A data fusion method and system based on logistics distribution knowledge graph

Abstract

The present invention relates to the field of logistics and distribution data fusion technology, and in particular to a data fusion method and system based on logistics and distribution knowledge graph. The method comprises the following steps: extracting raw data and performing independent representation learning and unified representation processing; extracting context information of the unified representation and performing dynamic perception embedding to obtain semantic enhancement features; integrating spatiotemporal and dependency features and physical characteristic features based on semantic enhancement features, using graph neural networks for message transmission and relationship inference, and finally determining a multi-modal relationship set; performing global entity clustering and local feature refinement matching based on semantic enhancement features, and performing hierarchical entity alignment in combination with multi-modal relationship sets to obtain an aligned entity set; constructing an entity filling graph based on the aligned entity set, and supplementing multi-dimensional relationship attributes, and finally generating a logistics and distribution knowledge graph. The present invention realizes a more comprehensive and intelligent logistics and distribution knowledge graph through logistics and distribution data fusion technology.

Description

Technical Field

[0001] The present invention relates to the technical field of logistics distribution data fusion, and in particular to a data fusion method and system based on a logistics distribution knowledge graph. Background Art

[0002] As an effective way of knowledge representation and management, knowledge graphs have been gradually introduced into the field of logistics and distribution. The data fusion method based on the logistics and distribution knowledge graph aims to integrate data from different sources and structures into a unified knowledge graph, thereby achieving data interconnection and deep knowledge mining. Although existing methods have made some progress in data fusion of logistics and distribution knowledge graphs, there are still some drawbacks that are difficult to overcome, especially in dealing with physical level associations and cross-source entity alignment:

[0003] 1. Difficulty in handling associations at the physical level: Existing knowledge graph construction methods focus more on associations at the information level, such as the association between orders and goods, and the association between vehicles and drivers. There is a lack of effective modeling and representation capabilities for the physical constraints and interactions involved in the logistics and distribution process, such as the stacking method of goods, the movement trajectory of vehicles, and the stress conditions of goods. Due to the lack of abstraction and integration of physical models, existing knowledge graphs are difficult to reason based on physical laws, such as predicting vehicle arrival time (ETA), evaluating the stability of cargo stacking, and analyzing the stress risks of special goods during transportation. This limits the application of knowledge graphs in logistics and distribution optimization and risk prediction.

[0004] 2. Difficulty in cross-source entity alignment: Logistics and distribution data comes from a wide range of sources, and there are significant differences in entity naming, attribute descriptions, and data formats used by different systems, which makes entity alignment more difficult. For example, the same goods have different names or codes in different systems. When dealing with cross-source entities, existing entity alignment methods often only focus on the attribute information of the entity itself, while ignoring the contextual information of the entity, such as the relationship between entities, the spatiotemporal information of entities, etc. This limits the accuracy of alignment. Summary of the invention

[0005] Based on this, it is necessary to provide a data fusion method and system based on the logistics distribution knowledge graph to solve at least one of the above technical problems.

[0006] To achieve the above purpose, a data fusion method based on logistics distribution knowledge graph includes the following steps:

[0007] Step S1: Obtaining original data of logistics distribution; performing independent representation learning of heterogeneous data on the original data of logistics distribution to obtain an independent feature vector set; performing unified feature representation processing on the independent feature vector set to obtain a unified feature representation;

[0008] Step S2: extracting context information from the unified feature representation and performing dynamic perception embedding to obtain dynamic context embedding data; performing multi-dimensional semantic fusion on the dynamic context embedding data to obtain semantic enhancement features;

[0009] Step S3: According to the semantic enhancement features, the spatiotemporal and dependent features are integrated and processed to obtain an enhanced feature map; according to the enhanced feature map, the physical motion model features are fused to obtain a motion model fusion map; according to the enhanced feature map, the cargo stacking model features are fused to obtain a stacking model fusion map; according to the enhanced feature map, the mechanical analysis features are fused to obtain a mechanical model fusion map; according to the motion model fusion map, the stacking model fusion map and the mechanical model fusion map, the graph neural network message transmission and relationship inference are performed to obtain a relationship inference score; according to the relationship inference score, multi-modal feature fusion and relationship determination are performed to obtain a multi-modal relationship set;

[0010] Step S4: performing global entity clustering analysis based on semantic enhancement features to obtain a global clustering entity set; performing local feature refinement matching on the global clustering entity set to obtain local matching entity pairs; performing hierarchical entity alignment based on the multimodal relationship set, the local matching entity pairs and the global clustering entity set to obtain an aligned entity set;

[0011] Step S5: construct an entity filling graph based on the aligned entity set to obtain an entity filling graph; supplement the entity filling graph with multi-dimensional relationship attributes based on the multi-modal relationship set to obtain an attributed relationship graph; generate a logistics distribution knowledge graph based on the attributed relationship graph to obtain a logistics distribution knowledge graph.

[0012] The present invention effectively solves the problems of data islands and data heterogeneity by cleaning, converting and preliminarily characterizing raw data from different sources and structures, extracts key features of data of different modalities, and lays a solid foundation for subsequent cross-modal fusion. By integrating pre-constructed logistics domain knowledge, the features of the preliminary characterization are deeply semantically enhanced, which significantly improves the accuracy and richness of the expression of logistics entity features, enabling the model to better understand the complex semantics of logistics scenarios. Using multiple technical means such as physical models and graph neural networks, the complex interaction patterns between logistics entities are finely identified and inferred from multiple dimensions, capturing the deep associations hidden behind the data. Through multi-level feature matching and fusion, entity alignment across data sources is achieved, effectively eliminating entity redundancy and ambiguity, and ensuring the uniqueness and accuracy of entities in the knowledge graph. The aligned entities and relationships are integrated to construct a high-quality logistics distribution knowledge graph, and the interpretability, information richness and query efficiency of the graph are improved through hierarchical organization, attribute supplementation and index construction. Therefore, the present invention provides a data fusion method based on the logistics and distribution knowledge graph, which innovatively integrates physical movement, cargo stacking and mechanical analysis models into the relationship extraction stage to achieve modeling and reasoning of physical level associations in the logistics and distribution process, thereby effectively solving the problem that existing methods are difficult to handle physical associations; at the same time, through multi-level feature fusion from unified feature representation, cross-modal alignment to semantic enhancement, and adopting a hierarchical entity alignment strategy combining global clustering and local refined matching, multi-modal relationship information is integrated, which significantly improves the accuracy and efficiency of cross-source entity alignment, and overcomes the difficulties of existing methods in cross-source entity alignment.

[0013] Preferably, step S1 comprises the following steps:

[0014] Step S11: obtaining original logistics distribution data; performing logistics data preprocessing on the original logistics distribution data to obtain preprocessed logistics data;

[0015] Step S12: Performing independent representation learning of heterogeneous data on the pre-processed logistics data to obtain an independent feature vector set;

[0016] Step S13: using the pre-built logistics domain knowledge base to perform cross-modal alignment on the independent feature vector sets to obtain an aligned feature space set;

[0017] Step S14: performing unified feature representation fusion on the aligned feature space set to obtain a unified feature representation.

[0018] The present invention effectively improves the quality and availability of data by collecting raw logistics data from different systems and performing data cleaning and format conversion. Data with duplicate, missing or incorrect formats is eliminated to ensure the accuracy of subsequent analysis and modeling. Converting heterogeneous data into a unified structured format lays the foundation for subsequent feature extraction and fusion, so that data from different sources can be effectively integrated and utilized. For example, the unified timestamp format facilitates subsequent time series analysis, and the unified geographic location coordinate standard facilitates spatial analysis. By adopting independent representation learning models for different types of data, the inherent characteristics of each data can be fully explored. For example, the word embedding model effectively captures the semantic information described in the order text, the time series encoder can learn the temporal dependency and spatial correlation of the vehicle trajectory, and the entity embedding model can capture the potential connection between the warehouse and the goods. The generated independent feature vector set provides high-quality input for subsequent cross-modal alignment, so that data of different modalities can be effectively associated at the semantic level. Using the logistics field knowledge base as a bridge, the semantic gap problem between heterogeneous data is effectively solved. Through the contrast learning strategy, semantically related different modal data are close to each other in the vector space, realizing the effective alignment of cross-modal data. For example, the embedding vector representing the text description of the same product and the knowledge base embedding vector of the product entity have higher similarity after alignment, which provides a more accurate semantic basis for subsequent knowledge fusion and relationship extraction. Through the weighted fusion method based on the attention mechanism, data information from different modalities is effectively integrated. The attention mechanism can learn the contribution of different modal data to a specific logistics object and perform adaptive weighted fusion, avoiding information loss or redundancy caused by simple splicing or averaging. The generated unified feature representation not only contains information from different sources such as orders, trajectories, and inventory, but also highlights key information, providing a more comprehensive and refined feature representation for subsequent semantic enhancement and relationship extraction.

[0019] Preferably, step S2 comprises the following steps:

[0020] Step S21: extracting context information from the unified feature representation using a pre-built logistics domain knowledge base to obtain a context information set;

[0021] Step S22: supplementing the unified feature representation with knowledge features according to the context information set and the pre-built logistics domain knowledge base to obtain knowledge fusion features;

[0022] Step S23: Performing dynamic context-aware embedding on the knowledge fusion features according to the context information set to obtain dynamic context embedding data;

[0023] Step S24: Perform multi-dimensional semantic fusion on the dynamic context embedding data to obtain semantic enhancement features.

[0024] The present invention effectively expands the semantic information of the unified feature representation by utilizing the knowledge base in the logistics field. Through the graph traversal algorithm, other entities and their relationships directly or indirectly related to the target entity are extracted, thereby providing richer background knowledge for subsequent semantic understanding. For example, extracting the commodity information, warehouse information and vehicle information related to the order helps to understand the context of the order more comprehensively, and provides a more reliable basis for subsequent relationship extraction and knowledge fusion. By querying the knowledge base, the key information missing in the unified feature representation, such as commodity categories, administrative divisions and vehicle types, is supplemented. These supplementary knowledge features enrich the semantics of the entity representation, so that the model can understand the meaning of the entity more accurately, and improve the performance of subsequent processing tasks, such as more accurate relationship extraction and more effective entity alignment. Through dynamic context-aware embedding, the entity representation can better adapt to different contexts. Using the attention mechanism of the Transformer encoder, the embedding representation of each entity is dynamically adjusted according to its context information, so that the same entity can have different representations in different scenarios, thereby more accurately expressing its meaning in the current context. For example, when the same vehicle performs different delivery tasks, its embedding representation will vary due to different tasks, which helps to improve the accuracy of downstream tasks. By integrating information such as timestamps and geographic locations, entity representations contain not only static attribute information and context information, but also dynamic spatiotemporal information, thereby more comprehensively describing the state and behavior of logistics entities. For example, order representations that integrate temporal and spatial information can better reflect the life cycle and geographic location information of orders, providing more powerful feature support for subsequent tasks such as route optimization and ETA prediction.

