Logistics behavior safety risk assessment system based on multi-source heterogeneous data combination

Abstract

The invention discloses a logistics behavior safety risk assessment system based on multi-source heterogeneous data combination, and the system comprises a multi-source data integration and preprocessing module which is used for accessing logistics business data of heterogeneous data sources, including logistics package data, package attribute data, personnel attribute data and communication attribute data, carrying out preprocessing operation on the data; the logistics behavior knowledge graph construction module is used for constructing a logistics behavior knowledge graph; the graph embedding representation learning module is used for learning entity and relation embedding representation of the logistics behavior knowledge graph through a TransE model; the risk assessment and prediction module is used for predicting a behavior risk probability through R-GCN and MLP models; and the risk decision and early warning module is used for triggering an early warning strategy according to the prediction result. According to the invention, through integration of multi-source information, the logistics behavior knowledge graph is constructed, and the graph neural network technology is applied, so that accurate prediction and real-time early warning of logistics behavior safety risks are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of logistics safety risk assessment, and in particular to a logistics behavior safety risk assessment system based on multi-source heterogeneous data joint. BACKGROUND

[0002] With the rapid development of e-commerce and logistics industry, the importance of logistics safety risk assessment is increasingly prominent. Traditional logistics safety monitoring mainly relies on manual sampling inspection and rule engine, which has problems such as low efficiency and insufficient coverage. Traditional methods lack the ability to deeply mine hidden associated risks (such as abnormal association of personnel, packages and communication), and have weak detection ability for abnormal behaviors across regions and multiple links, resulting in a high false negative rate of high-risk events.

[0003] Therefore, by integrating package data (such as delivery information, prohibited item identification), personnel risk level data (such as risk level, historical behavior) and communication risk level data, a logistics behavior knowledge graph covering the whole process can be constructed to accurately depict the complex association between entities. Combined with the representation learning ability of graph neural network, it can automatically mine the risk propagation pattern in the logistics network and realize dynamic perception from local anomaly to global risk. Compared with traditional methods, this technical solution has higher risk assessment accuracy and real-time performance, and supports active early warning of hidden risks, providing intelligent support for logistics safety supervision. SUMMARY

[0004] The purpose of the present application is to solve the problems of low detection efficiency and weak correlation analysis ability in existing logistics safety risk assessment technology, and to provide a logistics behavior safety risk assessment system based on multi-source heterogeneous data joint. By innovatively integrating multi-source information such as package data, personnel risk level data and communication risk level data, a logistics behavior knowledge graph is constructed, and graph neural network technology is applied to realize accurate prediction and real-time warning of logistics behavior safety risk. The system can deeply mine the complex association between entities in the logistics network, establish a risk assessment model with higher accuracy and timeliness, and provide intelligent decision support for logistics safety supervision.

[0005] To achieve the above purpose, the technical scheme provided by the present application is: a logistics behavior safety risk assessment system based on multi-source heterogeneous data joint, comprising:

[0006] A multi-source data integration and preprocessing module is used to access logistics business data from heterogeneous data sources, including logistics package data, package risk level data, personnel risk level data and communication risk level data, and to perform missing value filling and standardization preprocessing operations on the data to generate structured input data stream;

[0007] The logistics behavior knowledge graph construction module is configured to construct a logistics behavior knowledge graph with <head entity, relationship, tail entity> as a basic unit, and the input of the logistics behavior knowledge graph construction module is the output of the multi-source data integration and preprocessing module.

[0008] The graph embedding representation learning module is configured to learn entity and relationship embedding representations of the logistics behavior knowledge graph through a TransE model, and generate a low-dimensional vector matrix, and the input of the graph embedding representation learning module is the output of the logistics behavior knowledge graph construction module.

[0009] The risk assessment and prediction module is configured to realize message passing of node features through an R-GCN model, and output behavior risk probability in combination with an MLP model, and the input of the risk assessment and prediction module is the output of the graph embedding representation learning module.

