certain time range and their spatial relationships within the target data scope is called an element.

[0063] Temporal characteristics: SkyNet Big Data itself possesses rich temporal information. The metadata, datasets, and data items of SkyNet Big Data all contain abundant temporal information during the processes of data collection, processing, updating, storage, sharing, and distribution.

[0064] Spatial features: Spatial features refer to the location of spatial entities through a series of spatial coordinates. Spatial features also include the geometric characteristics of spatial entities, such as their shape, size, and spatial development patterns, as well as the spatial relationships used to describe the connections between spatial entities.

[0065] Morphological characteristics: The morphological characteristics of SkyNet Big Data are manifested in the external form and additional features of the data, describing the structure, format, storage, benchmarks, and other contents of SkyNet Big Data.

[0066] Source characteristics: The record of the data source is a series of data origins and the process that the data goes through from its generation to its final entry into the database (target knowledge graph).

[0067] Other characteristics: These mainly include data sharing service characteristics and data utilization characteristics. Among them, the sharing service characteristics mainly include service status and service methods, while the data utilization characteristics mainly include timeliness restrictions, scope restrictions, and accuracy restrictions.

[0068] The knowledge organization of objectives involves organizing various objectives based on the characteristics mentioned above. The structure of knowledge organization consists of these characteristic dimensions, such as time characteristics. From the perspective of time content, there are three aspects: first, event time (effective time), which refers to the time when a phenomenon or event occurs or exists in reality, such as the time a drone spends in Okinawa; second, entry time (transaction time), which refers to the time when an event is recorded, updated, or deleted in the database, such as the time the drone's stay in Okinawa is entered into the database; and third, recording time, which refers to the time range within which a phenomenon or event is observed, collected, and used to form data records or datasets. As an example, the knowledge organization of objectives might look like this: Figure 5 As shown, Figure 5 The target in the middle is a ship.

[0069] In step S20 above, the process of tracking and analyzing the target's activity status includes target identification and situation assessment.

[0070] Target recognition involves identifying the type of target based on input satellite remote sensing image data, and then further identifying the target model by combining a constructed knowledge graph and knowledge organization. For target recognition, this embodiment employs a densely connected network fusion of YOLOv3, referred to as D-YOLOv3. This method adds feature maps of the same size along the last dimension and then performs scale reduction to increase detailed information beneficial for target detection and recognition, while also reducing computational load.

[0071] Specifically, feature maps of the same size in YOLOv3 are passed across layers, stacking feature maps of the same size and dimension as input to the next convolutional layer with a stride of 2. Since the purpose of dense connections is to better utilize shallow features, and dense connections require feature maps to have the same dimension, traditional densely connected networks pass the output of each convolutional layer to every subsequent convolutional layer. However, directly passing the input of each layer in YOLOv3 to every subsequent layer would increase the number of parameters in the entire network. Therefore, this embodiment improves upon this by only connecting feature maps of the same scale and only passing them to the input of the last convolutional layer of that scale. This reduces the number of parameters, improves computational efficiency, and increases the accuracy of target recognition.

[0072] like Figure 6The diagram illustrates residual blocks in a densely connected network in an example. In this example, the output of the 5th convolutional layer is passed to the 10th layer, and the outputs of the 10th, 12th, 14th, 16th, 18th, 20th, and 22nd convolutional layers are passed to the 27th layer. The numbers on the curves in the diagram correspond to the outputs of the 5th, 10th, 12th, 14th, 16th, 18th, 20th, and 22nd convolutional layers, respectively. In the diagram, Res represents a residual block, and each residual block contains two convolutional layers, so the number of convolutional layers increases by 2 after each Res block. Conv represents a convolutional layer with a stride of 2.

[0073] Situation assessment is based on target identification results, historical and real-time situational data (situational data consists of regular data composed of individual points, such as ADS data of flight aircraft, which expresses an evolutionary process), and multi-source information data related to the target (i.e., network data). According to a set situational template, the current situation of a specified target is generated. Furthermore, the distribution of the specified target can be visualized on a 3D digital globe using labeled symbols, displaying the current situation of the specified target, and providing detailed data on the thematic situational analysis in a list format.