[0025] Preferably, step S3 comprises the following steps:

[0026] Step S31: generating candidate entity pairs based on the semantic enhancement features to obtain a candidate entity pair set; extracting spatiotemporal features based on the candidate entity pair set and the semantic enhancement features to obtain a spatiotemporal feature vector; extracting resource dependency features based on the candidate entity pair set and the semantic enhancement features to obtain a dependency feature vector;

[0027] Step S32: constructing an initial logistics relationship graph according to the candidate entity pair set and the semantic enhancement features to obtain an initial relationship graph; integrating the time, space and dependency features of the initial relationship graph according to the semantic enhancement features and the dependency feature vector to obtain an enhanced feature graph;

[0028] Step S33: performing physical motion model feature fusion according to the enhanced feature map to obtain a motion model fusion map;

[0029] Step S34: performing cargo stacking model feature fusion according to the enhanced feature map to obtain a stacking model fusion map;

[0030] Step S35: performing mechanical analysis feature fusion according to the enhanced feature map to obtain a mechanical model fusion map;

[0031] Step S36: performing graph neural network message transmission and relationship inference according to the motion model fusion graph, the stacking model fusion graph, and the mechanical model fusion graph to obtain a relationship inference score;

[0032] Step S37: performing time series-based relational pattern mining according to the candidate entity pair set and the semantic enhancement features to obtain a time relational pattern;

[0033] Step S38: Perform multi-modal feature fusion and relationship determination based on the relationship inference score and the temporal relationship pattern to obtain a multi-modal relationship set.

[0034] The present invention effectively narrows the scope of relationship extraction through heuristic rules, avoids calculating all entity pairs, and reduces the computational complexity. The extracted spatiotemporal feature vectors and resource dependency feature vectors characterize the potential connection between entity pairs from different dimensions, providing multiple bases for subsequent relationship inference. For example, features such as the distance between the order and the delivery point, the time when the vehicle arrives at the delivery point, and the weight of the order and the load capacity of the vehicle can effectively help the model determine whether there is a delivery relationship between the two. By integrating semantic enhancement features as node features, spatiotemporal features and resource dependency features as edge features into the graph, the model can simultaneously consider the semantic information of the entity itself and the association information between entities, thereby more comprehensively understanding the relationship between entities. By predicting the estimated time (ETA) of the vehicle arriving at the target location and its confidence, it provides strong evidence for judging whether the vehicle will arrive at a certain delivery point or warehouse, overcoming the limitations of traditional methods that only rely on historical data and statistical information, so that the model can make more accurate predictions. By evaluating the stacking stability, accessibility and space utilization of goods on the shelves, it provides a physical basis for judging whether the goods are stored on a specific shelf, so that the model can understand the physical constraints between the goods and the storage location, and improve the rationality of relationship extraction. By predicting the impact force, vibration intensity and tilt angle of the goods during transportation, it provides important safety assessment information for judging whether special goods are suitable for transportation by specific transportation tools, enabling the model to understand the physical risks in the transportation of goods and improving the reliability of relationship extraction. By performing message transmission on the graph that integrates the physical model information, the model can comprehensively consider the semantic information of the entity itself, the spatiotemporal and resource dependencies between entities, and the physical constraints, so as to more accurately predict the probability of different types of relationships between entities, and improve the accuracy and robustness of relationship extraction. By analyzing the temporal connection relationship between orders and trajectory points, the driving time of vehicles between distribution points and other information, it can effectively capture the temporal dependency and behavior patterns between entities, providing an important reference for the final relationship determination. By integrating the relationship scores predicted by the graph neural network and the temporal relationship patterns mined by time series analysis, the structured information and time series information are comprehensively utilized to improve the accuracy and reliability of relationship extraction. By setting the confidence threshold, high-quality relationships are screened out, and the final multi-modal relationship set provides key relationship information for building a high-quality knowledge graph.

[0035] Preferably, step S4 comprises the following steps:

[0036] Step S41: Preliminary screening of candidate entity pairs is performed on the semantic enhancement features to obtain preliminary candidate entity pairs;

[0037] Step S42: Calculate the global feature similarity of the preliminary candidate entity pairs to obtain a global similarity score;

[0038] Step S43: clustering the preliminary candidate entity pairs according to the global similarity scores to obtain a global clustered entity set;

[0039] Step S44: using semantic enhancement features to perform local feature refinement matching on the global clustering entity set to obtain local matching entity pairs;

[0040] Step S45: performing unified alignment objective function optimization according to the multi-modal relationship set, the local matching entity pairs and the global clustering entity set to obtain an optimized alignment result;

[0041] Step S46: Generate an aligned entity set according to the optimized alignment result to obtain an aligned entity set.

[0042] The present invention realizes entity alignment across data sources and eliminates redundancy and ambiguity through multi-level feature matching and fusion. The preliminary screening of candidate entity pairs is like an efficient "preliminary screener". Based on simple global features, such as name similarity, geographic location proximity, license plate number consistency, etc., potential matching entity pairs are quickly screened out, narrowing the scope of subsequent refined matching. The global feature similarity calculation (S42) is like a "preliminary evaluation" of the preliminarily screened entity pairs, calculating the similarity score of their global features, and providing a quantitative basis for subsequent clustering. Using clustering algorithms such as DBSCAN, it is like "grouping similar individuals together". Based on the global similarity score, cross-source entities that potentially point to the same real-world entity are clustered into clusters, which limits the scope for subsequent local refined matching. Using semantic enhancement features, the cosine similarity between entities in the cluster is calculated to further identify entity pairs with high semantic similarity, thereby improving the accuracy of matching. By comprehensively considering multi-modal relationships, local matching results and global clustering information, an entity alignment graph is constructed, and an optimization algorithm is used to find the optimal matching solution to ensure that one entity is matched with only one entity of another data source, avoiding one-to-many mismatches. The alignment entity set generation stage is like giving a "unified identity" to the matching entities, assigning a unified entity ID to each matching entity pair, thereby effectively eliminating entity redundancy and inconsistency across data sources, and laying a solid foundation for building a unified knowledge graph.

[0043] Preferably, step S5 comprises the following steps:

[0044] Step S51: perform entity traversal on the aligned entity set, and add each aligned entity to the pre-built initial graph to obtain an entity-filled graph;

[0045] Step S52: traverse the multi-mode relationship set, and add each relationship to the entity filling graph to obtain a relationship connection graph;

[0046] Step S53: applying hierarchical organization to the relationship connection map to obtain a hierarchical organization map;

[0047] Step S54: supplementing the hierarchical organization map with multi-dimensional relationship attributes according to the multi-mode relationship set to obtain an attributed relationship map;

[0048] Step S55: construct an index for the attributed relationship graph to obtain an index optimization graph;

[0049] Step S56: Generate a logistics distribution knowledge graph for the index optimization graph to obtain a logistics distribution knowledge graph.

[0050] The present invention preliminarily constructs the basic structure of the graph by adding each aligned entity as a node to the graph database and assigning it corresponding attribute information. According to the relationship information in the multi-modal relationship set, edges are created between related entity nodes to connect various entities to form a preliminary relationship network. According to the predefined logistics field ontology, the entities and relationships in the graph are layered to improve the interpretability and query efficiency of the graph. By adding multi-dimensional attribute information to the relationship edges, the relationship expression is made more complete and specific. By building an index based on commonly used query patterns, the query efficiency of the graph is improved, allowing users to quickly retrieve the required information. The constructed index optimization graph is stored to generate the final logistics distribution knowledge graph. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 The figure is a flowchart of the steps of a data fusion method based on the logistics distribution knowledge graph.

[0052] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0053] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0054] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0055] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0056] To achieve this, please refer to Figure 1 , a data fusion method based on logistics distribution knowledge graph, including the following steps:

[0057] Step S1: Obtaining original data of logistics distribution; performing independent representation learning of heterogeneous data on the original data of logistics distribution to obtain an independent feature vector set; performing unified feature representation processing on the independent feature vector set to obtain a unified feature representation;

[0058] Step S2: extracting context information from the unified feature representation and performing dynamic perception embedding to obtain dynamic context embedding data; performing multi-dimensional semantic fusion on the dynamic context embedding data to obtain semantic enhancement features;

[0059] Step S3: According to the semantic enhancement features, the spatiotemporal and dependent features are integrated and processed to obtain an enhanced feature map; according to the enhanced feature map, the physical motion model features are fused to obtain a motion model fusion map; according to the enhanced feature map, the cargo stacking model features are fused to obtain a stacking model fusion map; according to the enhanced feature map, the mechanical analysis features are fused to obtain a mechanical model fusion map; according to the motion model fusion map, the stacking model fusion map and the mechanical model fusion map, the graph neural network message transmission and relationship inference are performed to obtain a relationship inference score; according to the relationship inference score, multi-modal feature fusion and relationship determination are performed to obtain a multi-modal relationship set;

[0060] Step S4: performing global entity clustering analysis based on semantic enhancement features to obtain a global clustering entity set; performing local feature refinement matching on the global clustering entity set to obtain local matching entity pairs; performing hierarchical entity alignment based on the multimodal relationship set, the local matching entity pairs and the global clustering entity set to obtain an aligned entity set;

[0061] Step S5: construct an entity filling graph based on the aligned entity set to obtain an entity filling graph; supplement the entity filling graph with multi-dimensional relationship attributes based on the multi-modal relationship set to obtain an attributed relationship graph; generate a logistics distribution knowledge graph based on the attributed relationship graph to obtain a logistics distribution knowledge graph.

[0062] In the embodiment of the present invention, reference Figure 1 As shown, it is a schematic diagram of the steps of the data fusion method based on the logistics distribution knowledge graph of the present invention. In this example, the data fusion method based on the logistics distribution knowledge graph includes the following steps:

[0063] Step S1: Obtaining original data of logistics distribution; performing independent representation learning of heterogeneous data on the original data of logistics distribution to obtain an independent feature vector set; performing unified feature representation processing on the independent feature vector set to obtain a unified feature representation;

[0064] In the embodiment of the present invention, the original logistics data from the order management, transportation management and warehouse management systems are first collected, and pre-processing operations such as cleaning and format unification are performed to form structured pre-processed logistics data. Subsequently, for different types of data such as orders, vehicle trajectories and inventory, pre-trained word embedding models, Transformer-based time series encoders and entity embedding models are used to perform independent feature representation learning, and the heterogeneous data are converted into their respective vector spaces to obtain independent feature vector sets consisting of order embedding vectors, trajectory embedding vectors and inventory embedding vectors. Finally, using the pre-built logistics domain knowledge base, a contrastive learning strategy is used to perform cross-modal alignment of the independent feature vector set, and data of different modalities are mapped to the same semantic space. The feature vectors of the aligned different types of data are fused through a weighted fusion method based on the attention mechanism to generate a unified feature representation containing multi-source heterogeneous information.

[0065] Step S2: extracting context information from the unified feature representation and performing dynamic perception embedding to obtain dynamic context embedding data; performing multi-dimensional semantic fusion on the dynamic context embedding data to obtain semantic enhancement features;

[0066] In an embodiment of the present invention, a pre-built logistics domain knowledge base is used to extract contextual information around each logistics entity, including related entities and their attributes, through a graph traversal algorithm. Then, a Transformer-based encoder is used to dynamically context-aware embedding of knowledge fusion features. The encoder uses an attention mechanism to dynamically adjust the entity embedding representation according to the context information and generate dynamic context embedding data. Finally, by concatenating or attentive mechanisms to fuse dimensional information such as timestamps and geographic locations, and combining domain knowledge for deep semantic integration, semantically enhanced features containing multi-dimensional semantic information such as time, space, and text are obtained.