[0010] The risk decision and early warning module is configured to trigger an early warning strategy according to a prediction result, and the input of the risk decision and early warning module is the output of the risk assessment and prediction module.

[0011] Further, the multi-source data integration and preprocessing module comprises:

[0012] The data access unit is configured to collect logistics business data from heterogeneous data sources, and the logistics business data comprises:

[0013] The logistics package data field includes a mail barcode, a mailing time, a pickup person, a sender, a sender address, a sender phone number, a recipient, a recipient address, and a recipient phone number.

[0014] The package risk level data field includes a predefined package risk level classification, including two risk level identifiers of contraband and non-contraband.

[0015] The personnel risk level data field includes a predefined personnel risk level classification, including four risk level identifiers of general personnel, relatively important personnel, important personnel, and very important personnel.

[0016] The communication risk level data field includes a predefined communication risk level classification, including four risk level identifiers of a general phone number, a relatively important phone number, an important phone number, and a very important phone number.

[0017] The data preprocessing unit is configured to perform missing value filling, data standardization, and data verification operations on the data to generate a structured input data stream.

[0018] Further, the logistics behavior knowledge graph construction module performs the following operations:

[0019] 1) Selecting a mail barcode, a mailing time, a pickup person, a sender, a sender address, a sender phone number, a recipient, a recipient address, and a recipient phone number, a package risk level, a personnel risk level, and a communication risk level as entities.

[0020] 2) define the logistics event relationship and attribute association relationship, wherein the logistics event relationship includes "send out", "receive", "pick up", the attribute association relationship includes "sending address", "receiving address", "binding mobile phone number", "belonging to personnel risk level", "belonging to communication risk level", "belonging to prohibited article type", "belonging to pick-up time";

[0021] 3) build a triple of the form <head entity, relationship, tail entity>, store it in the Neo4j graph database, and generate a visual knowledge graph display interface.

[0022] Further, the graph embedding representation learning module performs the following operations:

[0023] 1) embedding representation unit, mapping each entity and relationship in the knowledge graph into a continuous vector representation in a vector space , wherein the head entity vector representation is , the relationship vector representation is , and the tail entity vector representation is ;

[0024] 2) scoring function unit, defining the following scoring function to measure the semantic association between entities and relationships:

[0025]

[0026] wherein is the operator of norm, represents the distance between the head entity vector plus the relationship vector and the tail entity vector ;

[0027] 3) negative sampling unit, for each positive sample , generate negative samples by replacing the head entity or tail entity: or , wherein , are randomly sampled from the entity set;

[0028] 4) loss function unit, using the following loss function for model training:

[0029]

[0030] wherein: represents the set of positive samples, represents the set of negative samples, represents the distance between positive and negative samples, head entity vector in negative samples, tail entity vector in negative samples, representing the non-negative part;

[0031] 5) Parameter updating unit, using stochastic gradient descent algorithm to minimize the loss function, iteratively updating the embedding matrix of entity and relation until convergence.

[0032] Further, the risk assessment and prediction module performs the following operations:

[0033] 1) Node feature initialization unit, taking the entity embedding matrix obtained by the graph embedding representation learning module as the initial node feature matrix , and taking the relation embedding matrix as the initial edge weight matrix ;

[0034] 2) Message passing unit based on R-GCN, as shown in the following formula:

[0035]

[0036] Wherein: denotes the neighbor set of node under the relation , denotes the set of all relations, denotes the corresponding learnable weight matrix of relation , denotes the weight matrix of self-loop connection (used to retain node information), denotes the normalization constant, denotes the Relu activation function, denotes the feature of node in the th layer, denotes the feature of the th node in the th layer, denotes the feature of node in the th layer.

