Target action area clustering prediction method based on sequential network model

Abstract

The invention discloses a target action area clustering prediction method based on a sequential network model. The method comprises the following steps: preprocessing original situation data; establishing a target behavior portrait based on the business rule and the specific situation data content, and determining a target behavior feature; mining the predicted target situation data by using a DBSCAN algorithm, and calculating behavior characteristics of all targets through a clustering method; establishing a target behavior prediction network model based on an LSTM structure based on the number of features in the feature set, and determining a training loss function of the network model; training a network model based on an SGD algorithm, and exporting the weight and the network structure of the model to a file for storage after training convergence; and importing the trained model weight, applying the trained model weight to real-time target data, and performing prediction output on the action or area of the target. The method has high prediction accuracy and real-time performance, and can quickly identify and predict the action trend of the target.

Description

Technical Field

[0001] The present invention belongs to the technical field of target situation data analysis, and particularly relates to a method for clustering and predicting target action areas based on a time series network model. Background Art

[0002] In target situation data analysis, predicting the behavior of targets is the key to improving the situation awareness ability and confrontation effect in the adversarial environment. Due to the uncertainty of the behavior patterns of the opposing targets. The opposing targets may adopt different tactics and strategies, and their behavior patterns will change with the changes in the adversarial environment and the situation of both sides in the confrontation. Therefore, accurately capturing and understanding the behavior patterns of the opposing targets is a challenging task. Secondly, the diversity and complexity of the adversarial environment data bring difficulties to the prediction analysis of target behavior. The data involved in the battlefield includes various types such as behavior data, historical records, and environmental factors. How to effectively integrate and utilize these data, extract key features, and establish a model for prediction is crucial for situation awareness analysis.

[0003] The existing research on predicting the target action intention based on adversarial situation data mainly focuses on two types of data: target historical behavior trajectory data and target characteristic data. Among them, for predicting the target action location based on historical trajectory data, usually a target recognition model, a target database, a geographical environment database, etc. are used to detect the specific location of the target during the action process in real time, and a hidden Markov model for accurately predicting the target action route and action purpose is established. Based on this model, information such as the potential behavior purpose of individual targets is inferred. In recent years, deep learning, with the advantages of being able to efficiently process multi-dimensional data, being flexible and highly versatile, has also begun to be applied to the research on predicting the behavior destination of individual targets, such as using technologies such as BP neural networks to improve the prediction accuracy of the behavior intention of individual targets.

[0004] Currently, driven by artificial intelligence and machine learning technologies, target situation analysis is particularly important in many fields such as confrontation, security, and cities. In complex scenarios, traditional analysis methods can no longer meet the requirements for accurately predicting and real-time responding to the target action area. Therefore, it has become urgent and necessary to analyze and predict the target action area by means of advanced data mining technologies and intelligent algorithms. Summary of the Invention

[0005] To achieve the above object, the method for clustering and predicting target action areas based on a time series network model disclosed in the present application includes the following steps: S1: Preprocess the original situation data, including data screening and data cleaning; S2: Establish a target behavior portrait based on business rules and specific situation data content, and determine target behavior characteristics; S3: Use the DBSCAN algorithm to mine the predicted target situation data, calculate the behavior characteristics of all targets through the clustering method, and the results after clustering form a feature data set about the targets for training the subsequent neural network model; S4: Establish a target behavior prediction network model based on the LSTM structure based on the number of features in the feature set, and determine the training loss function of the network model; S5: Train the network model based on the SGD algorithm, and export the weights and network structure of the model to a file for storage after training convergence; S6: Import the trained model weights, apply them to the real-time target data, and predict and output the actions or regions of the targets.

[0006] Further, the original situation data is described mathematically as: ; In the above formula, represents the spatio-temporal characteristic value of the th type of individual action of the target, represents the time period when the spatio-temporal characteristic of the th type of behavior of the target occurs, represents the key area number corresponding to the spatio-temporal characteristic of the th type of behavior of the target, represents the historical behavior frequency corresponding to the spatio-temporal characteristic of the th type of behavior of the target, respectively represent the longitude and latitude coordinates of the key area points corresponding to the spatio-temporal characteristic of the th type of action of the target, represents the action route number corresponding to the spatio-temporal characteristic of the th type of action of the target; The set of longitude and latitude coordinates of the center or key points of the historical action key areas of the target .