[0067] Step S3: According to the semantic enhancement features, the spatiotemporal and dependent features are integrated and processed to obtain an enhanced feature map; according to the enhanced feature map, the physical motion model features are fused to obtain a motion model fusion map; according to the enhanced feature map, the cargo stacking model features are fused to obtain a stacking model fusion map; according to the enhanced feature map, the mechanical analysis features are fused to obtain a mechanical model fusion map; according to the motion model fusion map, the stacking model fusion map and the mechanical model fusion map, the graph neural network message transmission and relationship inference are performed to obtain a relationship inference score; according to the relationship inference score, multi-modal feature fusion and relationship determination are performed to obtain a multi-modal relationship set;

[0068] In the embodiment of the present invention, firstly, candidate entity pairs are generated based on heuristic rules, and the spatiotemporal features and resource dependency features of the entity pairs are extracted. Then, an initial logistics relationship graph is constructed, and the spatiotemporal and dependency features are integrated into the features of the edge to obtain an enhanced feature graph. For the vehicle-location edge, the estimated time of arrival (ETA) information based on Kalman filter prediction is integrated; for the cargo-location edge, the stacking stability, accessibility and space utilization score based on the simulation of the physical engine are integrated; for the special cargo-transportation tool edge, the stacking stability, accessibility and space utilization score based on the simulation of the physical engine are integrated to form a stacking model fusion graph. Then, for the special cargo-transportation tool edge, a mechanical analysis model instance is constructed based on the road conditions of the transportation path, the motion state of the transportation tool and the properties of the special cargo, and the impact force, vibration intensity and tilt angle of the cargo during transportation are predicted, and these mechanical features are integrated into the features of the edge to obtain a mechanical model fusion graph. Finally, the corresponding fusion graph is selected according to the connected entity type, and multiple rounds of message transmission are performed using a graph convolutional network or a graph attention network. The feature information of the nodes and edges is comprehensively considered to predict the probability or confidence score of different types of relationships between the two entities connected by each edge. The relationship inference score is then weighted averaged or fused based on a neural network with the temporal relationship pattern mined based on the dynamic time warping algorithm, and the final multi-modal relationship set is output according to the set confidence threshold.

[0069] Step S4: performing global entity clustering analysis based on semantic enhancement features to obtain a global clustering entity set; performing local feature refinement matching on the global clustering entity set to obtain local matching entity pairs; performing hierarchical entity alignment based on the multimodal relationship set, the local matching entity pairs and the global clustering entity set to obtain an aligned entity set;

[0070] In the embodiment of the present invention, firstly, based on the name similarity, geographical location distance, license plate number consistency and consignee and consignee information similarity of entities such as warehouses, vehicles and orders, the similarity scores of entity pairs are calculated by using algorithms such as Jaccard similarity, Haversine distance and Jaro-Winkler distance, and the entities are clustered by using the DBSCAN algorithm, and entities similar in global features are divided into the same cluster to obtain a global cluster entity set. Then, for each entity cluster, the cosine similarity of the entity semantic enhancement feature vector is used for more refined matching, and cross-source entity pairs with high similarity at the semantic level are added to the local matching entity pairs. Finally, an entity alignment graph is constructed, and the matching edge weight is increased by using the local matching entity pairs, and the relationship information between entities in the multi-modal relationship set is combined, and the unified alignment objective function is optimized by using the Hungarian algorithm or the iterative local search algorithm to find the optimal entity matching solution, generate an optimized alignment result containing the final cross-source entity matching pairs and their matching confidence, and assign a unified entity ID to the matching entities to form an aligned entity set.

[0071] Step S5: construct an entity filling graph according to the aligned entity set to obtain an entity filling graph; supplement the entity filling graph with multi-dimensional relationship attributes according to the multi-modal relationship set to obtain an attributed relationship graph; generate a logistics distribution knowledge graph according to the attributed relationship graph to obtain a logistics distribution knowledge graph;

[0072] In the embodiment of the present invention, the aligned entity set is first traversed, a new node is created in the graph according to the unified entity ID, and the node attributes are set according to the original data to construct an entity-filled graph containing all aligned entities. Then, the multi-mode relationship set is traversed, and the corresponding node is found in the entity-filled graph according to the entity ID of the relationship connection, and a directed edge is created and the relationship type is set to form a relationship connection graph. Then, according to the predefined logistics domain ontology, the entities and relationships in the graph are classified into corresponding levels to realize the hierarchical organization of the graph. Subsequently, according to the attribute information contained in the multi-mode relationship set, the expected arrival time, actual arrival time, delivery distance, storage temperature and other attributes are added to the corresponding relationship edge in the graph, so that the relationship information is more complete. Finally, for the commonly used query conditions of entities and relationships, the B-tree index or full-text index and other technologies are used to build an index to optimize the query efficiency of the graph, and the final index-optimized graph is stored as a logistics distribution knowledge graph.

[0073] Preferably, step S1 comprises the following steps:

[0074] Step S11: obtaining original logistics distribution data; performing logistics data preprocessing on the original logistics distribution data to obtain preprocessed logistics data;

[0075] Step S12: Performing independent representation learning of heterogeneous data on the pre-processed logistics data to obtain an independent feature vector set;

[0076] Step S13: using the pre-built logistics domain knowledge base to perform cross-modal alignment on the independent feature vector sets to obtain an aligned feature space set;

[0077] Step S14: performing unified feature representation fusion on the aligned feature space set to obtain a unified feature representation.

[0078] In an embodiment of the present invention, the original data of logistics distribution is collected from multiple information systems, including at least order data generated by the order management system, vehicle trajectory data provided by the transportation management system, and inventory data recorded by the warehouse management system. The order data includes fields such as order number, product identification, longitude and latitude of the delivery address, longitude and latitude of the delivery address, and order timestamp. The vehicle trajectory data includes a unique vehicle identification, a sequence of geographic location coordinates, a corresponding timestamp sequence, and a vehicle status code. The inventory data includes a unique warehouse identification, a product identification, a storage location code, an inventory quantity, and an inventory timestamp. For this type of multi-source heterogeneous data, the preprocessing operation is started in the data access module, which is responsible for extracting raw data from each system interface and storing it in a distributed file system. Subsequently, the data cleaning module is started, and the original data in the distributed file system is read to perform the following operations: For order data, the order number format is verified by using the regular expression matching method, and the data records with missing longitude and latitude of the shipping and receiving addresses are removed; for vehicle trajectory data, the monotonicity verification method of the timestamp sequence is used to identify and remove disordered or repeated trajectory points, and the outlier detection algorithm based on the speed threshold is used to remove drifting trajectory points; for inventory data, the combined uniqueness constraint of the product identifier and the warehouse identifier is used to verify duplicate records, and the latest inventory status is retained based on the inventory timestamp. After the data is cleaned, the format is converted according to the predefined unified data model. For example, the timestamp is uniformly converted to the ISO 8601 format, and the geographic location coordinates are uniformly converted to the WGS84 standard. Finally, the converted structured data is stored in the data warehouse in Parquet format for use in subsequent steps.

[0079] The pre-processed logistics data stored in the data warehouse is loaded into the distributed computing framework. Independent representation learning models are used for different types of data. For the text description fields (such as product names) in the order data, a pre-trained word embedding model (such as Word2Vec or FastText) is used to map each word into a vector of fixed dimension, and then the order text description is converted into a single vector representation through average pooling operation. For vehicle trajectory data, the geographic coordinate sequence and timestamp sequence of each vehicle are input into the Transformer-based time series encoder, which uses the self-attention mechanism to capture the temporal dependency and spatial correlation between trajectory points, and finally generates a fixed-length trajectory embedding vector for each vehicle. For inventory data, the warehouse ID, product ID and inventory quantity are combined, and the embedding vectors of the warehouse and product are learned using an entity embedding model (such as Node2Vec) on the bipartite graph composed of the warehouse and the product. The above three independent embedding models are all trained offline and the model parameters are stored. In the representation learning stage, the pre-trained model parameters are directly loaded, and the pre-processed order, trajectory and inventory data are respectively input into the corresponding models to generate order embedding vectors, trajectory embedding vectors and inventory embedding vectors. These vectors constitute an independent set of feature vectors, each representing the semantic information of different types of data, but have not yet been aligned across modalities.

[0080] The pre-built logistics domain knowledge base is stored in the form of a graph database, which contains the core concepts in the logistics field (e.g., goods, warehouses, vehicles, customers) and the relationships between them (e.g., the storage relationship between goods and warehouses, and the service relationship between vehicles and distribution areas). Each entity and relationship in the knowledge base is pre-embedded in the vector space. The order embedding vector, trajectory embedding vector, and inventory embedding vector in the independent feature vector set are loaded into the cross-modal alignment module. This module uses a contrastive learning strategy to close the distance between semantically related different modal data in the vector space. Specifically, for an order, its order embedding vector is similar to the corresponding product entity embedding vector, shipping warehouse entity embedding vector, and geographic location embedding vector corresponding to the delivery address in the knowledge base. Triplet Loss is used as the loss function to minimize the distance between the order embedding vector and the embedding vector of its positive sample (e.g., the goods included in the order), while maximizing the distance between it and the embedding vector of the negative sample (e.g., irrelevant goods). For the trajectory embedding vector, the contrastive learning strategy is also used to make it close to the geographic location embedding vector of its service area and away from the embedding vectors of other areas. Through multiple rounds of iterative training, the representation of order, trajectory, and inventory embedding vectors is adjusted to align them with the corresponding entity embedding vectors in the knowledge base in the vector space. After the training is completed, an aligned feature space set is obtained, in which data of different modalities are mapped to the same semantic space and are comparable.

[0081] The aligned feature space set contains the aligned order embedding vectors, trajectory embedding vectors, and inventory embedding vectors. The unified feature representation fusion module receives these vectors as input and fuses the corresponding order embedding vector, the related trajectory embedding vector, and the involved inventory embedding vector for each logistics object (e.g., a specific order). The fusion operation adopts a weighted fusion method based on the attention mechanism. Specifically, for an order, its order embedding vector is used as the query vector (Query), and the related trajectory embedding vector and inventory embedding vector are used as key-value pairs (Key-Value). The correlation score between the order embedding vector and each trajectory embedding vector and inventory embedding vector is calculated through the attention mechanism, and the trajectory embedding vector and inventory embedding vector are weighted summed according to the score. The result of the weighted sum is concatenated with the order embedding vector to form a unified feature representation of the order. For orders without related trajectory or inventory information, its unified feature representation only contains the order embedding vector. Finally, the unified feature representation of all logistics objects constitutes a unified feature representation, which fuses multi-source heterogeneous data into the same vector space and retains the key information of different modal data.