[0037] 3) Risk prediction unit based on MLP, including the following steps:

[0038] Step 1: Define the logistics behavior vector: for a given logistics event relation ( ), the head entity is ( ), and the tail entity is ( ), the head entity embedding representation and the tail entity embedding representation are spliced to obtain the logistics behavior feature;

[0039] Step2: Map the risk level corresponding to the attribute of each entity to a numerical score, and take the average of the scores of the head entity and the tail entity as the behavioral risk weight, as shown in the following formula:

[0040]

[0041] wherein: represents the behavioral risk weight, represents the risk score of the head entity, represents the risk score of the tail entity;

[0042] Step3: Weight the logistics behavior vector with the behavioral risk weight, and then input it into the MLP to predict the risk probability:

[0043]

[0044] wherein: represents the logistics behavior vector, represents the multi-layer perception, represents the softmax function, represents the predicted risk probability vector of the behavior.

[0045] 4) Behavior risk label construction unit, set the behavior risk label according to the attributes of the head entity and the tail entity, and perform the following operations:

[0046] If the attribute of the tail entity is contraband, the behavior belongs to the high-risk category;

[0047] If the attribute of the tail entity is non-contraband and the attribute of the head entity is very important personnel, the behavior belongs to the high-medium risk category;

[0048] If the attribute of the tail entity is non-contraband and the attribute of the head entity is important personnel, the behavior belongs to the medium risk category;

[0049] If the attribute of the tail entity is non-contraband and the attribute of the head entity is relatively important personnel, the behavior belongs to the medium-low risk category;

[0050] If the attribute of the tail entity is non-contraband and the attribute of the head entity is general personnel, the behavior belongs to the low risk category;

[0051] 5) Loss function unit, adopts multi-class cross-entropy loss function, formula as follows:

[0052]

[0053] wherein, represents the total number of risk categories (there are five risk categories), represents the a true label of a first category of the i-th sample, a true label of a first category of the i-th sample, a true label of a first category of the i-th sample, a true label of a first category of the i-th sample, a true label of a first category of the i-th sample, denotes the total number of samples;

[0054] 6) a parameter updating unit, which adopts a stochastic gradient descent algorithm to minimize a loss function, updates model parameters, and converges until convergence;

[0055] Further, the risk decision and early warning module performs the following operations:

[0056] According to the risk probability vector output by the risk assessment and prediction module , the vector contains five-dimensional probability values, respectively corresponding to five risk categories of low risk, low-medium risk, medium risk, high-medium risk and high risk, and the risk category corresponding to the maximum probability value in the risk probability vector is selected as the final risk judgment result;

[0057] According to the risk level, differential early warning measures are performed: for low-risk and low-medium-risk behaviors, no active early warning is triggered; for medium-risk behaviors, a yellow early warning identifier is triggered; for high-medium-risk behaviors, an orange early warning identifier is triggered; and for high-risk behaviors, a red early warning identifier is triggered.

[0058] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0059] 1. By integrating multi-source heterogeneous information such as package data, personnel risk level data and communication risk level data, a knowledge graph covering all elements of logistics is constructed, breaking through the limitations of traditional single data source analysis. The knowledge graph can accurately depict the complex association relationships between multi-dimensional entities such as personnel, packages, locations and communication in the logistics network, providing more comprehensive data support for risk assessment.

[0060] 2. The graph neural network (R-GCN) combined with TransE embedding representation learning is adopted to effectively capture the topological structure and semantic association in the logistics behavior data. The model can adaptively learn the risk propagation pattern, realize deep mining of implicit associated risks, and significantly improve the accuracy and reliability of risk assessment.

[0061] 3. The system establishes a perfect risk assessment and early warning mechanism, realizes intelligent closed-loop management from risk identification, assessment to early warning decision through the organic combination of multi-level risk classification system and differentiated disposal strategy, and greatly improves the automation level and decision efficiency of logistics safety supervision. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1A framework diagram of the system of the present application.

[0063] Figure 2 A flowchart of the data import module.

[0064] Figure 3 A schematic diagram of the logistics behavior knowledge graph.

[0065] Figure 4 A schematic diagram of the TransE model in the graph embedding representation learning module.

