Cold chain logistics transportation demand forecasting method and system based on data mining

Abstract

The present invention provides a method and system for predicting cold chain logistics transportation demand based on data mining, which relates to the field of artificial intelligence technology, and effectively improves the prediction accuracy of cold chain logistics transportation demand. First, the conversation interaction data of the cold chain logistics transportation conversation is obtained, and the conversation interaction path data of multiple logistics knowledge label fields are parsed therefrom, and then these conversation interaction path data are encoded into multiple directed flow graph vector sets. By intercepting and blending these directed flow graph vector sets, conversation characterization vector data that can comprehensively characterize the cold chain logistics transportation conversation is generated. Finally, the present invention can accurately predict the logistics demand of cold chain logistics transportation based on the conversation characterization vector data, thereby facilitating the scheduling and optimization of cold chain logistics transportation, and significantly improving the efficiency and benefits of cold chain logistics transportation.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for predicting cold chain logistics transportation demand based on data mining. Background Art

[0002] As an important part of the modern logistics system, cold chain logistics plays a vital role in ensuring the quality and safety of perishable and changeable commodities such as food and medicine. However, the forecasting of cold chain logistics transportation demand has always been a complex and difficult problem to grasp accurately. Traditional forecasting methods often rely on historical transportation data, market trend analysis and experience judgment. These methods often fail to provide accurate and timely forecast results when faced with a changing market environment and complex transportation needs.

[0003] With the continuous development of data mining technology, related technologies have begun to try to use data mining methods to predict the demand for cold chain logistics transportation. However, when processing cold chain logistics transportation session data, existing data mining methods often have problems such as insufficient data analysis, insufficient feature extraction, and insufficient accuracy of prediction models, resulting in certain errors and uncertainties in the prediction results. Summary of the invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for predicting cold chain logistics transportation demand based on data mining, the method comprising:

[0005] Get the session interaction data of the cold chain logistics transportation session;

[0006] Parsing the session interaction path data of a plurality of logistics knowledge tag fields from the session interaction data, and encoding the parsed plurality of the session interaction path data into a plurality of directed flow graph vector sets;

[0007] Taking the first set number as the interception interval and the second set number as the moving interval, intercepting each of the directed flow graph vector sets to generate a first directed graph node vector fragment set corresponding to each of the directed flow graph vector sets;

[0008] Blending each of the first directed graph node vector fragment sets, and integrating a plurality of blended directed graph node set vectors generated by blending to generate session representation vector data of the cold chain logistics transportation session;

[0009] The logistics demand prediction result of the cold chain logistics transportation session is determined based on the session characterization vector data.

[0010] On the other hand, an embodiment of the present invention also provides a cold chain logistics transportation demand forecasting system based on data mining, including a processor and a machine-readable storage medium, the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0011] Based on the above aspects, the embodiment of the present application effectively improves the prediction accuracy of cold chain logistics transportation demand. First, the conversation interaction data of the cold chain logistics transportation conversation is obtained, and the conversation interaction path data of multiple logistics knowledge tag fields are parsed therefrom, and then these conversation interaction path data are encoded into multiple directed flow graph vector sets. By intercepting and blending these directed flow graph vector sets, conversation characterization vector data that can comprehensively characterize the cold chain logistics transportation conversation is generated. Finally, based on the conversation characterization vector data, the logistics demand of cold chain logistics transportation can be accurately predicted, thereby facilitating the scheduling and optimization of cold chain logistics transportation, and significantly improving the efficiency and benefits of cold chain logistics transportation. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the execution flow of the cold chain logistics transportation demand prediction method based on data mining provided in an embodiment of the present invention.

[0013] Figure 2 It is a schematic diagram of the hardware architecture of a cold chain logistics transportation demand prediction system based on data mining provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 It is a flow chart of a method for predicting cold chain logistics transportation demand based on data mining provided by an embodiment of the present invention. The method for predicting cold chain logistics transportation demand based on data mining is introduced in detail below.

[0015] Step S110, obtaining session interaction data of the cold chain logistics transportation session.

[0016] In this embodiment, the server establishes data connections with multiple clients related to cold chain logistics and transportation (such as the scheduling system of the logistics and transportation platform, the warehouse management system, the customer's ordering platform, etc.). These clients will generate a large amount of session interaction data in various links of cold chain logistics and transportation. For example, a factory acts as a customer and interacts with the cold chain logistics and transportation service provider through the ordering platform. The customer fills in the order information on the ordering platform, including the type of goods (such as Class A drugs), the starting point of transportation (a factory warehouse), the destination (wholesaler warehouse in the city), the expected transportation time (hope to be delivered within three days), the temperature requirement (0-4 degrees Celsius), etc. At the same time, the scheduling system of the logistics and transportation platform will also interact with the server, for example, informing the server of the vehicle scheduling arrangement, such as a refrigerated truck has departed from the garage to the factory warehouse to load goods, and the vehicle's refrigeration equipment is in good condition, and the estimated time of arrival at the warehouse. The warehouse management system will also send information to the server, such as the inventory status of the factory warehouse (currently there are 500 boxes of Class A drugs available for transportation), the loading progress of the goods (300 boxes of Class A drugs have been loaded), etc. The various information related to cold chain logistics and transportation collected from different clients constitutes the session interaction data. The server collects the information scattered in various clients through specific data interfaces and communication protocols, such as RESTful API based on HTTP protocol or message queues (such as RabbitMQ), to form a complete session interaction data set of cold chain logistics and transportation sessions. These data contain text information, digital information and other forms, and are the basic data source for subsequent operations such as logistics demand forecasting.

[0017] Step S120, parsing the session interaction path data of the multiple logistics knowledge tag fields from the session interaction data, and encoding the multiple parsed session interaction path data into multiple directed flow graph vector sets.

[0018] In this embodiment, the server first constructs a dictionary of logistics knowledge tag fields. This dictionary covers various knowledge tag fields involved in the field of cold chain logistics and transportation, such as "cargo information" (including cargo type, quantity, weight, etc.), "transportation starting point", "transportation end point", "transportation time requirements", "temperature requirements", "vehicle information" (including vehicle model, refrigeration equipment status, etc.), "warehouse information" (inventory, in and out of the warehouse, etc.). For each text fragment in the collected session interaction data, the server uses a text matching algorithm for processing. Taking the order information filled in by the customer on the ordering platform mentioned above as an example, for the text fragment "Class A drugs", the server will compare it with the knowledge tag fields in the dictionary one by one in a predefined order. Through comparison, it is found that "Class A drugs" belongs to the logistics knowledge tag field "cargo information". Then, for the entire cold chain logistics and transportation session, starting from the starting point, it is sorted in chronological order or logical order. Assuming that the session starts with the customer placing an order on the ordering platform, then in order, the text fragment related to "cargo information" is first determined as a node. For example, when the text fragment "Class A drugs" is identified, it is used as a node in the graph structure corresponding to "cargo information", and the location of the node in the session is recorded (such as this is the information corresponding to the first step of filling out an order) and its relationship with the previous related node (since this is the starting node and there is no previous related node, it can be marked as a starting node relationship). At the same time, the specific content of the logistics knowledge label field corresponding to the node is recorded (i.e., "Class A drugs").

[0019] As the conversation progresses, when the logistics transportation platform dispatch system informs the vehicle to leave for the warehouse, the text segment "A refrigerated truck has left the garage to go to the factory warehouse to load goods, and the vehicle's refrigeration equipment is in good condition, and the estimated time of arrival at the warehouse" is recognized. The information such as "a refrigerated truck" belongs to the logistics knowledge tag field "vehicle information", which is used as a node in the graph structure corresponding to "vehicle information". It appears after the customer places an order, and its relationship with the previous node (assuming that the previous node is the cargo information node in the customer's order) is based on the chronological order. Record the specific content corresponding to this node (vehicle model, refrigeration equipment status, and estimated arrival time, etc.).

[0020] When constructing the conversation interaction path corresponding to each logistics knowledge tag field, take "cargo information" as an example, and connect all the nodes related to cargo information in order. Assuming that there are nodes about cargo weight, quantity and other information in the future, connect them to obtain the conversation interaction path data including node order (cargo type first, then cargo quantity, weight, etc.), node relationship (precedence relationship based on time and logical order) and node specific content (such as Class A drugs, 500 boxes, total weight 1 ton, etc.).