[0082] Preferably, step S2 comprises the following steps:

[0083] Step S21: extracting context information from the unified feature representation using a pre-built logistics domain knowledge base to obtain a context information set;

[0084] Step S22: supplementing the unified feature representation with knowledge features according to the context information set and the pre-built logistics domain knowledge base to obtain knowledge fusion features;

[0085] Step S23: Performing dynamic context-aware embedding on the knowledge fusion features according to the context information set to obtain dynamic context embedding data;

[0086] Step S24: Perform multi-dimensional semantic fusion on the dynamic context embedding data to obtain semantic enhancement features.

[0087] In an embodiment of the present invention, a unified feature representation is loaded into a context information extraction module. The pre-built logistics domain knowledge base is stored in the form of a graph database, where nodes represent entities (e.g., orders, goods, warehouses, vehicles, geographic locations), and edges represent relationships between entities (e.g., orders contain goods, vehicles belong to warehouses, and orders are sent to geographic locations). For each logistics entity corresponding to the unified feature representation, the context information extraction operation is implemented through a graph traversal algorithm. For an order entity, the first-order and second-order neighbor nodes are traversed to extract the directly connected commodity entities, customer entities, shipping warehouse entities, shipping address geographic location entities, and other entities connected to these neighbor nodes. For example, the attributes of the commodity entity directly connected to the order (e.g., commodity category, weight, volume) are extracted; the identifier of the vehicle entity (the vehicle that performs the order delivery task) connected to the shipping warehouse entity of the order is extracted. For the vehicle entity, its current geographic location entity, the order entity corresponding to the delivery task being executed, and the warehouse entity to which it belongs are extracted. The extracted neighbor entities and their relationship types constitute the context information of the logistics entity. The context information of each logistics entity is encoded as an adjacency list or adjacency matrix and stored in association with the unified feature representation of the entity to form a context information set. Each element in the set contains the unified feature representation of a logistics entity and its corresponding context information.

[0088] The context information set and the pre-built logistics domain knowledge base are loaded into the knowledge feature supplementation module. For each logistics entity in the context information set and its corresponding unified feature representation and context information, the knowledge supplementation operation is performed based on the entity and relationship information in the knowledge base. For the order entity, if the unified feature representation lacks product category information, the knowledge base is queried to obtain the category attribute of the product based on the product entity extracted from the context information, and the category information is encoded as a one-hot vector or embedded vector and spliced ​​into the unified feature representation of the order. If the delivery address of the order only contains longitude and latitude coordinates, the administrative region information such as the province and city to which the address belongs is supplemented based on the administrative division information of the geographic location entity in the knowledge base, and this information is encoded and added to the feature representation of the order. For the vehicle entity, if the unified feature representation lacks vehicle type information, the relationship between the vehicle and the warehouse in the knowledge base is queried based on the warehouse entity to which it belongs, and the vehicle type is inferred (for example, a vehicle belonging to a cold chain warehouse is likely to be a cold chain vehicle), and the vehicle type is encoded and added to its feature representation. The supplemented knowledge features are spliced ​​with the original unified feature representation to form knowledge fusion features. This feature combines the original data information with the supplementary information from the knowledge base, enriching the semantics of entity representation.

[0089] The knowledge fusion features and context information set are loaded into the dynamic context-aware embedding module. The core of this module is a Transformer-based encoder that dynamically adjusts the embedding representation of each logistics entity's knowledge fusion feature using its context information. Specifically, for the knowledge fusion feature of an order, the feature representation of its corresponding context information (e.g., the included goods, shipping warehouse, and delivery address) is input into the attention layer of the Transformer encoder. The attention mechanism calculates the correlation score between the order feature and each context entity feature, and weighted sums the context entity features according to the score to obtain a context vector. The context vector is concatenated or weightedly fused with the knowledge fusion feature of the order to generate the dynamic context embedding of the order. For example, if the context information indicates that an order contains fragile items, the dynamic context-aware embedding adjusts the embedding representation of the order to make it more emphasized on the semantics of "fragile". For vehicle entities, the generation process of its dynamic context embedding is similar, taking into account context information such as its current location and the task being performed. In this way, the embedding representation of each logistics entity incorporates its specific context information, thereby more accurately expressing its meaning in the current scenario. The dynamic context embeddings of all logistics entities constitute the dynamic context embedding data.

[0090] The dynamic context embedding data is loaded into the multi-dimensional semantic fusion module. This module aims to further fuse semantic information in dimensions such as time and space. For the dynamic context embedding of each logistics entity, the fusion operation is as follows: For the order entity, extract its order timestamp and convert it into a time embedding vector (for example, through sinusoidal positional encoding). Extract the longitude and latitude of the delivery address of the order and convert it into a geographic location embedding vector (for example, through GeoHash encoding or learning). Concatenate the time embedding vector and the geographic location embedding vector with the dynamic context embedding vector of the order or fuse them using the attention mechanism to obtain the fused feature representation. For the vehicle entity, extract its current timestamp and geographic location coordinates, and perform similar embedding conversion and fusion operations. For the warehouse entity, extract its geographic location coordinates and establishment time and other information for encoding and fusion. In the fusion process, other dimensional information can also be considered, such as the priority of the order, the load capacity of the vehicle, etc. Ultimately, the multi-dimensional fusion features of each logistics entity constitute a semantic enhancement feature, which not only contains the basic attributes and contextual information of the entity, but also incorporates semantic information in dimensions such as time and space, thereby expressing the meaning of the logistics entity more comprehensively and accurately.

[0091] Preferably, step S3 comprises the following steps:

[0092] Step S31: generating candidate entity pairs based on the semantic enhancement features to obtain a candidate entity pair set; extracting spatiotemporal features based on the candidate entity pair set and the semantic enhancement features to obtain a spatiotemporal feature vector; extracting resource dependency features based on the candidate entity pair set and the semantic enhancement features to obtain a dependency feature vector;

[0093] Step S32: constructing an initial logistics relationship graph according to the candidate entity pair set and the semantic enhancement features to obtain an initial relationship graph; integrating the time, space and dependency features of the initial relationship graph according to the semantic enhancement features and the dependency feature vector to obtain an enhanced feature graph;

[0094] Step S33: performing physical motion model feature fusion according to the enhanced feature map to obtain a motion model fusion map;

[0095] Step S34: performing cargo stacking model feature fusion according to the enhanced feature map to obtain a stacking model fusion map;

[0096] Step S35: performing mechanical analysis feature fusion according to the enhanced feature map to obtain a mechanical model fusion map;

[0097] Step S36: performing graph neural network message transmission and relationship inference according to the motion model fusion graph, the stacking model fusion graph, and the mechanical model fusion graph to obtain a relationship inference score;

[0098] Step S37: performing time series-based relational pattern mining according to the candidate entity pair set and the semantic enhancement features to obtain a time relational pattern;

[0099] Step S38: Perform multi-modal feature fusion and relationship determination based on the relationship inference score and the temporal relationship pattern to obtain a multi-modal relationship set.

[0100] The semantic enhancement features are loaded into the candidate entity pair generation module. For all logistics entities, heuristic rules are used to generate candidate entity pairs. For order entities, candidate pairs are generated with distribution point entities within the specified radius of the same delivery address and vehicle entities within the specified time window. For vehicle entities, candidate pairs are generated with warehouse entities within the specified radius and order entities currently executing tasks. For warehouse entities, candidate pairs are generated with stored commodity entities. The generated entity pairs are stored in the candidate entity pair set. The spatiotemporal feature extraction module receives the candidate entity pair set and semantic enhancement features as input. For each candidate entity pair, its spatiotemporal features are extracted. For order-delivery point pairs, the distance between the order delivery address and the geographic location of the delivery point, and the difference between the estimated delivery time of the order and the current time are calculated as features. For vehicle-delivery point pairs, the timestamp of the vehicle arriving at the delivery point, the length of time the vehicle stays at the delivery point, and the current speed of the vehicle are extracted as features. For order-vehicle pairs, the difference between the order shipment time and the time the vehicle leaves the shipment warehouse, and the difference between the estimated delivery time of the order and the estimated time the vehicle arrives at the delivery address are extracted as features. The extracted spatiotemporal features are combined into a spatiotemporal feature vector of fixed dimension. The resource dependency feature extraction module receives the same input. For each candidate entity pair, its resource dependency is analyzed. For the order-vehicle pair, the weight, volume, required model of the order and the load, capacity, and model of the vehicle are extracted as features. For the warehouse-goods pair, the storage temperature requirements of the goods, the volume and the temperature zone of the warehouse, and the remaining capacity are extracted as features. The extracted resource dependency features are combined into a dependency feature vector of fixed dimension.

[0101] The candidate entity pair set and semantic enhancement features are loaded into the initial logistics relationship graph construction module. Each logistics entity in the graph is represented as a node, and the initial feature vector of the node is the corresponding semantic enhancement feature. If two entities exist in the candidate entity pair set, an undirected edge is created between them. After the initial relationship graph is constructed, the spatiotemporal and dependency feature integration module receives the initial relationship graph, the spatiotemporal feature vector, and the dependency feature vector as input. For each edge in the initial relationship graph, the two entities connected by the edge correspond to a candidate entity pair. Find the spatiotemporal feature vector and dependency feature vector corresponding to the candidate entity pair, and concatenate these two vectors as the feature vector of the edge. The nodes of the enhanced feature graph contain semantic enhancement features, and the edges contain spatiotemporal and resource dependency features.

[0102] The enhanced feature graph is loaded into the physical motion model feature fusion module. This module focuses on processing the edges connecting vehicle entities and location entities (delivery points, warehouses). First, all edges connecting vehicle nodes and location nodes are filtered out. For each such edge, the historical trajectory data of the connected vehicle is extracted, including the timestamp sequence and the corresponding geographic location coordinate sequence. The current state information of the vehicle is obtained, such as the current speed and driving direction. Road network data is obtained from the geographic information system, including the geometry of the road, speed limit information, traffic flow prediction, etc. The estimated time (ETA) of the vehicle to the target location and its confidence are predicted by combining the vehicle's historical trajectory, current state and road network information using state estimation algorithms such as Kalman filtering or particle filtering. The ETA prediction results, including the ETA value and confidence score, are encoded as a fixed-dimensional vector and integrated into the enhanced feature graph as additional features of the edge. In addition to the original spatiotemporal and dependency features, the vehicle-location edge in the motion model fusion graph also contains the ETA features predicted based on the physical motion model.

[0103] The enhanced feature graph is loaded into the cargo stacking model feature fusion module. This module processes the edges connecting cargo entities and warehouse or shelf entities. First, all edges connecting cargo nodes and warehouse or shelf nodes are filtered out. For each such edge, the physical properties of the connected cargo are extracted, such as length, width, height, weight, and fragility. The physical structure information of the warehouse or shelf is obtained, such as the size, load-bearing capacity, and accessibility of the shelf. A cargo stacking model is constructed, which evaluates the stacking stability, accessibility, and space utilization of cargo on a specific shelf based on a physical engine or heuristic rules. The evaluation results, such as stability score, accessibility score, and space utilization score, are encoded as vectors of fixed dimensions and incorporated into the enhanced feature graph as additional features of the edge. In addition to the original spatiotemporal and dependency features, the cargo-warehouse / shelf edges in the stacking model fusion graph also contain features based on the cargo stacking model evaluation.