[0066] Figure 5 A schematic diagram of the over R-GCN model and the MLP model structure in the risk assessment and prediction module. DETAILED DESCRIPTION

[0067] The present application will be further described in conjunction with the embodiments and the accompanying drawings, but the embodiments of the present application are not limited thereto.

[0068] As Figure 1 shown, the present embodiment provides a logistics behavior security risk assessment system based on multi-source heterogeneous data joint. It includes:

[0069] The multi-source data integration and preprocessing module is used to access the logistics business data of heterogeneous data sources, including logistics package data, package risk level data, personnel risk level data and communication risk level data, and to perform missing value filling and standardization preprocessing operations on the data to generate structured input data stream;

[0070] The logistics behavior knowledge graph construction module is used to construct a logistics behavior knowledge graph with <head entity, relationship, tail entity> as the basic unit, and the input is the output of the multi-source data integration and preprocessing module;

[0071] The graph embedding representation learning module is used to learn the entity and relationship embedding representation of the logistics behavior knowledge graph through the TransE model, and generate a low-dimensional vector matrix, and the input is the output of the logistics behavior knowledge graph construction module;

[0072] The risk assessment and prediction module is used to realize the message passing of node features through the R-GCN model, and outputs the behavior risk probability in combination with the MLP model, and the input is the output of the graph embedding representation learning module;

[0073] The risk decision and early warning module is used to trigger the early warning strategy according to the prediction result, and the input is the output of the risk assessment and prediction module.

[0074] Specifically, the multi-source data integration and preprocessing module includes Figure 2 as shown, including:

[0075] A data access unit is configured to collect logistics business data from heterogeneous data sources, wherein the logistics business data comprises:

[0076] Logistics package data fields: mail barcode, mailing time, pickup person, sender, sender address, sender phone number, recipient, recipient address, recipient phone number;

[0077] Package risk level data fields: predefined package risk level classification, including two types of risk level identification of contraband and non-contraband;

[0078] Personnel risk level data fields: predefined personnel risk level classification, including four types of risk level identification of general personnel, relatively important personnel, important personnel, and very important personnel;

[0079] Communication risk level data fields: predefined communication risk level classification, including four types of risk level identification of general phone number, relatively important phone number, important phone number, and very important phone number;

[0080] A data preprocessing unit is configured to perform missing value filling, data standardization, and data verification operations on the data to generate a structured input data stream.

[0081] Specifically, the logistics behavior knowledge graph construction module performs the following operations as shown in Figure 3

[0082] 1) Selecting mail barcode, mailing time, pickup person, sender, sender address, sender phone number, recipient, recipient address, recipient phone number, package risk level, personnel risk level, and communication risk level as entities;

[0083] 2) Defining logistics event relationships and attribute association relationships, wherein the logistics event relationships include "mail out", "receive", and "pickup", and the attribute association relationships include "sender location", "recipient location", "bound mobile phone number", "belonging to personnel risk level", "belonging to communication risk level", "belonging to contraband type", and "belonging to pickup time";

[0084] 3) Constructing triples in the form of <head entity, relationship, tail entity>, storing them in a Neo4j graph database, and generating a visual knowledge graph display interface.

[0085] Specifically, the graph embedding representation learning module learns entity and relationship embedding representations of the logistics behavior knowledge graph using a TransE model, and generates a low-dimensional vector matrix as shown in Figure 4

[0086] 1) An embedding representation unit maps each entity and relationship in the knowledge graph to ​​A continuous vector representation in a vector space, where the head entity vector is represented as , the relation vector is represented as , and the tail entity vector is represented as ;

[0087] 2) A scoring function unit defines the following scoring function to measure the semantic association between entities and relations:

[0088] where is the operator of norm, denotes the distance between the head entity vector plus the relation vector and the tail entity vector ;

[0089] 3) A negative sampling unit generates negative samples for each positive sample by replacing the head entity or the tail entity: or , where , are randomly sampled from the entity set;

[0090] 4) A loss function unit uses the following loss function for model training:

[0091]

[0092] where: denotes the set of positive samples, denotes the set of negative samples, denotes the distance hyperparameter between positive and negative samples, denotes the head entity vector in the negative sample, denotes the tail entity vector in the negative sample, denotes the non-negative part;

[0093] 5) A parameter update unit uses the stochastic gradient descent algorithm to minimize the loss function, iteratively updates the embedding matrix of entities and relations until convergence.