[0021] The node corresponding to each text segment is used as the vertex in the graph structure of the corresponding logistics knowledge label field. According to the node order and relationship in the conversation interaction path data, the connection relationship between the edges of the vertices is determined. For example, in the graph structure of "Cargo Information", there is a connection relationship between the "Class A Drugs" node and the "500 Boxes" node, and an initial weight value is assigned to this edge. This initial weight value is determined based on the distance between the nodes (the order of precedence in the conversation) and the degree of association between the nodes (the degree of association here is very high because the quantity is for Class A drugs). Assuming that the "Class A Drugs" node is numbered 1 and the "500 Boxes" node is numbered 2, the distance between them is close and the association is close. According to the pre-set algorithm (such as the algorithm with a higher weight for closer distance and higher association), a weight value of 0.8 is assigned.

[0022] The server will also detect isolated nodes in the initial graph structure. For example, when processing a session in a transportation link, a separate "special packaging requirements" text fragment may appear due to data transmission errors, which is not connected to other nodes in the current graph structure. If this isolated node is caused by data errors or incompleteness, the server will supplement the connection relationship based on its context information. For example, by checking the previous and subsequent session content, it is found that this special packaging requirement is for Class A drugs, and it will be connected to the "Class A drugs" node in the "Cargo Information". If this isolated node has independent information that is not associated with other nodes, such as a random coded information that is mistransmitted and irrelevant to transportation, the server determines that it is not necessary to retain the isolated node based on the pre-defined business integrity template, and re-optimizes the edge connection in the initial graph structure according to the logistics business logic. For example, if it is found that the relationship between the "number of spare tires" in a certain vehicle information and other nodes does not conform to the business logic (such as the number of spare tires in cold chain logistics is not closely related to the main logic of transportation), the edge connection may be adjusted or removed to generate an optimized graph structure.

[0023] The nodes in the optimized graph structure are numbered in the depth-first search order. For example, in the "Cargo Information" graph structure, the nodes are numbered in the order in which they appear in the session. For each edge, the encoding is performed according to the numbers of the two nodes connected by the edge and the weight value of the edge. For example, the "Class A Drugs" node mentioned above is numbered 1, the "500 Boxes" node is numbered 2, and the weight value of the edge is 0.8, which are combined into a coding vector [1, 2, 0.8]. Finally, the coding vectors of all edges are combined to generate a directed flow graph vector set. For example, for the graph structure corresponding to the logistics knowledge label field "Cargo Information", the generated directed flow graph vector set contains the coding vectors of all edges, which fully describe the relationship, order, weight and other information between the nodes related to the cargo information. The same operation is performed on other logistics knowledge label fields (such as "Transportation Start Point", "Transportation End Point", etc.), and multiple directed flow graph vector sets are obtained.

[0024] Step S130, intercepting each of the directed flow graph vector sets with a first set number as the interception interval and a second set number as the moving interval, to generate a first directed graph node vector fragment set corresponding to each of the directed flow graph vector sets.

[0025] Assume that the first set number is 5 and the second set number is 2. For a previously generated directed flow graph vector set (taking the directed flow graph vector set corresponding to "cargo information" as an example), this set contains multiple code vectors, indicating information such as the connection relationship between nodes related to cargo information. The server starts from the starting position of this vector set and takes 5 code vectors as a cut-off interval. For example, if this directed flow graph vector set contains 15 code vectors, then the first cut-off obtains fragments from the 1st to the 5th code vector. Then, according to the rule of moving interval of 2, the next cut-off is the fragment from the 3rd to the 7th code vector. Each cut-off fragment is a first directed graph node vector fragment set. This first directed graph node vector fragment set contains vector representations of information such as node relationships and weights within the cut-off interval. For other directed flow graph vector sets (such as directed flow graph vector sets corresponding to "vehicle information" and "transportation destination"), the cut-off operation is performed according to the same rules. After such a truncation operation, each directed flow graph vector set generates a corresponding first directed graph node vector fragment set. These first directed graph node vector fragment sets are the results of segmenting the original directed flow graph vector set according to specific rules, providing a data basis for subsequent blending operations.

[0026] Step S140, blending each of the first directed graph node vector fragment sets, and integrating the multiple blended directed graph node set vectors generated by the blending to generate the session representation vector data of the cold chain logistics transportation session.

[0027] First, the server determines the first deep learning network corresponding to each first directed graph node vector fragment set. Assume that for the first directed graph node vector fragment set corresponding to "cargo information", there is a dedicated first deep learning network A, which is pre-trained according to the characteristics and sample data of the cold chain logistics and transportation field. The first deep learning network A includes structures such as the graph self-attention layer, the feature fusion layer, and the spatial embedding representation layer.

[0028] The server loads the first directed graph node vector fragment set corresponding to the "cargo information" into the first deep learning network A. In network A, the vector fragment set is first processed by graph self-attention according to the graph self-attention layer. For example, for each node vector in the fragment set, the graph self-attention layer calculates the attention weight of each node to other nodes based on the relationship between the nodes and its own characteristics. If one node in the first directed graph node vector fragment set represents the weight of the goods and the other node represents the type of goods, the graph self-attention layer calculates the attention weight between the two nodes according to the business logic of cold chain logistics transportation (such as different types of goods may have different weight range standards). After graph self-attention processing, the first template graph self-attention processing result is generated.

[0029] Next, the first directed graph node vector fragment set corresponding to the "cargo information" is spatially embedded and represented according to the spatial embedding representation layer to generate a first template spatial embedding representation vector fragment set. This process is to represent the node vector in a specific space in order to better reflect the spatial relationship between the nodes. For example, the node vectors such as the weight and quantity of the goods are embedded and represented in a feature space related to cold chain logistics and transportation, so that in this space, the relationship between nodes such as weight and quantity can be more in line with the logistics business logic.

[0030] Then, the first template space embedding representation vector fragment set is fused according to the feature fusion layer. During the fusion process, the feature fusion layer will fuse the features of different nodes, for example, the features of the type of goods and the features of the quantity of goods, to generate a template fusion directed graph node set vector corresponding to each "goods information". This template fusion directed graph node set vector integrates various features and relationships of the nodes related to the goods information.

[0031] For other first directed graph node vector fragment sets (such as the sets corresponding to "vehicle information", "transportation destination", etc.), the same operation is performed according to their respective corresponding first deep learning networks to generate their respective corresponding fused directed graph node set vectors.

[0032] Next, the server determines the first impact factor of each intersection directed graph node set vector. For example, obtain the cold chain logistics application details, assuming that the cold chain logistics application details contain the service information of the interactive source service of the session interaction data. If a certain intersection directed graph node set vector is about "cargo information", and the service information of the interactive source service shows that this batch of goods is high-value perishable goods (such as imported fresh seafood), then the first impact factor of the intersection directed graph node set vector related to "cargo information" will be relatively high, because for high-value perishable goods, the importance of cargo information in the entire cold chain logistics transportation is higher.

[0033] Each intersecting directed graph node set vector is optimized based on this first impact factor to generate multiple optimized vector sets. For example, if the first impact factor of the intersecting directed graph node set vector of "cargo information" is 1.2 (assuming it is calculated based on the cold chain logistics application details through a specific algorithm), then each element in this vector is multiplied by 1.2 for optimization.

[0034] Finally, multiple optimized vector sets are integrated. All optimized vector sets such as "cargo information", "vehicle information", and "transportation destination" are integrated according to certain rules (such as weighted summation or splicing) to generate the session representation vector data of the cold chain logistics transportation session. This session representation vector data integrates the information of various aspects of the cold chain logistics transportation session (cargo, vehicle, transportation route, etc.) and represents the characteristics of the entire session in the form of a vector.

[0035] Step S150: determining a logistics demand prediction result of the cold chain logistics transportation session based on the session characterization vector data.

[0036] In this embodiment, the server loads the previously generated session representation vector data into the second deep learning network for logistics demand forecasting. The second deep learning network is also pre-trained based on the sample learning data. For example, the second deep learning network is a multi-layer neural network structure, including an input layer, multiple hidden layers, and an output layer. The session representation vector data is input into the input layer of the second deep learning network, and undergoes a series of complex calculations and transformations in the hidden layer.

[0037] Assume that in the hidden layer, analysis is performed based on various business rules and data features of cold chain logistics transportation. For example, based on comprehensive information such as cargo information (type, quantity, etc.), vehicle information (refrigeration capacity, transportation capacity, etc.), and transportation route (distance, transportation time requirements, etc.), feature extraction and analysis of data is performed in the hidden layer.

[0038] After being processed by the hidden layer, the first logistics demand prediction heat map is generated in the output layer. This heat map shows the prediction probability or importance of different logistics demand knowledge points in a visual way (here visualization is a representation method that is easy to understand, and actually exists in the form of data inside the server). For example, in the first logistics demand prediction heat map, the area corresponding to the logistics demand knowledge point of temperature requirement is displayed as a higher heat (indicating a higher probability or importance), because according to the cargo information (Class A drugs) in the session representation vector data, there are strict requirements for temperature in cold chain logistics.

