Intelligent logistics integrated management method and system

Abstract

The application discloses an intelligent logistics integrated management and control method and system, and belongs to the technical field of logistics information management, and comprises the following steps: acquiring logistics management and control data of a target logistics entity, and generating a logistics management and control data set; processing two types of data through a logistics level coding embedding model and a logistics time sequence signal embedding model respectively, and generating logistics level feature mapping values and logistics time sequence feature mapping values; performing a fusion operation on the two types of feature mapping values to obtain a logistics node fusion feature set; calling a federated hybrid model to process the logistics node fusion feature set, and generating a scheduling index set representing a scheduling priority of a target logistics node; if the scheduling priority meets a preset threshold, determining a target scheduling node and sending a scheduling instruction, and after the target scheduling node completes a task, registering the target scheduling node as a target management and control node; and obtaining and allocating a management and control path scheme according to a node type indicated by an operation analysis result of the target management and control node.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of logistics information management, and in particular to an intelligent logistics integrated management method and system. BACKGROUND

[0002] With the rapid development of the logistics industry, the logistics network is becoming increasingly complex, the number of logistics nodes involved is increasing, and the types of nodes are becoming more diversified. Traditional logistics management methods rely on manual scheduling or single-dimensional data for decision-making, which has problems such as low management efficiency, decision lag, and unreasonable resource allocation. In the prior art, logistics data collection and processing are often independent of each other, making it difficult to effectively integrate multi-source data, resulting in an incomplete and inaccurate assessment of the status of logistics nodes. At the same time, in the process of scheduling logistics nodes, there is a lack of a scientific priority evaluation mechanism, which cannot dynamically adjust the management and control strategy according to the actual situation of the nodes, affecting the operation efficiency of the entire logistics network. Therefore, there is an urgent need for an intelligent logistics integrated management method and system that can achieve deep integration of logistics data, accurately assess the status of nodes, and dynamically optimize management and control strategies. SUMMARY

[0003] To overcome the shortcomings of the prior art, the present application provides an intelligent logistics integrated management method and system, which acquires logistics management data of a target logistics entity and generates a logistics management data set. The logistics management data set includes logistics node coding data and logistics state continuous time series signal data. The logistics node coding data is processed by a logistics hierarchical coding embedding model to generate logistics hierarchical feature mapping values, and the logistics state continuous time series signal data is processed by a logistics time series signal embedding model to generate logistics time series feature mapping values. Fusion operations are performed on the two types of feature mapping values to obtain a logistics node fusion feature set. A federated hybrid model is called to process the logistics node fusion feature set to generate a scheduling index set representing the scheduling priority of the target logistics node. If the scheduling priority meets a predetermined threshold, a target scheduling node is determined and a scheduling instruction is sent, and after the target scheduling node completes the work, it is registered as a target management and control node. According to the node type indicated by the operation analysis result of the target management and control node, a management and control path scheme is obtained and allocated. This method realizes efficient management and control of logistics nodes and improves the efficiency of logistics operations.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0005] An intelligent logistics integrated management method, comprising:

[0006] Acquiring logistics management data of a target logistics entity and generating a logistics management data set. The logistics management data set includes logistics node coding data and logistics state continuous time series signal data.

[0007] Processing the logistics node coding data by a logistics hierarchical coding embedding model to generate logistics hierarchical feature mapping values, and simultaneously processing the logistics state continuous time series signal data by a logistics time series signal embedding model to generate logistics time series feature mapping values.

[0008] performing a fusion operation on the logistics level feature mapping value and the logistics time sequence feature mapping value to generate a logistics node fusion feature set;

[0009] calling a federated hybrid model to process the logistics node fusion feature set to generate a scheduling index set representing a scheduling priority of a target logistics node;

[0010] If the scheduling priority in the scheduling index set meets a preset scheduling threshold, a corresponding logistics node is determined as a target scheduling node and a scheduling instruction is sent to it, and after the target scheduling node completes the scheduling task, it is registered as a target management and control node, and the task completion record and equipment state of the node are updated synchronously;

[0011] According to the logistics node type indicated by the operation analysis result of the target management and control node, a management and control path scheme is matched or generated from a preset management and control path library and is assigned to the target management and control node.

[0012] Specifically, the logistics management and control data of the target logistics entity is obtained and a logistics management and control data set is generated, including:

[0013] The boundary and identification of the target logistics entity are determined, and a unique entity code is assigned to each target logistics entity; the target logistics entity is a unit with independent logistics operation function; the unit includes a warehouse, a transportation hub, and a distribution site; the entity code includes entity type, belonging area, and serial number information;

[0014] Collecting logistics node coding data with hierarchical relationship, including: extracting logistics node coding from the warehouse management system of the target logistics entity; the logistics node coding is constructed in five levels of entity-area-aisle-shelf-locus; extracting goods classification coding, the goods classification coding is constructed based on goods type, storage requirement, and specification, and the goods classification coding and the locus coding are associated through a locus-goods binding table to record the current storage goods type of each locus;

[0015] Collecting logistics state continuous time sequence signal data in real time through Internet of Things devices deployed in the target logistics entity; the logistics state continuous time sequence signal data includes time sequence of environmental parameters, goods throughput, location information, and equipment operation parameters;

[0016] Pretreating the logistics node coding data and the logistics state continuous time sequence signal data, and associating the pretreated logistics node coding data and the logistics state continuous time sequence signal data through entity coding and locus coding to generate a structured logistics management and control data set; each row in the logistics management and control data set corresponds to the coding of a logistics node and the associated logistics state continuous time sequence signal data.

[0017] Specifically, the generation of the logistics level feature mapping value comprises:

[0018] Obtaining logistics node coding data and preprocessing the same;

[0019] Converting the preprocessed logistics node coding data into a graph structure; in the graph structure conversion, each logistics node coding is taken as a node of a graph, and edges of the graph are constructed based on the hierarchical relationship of the logistics node coding;

[0020] Initializing a logistics level coding embedding model; the logistics level coding embedding model adopts a hyperbolic graph convolutional network structure, and the input is the constructed graph structure data, and the output is a fixed-dimension feature vector; the logistics level coding embedding model comprises a graph convolutional layer, a hyperbolic space mapping layer and a feature optimization layer, wherein the graph convolutional layer is used to aggregate neighbor features of a node, the hyperbolic space mapping layer is used to map the features to a hyperbolic space, and the feature optimization layer is used to improve the discrimination of the feature mapping value;

[0021] Inputting the graph structure data into the initialized logistics level coding embedding model for training to obtain a trained logistics level coding embedding model;

[0022] Inputting real-time preprocessed logistics node coding data into the trained logistics level coding embedding model to output an N-dimensional vector corresponding to each logistics node coding, i.e., a logistics level feature mapping value; the logistics level feature mapping value is located in a hyperbolic space and contains hierarchical association, cargo attribute and functional area information of the logistics node.

[0023] Specifically, the inputting of the graph structure data into the initialized logistics level coding embedding model comprises:

[0024] Weighted aggregation of all neighbor node features of each node is performed through the graph convolutional layer to obtain preliminary hierarchical aggregation features, wherein in the weighted aggregation process, the weight calculation combines the edge weight and the distance of the node in the hierarchy;

[0025] Through the hyperbolic space mapping layer, a Euclidean-hyperbolic conversion function is used to map the preliminary hierarchical aggregation features from Euclidean space to hyperbolic space to obtain a preliminary hierarchical feature vector in the hyperbolic space; the hyperbolic space is constructed by using a Poincare ball model, and the distance between any two points in the ball is calculated by using a Poincare distance formula; the calculation method of the Euclidean-hyperbolic conversion function is as follows: first, the modulus of a Euclidean space vector v is calculated, then the hyperbolic sine value of the modulus is calculated, then the hyperbolic sine value is multiplied by v, and finally the obtained result is divided by the modulus of v;

[0026] The two-class loss includes a hierarchy loss and a cargo type loss; the hierarchy loss is obtained by calculating the Poincare distance of the feature vectors of parent nodes and child nodes; and the cargo type loss is obtained by calculating the similarity of the feature vectors of all storage nodes corresponding to the same cargo classification code.