[0094] Specifically, the risk assessment and prediction module takes the entity and relation embedding matrix obtained by the graph embedding representation learning module as the initial node feature, uses the R-GCN model to realize the message passing of node features, and combines the MLP model to output the behavior risk probability, as shown in Figure 5 , the following operations are performed:

[0095] 1) A node feature initialization unit takes the entity embedding matrix obtained by the graph embedding representation learning module as the initial node feature matrix , the relation embedding matrix as the initial edge weight matrix ;

[0096] 2) R-GCN-based message passing unit, as shown in the following formula:

[0097]

[0098] wherein: denotes the neighbor set of node under the relation , denotes all relation sets, denotes the corresponding learnable weight matrix of relation , denotes the weight matrix of self-loop connection (for preserving node information), denotes the normalization constant, denotes the Relu activation function, denotes the feature of node in the layer, denotes the feature of the th node in the layer, denotes the feature of node in the layer.

[0099] 3) MLP-based risk prediction unit, including the following steps:

[0100] Step 1: Define the logistics behavior vector: for a given logistics event relation ( ), the head entity is ( ), and the tail entity is ( ), the head entity embedding representation is concatenated with the tail entity embedding representation to obtain the logistics behavior feature;

[0101] Step 2: Map the risk level corresponding to the attribute of each entity to a numerical score, and take the average of the scores of the head entity and the tail entity as the behavior risk weight, as shown in the following formula:

[0102]

[0103] wherein: denotes the behavior risk weight, denotes the risk score of the head entity, denotes the risk score of the tail entity;

[0104] Step 3: Weight the logistics behavior vector with behavioral risk weights, and then feed it into the MLP to predict the risk probability:

[0105]

[0106] in: Represents a logistics behavior vector. This represents a multilayer perceptron. This represents the softmax function. This represents the predicted risk probability vector for this behavior.

[0107] 4) Behavioral risk label construction unit: Set behavioral risk labels based on the attributes of the head and tail entities, and perform the following operations:

[0108] If the tail entity is classified as a contraband, then the behavior falls into the high-risk category.

[0109] If the tail entity's attribute is non-contraband and the head entity's attribute is very important personnel, then the behavior belongs to the medium-to-high risk category.

[0110] If the tail entity's attribute is non-contraband and the head entity's attribute is important personnel, then the behavior falls into the medium-risk category.

[0111] If the tail entity is classified as a non-contraband item and the head entity is classified as a relatively important person, then the behavior falls into the low-to-medium risk category.

[0112] If the tail entity's attribute is non-contraband and the head entity's attribute is general personnel, then the behavior falls into the low-risk category.

[0113] 5) Loss function unit: The cross-entropy loss function for multi-class classification is adopted, and the formula is as follows:

[0114]

[0115] in, This indicates the total number of risk categories (there are five risk categories). Indicates the first The first sample Real labels for each category Indicates the first The first sample Predicted values ​​for each category, Indicates the total number of samples;

[0116] 6) The parameter update unit uses the stochastic gradient descent algorithm to minimize the loss function and update the model parameters until convergence.

[0117] Specifically, the risk decision-making and early warning module is used to trigger and visualize early warning strategies based on the prediction results, and performs the following operations:

[0118] 1) According to the risk probability vector output by the risk assessment and prediction module , the vector contains five-dimensional probability values corresponding to five risk categories of low risk, low-medium risk, medium risk, high-medium risk and high risk, and the risk category corresponding to the maximum value in the risk probability vector is selected as the final risk judgment result;

[0119] 2) According to the risk level, difference warning measures are executed: for low-risk and low-medium-risk behaviors, no active warning is triggered; for medium-risk behaviors, a yellow warning mark is triggered; for high-medium-risk behaviors, an orange warning mark is triggered; for high-risk behaviors, a red warning mark is triggered.

