Cross-border e-commerce logistics intelligent matching method and device and storage medium

Abstract

The invention relates to a cross-border e-commerce logistics intelligent matching method and device and a storage medium. The method comprises the following steps: preprocessing order data, transport capacity data and distribution record data collected by logistics nodes to generate a logistics data set; inputting the logistics data set into a double-flow feature extraction network for feature parallel calculation, and generating a logistics time-space incidence matrix; performing feature decomposition and feature recombination on the logistics space-time incidence matrix to obtain a logistics dynamic feature sequence; inputting the logistics dynamic feature sequence into a mixed attention network for node correlation weight analysis, and generating a logistics matching feature vector; constructing a multi-objective optimization function based on the logistics matching feature vector, and solving to obtain a logistics resource allocation scheme; and decomposing and dynamically adjusting the logistics resource allocation scheme, and outputting a target matching result. According to the invention, the intelligent level and the operation efficiency of cross-border logistics distribution are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of logistics intelligent matching, and particularly relates to a cross-border e-commerce logistics intelligent matching method, device, and storage medium. Background Art

[0002] With the rapid development of cross-border e-commerce, the logistics distribution demand shows characteristics such as a large number of orders, a wide distribution range, and high timeliness requirements. The traditional manual matching method has been difficult to meet the efficient and accurate logistics distribution demand. Currently, cross-border e-commerce logistics faces problems such as low matching efficiency between orders and transport capacity resources, high distribution costs, and difficult guarantee of timeliness, and there is an urgent need for an intelligent logistics matching solution.

[0003] The existing logistics matching methods mainly have three problems: the spatio-temporal characteristics of logistics data are not fully extracted, making it difficult to effectively capture the dynamic change rules of order distribution and transport capacity resources; the comprehensive consideration of multi-dimensional constraint conditions is lacking, resulting in the matching results being difficult to meet the requirements of timeliness and economy simultaneously; the existing methods have insufficient response to the dynamic changes during the logistics execution process and cannot adjust the matching strategy in a timely manner. Summary of the Invention

[0004] The main purpose of the present invention is to provide a cross-border e-commerce logistics intelligent matching method, device, and storage medium, thereby improving the intelligent level and operation efficiency of cross-border logistics distribution.

[0005] To achieve the above purpose, the present invention provides a cross-border e-commerce logistics intelligent matching method, including the following steps: Preprocess the order data, transport capacity data, and distribution record data collected by logistics nodes to generate a logistics data set; Input the logistics data set into a dual-stream feature extraction network for feature parallel calculation to generate a logistics spatio-temporal correlation matrix; Perform feature decomposition and feature recombination on the logistics spatio-temporal correlation matrix to obtain a logistics dynamic feature sequence; Input the logistics dynamic feature sequence into a hybrid attention network for node correlation weight analysis to generate a logistics matching feature vector; Construct a multi-objective optimization function based on the logistics matching feature vector and solve to obtain a logistics resource allocation plan; Decompose and dynamically adjust the logistics resource allocation plan, and output a target matching result.

[0006] The present invention also provides a cross-border e-commerce logistics intelligent matching device, including: A preprocessing module, configured to preprocess the order data, transport capacity data, and distribution record data collected by logistics nodes to generate a logistics data set; A computing module, configured to input the logistics data set into a dual-stream feature extraction network for parallel feature computing to generate a logistics spatio-temporal correlation matrix; A decomposition module, configured to perform feature decomposition and feature recombination on the logistics spatio-temporal correlation matrix to obtain a logistics dynamic feature sequence; An analysis module, configured to input the logistics dynamic feature sequence into a hybrid attention network for node correlation weight analysis to generate a logistics matching feature vector; A solution module, configured to construct a multi-objective optimization function based on the logistics matching feature vector and solve to obtain a logistics resource allocation plan; An output module, configured to decompose and dynamically adjust the logistics resource allocation plan and output a target matching result.

[0007] The present invention also provides a computer device, including a memory and a processor, where a computer program is stored in the memory, and when the processor executes the computer program, the steps of the method described in any one of the above are implemented.

[0008] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method described in any one of the above are implemented.

[0009] In summary, the technical solution provided by the present invention realizes parallel processing of logistics spatio-temporal data through a dual-stream feature extraction network, improves the efficiency of feature extraction, and at the same time ensures the integrity of temporal features and spatial features; adopts a multi-scale wavelet decomposition and adaptive weight fusion method to realize fine decomposition and recombination of the logistics spatio-temporal correlation matrix, and accurately captures the periodic and trend features of logistics data; the designed hybrid attention network enhances the model's ability to model long-term and short-term dependencies and improves the richness of feature expression through dual-channel feature separation and multi-level attention calculation; the constructed multi-objective optimization model considers both timeliness and cost-effectiveness at the same time, and through the Lagrange multiplier method and iterative optimization algorithm, realizes balanced solution under multiple constraints; a complete dynamic adjustment mechanism is established, and through real-time feedback analysis and online update strategy, the matching plan can be optimized in a timely manner according to the actual execution situation; through multi-level feature extraction and optimization calculation, the present invention significantly improves the intelligent level and operation efficiency of cross-border logistics distribution. Description of the Drawings

[0010] Figure 1 is a schematic diagram of the steps of a cross-border e-commerce logistics intelligent matching method in an embodiment of the present invention; Figure 2 is a structural block diagram of a cross-border e-commerce logistics intelligent matching device in an embodiment of the present invention; Figure 3 is a structural schematic diagram of a computer device in an embodiment of the present invention.

[0011] The realization, functional features and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. Specific embodiments

[0012] In order to make the object, technical solution and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0013] Refer to Figure 1 , this embodiment provides a cross-border e-commerce logistics intelligent matching method, including the following steps: S1, preprocess the order data, transport capacity data and distribution record data collected by the logistics nodes to generate a logistics data set; Among them, the order data collected by the logistics nodes is subjected to feature decomposition, key attributes are extracted from the original order data, and a feature set that has a substantial impact on the logistics system is formed. During the feature decomposition process of the order data, an order feature sequence including order number, order origin, order destination, order weight, order timeliness, and order price is obtained, reflecting the basic attributes of the order and its potential impact on logistics operations. At the same time, the capacity data collected by the logistics nodes is subjected to feature decomposition, and the extracted features include carrier number, transportation tool type, load capacity, transportation timeliness, and transportation cost, etc. By decomposing these features, the capabilities and limitations of each carrier and transportation tool in the logistics network are evaluated, and these data contribute to the rational allocation of logistics resources and the optimal matching of transportation capacity. The distribution record data collected by the logistics nodes is subjected to feature decomposition, and a distribution feature sequence including historical distribution time, historical distribution path, historical distribution cost, and historical service evaluation is extracted, effectively reflecting the past logistics execution situation, including logistics service quality and operation efficiency, etc. The missing values in the order feature sequence, capacity feature sequence, and distribution feature sequence are input into the time series interpolation model. Through this time series interpolation model, the missing data sequence is completed based on the data mean value within adjacent time windows. This method effectively fills the gaps in the data by utilizing the temporal continuity and data trend, making the completed data more reasonable and conforming to the actual time evolution law. After completing the filling of the missing values, the standard deviation of the completed data sequence is calculated. In this way, the volatility of the data is measured, and outliers are identified. By setting a threshold interval, the outliers exceeding the threshold interval are replaced by the median, effectively reducing the impact of outliers on the overall data distribution, and obtaining a more stable and reliable cleaned data sequence. The cleaned data sequence is subjected to normalization calculation, and the data of different features are scaled to the same dimension range to eliminate the dimension differences between features, thereby improving the calculation efficiency and accuracy of the model, and obtaining a normalized feature sequence. According to the order features and capacity features, the normalized feature sequence is matrix reconstructed to generate a logistics data set including an order feature matrix and a capacity feature matrix. The dimension of the order feature matrix is m×n, where m represents the number of orders and n represents the feature dimension of the orders, while the dimension of the capacity feature matrix is p×q, where p represents the number of carriers and q represents the dimension of the capacity features. Through matrix reconstruction, these logistics feature data are stored and processed more efficiently, and a standardized input form is provided for subsequent feature extraction and matching algorithms.