[0007] Further, the data is screened against the target intelligence information, the geographical location of the key area, and the relevant business rules, all the target situation data within all historical periods is screened out, the behavior data related to the non-concerned targets needs to be excluded, the data records with the target ID number are retained, and the data records without a clear corresponding target ID are excluded together; The data cleaning process removes duplicate fields within the behavior records to ensure the uniqueness and mutual exclusivity among each behavior record. Moreover, all types of information within each behavior record should remain complete; otherwise, it is regarded as abnormal data and directly excluded. Meanwhile, behavior information that is irrelevant to the spatio-temporal characteristic analysis of the individual's target behavior is excluded.

[0008] Furthermore, step S3 also includes: S31: Mark the initial state of all historical behavior records of each target as True; S32: Randomly extract one behavior record from them , and change its mark to False: Taking as the center point to draw a circle, and traverse all the objects marked as True in the behavior data set corresponding to the target of this behavior record ; S33: Judge the attributes of this behavior record : If the historical behavior frequency within the circle centered at with a radius of , then mark this behavior record as the cluster center point and create the corresponding cluster , take the average value of the target record time within the cluster as the standard time of this cluster , then randomly extract the next behavior record , and transfer to S32 until all the historical behavior records of the target have been traversed, that is, all records are marked as False, and transfer to S34; S34: If the total number of center points marked for the target is 0, it means that the target has no fixed behavior pattern, and directly transfer to S38. If the total number of center points marked for the target is not 0, then transfer to S35; S35: For any two center points in a certain cluster, if the time interval between their records is less than EPS, then connect the two points. The connectable center points and all the points within their neighborhoods form a combined cluster; if a center point is not connected to any other center points, then the center point and all the points within its neighborhood form a separate cluster: The historical behavior records that do not form a cluster are recorded as noise points, indicating the random behavior of the individual target; S36: Output all the clusters generated by the individual target : ; S37: Match the key areas within all the clusters with the coordinate set to obtain the spatio-temporal characteristics of the individual's behavior, where is the track segment number generated according to the set of target action key points:

[0009] Furthermore, integrate and output the target of the individual behavior spatio-temporal characteristic matrix; ; S38: Input the historical behavior record set of the next target ID, transfer to S31, until all target IDs are traversed, and the algorithm ends.

[0010] Furthermore, the target behavior prediction network model based on the LSTM structure includes an encoding network, an LSTM network, and a decoding network; The encoding network structure is composed of 2 layers of DNN neural networks, and the input of the model is the feature vector of the above target: ; Output the encoded intermediate vector , as the input of the subsequent LSTM network layer; The LSTM network analyzes the input using the time series behavior data of the target; is the input information at time t, is the hidden vector of the network input at time t. The LSTM network recursively calls itself and passes the information at time t - 1 to time t, and uses the feature vector given by the encoding network as and inputs them into the LSTM network in sequence. The LSTM network finally integrates to obtain an encoded representation c of the current target state, and this encoded representation is used as the input of the subsequent decoding layer network; The decoding network is composed of a fully connected network, and its output is the input of the LSTM network, and outputs the probability distribution of the next key area or key point of the target . The decoding network generates a specific label probability distribution according to the encoded representation c of the current target state, and selects the label with the highest probability as the prediction result.

[0011] Furthermore, each time the target behavior feature prediction model makes a prediction, it inputs the previous feature vectors into the network model in chronological order, and the model finally gives the probability distribution of the next key area or key point of the target , and uses the actual key points or key areas in the dataset of the One - Hot encoding representation as the label to calculate the loss function: ; During training, the SGD optimizer is used to train on the feature set of the target to obtain the parameters of the model.