[0027] Specifically, the logistics time sequence feature mapping value is generated, including:

[0028] Obtain logistics state continuous time sequence signal data and preprocess the same;

[0029] Time sequence segmentation and standardization are performed on the preprocessed logistics state continuous time sequence signal data to obtain a standardized logistics state continuous time sequence;

[0030] Initialize a logistics time sequence signal embedding model; the logistics time sequence signal embedding model adopts a bidirectional long short-term memory network structure, and the input is the preprocessed logistics state continuous time sequence signal data, and the output is a fixed-dimension feature vector; the logistics time sequence signal embedding model includes a forward LSTM layer, a backward LSTM layer, a bidirectional feature fusion layer and an attention layer, wherein the forward LSTM layer captures forward dependence of the time sequence, the backward LSTM layer captures reverse dependence, the bidirectional feature fusion layer integrates bidirectional information, and the attention layer highlights key time sequence features;

[0031] The standardized logistics state continuous time sequence is input into the initialized logistics time sequence signal embedding model to obtain an N-dimensional logistics time sequence feature mapping value.

[0032] Specifically, the standardized logistics state continuous time sequence is input into the initialized logistics time sequence signal embedding model to obtain an N-dimensional logistics time sequence feature mapping value, including:

[0033] The forward LSTM layer starts iteration from the starting time step of the logistics state continuous time sequence, combines the current time step input and the previous step hidden state at each step, outputs the forward time sequence feature state of the current time step, and records the forward dynamic features of the time sequence;

[0034] The backward LSTM layer reverses the logistics state continuous time sequence, starts iteration from the starting time step of the reversed sequence, combines the current time step input and the previous step hidden state at each step, outputs the reverse time sequence feature state of the current time step, and records the reverse dynamic features of the time sequence;

[0035] The bidirectional feature fusion layer adopts a vector splicing and linear transformation manner to combine the forward time sequence feature state and the reverse time sequence feature state into an M-dimensional bidirectional feature vector, and then enhances the non-linear expression ability of the features through a ReLU activation function to obtain a preliminary time sequence feature vector;

[0036] The attention layer assigns weights according to the importance of the logistics state continuous time sequence signal data, and obtains an attention weight;

[0037] The preliminary time sequence feature vector is multiplied by the attention weight element by element to obtain an N-dimensional logistics time sequence feature mapping value.

[0038] Specifically, the logistics node fusion feature set is generated, including:

[0039] A hyperbolic-Euclidean conversion function is used to project the logistics level feature mapping value in the hyperbolic space to the origin of the Euclidean space to generate a Euclidean space feature mapping value; the calculation method of the hyperbolic-Euclidean conversion function is: first, calculate the modulus of the hyperbolic space vector u, then calculate the inverse hyperbolic tangent value of the modulus, then multiply the inverse hyperbolic tangent value by u, and finally divide the result by the modulus of u;

[0040] The logistics time sequence feature mapping value is mapped to the same dimension and distribution space as the Euclidean space feature mapping value through a linear transformation layer to generate an adaptive time sequence feature mapping value;

[0041] A fusion weight predictor is constructed to output level feature weights and time sequence feature weights; the fusion weight predictor is constructed based on a logistic regression model;

[0042] The Euclidean space feature mapping value and the adaptive time sequence feature mapping value are respectively weighted and summed according to the level feature weight and the time sequence feature weight to generate a logistics feature fusion vector; the logistics feature fusion vector is N-dimensional;

[0043] A Euclidean-hyperbolic conversion function is used to map the logistics feature fusion vector from the Euclidean space back to the hyperbolic space with the origin of the hyperbolic space as the benchmark to generate an alignment feature vector;

[0044] All alignment feature vectors corresponding to the logistics nodes are collected and stored in the format of logistics node code-alignment feature vector-generation timestamp to generate a logistics node fusion feature set.

[0045] Specifically, the federated hybrid model is composed of a local XGBoost model of each logistics node and a cloud federated aggregation layer; the local XGBoost model processes local sub-features in the logistics node fusion feature set to output local scheduling features; the local scheduling features are encrypted and transmitted to the cloud federated aggregation layer; the cloud federated aggregation layer aggregates global features through a federated averaging algorithm to finally generate a scheduling index set, which includes node cargo backlog rate, equipment utilization rate, order overtime rate and transportation cost coefficient.

[0046] Specifically, the operation analysis result is generated based on application module logs of a warehouse control system log management module, system operation logs and extended warehouse management interface records of an interface log management module; the operation analysis result includes a temperature and humidity compliance rate of a warehousing node, a device failure rate of a transportation node and a task delay rate of a distribution node.

[0047] The intelligent logistics integrated management and control system comprises a data acquisition module, a feature mapping module, a feature fusion module, a scheduling evaluation module, a node scheduling module and a scheme matching module.

[0048] The data acquisition module is configured to acquire various types of management and control data of a target logistics entity, and generate a structured logistics management and control data set after processing.

[0049] The feature mapping module is configured to perform feature extraction on logistics node coding data and logistics state time sequence signal data, and generate feature mapping values reflecting node hierarchical attributes and time sequence dynamics.

[0050] The feature fusion module is configured to fuse logistics hierarchical feature mapping values and logistics time sequence feature mapping values, generate aligned feature vectors and form a logistics node fusion feature set.

[0051] The scheduling evaluation module is configured to calculate scheduling indexes by using a federated hybrid model based on the logistics node fusion feature set, and evaluate node scheduling priorities.

[0052] The node scheduling module is configured to perform scheduling operations on a target scheduling node, track job completion conditions and update node states.

[0053] The scheme matching module is configured to match or generate adaptive management and control path schemes according to the type of a target management and control node.

[0054] Compared with the prior art, the present application has the following advantages:

[0055] 1. The intelligent logistics integrated management and control method is proposed, which integrates logistics node coding data and time sequence signal data collected by the Internet of Things, performs feature extraction by using hierarchical coding embedding and time sequence signal embedding models respectively, realizes effective alignment of multi-dimensional features through fusion operations, generates scheduling priority indexes by using a federated hybrid model, accurately identifies logistics nodes that need to be controlled, and improves the efficiency and accuracy of logistics node scheduling.

[0056] 2.The intelligent logistics integrated management method is proposed, which dynamically matches or generates a management and control path scheme suitable for different node types, combines real-time updates of task completion records and equipment states, forms a complete management and control closed loop from data collection, analysis and decision-making to execution feedback, not only realizes intelligent and integrated management and control of logistics nodes, but also helps to optimize the overall logistics network resource allocation, reduce operating costs, and improve the overall operation efficiency and service quality of the logistics system. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 It is a schematic diagram of the intelligent logistics integrated management method of the present application.

[0058] Figure 2 It is a principle flow chart of the intelligent logistics integrated management method of the present application.

[0059] Figure 3 It is an architecture diagram of the intelligent logistics integrated management system of the present application. DETAILED DESCRIPTION

[0060] Embodiment 1

[0061] Please refer to Figure 1 and Figure 2 An embodiment provided by the present application is an intelligent logistics integrated management method, which comprises S1-S6 and the following steps:

[0062] S1: Obtain the logistics management data of the target logistics entity and generate a logistics management data set; the logistics management data set comprises logistics node coding data and logistics state continuous time series signal data;

[0063] Further, the logistics node coding data comprises logistics network point level coding and cargo classification coding; the logistics state continuous time series signal data comprises but is not limited to logistics transportation location time series data, warehouse environment temperature and humidity time series data, and cargo loading and unloading efficiency time series data;

[0064] S2: Process the logistics node coding data through a logistics level coding embedding model to generate logistics level feature mapping values, and at the same time, process the logistics state continuous time series signal data through a logistics time series signal embedding model to generate logistics time series feature mapping values;

[0065] S3: Perform fusion operation on the logistics level feature mapping values and the logistics time series feature mapping values to generate a logistics node fusion feature set;

[0066] S4: Call a federated hybrid model to process the logistics node fusion feature set to generate a scheduling index set representing the scheduling priority of the target logistics node;

[0067] The federal hybrid model is composed of a local XGBoost model of each logistics node and a cloud federated aggregation layer. The local XGBoost model first processes local sub-features in the logistics node fusion feature set to output local scheduling features. The local scheduling features are encrypted and transmitted to the cloud federated aggregation layer. The cloud federated aggregation layer aggregates global features through a federated averaging algorithm to finally generate a scheduling index set, which includes node cargo backlog rate, equipment utilization rate, order overtime rate, and transportation cost coefficient.