[0074] like Figure 7 As shown, the situation assessment process includes the following steps:

[0075] S201. Select search criteria such as region, country, and time range to filter out targets that meet the search criteria from knowledge graphs and knowledge organizations, respectively.

[0076] S202, Obtain the historical status of the targets selected above, real-time news information about the targets, introductory information about the targets, past activity routes of the targets, and information on the targets' intentions from the knowledge graph and knowledge organization.

[0077] S203, analyze the target's activity trajectory data and activity status data within the time range (selected in step S201). The activity trajectory data refers to the target's historical usual route, while the activity status data refers to the places the target has already passed through and where it is currently parked. Simply put, one is historical data and the other is real-time data.

[0078] S204: Based on the activity trajectory data and activity status data obtained in step S203, generate and display a target monitoring situation report. The report includes: country, target name, target category, target level, location, start time, end time, and duration. The target monitoring situation report can also be exported to a local computer or saved to a database.

[0079] In step S30 above, the processing of analyzing the target's activity patterns based on the tracking and analysis results of the target's activity status, knowledge graph, and knowledge organization includes extracting target activity trajectory information and deducing target activity patterns.

[0080] For the extraction of target activity trajectory information, this embodiment adopts a high-speed multi-kernel convolutional neural network (HMKN) extraction model based on small samples to solve the problem of large corpus size.

[0081] Specifically, such as Figure 8 As shown, the text data in the target's historical activity dataset is first converted into vector form using word embedding to obtain sentence representation vectors, which are then used as input to the sentence embedding module. Next, in the bag embedding module, the representation vectors of all sentences containing the same entity pairs are placed into a bag, and the bag's representation vector is obtained through an attention mechanism. Subsequently, in the group embedding module, the representation vectors of all bags classified as the same relation type are placed into a group, and the group's representation vector is obtained through a self-attention mechanism. The historical activity dataset refers to the target's past activity routes and frequently visited resupply points during its journey, obtained from the target's activity status tracking analysis, knowledge graph, and knowledge organization.

[0082] refer to Figure 8 ,set up This represents the i-th packet in a group. This represents the representation vector of the j-th sentence in the i-th packet, with a sentence length of . , , Representing the dimension, the matrix As the input to the sentence in the high-speed multi-kernel convolutional neural network model, S is the representation vector of the sentence after feature extraction, B is the representation vector of the bag, and G is the representation vector of the group.

[0083] By using word embedding, the text data in the target's historical activity dataset is transformed into vector form to obtain sentence representation vectors. Specifically, the process could involve mapping each word of each sentence in the target's historical activity dataset to a word embedding vector. The word representation vectors are 3D, and position features (PFs) are applied to HMKN. For each word in the sentence, PFs describe the relative distance between the current word and the preceding and following entity words, generating two... The three vectors enrich the semantic features of the target entity. Finally, these three vectors are concatenated into a single vector. A dimensional vector serves as the representation vector for a word.

[0084] For the sentence feature extraction part, use Figure 9The sentence feature extraction module shown is a high-speed multi-kernel convolutional neural network, and the word representation vectors form a matrix. Feature extraction was performed using three convolutional kernels of different sizes, and the resulting convolutional matrices were obtained respectively. , , Adding these three output matrices together gives... and will As the result of the nonlinear transformation of the word representation vector, a gating mechanism, namely two nonlinear transformation layers, transformation gate T and pass gate C, is used to weight and sum the linear and nonlinear transformation results of the word vector features, and the sum is used as the sentence representation vector.

[0085] For input Its conversion gate is:

[0086]

[0087] Where b represents a group and w represents a matrix.

[0088] The result output by high-speed multi-kernel convolution is:

[0089]

[0090] For ease of calculation, we define C = 1 - T, that is, the above formula is:

[0091]

[0092] in , T and C are ensured to have the same dimensions through operations such as completion or linear transformation, making the formula application highly flexible. When T=0, This means that the original information is preserved without undergoing multi-kernel convolution transformation. When T=1, The original information undergoes a multi-kernel convolution transformation, and no original information is retained. This is equivalent to the function of a regular convolutional neural network.