[0039] Based on this first logistics demand prediction heat map, the logistics demand knowledge points of the cold chain logistics transportation session are determined. For example, if the heat of the area corresponding to the transportation time requirement shown in the first logistics demand prediction heat map is also very high, then it is determined that the transportation time requirement is an important logistics demand knowledge point of this cold chain logistics transportation session. The server can provide decision support for the arrangement of cold chain logistics transportation based on these determined logistics demand knowledge points, such as arranging suitable vehicles, planning the best transportation route, and ensuring that the goods are delivered within the specified temperature and time.

[0040] Based on the above steps, the embodiment of the present application effectively improves the prediction accuracy of cold chain logistics transportation demand. First, the conversation interaction data of the cold chain logistics transportation conversation is obtained, and the conversation interaction path data of multiple logistics knowledge tag fields are parsed therefrom, and then these conversation interaction path data are encoded into multiple directed flow graph vector sets. By intercepting and blending these directed flow graph vector sets, conversation characterization vector data that can comprehensively characterize the cold chain logistics transportation conversation is generated. Finally, based on the conversation characterization vector data, the logistics demand of cold chain logistics transportation can be accurately predicted, thereby facilitating the scheduling and optimization of cold chain logistics transportation, and significantly improving the efficiency and benefits of cold chain logistics transportation.

[0041] For example, in a possible implementation, step S120 includes:

[0042] Step S121, pre-constructing a dictionary of logistics knowledge label fields, wherein the dictionary contains knowledge label fields involved in the field of cold chain logistics and transportation.

[0043] For example, it includes cargo-related information such as "cargo type", "cargo quantity", "cargo weight", "cargo value", "cargo shelf life", etc.; transportation tool-related information such as "refrigerated truck model", "refrigeration capacity of refrigerated truck", "refrigerated truck volume", "transport ship type" (if water transportation is involved), etc.; transportation route-related information such as "transportation starting point", "transportation end point", "passing stations", "transportation distance", etc.; and warehousing-related information such as "warehouse temperature requirements", "warehouse inventory capacity", "warehouse location" and many other knowledge label fields.

[0044] Step S122: for each text segment in the conversation interaction data, a text matching algorithm is used to compare the text segments with the knowledge tag fields in the dictionary one by one in a predefined order, and the logistics knowledge tag field to which each text segment belongs is output according to the comparison result.

[0045] For each text segment in the collected conversation interaction data, the server uses a text matching algorithm to compare it with the knowledge tag fields in the dictionary one by one in strict accordance with the predefined order. For example, in the cold chain logistics transportation conversation interaction data, there is a text segment of "100 boxes of Class A drugs". The server will compare it with the knowledge tag fields in the dictionary in a predetermined order. Through the comparison, it is found that "100 boxes" belongs to the logistics knowledge tag field of "quantity of goods" and "Class A drugs" belongs to the logistics knowledge tag field of "type of goods". Then, for each cold chain logistics transportation conversation, starting from the starting point of the conversation, it is carefully sorted according to the time sequence or logical sequence of the cold chain logistics transportation conversation interaction. Assume that a cold chain logistics transportation conversation starts with the customer entering the order information on the ordering platform. In this process, when a text segment related to a logistics knowledge tag field is identified, this text segment is used as a node of the graph structure corresponding to the logistics knowledge tag field. For example, when the text segment "Class A drugs" is identified, it is used as a node of the graph structure corresponding to "type of goods". At this time, the server will accurately record the location of the node in the cold chain logistics transport session. Since this is the starting part of the order information input, its accurate sequence position will be marked, and the relationship with the previous related node will be determined. Since it is the starting node, it has no relationship with the previous node based on time sequence or business logic, and can be specially marked as the starting node relationship. And for this node, the specific content of its corresponding logistics knowledge label field, that is, "Class A drugs", is also recorded.

[0046] Step S123, for each cold chain logistics transport session, starting from the starting point of the cold chain logistics transport session, the cold chain logistics transport session is sorted according to the time sequence or logical sequence of the cold chain logistics transport session interaction. During the sorting process, when a text segment related to a logistics knowledge label field is identified, the text segment is used as a node of the graph structure corresponding to the logistics knowledge label field, and the position of the node in the cold chain logistics transport session and the relationship with the previous related node are recorded, and the relationship includes a sequential relationship based on the time sequence or a causal relationship based on business logic, etc. And, for each node, the specific content of the logistics knowledge label field corresponding to the node is recorded at the same time.

[0047] Step S124, when constructing the conversation interaction path corresponding to the logistics knowledge tag field, all nodes are connected in sequence to obtain conversation interaction path data including node sequence, node relationship and node specific content.

[0048] When constructing the conversation interaction path corresponding to the logistics knowledge tag field, the server connects all related nodes in order. For example, for the logistics knowledge tag field "Cargo Type", if the node "Class B Drugs" is subsequently identified, the two nodes "Class A Drugs" and "Class B Drugs" are connected in order, so that the conversation interaction path data containing the node order (first "Class A Drugs" and then "Class B Drugs"), the node relationship (the relationship based on the time sequence) and the specific content of the node ("Class A Drugs" and "Class B Drugs") are obtained.

[0049] Step S125, taking the node corresponding to each text fragment as the vertex in the graph structure of the corresponding logistics knowledge label field, determining the connection relationship of the edges between the vertices according to the node order and relationship in the conversation interaction path data, and assigning an initial weight value to each edge, to obtain an initial graph structure including nodes and preliminary connection relationships, wherein the initial weight value is determined based on the distance between the nodes and the degree of association between the nodes.

[0050] Next, the server uses the node corresponding to each text segment as the vertex in the graph structure of the corresponding logistics knowledge tag field. According to the node order and relationship in the session interaction path data, the connection relationship of the edge between the vertices is determined. For example, in the graph structure of "goods type", if the "Class A drug" node is followed by the "Class B drug" node, then there is an edge between the two vertices. For each edge, the server shall assign an initial weight value respectively, so as to obtain an initial graph structure containing nodes and preliminary connection relationships. This initial weight value is determined based on the distance between nodes and the degree of association between nodes. If the distance between two nodes in the session is very close and the degree of association is very high, then the weight value will be larger. For example, in the graph structure of "goods type", if "Class A drug" and "Class B drug" are goods types that appear successively in the same order, the distance between them is close and the degree of association is high. Because they both belong to the category of goods and are under the same order, according to the pre-set weight calculation rules, a higher weight value, such as 0.8, may be assigned.

[0051] Step S126, detecting isolated nodes in the initial graph structure, wherein the isolated nodes refer to nodes that have no connection relationship with other nodes in the initial graph structure.

[0052] Step S127, if the isolated node is an isolated node caused by data error or incompleteness, the connection relationship is supplemented according to the context information of the isolated node. If the isolated node has no independent information associated with other nodes, it is determined whether the isolated node needs to be retained based on a pre-defined business integrity template, and the connection of the edges in the initial graph structure is re-optimized according to the logistics business logic to generate an optimized graph structure.

[0053] After that, the server will detect isolated nodes in the initial graph structure. An isolated node is a node that has no connection relationship with other nodes in the initial graph structure. For example, in the graph structure constructed by processing the session data of a certain transportation link, a node corresponding to a separate "special packaging requirements" text fragment appears, and this node has no connection relationship with other nodes in the current graph structure. If this isolated node is an isolated node caused by data errors or incompleteness, the server will supplement the connection relationship based on the context information of the isolated node. Assuming that this "special packaging requirements" node appears in the conversation part related to "Class A drugs" goods, by checking the previous and subsequent conversation contents, it is found that this special packaging requirement is for Class A drugs, then it will be connected to the "Class A drugs" node in "Cargo Type". If this isolated node has no independent information associated with other nodes, such as a node corresponding to a random coded information that is mistransmitted and has nothing to do with cold chain logistics transportation, the server will determine whether the isolated node needs to be retained based on the pre-defined business integrity template. If according to the business integrity template, this node has no value for the entire cold chain logistics transportation business logic, then it is determined that the node does not need to be retained. And re-optimize the edge connections in the initial graph structure according to the logistics business logic. For example, if it is found that the relationship between the "spare tire quantity" node in a certain vehicle information and other nodes does not conform to the business logic (in cold chain logistics, the number of spare tires is not closely related to the main logic of transportation), the server may adjust the edge connection or remove this edge to generate an optimized graph structure.