[0104] The enhanced feature graph is loaded into the mechanical analysis model feature fusion module. This module processes the edges connecting special cargo (fragile goods, dangerous goods) and transport entities. First, all edges connecting special cargo nodes and transport nodes are filtered out. For each such edge, the attribute information of the special cargo is extracted, such as fragility level, dangerous goods category, and packaging method. The motion state data of the transport vehicle, such as acceleration and vibration frequency, is obtained. The road condition information of the transport path, such as bumpiness and slope, is obtained. A mechanical analysis model is constructed, which is based on finite element analysis or multibody dynamics methods to predict the mechanical effects such as impact force, vibration intensity, and tilt angle on the cargo during transportation. The prediction results, such as maximum impact force, vibration frequency peak, and maximum tilt angle, are encoded as vectors of fixed dimensions and integrated into the enhanced feature graph as additional features of the edge. In addition to the original spatiotemporal and dependency features, the special cargo-transportation tool edge in the mechanical model fusion graph also contains features predicted based on the mechanical analysis model.

[0105] The motion model fusion graph, stacking model fusion graph, and mechanical model fusion graph are loaded into the graph neural network message passing and relationship inference module. According to the type of connected entities, the corresponding fusion graph is selected for processing. For example, for the edge connecting the vehicle and the location, the motion model fusion graph is used; for the edge connecting the goods and the warehouse / shelf, the stacking model fusion graph is used; for the edge connecting the special goods and the means of transport, the mechanical model fusion graph is used; for other types of edges, the enhanced feature graph is used. A graph convolutional network (GCN) or a graph attention network (GAT) is used as the graph neural network model. The model enables each node to aggregate feature information from its neighboring nodes and edges through a multi-layer message passing mechanism. During the message passing process, the features of the edge (including spatiotemporal, dependency, and physical model fusion features) participate in the update of the node features. Finally, the graph neural network model predicts the probability or confidence score of different types of relationships between the two entities connected by each edge. The relationship type is predefined, such as "distribution", "storage", and "transportation". The relationship inference score indicates the confidence level that a specific relationship exists between the entity pair.

[0106] The candidate entity pair set and semantic enhancement features are loaded into the time series-based relational pattern mining module. Time series analysis is performed on entity pairs containing time attributes in the candidate entity pair set, such as order-trajectory point pairs and vehicle-delivery point pairs. For order-trajectory point pairs, the estimated delivery time of the order and the timestamp sequence of the trajectory points are extracted, and the dynamic time warping (DTW) algorithm is used to calculate the time similarity between the two, and the binary feature of whether the trajectory point is within the estimated delivery time window of the order is extracted. For vehicle-delivery point pairs, the timestamp sequence of the vehicle arriving at the delivery point and the timestamp of the vehicle leaving the previous delivery point are extracted, and statistical features such as the average driving time and driving time variance of the vehicle between the delivery points are calculated. The mined temporal relational patterns are encoded as vectors or discrete features of fixed dimensions.

[0107] The relationship inference scores and temporal relationship patterns are loaded into the multimodal feature fusion and relationship determination module. For each candidate entity pair, the corresponding relationship inference scores and temporal relationship patterns are fused. The fusion operation uses a weighted average or a neural network-based fusion model. The fusion model learns the importance of different types of features and outputs the final relationship type prediction and its confidence. For example, if the graph neural network predicts that there is a high probability of a "delivery" relationship between order A and vehicle B, and the time series analysis shows that the trajectory of vehicle B passes near the delivery address of order A within the delivery time window of order A, the fusion model will further increase the confidence of the "delivery" relationship. Set a confidence threshold, and store the relationship predictions with confidence higher than the threshold as the final multimodal relationship in the multimodal relationship set. Each relationship contains two connected entities, a relationship type, and a confidence score.

[0108] Preferably, step S33 includes the following steps:

[0109] Step S331: filter the vehicle-position edges on the enhanced feature graph to obtain a vehicle-position edge set;

[0110] Step S332: extracting vehicle historical trajectories from the vehicle-position edge set to obtain vehicle historical trajectory data;

[0111] Step S333: acquiring the current vehicle state from the vehicle-position edge set to obtain the current vehicle state information;

[0112] Step S334: Acquire road network information for the vehicle-position edge set to obtain road network information data;

[0113] Step S335: predicting the estimated time for the vehicle to reach the target location based on the vehicle-location edge set, the vehicle historical trajectory data, the current vehicle status information, and the road network information data, and performing a confidence evaluation to obtain an ETA prediction result;

[0114] Step S336: performing ETA feature fusion according to the vehicle-position edge set and the ETA prediction result to obtain ETA fused edge features;

[0115] Step S337: Generate a motion model fusion map based on the enhanced feature map and the ETA fusion edge features to obtain a motion model fusion map.

[0116] In an embodiment of the present invention, the enhanced feature graph is loaded into a vehicle-location edge screening module. The module traverses all edges in the enhanced feature graph and checks the types of the two nodes connected by each edge. If the two nodes connected by an edge are a vehicle entity and a location entity (distribution point or warehouse), respectively, the edge is determined to be a vehicle-location edge and is added to the vehicle-location edge set. The type information of the location entity is stored in its node attributes, and the type information of the vehicle entity is also stored in its node attributes. The screening operation is based on the judgment of the type attributes of the two nodes connected by the edge. The vehicle-location edge set contains all edges connecting vehicles and locations in the enhanced feature graph, and each edge records the connected vehicle node identifier and location node identifier.

[0117] The vehicle-location edge set is loaded into the vehicle historical trajectory extraction module. For each edge in the vehicle-location edge set, the historical trajectory data of the vehicle node connected to it is extracted. The identifier of the vehicle node is used to query the time series database that stores the vehicle trajectory data. The query operation is based on the vehicle identifier and the current timestamp to retrieve the geographic location coordinate sequence and the corresponding timestamp sequence of the vehicle in the past period of time. The extracted trajectory data is arranged in timestamp order and stored as the attributes of the vehicle-location edge. The vehicle historical trajectory data contains at least the trajectory point information within the past hour, and the accuracy of the trajectory point is not less than 10 meters.

[0118] The vehicle-location edge set is loaded into the current vehicle status acquisition module. For each edge in the vehicle-location edge set, the current status information of the vehicle node connected to it is obtained. The current vehicle status information is obtained by querying the real-time vehicle monitoring system. The query operation is based on the vehicle identifier to obtain the current geographic location coordinates, current speed, driving direction, and vehicle status code (e.g., driving, stopped, loading). The obtained current vehicle status information is stored as an attribute of the vehicle-location edge.

[0119] The vehicle-location edge set is loaded into the road network information acquisition module. For each edge in the vehicle-location edge set, the road network information between the current location of the connected vehicle and the target location (distribution point or warehouse) is obtained. The road network information is obtained by calling the API of the map service. The API call needs to provide the longitude and latitude coordinates of the starting location and the longitude and latitude coordinates of the target location. The map service returns road network data containing at least the following information: the longitude and latitude coordinate sequence of the planned driving path, the total length of the path, the estimated driving time, the road types included in the path, the speed limit information of each road section, and the real-time traffic flow information. The obtained road network information data is stored as the attribute of the vehicle-location edge.

[0120] The vehicle-location edge set, the vehicle's historical trajectory data, the current vehicle state information, and the road network information data are loaded into the estimated time prediction module. For each edge in the vehicle-location edge set, the Kalman filter algorithm is used to predict the estimated time of the vehicle's arrival at the target location. The state vector of the Kalman filter algorithm includes the vehicle's position and speed. The prediction process integrates the vehicle's historical trajectory data, current speed, and speed limit information and traffic flow information in the road network information. The traffic flow information in the road network information is converted into a speed attenuation factor to adjust the speed prediction value. The prediction process outputs the estimated time point when the vehicle arrives at the target location. At the same time, the confidence of the ETA prediction is evaluated based on the covariance matrix of the state estimation during the prediction process. The confidence score ranges from 0 to 1, and the higher the value, the higher the confidence of the prediction result. The ETA prediction result includes the estimated arrival time point and the confidence score.

[0121] The vehicle-location edge set and ETA prediction results are loaded into the ETA feature fusion module. For each edge in the vehicle-location edge set, the corresponding ETA prediction result (estimated arrival time point and confidence score) is encoded as a vector of fixed dimension. The encoding method can be to convert the estimated arrival time point into the number of seconds from the current time and use the confidence score directly as another dimension. The encoded ETA feature vector is concatenated with the original feature vector of the vehicle-location edge (including spatiotemporal and dependency features) to form the ETA fused edge feature.

[0122] The enhanced feature graph and ETA fused edge features are loaded into the motion model fusion graph generation module. This module copies the structure and node features of the enhanced feature graph. For each vehicle-position edge in the enhanced feature graph, its corresponding ETA fused edge feature is used to replace the original edge feature. For non-vehicle-position edges, their edge features remain unchanged. The final generated motion model fusion graph has the same structure as the enhanced feature graph, but the features of the edges connecting vehicles and positions are fused with ETA information predicted based on the physical motion model.

[0123] Preferably, step S34 includes the following steps:

[0124] Step S341: filtering the goods-location edges on the enhanced feature graph to obtain a goods-location edge set;

[0125] Step S342: Acquire the physical attributes of the goods on the goods-location edge set to obtain the goods attribute data;

[0126] Step S343: Acquire the shelf physical structure of the goods-location edge set to obtain shelf structure data;

[0127] Step S344: constructing a cargo stacking model according to the cargo attribute data and the shelf structure data to obtain a cargo stacking model instance;

[0128] Step S345: using the cargo stacking model instance to evaluate the stacking characteristics of the cargo-location edge set to obtain a stacking evaluation result;

[0129] Step S346: performing stacking feature fusion according to the stacking evaluation result and the goods-location edge set to obtain a stacking fusion edge feature;

[0130] Step S347: Generate a stacked model fusion graph based on the enhanced feature graph and the stacked fusion edge features to obtain a stacked model fusion graph.

[0131] In an embodiment of the present invention, the enhanced feature graph is loaded into the cargo-location edge screening module. The module traverses all the edges in the enhanced feature graph and checks the types of the two nodes connected by each edge. If the two nodes connected by an edge are a cargo entity and a location entity (warehouse or shelf), respectively, the edge is determined to be a cargo-location edge and is added to the cargo-location edge set. The type information of the cargo entity and the type information of the location entity (warehouse or shelf) are stored in its node attributes. The screening operation is based on an exact match judgment of the type attributes of the two nodes connected by the edge. The cargo-location edge set contains all the edges connecting cargo and warehouses or shelves in the enhanced feature graph, and each edge records the connected cargo node identifier and location node identifier.