[0093]

[0094] Then, based on the positions of the two target entities in the sentence, the output of the high-speed multi-kernel convolutional neural network is... The sentence is divided into three parts, and three max-pooling layers are used to extract features from each part to obtain the sentence representation vector. Let the representation vector of all sentences in a package be a matrix. The relation represents a vector as a matrix. Where h is the number of relation types, Indicates a group, Let be the number of sentences in the i-th packet.

[0095] After completing the word vector embedding, the relation representation vector will be... Incorporating all relationships into the attention mechanism calculation as query vectors can effectively improve the final classification performance compared to considering only one relationship.

[0096] First, calculate the matching score between the sentences in the package and each relation: , where r k Relational representation vector The k-th row vector, Let be the representation vector of the j-th sentence in the i-th bag within the group. To ultimately compute the bag representation vector using the attention mechanism, different sentences within a bag need to be assigned different weights. The weights of sentences within a bag are defined as follows: Weight This reflects the relevance between the j-th sentence and the k-th relation in the i-th packet. The representation vector of the i-th packet in each group is: The representation vector b of all packages i Form a matrix , where k∈{1,2,...,k} represents the relation index.

[0097] After completing the packet embedding representation, to address inter-packet noise, an inter-packet attention mechanism is introduced to reduce the weight of noisy packets. If all packets expressing the same relationship are grouped together, the representation vectors of all packets within a group should have a relatively high similarity. Therefore, during the calculation, for each packet within a group, its similarity is calculated with all other packets in the group, and the similarity score is added to the packet weight calculation. Furthermore, in calculating the sum of packet weights to obtain the group's representation vector, attention weights are only calculated for positive packets, ignoring packets marked as unrelated.

[0098] During attention calculation, a self-attention algorithm is used, meaning the attention weights for different packets within a group are calculated using the representation vectors of all packets within that group. Specifically, each packet's own self-matching score is calculated first.

[0099]

[0100]

[0101] Each x ik This reflects the confidence level between the i-th package and the k-th relation. The weight of the package within the group is defined as follows: Therefore, the final representation vector of the group can be obtained as follows: All gk Form a matrix g k yes The kth row.

[0102] In the training phase of the high-speed multi-kernel convolutional neural network extraction model, each group is used as the basic unit for model training. The score for classifying a group into a specific category label is calculated using the relation vector matrix and the group's representation vector. k : , where d k This is the bias term, which is ultimately processed by a softmax classifier to obtain the probability of classifying a group as a certain relation. The objective function of the entire high-speed multi-kernel convolutional neural network extraction model is set as follows: , where T is the set of all training samples, This is the set of parameters for the model, including word embedding matrices, positional feature embedding matrices, convolutional neural network weight matrices, and relation embedding matrices. The objective function of the model is optimized using minimum gradient descent. The parameters of the model are then determined by the descent.

[0103] For reasoning about the patterns of target activities, this embodiment proposes a new architecture based on convolutional neural networks to solve the knowledge completion task, namely, a knowledge completion model based on convolutional neural networks. The idea is to treat a triple as an "image" and use filters of 1x3, 3x3 and 5x3 shapes to extract the overall feature information in the triple.

[0104] Specifically, such as Figure 10 As shown, the vector representation of the triples is initialized by concatenating all vectors (head entity h, relation r, and tail entity t) to form matrix A, and this matrix representation is input into a convolutional layer. Then, in the convolutional layer, row-wise convolutions are performed on the triple matrix using filters of 1x3, 3x3, and 5x3 shapes to generate feature extractions of the triples. At the same time, all feature extractions are expanded and input into two fully connected layers. Finally, a score is calculated in the fully connected layer. This score represents the reasonableness of the triple. That is, if it is a triple in the training samples, then the score of the triple should be as large as possible, and conversely, the score of a fake sample should be as small as possible.