[0068] Further, the local XGBoost model uses 500 decision trees, a learning rate of 0.1, and a maximum depth of 6. The objective function is a multi-classification loss function. During model training, the aligned feature vector in the logistics node fusion feature set is input, and the manually labeled scheduling priority level is used as the label. After training, the local scheduling features are output, including preliminary calculation results such as node cargo backlog rate and equipment utilization rate. The local scheduling features are encrypted and transmitted to the cloud federated aggregation layer through a homomorphic encryption algorithm such as Paillier algorithm to ensure data privacy and security.

[0069] Further, the federated averaging algorithm of the cloud federated aggregation layer executes the following process: collecting the local scheduling features encrypted and transmitted by each logistics node, setting weights for feature values of the same dimension according to node data volume, and the larger the data volume, the higher the weight; calculating the weighted average to obtain global scheduling features, and then converting the global scheduling features into a scheduling index set through a preset index conversion formula, such as the node cargo backlog rate being equal to the percentage of the backlog cargo volume and the total capacity; the generated scheduling index set is stored according to the logistics node code for subsequent scheduling priority judgment.

[0070] S5: If the scheduling priority in the scheduling index set meets the preset scheduling threshold, the corresponding logistics node is determined as the target scheduling node and a scheduling instruction is sent to it. After the target scheduling node completes the scheduling task, it is registered as the target management and control node, and the task completion record and equipment state of the node are updated synchronously;

[0071] Further, the preset scheduling threshold is set by the following method: extracting historical scheduling data of the past 6 months, and counting the scheduling priority distribution of different node types; using the percentile method to set the scheduling threshold for each node type, such as taking the 90th percentile value for warehouse nodes and the 85th percentile value for transportation nodes; adjusting the threshold by ±10% dynamically according to the monthly logistics peak fluctuation, and using the adjusted threshold for the next round of scheduling priority comparison.

[0072] Further, the scheduling instruction includes a list of goods to be processed, a resource allocation scheme, and a job timing arrangement, which is sent to the control terminal of the target scheduling node through an encrypted communication channel.

[0073] Further, the task completion record includes job start time, completion time, processed goods quantity and abnormal event log, and the equipment state includes equipment operation parameter, energy consumption data and fault warning information.

[0074] S6: According to the logistics node type indicated by the operation analysis result of the target management and control node, a management and control path scheme is matched or generated from a preset management and control path library and is allocated to the target management and control node.

[0075] The operation analysis result is generated based on the application module log of the warehouse control system log management module, the system operation log and the extended warehouse management interface record of the interface log management module; and the operation analysis result includes the temperature and humidity compliance rate of the warehousing node, the equipment failure rate of the transportation node and the task delay rate of the distribution node.

[0076] Further, the logistics node type includes a warehousing node, a transportation node, a distribution node, a transportation hub node, a temporary transfer node and a cross-border customs node, and each type of node corresponds to a preset management and control path template.

[0077] Further, the preset management and control path library stores standard operation processes, resource allocation schemes and abnormal processing mechanisms among different node types; and the matching process is based on node type, current load and task characteristics for multi-dimensional retrieval.

[0078] Further, the process of matching the management and control path scheme from the preset management and control path library is that the preset management and control path library stores standard schemes classified by logistics node type, each scheme including an operation process, a resource allocation standard and an abnormal processing mechanism; in the matching process, first, the same type of scheme in the library is filtered according to the type of the target management and control node, then the parameters in the scheme are adjusted in combination with the operation analysis result, and finally the scheme is allocated to the target management and control node.

[0079] Further, when there is no matching management and control path scheme in the management and control path library, a new management and control path scheme is generated based on a genetic algorithm, the new scheme is verified for feasibility after simulation test, and then the new scheme is allocated to the target management and control node; the genetic algorithm takes the highest operation efficiency and the lowest cost as the optimization target, wherein the genetic algorithm is a prior art content in the field and is not the creative scheme of the present application, and thus is not described herein.

[0080] Further, after the management and control path scheme is allocated, the execution progress of the scheme of the target management and control node is monitored in real time, execution data is fed back to the logistics state continuous time sequence signal data acquisition link, and a management and control closed loop is formed.

[0081] The logistics management and control data of the target logistics entity are acquired and a logistics management and control data set is generated, including:

[0082] S1.1: Determine the boundary and identification of the target logistics entity, and assign a unique entity code to each target logistics entity; the target logistics entity is a unit with independent logistics operation function; the unit includes warehouse, transportation hub, distribution site; the entity code includes entity type, belonging area, serial number information;

[0083] S1.2: Collect logistics node coding data with hierarchical relationship, specifically including: extracting logistics node coding from the warehouse management system of the target logistics entity; the logistics node coding is constructed in five levels of entity-area-aisle-shelf-locus; extract the cargo classification coding, which is constructed based on cargo type, storage requirement and specification, and the cargo classification coding is associated with the locus coding through the locus-cargo binding table to record the current storage cargo type of each locus;

[0084] Further, when the logistics node coding is constructed in five levels of entity-area-aisle-shelf-locus, the five-level hierarchy forms a parent-child node relationship, for example, entity is the parent node of area, and area is the child node of entity; area is also the parent node of aisle, and aisle is the child node of area, and so on. Aisle-shelf-locus also follows the same parent-child level logic, that is, the upper level coding node is the parent node of the lower level coding node, and the lower level is the child node of the upper level.

[0085] S1.3: Collect logistics state continuous time series signal data in real time through Internet of Things devices deployed in the target logistics entity; the logistics state continuous time series signal data includes time series of environmental parameters, cargo throughput, location information and device operation parameters; the Internet of Things devices include GPS locator, cargo sensor, warehouse temperature and humidity monitor;

[0086] S1.4: Preprocess the logistics node coding data and logistics state continuous time series signal data, and associate the preprocessed logistics node coding data and logistics state continuous time series signal data through entity coding and locus coding to generate a structured logistics control data set; each row in the logistics control data set corresponds to the coding of a logistics node and the associated logistics state continuous time series signal data.

[0087] Further, the preprocessing of the logistics node coding data and the logistics state continuous time series signal data includes: performing format verification on the logistics node coding data, eliminating abnormal data with inconsistent coding bit number and invalid area identification, and converting the verified coding data into standard string format; the logistics state continuous time series signal data is processed by 3 The criterion eliminates abnormal values in the environmental parameters and cargo throughput, linear interpolation is used to complete the missing time series data, and then the data is mapped to the interval [0, 1] through Z-score standardization; the two types of preprocessed data are associated through entity coding and storage location coding to establish an associated index, ensuring that each row of the logistics control data set corresponds to the coding and associated time series signal number of a unique logistics node, wherein, 3 The criterion and Z-score standardization are prior art in the art and are not the inventive scheme of the present application, and will not be described here.

[0088] Further, because the logistics node coding is constructed in a five-level hierarchy of entity-area-aisle-shelf-storage location, it can be seen that the storage location coding is the last level in the five-level hierarchy of logistics node coding and is an inherent part of the logistics node coding system, which can be obtained by extracting the logistics node coding.