[0132] The cargo-location edge set is loaded into the cargo physical property acquisition module. For each edge in the cargo-location edge set, the physical properties of the cargo node connected to it are extracted. The identifier of the cargo node is used to query the database storing cargo attribute data. The query operation is based on the cargo identifier to retrieve the length, width, height, weight, shape coefficient and fragility level of the cargo. The shape coefficient is used to describe the regularity of the shape of the cargo, and the value range is 0 to 1. The closer to 1, the more regular the shape. The fragility level is divided into level 1 (not fragile), level 2 (slightly fragile), and level 3 (highly fragile). The acquired cargo attribute data is stored as the attributes of the cargo-location edge.

[0133] The goods-location edge set is loaded into the shelf physical structure acquisition module. For each edge in the goods-location edge set, if the location node it connects to is a shelf, the physical structure information of the shelf is obtained. The identifier of the shelf node is used to query the database storing the shelf structure data. The query operation is based on the shelf identifier to retrieve the number of layers of the shelf, the upper limit of each layer, the internal dimensions of each layer (length, width, height), and whether there are divider columns. The obtained shelf structure data is stored as the attribute of the goods-location edge. If the connected location node is a warehouse, skip this step.

[0134] The cargo attribute data and shelf structure data are loaded into the cargo stacking model construction module. For each cargo-location edge, if the location node it connects to is a shelf, a cargo stacking model instance of the cargo on the shelf is constructed. The stacking model simulates the placement process of cargo on the shelf based on the physical engine. The simulation process takes into account the size, weight, shape factor of the cargo, as well as the upper load limit and internal dimensions of the shelf. The model outputs the stability score, accessibility score and space utilization score of the cargo on the shelf. The stability score is calculated based on the relationship between the center of gravity of the cargo and the supporting surface, ranging from 0 to 1, and the higher the more stable. The accessibility score is calculated based on the free space around the simulated cargo, ranging from 0 to 1, and the higher the easier it is to access. The space utilization score is calculated based on the proportion of shelf space occupied by the cargo, ranging from 0 to 1, and the higher the space utilization.

[0135] The cargo stacking model instance and the cargo-location edge set are loaded into the stacking characteristic evaluation module. For each edge in the cargo-location edge set, if the location node it is connected to is a shelf, the cargo stacking model instance constructed in step S344 is used to evaluate the stacking characteristics of the cargo on the shelf. The evaluation results include stability score, accessibility score and space utilization score. If the connected location node is a warehouse, a simplified evaluation method is used, only considering whether the volume of the cargo is less than the remaining capacity of the warehouse, and outputting a binary result (yes / no). The stacking evaluation result is stored as an attribute of the cargo-location edge.

[0136] The goods-location edge set and the stacking evaluation results are loaded into the stacking feature fusion module. For each edge in the goods-location edge set, the corresponding stacking evaluation result is encoded as a vector of fixed dimension. The encoding method can be to directly use the stability score, accessibility score and space utilization score as the three dimensions of the vector. If the connected location node is a warehouse, the binary evaluation result is encoded as 0 or 1. The encoded stacking feature vector is spliced ​​with the original feature vector of the goods-location edge (including spatiotemporal and dependency features) to form a stacking fusion edge feature.

[0137] The enhanced feature graph and stacked fusion edge features are loaded into the stacked model fusion graph generation module. This module copies the structure and node features of the enhanced feature graph. For each item-location edge in the enhanced feature graph, the original edge feature is replaced with its corresponding stacked fusion edge feature. For non-item-location edges, their edge features remain unchanged. The resulting stacked model fusion graph has the same structure as the enhanced feature graph, but the features of the edges connecting items and warehouses / shelves are fused with information based on the evaluation of the item stacking model.

[0138] Preferably, step S35 includes the following steps:

[0139] Step S351: filtering the special cargo-transportation tool edges on the enhanced feature graph to obtain a special cargo-transportation edge set;

[0140] Step S352: acquiring special cargo attributes from the special cargo-transport edge set to obtain special cargo attribute data;

[0141] Step S353: acquiring the transport tool movement state for the special cargo-transportation edge set to obtain the transport tool movement data;

[0142] Step S354: Acquire the transportation path and road condition of the special cargo-transport edge set to obtain the transportation path road condition data;

[0143] Step S355: constructing a mechanical analysis model according to the transportation route road condition data, the transportation tool movement data and the special cargo attribute data to obtain a mechanical analysis model instance;

[0144] Step S356: using the mechanical analysis model instance to predict the stress condition of the special cargo-transportation edge set, and obtaining the stress prediction result;

[0145] Step S357: performing mechanical feature fusion according to the force prediction result and the special cargo-transport edge set to obtain mechanical fusion edge features;

[0146] Step S358: Generate a mechanical model fusion graph based on the enhanced feature graph and the mechanical fusion edge features to obtain a mechanical model fusion graph.

[0147] In an embodiment of the present invention, an enhanced feature graph is loaded into a special cargo-transportation tool edge screening module. The module traverses all edges in the enhanced feature graph and checks the type attributes of the two nodes connected by each edge. If the two nodes connected by an edge are a special cargo entity and a transportation tool entity, respectively, the edge is determined to be a special cargo-transportation tool edge and is added to the special cargo-transportation edge set. The type attribute of the special cargo entity includes subtype identifiers such as fragile goods and dangerous goods. The type attribute of the transportation tool entity includes identifiers such as trucks, trains, ships, and airplanes. The screening operation is based on an exact match judgment of the type attributes of the two nodes connected by the edge. For example, if the node types connected by an edge are "fragile goods" and "trucks", it meets the screening conditions. The special cargo-transportation edge set contains all edges connecting special cargoes and transportation tools in the enhanced feature graph, and each edge records the connected special cargo node identifier and transportation tool node identifier.

[0148] The special cargo-transport edge set is loaded into the special cargo attribute acquisition module. For each edge in the special cargo-transport edge set, the attribute information of the special cargo node connected to it is extracted. The identifier of the special cargo node is used to query the database storing the special cargo attribute data. The query operation is based on the special cargo identifier to retrieve the fragility level, dangerous goods category, packaging material strength, center of gravity position and fixing method description of the cargo. The fragility level is divided into three levels: high, medium and low. The dangerous goods category is classified according to the United Nations dangerous goods numbering standard. The strength of the packaging material is expressed as the compressive strength value (unit: Pascal). The center of gravity position is expressed as the offset relative to the geometric center of the cargo (unit: meter). The fixing method is described as text information, such as "using shock pads" and "using binding fixation". The acquired special cargo attribute data is stored as the attributes of the special cargo-transport tool edge.

[0149] The special cargo-transport edge set is loaded into the transport motion state acquisition module. For each edge in the special cargo-transport edge set, the motion state data of the transport node connected to it is extracted. The identifier of the transport node is used to query the real-time transport monitoring system. The query operation retrieves the instantaneous acceleration (acceleration values ​​of three axes, unit: m / s²), angular velocity (angular velocity values ​​of three axes, unit: radians / second) and attitude angle (roll angle, pitch angle, yaw angle, unit: degree) of the transport based on the transport identifier and the current timestamp. The acquired transport motion data is stored as the attribute of the special cargo-transport edge. If the real-time data cannot be obtained, the average acceleration, average angular velocity and attitude angle of the transport in the past specified time window are used as substitute data.

[0150] The special cargo-transport edge set is loaded into the transport path and road condition acquisition module. For each edge in the special cargo-transport edge set, the planned transport path of the connected transport and the road condition information of the path are obtained. The transport path is obtained by querying the transport management system and contains a series of geographic coordinate points. The road condition information is obtained by calling the API of the map service. The API call needs to provide the sequence of geographic coordinates of the transport path. The map service returns the road grade, flatness score, congestion index and weather forecast information for the next period of time for each section on the path. The flatness score is calculated from the road sensor data. The higher the value, the smoother the road surface. The congestion index is calculated from the real-time traffic flow data. The higher the value, the more congested it is. The obtained transport path road condition data is stored as the attribute of the special cargo-transport tool edge.

[0151] The transportation path road condition data, transportation tool movement data, and special cargo attribute data are loaded into the mechanical analysis model construction module. For each special cargo-transport tool edge, a mechanical analysis model instance for this transportation scenario is constructed. The model construction is based on multi-body dynamics theory, which simplifies special cargo into a rigid body with mass, center of gravity, and moment of inertia, and the chassis of the transportation tool is simplified into a moving support platform. According to the transportation path road condition data, the excitation of the support platform in the vertical and horizontal directions is simulated. According to the transportation tool movement data, the acceleration and angular velocity of the support platform are set. According to the special cargo attribute data, the mass, center of gravity position, moment of inertia, and connection constraints with the support platform (for example, friction coefficient) of the cargo are set. The mechanical analysis model instance can simulate the gravity, inertia, friction, and excitation force from the support platform that the cargo is subjected to during transportation.

[0152] The mechanical analysis model instance and the special cargo-transport edge set are loaded into the force condition prediction module. For each edge in the special cargo-transport edge set, the mechanical analysis model instance constructed in step S355 is used to predict the maximum impact force, maximum vibration frequency and maximum tilt angle of the special cargo during transportation. The prediction process is implemented by a numerical simulation method, which simulates the force conditions at different time points during transportation and records the maximum value. The maximum impact force is the peak value of the resultant force at the center of gravity of the cargo (unit: Newton). The maximum vibration frequency is the peak frequency of the acceleration spectrum of the cargo surface (unit: Hertz). The maximum tilt angle is the maximum tilt angle of the cargo relative to the horizontal plane (unit: degree). The force prediction results are stored as attributes of the special cargo-transport tool edge.

[0153] The special cargo-transport edge set and the force prediction results are loaded into the mechanical feature fusion module. For each edge in the special cargo-transport edge set, the corresponding force prediction result (maximum impact force, maximum vibration frequency, maximum tilt angle) is encoded as a vector of fixed dimension. The encoding method can directly use the three predicted values ​​as the three dimensions of the vector. The encoded mechanical feature vector is spliced ​​with the original feature vector (including spatiotemporal and dependent features) of the special cargo-transport tool edge to form a mechanical fusion edge feature.

[0154] The enhanced feature graph and mechanical fusion edge features are loaded into the mechanical model fusion graph generation module. This module copies the structure and node features of the enhanced feature graph. For each special cargo-transportation tool edge in the enhanced feature graph, the original edge feature is replaced with its corresponding mechanical fusion edge feature. For non-special cargo-transportation tool edges, their edge features remain unchanged. The final generated mechanical model fusion graph has the same structure as the enhanced feature graph, but the features of the edges connecting special cargo and transportation tools incorporate the force information predicted by the mechanical analysis model.

[0155] Preferably, step S4 comprises the following steps:

[0156] Step S41: Preliminary screening of candidate entity pairs is performed on the semantic enhancement features to obtain preliminary candidate entity pairs;

[0157] Step S42: Calculate the global feature similarity of the preliminary candidate entity pairs to obtain a global similarity score;

[0158] Step S43: clustering the preliminary candidate entity pairs according to the global similarity scores to obtain a global clustered entity set;

[0159] Step S44: using semantic enhancement features to perform local feature refinement matching on the global clustering entity set to obtain local matching entity pairs;

[0160] Step S45: performing unified alignment objective function optimization according to the multi-modal relationship set, the local matching entity pairs and the global clustering entity set to obtain an optimized alignment result;

[0161] Step S46: Generate an aligned entity set according to the optimized alignment result to obtain an aligned entity set.