[0012] Compared with the existing target situation analysis methods, the target action area clustering prediction method based on the time series network model proposed by the present invention has higher prediction accuracy and real-time performance, and can quickly identify and predict the action trends of targets in massive data, so as to make more scientific and reasonable decisions. Description of the Drawings

[0013] Figure 1 is a flowchart of an embodiment of the present invention; Figure 2 is a processing flowchart for analyzing the target action intention in an embodiment of the present invention; Figure 3 is a flowchart of the DBSCAN mining algorithm used in an embodiment of the present invention; Figure 4 is the encoded network data structure provided by an embodiment of the present invention; Figure 5 is the LSTM network data structure provided by an embodiment of the present invention; Figure 6 is the decoded network data structure provided by an embodiment of the present invention. Detailed Embodiments

[0014] The present invention will be further described below with reference to the accompanying drawings, but the present invention is not limited in any way. Any transformation or replacement made based on the teachings of the present invention falls within the protection scope of the present invention.

[0015] Referring to Figure 1 and Figure 2 , the target action area clustering prediction method based on the time series network model disclosed in the present application includes the following steps: S1: Preprocess the original situation data, including data screening and data cleaning; S2: Establish a target behavior portrait based on business rules and specific situation data content, and determine the target behavior characteristics; S3: Use the DBSCAN algorithm to mine the predicted target situation data, calculate the behavior characteristics of all targets through the clustering method, and the result after clustering forms a feature data set about the targets for training the subsequent neural network model; S4: Establish a target behavior prediction network model based on the LSTM structure based on the number of features in the feature set, and determine the training loss function of the network model; S5: Train the network model based on the SGD algorithm, and export the weights and network structure of the model to a file for storage after training convergence; S6: Import the trained model weights, apply them to the real-time target data, and predict and output the actions or areas of the targets.

[0016] In one embodiment, the process of predicting the target behavior intention based on the LSTM neural network is as follows: Input data: the current behavior state sequence of a certain target to be predicted , and the longitude and latitude coordinate sets of the historical action key areas or key points of the target , the target ; The meanings of the fields of the behavior state information are as follows: (1) , indicating the action time number of the target , (2) , indicating the action area number of the target , (3) , indicating the action track number of the target , ; Output result: , the possible action area number of the target in the future. The overall data processing process can be seen in Figure 2 .

[0017] The spatio-temporal portrait of the target situation is the tagging of the land, sea and air situation data information related to the target situation. Tags are usually highly refined feature systems stipulated by humans. The tag set corresponding to each target can abstract the overall information of the target and jointly constitute an overall description of the target. In other words, the target portrait is a data analysis tool that can accurately mine target features and depict target behaviors. Building a target portrait generally can be divided into three steps: target analysis, system construction, and portrait establishment. The core work is to establish a tag system that reflects target attributes. Applying this concept to the analysis of air situation, the concept of spatio-temporal portrait of target behavior appears. The spatio-temporal portrait of target behavior refers to the tagging of target dynamic behavior information.

[0018] The behavior state information of the target mainly relies on the land, sea and air situation information system and the geographical location information of relevant key areas and key points. Through the land, sea and air situation information system, the dynamic behavior data of a single target can be obtained and stored in real time, that is, the longitude and latitude corresponding to each target ID number or the action records at key points; Based on environmental geographical data combined with the deployment information of both sides of the combat, it can be used to effectively quantify the characteristics of the target action points.

[0019] The above-mentioned situation information cannot be directly used for data analysis and model training, and it is impossible to intuitively analyze the spatio-temporal laws of target behaviors based on this. Therefore, it is necessary to perform tagging processing on the land, sea and air situation data and the geographical location information of key regions or key points, that is, to construct the spatio-temporal portrait of target behaviors under the land, sea and air situation data. Correspondingly, constructing the spatio-temporal portrait of a target individual mainly includes three links: preprocessing of target behavior information, extraction of spatio-temporal characteristic values of target behaviors, and visual expression of target behavior portraits.

[0020] The spatio-temporal characteristics of target behaviors in the adversarial environment situation data refer to the statistical characteristic regularity presented by a single target individual when acting in the same time period and the same spatial interval in history. There are often significant differences in the spatio-temporal characteristics of different target individuals' behaviors. For the convenience of quantitative research and intuitive expression, the spatio-temporal characteristics of target individuals based on the real-time monitoring data of various reconnaissance equipment and the geographical environment information of key regions can be described by mathematical methods as follows: ; In the above formula, represents the spatio-temporal characteristic value of the th type of individual action of target , represents the time period (generally accurate to minutes, specifically defined according to the business scenario) when the th type of spatio-temporal characteristics of target occurs, represents the key region number corresponding to the th type of spatio-temporal characteristics of target , represents the historical behavior frequency corresponding to the th type of spatio-temporal characteristics of target , respectively represent the longitude and latitude coordinates of the key region point corresponding to the th type of spatio-temporal characteristics of target action, represents the action route number corresponding to the th type of spatio-temporal characteristics of target action.

