Target person dynamic risk early warning method based on trajectory prediction and related equipment
By constructing knowledge graphs and deep learning models, the future behavior and risks of target individuals are predicted, solving the problem of delayed early warning in existing technologies and achieving efficient and accurate risk warning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TIANJIN ZHONGHUAN SYST ENG CO LTD
- Filing Date
- 2022-12-09
- Publication Date
- 2026-06-23
AI Technical Summary
Existing technologies are insufficient for predicting potential risks in the management of target personnel, resulting in a certain lag in early warning and an inability to prevent problems before they occur.
This study employs a trajectory prediction-based approach. By acquiring multi-source heterogeneous data on target individuals, a knowledge graph is constructed. This graph is then combined with a deep learning encoder-decoder framework and a long short-term memory network to predict the future behavior and risks of target individuals. Finally, a risk assessment model is used to output a risk score.
It enables effective early warning of potential risks to target personnel, improves the timeliness and accuracy of early warning, and solves the problem of lag in traditional early warning technology solutions.
Smart Images
Figure CN116128282B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of machine learning technology, and in particular to a method and related equipment for dynamic risk warning of target personnel based on trajectory prediction. Background Technology
[0002] With the rapid development of smart cities and the growing urban population, passive urban security solutions have become obsolete, and the tracking of at-risk individuals, as an important part of urban security, faces new challenges.
[0003] Currently, in terms of target personnel management, the analysis of massive amounts of information on risky personnel mainly relies on fixed tactics or manual analysis, which makes it difficult to predict potential risks. This results in risky behaviors occurring before warnings are issued, and warnings have a certain lag, failing to prevent problems before they occur.
[0004] Therefore, there is an urgent need for a predictive and early warning method to achieve effective early warning and control of potential risks to target personnel. Summary of the Invention
[0005] In view of this, the purpose of this application is to propose a dynamic risk early warning method for target personnel based on trajectory prediction, including:
[0006] Obtain data information of the target personnel;
[0007] The data information is divided into basic information and historical time-series feature information;
[0008] A target personnel knowledge graph is constructed based on the basic information, and a target personnel feature value list is determined based on the target personnel knowledge graph.
[0009] The historical time series feature information and the target personnel feature values are merged to obtain a list of historical time series feature values, which is then input into a pre-built time series prediction model. The time series prediction model outputs a list of future time series feature values, which includes both the predicted historical time series feature information and the predicted target personnel feature values.
[0010] Based on the list of future time period feature values, the risk score of the target person is obtained through a pre-constructed risk assessment model.
[0011] Based on the same inventive concept, this application also provides a target personnel dynamic risk early warning device based on trajectory prediction, comprising:
[0012] The acquisition module is configured to acquire data information of the target personnel.
[0013] The segmentation module is configured to divide the data information into basic information and historical time-series feature information;
[0014] The construction module is configured to construct a target personnel knowledge graph based on the basic information, and determine a target personnel feature value list based on the target personnel knowledge graph;
[0015] The merging module is configured to merge the historical time-series feature information and the target personnel feature values to obtain a list of historical time period feature values;
[0016] The prediction module is configured to input the historical time series feature information into a pre-built time series prediction model, and output a list of feature values for future time periods through the time series prediction model.
[0017] The assessment module is configured to obtain the risk score of the target person based on the list of feature values for the future time period through a pre-built risk assessment model.
[0018] Based on the same inventive concept, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.
[0019] Based on the same inventive concept, this disclosure also provides a non-transitory computer-readable storage medium that stores computer instructions for causing a computer to perform the method described above.
[0020] As described above, the target personnel dynamic risk early warning method and related equipment based on trajectory prediction provided in this application, based on a deep learning encoder-decoder framework, predicts the temporal attributes of target personnel, transforming the problem of predicting future behavior of target personnel into a problem of "predicting a variable-length output given a variable-length input." A Long Short-Term Memory (LSTM) deep neural network is applied to encode, parse, and decode the variable-length input to generate a variable-length output. Based on the temporal change pattern of given characteristics of target personnel over a historical period, the method predicts the future trend of these characteristics and provides early warnings of the dynamic risks of target personnel based on the predicted information. Combining the predicted dynamic information of target personnel with temporal characteristics and other basic information of the target personnel extracted from big data, each target personnel is treated as a sample. Labeled data for each category of target personnel is generated based on risk indicators. A target personnel dynamic risk early warning model is constructed based on a machine learning classification algorithm. The labeled data is used as the training set to train the target personnel dynamic risk early warning model, ultimately outputting the target personnel dynamic risk early warning model. The model's output is the target personnel risk score, and the warning threshold can be adjusted according to the actual situation to ultimately determine whether to issue a warning. This application predicts the time-series characteristics of target personnel and, in conjunction with the basic information of the target personnel, anticipates their potential risks in future periods. It takes into account the timeliness, accuracy, and interpretability of the prediction and early warning results, and solves the problem of lag in traditional early warning technology solutions.