[0162] In an embodiment of the present invention, semantic enhancement features are loaded into the preliminary screening module for candidate entity pairs. For all logistics entities from different data sources, preliminary matching is performed based on predefined global features. For warehouse entities, warehouse entities from different warehousing systems are added to the preliminary candidate entity pairs if the Jaccard similarity of their names is higher than 0.8 or the Haversine distance of their geographical locations is less than 1 km. For vehicle entities, vehicle entities from the transportation management system and the vehicle monitoring system are added to the preliminary candidate entity pairs if their license plate numbers are exactly the same. For order entities, order entities from the order management system and the logistics tracking system are added to the preliminary candidate entity pairs if their order numbers are exactly the same or the Jaro-Winkler distances of the consignee's name and the delivery address are both higher than 0.9. The preliminary candidate entity pairs store cross-source entity pairs that are considered to point to the same real-world entity.

[0163] The preliminary candidate entity pairs are loaded into the global feature similarity calculation module. For each preliminary candidate entity pair, the similarity score of its global features is calculated. For the warehouse entity pair, the Jaccard similarity of its name and the Haversine distance of its geographical location are calculated, and the two similarity values ​​are weighted averaged, with weights of 0.6 and 0.4 respectively. For the vehicle entity pair, if the license plate numbers are the same, the similarity score is 1, otherwise it is 0. For the order entity pair, the Jaro-Winkler distance of the consignee's name, the Jaro-Winkler distance of the delivery address, and the absolute difference in order time (unit: hours) are calculated, and the three similarity values ​​are weighted averaged, with weights of 0.3, 0.5 and 0.2 respectively. The calculated global similarity score is assigned to the corresponding preliminary candidate entity pair.

[0164] The preliminary candidate entity pairs and their global similarity scores are loaded into the entity clustering module. The density-based clustering algorithm DBSCAN is used to cluster the entities. Each entity is regarded as a node in the graph. If two entities are in the preliminary candidate entity pair and their global similarity scores are higher than the set threshold (for example, 0.7), there is an edge between the two entities. The DBSCAN algorithm builds density-reachable clusters based on connected entities. Entities in each cluster have high similarity in global features and are considered to be potentially the same entity. The clustering results form a global clustering entity set, each element of which is a cluster containing multiple cross-source entities.

[0165] The global clustering entity set and semantically enhanced features are loaded into the local feature refined matching module. For each cluster in the global clustering entity set, the semantically enhanced features of the entities it contains are used for more refined matching. For the warehouse entity cluster, the cosine similarity of the semantically enhanced feature vectors of any two warehouse entities in the cluster is calculated. For the vehicle entity cluster, the cosine similarity of the semantically enhanced feature vectors of any two vehicle entities in the cluster is calculated. For the order entity cluster, the cosine similarity of the semantically enhanced feature vectors of any two order entities in the cluster is calculated. If the cosine similarity of the semantically enhanced feature vectors of two entities is higher than the set threshold (for example, 0.85), the two entities are added to the local matching entity pairs. Local matching entity pairs store cross-source entity pairs that are highly similar at the semantic level.

[0166] The multimodal relationship set, local matching entity pairs, and global cluster entity set are loaded into the unified alignment objective function optimization module. An entity alignment graph is constructed, where nodes are all entities to be aligned and edges represent matching relationships between entities. Initially, there are potential matching edges between entities from the same global cluster. Using the information of local matching entity pairs, matching edges and their weights are added, and the weight value is equal to the cosine similarity of the semantically enhanced feature vectors of the local matching entity pairs. A unified alignment objective function is constructed, and the goal is to maximize the sum of the matching edge weights of the graph while satisfying the following constraint: an entity can only match one entity from another data source. The Hungarian algorithm or iterative local search algorithm is used to optimize the objective function to find the optimal entity matching solution. Optimization results Output optimized alignment results, including the final cross-source entity matching pairs and their matching confidence (for example, the semantic similarity score for local matching).

[0167] The optimized alignment results are loaded into the aligned entity set generation module. According to the cross-source entity matching pairs determined in the optimized alignment results, a unified entity ID is assigned to each matching entity pair. Entities from different data sources but judged to point to the same real-world entity are assigned the same unified entity ID. For entities that fail to match successfully, a new unique entity ID is assigned to them. All entities assigned a unified entity ID constitute an aligned entity set. The aligned entity set eliminates entity redundancy and inconsistency across data sources, laying the foundation for building a unified knowledge graph.

[0168] Preferably, step S5 comprises the following steps:

[0169] Step S51: perform entity traversal on the aligned entity set, and add each aligned entity to the pre-built initial graph to obtain an entity-filled graph;

[0170] Step S52: traverse the multi-mode relationship set, and add each relationship to the entity filling graph to obtain a relationship connection graph;

[0171] Step S53: applying hierarchical organization to the relationship connection map to obtain a hierarchical organization map;

[0172] Step S54: supplementing the hierarchical organization map with multi-dimensional relationship attributes according to the multi-mode relationship set to obtain an attributed relationship map;

[0173] Step S55: construct an index for the attributed relationship graph to obtain an index optimization graph;

[0174] Step S56: Generate a logistics distribution knowledge graph for the index optimization graph to obtain a logistics distribution knowledge graph.

[0175] In an embodiment of the present invention, the aligned entity set is loaded into the entity filling module. The pre-built initial graph is stored in a graph database, and the initial state is empty. The entity filling module traverses each entity in the aligned entity set. For each entity, a new node is created in the graph according to its unified entity ID. The attributes of the node are set according to the information of the entity in the original data, including entity type (e.g., order, vehicle, warehouse), name, description, and other related attributes. For example, for an order entity corresponding to a unified entity ID, its node attributes include order number, order time, shipping address, etc. After the traversal is completed, all aligned entities are added to the graph to form an entity filling graph. The nodes of the entity filling graph represent the core concepts in the logistics field, but no connection has been established between the nodes.

[0176] The multimodal relationship set and entity-filled graph are loaded into the relationship connection module. The relationship connection module traverses each relationship in the multimodal relationship set. For each relationship, find the corresponding node in the entity-filled graph based on the unified entity ID of the two entities it connects. Create a directed edge between the two nodes, and the edge type is set according to the relationship type in the multimodal relationship set (for example, "distribution", "storage", "contains"). After the traversal is completed, all multimodal relationships are added to the graph, connecting the relevant entity nodes to form a relationship connection graph. The relationship connection graph preliminarily constructs the associations between entities in the logistics field.

[0177] The relational connection graph is loaded into the hierarchical organization module. The hierarchical organization module hierarchizes the entities and relationships in the graph according to the predefined logistics domain ontology. The ontology defines the type hierarchy of entities (e.g., geographic location > province > city > distribution point) and the semantic hierarchy of relationships (e.g., physical connection > logical dependency). The module traverses each entity and relationship in the graph and classifies them into the corresponding hierarchy defined in the ontology according to their type and attributes. For example, warehouse entities are classified into the "geographic location" and "facility" hierarchies, order entities are classified into the "business object" hierarchy, and "distribution" relationships are classified into the "logistics activity" hierarchy. By adding hierarchical attributes to nodes and edges, the hierarchical organization of the graph is realized to form a hierarchical organization graph. Hierarchical organization improves the interpretability and query efficiency of the graph.

[0178] The multimodal relationship set and the hierarchical organization graph are loaded into the relationship attribute supplementation module. The relationship attribute supplementation module traverses each relationship in the multimodal relationship set. For each relationship, according to the attribute information it contains, add attributes to the corresponding edge in the hierarchical organization graph. For example, for a "distribution" relationship, its multimodal relationship set contains attributes such as "estimated arrival time", "actual arrival time", and "distribution distance", which are added to the attribute list of the "distribution" relationship edge in the graph. For the "storage" relationship, it contains attributes such as "storage temperature" and "storage time". By supplementing multi-dimensional relationship attributes, the relationship information in the graph is made more complete and rich, forming an attributed relationship graph.

[0179] The attributed relationship graph is loaded into the index building module. The index building module analyzes the attributes of entities and relationships in the graph and builds indexes based on common query patterns to improve the query efficiency of the graph. For nodes, indexes are built for attributes such as entity ID, entity type, and name. For edges, indexes are built for relationship types and commonly used relationship attributes (for example, the "estimated arrival time" of the "delivery" relationship). Index building is achieved using technologies such as B-tree indexing or full-text indexing. After the construction is completed, an index-optimized graph is formed, which supports efficient attribute queries and relationship queries.

[0180] The index optimization graph is loaded into the knowledge graph generation module. The knowledge graph generation module stores the index optimization graph as the final logistics and distribution knowledge graph. The storage format can adopt the native format of the graph database, or the graph data can be exported as RDF or other graph data exchange formats. The generated logistics and distribution knowledge graph contains multi-dimensional entities and their rich attribute information in the logistics scenario, as well as various types of relationship connections between entities. It has good interpretability and efficient query capabilities through hierarchical organization and index construction, and can serve as the data foundation for downstream intelligent logistics applications.

[0181] Preferably, the present invention further provides a data fusion system based on a logistics distribution knowledge graph, which is used to execute the data fusion method based on a logistics distribution knowledge graph as described above. The data fusion system based on a logistics distribution knowledge graph comprises:

[0182] The data normalization module is used to obtain the original data of logistics distribution; perform independent representation learning of heterogeneous data on the original data of logistics distribution to obtain an independent feature vector set; perform unified feature representation processing on the independent feature vector set to obtain a unified feature representation;

[0183] The semantic enhancement module is used to extract context information from the unified feature representation and perform dynamic perception embedding to obtain dynamic context embedding data; perform multi-dimensional semantic fusion on the dynamic context embedding data to obtain semantic enhancement features;

[0184] The relationship extraction module is used to integrate the spatiotemporal and dependent features according to the semantic enhancement features to obtain an enhanced feature map; to fuse the physical motion model features according to the enhanced feature map to obtain a motion model fusion map; to fuse the cargo stacking model features according to the enhanced feature map to obtain a stacking model fusion map; to fuse the mechanical analysis features according to the enhanced feature map to obtain a mechanical model fusion map; to perform graph neural network message transmission and relationship inference according to the motion model fusion map, the stacking model fusion map and the mechanical model fusion map to obtain a relationship inference score; to perform multi-modal feature fusion and relationship determination according to the relationship inference score to obtain a multi-modal relationship set;

[0185] The knowledge fusion module is used to perform global entity clustering analysis based on semantic enhancement features to obtain a global clustering entity set; perform local feature refinement matching on the global clustering entity set to obtain local matching entity pairs; perform hierarchical entity alignment based on the multi-modal relationship set, local matching entity pairs and the global clustering entity set to obtain an aligned entity set;

[0186] The graph construction module is used to construct an entity filling graph according to the aligned entity set to obtain an entity filling graph; to supplement the entity filling graph with multi-dimensional relationship attributes according to the multi-modal relationship set to obtain an attributed relationship graph; and to generate a logistics distribution knowledge graph according to the attributed relationship graph to obtain a logistics distribution knowledge graph.