[0014] S2. Input the logistics data set into a two-stream feature extraction network for parallel feature calculation to generate a logistics spatio-temporal correlation matrix; Specifically, the order feature matrix in the logistics data set is input into the temporal feature branch of the dual-stream feature extraction network for processing to capture the temporal dynamics of order features. This branch contains three layers of temporal convolutional modules, and each layer of temporal convolutional module consists of a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. The combination of these layers effectively captures the changing patterns of order features over time and enhances the model's sensitivity to order demand fluctuations. In the one-dimensional convolutional layer, the convolutional kernel sizes are set to 3, 5, and 7 respectively. These convolutional kernels of different sizes are used to extract order features at different time scales to capture the time-dependent relationships from short-term to long-term, resulting in an order temporal feature tensor. At the same time, the capacity feature matrix in the logistics data set is input into the spatial feature branch of the dual-stream feature extraction network for processing. This spatial feature branch is designed to analyze the spatial distribution characteristics of capacity features and contains three layers of spatial convolutional modules. Each layer of spatial convolutional module consists of a two-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. Through the two-dimensional convolutional layer, the spatial distribution features of capacity in the logistics network are captured, the mutual influence relationships between various logistics nodes are extracted, and the robustness and non-linear expression ability of feature extraction are improved through the batch normalization layer and the ReLU activation function. The convolutional kernel size of each layer is 3×3, and this design effectively captures the local correlation of capacity distribution among carriers, generating a capacity spatial feature tensor. Temporal attention calculation is performed on the order temporal feature tensor to obtain a temporal weight matrix. According to the changes in order features at different times, different weights are assigned to different features to highlight the features that have a greater impact on logistics matching. Similarly, spatial attention calculation is performed on the capacity spatial feature tensor to obtain a spatial weight matrix. The role of spatial attention is to identify the importance of capacity features at different spatial positions, thereby highlighting the features that are more critical for logistics resource allocation and distribution path planning. Through the tensor multiplication operation of the temporal weight matrix and the spatial weight matrix, an initial association matrix is generated to describe the preliminary association relationship between order features and capacity features. The initial association matrix is input into the feature fusion layer. The feature fusion layer contains two fully connected layers and a Tanh activation function. Through the non-linear transformation of the fully connected layer and the output of the Tanh activation function, the complex relationship between orders and capacity can be effectively fused to obtain a fused feature matrix, which more accurately expresses the association and interaction between nodes in the logistics network. Dimension reshaping and normalization processing are performed on the fused feature matrix to generate a logistics spatio-temporal association matrix. The purpose of dimension reshaping is to convert the fused feature matrix into a structural form with clear node association meaning, enabling it to intuitively represent the relationship between order nodes and capacity nodes in the logistics network. The normalization processing is to eliminate the dimensional differences between feature values, making the elements in the association matrix have a relatively consistent scale. The finally obtained logistics spatio-temporal association matrix contains the capacity distribution relationship between logistics nodes and the delivery relationship of orders, comprehensively describing the association and interaction between nodes in the logistics network.In this matrix, the matrix element aij represents the association strength between the i-th order node and the j-th transportation capacity node, and this association strength is used to evaluate the matching degree between order demand and transportation capacity supply.

[0015] S3. Perform eigenvalue decomposition and eigenrecombination on the logistics spatio-temporal association matrix to obtain the logistics dynamic feature sequence; It should be noted that for the multi-scale wavelet decomposition of the logistics spatio-temporal correlation matrix, by using the Daubechies wavelet basis function to calculate the wavelet coefficients, the periodic characteristics of different time scales hidden in the logistics data are effectively extracted. Using the Daubechies wavelet basis function for wavelet transform can decompose the complex logistics spatio-temporal correlation matrix into periodic characteristic matrices of different frequency components. These characteristic matrices respectively describe the daily cycle, weekly cycle, and monthly cycle in the dynamic characteristics of logistics, capturing the fluctuations and change characteristics of different cycles in the logistics network. Perform singular value decomposition operation on the periodic characteristic matrix, and obtain the main period component matrix by calculating the eigenvalues and eigenvectors. Extract the characteristic components that have the greatest impact on the dynamic changes of logistics from the periodic characteristic matrix, simplify the representation of the characteristics, and retain the most representative periodic change information. On this basis, input the original logistics spatio-temporal correlation matrix into the moving average filter, and obtain the trend characteristic matrix by calculating the weighted average within the sliding time window. The role of the moving average filter is to smooth the short-term fluctuations in the logistics data and reveal the more significant trend characteristics in the data. These trend characteristics reflect the overall change trend of the logistics network in the long-term operation. Perform least squares fitting calculation on the trend characteristic matrix to construct a polynomial regression model to extract the long-term trend characteristics. Through this method, the long-term change trend of the logistics system is fitted to obtain the long-term trend function. Based on this long-term trend function, a trend component matrix is generated to reflect the behavioral characteristics of the logistics network in the long-term evolution. The long-term trend reflects the regular changes shown by the logistics system during continuous operation. Calculate the variance contribution degree of each cycle component in the main cycle component matrix. The variance contribution degree is used to measure the influence degree of different cycle components on the overall logistics characteristic changes. Through eigenvalue variance normalization, the cycle characteristic weight vector is obtained, which reflects the relative importance of each cycle component in the dynamic changes of the logistics system. At the same time, perform time series correlation analysis on the trend component matrix. By calculating the autocorrelation coefficient, the trend characteristic weight vector is obtained. The autocorrelation coefficient can measure the similarity between trend characteristics at different time points, thereby effectively evaluating the internal structure and regularity of the trend characteristics. Input the cycle characteristic weight vector and the trend characteristic weight vector into the adaptive weight fusion device, and obtain the fusion feature tensor through dynamic weighted calculation. The adaptive weight fusion device includes a fully connected layer and a Softmax normalization layer. The fully connected layer is used for linear transformation and non-linear mapping of the input features to enhance the interaction relationship between the features, while the Softmax normalization layer is used for normalizing the weights to ensure that the sum of all weights is 1, making the weight calculation process more reasonable and stable. Through this step, the fusion feature tensor contains the comprehensive information of periodic characteristics and trend characteristics and has good dynamic adaptability, and can dynamically adjust the weight distribution of the features according to the real-time changes of the logistics network. Perform time series recombination on the fusion feature tensor, and obtain the logistics dynamic feature sequence by splicing the feature vectors in chronological order.The eigenvectors of different time steps in the fused feature tensor are concatenated in sequence to form a complete dynamic feature sequence, reflecting the dynamic association and evolution process between nodes in the logistics system.

[0016] S4. Input the logistics dynamic feature sequence into the hybrid attention network for node correlation weight analysis to generate a logistics matching eigenvector; Specifically, the logistics dynamic feature sequence is input into the first feature separation module, and a gated recurrent unit is used to perform dual-channel feature extraction on the feature sequence for the main path branch and the bypass branch. The design of the dual-channel structure enables the network to process the main feature information and the secondary feature information separately, retaining both the most important dynamic characteristics in the logistics system and the secondary information that is supplementary to the matching. Through feature separation, a first feature matrix and a second feature matrix are obtained. Multilayer self-attention calculation is performed on the first feature matrix to comprehensively capture the complex dependencies between spatial nodes in the logistics network. Short-range spatial dependencies are calculated through the first attention head, that is, the mutual influence between relatively close nodes is modeled to identify the interactions between other nodes adjacent to or directly associated with each logistics node. Long-range spatial dependencies are calculated through the second attention head to analyze the correlations between remote nodes, thereby capturing the long-distance association characteristics spanning multiple logistics nodes. By combining these two attention mechanisms at different scales, a multi-scale spatial attention vector is obtained. At the same time, hierarchical temporal attention calculation is performed on the second feature matrix, which is divided into two levels: the first level is used to extract local temporal patterns, targeting short-term fluctuations and feature changes in the time series, which helps to understand the dynamic behavior characteristics of the logistics system within a relatively short time range; the second level is used to extract global temporal patterns, capturing long-term trends and regularities to form a comprehensive global temporal feature. Through hierarchical attention calculation, a hierarchical temporal attention vector is obtained, which fully reflects the dynamic characteristics of the logistics system at different times. The multi-scale spatial attention vector is input into the first feature enhancement module to enhance the spatial features. This feature enhancement module includes a feed-forward neural network and a skip connection. The feed-forward neural network performs a non-linear transformation on the features to improve the expressive power of the features; while the skip connection, through the way of residual learning, enables the network to better capture complex feature information and at the same time alleviates the problem of gradient vanishing. Through the design of this structure, spatially enhanced features are obtained. The hierarchical temporal attention vector is input into the second feature enhancement module for enhancement. The second feature enhancement module includes a self-calibrating gating unit and an adaptive normalization layer, where the self-calibrating gating unit selectively enhances or suppresses specific temporal features through a gating mechanism to perform more refined adjustment on the temporal information; the adaptive normalization layer normalizes the features to make the scales of different features consistent, thereby improving the training stability and generalization ability of the model. The obtained temporally enhanced features can better reflect the dynamic change characteristics of the time dimension in the logistics system. Cross-modal feature fusion is performed on the spatially enhanced features and the temporally enhanced features to generate a fused feature tensor through an attention-guided feature recalibration mechanism. By combining the important information in the spatial and temporal features, the fused feature tensor can capture the dynamic changes of the logistics system in both the spatial and temporal dimensions simultaneously.In this process, the attention-guided feature recalibration mechanism effectively weights the features of different modalities, thereby enhancing the influence of key features and suppressing redundant information. The fused feature tensor is input into the feature compression network for feature compression and feature calibration. The role of the feature compression network is to reduce the dimension of the fused features to reduce the data volume and improve the computational efficiency while retaining the most representative feature information; while feature calibration is used to standardize the features to ensure good stability and interpretability of the features in subsequent calculations, and finally generate a logistics matching feature vector, which is used to evaluate the matching degree between logistics nodes.