[0021] This application utilizes knowledge graph technology to mine the relationships between target individuals and their related personnel, uncovering deeper, hidden factors. Furthermore, compared to traditional single-time-series coding prediction methods, this application combines highly variable time-series attributes with relatively stable basic attributes. This not only increases the dimensionality and improves accuracy but also allows for the prediction and updating of the relatively stable basic attributes. This application predicts the time-series characteristics of target individuals and, combined with their basic information, anticipates their potential risks in future time periods, balancing the timeliness, accuracy, and interpretability of the prediction and early warning results, thus solving the problem of lag in traditional early warning technologies. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1This is a flowchart illustrating the dynamic risk warning method for target personnel based on trajectory prediction, as described in an embodiment of this application.
[0024] Figure 2 This is a schematic diagram of the first knowledge graph in an embodiment of this application;
[0025] Figure 3 This is a schematic diagram illustrating the knowledge graph completion process in an embodiment of this application.
[0026] Figure 4 This is a schematic diagram of the prediction process of the time series prediction model in an embodiment of this application;
[0027] Figure 5 This is a structural diagram of the target personnel dynamic risk early warning method device based on trajectory prediction according to an embodiment of this application;
[0028] Figure 6 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation
[0029] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.
[0030] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0031] As described in the background technology, based on historical experience in managing target personnel, their behavior exhibits certain periodicity and regularity. Inspired by this, predicting the movement trajectory of target personnel and integrating historical big data information for feature analysis will drive the transformation from traditional post-event analysis and judgment to pre-event prediction and early warning, thereby achieving effective early warning and control of potential risks to target personnel.
[0032] The embodiments of this application will be described in detail below with reference to the accompanying drawings.
[0033] This application provides a dynamic risk early warning method for target personnel based on trajectory prediction, referencing... Figure 1 It includes the following steps:
[0034] Step S101: Obtain data information of the target personnel.
[0035] Acquire massive amounts of information resources related to the target personnel, including multi-source heterogeneous data such as basic information, historical information, daily activity information, vehicle checkpoint information, civil aviation information, railway information, passenger transport information, accommodation information, violation information, and case and incident information.
[0036] Step S102: Divide the data information into basic information and historical time-series feature information. In this embodiment, the basic information mainly includes factors related to personal and family circumstances. Historical time-series feature information may have attributes with different time-series dimensions, such as periodicity, continuous growth, continuous decline, or irregular changes. By analyzing historical time-series feature information, the trajectory of a target individual over a future period can be predicted. When the periodic historical time-series feature information changes, for example, if the periodicity is disrupted, it indicates that the target individual's characteristic has become abnormal. This allows for continuous monitoring of the target individual and early prevention of potential risky behaviors.
[0037] Step S103: Construct a target personnel knowledge graph based on the basic information, and determine a target personnel feature value list based on the target personnel knowledge graph.
[0038] The construction of a knowledge graph refers to the process of extracting knowledge elements from raw data using a series of automatic or semi-automatic techniques and storing them in a knowledge base. The key technology of knowledge graphs is knowledge extraction, also known as triplet extraction. This technology can extract knowledge elements such as entities, relationships, and attributes from publicly available semi-structured, unstructured, and third-party structured databases. In this embodiment, a target personnel knowledge graph is constructed based on the aforementioned basic information. Using the constructed target personnel knowledge graph, the feature values of pre-labeled target personnel with missing values are filled in, resulting in a list of target personnel feature values.
[0039] Step S104: Merge the historical time series feature information and basic information one by one into a time series state vector in chronological order, then combine them into a list of historical time series feature values, input them into a pre-built time series prediction model, and output a list of feature values for future time periods through the time series prediction model.