[0021] The spatio-temporal behavior characteristics of a single target may also be different each time it executes a task. Therefore, there may be multiple types of spatio-temporal characteristics of individual behaviors in the historical action records of the target. At this time, the spatio-temporal characteristic matrix of target can be expressed as follows: ; The sources of target behavior information in the adversarial situation scenario include two major categories: the situation information obtained and integrated by the target reconnaissance and detection system and the corresponding geographical location information.

[0022] The target situation data mainly includes the complete information of individual targets, including track information, the type of the target, the action direction of the target, etc. Since data anomalies or incorrect storage may occur during the storage of massive data, and data redundancy that is not very relevant to the study of the behavior patterns of individual targets will also affect the data analysis efficiency, it is necessary to preprocess the original situation data set. In one embodiment, the specific preprocessing rules are as follows: (1) Data screening: With reference to the target intelligence information, the geographical locations of key areas, and relevant business rules, all target situation data within all historical periods are screened out, and the behavior data related to some non-concerned targets such as civilian targets needs to be excluded; the data records with target ID numbers are retained, and the data records without clear corresponding target IDs are excluded together; (2) Data cleaning: If all fields within two or more behavior records are repeated, the corresponding duplicate data needs to be cleaned to ensure the uniqueness and difference between each behavior record; and all types of information within each behavior record must be complete, otherwise it is regarded as abnormal data and directly excluded; at the same time, behavior information that has nothing to do with the spatio-temporal characteristics analysis of individual target behaviors, such as device numbers, also needs to be excluded.

[0023] In summary, the effective data fields for the acquisition requirements of target behavior information for situation data are shown in Table 1. By matching and connecting these two types of data, the spatio-temporal characteristics of the individual behaviors of the behavior targets can be completely represented; Table 1 Effective data fields:

[0024] To accurately construct the spatio-temporal portrait of individual behaviors, it is particularly important to effectively extract reliable spatio-temporal characteristic values of target individuals. Since the number of spatio-temporal characteristics of target individuals is unknown and uncertain, the clustering method that requires the number of categories to be given in advance is not suitable for the extraction of spatio-temporal characteristic values of individual behaviors. The density clustering method DBSCAN (Density-Based Spatial Clustering of Application with Noise) based on high-density connection regions does not require the initial core or the number of clusters to be set in advance. Here, we use the DBSCAN algorithm to extract the spatio-temporal characteristic values of target individuals, which are used as the input for the subsequent spatio-temporal action prediction model.

[0025] The DBSCAN algorithm is a density-based clustering method with noise. It divides the spatial range with sufficient density into clusters, can aggregate clusters of any shape in the data space with noise, and defines a cluster as the largest set of density-connected points. That is, this algorithm assumes that the categories of research objects can be determined by their degree of distribution tightness, and the research objects with similarity will be divided into one cluster.

[0026] This algorithm has two global parameters: 1) The maximum density distance EPS, which refers to the radius of the adjacent area around the target point object; 2) The minimum number of similar points MinPts, which refers to the minimum number of points that should be included inside the adjacent area around the target point object. The value ranges of both can be adjusted according to the changes in the research scenario and there are no fixed values. Based on these two parameters, all the points in the studied spatial database can be classified into three categories: 1) Central points, which refer to the point objects whose number of other point objects in their neighborhood is greater than or equal to MinPts; 2) Edge points, which refer to the point objects whose number of other point objects in their neighborhood is less than MinPts and which are themselves located within the neighborhood of a certain central point; 3) Noise points, which refer to the point objects that neither belong to the central points nor the edge points.