[0017] S5. Construct a multi-objective optimization function based on the logistics matching feature vector and solve to obtain a logistics resource allocation plan; Among them, an aging optimization function is constructed based on the logistics matching feature vector. The goal of aging optimization is to minimize the difference between the actual delivery time and the expected delivery time of orders as much as possible. By calculating the sum of squared differences between the expected delivery time and the actual delivery time of orders, an aging loss function is obtained to quantify the severity of order delays in the logistics system. At the same time, a cost optimization function is constructed based on the logistics matching feature vector to minimize the overall logistics cost as much as possible on the premise of ensuring logistics aging. The cost optimization function obtains a cost loss function by calculating the weighted sum of transportation costs, warehousing costs, and labor costs. In this way, the impact of various costs in logistics operations can be accurately evaluated and incorporated into the optimization objectives. The aging loss function and the cost loss function are weighted and combined to generate a multi-objective optimization function by setting a balance factor. The role of the balance factor is to adjust the trade-off relationship between aging and cost, so that the optimization result can achieve a reasonable balance between aging and cost. To ensure the feasibility of the optimization result, capacity constraints are added to the multi-objective optimization function. The capacity constraints include the maximum load capacity of the carrier, the maximum service radius, and the time window limit. These constraints effectively limit the scope of logistics resource allocation and avoid unrealistic allocation schemes. Based on the constraints, a capacity constraint matrix is generated to constrain the variables in the multi-objective optimization function and ensure that the final logistics resource allocation scheme conforms to the capacity limitations in reality. The multi-objective optimization function and the capacity constraint matrix are input into the Lagrange multiplier method solver, and the corresponding dual problem is obtained by constructing a Lagrangian function. The introduction of the Lagrangian function can transform the original optimization problem into a dual problem, making the problem-solving more flexible and operable when dealing with complex constraints. The dual problem is solved by subgradient iteration, and the optimal solution is gradually approximated. In this process, the iterative update formula is obtained by calculating the KKT (Karush-Kuhn-Tucker) conditions. The iterative update formula includes two parts: the update of the original variables and the update of the dual variables. The update of the original variables is used to adjust the specific logistics resource allocation scheme, while the update of the dual variables is used to adjust the penalty term of the capacity constraint to ensure that the final allocation scheme meets both the optimization objectives and the constraint conditions. Optimization calculations are performed according to the iterative update formula, and the values of the variables are gradually updated in a cyclic iterative manner. In each iteration, it is judged whether to continue the iteration according to the current objective function value. By setting a convergence threshold, the iteration stops when the change in the objective function value is less than this threshold, and the optimal solution of the multi-objective optimization problem is obtained. This optimal solution is a comprehensive balance of logistics aging and cost, and the optimal resource allocation scheme obtained while meeting the capacity constraints. Substitute the optimal solution into the decision variable mapping function to obtain the specific logistics resource allocation scheme through calculation, including the order-carrier matching matrix and the transportation path matrix.The matrix element Bmn in the order-carrier matching matrix represents the probability that the m-th order is assigned to the n-th carrier. This probability representation makes the solution flexible, and the optimal carrier can be selected according to the real-time situation during the actual execution process. At the same time, the matrix element Ckl in the transportation path matrix represents the connection relationship from the k-th node to the l-th node in the path. This matrix can clearly describe the path planning of order delivery, ensure the optimal route selection in actual transportation, and reduce transportation time and costs.

[0018] S6. Decompose and dynamically adjust the logistics resource allocation solution, and output the target matching result.

[0019] Specifically, the task decomposition process is carried out on the logistics resource allocation plan, refining the complex resource scheduling problem into executable specific tasks. The orders are grouped according to the destination and time limit requirements of the orders, and an order grouping matrix is obtained. Based on the order grouping matrix, a carrier matching model is constructed. Specifically, various factors such as the load capacity, route coverage, and historical service quality of the carriers are considered, and a comprehensive score is given to each carrier to obtain a carrier score vector. The carrier score vector is sorted and screened, and the optimal carrier for each order group is determined to generate a carrier allocation instruction set. This instruction set includes the scheduling priority of the carrier, the loading time window, and the service level requirements. Based on the carrier allocation instruction set and the geographical location information of each node in the order set, a path planning graph is constructed, and the optimal path between nodes is calculated to obtain an initial path plan. The initial path plan includes the path sequence and the estimated passing time. Through this planning, on the premise of ensuring the time limit, the path selection is optimized as much as possible, the total mileage and transportation cost of vehicle driving are reduced, and the logistics operation efficiency is improved. Multi-objective optimization is carried out on the initial path plan, including the joint optimization of factors such as the total path length, time window constraint, and service quality. Through multi-objective optimization, a transportation path instruction set is obtained. The transportation path instruction set includes detailed passing points, the estimated arrival time of each node, and corresponding operation requirements. During the logistics execution process, the actually collected delivery time, path change situation, service evaluation, etc. are input into the feedback analysis module, and these data are dynamically analyzed through a sliding time window, and the deviation values of various indicators are calculated to obtain an execution deviation matrix. The generation of the execution deviation matrix enables the logistics system to monitor the deviation situation in the execution process in real time. Based on the execution deviation matrix, the carrier allocation instruction set and the transportation path instruction set are updated online, and the carrier selection strategy and path planning strategy are dynamically adjusted to obtain an optimized instruction set. The dynamic adjustment mechanism can ensure that various uncertain factors are timely responded to during the logistics execution process, such as changes in traffic conditions and the temporary inability of carriers to fulfill contracts, thereby effectively improving the flexibility and response ability of the logistics system. Instruction parsing and execution status tracking are carried out on the optimized instruction set to ensure that each instruction can be accurately executed. Through a real-time monitoring algorithm, the progress and abnormal situations of instruction execution are recorded to ensure the stable operation of the logistics system. When an abnormal situation is monitored, the system responds in a timely manner, adjusts the logistics resources and execution plan to minimize the impact of the abnormality on the overall logistics operation. Through these operations, the target matching result is finally output to ensure that the logistics resources are optimally allocated, the distribution path is the most efficient, and the logistics service quality meets the expected standard.

[0020] In one example, the order data, transport capacity data, and distribution record data collected from logistics nodes are preprocessed to generate a logistics data set, including: Perform feature decomposition on the order data collected by the logistics node to obtain an order feature sequence containing order number, order source, order destination, order weight, order timeliness, and order price; perform feature decomposition on the transport capacity data collected by the logistics node to obtain a transport capacity feature sequence containing carrier number, transport vehicle type, load capacity, transport timeliness, and transport cost; perform feature decomposition on the distribution record data collected by the logistics node to obtain a distribution feature sequence containing historical distribution time, historical distribution route, historical distribution cost, and historical service evaluation; Input the missing values in the order feature sequence, transport capacity feature sequence, and distribution feature sequence into the time series interpolation model, obtain the completed data sequence by calculating the mean value of the data within adjacent time windows, calculate the standard deviation of the completed data sequence, and replace the outliers beyond the threshold interval with the median to obtain the cleaned data sequence; Perform normalization calculation on the cleaned data sequence to obtain a normalized feature sequence, and perform matrix reconstruction on the normalized feature sequence according to the order features and transport capacity features to generate a logistics data set containing an order feature matrix and a transport capacity feature matrix, where the dimension of the order feature matrix is m×n, m represents the number of orders, n represents the order feature dimension, and the dimension of the transport capacity feature matrix is p×q, p represents the number of carriers, and q represents the transport capacity feature dimension.