[0040] Step S105: Input the historical time series feature value list into the pre-built time series prediction model, and output the future time series feature information, the future time period feature value list.
[0041] Specifically, the time series prediction model is pre-trained. By inputting a list of historical time series feature values for m days into the time series prediction model, the model can output a list of future time period feature values for the next n days. Different training sets can be used to train and obtain future time period feature value lists for different future numbers of days. The specific prediction time series model can be adjusted and trained according to the actual situation.
[0042] Step S106: Based on the list of future time period feature values, obtain the risk score of the target person through a pre-built risk assessment model.
[0043] The system inputs a list of future time-period characteristic values into a risk assessment model for calculation. The model then outputs a risk score for the target individual, with a score of 0 representing low risk and a score of 1 representing high risk, indicating that the individual should be given special attention. This improves the timeliness and accuracy of early warnings for target individuals and solves the problem of delayed warnings.
[0044] In some embodiments, the step of constructing a target personnel knowledge graph based on the basic information and determining a target personnel feature value list based on the target personnel knowledge graph includes:
[0045] A risk assessment indicator system is constructed based on the basic information; the basic information is labeled with data according to the risk assessment indicator system to obtain an initial feature value list; based on the initial feature value list, a first entity-relation triplet is determined through entity extraction and relation extraction, and an initial knowledge graph is constructed based on the first entity-relation triplet; based on the initial knowledge graph, all entities in the initial knowledge graph are normalized through entity alignment to obtain a first knowledge graph; the first knowledge graph is completed based on a preset abstract rule base to obtain the target personnel knowledge graph; the initial feature value list is completed based on the target personnel knowledge graph to obtain the target personnel feature value list.
[0046] Specifically, in this embodiment, a risk assessment indicator system is constructed for different categories of target personnel based on historical information and industry expert recommendations. This system covers the basic information of the target personnel, which mainly includes factors related to the individual and family circumstances of the target personnel, as well as specific risk assessment indicators for different categories of target personnel. The specific risk assessment indicator system is shown in Table 1.
[0047] Table 1 Risk Assessment Indicator System
[0048]
[0049] Based on the risk assessment indicator system for the target personnel that has been constructed above, the target personnel data is labeled, as shown in Table 2, which illustrates the labeled data format. Since some indicator values may be missing in the collected data, it is necessary to fill in the missing values as much as possible in the subsequent process.
[0050] Table 2 List of Initial Eigenvalues
[0051]
[0052] For the initial feature value list in this embodiment, an initial knowledge graph is constructed through entity extraction and relation extraction. For tabular data, preliminary knowledge representation can be performed after data integration. If textual data exists, knowledge extraction is required, namely entity extraction, relation extraction, and attribute extraction, before preliminary knowledge representation.
[0053] Knowledge extraction technology is mainly divided into two parts: entity extraction, also known as named entity recognition, which identifies entities in given data, and relation extraction. A more detailed division divides element extraction into three parts: in addition to entity recognition and relation extraction, attribute extraction is also included. Since an entity's attributes can be viewed as a special noun relationship between the entity and its attribute value, we consider attribute extraction as a special case of relation extraction, and its processing method and algorithm are consistent with relation extraction. The entity types defined in this embodiment are shown in Table 3.
[0054] Table 3 Entity Types
[0055] Entity type describe PERSON Task LOC address TIME time PHONE_NUMBER Phone number ID ID number ORGNIZATION Organization CARD_ID Bank card number LICENCE_PLATE license plate number EMAIL Mail …… ……
[0056] The defined relationship types mainly include relationships between people and modification relationships between various entities.
[0057] This application employs a deep learning bidirectional long short-term memory network (Bi-LSTM) combined with a conditional random field (CRF) model to learn from labeled text sequences, constructing an algorithm model for named entity recognition, and extracting entity elements from the aforementioned evaluation index system.
[0058] LSTM is an algorithm model for processing serialized data. It solves the problem of recurrent neural networks' dependence on long-range sequence information and effectively addresses issues such as vanishing and exploding gradients that occur in recurrent neural networks. It has also achieved good results in entity extraction tasks. CRF is an undirected graphical probabilistic model for labeling sequence data. Given a set of input sequences, it can generate a conditional probability distribution of another set of output sequences, maximizing the probability of the target labeled sequence and achieving sequence labeling of the data to be labeled.