[0054] Step S128, number the nodes in the optimized graph structure according to the depth-first search order or the breadth-first search order, and, for each edge, encode according to the numbers of the two nodes connected by the edge and the weight value of the edge, specifically combining the starting node number, the ending node number and the weight value of the edge into a coding vector.

[0055] Step S129: combine the encoding vectors of all edges to generate a directed flow graph vector set.

[0056] Finally, the server numbers the nodes in the optimized graph structure in the depth-first search order. For example, in the "goods type" graph structure, the nodes are numbered in the order in which they appear in the session. For each edge, the encoding is performed according to the numbers of the two nodes connected by the edge and the weight value of the edge. For example, the node number of the "Class A drug" mentioned above is 1, the node number of the "Class B drug" is 2, and the weight value of the edge is 0.8, then the start node number, the end node number and the weight value of the edge are combined into a coding vector [1, 2, 0.8]. Combining the coding vectors of all edges generates a directed flow graph vector set. This directed flow graph vector set fully describes the relationship, order, weight and other information between the nodes in the graph structure corresponding to the logistics knowledge label field. For the graph structures corresponding to other logistics knowledge label fields, the server also operates according to the same process to generate multiple directed flow graph vector sets, which will provide an important data basis for the subsequent processing of cold chain logistics transportation sessions.

[0057] In a possible implementation, step S140 includes:

[0058] Step S141: determine a first deep learning network corresponding to each of the first directed graph node vector fragment sets.

[0059] Step S142: Load the first directed graph node vector fragment set corresponding to each of the directed flow graph vector sets into the corresponding first deep learning network for blending, and generate a blended directed graph node set vector corresponding to each of the directed flow graph vector sets.

[0060] Step S143: determining a first impact factor of each of the blended directed graph node set vectors.

[0061] Step S144: Optimize each of the fusion directed graph node set vectors according to the first impact factor to generate a plurality of optimized vector sets.

[0062] Step S145 , integrating the multiple optimized vector sets to generate session representation vector data of the cold chain logistics transportation session.

[0063] In a possible implementation manner, the step of determining the first impact factor includes:

[0064] The cold chain logistics application details are obtained, where the cold chain logistics application details include service information of the interactive source service of the session interactive data.

[0065] Based on the association between each of the logistics knowledge label fields and the cold chain logistics application details, a first influencing factor of each of the intersecting directed graph node set vectors is determined.

[0066] In this embodiment, since cold chain logistics transportation involves multiple different logistics knowledge label fields, the first directed graph node vector fragment set corresponding to each field has its own specific first deep learning network. For example, for the first directed graph node vector fragment set corresponding to the logistics knowledge label field "type of goods", the server determines that the corresponding first deep learning network is network A. This network A is pre-trained based on a large amount of cold chain logistics transportation related data and corresponding business logic. It has a specific network structure and parameters, and is specifically used to process vector fragment sets related to "type of goods". Similarly, for the first directed graph node vector fragment set corresponding to the "transportation starting point", there is a corresponding network B, and each network is optimized for a specific logistics knowledge label field.

[0067] Next, the server loads the first directed graph node vector fragment set corresponding to each directed flow graph vector set into the corresponding first deep learning network for blending. Taking the first directed graph node vector fragment set corresponding to "goods type" loaded into network A as an example, multiple processing layers are included inside network A. These processing layers will perform complex calculations and processing on the input vector fragment set. For example, the convolution layer in network A will perform convolution operations on each element in the vector fragment set to extract local features. During the blending process, network A will comprehensively consider the relationship between each node vector in the first directed graph node vector fragment set. Assuming that this fragment set contains vector information of nodes related to goods types such as "Class A drugs" and "Class B drugs", network A will perform blending operations based on the mutual relationship between goods types in cold chain logistics transportation (such as the differences and connections between Class A drugs and Class B drugs in terms of storage temperature, transportation requirements, etc.). Through this process, a blended directed graph node set vector corresponding to each directed flow graph vector set is generated. For the directed flow graph vector set corresponding to "goods type", after the fusion operation of network A, a fusion directed graph node set vector that integrates the characteristics and relationships of the node vectors related to "goods type" is obtained.

[0068] Then, obtain the cold chain logistics application details, which include the service information of the interactive source service of the session interaction data. For example, the service information of the interactive source service may include the origin, destination, special requirements of the customer, etc. If the drugs are transported from a specific pharmaceutical factory (such as refrigerated transportation), the drugs of this pharmaceutical factory may have special transportation requirements (such as sensitivity to temperature and humidity). Determine the first impact factor based on the association between each logistics knowledge tag field and the cold chain logistics application details. For the intersection directed graph node set vector corresponding to "cargo type", if this batch of goods is a high-value Class A drug and the transportation distance is long, according to the pre-set association evaluation algorithm, its first impact factor may be determined to be 1.2. This is because the high value and long-distance transportation make the cargo type more important in the entire cold chain logistics transportation, requiring more attention and resource allocation.

[0069] Still taking the intersecting directed graph node set vector corresponding to "cargo type" as an example, if its first impact factor is 1.2, the server will multiply each element in this vector by 1.2 for optimization. This optimization process adjusts the representation of the vector according to the actual needs of cold chain logistics and the importance of each logistics knowledge label field, so that the optimized vector can better reflect the importance of each factor in the current cold chain logistics application scenario. Through this operation, multiple optimized vector sets are generated. Each optimized vector set corresponds to a logistics knowledge label field, such as "cargo type", "transportation starting point", etc., and these optimized vector sets have been adjusted based on the first impact factor, which is more in line with the actual situation of cold chain logistics.

[0070] Finally, the server integrates multiple optimization vector sets to generate session representation vector data for the cold chain logistics transport session. This integration process is to combine each optimization vector set according to certain rules. For example, for the optimization vector sets corresponding to multiple logistics knowledge tag fields such as "cargo type", "transportation starting point", and "transportation vehicle", the server may integrate them in a weighted summation manner. Assuming that the weight of the "cargo type" optimization vector set is 0.3, the weight of the "transportation starting point" optimization vector set is 0.2, and the weight of the "transportation vehicle" optimization vector set is 0.5 (these weights are determined based on the cold chain logistics transport business logic and experience), each optimization vector set is multiplied by the corresponding weight and summed up, and the result is the session representation vector data for the cold chain logistics transport session. This session representation vector data integrates the information of various aspects of the cold chain logistics transport session (cargo, transportation starting point, transportation vehicle, etc.), and represents the characteristics of the entire cold chain logistics transport session in a unified vector form, providing an important data basis for subsequent operations such as logistics demand forecasting.

[0071] In a possible implementation, after step S120, it also includes: intercepting each of the directed flow graph vector sets with a third set number as the interception interval and a fourth set number as the moving interval, to generate a second directed graph node vector fragment set corresponding to each of the directed flow graph vector sets.

[0072] Step S140 may further include:

[0073] Each of the first directed graph node vector fragment sets is blended, and the multiple first blended directed graph node subset vectors generated by the blending are integrated to generate first session characterization vector sub-data, and each of the second directed graph node vector fragment sets is blended, and the multiple second blended directed graph node subset vectors generated by the blending are integrated to generate second session characterization vector sub-data.

[0074] The first session characterization vector sub-data and the second session characterization vector sub-data are integrated to generate the session characterization vector data of the cold chain logistics transportation session.

[0075] In this embodiment, first, each directed flow graph vector set is intercepted with the third set number as the interception interval and the fourth set number as the moving spacing, so as to generate a second directed graph node vector fragment set corresponding to each directed flow graph vector set. Assume that the third set number is 8 and the fourth set number is 3. For a certain directed flow graph vector set, such as the directed flow graph vector set corresponding to "cargo information", this directed flow graph vector set contains a series of vectors representing node relationships, weights and other information related to cargo information. The server starts from the starting position of this directed flow graph vector set and takes 8 vectors as a interception interval. For example, if this directed flow graph vector set has a total of 30 vectors, then the first interception will obtain fragments from the 1st to the 8th vector. Then according to the rule of moving spacing of 3, the next interception will be fragments from the 4th to the 11th vector, and so on. The fragment obtained by each interception is a second directed graph node vector fragment set. This process will operate on each directed flow graph vector set. Regardless of the directed flow graph vector set corresponding to "cargo information", "transportation starting point" or "transportation vehicle", the corresponding second directed graph node vector fragment set will be generated according to the same rules.

[0076] Next, the server will blend each first directed graph node vector fragment set, and integrate the multiple first blended directed graph node subset vectors generated by the blending to generate the first session representation vector sub-data, and at the same time blend each second directed graph node vector fragment set, and integrate the multiple second blended directed graph node subset vectors generated by the blending to generate the second session representation vector sub-data.