[0021] In this example, perform feature decomposition on the order data collected by the logistics node. The order data collected in the logistics system contains various information, such as order number, order source and destination, order weight, estimated delivery timeliness, and order price. Decompose this data to obtain a standardized and easy-to-process order feature sequence in the form of: ; where, represents the number of the th order, represents the order source, represents the order destination, represents the order weight, represents the order timeliness, and represents the order price. By extracting this information for each order, an order feature sequence is formed, containing the key attributes of all orders. Perform feature decomposition on the transport capacity data collected by the logistics node. The transport capacity data reflects the transport capacity of the carrier, including information such as the carrier number, type of transport vehicle, load capacity, transport timeliness, and transport cost. The transport capacity feature sequence is expressed as: ; where, represents the The number of a carrier, Indicates the type of transportation vehicle (such as truck, cargo ship, air freight, etc.), Indicates the load capacity of the transportation vehicle, Indicates the timeliness of transportation, Indicates the transportation cost. By decomposing the transportation capacity of each carrier, we can better understand the transportation capabilities and service characteristics of different carriers, so as to make an optimized choice in the allocation of logistics resources. Feature decomposition is carried out on the distribution record data collected at logistics nodes. The distribution record data includes historical distribution time, distribution path, distribution cost, and service evaluation and other information. Through feature extraction of these data, a historical distribution feature sequence is obtained: ; Among them, Indicates the historical distribution time of the th distribution, Indicates the distribution path, Indicates the distribution cost, Indicates the customer's evaluation of the service (such as satisfaction score). These data can reflect the past operation of the logistics system. In the data preprocessing stage, the missing values in the order feature sequence, transportation capacity feature sequence, and distribution feature sequence are processed. A time series interpolation model is used to fill in the missing values. The missing values are estimated by calculating the mean value of the data within adjacent time windows. Suppose there is a time window , and the mean value within the window is expressed as: ; Among them, Is the number of data within the window, Is the data at time . In this way, a completed data sequence is obtained to ensure that the feature values are continuous and valid in the time dimension. For the completed data sequence, outlier detection and processing are carried out. By calculating the standard deviation of the data sequence to judge whether there are outliers, calculate the mean value and standard deviation of the data sequence: ; Set a threshold interval to judge which data are outliers, where is a control parameter, usually taking values between 2 and 3. For outliers outside this interval, the median is used for replacement. Let be the median of the data, and the replacement rule for outliers is: ; By processing outliers through this method, a cleaned data sequence is obtained. The cleaned data is normalized to eliminate the dimensional differences between different features and improve the performance and convergence speed of the model. The formula for normalization is expressed as: ; where and represent the minimum and maximum values of the data respectively. The normalized feature value is compressed into the interval [0, 1] to ensure that all features have the same scale, facilitating subsequent feature fusion and processing. After normalization, the order features and transportation capacity features are matrix-reconstructed to generate a logistics data set containing an order feature matrix and a transportation capacity feature matrix. The dimension of the order feature matrix is , where represents the number of orders, represents the dimension of the order features ( , because the order features include number, source, destination, weight, timeliness, and price). The order feature matrix is expressed as: ; Similarly, the dimension of the transportation capacity feature matrix is , where represents the number of carriers, represents the dimension of the transportation capacity features ( , because the transportation capacity features include number, tool type, load capacity, timeliness, and cost). The transportation capacity feature matrix is expressed as: ; By matrix-reconstructing the order features and transportation capacity features, a complete logistics data set is obtained.

[0022] In an example, the logistics data set is input into a two-stream feature extraction network for feature parallel computing to generate a logistics spatio-temporal correlation matrix, including: The order feature matrix in the logistics data set is input into the temporal feature branch in the two-stream feature extraction network for processing. The temporal feature branch contains three layers of temporal convolution modules. Each layer of the temporal convolution module contains a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. The kernel sizes are 3, 5, and 7 respectively, to obtain an order temporal feature tensor; The transportation capacity feature matrix in the logistics data set is input into the spatial feature branch in the two-stream feature extraction network for processing. The spatial feature branch contains three layers of spatial convolution modules. Each layer of the spatial convolution module contains a two-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. The kernel size is 3×3, to obtain a transportation capacity spatial feature tensor; Perform temporal attention calculation on the order temporal feature tensor to obtain a temporal weight matrix, and perform spatial attention calculation on the transportation capacity spatial feature tensor to obtain a spatial weight matrix; Perform tensor multiplication on the temporal weight matrix and the spatial weight matrix to obtain an initial correlation matrix, and input the initial correlation matrix into the feature fusion layer. The feature fusion layer includes two fully connected layers and a Tanh activation function, and obtains a fused feature matrix through non-linear transformation; Perform dimension reshaping and normalization on the fused feature matrix to generate a logistics spatio-temporal correlation matrix. The logistics spatio-temporal correlation matrix includes the transportation capacity distribution relationship and order delivery relationship between nodes. The matrix element aij of the logistics spatio-temporal correlation matrix represents the correlation strength between the i-th order node and the j-th transportation capacity node.

[0023] In this example, the order feature matrix in the logistics dataset is input into the temporal feature branch of the two-stream feature extraction network for processing. For the order feature matrix , its dimension is , where represents the number of orders, represents the feature dimension of the order, including order number, order source, order destination, order weight, order timeliness, order price, etc. Input into the temporal feature branch. The temporal feature branch includes three layers of temporal convolution modules. The role of each layer of temporal convolution module is to extract the dynamic change characteristics of the order in the time dimension. The first layer of temporal convolution module of the temporal feature branch consists of a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. The convolutional kernel size of the one-dimensional convolutional layer is 3, indicating that the features of adjacent 3 time moments are convolved on the time axis. Assume that the weight of the convolutional kernel is , and the input data is , the convolutional output is expressed as: ; Among them, * represents the convolution operation, represents the bias term, is the feature obtained after the first layer of convolution. The batch normalization layer normalizes to reduce the numerical differences between different samples. The batch normalization formula is: ; Among them, and are the mean and standard deviation of the batch data respectively, is a small value used to prevent the denominator from being zero. Use the ReLU activation function to non-linearly activate the normalized feature, and the formula is: ; The purpose of ReLU activation is to enhance the non - linear representation ability of the model, enabling the extracted features to better express the complex characteristics of the data. Similarly, the second - layer and third - layer temporal convolutional modules use convolutional kernels of sizes 5 and 7 respectively to perform larger - range temporal feature extraction layer by layer, obtaining the order temporal feature tensor . Meanwhile, the capacity feature matrix in the logistics dataset is input into the spatial feature branch of the dual - stream feature extraction network for processing. The dimension of the capacity feature matrix is , where represents the number of carriers, represents the dimension of carrier features, including carrier number, transport vehicle type, load capacity, transport timeliness, and transport cost, etc. The spatial feature branch contains three layers of spatial convolutional modules, and each layer of spatial convolutional module consists of a two - dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. The size of the convolutional kernel of the two - dimensional convolutional layer is , which is used to extract local feature relationships in the two - dimensional space of the carrier feature matrix. Assuming the weight of the convolutional kernel is , and the input data is , the convolution operation is expressed as: ; where, * represents two - dimensional convolution, represents the bias term, is the feature after convolution. The batch normalization layer is used to normalize , and then the ReLU activation function is used for non - linear activation to improve the spatial feature expression ability of the model. Through the stacking of multiple layers of convolutions, the spatial relationships in the carrier feature matrix are gradually extracted, obtaining the capacity spatial feature tensor . The order temporal feature tensor and the capacity spatial feature tensor are subjected to attention calculation. For the order temporal feature tensor, temporal attention calculation is used to capture the correlation between each time node, obtaining the temporal weight matrix . The formula for temporal attention is expressed as: ; where, and are the query matrix and key matrix generated from through linear transformation respectively, represents the feature dimension, and the softmax function is used to normalize the weights. The temporal weight matrix reflects the importance distribution of order features in the time dimension. For the capacity spatial feature tensor, spatial attention calculation is used to analyze the mutual influence between different carriers, obtaining the spatial weight matrix The calculation of spatial attention is similar to that of temporal attention and is expressed as: ; where and are the query matrix and key matrix generated from through linear transformation. The spatial weight matrix is used to reflect the interaction relationship between different carriers in the spatial dimension. Perform tensor multiplication on the temporal weight matrix and the spatial weight matrix to combine the correlation information between orders and transportation capacities and obtain the initial correlation matrix , and the formula is: ; where · represents the tensor multiplication operation, reflects the preliminary correlation relationship between order features and transportation capacity features. Input the initial correlation matrix into the feature fusion layer, which contains two fully connected layers and a Tanh activation function. The first fully connected layer performs a linear transformation on the initial correlation matrix, and the formula is: ; where and are the weight matrix and bias term of the fully connected layer respectively, is the feature after the first linear transformation. Use the Tanh activation function for non-linear transformation, and the formula is: ; The role of the Tanh function is to compress the feature values into the range of [-1, 1], thereby increasing the non-linear expression ability of the model. Use the second fully connected layer to perform a second linear transformation on to obtain the fused feature matrix : ; where and are the weight matrix and bias term of the second fully connected layer respectively. Through the processing of the feature fusion layer, the order features and transportation capacity features are effectively fused to obtain a fused feature matrix containing multi-dimensional correlation information. Perform dimension reshaping and normalization on the fused feature matrix to generate the logistics spatio-temporal correlation matrix . The purpose of dimension reshaping is to adjust the fused feature matrix into a structured form to facilitate representing the correlation between logistics nodes. The purpose of normalization is to eliminate the scale differences between different features, making the elements of the logistics spatio-temporal correlation matrix numerically consistent. The finally generated logistics spatio-temporal correlation matrix It includes the transport capacity distribution relationship and order delivery relationship between nodes, where the matrix elements represent the th order node and the th transport capacity node, and the correlation strength therebetween.