[0059] This application employs a deep learning bidirectional long short-term memory network (BiLSTM-CRF) for entity extraction. The model first uses a Look-up layer to obtain word vector mappings for each word, and a Dropout layer is used to prevent overfitting. The second layer is a bidirectional LSTM (BiLSTM) layer. Word vectors are fed into the BiLSTM layer, which learns contextual information and outputs the score probability of each word corresponding to each label. The bidirectional structure effectively discovers structural relationships within the text. Finally, the output of the BiLSTM layer is used as input to the CRF layer, which learns the sequence dependencies between labels to obtain the final prediction result. The CRF layer can correct the output of the BiLSTM layer by learning the transition probabilities between labels in the dataset, thus ensuring the reasonableness of the predicted labels. In the BiLSTM-CRF algorithm, the BiLSTM layer learns the contextual information of the sequence, and the CRF layer learns the dependencies between labels. Through the combination of these two, the algorithm can ultimately achieve accurate extraction of entity elements such as people, places, and organizations.
[0060] After entity extraction, further extraction of relationships between entity pairs such as "person-person" and "person-event" is required. This application uses the Bert-Chinese pre-trained language model for relationship classification to achieve relationship extraction. The target personnel data contains 12 common relationship categories: parents, spouses, teachers and students, siblings, colleagues, partners, lovers, grandparents and grandchildren, friends, relatives, superiors and subordinates, and others. Through entity and relationship extraction, a set of "entity-relationship-entity" triples is ultimately output as the initial knowledge graph.
[0061] Based on the initial knowledge graph obtained in the previous step, this step requires us to perform entity alignment on entity nodes in the extracted "entity-relationship-entity" triplet set that have different names but actually point to the same entity. For example, in "Li San-husband and wife-Wang Wu" and "Li Mou-husband and wife-Wang Wu", both "Li San" and "Li Mou" refer to the entity "Li San". We need to normalize these different expressions through entity alignment to ultimately construct a complete first knowledge graph, such as... Figure 2 As shown. Entity alignment is the process of determining whether two entities refer to the same object in the real world. The purpose of entity alignment is to discover entities with different names but representing the same thing in the real world, merge these entities, identify each entity with a unique identifier, and finally add the entity to the corresponding knowledge graph. This application adopts an entity alignment method based on similarity propagation, modeling the entity alignment problem as an optimization problem with a global matching scoring objective function. This problem belongs to the binary classification problem, and an approximate solution can be obtained through a greedy optimization algorithm.
[0062] This embodiment uses a hybrid reasoning algorithm to complete the first knowledge graph, thereby obtaining more complete target personnel information, preparing for subsequent completion of missing values in the initial feature value list of the original target personnel. The hybrid reasoning algorithm first embeds the first knowledge graph into a vector representation, embedding nodes as vectors and edge relationships as matrices. An initial instance rule base is constructed based on the first knowledge graph and a predefined "abstract rule base," and reasonable instance rules are selected from the initial instance rule base based on the edge relationship embedding matrix to form the final reasonable instance rule base. Based on the obtained reasonable instance rule base, new triples are inferred according to the triple inference rules corresponding to each instance rule, thus completing the completion of the first knowledge graph and obtaining the target personnel knowledge graph. Based on the obtained target personnel knowledge graph, the initial feature value list is completed to obtain the target personnel feature value list.
[0063] In some embodiments, the step of completing the first knowledge graph based on a preset abstract rule base to obtain the target person knowledge graph includes:
[0064] Traverse all second entity relation triples in the first knowledge graph, and merge the instance rules corresponding to all second entity relation triples that satisfy the abstract rule base as an initial instance rule base; construct negative example entity relation triples based on the second entity relation triples; input the second entity relation triples and the negative example entity relation triples into a pre-constructed knowledge graph embedding model, and output the vector embedding representation of the entity and the matrix embedding representation of the relation; calculate the confidence score of each instance rule in the initial instance rule base based on the vector embedding representation of the entity and the matrix embedding representation of the relation; merge all instance rules whose confidence scores exceed a preset confidence threshold as a reasonable instance rule base; infer supplementary entity relation triples based on the second entity relation triples and the reasonable instance rule base; add the supplementary entity relation triples to the first knowledge graph to obtain the target person knowledge graph.