[0027] This application applies the DBSCAN algorithm to the extraction of spatio-temporal characteristic values of target individual behaviors, mainly based on the land, sea and air situation data, and taking the spatio-temporal behaviors of individual targets as the research object. In this method, EPS is used to represent the ideal time interval for the target to move between key points or key regions, to measure whether the movement of the target between certain different key points (or key regions) has time regularity; MinPts represents the minimum number of activities of the target in a certain time period and a certain key region, and is used to determine whether the target behavior has spatial regularity. If the historical behavior frequency of the target in a certain key region within a certain time period then it is considered that the behavior of the target in this time period and this region is random. On the contrary, if then it is considered that the target moves according to a fixed rule in this time period and this spatial region.

[0028] Input data: (1) The ideal time interval EPS and the minimum number of behaviors MinPts; (2) The set of longitude and latitude coordinates of the centers of key regions or key points in the situation data ; (3) The target set in the situation data: ; (4) All historical behavior records of the target , where 、 are the action time and the key point (key region) number corresponding to the th behavior record of the target , .

[0029] Output result: The spatio-temporal characteristic matrix of the behavior of the target ; ; .

[0030] Reference Figure 3 , in one embodiment, the processing flow of the DBSCAN target individual action spatio-temporal characteristic value extraction method improved based on the above land, sea and air confrontation situation data is as follows: STEP 1: Mark all historical behavior records of each target The initial state is True; STEP 2: Randomly extract one behavior record from them , change its mark to False: Take as the center to draw a circle, and traverse all objects marked as True in the behavior dataset corresponding to this behavior record ; STEP 3: Judge the attributes of this behavior record . If the historical behavior frequency within the circle centered on with as the radius , then mark this behavior record as the cluster center point and create the corresponding cluster , take the average value of the target record time within the cluster as the standard time of this cluster , then randomly extract the next behavior record , and transfer to STEP 2 until all historical behavior records of the target are traversed (that is, all records are marked as False), and transfer to STEP 4; STEP 4: If the total number of center points marked by the target is 0, it means that the target has no fixed behavior pattern, and directly transfer to STEP 8: If the total number of center points marked by the target is not 0, then transfer to STEP 5; STEP 5: For any two center points in a certain cluster, if the time interval between their records is less than EPS, then connect the two points. The connectable center points and all points within their neighborhoods form a combined cluster; if a center point is not connected to any other center points, then the center point and all points within its neighborhood form a cluster by themselves: The historical behavior records that do not form a cluster are recorded as noise points, indicating the random behavior of individual targets; STEP6: Output all clusters generated by the individual target ; STEP7: Match the key areas within all clusters with the coordinate set to obtain the spatio-temporal characteristics of individual behaviors, where is the track segment number generated according to the set of target action key points: ; Furthermore, integrate and output the spatio-temporal characteristic matrix of the target ; STEP 8: Input the historical behavior record set of the next target ID, and transfer to STEP1. Until all target IDs are traversed, the algorithm ends.

[0031] The target individual spatio-temporal behavior feature prediction model is a model used to analyze and predict the behavior patterns of individuals within a specific spatio-temporal range. It mainly uses deep learning models (such as DNN networks, LSTM neural networks, etc.) to construct a prediction model, and uses historical data to train the model. After the model training is completed, real-time data is input into the model to predict the behavior patterns of the target individual. In one embodiment, the construction of the present application based on the target individual spatio-temporal behavior feature prediction model includes: (1) Encoding network construction The structure of the prediction network model mainly consists of an encoding layer - an LSTM network layer - a decoding layer. The network structure of the encoding layer is as Figure 4 shown, mainly consisting of 2 layers of DNN neural networks. The input of the model is the feature vector of the above target: ; Output the encoded intermediate vector , as the input of the subsequent LSTM network layer.

[0032] (2) LSTM network layer construction The LSTM network is a special recurrent neural network. Different from general feedforward neural networks, the LSTM can analyze the input using the time series behavior data of the target; Figure 5 Among them, is the input information at time t, is the hidden vector of the network input at time t. We can see that the network will recursively call itself and transfer the information at time t - 1 to time t; This network has a certain memory ability and is suitable for processing time series data. When using this network, the feature vector given by the encoding network is used as input into the network in sequence. The network finally integrates to obtain an encoded representation c of the current target state, and this encoded representation will be used as the input of the subsequent decoding layer network.