[0120] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited by the above embodiments, and any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present application are equivalent replacement methods and are included in the protection scope of the present application.

Claims

1. A logistics behavior safety risk assessment system based on multi-source heterogeneous data fusion, characterized in that, include: The multi-source data integration and preprocessing module is used to access logistics business data from heterogeneous data sources, including logistics parcel data, parcel risk level data, personnel risk level data, and communication risk level data, and to perform missing value filling and standardization preprocessing operations on the data to generate a structured input data stream. The Logistics Behavior Knowledge Graph Construction Module is used to construct a logistics behavior knowledge graph with <head entity, relation, tail entity> as the basic unit. Its input is the output of the multi-source data integration and preprocessing module. The graph embedding representation learning module is used to learn the entity and relation embedding representations of the logistics behavior knowledge graph through the TransE model, and generate a low-dimensional vector matrix. Its input is the output of the logistics behavior knowledge graph construction module. The risk assessment and prediction module is used to realize the message passing of node features through the R-GCN model and output the behavioral risk probability by combining the MLP model. Its input is the output of the graph embedding representation learning module. The risk decision-making and early warning module is used to trigger early warning strategies based on the prediction results. Its input is the output of the risk assessment and prediction module.

2. The logistics behavior safety risk assessment system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The multi-source data integration and preprocessing module includes: The data access unit is used to collect logistics business data from heterogeneous data sources, the logistics business data including: Logistics parcel data fields: mail barcode, pickup and drop-off time, pickup person, sender, sender's address, sender's phone number, recipient, recipient's address, recipient's phone number; Package risk level data field: Predefined package risk level classification, including two risk level labels: prohibited items and non-prohibited items; Personnel risk level data field: Predefined personnel risk level classification, including four risk level identifiers: general personnel, relatively important personnel, important personnel, and very important personnel; Communication risk level data field: Predefined communication risk level classification, including four risk level identifiers: general telephone numbers, relatively important telephone numbers, important telephone numbers, and very important telephone numbers; The data preprocessing unit is used to perform missing value imputation, data standardization, and data validation operations on the data to generate a structured input data stream.

3. The logistics behavior safety risk assessment system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The logistics behavior knowledge graph construction module performs the following operations: 1) Select the mail barcode, collection time, recipient, sender, sender address, sender phone number, recipient, recipient address, recipient phone number, package risk level, personnel risk level, and communication risk level as entities; 2) Define logistics event relationships and attribute associations. Logistics event relationships include "sent", "received", and "pickup". Attribute associations include "sender location", "receiver location", "bound mobile phone number", "belongs to personnel risk level", "belongs to communication risk level", "belongs to prohibited item type", and "belongs to pickup time". 3) Construct triples in the form of <head entity, relation, tail entity>, store them in the Neo4j graph database, and generate a visual knowledge graph display interface.

4. The logistics behavior safety risk assessment system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The graph embedding representation learning module uses the TransE model to learn the entity and relation embedding representations of the logistics behavior knowledge graph, generates a low-dimensional vector matrix, and performs the following operations: 1) Embedded representation units map each entity and relation in the knowledge graph to... A continuous vector representation in a 3D vector space, where the head entity vector is represented as... The relation vector is represented as The tail entity vector is represented as ; 2) Scoring function unit, defining the following scoring function to measure the semantic association between entities and relations: ; in for Operators of norms Represents the head entity vector Add relation vector Post and tail entity vectors The distance between them; 3) Negative sampling unit, for each positive sample Generate by replacing the head or tail entity. One negative sample: or ,in , Obtained by random sampling from the entity set; 4) Loss function unit: The following loss function is used for model training: ； in: Describes the set of positive samples. Represents the set of negative samples. This represents the hyperparameter indicating the distance between positive and negative samples. This represents the head entity vector in the negative sample. This represents the tail entity vector in the negative sample. This indicates taking the non-negative part; 5) The parameter update unit uses the stochastic gradient descent algorithm to minimize the loss function and iteratively updates the embedding matrices of entities and relations until convergence.