[0089] Further, the preprocessed logistics node coding data and the logistics state continuous time series signal data are associated through entity coding and storage location coding to generate a structured logistics control data set, including:

[0090] (1) Clearly associate the mapping relationship of the core identifier: first, extract the entity coding and storage location coding corresponding to each logistics node from the preprocessed logistics node coding data, wherein the entity coding is the unique identifier of the target logistics entity, such as the exclusive coding of the warehouse or transportation hub, and the storage location coding is the identifier of the specific storage location within the entity, such as the five-level coding of the area-aisle-shelf-storage location in the warehouse, and the one-to-one correspondence between each logistics node coding and entity coding and storage location coding needs to be confirmed one by one to ensure that each node has clear entity-storage location dual identification and avoid missing or mismatched nodes;

[0091] (2) The preprocessed logistics state continuous time series signal data is classified according to the source of collection, because the logistics state continuous time series signal data is collected by the Internet of Things devices deployed in the target logistics entity, each Internet of Things device is pre-bound with the entity coding of the corresponding entity, such as the temperature and humidity sensor of the warehouse being bound with the entity coding of the warehouse, and the collection location corresponding to the storage location coding, such as the goods recognition device on the shelf being bound with the storage location coding of the shelf, then, according to the binding information of the device, the corresponding entity coding and storage location coding are added to each preprocessed logistics state continuous time series signal data, so that the time series data originally containing only time and value is associated with the logistics node coding to obtain the logistics state continuous time series signal data with added entity coding and storage location coding;

[0092] (3) Create a temporary association index table with entity code + location code as the joint primary key. First, enter the logistics node code, entity code, and location code information in the logistics node code data into the index table one by one to form a mapping entry of the joint primary key-logistics node code. Then, group the logistics state continuous time series signal data with added entity code and location code according to the joint primary key. Match all logistics state continuous time series signal data with added entity code and location code under the same joint primary key with the corresponding logistics node code in the association index table to ensure that each joint primary key can be associated with a unique logistics node code and corresponding logistics state continuous time series signal data with added entity code and location code, thereby obtaining the matched data set;

[0093] (4) Check the matched data set to check if there are any abnormal situations where the joint primary key is only associated with the logistics node code but has no time series signal data or only has time series signal data but has no corresponding logistics node code. For the case where the joint primary key is only associated with the logistics node code but has no time series signal data, it is necessary to confirm whether the Internet of Things device is running normally. If the device is faulty, additional data collection or data missing markers are added. For the case where only time series signal data is available but no corresponding logistics node code is available, it is necessary to check whether the entity code and location code of the time series signal data are correct. If the code is incorrect, it is corrected and matched again. If the code is correct but there is no corresponding node code, it is determined as invalid data and removed to ensure that each group of associated data is complete and accurate, thereby obtaining the checked associated data;

[0094] (5) According to the rule that one row corresponds to one logistics node, the checked associated data is arranged in a structured format, i.e., the first column of each row of data is the logistics node code, and the subsequent columns are the entity code, location code, and associated logistics state continuous time series signal data of the node. At the same time, a data generation timestamp is added to each row of data to record the integration time of the node data. Finally, a structured logistics management data set is formed to ensure that the complete data of a single node can be directly located during subsequent model processing.

[0095] The generated logistics hierarchical feature mapping value includes:

[0096] A1: Obtain logistics node code data and perform preprocessing;

[0097] A2: Perform graph structure conversion on the preprocessed logistics node code data. In the graph structure conversion, each logistics node code is taken as a node of the graph, and the edges of the graph are constructed based on the hierarchical relationship of the logistics node codes;

[0098] Further, in the graph structure conversion, when the edges of the five-level hierarchy graph are constructed based on the logistics node encoding, the entity layer node and the area layer node, the area layer node and the lane layer node, the lane layer node and the shelf layer node, and the shelf layer node and the storage location layer node are respectively established as parent-child edges, and the weight of the edge is set according to the frequency of the goods flow between the upper and lower level nodes, and the higher the flow frequency, the greater the weight, and the weight value is used for feature aggregation of the graph convolution layer after normalization processing.

[0099] A3: initializing a logistics hierarchical encoding embedding model; the logistics hierarchical encoding embedding model adopts a hyperbolic graph convolution network structure, and the input is the constructed graph structure data, and the output is a fixed dimension feature vector; the logistics hierarchical encoding embedding model comprises a graph convolution layer, a hyperbolic space mapping layer, and a feature optimization layer, wherein the graph convolution layer is used to aggregate neighbor features of nodes, the hyperbolic space mapping layer is used to map features to a hyperbolic space, and the feature optimization layer is used to improve the discrimination of feature mapping values, and the hyperbolic graph convolution network structure is prior art content in the field and is not the inventive scheme of the present application, and will not be described here;

[0100] A4: inputting the graph structure data into the initialized logistics hierarchical encoding embedding model for training to obtain a trained logistics hierarchical encoding embedding model;

[0101] A5: inputting the real-time preprocessed logistics node encoding data into the trained logistics hierarchical encoding embedding model to output an N-dimensional vector corresponding to each logistics node encoding, that is, a logistics hierarchical feature mapping value; the logistics hierarchical feature mapping value is located in a hyperbolic space and contains hierarchical association of logistics nodes, goods attributes, and functional area information.

[0102] Further, the specific steps of A5 include:

[0103] (1) confirming the preprocessing integrity of the real-time logistics node encoding data: before inputting the real-time logistics node encoding data into the model, it is checked whether the real-time collected logistics node encoding data has completed the preprocessing described above, including format verification and conversion to a standard string format, and it is confirmed that each encoding can be split into five-level hierarchical information of entity-area-lane-shelf-storage location, and there is no missing or mismatched hierarchical field;

[0104] (2) converting the encoding data in the standard string format into an input format recognizable by the trained logistics hierarchical encoding embedding model, wherein the trained logistics hierarchical encoding embedding model needs to receive graph structure data as input, so it is necessary to convert the logistics node encoding data in the standard string format into graph structure data according to the step A2 to ensure that the model can recognize the hierarchical association between nodes;

[0105] (3) Normalizing and dimensionally adapting the graph structure data, wherein the dimensionally adapting refers to checking the feature dimension of the nodes in the graph structure, such as the number of hierarchical fields after encoding splitting, and if the feature dimension of the real-time data does not match the input dimension during model training, the feature completion or dimension compression, such as merging similar hierarchical fields, is needed to adjust to the adaptive dimension, so as to ensure that the data dimension input into the model is consistent with that during training, and to avoid errors of the model due to dimension mismatch;

[0106] (4) Inputting the adapted graph structure data into the trained logistics hierarchical coding embedding model and starting feature extraction, specifically including: the trained logistics hierarchical coding embedding model first aggregates the neighbor node features of each node by the graph convolution layer, wherein the aggregation combines the edge weight and the distance of the node in the hierarchy to generate the preliminary hierarchical aggregation features of each node; then through the hyperbolic space mapping layer, the preliminary aggregation features are mapped from Euclidean space to hyperbolic space by using the Euclidean-hyperbolic conversion function to obtain the preliminary hierarchical feature vector in the hyperbolic space; finally, through the feature optimization layer, the preliminary hierarchical feature vector is fine-tuned combined with the hierarchical loss and the cargo type loss optimization rules determined during training to improve the discrimination of the features and ensure that the feature vector can accurately reflect the hierarchical association of the node;

[0107] (5) Extracting the N-dimensional vector output by the trained logistics hierarchical coding embedding model and verifying the feature effectiveness, including: after the logistics hierarchical coding embedding model is processed, the N-dimensional feature vector corresponding to each logistics node coding, i.e., the logistics hierarchical feature mapping value, is output, which needs to be checked first whether the dimension of the feature vector is the preset N-dimensional, and if the dimension is abnormal, the data processing is re-input; then the feature similarity check is performed, i.e., the similarity of the feature vectors of the nodes in the same hierarchy is calculated, and if the similarity meets the expectation, i.e., the similarity of the nodes in the same hierarchy is higher than that of the nodes in different hierarchies, it is determined that the feature vector is effective; if the similarity is abnormal, it is analyzed whether there is a problem with the model input data or parameters, and after adjustment, the processing is reprocessed, and finally the effective logistics hierarchical feature mapping value is determined for fusion operation with the logistics time sequence feature mapping value.