[0024] In one example, the logistics spatio-temporal correlation matrix is subjected to eigen decomposition and eigen recombination to obtain a logistics dynamic feature sequence, including: Perform multi-scale wavelet decomposition on the logistics spatio-temporal correlation matrix, calculate wavelet coefficients by using Daubechies wavelet basis functions, and obtain a periodic feature matrix including daily, weekly, and monthly cycles; Perform singular value decomposition operation on the periodic feature matrix, obtain the main periodic component matrix by calculating eigenvalues and eigenvectors, input the logistics spatio-temporal correlation matrix into a moving average filter, and obtain the trend feature matrix by calculating the weighted average within a sliding time window; Perform least squares fitting calculation on the trend feature matrix, obtain a long-term trend function by constructing a polynomial regression model, and generate a trend component matrix according to the long-term trend function; Calculate the variance contribution degree of each periodic component in the main periodic component matrix, obtain the periodic feature weight vector through eigen variance normalization, perform time series correlation analysis on the trend component matrix, and obtain the trend feature weight vector by calculating the autocorrelation coefficient; Input the periodic feature weight vector and the trend feature weight vector into an adaptive weight fusion device, obtain a fusion feature tensor through dynamic weighted calculation, and the adaptive weight fusion device includes a fully connected layer and a Softmax normalization layer; Perform time series recombination on the fusion feature tensor, and obtain a logistics dynamic feature sequence by splicing feature vectors in chronological order.

[0025] In this example, multi-scale wavelet decomposition is performed on the logistics spatio-temporal correlation matrix to reveal its periodic features at different time scales. The Daubechies wavelet basis function is used to perform wavelet decomposition on the logistics spatio-temporal correlation matrix to calculate wavelet coefficients at different time scales. These wavelet coefficients reflect the features of different cycles (such as daily, weekly, and monthly cycles) in the logistics data. Assume that the original logistics spatio-temporal correlation matrix is , and through the Daubechies wavelet transform, we get: where DWT( ) represents the discrete wavelet transform, is the wavelet basis function, respectively represent the wavelet coefficients of the daily, weekly, and monthly cycles, and these wavelet coefficients form a periodic feature matrix containing different periodic features . Perform singular value decomposition on the periodic characteristic matrix to extract the most important part of the periodic characteristics. Find the main influencing factors from the complex periodic data to simplify subsequent analysis. Assume that the periodic characteristic matrix is , then the singular value decomposition is expressed as: ; in, and are the left singular vector matrix and the right singular vector matrix, respectively. is a diagonal matrix of singular values. By selecting the largest singular value and its corresponding eigenvector, we get the main periodic component matrix , contains the most important periodic information. At the same time, in order to analyze the trend characteristics in the logistics space-time correlation matrix, Input to the moving average filter. The moving average filter smoothes short-term fluctuations by calculating the weighted average of the data in the sliding time window and reveals the long-term trend characteristics. Suppose the sliding time window size is , then the trend characteristic matrix The calculation formula is: ; in, Indicates at time The logistics data characteristics at that time are used to obtain a smooth trend feature matrix through sliding average to eliminate the influence of noise. In order to extract long-term trend characteristics, the trend feature matrix Perform least squares fitting calculation. By constructing a polynomial regression model to fit the long-term trend, a trend function that can accurately describe the long-term changes in logistics data is obtained. Assume that the fitted polynomial is , the coefficients of the polynomial are solved by the least squares method. The goal of the least squares fitting is to minimize the sum of squared errors, and the formula is: ; Long-term trend function obtained by least squares fitting , generate the trend component matrix , which reflects the evolution characteristics of the logistics system over a longer period of time. The variance contribution of each period component in the main period component matrix is ​​calculated to quantify the impact of each period on the overall logistics characteristics. The variance contribution is calculated by calculating the variance of each period characteristic. Assuming that the variance of a period component is , then the periodic feature weight vector By normalizing the feature variance, we get: ; in, represents the number of periodic components, Indicates the contribution weight of each period in the overall change. At the same time, in order to perform a time series correlation analysis on the trend component matrix, the autocorrelation coefficient between each time point is calculated. The autocorrelation coefficient reflects the similarity of trend characteristics between different time points, and the formula is: ; Among them, Indicates the time lag, Is the mean of the trend component matrix. By calculating the autocorrelation coefficient, the trend characteristic weight vector Is obtained, which is used to describe the correlation of long-term trend characteristics. The periodic characteristic weight vector And the trend characteristic weight vector Are input into the adaptive weight fusion device, which includes a fully connected layer and a Softmax normalization layer. The role of the fully connected layer is to perform a linear combination of the input features, and the formula is: ; Among them, And Are the weight matrix and the bias term respectively. The Softmax layer is used to normalize the fused features to ensure that the weights are numerically interpretable, and the formula is: ; Among them, Is the normalized weight. Through this step, the fused feature tensor Is obtained, which combines periodic and trend information. The fused feature tensor Is reorganized in time series. The feature vectors at different time steps are concatenated in chronological order to form a complete physical flow dynamic feature sequence. Let Represent the fused feature vector at time , then the physical flow dynamic feature sequence is expressed as: ; Through time series reorganization, the fused feature tensor is arranged in chronological order to generate a physical flow dynamic feature sequence containing rich spatio-temporal information. This feature sequence reflects the dynamic changes of the physical flow system in different time periods and includes a comprehensive description of periodicity and trend.

[0026] In an example, the physical flow dynamic feature sequence is input into the hybrid attention network for node correlation weight analysis to generate a physical flow matching feature vector, including: The physical flow dynamic feature sequence is input into the first feature separation module, and the gated recurrent unit is used to perform two-channel feature extraction on the feature sequence for the main path branch and the bypass branch to obtain the first feature matrix and the second feature matrix; Perform multi - layer self - attention calculation on the first feature matrix. Calculate short - range spatial dependence through the first attention head and long - range spatial dependence through the second attention head to obtain a multi - scale spatial attention vector; Perform hierarchical temporal attention calculation on the second feature matrix. Extract local temporal patterns through the first layer and global temporal patterns through the second layer to obtain a hierarchical temporal attention vector; Input the multi - scale spatial attention vector into the first feature enhancement module. The first feature enhancement module contains a feed - forward neural network and a skip connection, and obtain spatial enhanced features through residual learning; Input the hierarchical temporal attention vector into the second feature enhancement module. The second feature enhancement module contains a self - calibration gating unit and an adaptive normalization layer, and obtain temporal enhanced features through the gating mechanism; Perform cross - modal feature fusion on the spatial enhanced features and the temporal enhanced features. Obtain a fused feature tensor through an attention - guided feature recalibration mechanism, and input the fused feature tensor into a feature compression network for feature compression and feature calibration to generate a logistics matching feature vector.