5. The logistics behavior safety risk assessment system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The risk assessment and prediction module uses the entity and relation embedding matrices obtained by the graph embedding representation learning module as initial node features, utilizes the R-GCN model to realize message passing of node features, and combines the MLP model to output behavioral risk probabilities, and performs the following operations: 1) The node feature initialization unit uses the entity embedding matrix obtained from the graph embedding representation learning module as the initial node feature matrix. The relation embedding matrix serves as the initial edge weight matrix. ; 2) The message passing unit based on R-GCN is shown in the following formula: ； in: Indicates a relationship Next node The neighborhood group, Represents the set of all relations. Representing relations The corresponding learnable weight matrix, The weight matrix represents the self-loop connection (used to preserve node information). Represents the normalization constant. This represents the ReLU activation function. Represents a node In the Features of the layer Indicates the first Nodes In the Features of the layer Represents a node In the Features of the layer; 3) MLP-based risk prediction unit, including the following steps: Step 1: Define the logistics behavior vector: for a given logistics event relationship ( Its head entity is ( ), the tail entity is ( The head entity embedding representation and the tail entity embedding representation are concatenated to obtain the logistics behavior features; Step 2: Map the risk level corresponding to each entity's attributes to a numerical score, and take the average score of the head entity and the tail entity as the behavioral risk weight, as shown in the following formula: ； in: Indicates the behavioral risk weight. This represents the risk score of the head entity. This represents the risk score of the tail entity; Step 3: Weight the logistics behavior vector with behavioral risk weights, and then feed it into the MLP to predict the risk probability: ； in: Represents a logistics behavior vector. This represents a multilayer perceptron. This represents the softmax function. This represents the predicted risk probability vector for this behavior; 4) Behavioral risk label construction unit: Set behavioral risk labels based on the attributes of the head and tail entities, and perform the following operations: If the tail entity is classified as a contraband, then the behavior falls into the high-risk category. If the tail entity's attribute is non-contraband and the head entity's attribute is very important personnel, then the behavior belongs to the medium-to-high risk category. If the tail entity's attribute is non-contraband and the head entity's attribute is important personnel, then the behavior falls into the medium-risk category. If the tail entity is classified as a non-contraband item and the head entity is classified as a relatively important person, then the behavior falls into the low-to-medium risk category. If the tail entity's attribute is non-contraband and the head entity's attribute is general personnel, then the behavior falls into the low-risk category. 5) Loss function unit: The cross-entropy loss function for multi-class classification is adopted, and the formula is as follows: ； in, This indicates the total number of risk categories (there are five risk categories). Indicates the first The first sample Real labels for each category Indicates the first The first sample Predicted values ​​for each category, Indicates the total number of samples; 6) The parameter update unit uses the stochastic gradient descent algorithm to minimize the loss function and update the model parameters until convergence.

6. The logistics behavior safety risk assessment system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The risk decision-making and early warning module is used to trigger and visualize early warning strategies based on prediction results, and performs the following operations: 1) Based on the risk probability vector output by the risk assessment and prediction module The vector contains probability values ​​in five dimensions, corresponding to five risk categories: low risk, low-medium risk, medium risk, medium-high risk, and high risk. The risk category corresponding to the highest probability value in the risk probability vector is selected as the final risk judgment result. 2) Implement differentiated early warning measures based on risk level: No proactive early warning will be triggered for low-risk and low-to-medium-risk behaviors; For medium-risk behaviors, a yellow alert is triggered; For behaviors that are of medium to high risk, an orange alert will be triggered. High-risk behaviors trigger a red alert.

Images