[0108] The inputting of the graph structure data into the initialized logistics hierarchical coding embedding model includes:

[0109] A4.1: Weighted aggregation of all neighbor node features of each node by the graph convolution layer to obtain preliminary hierarchical aggregation features, wherein the weight calculation combines the edge weight and the distance of the node in the hierarchy during the weighted aggregation process;

[0110] A4.2: Through the hyperbolic space mapping layer, the preliminary hierarchical aggregation features are mapped from the Euclidean space to the hyperbolic space by using a Euclidean-hyperbolic conversion function to obtain the preliminary hierarchical feature vector in the hyperbolic space; the hyperbolic space is constructed by using a Poincare ball model, and the distance between any two points in the ball is calculated by a Poincare distance formula; the calculation method of the Euclidean-hyperbolic conversion function is as follows: first, the module length of the Euclidean space vector v is calculated, then the hyperbolic sine value of the module length is calculated, then the hyperbolic sine value is multiplied by v, and finally the obtained result is divided by the module length of v;

[0111] Further, the Poincare distance is used to calculate the hierarchical loss of the feature vectors of the parent node and the child node, and the Poincare distance formula is prior art content in the field and is not the inventive scheme of the present application, and will not be described here.

[0112] Further, the calculation process of the hierarchical loss is as follows: for each child node, the hyperbolic space feature vector of its parent node is extracted, the distance between the two is calculated by the Poincare distance formula, and the mean of the distance is taken as the hierarchical loss value; the calculation process of the cargo type loss is as follows: for all the nodes corresponding to the same cargo classification code, the cosine similarity of the hyperbolic space feature vectors of the nodes is calculated, and the mean of the similarity is subtracted from 1 to obtain the cargo type loss value; during model training, the weighted sum of the hierarchical loss and the cargo type loss is minimized by the gradient descent method until the loss value converges to a preset threshold, and the preset threshold in the present application is set to 0.01, wherein the gradient descent method is prior art content in the field and is not the inventive scheme of the present application, and will not be described here.

[0113] A4.3: Through the feature optimization layer, the preliminary hierarchical feature vector is optimized by minimizing two types of losses; the two types of losses include a hierarchical loss and a cargo type loss; the hierarchical loss is obtained by calculating the Poincare distance of the feature vectors of the parent node and the child node; the cargo type loss is obtained by calculating the similarity of the feature vectors of all the nodes corresponding to the same cargo classification code.

[0114] The generated logistics time sequence feature mapping value includes:

[0115] B1: Obtain logistics state continuous time sequence signal data and perform preprocessing;

[0116] B2: Time sequence segmentation and standardization are performed on the preprocessed logistics state continuous time sequence signal data to obtain a standardized logistics state continuous time sequence;

[0117] Further, the specific process of the time sequence segmentation and standardization is: the logistics state continuous time sequence signal data is segmented into time sequence segments with each 30 minutes as a time window, and each segment contains 30 time step signal data; for each time sequence segment, the mean and standard deviation of each signal dimension are calculated; the logistics state continuous time sequence is standardized according to the mean and standard deviation to obtain a standardized logistics state continuous time sequence, which is used as input of the logistics time sequence signal embedding model, wherein the mean and standard deviation are prior art contents in the field and are not the creative schemes of the present application, and are not described herein.

[0118] B3: initializing a logistics time sequence signal embedding model; the logistics time sequence signal embedding model adopts a bidirectional long short-term memory network structure, and the input is the preprocessed logistics state continuous time sequence signal data, and the output is a fixed-dimensional feature vector; the logistics time sequence signal embedding model comprises a forward LSTM layer, a backward LSTM layer, a bidirectional feature fusion layer and an attention layer, wherein the forward LSTM layer captures forward dependence of the time sequence, the backward LSTM layer captures reverse dependence, the bidirectional feature fusion layer integrates bidirectional information, and the attention layer highlights key time sequence features, and the long short-term memory network structure is a prior art content in the field and is not the creative scheme of the present application, and is not described herein;

[0119] B4: inputting the standardized logistics state continuous time sequence into the initialized logistics time sequence signal embedding model to obtain N-dimensional logistics time sequence feature mapping values.

[0120] The inputting of the standardized logistics state continuous time sequence into the initialized logistics time sequence signal embedding model to obtain N-dimensional logistics time sequence feature mapping values comprises:

[0121] B4.1: the forward LSTM layer starts iteration from the starting time step of the time sequence, combines the current time step input and the previous step hidden state at each step, outputs the forward time sequence feature state of the current time step, and records the forward dynamic feature of the time sequence;

[0122] B4.2: the backward LSTM layer reverses the time sequence, starts iteration from the starting time step of the reversed sequence, combines the current time step input and the previous step hidden state at each step, outputs the reverse time sequence feature state of the current time step, and records the reverse dynamic feature of the time sequence;

[0123] B4.3: the bidirectional feature fusion layer adopts a vector splicing and linear transformation manner to combine the forward time sequence feature state and the reverse time sequence feature state into an M-dimensional bidirectional feature vector, and then enhances the nonlinear expression ability of the feature through a ReLU activation function to obtain a preliminary time sequence feature vector, wherein the ReLU activation function is a prior art content in the field and is not the creative scheme of the present application, and is not described herein;

[0124] B4.4: The attention layer assigns weights according to the importance of the logistics state continuous time sequence signal data, and obtains an attention weight;

[0125] Further, the calculation process of the attention weight is: inputting the preliminary time sequence feature vector into a single hidden layer perceptron, the number of hidden units of which is 64, and the activation function is ReLU, to obtain an evaluation score vector; the score vector is normalized by using a softmax function to obtain the attention weight of each time step; when the weight is assigned, the time steps with high cargo throughput fluctuation and equipment operating parameters exceeding the threshold value correspond to higher weights, so as to ensure that the logistics time sequence feature mapping value highlights the key time sequence information.

[0126] B4.5: Multiply the preliminary time sequence feature vector and the attention weight element by element to obtain an N-dimensional logistics time sequence feature mapping value.

[0127] The logistics node fusion feature set is generated, including:

[0128] S3.1: A hyperbolic-Euclidean conversion function is used to project the logistics level feature mapping value in the hyperbolic space to the origin of the Euclidean space to generate a Euclidean space feature mapping value; the calculation method of the hyperbolic-Euclidean conversion function is: first, calculate the modulus of the hyperbolic space vector u, then calculate the inverse hyperbolic tangent value of the modulus, then multiply the inverse hyperbolic tangent value by u, and finally divide the result by the modulus of u;

[0129] S3.2: Map the logistics time sequence feature mapping value to the same dimension and distribution space as the Euclidean space feature mapping value through a linear transformation layer to generate an adaptive time sequence feature mapping value;

[0130] Further, the specific steps of S3.2 include:

[0131] (1) Extract the attribute information from the generated Euclidean space feature mapping value; the attribute information includes the number of target dimensions and distribution characteristics; the number of target dimensions is the number of feature vector elements contained in each Euclidean space feature mapping value; the distribution characteristics include mean, standard deviation, maximum value and minimum value;

[0132] (2) initialize parameters of the linear transformation layer; the parameters of the linear transformation layer include a weight matrix and a bias term, and need to be initialized based on a target dimension and distribution characteristics, wherein the number of rows of the weight matrix is set as the dimension of the Euclidean space feature mapping value, and the number of columns is set as the dimension of the logistics time sequence feature mapping value, to ensure that the logistics time sequence feature mapping value can be converted to the target dimension after matrix multiplication; the initial value of the weight matrix uses the Xavier initialization method, that is, a random initial value is calculated according to the input and output dimensions; the dimension of the bias term is consistent with the target dimension, and the initial value is set as the mean value of the Euclidean space feature mapping value;