[0077] For the operation of the first directed graph node vector fragment set, the server determines the corresponding first deep learning network for each first directed graph node vector fragment set. For example, for the first directed graph node vector fragment set corresponding to "cargo type", there is a corresponding deep learning network A. This fragment set is loaded into network A for fusion. Inside network A, there are specific structures and algorithms to handle this fusion process. Each layer in network A will operate on the input vector fragments. For example, in the convolution layer, the elements in the vector will be convolved to extract local features. In the fully connected layer, the features of different nodes will be processed comprehensively. Through this process, multiple first fused directed graph node subset vectors are generated. Then these first fused directed graph node subset vectors are integrated according to certain rules to generate first session representation vector sub-data. This integration process may involve operations such as weighted summation or splicing, depending on the pre-set business logic and algorithm. For example, if there are three first fusion directed graph node subset vectors, namely vector A, vector B and vector C, and their weights are set according to their importance in the types of goods in cold chain logistics transportation, assuming that the weight of vector A is 0.3, the weight of vector B is 0.4, and the weight of vector C is 0.3, then they are integrated by weighted summation to obtain the first session representation vector sub-data.

[0078] The same process occurs for the second directed graph node vector fragment set. Taking the second directed graph node vector fragment set corresponding to the "transportation starting point" as an example, there is a corresponding deep learning network B. This fragment set is loaded into network B for fusion. Network B will process the input vector fragments according to its own structure and algorithm. For example, in the hidden layer of network B, the node information in the vector fragment will be fused according to the business logic related to the transportation starting point, such as the differences in the requirements for cold chain equipment at the transportation starting points in different regions. Through this fusion process, multiple second fused directed graph node subset vectors are generated. Then, according to similar integration rules, such as setting weights according to the importance of these subset vectors in the information related to the transportation starting point, they are integrated to generate the second session representation vector sub-data.

[0079] Finally, the server integrates the first session characterization vector sub-data with the second session characterization vector sub-data to generate the session characterization vector data of the cold chain logistics transportation session. This integration process is also based on pre-set rules. For example, if the first session characterization vector sub-data mainly focuses on the information of the goods themselves, such as the type and quantity of goods, and the second session characterization vector sub-data focuses on the information related to transportation, such as the transportation starting point and transportation vehicle. According to the overall importance ratio of goods and transportation in the cold chain logistics transportation business, it is assumed that the weight of the goods information is 0.4 and the weight of the transportation related information is 0.6. The server multiplies the first session characterization vector sub-data by 0.4 and the second session characterization vector sub-data by 0.6, and then adds the two results to obtain the final session characterization vector data of the cold chain logistics transportation session. This session characterization vector data integrates the information obtained by processing the cold chain logistics transportation session from different angles (the first directed graph node vector fragment set and the second directed graph node vector fragment set), and represents the characteristics of the entire cold chain logistics transportation session in a comprehensive vector form, which provides a complete data basis for subsequent operations such as logistics demand forecasting based on this session characterization vector data.

[0080] In a possible implementation, step S150 includes:

[0081] Step S151: Load the session representation vector data into the second deep learning network to perform logistics demand prediction and generate a first logistics demand prediction heat map.

[0082] In this embodiment, the second deep learning network is pre-trained with a large amount of cold chain logistics and transportation related data, has a specific network structure and parameters, and can effectively analyze and process the input session representation vector data. When the session representation vector data is input into the second deep learning network, each layer in the second deep learning network starts to work. For example, the input layer receives the session representation vector data and then passes it to the hidden layer. The hidden layer contains multiple neurons, which perform complex calculations on the input data according to pre-set weights and activation functions. In this process, the second deep learning network analyzes according to the business logic and data characteristics of cold chain logistics and transportation. Assuming that the session representation vector data contains information such as the type, quantity, transportation starting point, and transportation vehicle of the goods, the network will extract and analyze the features of the data in the hidden layer based on this information. For the type of goods, if it is a Class A drug, the network will analyze its potential demand for temperature, humidity, transportation time, etc. based on the relevant data and experience of the previous transportation of Class A drugs (these data have been learned during the network training process). For the transportation vehicle, if it is a certain model of refrigerated truck, the second deep learning network will consider the impact of its refrigeration capacity, volume, and other factors on logistics demand. After multiple layers of processing in the hidden layer, the first logistics demand prediction heat map is generated in the output layer.

[0083] This first logistics demand forecast heat map shows the prediction probability or importance of different logistics demand knowledge points in a data visualization way (the visualization here is more from the perspective of data representation, which is easy to understand its meaning). For example, in the heat map, the area corresponding to the logistics demand knowledge point of temperature requirement may show a higher heat value. This is because according to the type of goods in the session representation vector data (such as Class A drugs), there are strict requirements for temperature in cold chain logistics, so the network judges that the temperature requirement is more important in this cold chain logistics transportation session based on the knowledge obtained through training and the current input data. Similarly, if the quantity of goods is large, the area corresponding to the transportation time requirement may also show a higher heat, because a large number of goods may require a more reasonable transportation time arrangement to ensure the freshness and quality of the goods.

[0084] Step S152: determining the logistics demand knowledge points of the cold chain logistics transportation session based on the first logistics demand prediction heat map.

[0085] In this embodiment, if the regional heat value corresponding to a logistics demand knowledge point in the first logistics demand prediction heat map exceeds a preset threshold, then this knowledge point is determined as an important logistics demand knowledge point for the cold chain logistics transportation session. For example, in addition to the temperature requirements and transportation time requirements mentioned above, if the regional heat corresponding to the cargo packaging requirements shown in the first logistics demand prediction heat map is also high, then the cargo packaging requirements are also determined as the logistics demand knowledge point for this cold chain logistics transportation session. The server can provide decision support for the actual arrangement of cold chain logistics transportation based on these determined logistics demand knowledge points. For example, adjust the refrigeration parameters of the refrigerated truck according to the temperature requirements, plan the best transportation route according to the transportation time requirements, and ensure the safety of the cargo during transportation according to the cargo packaging requirements. In this way, through the analysis of the first logistics demand prediction heat map, the key logistics demand knowledge points in the cold chain logistics transportation session can be accurately determined, thereby effectively guiding each link of the cold chain logistics transportation.

[0086] In a possible implementation, the method further includes:

[0087] Step S101, obtaining a sample conversation interaction data sequence, wherein the sample conversation interaction data sequence includes template conversation interaction data of multiple template cold chain logistics transportation conversations and prior logistics demand knowledge point data of each template cold chain logistics transportation conversation.

[0088] For example, the sample conversation interaction data sequence includes template conversation interaction data for different types of goods (such as Class A drugs, Class B drugs, etc.) in different transportation scenarios (such as from factory to wholesaler, from wholesaler to retailer, etc.). For a template cold chain logistics transportation conversation about the transportation of Class A drugs from factory to wholesaler, its template conversation interaction data may include detailed information of the goods (such as the type is A drug (A drug is a certain drug under Class A drugs), the quantity is 100 boxes, the weight is 1 ton, etc.), the starting point of transportation (the specific address of a factory warehouse), the end point of transportation (the specific address of a wholesaler warehouse), the transportation time requirement (hope to be delivered within 3 days), the temperature requirement (0 - 4 degrees Celsius), etc. At the same time, each template cold chain logistics transportation conversation has corresponding prior logistics demand knowledge point data. For example, for this conversation about the transportation of Class A drugs, the prior logistics demand knowledge point data may clearly indicate that the temperature requirement is a very critical factor, the transportation time requirement is also very important, and it may also include some special packaging requirements (such as using breathable but moisturizing packaging materials).

[0089] Step S102, parsing the template session interaction path data of multiple logistics knowledge tag fields from the template session interaction data, and encoding the parsed multiple template session interaction path data into multiple template directed flow graph vector sets.