[0065] This embodiment uses hybrid reasoning based on knowledge graphs to complete the knowledge graph, thereby improving the target personnel information and constructing a more comprehensive labeled data sample. Generally speaking, target personnel information may have a certain degree of deliberate or unintentional concealment, which leads to information gaps. These missing information are difficult to obtain directly through manual reasoning from a large amount of sample information. These concealments are reflected in the knowledge graph as path relationships in the graph structure. This application uses a hybrid reasoning method in knowledge reasoning to complete the entity and entity edge relationships of the implicit relation variables and features in the existing knowledge graph. Then, based on the completed knowledge graph, it further completes the missing values in the initial feature value list.
[0066] Specifically, the first knowledge graph is embedded. The embedding model can be obtained by training common knowledge graph embedding models such as ANALOGY and RESCAL, and by minimizing the following loss function.
[0067]
[0068] Where L is the loss function, n is the total number of input triples, σ(·) is the sigmoid function, and v s v o ∈R 1×m M represents the embedding representation of topic s and object o in a triple. r ∈R m×m This represents the matrix embedding of relation r in a triple, where m is the embedding dimension and l is the matrix embedding. sro The tag corresponding to the embedded triple.
[0069] The input to embedding learning is a set of triples and their corresponding labels:
[0070] I = {(s, r, o), l} sro )|(s, r, o)∈G∪G neg}
[0071] The possible values for the tag are defined as follows:
[0072]
[0073] The triples in the set include the entity-relation-entity triples (s, r, o) ∈ G from the first knowledge graph and the constructed negative triples (s, r, o) ∈ G. neg The negative example of a triple can be obtained by replacing s and o in the triple of the first knowledge graph with any entity in the first knowledge graph, or by replacing relation r with any relation in the first knowledge graph. Figure 3 For example, (Target Person 1, Residence, Beijing) ∈ G belongs to the triple in the first knowledge graph, and (Target Person 2, Couple, Beijing) ∈ G neg This is a negative example of a triplet.
[0074] As shown in the figure, all second entity relation triples in the first knowledge graph are traversed, and the instance rules corresponding to all second entity relation triples that satisfy the abstract rule base are merged as the initial instance rule base. The abstract rule base is shown in Table 4.
[0075] Table 4 Abstract Rule Base
[0076]
[0077] For a given abstract rule defined in Table 4 above, all edge relations in the graph are traversed. If an instance that conforms to the abstract rule exists, it is added to the initial instance rule base. Taking the abstract rule "symmetric attribute" as an example, the triplet reasoning of instances in the graph that satisfy this abstract rule is represented as follows:
[0078] (Target Person 1, Couple, Target Person 2) → (Target Person 2, Couple, Target Person 1)
[0079] The corresponding instance rule is: symmetric attribute (couple), and the instance is placed in the initial instance rule base.
[0080] Table 5 lists some of the instance rules in the initial instance rule base.
[0081] Table 5 Initial Instance Rule Base
[0082] Initial instance rule base Symmetrical attributes (couples) Symmetrical attribute (friend) Passing attributes (friends) Equivalent attributes (Born in, Date of Birth) Reversible attributes (children, parents) Reversible attributes (children, mother) Reversible attributes (children, father) The attribute chain includes ((spouse, place of residence), place of residence) ……
[0083] Since some instance rules in the initial instance rule base obtained above are illogical—for example, "friend" may not satisfy the transitive attribute requirement in certain situations—we need to further filter the initial instance rule base to obtain a more general and reasonable instance rule base. Given the embedding representations of all relations in the first knowledge graph and the above initial instance rule base, in order to filter out a reasonable instance rule base, we need to provide the confidence score of each instance in the initial instance rule base for its corresponding abstract rule. The formula for calculating this confidence score is as follows:
[0084]
[0085] Where ||·||F is the Frobenius norm, used to measure the similarity between two matrices. and These represent the equivalence symbols in column 3 of Table 4. The parts on both sides of the equals sign on the right. For example, for "symmetric properties", the matrix... matrix Confidence score s a ∈[0,1], in practical applications, 0.9 is selected as the threshold, and finally all confidence scores satisfy s a The output consists of a reasonable instance rule base composed of instance rules with a value of ≥0.9, as shown in Table 6.