[0133] (3) check whether the dimension of the logistics time sequence feature mapping value matches the number of columns of the weight matrix, if not, adjust through feature pruning or feature expansion to obtain the logistics time sequence feature mapping value after dimension verification;

[0134] (4) perform matrix multiplication operation on the logistics time sequence feature mapping value after dimension verification and the initialized weight matrix to obtain a preliminary transformed feature vector, and then perform addition operation on the preliminary feature vector and the bias term to complete linear transformation and generate a preliminary adaptive time sequence feature mapping value;

[0135] (5) check the dimension of the generated preliminary adaptive time sequence feature mapping value, confirm whether it is completely consistent with the dimension of the Euclidean space feature mapping value, if there is a dimension deviation, adjust the dimension of the weight matrix again and calculate again, and compare the distribution characteristics with the distribution characteristics of the Euclidean space feature mapping value, calculate the mean difference and the standard deviation, if the mean difference exceeds 5% of the mean value of the Euclidean space feature, or the standard deviation exceeds 10% of the standard deviation of the Euclidean space feature, adjust the value of the bias term, repeat the linear transformation process until the dimension and distribution of the adaptive time sequence feature mapping value meet the preset requirements;

[0136] (6) when the dimension and distribution of the adaptive time sequence feature mapping value meet the requirements, lock the weight matrix and bias term parameters of the linear transformation layer to avoid parameter changes in subsequent operations, and determine the feature vector generated at this time through linear transformation as the final adaptive time sequence feature mapping value, and add dimension identifier and distribution verification label to the final adaptive time sequence feature mapping value; the distribution verification label includes target dimension, mean value and standard deviation.

[0137] S3.3: construct a fusion weight predictor to output hierarchical feature weights and time sequence feature weights; the fusion weight predictor is constructed based on a logistic regression model;

[0138] Further, the training process of the fusion weight predictor is: taking the logistics level feature mapping value, the logistics time sequence feature mapping value and the corresponding scheduling effect evaluation result in the historical data as the training sample; inputting the Euclidean space feature mapping value and the adaptive time sequence feature mapping value into a logistic regression model, taking the scheduling effect evaluation result as the label, training the model to learn the level feature weight; after the training is completed, inputting the real-time Euclidean space feature mapping value and the adaptive time sequence feature mapping value, and outputting the corresponding weight for weighted summation to generate the logistics feature fusion vector.

[0139] S3.4: The Euclidean space feature mapping value and the adaptive time sequence feature mapping value are respectively weighted and summed according to the level feature weight and the time sequence feature weight to generate a logistics feature fusion vector; the logistics feature fusion vector is N-dimensional.

[0140] It should be understood that the logistics feature fusion vector is generated by the logistics time sequence feature mapping value and the logistics level feature mapping value through the fusion operation, and in the present application, the fusion operation selects element-level addition, and the premise is that the input logistics time sequence feature mapping value and the logistics level feature mapping value are consistent in dimension, so the logistics feature fusion vector generated after fusion is also N-dimensional.

[0141] S3.5: A Euclidean-hyperbolic conversion function is used to map the logistics feature fusion vector from the Euclidean space back to the hyperbolic space with the hyperbolic space origin as the benchmark to generate an alignment feature vector.

[0142] Further, the specific steps of S3.5 include:

[0143] (1) Obtain the logistics feature fusion vector and perform preprocessing;

[0144] (2) Define the basic model and origin benchmark of the hyperbolic space;

[0145] Since the mapping needs to be based on the hyperbolic space origin, the specific model of the hyperbolic space needs to be determined first, and the present application uses the Poincare ball model, which is a common representation form of the hyperbolic space, with the origin as the center of the ball and the radius parameter set to 1. The coordinates of the hyperbolic space origin are defined as zero vectors with the same dimension as the logistics feature fusion vector, i.e. the element value of each dimension is 0, as the position benchmark in the mapping process.

[0146] (3) According to the mathematical properties of the Poincare ball model, derive the Euclidean-hyperbolic conversion function that adapts to the current logistics feature fusion vector: the function needs to satisfy the mapping relationship of Euclidean space vector input-hyperbolic space vector output, and the output vector needs to be located within the Poincare ball, i.e. the modulus is less than 1;

[0147] (4) Verify the effectiveness of the derived Euclidean-hyperbolic conversion function by test vectors, including: selecting multiple known Euclidean space vectors, such as unit vectors and zero vectors, substituting them into the Euclidean-hyperbolic conversion function to calculate hyperbolic space vectors, checking whether the hyperbolic space vectors have a length less than 1 and can accurately reflect the relative position relationship of the original vectors, such as the larger the original vector length, the farther the converted hyperbolic space vector from the origin, if the verification fails, adjusting the function parameters, such as modifying the length calculation coefficient, until the function meets the space mapping requirements;

[0148] (5) Sum the square values of each element in the preprocessed logistics feature fusion vector, and then take the square root of the sum to obtain the Euclidean length of the logistics feature fusion vector;

[0149] (6) Substitute the preprocessed logistics feature fusion vector into the verified Euclidean-hyperbolic conversion function, calculate the hyperbolic sine value of the vector length, multiply the hyperbolic sine value with each element of the logistics feature fusion vector respectively to obtain an intermediate vector, and divide each element of the intermediate vector by the length of the preprocessed logistics feature fusion vector to obtain a preliminary hyperbolic space vector, that is, a preliminary alignment feature vector;

[0150] (7) Check the spatial legality and accuracy of the preliminary alignment feature vector, and finally generate an alignment feature vector that meets the spatial requirements and accurately reflects the relationship of the original vector.

[0151] Further, the spatial legality verification includes: calculating the hyperbolic length of the preliminary alignment feature vector, checking whether it is less than the Poincare ball radius, if it is greater than it, adjusting the length coefficient in the conversion function, recalculating until it meets the requirements; The hyperbolic length is calculated based on the distance formula of the Poincare ball model to realize the distance from the vector to the origin.

[0152] Further, the accuracy verification includes: selecting two logistics feature fusion vectors with known relative relationship, such as the length of A vector is twice the length of B vector, checking whether the distance from their converted hyperbolic space vectors to the origin meets the expectation, that is, the hyperbolic distance of A vector should be greater than B vector, and the proportion should be consistent with the original length proportion, if it does not meet the requirement, modify the parameters of the conversion function, such as adjusting the base of the hyperbolic function, finally generate an alignment feature vector that meets the spatial requirements and accurately reflects the relationship of the original vector.

[0153] S3.6: Collect all the alignment feature vectors corresponding to the logistics nodes, store them in the format of logistics node code-alignment feature vector-generation timestamp, and generate a logistics node fusion feature set.

[0154] Embodiment 2

[0155] Please refer to Figure 3 , the present application provides another embodiment: an intelligent logistics integrated management system, comprising:

[0156] The data acquisition module, the feature mapping module, the feature fusion module, the scheduling evaluation module, the node scheduling module, and the scheme matching module;

[0157] The data acquisition module is configured to acquire various types of management and control data of a target logistics entity, and generate a structured logistics management and control data set after processing.

[0158] The feature mapping module is configured to perform feature extraction on the logistics node coding data and the logistics state time sequence signal data, and generate feature mapping values reflecting the node hierarchical attributes and time sequence dynamics.

[0159] The feature fusion module is configured to fuse the logistics hierarchical feature mapping values and the logistics time sequence feature mapping values, generate an aligned feature vector with both hierarchical attributes and dynamic characteristics, and form a logistics node fusion feature set.

[0160] The scheduling evaluation module is configured to calculate scheduling indicators based on the logistics node fusion feature set through a federated hybrid model, and evaluate the node scheduling priority.