[0090] For example, for the template session interaction data of the above-mentioned Class A drug transportation, the server constructs a dictionary of logistics knowledge label fields, which includes label fields such as cargo information, transportation starting point, transportation end point, transportation time requirements, temperature requirements, and packaging requirements. Then, for each text segment in the template session interaction data, the text matching algorithm is used to compare it with the knowledge label fields in the dictionary one by one in a predefined order. For example, the text segment "100 boxes of Class A drugs" is determined to belong to the logistics knowledge label field of cargo information through comparison. After that, for each template cold chain logistics transportation session, it is sorted in chronological order or logical order from the starting point. When a text segment related to a logistics knowledge label field is identified, it is used as a node of the graph structure corresponding to the logistics knowledge label field. For example, when the node "100 boxes of Class A drugs" is identified, its appearance position in the template cold chain logistics transportation session (assuming it is at the beginning of the order information filling) and its relationship with the previous related node (if it is a starting node, it is marked as a starting node relationship) are recorded, and the specific content of the logistics knowledge label field corresponding to the node (i.e., 100 boxes of Class A drugs) is recorded. When constructing the template conversation interaction path corresponding to the logistics knowledge label field, all nodes are connected in order. For example, under the label field of cargo information, if there are nodes about the weight of drug A, they are connected in order to obtain template conversation interaction path data containing node order, node relationship and node specific content. Then, the node corresponding to each text fragment is used as the vertex in the graph structure of the corresponding logistics knowledge label field, and the connection relationship of the edge between the vertices is determined according to the node order and relationship in the template conversation interaction path data, and an initial weight value is assigned to each edge. This initial weight value is determined based on the distance between the nodes and the degree of association between the nodes. For example, in the graph structure of cargo information, the distance between the "100 boxes of drug A" node and the "1 ton weight" node is close and the degree of association is high, and a higher weight value is assigned according to the set algorithm. After that, the isolated nodes in the initial graph structure are detected. If the isolated nodes are caused by data errors or incompleteness, the connection relationship is supplemented according to their context information; if there is no independent information associated with other nodes, it is determined whether to retain the node based on the pre-defined business integrity template, and the connection of the edge is re-optimized according to the logistics business logic to generate an optimized graph structure. Finally, the nodes in the optimized graph structure are numbered in depth-first search order. For each edge, it is encoded according to the numbers of the two connected nodes and the weight value of the edge. The encoding vectors of all edges are combined to generate a template directed flow graph vector set.

[0091] Step S103, taking the first set number as the truncation interval and the second set number as the moving interval, truncate each of the template directed flow graph vector sets to generate a first template directed graph node vector fragment set corresponding to each of the template directed flow graph vector sets.

[0092] Step S104, loading the first template directed graph node vector fragment set corresponding to each of the template directed flow graph vector sets into the corresponding first deep learning network for blending, and integrating the template blending directed graph node set vectors corresponding to each of the template directed flow graph vector sets generated by blending to generate template session representation vector data.

[0093] Assume that the first set number is 5 and the second set number is 2. For the previously generated template directed flow graph vector set corresponding to the transport of Class A drugs, if this template directed flow graph vector set contains 20 code vectors, then starting from the starting position, with 5 code vectors as a cut interval, the first cut is to obtain fragments from the 1st to the 5th code vector, and then according to the rule of moving spacing of 2, the next cut is from the 3rd to the 7th code vector fragment, and so on, thus generating the first template directed graph node vector fragment set.

[0094] For the first template directed graph node vector fragment set corresponding to the transport of Class A drugs, it is loaded into the corresponding first deep learning network. This first deep learning network has a specific structure, such as a graph self-attention layer, a feature fusion layer, etc. In the network, the first template directed graph node vector fragment set is firstly processed by the graph self-attention layer. The graph self-attention layer calculates the attention weight of each node to other nodes based on the relationship between the nodes and its own characteristics. For example, in the node vector fragment related to the cargo information, if there are nodes representing the quantity and weight of drug A, the graph self-attention layer calculates the attention weight between the two nodes based on the relationship between the quantity and weight in the cold chain logistics transportation (such as different quantities of drug A may correspond to different weight range standards). After the graph self-attention processing, the processing results are fused according to the feature fusion layer to generate the template fusion directed graph node set vector corresponding to each template directed flow graph vector set. After that, these template fusion directed graph node set vectors are integrated. For example, different template fusion directed graph node set vectors may be combined by weighted summation or splicing to generate template session representation vector data.

[0095] Step S105: load the template session representation vector data into the second deep learning network for logistics demand prediction, generate a second logistics demand prediction heat map, and calculate the prediction error parameter based on the second logistics demand prediction heat map and the corresponding prior logistics demand knowledge point data.

[0096] Step S106: training a plurality of the first deep learning networks and the second deep learning network according to the prediction error parameter.

[0097] In this embodiment, in the hidden layer of the network, analysis and processing are performed based on comprehensive information such as cargo information and transportation requirements, and a second logistics demand forecast heat map is generated in the output layer. This second logistics demand forecast heat map shows the prediction probability or importance of different logistics demand knowledge points. Then, the second logistics demand forecast heat map is compared with the corresponding prior logistics demand knowledge point data to calculate the prediction error parameter. For example, if the prior logistics demand knowledge point data indicates that temperature requirements are very important factors, and the second logistics demand forecast heat map shows that the prediction importance of temperature requirements is low, then a large prediction error will occur.

[0098] The multiple first deep learning networks and the second deep learning network are trained according to the prediction error parameters. If the prediction error is large, it means that the parameters of the network need to be adjusted. According to the prediction error parameters, the weights and other parameters in the multiple first deep learning networks and the second deep learning network are adjusted by means of back propagation algorithms, etc., so that the network can more accurately predict the logistics demand knowledge points in subsequent predictions, continuously improve the accuracy and reliability of the network, and thus better adapt to the logistics demand prediction tasks in the cold chain logistics transportation session.

[0099] In a possible implementation, after step S103, the method further includes: intercepting each of the template directed flow graph vector sets with a third set number as the interception interval and a fourth set number as the moving interval, and generating a second template directed graph node vector fragment set corresponding to each of the template directed flow graph vector sets.

[0100] Step S104 includes:

[0101] Step S1041, load the first template directed graph node vector fragment set corresponding to each of the template directed flow graph vector sets into the corresponding first deep learning network for blending, and integrate the first template blended directed graph node subset vectors corresponding to each of the template directed flow graph vector sets generated by blending to generate the first template session representation vector sub-data.

[0102] Step S1042, loading the second template directed graph node vector fragment set corresponding to each of the template directed flow graph vector sets into the corresponding first deep learning network for blending, and integrating the second template blended directed graph node subset vectors corresponding to each of the template directed flow graph vector sets generated by blending to generate second template session representation vector sub-data.

[0103] Step S1043: Integrate the first template session representation vector sub-data and the second template session representation vector sub-data to generate the template session representation vector data of the template cold chain logistics transportation session.

[0104] Assume that the third set quantity is 8 and the fourth set quantity is 3. Also, take the template directed flow graph vector set corresponding to the transportation of Class A drugs as an example. If this set contains 30 code vectors, starting from the starting position, with 8 code vectors as a cut interval, the first cut is the fragment from the 1st to the 8th code vector, and then according to the rule of moving interval of 3, the next cut is the fragment from the 4th to the 11th code vector, and so on, to generate the second template directed graph node vector fragment set.

[0105] For the first template directed graph node vector fragment set corresponding to the transport of Class A drugs, after being loaded into the first deep learning network, the first template fusion directed graph node subset vector is generated through network processing, and then integrated to obtain the first template session representation vector sub-data. For the second template directed graph node vector fragment set, it is also loaded into the first deep learning network, and after processing, integrated to obtain the second template session representation vector sub-data.

[0106] Finally, according to pre-set rules, the first template session characterization vector sub-data and the second template session characterization vector sub-data may be assigned certain weights respectively (such as the first template session characterization vector sub-data weight is 0.4, and the second template session characterization vector sub-data weight is 0.6), and then they are weighted and summed according to the weights to obtain the template session characterization vector data of the template cold chain logistics transportation session.

[0107] In a possible implementation, each of the first deep learning networks includes a graph self-attention layer and a feature fusion layer, and step S104 may further include:

[0108] According to the graph self-attention layer, graph self-attention processing is performed on the first template directed graph node vector fragment set corresponding to each template directed flow graph vector set to generate a first template graph self-attention processing result.

[0109] The first template graph self-attention processing results corresponding to each template directed flow graph vector set are fused according to the feature fusion layer to generate a template fusion directed graph node set vector corresponding to each template directed flow graph vector set.

[0110] The template fusion directed graph node set vectors corresponding to each of the template directed flow graph vector sets are integrated to generate template session representation vector data.

[0111] In a possible implementation, each of the first deep learning networks further includes a spatial embedding representation layer, and the graph self-attention layer performs graph self-attention processing on the first template directed graph node vector fragment set corresponding to each template directed flow graph vector set, and after generating the first template graph self-attention processing result, it also includes:

[0112] The spatial embedding representation layer is used to perform spatial embedding representation on the first template directed graph node vector fragment set corresponding to each template directed flow graph vector set to generate a first template spatial embedding representation vector fragment set.