[0086] Table 6. Reasonable Instance Rule Base
[0087] Reasonable instance rule base Symmetrical attributes (couples) Symmetrical attribute (friend) Equivalent attributes (Born in, Date of Birth) Reversible attributes (children, parents) The attribute chain includes ((spouse, place of residence), place of residence) ……
[0088] Based on the obtained reasonable instance rule base, supplementary entity relation triples can be inferred according to the triple inference rules corresponding to each instance rule. For example, the triple inference corresponding to the instance rule "((spouse, place of residence), place of residence)" is expressed as follows:
[0089] (x0, r1, x1), (x1, r2, x2) → (x0, r2, x2)
[0090] Therefore, from the triples (target person 2, spouse, target person 1) and (target person 1, place of residence, Beijing) in the first knowledge graph, the supplementary entity relation triple (target person 2, place of residence, Beijing) can be inferred.
[0091] Figure 3 The target personnel knowledge graph (including dashed lines) is a knowledge graph completed by performing knowledge reasoning on the first knowledge graph (excluding dashed lines). In the original initial feature value list, the value of the "living environment" item for target personnel 2 is missing. By using the relationship between the target personnel and their wife in the knowledge graph, as well as the residence and security environment of their wife (target personnel 1), the "living environment" of the target personnel can be inferred, thus obtaining more comprehensive information about the target personnel.
[0092] Based on the completed target personnel knowledge graph, the corresponding completed values are mapped back to the initial feature value list, missing values are filled in, and some descriptive features are quantified and converted into numerical values. Taking "economic status" as an example, "good -2, medium -1, poor -0" is used. Finally, a complete list of target personnel feature values is output, as shown in Table 7.
[0093] Table 7 List of Target Personnel Characteristic Values
[0094]
[0095]
[0096] In some embodiments, the time-series prediction model includes an encoder and a decoder, both of which are recurrent neural networks. The step of inputting the historical time-series feature information into the pre-built time-series prediction model and outputting a list of feature values for future time periods through the time-series prediction model includes:
[0097] The historical time-series feature information is input into the encoder, and the encoder outputs an encoded state vector; the encoded state vector is input into the decoder, and the decoder outputs a list of feature values for future time periods.
[0098] Specifically, in this embodiment, due to the uncertainty of the number of visit points of the target personnel within a given time period, both the input and output are variable-length sequences. Therefore, an encoder-decoder framework is adopted. The encoder and decoder are two recurrent neural networks corresponding to the input and output sequences, respectively. In this embodiment, the RNN network used to analyze and process the variable-length input is the encoder, and the network used to generate the variable-length output is the decoder. The network framework composed of these two is the encoder-decoder. The list of historical time-series feature values of the instance is used as the input of the encoder to obtain the encoded state vector. Then, the encoded state vector is input into the decoder to obtain the list of feature values for future time periods.
[0099] Specifically: the historical state vector synthesized from the feature values of the starting trajectory point and the target person is encoded, and the encoded state vector is input into the input layer of the decoder. The first predicted state vector is obtained by outputting the decoder.
[0100] Except for the first predicted state vector, each remaining predicted state vector is obtained by the following operation: the previous predicted state vector of the predicted state vector is encoded and input into the input layer of the decoder, and the predicted state vector is obtained by the output of the decoder.
[0101] In response to the detection of the end trajectory point, the output of the predicted state vector is stopped, and all the output predicted state vectors are sequentially combined to form a list of feature values for the future time period. Compared with using only the vector obtained by the trajectory point encoding for prediction, the combined vector has richer dimensions, higher accuracy, and can predict and update the relatively stable feature values of the target personnel.
[0102] Specifically, such as Figure 4 As shown, in this embodiment, the target person's historical time-series feature information is [home, shopping mall, school]. The target person's basic information is (male, 45 years old, average security, poor economic situation). The time-series feature information and basic information are combined to obtain a list of historical time-series feature values: [home, male, 45 years old, average security, poor economic situation], [shopping mall, male, 45 years old, average security, poor economic situation], [school, male, 45 years old, average security, poor economic situation]]. This list of historical time-series feature values is used as the input sequence and encoded by an encoder to obtain a latent vector as the encoding state vector. At the same time, the start character is... <start>The embedded state vector, along with the encoded state vector, is input into the decoder. The argmax function in the Softmax layer predicts the next predicted state vector [coffee shop, male, 45 years old, average security, poor economy]. This state vector is then encoded and input into the decoder for prediction. The argmax function predicts the next predicted state vector [square, male, 45 years old, average security, poor economy], and so on, obtaining predicted state vectors [square, male, 45 years old, average security, poor economy], [home, male, 45 years old, average security, poor economy], etc., until […]. <end>[Male, 45 years old, average security, poor economy], prediction ends, future time sequence characteristic information obtained [[Coffee shop, male, 45 years old, average security, poor economy], [Square, male, 45 years old, average security, poor economy], [Home, male, 45 years old, average security, poor economy]].