[0027] In this example, input the logistics dynamic feature sequence into the first feature separation module, and use a gated recurrent unit (GRU) to perform two - channel feature extraction on the feature sequence. Assume the logistics dynamic feature sequence is , where represents the feature vector at time , and input it into the first feature separation module. This module performs two - channel feature extraction on the main path branch and the bypass branch through a gated recurrent unit. The main path branch is used to capture the main feature information in the feature sequence, while the bypass branch is used to capture secondary features that are supplementary to subsequent decisions. Specifically, the update formula of the GRU is as follows: ; where, represents the hidden state of the main path branch at time , represents element - wise multiplication, is the update gate, is the candidate hidden state. In this way, the GRU adaptively decides which feature information to retain and update according to the input feature sequence, thus effectively extracting the first feature matrix and the second feature matrix . Perform multi - layer self - attention calculation on the first feature matrix to capture the spatial dependence between features. The multi - layer self - attention mechanism processes short - range and long - range dependencies through different attention heads. For the first attention head, it is used to calculate short - range spatial dependence. Assume If they are the query, key, and value matrices respectively, the attention calculation formula for short-range dependencies is as follows: ; where softmax is used to normalize the attention weights, is the scaling factor, which is used to avoid the vanishing gradient caused by too large dot product values. is the short-range dependency vector obtained through the first attention head, which mainly focuses on the spatial interaction between adjacent nodes. For the second attention head, it is used to calculate the long-range spatial dependency. The calculation method of the long-range dependency is similar to that of the short-range, but it focuses on the mutual influence between distant nodes. The calculation formula is: ; Through the multi-layer calculation of the first and second attention heads, a multi-scale spatial attention vector is obtained, which can capture the dependencies of features at different spatial scales. At the same time, hierarchical temporal attention calculation is performed on the second feature matrix to extract the feature relationships in the time dimension. Local temporal patterns are extracted through the first level to capture short-term dynamic changes. Assuming that the query, key, and value matrices of local time series are respectively, the calculation formula of local temporal attention is: ; This process helps to identify the change trend of features in a relatively short time, so as to extract local temporal patterns. Global temporal patterns are extracted through the second-level calculation, focusing on the overall changes over a longer time span. The calculation formula of global temporal attention is: ; Through the two-level temporal attention calculation, a hierarchical temporal attention vector is obtained, which contains local and global time features, enabling the model to better understand the evolution law of features over time. The multi-scale spatial attention vector is input into the first feature enhancement module. The first feature enhancement module contains a feed-forward neural network and a skip connection (residual connection). The feed-forward neural network is used to process the attention vector so that it can capture higher-level abstract features. The formula is: ; where and are the weight matrix and bias term respectively. The skip connection is used to maintain the combination of the original features and the enhanced features through residual learning, thereby preventing the problem of feature degradation and obtaining spatially enhanced features: ; In this way, the spatially enhanced features effectively combine the original features with the features processed by the feed-forward network, enhancing the feature representation ability. At the same time, the hierarchical temporal attention vector is input into the second feature enhancement module. The second feature enhancement module includes a self-calibrating gating unit and an adaptive normalization layer. The role of the self-calibrating gating unit is to selectively enhance or suppress specific temporal features through a gating mechanism, and the formula is: ; where, represents the enhanced feature at time , and is the update gate output by the gating unit. The adaptive normalization layer normalizes the enhanced features, making the features consistent between different batches, and finally obtains the temporally enhanced feature . Cross-modal feature fusion is performed on the spatially enhanced feature and the temporally enhanced feature . In order to effectively fuse these two different modal features, an attention-guided feature recalibration mechanism is used to enhance important features and suppress redundant features through weight allocation. Assuming the fused feature is , the fusion process is expressed as: ; where, Attention represents the attention-based fusion operation, and the features are recalibrated by calculating the correlation of the fused features. The fused feature tensor is input into the feature compression network for feature compression and calibration to reduce the dimension of the features and remove redundant information. The formula of the feature compression network is: ; where, and are the weight matrix and bias term of the feature compression network respectively. After feature compression, the dimension of the features is effectively reduced, the computational efficiency of the model is improved, and at the same time, the most representative feature information is retained. The feature vector output by the feature compression network is the logistics matching feature vector , which contains the spatial and temporal features after cross-modal fusion and has been compressed and calibrated to ensure its suitability for subsequent matching and optimization tasks.

[0028] In one example, a multi-objective optimization function is constructed based on the logistics matching feature vector, and the logistics resource allocation scheme is obtained by solving, including: Construct a timeliness optimization function based on the logistics matching feature vector, obtain the timeliness loss function by calculating the sum of squares of the differences between the expected delivery time and the actual delivery time of the order, and construct a cost optimization function based on the logistics matching feature vector, and obtain the cost loss function by calculating the weighted sum of transportation cost, warehousing cost and labor cost; The aging loss function and the cost loss function are weighted and combined. By setting a balance factor, a multi-objective optimization function is obtained, and capacity constraints are added to the multi-objective optimization function, including the maximum load of the carrier, the maximum service radius, and time window restrictions, to generate a capacity constraint matrix. The multi-objective optimization function and the capacity constraint matrix are input into the Lagrange multiplier method solver. By constructing a Lagrangian function, a dual problem is obtained, and the dual problem is solved by subgradient iteration. By calculating the KKT conditions, an iterative update formula is obtained, which includes two parts: the update of the primal variable and the update of the dual variable. Optimization calculations are performed according to the iterative update formula. By setting a convergence threshold, cyclic iteration is carried out until the change in the objective function value is less than the convergence threshold, and then the iteration stops to obtain the optimal solution. The optimal solution is substituted into the decision variable mapping function, and a logistics resource allocation plan is obtained through calculation, including an order-carrier matching matrix and a transportation path matrix. The matrix element Bmn in the order-carrier matching matrix represents the probability that the m-th order is assigned to the n-th carrier, and the matrix element Ckl in the transportation path matrix represents the connection relationship from the k-th node to the l-th node in the path.

[0029] In this example, an aging optimization function is constructed based on the logistics matching feature vector. Assume that the logistics matching feature vector is , and the expected delivery time for each order is , and the actual delivery time is . The goal of aging optimization is to minimize the difference between the actual delivery time and the expected delivery time as much as possible, so as to improve the timeliness of distribution. The aging loss function is constructed by the sum of squares of the differences, and the formula is as follows: ; where represents the aging loss function, and represents the number of orders. This loss function reflects the aging deviation of all orders. The smaller this value is, the better the overall aging performance of the logistics system. At the same time, a cost optimization function is constructed based on the logistics matching feature vector. The costs in the logistics system mainly include transportation costs, warehousing costs, and labor costs. Assume that the transportation cost is , the warehousing cost is and the labor cost is , then the cost loss function is expressed in the form of a weighted sum as: ; where They are the weights of transportation cost, warehousing cost, and labor cost respectively, reflecting the relative importance of different costs in the overall loss. By adjusting these weights, the cost allocation strategy can be dynamically optimized under different logistics scenarios. The timeliness loss function and the cost loss function are weighted and combined to obtain a multi-objective optimization function. To achieve a reasonable balance between timeliness and cost, a balance factor is introduced to control the importance of timeliness and cost. The multi-objective optimization function is expressed as: ; where, is the balance factor. When is larger, the weight of timeliness is larger; while when is smaller, the weight of cost is larger. In this way, a balance point between timeliness and cost can be found according to business requirements. To ensure that the final obtained logistics plan is feasible, capacity constraints are added to the multi-objective optimization function. The capacity constraints include the maximum load of the carrier, the maximum service radius, and the time window limit. Assume that the maximum load of the carrier is , the maximum service radius is , and the time window limit is , then these constraints construct a capacity constraint matrix: ; where, represents the number of orders carried by the th carrier, represents the distance between order and carrier , represents the time arrangement between the order and the carrier. The capacity constraint matrix ensures that in the actual transportation process, the allocation of all orders meets the capacity requirements of the carrier. The multi-objective optimization function and the capacity constraint matrix are input into the Lagrange multiplier method solver, and the constrained optimization problem is transformed into an unconstrained dual problem by constructing a Lagrangian function. The Lagrangian function is expressed as: ; where, represents the decision variable (such as the matching situation between the order and the carrier), is the Lagrange multiplier, is the constant term of the constraint. The dual problem is obtained by taking the extreme value of the Lagrangian function, and the dual problem is solved by the method of subgradient iteration. Using the KKT conditions (Karush-Kuhn-Tucker conditions), the iterative update formula is obtained, which is divided into two parts: the update of the primal variable and the update of the dual variable: Update of the primal variable: ; Dual variable update: ; where and are the learning rates that control the step size of each iteration. Through continuous iterative updates, the optimal solution is gradually approximated. Optimization calculations are performed according to the iterative update formula, and a loop iteration method is adopted until the change in the objective function value is less than the set convergence threshold to stop the iteration, and finally the optimal solution is obtained: ; When the change in the objective function satisfies this condition, it means that an optimal solution that meets the constraint conditions has been found, and the iteration process is stopped. Substitute the obtained optimal solution into the decision variable mapping function to calculate the final logistics resource allocation plan. This plan includes an order-carrier matching matrix and a transportation path matrix. The order-carrier matching matrix is expressed as , where the matrix element represents the probability that the rd order is assigned to the th carrier. This probability-based matching method provides greater flexibility to select the optimal carrier according to dynamic situations during the actual execution process. The transportation path matrix is expressed as , where the matrix element represents the connection relationship between the th node and the th node in the transportation path. If , it means that there is a direct transportation path from the th node to the th node; if , it means that there is no direct path. In this way, the connection situation between different nodes in the logistics system is clearly described, providing a basis for path planning in actual transportation.

[0030] In an example, the logistics resource allocation plan is decomposed and dynamically adjusted to output the target matching results, including: Perform task deconstruction on the logistics resource allocation plan, and group the orders according to the order destination and time limit requirements to obtain an order grouping matrix; Build a carrier matching model based on the order grouping matrix, and comprehensively evaluate the load capacity, line coverage, and historical service quality of the carriers to obtain a carrier score vector; Sort and screen the carrier score vector, and determine the optimal carrier for each order group to generate a carrier allocation instruction set. The carrier allocation instruction set includes carrier scheduling priority, loading time window, and service level requirements; Construct a path planning graph based on the carrier allocation instruction set and the geographical location information of each node in the order set, calculate the optimal path between nodes, and obtain an initial path plan, which includes a path sequence and an estimated travel time; Perform multi-objective optimization on the initial path plan, and jointly optimize the total path length, time window constraint, and service quality to obtain a transportation path instruction set, which includes detailed waypoints, arrival times, and operation requirements; Input the actual delivery time, path change, and service evaluation collected during the logistics execution process into the feedback analysis module, calculate the deviation values of various indicators through a sliding time window, and obtain an execution deviation matrix; Online update the carrier allocation instruction set and the transportation path instruction set based on the execution deviation matrix, and dynamically adjust the carrier selection strategy and path planning strategy to obtain an optimized instruction set; Parse the optimized instruction set and track the execution status, record the instruction execution progress and abnormal conditions through a real-time monitoring algorithm, and output the target matching result.