[0187] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0188] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A data fusion method based on logistics distribution knowledge graph, characterized in that: The following steps are involved: Step S1: Obtaining the original data of logistics distribution; performing independent representation learning of heterogeneous data on the original data of logistics distribution to obtain an independent feature vector set; Performing unified feature representation processing on the independent feature vector set to obtain a unified feature representation; Step S2: extracting context information from the unified feature representation and performing dynamic perception embedding to obtain dynamic context embedding data; performing multi-dimensional semantic fusion on the dynamic context embedding data to obtain semantic enhancement features; Step S3: According to the semantic enhancement feature, the spatiotemporal and dependent features are integrated to obtain an enhanced feature map; According to the enhanced feature map, the physical motion model feature fusion is performed to obtain the motion model fusion map; according to the enhanced feature map, the cargo stacking model feature fusion is performed to obtain the stacking model fusion map; according to the enhanced feature map, the mechanical analysis feature fusion is performed to obtain the mechanical model fusion map; according to the motion model fusion map, the stacking model fusion map and the mechanical model fusion map, the graph neural network message transmission and relationship inference are performed to obtain the relationship inference score; Perform multi-modal feature fusion and relationship determination based on the relationship inference score to obtain a multi-modal relationship set; Step S4: performing global entity clustering analysis based on semantic enhancement features to obtain a global clustering entity set; performing local feature refinement matching on the global clustering entity set to obtain local matching entity pairs; performing hierarchical entity alignment based on the multimodal relationship set, the local matching entity pairs and the global clustering entity set to obtain an aligned entity set; Step S5: construct an entity filling graph according to the aligned entity set to obtain an entity filling graph; supplement the entity filling graph with multi-dimensional relationship attributes according to the multi-modal relationship set to obtain an attributed relationship graph; The logistics distribution knowledge graph is generated according to the attributed relationship graph to obtain the logistics distribution knowledge graph.

2. The data fusion method based on the logistics distribution knowledge graph according to claim 1 is characterized in that: Step S1 includes the following steps: Step S11: obtaining original logistics distribution data; performing logistics data preprocessing on the original logistics distribution data to obtain preprocessed logistics data; Step S12: Performing independent representation learning of heterogeneous data on the pre-processed logistics data to obtain an independent feature vector set; Step S13: using the pre-built logistics domain knowledge base to perform cross-modal alignment on the independent feature vector sets to obtain an aligned feature space set; Step S14: performing unified feature representation fusion on the aligned feature space set to obtain a unified feature representation.

3. The data fusion method based on logistics distribution knowledge graph according to claim 1 is characterized in that: Step S2 includes the following steps: Step S21: extracting context information from the unified feature representation using a pre-built logistics domain knowledge base to obtain a context information set; Step S22: supplementing the unified feature representation with knowledge features according to the context information set and the pre-built logistics domain knowledge base to obtain knowledge fusion features; Step S23: dynamically context-aware embedding is performed on the knowledge fusion features according to the context information set to obtain dynamic context embedding data; Step S24: Perform multi-dimensional semantic fusion on the dynamic context embedding data to obtain semantic enhancement features.

4. The data fusion method based on logistics distribution knowledge graph according to claim 1 is characterized in that: Step S3 includes the following steps: Step S31: generating candidate entity pairs based on the semantic enhancement features to obtain a candidate entity pair set; extracting spatiotemporal features based on the candidate entity pair set and the semantic enhancement features to obtain a spatiotemporal feature vector; extracting resource dependency features based on the candidate entity pair set and the semantic enhancement features to obtain a dependency feature vector; Step S32: constructing an initial logistics relationship graph according to the candidate entity pair set and the semantic enhancement features to obtain an initial relationship graph; integrating the time, space and dependency features of the initial relationship graph according to the semantic enhancement features and the dependency feature vector to obtain an enhanced feature graph; Step S33: performing physical motion model feature fusion according to the enhanced feature map to obtain a motion model fusion map; Step S34: performing cargo stacking model feature fusion according to the enhanced feature map to obtain a stacking model fusion map; Step S35: performing mechanical analysis feature fusion according to the enhanced feature map to obtain a mechanical model fusion map; Step S36: performing graph neural network message transmission and relationship inference according to the motion model fusion graph, the stacking model fusion graph, and the mechanical model fusion graph to obtain a relationship inference score; Step S37: performing time series-based relational pattern mining according to the candidate entity pair set and the semantic enhancement features to obtain a time relational pattern; Step S38: Perform multi-modal feature fusion and relationship determination based on the relationship inference score and the temporal relationship pattern to obtain a multi-modal relationship set.

5. The data fusion method based on the logistics distribution knowledge graph according to claim 4 is characterized in that: Step S33 includes the following steps: Step S331: filter the vehicle-position edges on the enhanced feature graph to obtain a vehicle-position edge set; Step S332: extracting vehicle historical trajectories from the vehicle-position edge set to obtain vehicle historical trajectory data; Step S333: acquiring the current vehicle state from the vehicle-position edge set to obtain the current vehicle state information; Step S334: Acquire road network information for the vehicle-position edge set to obtain road network information data; Step S335: predicting the estimated time for the vehicle to reach the target location based on the vehicle-location edge set, the vehicle historical trajectory data, the current vehicle status information, and the road network information data, and performing a confidence evaluation to obtain an ETA prediction result; Step S336: performing ETA feature fusion according to the vehicle-position edge set and the ETA prediction result to obtain ETA fused edge features; Step S337: Generate a motion model fusion map based on the enhanced feature map and the ETA fusion edge features to obtain a motion model fusion map.

6. The data fusion method based on the logistics distribution knowledge graph according to claim 4 is characterized in that: Step S34 includes the following steps: Step S341: filtering the goods-location edges on the enhanced feature graph to obtain a goods-location edge set; Step S342: Acquire the physical attributes of the goods on the goods-location edge set to obtain the goods attribute data; Step S343: Acquire the shelf physical structure of the goods-location edge set to obtain shelf structure data; Step S344: constructing a cargo stacking model according to the cargo attribute data and the shelf structure data to obtain a cargo stacking model instance; Step S345: using the cargo stacking model instance to evaluate the stacking characteristics of the cargo-location edge set to obtain a stacking evaluation result; Step S346: performing stacking feature fusion according to the stacking evaluation result and the goods-location edge set to obtain a stacking fusion edge feature; Step S347: Generate a stacked model fusion graph based on the enhanced feature graph and the stacked fusion edge features to obtain a stacked model fusion graph.

7. The data fusion method based on logistics distribution knowledge graph according to claim 4 is characterized in that: Step S35 includes the following steps: Step S351: filtering the special cargo-transportation tool edges on the enhanced feature graph to obtain a special cargo-transportation edge set; Step S352: acquiring special cargo attributes from the special cargo-transport edge set to obtain special cargo attribute data; Step S353: acquiring the transport tool movement state for the special cargo-transportation edge set to obtain the transport tool movement data; Step S354: acquiring the transportation path and road condition of the special cargo-transport edge set to obtain the transportation path road condition data; Step S355: constructing a mechanical analysis model according to the transportation route road condition data, transportation tool movement data and special cargo attribute data to obtain a mechanical analysis model instance; Step S356: using the mechanical analysis model instance to predict the stress condition of the special cargo-transportation edge set, and obtaining the stress prediction result; Step S357: performing mechanical feature fusion according to the force prediction result and the special cargo-transport edge set to obtain mechanical fusion edge features; Step S358: Generate a mechanical model fusion graph based on the enhanced feature graph and the mechanical fusion edge features to obtain a mechanical model fusion graph.

8. The data fusion method based on logistics distribution knowledge graph according to claim 1 is characterized in that: Step S4 includes the following steps: Step S41: Preliminary screening of candidate entity pairs is performed on the semantic enhancement features to obtain preliminary candidate entity pairs; Step S42: Calculate the global feature similarity of the preliminary candidate entity pairs to obtain a global similarity score; Step S43: clustering the preliminary candidate entity pairs according to the global similarity scores to obtain a global clustered entity set; Step S44: using semantic enhancement features to perform local feature refinement matching on the global clustering entity set to obtain local matching entity pairs; Step S45: performing unified alignment objective function optimization according to the multi-modal relationship set, the local matching entity pairs and the global clustering entity set to obtain an optimized alignment result; Step S46: Generate an alignment entity set according to the optimized alignment result to obtain an alignment entity set.

9. The data fusion method based on logistics distribution knowledge graph according to claim 1 is characterized in that: Step S5 includes the following steps: Step S51: perform entity traversal on the aligned entity set, and add each aligned entity to the pre-built initial graph to obtain an entity-filled graph; Step S52: traverse the multi-mode relationship set, and add each relationship to the entity filling graph to obtain a relationship connection graph; Step S53: applying hierarchical organization to the relationship connection map to obtain a hierarchical organization map; Step S54: supplementing the hierarchical organization map with multi-dimensional relationship attributes according to the multi-mode relationship set to obtain an attributed relationship map; Step S55: construct an index for the attributed relationship graph to obtain an index optimization graph; Step S56: Generate a logistics distribution knowledge graph for the index optimization graph to obtain a logistics distribution knowledge graph.

10. A data fusion system based on logistics distribution knowledge graph, characterized in that: Used to execute the data fusion method based on the logistics distribution knowledge graph as claimed in claim 1, the data fusion system based on the logistics distribution knowledge graph comprises: The data normalization module is used to obtain the original data of logistics distribution; perform independent representation learning of heterogeneous data on the original data of logistics distribution to obtain an independent feature vector set; perform unified feature representation processing on the independent feature vector set to obtain a unified feature representation; The semantic enhancement module is used to extract context information from the unified feature representation and perform dynamic perception embedding to obtain dynamic context embedding data; perform multi-dimensional semantic fusion on the dynamic context embedding data to obtain semantic enhancement features; The relationship extraction module is used to integrate the spatiotemporal and dependent features according to the semantic enhancement features to obtain an enhanced feature map; to fuse the physical motion model features according to the enhanced feature map to obtain a motion model fusion map; to fuse the cargo stacking model features according to the enhanced feature map to obtain a stacking model fusion map; to fuse the mechanical analysis features according to the enhanced feature map to obtain a mechanical model fusion map; to perform graph neural network message transmission and relationship inference according to the motion model fusion map, the stacking model fusion map and the mechanical model fusion map to obtain a relationship inference score; to perform multi-modal feature fusion and relationship determination according to the relationship inference score to obtain a multi-modal relationship set; The knowledge fusion module is used to perform global entity clustering analysis based on semantic enhancement features to obtain a global clustering entity set; perform local feature refinement matching on the global clustering entity set to obtain local matching entity pairs; perform hierarchical entity alignment based on the multi-modal relationship set, local matching entity pairs and the global clustering entity set to obtain an aligned entity set; The graph construction module is used to construct an entity filling graph according to the aligned entity set to obtain an entity filling graph; to supplement the entity filling graph with multi-dimensional relationship attributes according to the multi-modal relationship set to obtain an attributed relationship graph; and to generate a logistics distribution knowledge graph according to the attributed relationship graph to obtain a logistics distribution knowledge graph.

Images