[0161] The node scheduling module is configured to perform scheduling operations on a target scheduling node, track job completion status and update node state, and form a scheduling closed loop.

[0162] The scheme matching module is configured to match or generate an adaptive management and control path scheme according to the type of a target management and control node, and realize fine management and control of the node.

[0163] The feature mapping module includes an encoding embedding unit and a signal embedding unit.

[0164] The encoding embedding unit is configured to load a logistics hierarchical coding embedding model, convert the preprocessed logistics node coding data into a graph structure, aggregate neighbor features through a graph convolution layer, convert spatial dimensions through a hyperbolic space mapping layer, improve discrimination through a feature optimization layer, and finally output logistics hierarchical feature mapping values in a hyperbolic space, retaining the hierarchical association and attribute information of the node.

[0165] The signal embedding unit is configured to load a logistics time sequence signal embedding model, perform time sequence segmentation and feature extraction on the standardized time sequence signal data, capture bidirectional time sequence dependencies through a forward / backward LSTM layer, integrate information through a bidirectional feature fusion layer, highlight key time sequence features through an attention layer, and generate logistics time sequence feature mapping values reflecting dynamic states.

[0166] The feature fusion module includes a space conversion unit, a weight fusion unit, and a feature construction unit.

[0167] a space conversion unit, configured to project the hierarchical feature mapping value in a hyperbolic space to a Euclidean space through a hyperbolic-Euclidean conversion function, and adjust the time sequence feature mapping value to a matching dimension and distribution through a linear transformation layer, so that the two types of features can be fused in the same space dimension;

[0168] a weight fusion unit configured to construct a fusion weight predictor based on a logistic regression model, output fusion weights of the hierarchical features and the time sequence features, perform weighted summation on the two types of features according to the weights, generate a logistics feature fusion vector, and map the fusion vector back to the hyperbolic space through a Euclidean-hyperbolic conversion function to obtain an aligned feature vector;

[0169] a feature construction unit configured to collect the aligned feature vectors of all logistics nodes in a node code-feature vector-time stamp format, and store the aligned feature vectors to form a logistics node fusion feature set.

[0170] The scheduling evaluation module includes a feature processing unit, a federal aggregation unit and a priority determination unit.

[0171] The feature processing unit is configured to deploy a local XGBoost model at each logistics node, process local sub-features in the fusion feature set, and output local scheduling features containing node load, cargo backlog and other information, so as to ensure local data privacy and security.

[0172] The federal aggregation unit is configured to receive the local scheduling features transmitted by each node in an encrypted manner, aggregate global features through a federal average algorithm, generate a scheduling index set containing node cargo backlog rate, device utilization rate, order overtime rate and transportation cost coefficient, and realize global evaluation across nodes.

[0173] The priority determination unit is configured to convert the scheduling index set into a scheduling priority value, compare the scheduling priority value with a preset scheduling threshold, and select a target scheduling node that meets the threshold condition.

[0174] The node scheduling module includes an instruction generation unit, a node registration unit and a state synchronization unit.

[0175] The instruction generation unit is configured to generate a scheduling instruction containing a to-be-processed order list, a resource allocation scheme and a job time window for the target scheduling node, and send the scheduling instruction to a node management terminal through an encrypted channel, so as to ensure accurate transmission of the instruction.

[0176] The node registration unit is configured to monitor the job progress of the target scheduling node in real time, register the target scheduling node as a target management node after the target scheduling node completes the scheduling job, and record task completion records such as job start / completion time and processed cargo quantity.

[0177] A state synchronization unit is configured to collect real-time device states of the target management node from the Internet of Things device, and store the states in the database synchronously with the task completion record, so as to ensure the timeliness and completeness of the node state data.

[0178] The scheme matching module comprises an operation analysis unit, a path searching unit, a scheme generating unit and a scheme assigning unit.

[0179] The operation analysis unit is configured to analyze operation indexes of the target management node based on warehouse control system logs, interface records and other data, and determine the node type.

[0180] The path searching unit is configured to search a matched standard management path scheme from a preset management path library according to the node type, the current load and the task characteristics, wherein the scheme comprises a work flow, resource allocation and an exception handling mechanism.

[0181] The scheme generating unit is configured to generate a new scheme by using a genetic algorithm if there is no matched scheme in the library, and provide the new scheme and the searched standard scheme to the scheme assigning unit after verifying the feasibility of the new scheme by simulation test.

[0182] The scheme assigning unit is configured to assign the matched or generated management path scheme to the target management node, monitor the execution progress of the scheme in real time, and feed back the execution data to the data collection link to form a complete management closed loop.

[0183] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above-described specific embodiments, and the above-described specific embodiments are only illustrative but not restrictive, and the person skilled in the art can make changes, modifications, replacements and variations to the above-described embodiments without departing from the purpose of the present application and the protected scope, and these are all within the protection of the present application.