[0105] The score is calculated as follows: f(h,r,t)=g(vec(g(A*Ω))·w1)·w2

[0106] Where A represents the matrix representation of the head entity h, relation r, and tail entity t in the triple, * represents convolution, Ω represents a set of filters, g represents the ReLU function, vec represents vectorization, · represents dot product, and w1 and w2 both represent matrices.

[0107] The above method obtains a reasonable representation of triples through backpropagation training. Based on this representation, it predicts the tail entity given the head entity and relation, or predicts the head entity given the relation and tail entity. That is, it obtains a score by combining all candidate entities and the given entity and relation using the above method. The entity with the highest score is the correct entity for predicting the target's activity pattern; the correct entity refers to the next place the target will go. This method not only predicts the target's acquisition pattern but also achieves high accuracy.

[0108] In step S40 above, the process of generating a target profile and visualizing it based on the target's pattern analysis results, activity status tracking analysis results, knowledge graph, and knowledge organization includes target profile generation and target profile visualization.

[0109] Based on the SkyNet big data target profile generation and display technology, the system uses data tracking to characterize the target profile. The system tags the target and continuously refines the target profile features through iterative updates of SkyNet big data. It constructs a target model with SkyNet big data characteristics and uses collaborative filtering and artificial intelligence algorithms to characterize the target profile in multiple dimensions. This achieves accurate definition of target labels based on target categories. During use, the target profile is gradually clarified, and effective data tracking is implemented, along with trigger collection and knowledge data weighting based on the data tracking.

[0110] The target profile visualization displays the target's basic information and movement trajectory in a table format. By setting the start and end times of the target trajectory data, the trajectory data within that time period is displayed. The target's position map and heat map are displayed in a graph format, and the target trajectory point details are displayed in a list format.

[0111] like Figure 11 As shown, as an example of an implementable method, the generation and visualization of a target profile includes the following steps:

[0112] S401, Select Target. For example, select a target by its name, or select a target directly from the target tree control.

[0113] S402 retrieves basic information about the target (such as target name, country, aircraft number, etc.), radar equipment, target threat, target location, etc., as well as historical data, actual radar situation, target image slices, etc. from the target's knowledge graph and knowledge organization to construct a target profile.

[0114] S403 acquires target data in real time, including the latest detected latitude and longitude of the target and alarm information. Alarm details can be viewed, and real-time detection data is displayed in a statistical chart format.

[0115] S404 sets the start and end times of the target trajectory data. Based on the target's pattern analysis results and activity status tracking analysis results, it queries the trajectory data within this time period and displays the target's position map and heat map in the form of a graph, and displays the target trajectory point details data in the form of a list.

[0116] S405 sets the start and end times of the target's real-time radar data, queries the real-time radar data within that time period, and displays it in a list.

[0117] S406: Set the start and end times for the target image slice data, query the database for the target image slice data within that time period, and display it in the form of a graph.