[0113] The step of blending the first template graph self-attention processing results corresponding to each template directed flow graph vector set according to the feature blending layer to generate a template blending directed graph node set vector corresponding to each template directed flow graph vector set includes:

[0114] The first template space embedding representation vector fragment set corresponding to each template directed flow graph vector set is blended according to the feature blending layer to generate a template blending directed graph node set vector corresponding to each template directed flow graph vector set.

[0115] In this embodiment, when the server loads the first template directed graph node vector fragment set corresponding to each template directed flow graph vector set into the corresponding first deep learning network for blending, and generates template session representation vector data, the operation of the graph self-attention layer is first involved. For each first template directed graph node vector fragment set corresponding to a template directed flow graph vector set, for example, when processing the first template directed graph node vector fragment set corresponding to a template directed flow graph vector set related to the cold chain transportation of Class A drugs, the graph self-attention layer will perform graph self-attention processing on this fragment set based on the business logic relationship in the cold chain logistics transportation. In this fragment set, node vectors related to cargo information such as the quantity, weight, and variety of goods may be included, as well as transportation-related node vectors such as the transportation starting point and estimated arrival time. The graph self-attention layer calculates the attention weight based on the characteristics of the node itself and the relationship between the nodes. For example, for the node "100 boxes of A medicine" and the node "1 ton weight" in the cargo information, since there is a certain correlation between the quantity and weight of A medicine in cold chain logistics (different quantities of A medicine will correspond to different total weights, and the weight will affect the transportation cost, the choice of transportation tools, etc.), the graph self-attention layer will calculate the attention weight between the two nodes based on this inherent logical relationship. If in the transportation-related nodes, the transportation starting point is a small warehouse close to the factory and the estimated arrival time is short, the graph self-attention layer will take into account this relationship between time and place and calculate the attention weight between the two nodes of the transportation starting point and the estimated arrival time. Through such calculations, the importance of each node in the entire fragment set is re-evaluated, thereby generating the first template graph self-attention processing result. This result contains the node vector information after the attention weight is adjusted, which can better reflect the relative importance of each node under the cold chain logistics transportation business logic.

[0116] If the first deep learning network also includes a spatial embedding representation layer, then after the graph self-attention layer generates the first template graph self-attention processing result, the spatial embedding representation layer will perform spatial embedding representation on the first template directed graph node vector fragment set corresponding to each template directed flow graph vector set to generate the first template spatial embedding representation vector fragment set. In the context of cold chain logistics transportation, the representation of node vectors corresponding to different logistics knowledge label fields in space has special significance. For example, node vectors related to cargo information (such as the quantity and weight of drug A) and node vectors related to transportation (such as the transportation starting point, estimated arrival time, etc.) can be embedded in a feature space related to cold chain logistics transportation. The dimension of this feature space may be determined based on various factors in cold chain logistics transportation, such as the physical properties of the cargo, the geographical space of transportation, time requirements, etc. For the node vectors in the cargo information, the spatial embedding representation layer will embed and represent these node vectors in the feature space based on factors such as the type characteristics, quantity and weight of drug A (affecting the space occupied by storage and transportation). Similarly, for transport-related node vectors, spatial embedding representation will be performed based on factors such as the geographical location of the transport starting point (climate, traffic conditions, etc. in different regions), estimated arrival time (related to transport distance, transport speed, etc.). The first template spatial embedding representation vector fragment set generated in this way can better reflect the spatial relationship of different types of node vectors in the large framework of cold chain logistics and transportation, which is helpful for subsequent feature fusion operations.

[0117] Next, the first template space embedding representation vector fragment set corresponding to each template directed flow graph vector set is fused according to the feature fusion layer to generate a template fusion directed graph node set vector corresponding to each template directed flow graph vector set. In this fusion process, the feature fusion layer will comprehensively consider various information in the first template space embedding representation vector fragment set. Taking the cold chain transportation of Class A drugs as an example, in the first template space embedding representation vector fragment set, different types of node vectors such as cargo information and transportation information have specific representations in the feature space. The feature fusion layer will blend factors such as the quantity and weight of A drugs in the cargo information with factors such as the transportation starting point and expected arrival time in the transportation information. For example, a drug that is prone to expiration may need to depart from a transportation starting point that is closer to the factory and has convenient transportation, and be delivered within a shorter expected arrival time. The feature fusion layer will fuse the relevant node vector features according to this inherent connection. Through this fusion operation, a template fusion directed graph node set vector corresponding to each template directed flow graph vector set is generated. This vector set integrates information on goods and transportation, and this information is the result of spatial embedding representation and feature fusion processing, which is more in line with the actual business logic of cold chain logistics and transportation.

[0118] Finally, the template intersection directed graph node set vectors corresponding to each template directed flow graph vector set are integrated to generate template session representation vector data. For different template directed flow graph vector sets (which may correspond to different types of cold chain logistics transportation scenarios, such as the transportation of Class A drugs, Class B drugs, etc.), their respective template intersection directed graph node set vectors contain comprehensive information such as goods and transportation in their respective scenarios. During the integration process, the server will operate according to pre-set rules. For example, if there are two template directed flow graph vector sets, one is related to the transportation of Class A drugs, and the other is related to the transportation of Class B drugs, their corresponding template intersection directed graph node set vectors may be integrated. The weights may be determined based on factors such as the common proportion and importance of these two types of goods in cold chain logistics transportation. Assuming that the weight of Class A drug transportation in the overall cold chain logistics transportation is 0.4, and the weight of Class B drug transportation is 0.6, then the template intersection directed graph node set vector corresponding to Class A drug transportation is multiplied by 0.4, and the template intersection directed graph node set vector corresponding to Class B drug transportation is multiplied by 0.6, and then the two results are summed or concatenated to finally generate the template session representation vector data. This template session representation vector data integrates various information such as goods and transportation corresponding to each template directed flow graph vector set, and represents the characteristics of the entire template cold chain logistics transportation session in a unified form, providing an effective data basis for subsequent operations such as logistics demand prediction in the second deep learning network.

[0119] Figure 2 The hardware structure of the cold chain logistics transportation demand prediction system 100 based on data mining for implementing the cold chain logistics transportation demand prediction method based on data mining provided by an embodiment of the present invention is shown as follows: Figure 2 As shown, the cold chain logistics transportation demand prediction system 100 based on data mining may include a processor 110, a machine-readable storage medium 120, a bus 130 and a communication unit 140.

[0120] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the cold chain logistics transportation demand forecasting system 100 based on data mining uses to execute or use to complete the exemplary method described in the present invention.

[0121] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the cold chain logistics transportation demand prediction method based on data mining in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0122] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned cold chain logistics transportation demand prediction system 100 based on data mining. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0123] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer executable instructions are preset. When the processor executes the computer executable instructions, the cold chain logistics transportation demand prediction method based on data mining as mentioned above is implemented.