Claims

1. An intelligent logistics integrated management method, characterized in that, The method comprises the following steps: acquiring logistics management data of a target logistics entity and generating a logistics management data set; the logistics management data set comprises logistics node coding data and logistics state continuous time sequence signal data; processing the logistics node coding data through a logistics level coding embedding model to generate logistics level feature mapping values, and simultaneously processing the logistics state continuous time sequence signal data through a logistics time sequence signal embedding model to generate logistics time sequence feature mapping values; performing a fusion operation on the logistics level feature mapping values and the logistics time sequence feature mapping values to generate a logistics node fusion feature set; calling a federated hybrid model to process the logistics node fusion feature set to generate a scheduling index set representing scheduling priorities of target logistics nodes; if the scheduling priorities in the scheduling index set meet a preset scheduling threshold, a corresponding logistics node is determined as a target scheduling node and a scheduling instruction is sent to the target scheduling node, and after the target scheduling node completes a scheduling task, the target scheduling node is registered as a target management node and a task completion record and a device state of the target management node are updated synchronously; according to a logistics node type indicated by an operation analysis result of the target management node, a management path scheme is matched or generated from a preset management path library and is allocated to the target management node. 2.The intelligent logistics integrated management method of claim 1, wherein, The method of acquiring logistics management data of a target logistics entity and generating a logistics management data set comprises the following steps: determining boundaries and identities of target logistics entities and allocating unique entity codes to each target logistics entity; the target logistics entity is a unit with independent logistics operation functions; the unit comprises a warehouse, a transportation hub and a distribution site; the entity code comprises entity type, belonging area and serial number information; collecting logistics node coding data with hierarchical relationships, which comprises extracting logistics node coding from a warehouse management system of a target logistics entity; the logistics node coding is constructed in a five-level hierarchy of entity-area-passage-shelf-locus; cargo classification coding is extracted, which is constructed based on cargo type, storage requirements and specifications, and the cargo classification coding is associated with the locus coding through a locus-cargo binding table to record the current storage cargo type of each locus; collecting logistics state continuous time sequence signal data in real time through Internet of Things devices deployed in a target logistics entity; the logistics state continuous time sequence signal data comprises time sequences of environmental parameters, cargo throughput, location information and device operation parameters; preprocessing the logistics node coding data and the logistics state continuous time sequence signal data, and associating the preprocessed logistics node coding data and the logistics state continuous time sequence signal data through entity coding and locus coding to generate a structured logistics management data set; each row in the logistics management data set corresponds to the coding of a logistics node and associated logistics state continuous time sequence signal data. 3.The intelligent logistics integrated management method of claim 2, wherein, The method of generating logistics level feature mapping values comprises the following steps: acquiring logistics node coding data and preprocessing it; performing graph structure conversion on the preprocessed logistics node coding data; in the graph structure conversion, each logistics node coding is taken as a node of a graph, and edges of the graph are constructed based on the hierarchical relationships of the logistics node coding; The initialization of the logistics hierarchical coding embedding model; the logistics hierarchical coding embedding model adopts a hyperbolic graph convolutional network structure, the input is the constructed graph structure data, and the output is a fixed dimension feature vector; the logistics hierarchical coding embedding model includes a graph convolution layer, a hyperbolic space mapping layer, and a feature optimization layer, wherein the graph convolution layer is used to aggregate the neighbor features of the nodes, the hyperbolic space mapping layer is used to map the features to the hyperbolic space, and the feature optimization layer is used to improve the discrimination of the feature mapping values; The graph structure data is input into the initialized logistics hierarchical coding embedding model for training to obtain a trained logistics hierarchical coding embedding model; The real-time preprocessed logistics node coding data is input into the trained logistics hierarchical coding embedding model, and an N-dimensional vector corresponding to each logistics node coding is output, which is a logistics hierarchical feature mapping value; the logistics hierarchical feature mapping value is located in the hyperbolic space and includes the hierarchical correlation of the logistics node, the cargo attribute, and the functional area information. 4.The intelligent logistics integrated management method of claim 3, wherein, The input of the graph structure data into the initialized logistics hierarchical coding embedding model includes: The graph convolution layer is used to weight and aggregate all neighbor node features of each node to obtain preliminary hierarchical aggregation features, wherein the weight calculation combines the edge weight and the distance of the node in the hierarchy during the weight aggregation process; The hyperbolic space mapping layer is used to map the preliminary hierarchical aggregation features from the Euclidean space to the hyperbolic space by using a Euclidean-hyperbolic conversion function to obtain a preliminary hierarchical feature vector in the hyperbolic space; the hyperbolic space is constructed by using a Poincare ball model, and the distance between any two points in the ball is calculated by using a Poincare distance formula; the calculation method of the Euclidean-hyperbolic conversion function is as follows: first, the modulus of the Euclidean space vector v is calculated, then the hyperbolic sine value of the modulus is calculated, then the hyperbolic sine value is multiplied by v, and finally the result is divided by the modulus of v; The feature optimization layer is used to minimize two types of losses to optimize the preliminary hierarchical feature vector; the two types of losses include a hierarchical loss and a cargo type loss; the hierarchical loss is obtained by calculating the Poincare distance between the feature vectors of the parent node and the child node; and the cargo type loss is obtained by calculating the similarity of the feature vectors of all storage location nodes corresponding to the same cargo classification code. 5.The intelligent logistics integrated management method of claim 4, wherein, The generation of the logistics time sequence feature mapping value includes: Obtaining logistics state continuous time sequence signal data and preprocessing the same; Performing time sequence segmentation and standardization on the preprocessed logistics state continuous time sequence signal data to obtain standardized logistics state continuous time sequence data; The initialization of the logistics time sequence signal embedding model; the logistics time sequence signal embedding model adopts a bidirectional long short-term memory network structure, the input is the preprocessed logistics state continuous time sequence signal data, and the output is a fixed dimension feature vector; the logistics time sequence signal embedding model includes a forward LSTM layer, a backward LSTM layer, a bidirectional feature fusion layer, and an attention layer, wherein the forward LSTM layer captures the forward dependence of the time sequence, the backward LSTM layer captures the reverse dependence, the bidirectional feature fusion layer integrates the bidirectional information, and the attention layer highlights the key time sequence features; The standardized logistics state continuous time sequence is input into the initialized logistics time signal embedding model to obtain N-dimensional logistics time sequence feature mapping values. 6.The intelligent logistics integrated management method of claim 5, wherein, The standardized logistics state continuous time sequence is input into the initialized logistics time signal embedding model to obtain N-dimensional logistics time sequence feature mapping values, which includes: The forward LSTM layer starts iteration from the starting time step of the logistics state continuous time sequence, combines the current time step input and the previous hidden state at each step, outputs the forward time sequence feature state of the current time step, and records the forward dynamic features of the logistics state continuous time sequence; The backward LSTM layer reverses the logistics state continuous time sequence and starts iteration from the starting time step of the reversed sequence, combines the current time step input and the previous hidden state at each step, outputs the reverse time sequence feature state of the current time step, and records the reverse dynamic features of the logistics state continuous time sequence; The bidirectional feature fusion layer adopts a vector splicing and linear transformation manner to combine the forward time sequence feature state and the reverse time sequence feature state into an M-dimensional bidirectional feature vector, and then enhances the non-linear expression ability of the features through a ReLU activation function to obtain a preliminary time sequence feature vector; The attention layer allocates weights according to the importance of the logistics state continuous time signal data to obtain attention weights; The preliminary time sequence feature vector and the attention weights are multiplied element by element to obtain N-dimensional logistics time sequence feature mapping values. 7.The intelligent logistics integrated management method of claim 6, wherein, The logistics node fusion feature set is generated, which includes: A hyperbolic-Euclidean conversion function is used to project the logistics level feature mapping values in the hyperbolic space to the origin of the Euclidean space to generate Euclidean space feature mapping values; the calculation method of the hyperbolic-Euclidean conversion function is: first, calculate the modulus of the hyperbolic space vector u, then calculate the inverse hyperbolic tangent value of the modulus, then multiply the inverse hyperbolic tangent value by u, and finally divide the result by the modulus of u; The logistics time sequence feature mapping values are mapped to the same dimension and distribution space as the Euclidean space feature mapping values through a linear transformation layer to generate adaptive time sequence feature mapping values; A fusion weight predictor is constructed to output level feature weights and time sequence feature weights; the fusion weight predictor is constructed based on a logistic regression model; The Euclidean space feature mapping values and the adaptive time sequence feature mapping values are weighted and summed according to the level feature weights and the time sequence feature weights, respectively, to generate a logistics feature fusion vector; the logistics feature fusion vector is N-dimensional; A Euclidean-hyperbolic conversion function is used to map the logistics feature fusion vector from the Euclidean space back to the hyperbolic space based on the origin of the hyperbolic space to generate an aligned feature vector; All logistics node corresponding aligned feature vectors are collected and stored in the format of logistics node code-aligned feature vector-generation timestamp to generate a logistics node fusion feature set. 8.The intelligent logistics integrated management method of claim 7, wherein, The federal hybrid model is composed of a local XGBoost model of each logistics node and a cloud federated aggregation layer, the local XGBoost model first processes the local sub-features in the logistics node fusion feature set, and outputs the local scheduling features; the local scheduling features are encrypted and transmitted to the cloud federated aggregation layer, the cloud federated aggregation layer aggregates global features through a federated average algorithm, and finally generates a scheduling index set, the scheduling index set includes node cargo backlog rate, equipment utilization rate, order overtime rate and transportation cost coefficient. 9.The intelligent logistics integrated management method of claim 8, wherein, The operation analysis result is generated based on the application module log, system operation log of the warehouse control system log management module and the extended warehouse management interface record of the interface log management module; the operation analysis result includes the temperature and humidity compliance rate of the warehouse node, the equipment failure rate of the transportation node and the task delay rate of the distribution node.

10. An intelligent logistics integrated management system for implementing the intelligent logistics integrated management method according to any one of claims 1-9, characterized in that, Comprise: a data acquisition module, a feature mapping module, a feature fusion module, a scheduling evaluation module, a node scheduling module and a scheme matching module; The data acquisition module is used for acquiring various control data of the target logistics entity, and generates a structured logistics control data set after processing; The feature mapping module is used for feature extraction of logistics node coding data and logistics state time sequence signal data, and generates feature mapping values reflecting node hierarchical attributes and time sequence dynamics; The feature fusion module is used for fusing logistics hierarchical feature mapping values and logistics time sequence feature mapping values, generating aligned feature vectors, and forming a logistics node fusion feature set; The scheduling evaluation module calculates the scheduling index based on the logistics node fusion feature set through the federal hybrid model, and evaluates the node scheduling priority; The node scheduling module is used for executing scheduling operation on the target scheduling node, tracking job completion and updating node state; The scheme matching module is used for matching or generating an adaptive control path scheme according to the type of the target control node.

Images