[0103] In some embodiments, the risk assessment model is pre-trained using the Extreme Gradient Boosting (XGBoost) algorithm.
[0104] Each person is treated as a sample. For different target groups, corresponding risk indicator data are used as features for each sample. 70% of the total sample is used as the test set, and 30% as the validation set, which is input into the regression model. The Extreme Gradient Boosting (XGBoost) algorithm is applied to train the model. Besides its high accuracy, this algorithm is also very fast because the entire model is built in C++. Furthermore, the model has undergone many optimizations, such as highly utilizing multi-core CPU parallel processing. XGBoost is essentially a gradient boosting decision tree (GBDT) algorithm in machine learning, but it maximizes speed and efficiency, and can be considered an engineering implementation of the GBDT algorithm. During training, the XGBoost algorithm's optimization parameters are: learning rate 0.1–0.3, maximum tree depth 5–10, sample sampling ratio 0.7–1, sample attribute sampling ratio 0.7–1, number of iterations 100–1000, and regularization term weight 5–10. This embodiment ultimately outputs the trained XGBoost model as a risk assessment model. The feature value list is input into the risk assessment model for calculation. The model then outputs the risk classification of the target personnel: a risk score of 0 represents no risk, a risk score of 1 represents low risk, 2 represents medium risk, and 3 represents high risk.
[0105] It should be noted that the method in this embodiment can be executed by a single device, such as a computer or server. The method can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the method described.
[0106] It should be noted that the above description describes some embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0107] Based on the same inventive concept, and corresponding to any of the above embodiments, this application also provides a target personnel dynamic risk warning device based on trajectory prediction.
[0108] refer to Figure 5 The target personnel dynamic risk early warning device based on trajectory prediction includes:
[0109] The acquisition module 501 is configured to acquire data information of the target personnel.
[0110] The segmentation module 502 is configured to divide the data information into basic information and historical time-series feature information;
[0111] Construction module 503 is configured to construct a target personnel knowledge graph based on the basic information, and determine a target personnel feature value list based on the target personnel knowledge graph;
[0112] The merging module 504 is configured to synthesize the historical time series feature information and basic information into a historical time series feature value list;
[0113] The prediction module 505 is configured to input the historical time series feature value list into a pre-built time series prediction model and output a future time series feature value list.
[0114] The assessment module 506 is configured to obtain the risk score of the target person based on the list of feature values for the future time period through a pre-built risk assessment model.
[0115] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware.
[0116] The apparatus described above is used to implement the corresponding trajectory prediction-based dynamic risk warning method for target personnel in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be elaborated here.
[0117] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the target personnel dynamic risk warning method based on trajectory prediction as described in any of the above embodiments.
[0118] Figure 6 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.
[0119] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.
[0120] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.
[0121] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.
[0122] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).
[0123] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.
[0124] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.
[0125] The electronic devices described above are used to implement the corresponding trajectory prediction-based dynamic risk warning method for target personnel in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be elaborated here.
[0126] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the target personnel dynamic risk warning method based on trajectory prediction as described in any of the above embodiments.
[0127] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium, which can be used to store information accessible by a computing device. The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the target personnel dynamic risk warning method based on trajectory prediction as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.
[0128] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application (including the claims) is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in the details for the sake of brevity.
[0129] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.
[0130] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.