[0033] (3) Decoding layer network construction Refer to Figure 6 , the decoding layer network consists of another fully connected network, and its output is the input of the LSTM network, outputting the probability distribution of the next key area or key point of the target , the main goal of this network is to generate a specific label probability distribution according to the encoded representation c of the target's current state. During actual inference, the label with the highest probability is selected as the prediction result.

[0034] In one embodiment, each time the target individual spatio-temporal behavior feature prediction model makes a prediction, the previous feature vectors are sequentially input into the network model in chronological order, and the model finally gives the probability distribution of the next key regions or key points of the target. During training, the actual key points or key regions in the dataset are used in the form of One-Hot encoding representation as labels to calculate the loss function: ; During training, the SGD optimizer is used to train on the feature set of the target to obtain the parameters of the model.

[0035] Compared with the existing target situation analysis methods, the target action area clustering prediction method based on the time series network model proposed by the present invention has higher prediction accuracy and real-time performance, and can quickly identify and predict the action trends of the target in massive data, so as to make more scientific and reasonable decisions.

[0036] As used herein, the term "preferred" is intended to be used as an example, illustration, or exemplification. Any aspect or design described herein as "preferred" should not necessarily be construed as more advantageous than other aspects or designs. On the contrary, the use of the term "preferred" is intended to present concepts in a specific manner. As used in this application, the term "or" is intended to mean an inclusive "or" rather than an exclusive "or". That is, unless otherwise specified or clear from the context, "X uses A or B" is intended to naturally include any one of the permutations. That is, if X uses A; X uses B; or X uses both A and B, then "X uses A or B" is satisfied in any of the foregoing examples.

[0037] Moreover, although the present disclosure has been shown and described with respect to one or more implementations, those skilled in the art will envision equivalent variations and modifications based on reading and understanding this specification and the drawings. The present disclosure includes all such modifications and variations and is limited only by the scope of the appended claims. In particular, with respect to the various functions performed by the above-described components (e.g., elements, etc.), the terms used to describe such components are intended to correspond to any component (unless otherwise indicated) that performs the specified function of the component (e.g., it is functionally equivalent), even if structurally different from the disclosed structure that performs the functions in the exemplary implementations of the present disclosure shown herein. Additionally, although a particular feature of the present disclosure has been disclosed with respect to only one of several implementations, such feature may be combined with one or other features of other implementations as may be desired and advantageous for a given or particular application. Also, insofar as the terms "comprising", "having", "containing", or any variation thereof are used in a particular embodiment or claim, such terms are intended to include in a manner similar to the term "including".

[0038] In the embodiments of the present invention, each functional unit can be integrated into a processing module, or each unit can exist physically alone, or multiple or more than multiple units can be integrated into one module. The above-mentioned integrated module can be implemented in the form of hardware or in the form of a software functional module. When the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The above-mentioned storage medium can be a read-only memory, a magnetic disk or an optical disc, etc. The above-mentioned devices or systems can execute the storage methods in the corresponding method embodiments.

[0039] In summary, the above embodiments are an implementation manner of the present invention, but the implementation manner of the present invention is not limited by the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A method for clustering and predicting target action areas based on a temporal network model, characterized in that, It includes the following steps: S1: Preprocess the original situation data, including data screening and data cleaning; S2: Establish a target behavior portrait based on business rules and specific situation data content, and determine target behavior characteristics; S3: Use the DBSCAN algorithm to mine the predicted target situation data, calculate the behavior characteristics of all targets through the clustering method, and the result after clustering forms a feature data set about the target for training the subsequent neural network model; S4: Establish a target behavior prediction network model based on the LSTM structure based on the number of features in the feature set, and determine the training loss function of the network model; S5: Train the network model based on the SGD algorithm, and export the weights and network structure of the model to a file for storage after training convergence; S6: Import the trained model weights, apply them to real-time target data, and predict and output the actions or regions of the targets.

2. The method for clustering and predicting a target action area based on a temporal network model according to claim 1, wherein The original situation data is described mathematically as: ； In the above formula, represents the spatio-temporal characteristic value of the nth type of individual action of the target, represents the time period during which the spatio-temporal characteristic of the nth type of behavior of the target occurs, represents the key area number corresponding to the spatio-temporal characteristic of the nth type of behavior of the target, represents the historical behavior frequency corresponding to the spatio-temporal characteristic of the nth type of behavior of the target, respectively represent the longitude and latitude coordinates of the key area points corresponding to the spatio-temporal characteristic of the nth type of action of the target, represents the action route number corresponding to the spatio-temporal characteristic of the nth type of action of the target; Target The set of longitude and latitude coordinates of the key area center or key points of historical actions .