[0118] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1.A method for intelligent analysis of target activity regularity based on big data of Tianwang, characterized in that, The method comprises the following steps: constructing a knowledge graph and a knowledge organization of the target based on the sky net big data; acquiring real-time or non-real-time situation data of the target, and tracking and analyzing the activity situation of the target based on the knowledge graph and the knowledge organization; analyzing the activity rule of the target based on the tracking analysis result of the activity situation of the target, the knowledge graph and the knowledge organization; the tracking and analyzing of the activity situation of the target comprises target identification and situation analysis, and the target identification adopts a dense connection network to fuse a YOLOv3 model, connects feature maps of the same size, and only transmits to the input end of the last convolution layer of the same scale for addition; the processing of the situation analysis comprises the following steps: S201, setting a search condition, and screening the target meeting the search condition from the knowledge graph and the knowledge organization respectively; S202, acquiring the historical situation data of the screened target from the knowledge graph and the knowledge organization; S203, analyzing the activity track data and the activity state data of the target in an input time period; S204, generating a tracking special situation report of the target based on the activity track data and the activity state data obtained in step S203, and displaying the tracking special situation report; generating a target portrait and visualizing and displaying the target portrait based on the rule analysis result of the target, the tracking analysis result of the activity situation of the target, the knowledge graph and the knowledge organization; the generation and visualizing and displaying of the target portrait comprise the following steps: S401, selecting a target; S402, acquiring the basic information, radar equipment, target threat and target position of the selected target from the knowledge graph and the knowledge organization, and historical data, real detection radar condition and target image slice, and constructing a target portrait based on the information; S403, acquiring the data of the target in real time, wherein the data comprises the latest detection latitude and longitude of the target and alarm information; S404, setting the start time and the end time of the target track data, querying the track data in the time period, and displaying the position map and the heat map of the target in the form of a graph, and displaying the track point detail data of the target in the form of a list; S405, setting the start time and the end time of the target real detection radar data, querying the real detection radar data in the time period, and displaying the real detection radar data in the form of a list; S406, setting the start time and the end time of the target image slice data, querying the image slice data of the target in the time period from the database, and displaying the image slice data in the form of a graph. 2.The method of claim 1, wherein, The step of constructing the knowledge graph and the knowledge organization of the target based on the sky net big data comprises: S101, collecting remote sensing image data and network open source data of the target; S102, mapping and merging the remote sensing image data and the network open source data based on the target; S103, encapsulating and extending the text description of the target and the related satellite image data based on the OWL language; S104, constructing the knowledge graph and the knowledge organization of the target based on the data processed in steps S102 and S103 respectively; the knowledge graph is centered on a specified target, and comprises full-factor information of the specified target; the knowledge organization is composed of various targets, and comprises factor characteristics, time characteristics, spatial characteristics, morphological characteristics and source characteristics of each target. 3.The method of claim 1, wherein, The analyzing of the activity rule of the target comprises target activity track information extraction and target activity rule inference; the target activity track information extraction adopts a high-speed multi-core convolutional neural network extraction model based on a small sample, which comprises a sentence embedding module, a package embedding module and a group embedding module; Firstly, the text data in the historical activity data set of the target is converted into a vector form by a word embedding method, and a sentence representation vector is obtained, and the representation vector is taken as the input of the sentence embedding module; then, in the package embedding module, the representation vectors of all sentences containing the same entity pair are put into a package, and the representation vector of the package is obtained through an attention mechanism; subsequently, in the group embedding module, the representation vectors of all packages classified as the same relationship type are put into a group, and the representation vector of the group is obtained through a self-attention mechanism. 4.The method of claim 3, wherein, In the training phase of the high-speed multi-core convolutional neural network extraction model, each group is taken as a basic unit for model training, and the score o of classifying a certain group into a certain category label is calculated through the relationship vector matrix and the representation vector of the group k : ； wherein r k is the kth row vector of the relation embedding matrix, g k is the vector representation of the group, d k is the bias term, and finally the score function is processed by a softmax classifier to obtain the probability of classifying a certain group into a certain relation: The objective function of the entire high-speed multi-core convolutional neural network extraction model is set as: wherein T is the set of all training samples, is the parameter set of the high-speed multi-core convolutional neural network extraction model, the parameters including the word embedding matrix, the position feature embedding matrix, the convolutional neural network weight matrix and the relation embedding matrix, and the objective function of the high-speed multi-core convolutional neural network extraction model is made to decrease by the minimum gradient descent method to further determine the model parameters. 5.The method of claim 3, wherein, The target activity rule inference adopts a knowledge completion model based on a convolutional neural network, first, the vector representation of the triplets is initialized, all vectors are spliced together to form a matrix A, and the matrix A is input to the convolution layer; then, in the convolution layer, 1x3, 3x3 and 5x3 filters are used to perform row-by-row convolution on the triplet matrix to generate triplet feature extraction, and the feature extraction is input to two fully connected layers; finally, in the fully connected layer, a score representing the rationality of the triplet is calculated, the maximum score is selected as the correct entity of the target activity rule, and the score is calculated in the following manner: f(h,r,t)=g(vec(g(A*Ω))·w1)·w2, A represents the matrix representation of the head entity h, the relationship r and the tail entity t in the triplet, * represents convolution, Ω represents a group of filters, g represents a Relu function, vec represents vectorization, · represents dot product, and w1 and w2 represent matrices.

Images