[0131] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.< / end> < / start>
Claims
1. A dynamic risk early warning method for target personnel based on trajectory prediction, characterized in that, include: Obtain data information of the target personnel; The data information is divided into basic information and historical time-series feature information; A target personnel knowledge graph is constructed based on the basic information, and a target personnel feature value list is determined based on the target personnel knowledge graph. The historical time series feature information and the target personnel feature values are merged to obtain a list of historical time series feature values, which is then input into a pre-built time series prediction model. The time series prediction model outputs a list of future time series feature values, which includes both the predicted historical time series feature information and the predicted target personnel feature values. Based on the list of future time period feature values, the risk score of the target person is obtained through a pre-constructed risk assessment model. Constructing a target personnel knowledge graph based on the basic information, and determining a target personnel feature value list based on the target personnel knowledge graph, including: constructing a risk assessment indicator system based on the basic information; The basic information is labeled according to the risk assessment index system to obtain an initial feature value list; based on the initial feature value list, the first entity relation triplet is determined by entity extraction and relation extraction; and an initial knowledge graph is constructed based on the first entity relation triplet. Based on the initial knowledge graph, all entities in the initial knowledge graph are normalized through entity alignment to obtain the first knowledge graph. The first knowledge graph is completed based on a preset abstract rule base to obtain the target personnel knowledge graph; The initial feature value list is completed based on the target personnel knowledge graph to obtain the target personnel feature value list; The time series prediction model includes an encoder and a decoder, both of which are recurrent neural networks. The step of inputting the historical time series feature value list into the pre-built time series prediction model and outputting the future time period feature value list through the time series prediction model includes: inputting the historical time series feature value list into the encoder and outputting a state vector through the encoder. The state vector is input into the decoder, and the decoder outputs the list of feature values for the future time period; The historical time series feature value list and the future time series feature value list both contain multiple state vectors. Each vector is composed of trajectory points and target personnel feature values representing time series feature information. The state vectors are input into the decoder, and the decoder outputs the predicted time series feature information, including: encoding the first vector in the historical time series feature value list and inputting it into the input layer of the decoder, and the decoder outputs the first predicted state vector. Except for the first predicted state vector, each of the remaining predicted state vectors is obtained through the following operation: The previous predicted state vector of the predicted state vector is encoded and input into the input layer of the decoder, and the predicted state vector is obtained by outputting the decoder. In response to the determination that the end trajectory point character has been detected, the output of the predicted state vector is stopped, and all the output predicted state vectors are sequentially merged to form a list of feature values for future time periods. The risk assessment model is pre-trained using the Extreme Gradient Boosting (XGBoost) algorithm.
2. The method according to claim 1, characterized in that, The step of determining the first entity-relation triplet based on the initial feature value list through entity extraction and relation extraction includes: performing entity extraction on the initial feature value list through a bidirectional long short-term memory network + conditional random field, and performing relation extraction on the initial feature value list through a pre-trained model BERT to determine the first entity-relation triplet.
3. The method according to claim 1, characterized in that, The step of completing the first knowledge graph based on the preset abstract rule base to obtain the target personnel knowledge graph includes: traversing all second entity relation triples in the first knowledge graph and merging the instance rules corresponding to all second entity relation triples that satisfy the abstract rule base as an initial instance rule base; Construct a negative instance entity relation triplet based on the second entity relation triplet; The second entity relation triplet and the negative example entity relation triplet are input into a pre-constructed knowledge graph embedding model, and the vector embedding representation of the entity and the matrix embedding representation of the relation are output. Based on the vector embedding representation of the entity and the matrix embedding representation of the relationship, the confidence score of each instance rule in the initial instance rule base is calculated. All instance rules whose confidence scores exceed a preset confidence threshold are merged into a reasonable instance rule library; Based on the second entity relation triplet, a supplementary entity relation triplet is obtained by reasoning according to the reasonable instance rule base; The supplementary entity relation triples are added to the first knowledge graph to obtain the target personnel knowledge graph.
4. The method according to claim 3, characterized in that, The step of constructing a negative example entity relation triplet based on the second entity relation triplet includes: replacing the entity in the second entity relation triplet with any entity in the first knowledge graph, and / or replacing the relation in the second entity relation triplet with any relation in the first knowledge graph, to obtain the negative example entity relation triplet.
5. An apparatus employing the target personnel dynamic risk early warning method based on trajectory prediction as described in claim 1, characterized in that, include: The acquisition module is configured to acquire data information of the target personnel. The segmentation module is configured to divide the data information into basic information and historical time-series feature information; The construction module is configured to construct a target personnel knowledge graph based on the basic information, and determine a target personnel feature value list based on the target personnel knowledge graph; The merging module is configured to merge the historical time-series feature information and the target personnel feature values to obtain a list of historical time period feature values; The prediction module is configured to input the historical time series feature information into a pre-built time series prediction model, and output a list of feature values for future time periods through the time series prediction model. The assessment module is configured to obtain the risk score of the target person based on the list of feature values for the future time period through a pre-built risk assessment model.
6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 4.