[0031] In this example, perform task decomposition on the logistics resource allocation plan, decompose the complex logistics resource allocation problem into executable independent tasks, and group the orders according to the destination and timeliness requirements of the orders. Assume the order set is , each order has different destinations and timeliness requirements. Construct an order grouping matrix , whose element indicates whether order belongs to the th order group. If it belongs, then , otherwise it is 0. Through grouping, orders with similar delivery requirements are grouped together to facilitate subsequent resource scheduling and path planning. Based on the order grouping matrix , construct a carrier matching model to determine the most suitable carrier to complete the delivery tasks of each order group. For each carrier , consider its load capacity , line coverage and historical service quality . By comprehensively scoring these factors, obtain a scoring vector for each carrier, where each element represents the comprehensive score of the th carrier. The scoring formula is as follows: ; Among them, They represent the weights of load capacity, line coverage, and service quality respectively. By adjusting these weights, the strategy for selecting carriers can be flexibly optimized according to specific business requirements. Sort and filter the carrier score vectors to select the optimal carrier for each order group. In this way, a most suitable carrier is determined for each order group, and a carrier allocation instruction set is generated. The carrier allocation instruction set includes the scheduling priority of the carrier, the loading time window, and the service level requirements. Let the carrier allocation instruction set be , and its element represents the detailed information of the th order group assigned to the th carrier, including the scheduling priority , the loading time window , and the service level requirements . The generation of the instruction set can ensure that each order group can obtain the best transportation service, thus improving the overall logistics operation efficiency. Based on the carrier allocation instruction set and the geographical location information of each node in the order set, a path planning graph is constructed. Let the node set be , and each node represents a geographical location. Using the information of these nodes and the transportation capacity of the carrier, a path planning graph is constructed, where the weight of the edge represents the transportation cost or time between node and node . Calculate the optimal path between nodes through the shortest path algorithm (such as Dijkstra algorithm) to obtain the initial path plan . The initial path plan includes the path sequence and the estimated passing time. The path sequence represents all the nodes passed from the starting point to the end point, and the estimated passing time is used to estimate the distribution timeliness. Perform multi-objective optimization on the initial path plan to improve logistics efficiency and service quality. Multi-objective optimization needs to consider multiple factors such as the total path length, time window constraint, and service quality at the same time. Assume that the total path length is , the time window constraint is , and the service quality is , then the objective function of multi-objective optimization is expressed as: ; where is the weight of different objectives, and is the multi-objective optimization function. By weighting these objectives, balance the relationship between path length, time window, and service quality to obtain the optimal transportation path plan. The obtained transportation path instruction set It contains detailed transit points, arrival times, and operational requirements to ensure that each carrier can deliver according to the best route. During the logistics execution process, the actual delivery time, route changes, and customer service evaluations are continuously collected and input into the feedback analysis module to monitor and analyze the effect of logistics execution. Through the sliding time window method, the deviation value of each indicator is calculated to obtain the execution deviation matrix . The elements of the execution deviation matrix Indicates order In time The execution deviation on , the calculation formula of the deviation is: ; in, Indicates order In time The actual delivery time on Indicates the expected delivery time. By calculating the deviation, the problems in logistics execution are clearly identified. Based on the execution deviation matrix, the carrier allocation instruction set and the transportation path instruction set are updated online. By dynamically adjusting the carrier selection strategy and path planning strategy, the optimized instruction set is obtained. and The optimized instruction set takes into account the real-time logistics execution status to ensure that the logistics system can be adjusted in time to achieve the best operation effect under the influence of uncertain factors (such as traffic jams, weather changes, etc.). The optimized instruction set is parsed and the execution status is tracked. In order to ensure that each instruction can be executed correctly, the execution progress and abnormal situations of the instructions are recorded through the real-time monitoring algorithm. Let the execution status matrix be , whose elements Indicates order In time The execution status can be 0 (not started), 1 (in progress) or 2 (completed). Through real-time monitoring, abnormal situations in the logistics execution process, such as route changes, delays, etc., are captured. These abnormal situations are handled in a timely manner to effectively reduce the negative impact on the overall logistics operation. Through the above processes and algorithms, the target matching results are output to ensure the efficient use of logistics resources and the timeliness of distribution.

[0032] Reference Figure 2 This embodiment provides a cross-border e-commerce logistics intelligent matching device, including: Preprocessing module 1, used to preprocess the order data, transportation data and delivery record data collected by the logistics node to generate a logistics data set; Computation module 2 is used to input the logistics data set into the dual-stream feature extraction network for feature parallel calculation to generate a logistics spatiotemporal correlation matrix; A decomposition module 3, which is used to perform eigen - decomposition and eigen - recombination on the logistics spatio - temporal correlation matrix to obtain a logistics dynamic feature sequence; An analysis module 4, which is used to input the logistics dynamic feature sequence into a hybrid attention network for node correlation weight analysis to generate a logistics matching feature vector; A solution module 5, which is used to construct a multi - objective optimization function based on the logistics matching feature vector and solve to obtain a logistics resource allocation plan; An output module 6, which is used to decompose and dynamically adjust the logistics resource allocation plan and output a target matching result.

[0033] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to that described in the above method embodiment, and details will not be elaborated here.

[0034] Refer to Figure 3 , in the embodiment of the present invention, a computer device is further provided. This computer device can be a server, and its internal structure can be as Figure 3 shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface, and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non - volatile storage medium and an internal memory. The non - volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non - volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above - mentioned method is implemented.

[0035] Those skilled in the art can understand that Figure 3 the structure shown in

[0036] is only a block diagram of a part of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0037] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium provided by the present invention and used in the embodiments can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0038] It should be noted that in this article, the terms "include", "comprise", or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that includes a series of elements includes not only those elements but also other elements not expressly listed, or elements that are inherent to such process, apparatus, article, or method. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, apparatus, article, or method that includes such element.

[0039] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent structural or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A cross-border e-commerce logistics intelligent matching method, characterized in that: The following steps are involved: Pre-process the order data, transportation capacity data, and delivery record data collected by logistics nodes to generate logistics data sets; Inputting the logistics data set into a dual-stream feature extraction network for parallel feature calculation to generate a logistics spatiotemporal association matrix; Performing feature decomposition and feature recombination on the logistics spatiotemporal correlation matrix to obtain a logistics dynamic feature sequence; Inputting the logistics dynamic feature sequence into a hybrid attention network to perform node correlation weight analysis to generate a logistics matching feature vector; Constructing a multi-objective optimization function based on the logistics matching feature vector, and solving it to obtain a logistics resource allocation plan; The logistics resource allocation plan is decomposed and dynamically adjusted, and the target matching result is output.

2. The cross-border e-commerce logistics intelligent matching method according to claim 1 is characterized in that: The order data, transportation capacity data and delivery record data collected by the logistics nodes are pre-processed to generate a logistics data set, including: The order data collected by the logistics nodes are feature decomposed to obtain an order feature sequence including order number, order source, order destination, order weight, order time and order price; the transportation capacity data collected by the logistics nodes are feature decomposed to obtain a transportation capacity feature sequence including carrier number, transportation tool type, load capacity, transportation time and transportation cost; the distribution record data collected by the logistics nodes are feature decomposed to obtain a distribution feature sequence including historical distribution time, historical distribution route, historical distribution cost and historical service evaluation; Input the missing values ​​in the order feature sequence, the transport feature sequence and the delivery feature sequence into the time series interpolation model, obtain the completed data sequence by calculating the mean value of the data in the adjacent time windows, calculate the standard deviation of the completed data sequence, and replace the median of the abnormal values ​​exceeding the threshold interval by setting the threshold interval to obtain the cleaned data sequence; The cleaned data sequence is normalized to obtain a normalized feature sequence, and the normalized feature sequence is matrix reconstructed according to the order characteristics and capacity characteristics to generate a logistics data set including an order feature matrix and a capacity feature matrix, wherein the dimension of the order feature matrix is ​​m×n, m represents the order quantity, and n represents the order feature dimension; the dimension of the capacity feature matrix is ​​p×q, p represents the number of carriers, and q represents the capacity feature dimension.