3. The method for clustering and predicting target action areas based on a temporal network model according to claim 1, wherein, The data screening compares with the target intelligence information, the geographical location of the key area and relevant business rules, screens out all target situation data within all historical periods, eliminates the behavior data related to non-concerned targets, retains the data records with target ID numbers, and eliminates the data records without corresponding explicit target IDs; The data cleaning cleans the duplicate fields in the behavior records to ensure the mutual exclusivity and uniqueness between each behavior record; and all types of information in each behavior record are kept complete, otherwise it is regarded as abnormal data and directly eliminated; at the same time, eliminate the behavior information irrelevant to the spatio-temporal characteristic analysis of individual target behaviors.

4. The method for clustering and predicting target action regions based on a temporal network model according to claim 2, wherein Step S3 also includes: S31: Mark all historical behavior records of each target The initial state is True; S32: Randomly extract one behavior record from them , change its label to False: Take as the center to draw a circle, and traverse all the objects labeled as True in the behavior dataset corresponding to the target of this behavior record ; S33: Determine the attribute of this behavior record : If the historical behavior frequency within the circle centered at with a radius of , then mark this behavior record as the cluster center point and create the corresponding cluster , and use the average value of the target record time within the cluster as the standard time of this cluster , then randomly select the next behavior record , and transfer to S32 until all historical behavior records are traversed, that is, all behavior records are marked as False, and transfer to S34; S34: If the target has a total of 0 marked center points, it means that the target has no fixed behavior pattern and directly proceeds to S38. If the target has a total of marked center points that is not 0, then proceed to S35; S35: For any two center points in a certain cluster, if the time interval between their records is less than EPS, then connect the two points, and form a combined cluster with the connected center points and all points in their neighborhoods; if a center point is not connected to any other center point, then the center point and all points in its neighborhood form a cluster by themselves: the historical behavior records that do not form a cluster are recorded as noise points, indicating the random behavior of individual targets; S36: Output individual target All generated clusters: ； S37: All clusters within the key area are correspondingly matched with the coordinate set to obtain the spatio-temporal characteristics of individual behaviors, where is the track segment number generated according to the set of target action key points: ； Furthermore, integrate and output the target individual behavior spatio-temporal characteristic matrix; ； S38: Input the historical behavior record set of the next target ID, and transfer to S31 until all target IDs are traversed and the algorithm ends.

5. The method for clustering and predicting a target action area based on a temporal network model according to claim 4, wherein The target behavior prediction network model based on the LSTM structure includes an encoding network, an LSTM network and a decoding network; The encoding network structure consists of 2 layers of DNN neural networks, and the input of the model is the feature vector of the target: ； Output the encoded intermediate vector , as the input to the subsequent LSTM network layer; The LSTM network analyzes the input using the time series behavior data of the target; is the input information at time t, is the hidden vector of the network input at time t. The LSTM network recursively calls itself and passes the information at time t−1 to time t, and uses the feature vector given by the encoding network as and inputs them into the LSTM network in sequence. The LSTM network finally integrates an encoded representation c that reaches the current target state, and this encoded representation is used as the input for the subsequent decoding layer network; The decoding network consists of a fully connected network, and its output is the input of the LSTM network, outputting the probability distribution of the next key region or key point of the target. Based on the encoded representation c of the current state of the target, the decoding network generates a specific label probability distribution and selects the label with the highest probability as the prediction result.

6. The method for clustering and predicting a target action area based on a temporal network model according to claim 5, wherein Each time the target behavior feature prediction model makes a prediction, it sequentially inputs the previous feature vectors into the network model in chronological order, and the model finally gives the probability distribution of the next key regions or key points of the target. , and during training, the actual key points or key regions in the dataset are used represented by the One-Hot encoding as the label to calculate the loss function: ； During training, the SGD optimizer is used to train on the feature set of the target to obtain the parameters of the model.

Images