[0124] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A method for forecasting cold chain logistics transportation demand based on data mining, characterized in that: The method comprises: Get the session interaction data of the cold chain logistics transportation session; Parsing the session interaction path data of a plurality of logistics knowledge tag fields from the session interaction data, and encoding the parsed plurality of the session interaction path data into a plurality of directed flow graph vector sets; Taking the first set number as the interception interval and the second set number as the moving interval, intercepting each of the directed flow graph vector sets to generate a first directed graph node vector fragment set corresponding to each of the directed flow graph vector sets; Blending each of the first directed graph node vector fragment sets, and integrating a plurality of blended directed graph node set vectors generated by blending to generate session representation vector data of the cold chain logistics transportation session; Determining a logistics demand prediction result of the cold chain logistics transportation session according to the session characterization vector data; The step of parsing the session interaction path data of the plurality of logistics knowledge tag fields from the session interaction data, and encoding the parsed plurality of the session interaction path data into a plurality of directed flow graph vector sets comprises: Pre-building a dictionary of logistics knowledge label fields, wherein the dictionary contains knowledge label fields involved in the field of cold chain logistics and transportation; For each text segment in the conversation interaction data, a text matching algorithm is used to compare the text segments with the knowledge tag fields in the dictionary one by one in a predefined order, and the logistics knowledge tag field to which each text segment belongs is output according to the comparison result; For each cold chain logistics transport session, starting from the starting point of the cold chain logistics transport session, the cold chain logistics transport session interactions are sorted in chronological order or logical order. During the sorting process, when a text segment related to a logistics knowledge label field is identified, the text segment is used as a node of the graph structure corresponding to the logistics knowledge label field, and the location of the node in the cold chain logistics transport session and its relationship with the previous related node are recorded, wherein the relationship includes a sequential relationship based on chronological order or a causal relationship based on business logic; and, for each node, the specific content of the logistics knowledge label field corresponding to the node is recorded at the same time; When constructing the conversation interaction path corresponding to the logistics knowledge tag field, all nodes are connected in order to obtain conversation interaction path data including node order, node relationship and node specific content; The node corresponding to each text segment is used as a vertex in the graph structure of the corresponding logistics knowledge label field. The connection relationship of the edges between the vertices is determined according to the node sequence and relationship in the conversation interaction path data. An initial weight value is assigned to each edge to obtain an initial graph structure including nodes and preliminary connection relationships. The initial weight value is determined based on the distance between the nodes and the degree of association between the nodes. Detecting isolated nodes in the initial graph structure, wherein the isolated nodes refer to nodes that have no connection relationship with other nodes in the initial graph structure; If the isolated node is caused by data error or incompleteness, the connection relationship is supplemented according to the context information of the isolated node. If the isolated node has no independent information associated with other nodes, it is determined whether the isolated node needs to be retained based on a predefined business integrity template, and the connection of the edges in the initial graph structure is re-optimized according to the logistics business logic to generate an optimized graph structure. Numbering the nodes in the optimized graph structure according to a depth-first search order or a breadth-first search order, and, for each edge, encoding according to the numbers of two nodes connected by the edge and the weight value of the edge, specifically combining the start node number, the end node number and the weight value of the edge into a coding vector; Combine the encoding vectors of all edges to generate a directed flow graph vector set; The step of blending each of the first directed graph node vector fragment sets and integrating the multiple blended directed graph node set vectors generated by the blending to generate the session representation vector data of the cold chain logistics transportation session includes: Determine a first deep learning network corresponding to each of the first directed graph node vector fragment sets; Loading the first directed graph node vector fragment set corresponding to each of the directed flow graph vector sets into the corresponding first deep learning network for blending, and generating a blended directed graph node set vector corresponding to each of the directed flow graph vector sets; Determine a first impact factor of each of the blended directed graph node set vectors; Optimizing each of the fusion directed graph node set vectors according to the first influencing factor to generate a plurality of optimized vector sets; Integrating the multiple optimized vector sets to generate session representation vector data of the cold chain logistics transportation session; The determining of the logistics demand prediction result of the cold chain logistics transportation session according to the session characterization vector data includes: Loading the session representation vector data into a second deep learning network to perform logistics demand prediction, and generating a first logistics demand prediction heat map; The logistics demand knowledge points of the cold chain logistics transportation session are determined based on the first logistics demand prediction heat map.

2. The method for predicting cold chain logistics transportation demand based on data mining according to claim 1 is characterized in that: The step of determining the first impact factor comprises: Acquire cold chain logistics application details, where the cold chain logistics application details include service information of the interactive source service of the session interactive data; Based on the association between each of the logistics knowledge label fields and the cold chain logistics application details, a first influencing factor of each of the intersecting directed graph node set vectors is determined.

3. The method for predicting cold chain logistics transportation demand based on data mining according to claim 1 is characterized in that: After parsing the session interaction path data of the plurality of logistics knowledge tag fields from the session interaction data and encoding the parsed plurality of the session interaction path data into a plurality of directed flow graph vector sets, the method further includes: Taking the third set number as the interception interval and the fourth set number as the moving interval, intercepting each of the directed flow graph vector sets to generate a second directed graph node vector fragment set corresponding to each of the directed flow graph vector sets; The step of blending each of the first directed graph node vector fragment sets and integrating the multiple blended directed graph node set vectors generated by the blending to generate the session representation vector data of the cold chain logistics transportation session includes: Blending each of the first directed graph node vector fragment sets, and integrating the multiple first blended directed graph node subset vectors generated by the blending to generate first session representation vector sub-data, and blending each of the second directed graph node vector fragment sets, and integrating the multiple second blended directed graph node subset vectors generated by the blending to generate second session representation vector sub-data; The first session characterization vector sub-data and the second session characterization vector sub-data are integrated to generate the session characterization vector data of the cold chain logistics transportation session.

4. The method for predicting cold chain logistics transportation demand based on data mining according to claim 1 is characterized in that: The method further comprises: Acquire a sample session interaction data sequence, wherein the sample session interaction data sequence includes template session interaction data of multiple template cold chain logistics transportation sessions and prior logistics demand knowledge point data of each template cold chain logistics transportation session; Parsing template session interaction path data of a plurality of logistics knowledge tag fields from the template session interaction data, and encoding the parsed plurality of template session interaction path data into a plurality of template directed flow graph vector sets; Taking the first set number as the interception interval and the second set number as the moving interval, intercepting each of the template directed flow graph vector sets to generate a first template directed graph node vector fragment set corresponding to each of the template directed flow graph vector sets; Loading the first template directed graph node vector fragment set corresponding to each of the template directed flow graph vector sets into the corresponding first deep learning network for blending, and integrating the template blended directed graph node set vectors corresponding to each of the template directed flow graph vector sets generated by blending to generate template session representation vector data; Loading the template session representation vector data into a second deep learning network to perform logistics demand prediction, generating a second logistics demand prediction heat map, and calculating a prediction error parameter based on the second logistics demand prediction heat map and corresponding prior logistics demand knowledge point data; A plurality of the first deep learning networks and the second deep learning networks are trained according to the prediction error parameter.

5. The method for predicting cold chain logistics transportation demand based on data mining according to claim 4 is characterized in that: After parsing the template session interaction path data of the plurality of logistics knowledge tag fields from the template session interaction data, and encoding the parsed plurality of the template session interaction path data into a plurality of template directed flow graph vector sets, the method further comprises: Taking the third set number as the interception interval and the fourth set number as the moving interval, intercepting each of the template directed flow graph vector sets to generate a second template directed graph node vector fragment set corresponding to each of the template directed flow graph vector sets; The step of loading the first template directed graph node vector fragment set corresponding to each of the template directed flow graph vector sets into the corresponding first deep learning network for blending, and integrating the template blended directed graph node set vectors corresponding to each of the template directed flow graph vector sets generated by blending to generate template session representation vector data includes: Loading the first template directed graph node vector fragment set corresponding to each of the template directed flow graph vector sets into the corresponding first deep learning network for blending, and integrating the first template blended directed graph node subset vectors corresponding to each of the template directed flow graph vector sets generated by blending to generate first template session representation vector sub-data; Loading the second template directed graph node vector fragment set corresponding to each template directed flow graph vector set into the corresponding first deep learning network for blending, and integrating the second template blended directed graph node subset vectors corresponding to each template directed flow graph vector set generated by blending to generate second template session representation vector sub-data; The first template session representation vector sub-data and the second template session representation vector sub-data are integrated to generate the template session representation vector data of the template cold chain logistics transportation session.

6. The method for predicting cold chain logistics transportation demand based on data mining according to claim 4 is characterized in that: Each of the first deep learning networks includes a graph self-attention layer and a feature fusion layer, wherein the first template directed graph node vector fragment set corresponding to each of the template directed flow graph vector sets is loaded into the corresponding first deep learning network for fusion, and the template fusion directed graph node set vectors corresponding to each of the template directed flow graph vector sets generated by fusion are integrated to generate template session representation vector data, including: Performing graph self-attention processing on the first template directed graph node vector fragment set corresponding to each template directed flow graph vector set according to the graph self-attention layer to generate a first template graph self-attention processing result; The first template graph self-attention processing results corresponding to each template directed flow graph vector set are blended according to the feature blending layer to generate a template blending directed graph node set vector corresponding to each template directed flow graph vector set; The template fusion directed graph node set vectors corresponding to each of the template directed flow graph vector sets are integrated to generate template session representation vector data.

7. The method for predicting cold chain logistics transportation demand based on data mining according to claim 6 is characterized in that: Each of the first deep learning networks also includes a spatial embedding representation layer, and after performing graph self-attention processing on the first template directed graph node vector fragment set corresponding to each template directed flow graph vector set according to the graph self-attention layer and generating the first template graph self-attention processing result, it also includes: Performing spatial embedding representation on the first template directed graph node vector fragment set corresponding to each template directed flow graph vector set according to the spatial embedding representation layer to generate a first template spatial embedding representation vector fragment set; The step of blending the first template graph self-attention processing results corresponding to each template directed flow graph vector set according to the feature blending layer to generate a template blending directed graph node set vector corresponding to each template directed flow graph vector set includes: The first template space embedding representation vector fragment set corresponding to each template directed flow graph vector set is blended according to the feature blending layer to generate a template blending directed graph node set vector corresponding to each template directed flow graph vector set.

8. A cold chain logistics transportation demand forecasting system based on data mining, characterized in that: The cold chain logistics transportation demand prediction system based on data mining includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the cold chain logistics transportation demand prediction method based on data mining as described in any one of claims 1 to 7 above.

Images