3. The cross-border e-commerce logistics intelligent matching method according to claim 2 is characterized in that: The step of inputting the logistics data set into a dual-stream feature extraction network for parallel feature calculation to generate a logistics spatiotemporal association matrix includes: The order feature matrix in the logistics data set is input into the time series feature branch in the dual-stream feature extraction network for processing. The time series feature branch includes three layers of time series convolution modules. Each time series convolution module includes a one-dimensional convolution layer, a batch normalization layer, and a ReLU activation function. The convolution kernel sizes are 3, 5, and 7, respectively, to obtain the order time series feature tensor. Input the capacity feature matrix in the logistics data set into the spatial feature branch in the dual-stream feature extraction network for processing, wherein the spatial feature branch comprises three-layer spatial convolution modules, each of which comprises a two-dimensional convolution layer, a batch normalization layer and a ReLU activation function, and the convolution kernel size is 3×3, to obtain a capacity spatial feature tensor; Performing a temporal attention calculation on the order temporal feature tensor to obtain a temporal weight matrix, and performing a spatial attention calculation on the capacity spatial feature tensor to obtain a spatial weight matrix; Performing a tensor multiplication operation on the temporal weight matrix and the spatial weight matrix to obtain an initial correlation matrix, and inputting the initial correlation matrix into a feature fusion layer, wherein the feature fusion layer includes two fully connected layers and a Tanh activation function, and obtains a fused feature matrix through nonlinear transformation; The fusion feature matrix is ​​dimensionally reshaped and normalized to generate a logistics space-time association matrix, which includes the capacity distribution relationship and order delivery relationship between nodes. The matrix element aij of the logistics space-time association matrix represents the association strength between the i-th order node and the j-th capacity node.

4. The cross-border e-commerce logistics intelligent matching method according to claim 3 is characterized in that: The feature decomposition and feature recombination of the logistics spatiotemporal association matrix to obtain a logistics dynamic feature sequence includes: Performing multi-scale wavelet decomposition on the logistics space-time correlation matrix, and calculating the wavelet coefficients by using Daubechies wavelet basis functions to obtain a periodic characteristic matrix containing daily cycles, weekly cycles and monthly cycles; Performing singular value decomposition operation on the periodic characteristic matrix, obtaining the main period component matrix by calculating the eigenvalues ​​and eigenvectors, and inputting the logistics spatiotemporal correlation matrix into a moving average filter, and obtaining the trend characteristic matrix by calculating the weighted average value within the sliding time window; Performing least square fitting calculation on the trend characteristic matrix, obtaining a long-term trend function by constructing a polynomial regression model, and generating a trend component matrix according to the long-term trend function; Calculating the variance contribution of each period component in the main period component matrix, obtaining a period feature weight vector by normalizing the feature variance, and performing time series correlation analysis on the trend component matrix, obtaining a trend feature weight vector by calculating the autocorrelation coefficient; Inputting the period feature weight vector and the trend feature weight vector into an adaptive weight fusion device, and obtaining a fusion feature tensor through dynamic weighted calculation, wherein the adaptive weight fusion device includes a fully connected layer and a Softmax normalization layer; The fused feature tensor is reorganized in time sequence, and a logistics dynamic feature sequence is obtained by splicing feature vectors in time sequence.

5. The cross-border e-commerce logistics intelligent matching method according to claim 4 is characterized in that: The step of inputting the logistics dynamic feature sequence into a hybrid attention network to perform node correlation weight analysis to generate a logistics matching feature vector includes: The logistics dynamic feature sequence is input into a first feature separation module, and dual-channel feature extraction of the main branch and the bypass branch is performed on the feature sequence through a gated cycle unit to obtain a first feature matrix and a second feature matrix; Performing multi-layer self-attention calculation on the first feature matrix, calculating short-range spatial dependencies through the first attention head, and calculating long-range spatial dependencies through the second attention head, to obtain a multi-scale spatial attention vector; Performing hierarchical temporal attention calculation on the second feature matrix, extracting local temporal patterns through the first level, extracting global temporal patterns through the second level, and obtaining a hierarchical temporal attention vector; Inputting the multi-scale spatial attention vector into a first feature enhancement module, wherein the first feature enhancement module comprises a feedforward neural network and a skip connection, and obtains spatial enhancement features through residual learning; Inputting the hierarchical temporal attention vector into a second feature enhancement module, wherein the second feature enhancement module comprises a self-calibration gating unit and an adaptive normalization layer, and obtaining temporal enhancement features through a gating mechanism; The spatial enhancement features and the temporal enhancement features are cross-modal fused, and a fused feature tensor is obtained through an attention-guided feature recalibration mechanism. The fused feature tensor is input into a feature compression network for feature compression and feature calibration to generate a logistics matching feature vector.

6. The cross-border e-commerce logistics intelligent matching method according to claim 5 is characterized in that: The multi-objective optimization function is constructed based on the logistics matching feature vector, and the logistics resource allocation scheme is obtained by solving the multi-objective optimization function, including: A time efficiency optimization function is constructed based on the logistics matching feature vector, and a time efficiency loss function is obtained by calculating the sum of squares of the difference between the expected delivery time of the order and the actual delivery time. A cost optimization function is constructed based on the logistics matching feature vector, and a cost loss function is obtained by calculating the weighted sum of transportation cost, warehousing cost and labor cost. The time loss function and the cost loss function are weightedly combined to obtain a multi-objective optimization function by setting a balancing factor, and capacity constraints are added to the multi-objective optimization function, including the maximum load of the carrier, the maximum service radius and the time window limit, to generate a capacity constraint matrix; Input the multi-objective optimization function and the capacity constraint matrix into the Lagrange multiplier method solver, obtain the dual problem by constructing the Lagrangian function, perform sub-gradient iterative solution on the dual problem, and obtain the iterative update formula by calculating the KKT condition, wherein the iterative update formula includes two parts: the original variable update and the dual variable update; Perform optimization calculation according to the iterative update formula, perform cyclic iteration by setting a convergence threshold, and stop iteration when the objective function value changes less than the convergence threshold to obtain the optimal solution; The optimal solution is substituted into the decision variable mapping function, and the logistics resource allocation plan is obtained by calculation, including the order-carrier matching matrix and the transportation path matrix. The matrix element Bmn in the order-carrier matching matrix represents the probability of the mth order being assigned to the nth carrier, and the matrix element Ckl in the transportation path matrix represents the connection relationship from the kth node to the lth node in the path.

7. The cross-border e-commerce logistics intelligent matching method according to claim 6 is characterized in that: Decomposing and dynamically adjusting the logistics resource allocation plan and outputting the target matching result include: Performing task deconstruction processing on the logistics resource allocation plan, and grouping orders according to order destinations and timeliness requirements to obtain an order grouping matrix; A carrier matching model is constructed based on the order grouping matrix, and a comprehensive score is given to the carrier's load capacity, route coverage, and historical service quality to obtain a carrier score vector; Sorting and screening the carrier score vectors, determining the optimal carrier for each order group, and generating a carrier allocation instruction set, wherein the carrier allocation instruction set includes carrier scheduling priority, loading time window, and service level requirements; Building a path planning diagram based on the carrier allocation instruction set and the geographical location information of each node in the order set, and calculating the optimal path between the nodes to obtain an initial path plan, wherein the initial path plan includes a path sequence and an estimated travel time; Performing multi-objective optimization on the initial path plan, and jointly optimizing the total path length, time window constraints, and service quality to obtain a transportation path instruction set, wherein the transportation path instruction set includes detailed waypoints, arrival times, and operation requirements; The actual delivery time, route changes and service evaluation collected during the logistics execution process are input into the feedback analysis module, and the deviation value of each indicator is calculated through the sliding time window to obtain the execution deviation matrix; Based on the execution deviation matrix, the carrier allocation instruction set and the transportation path instruction set are updated online, and the carrier selection strategy and the path planning strategy are dynamically adjusted to obtain an optimized instruction set; The optimized instruction set is parsed and the execution status is tracked, the instruction execution progress and abnormal conditions are recorded through a real-time monitoring algorithm, and the target matching result is output.

8. A cross-border e-commerce logistics intelligent matching device, characterized in that: For implementing the steps of the method according to any one of claims 1 to 7, the device comprises: The preprocessing module is used to preprocess the order data, transportation data and delivery record data collected by the logistics nodes to generate a logistics data set; A calculation module, used for inputting the logistics data set into a dual-stream feature extraction network for parallel feature calculation to generate a logistics spatiotemporal association matrix; A decomposition module, used for performing feature decomposition and feature recombination on the logistics spatiotemporal association matrix to obtain a logistics dynamic feature sequence; An analysis module, used for inputting the logistics dynamic feature sequence into a hybrid attention network to perform node correlation weight analysis and generate a logistics matching feature vector; A solution module, used for constructing a multi-objective optimization function based on the logistics matching feature vector, and solving to obtain a logistics resource allocation plan; The output module is used to decompose and dynamically adjust the logistics resource allocation plan and output the target matching result.

9. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images