Cross-border logistics path optimization method and system based on multi-modal logistics information

Abstract

The invention relates to the technical field of cross-border logistics, in particular to a cross-border logistics path optimization method and system based on multi-modal logistics information, and the method comprises the steps: collecting multi-modal logistics information associated with an order, forming a multi-modal evidence package, and generating consistency measurement based on the multi-modal evidence package; feature coding and fusion are carried out on the multi-modal evidence packet, and a feasible constraint set is constructed and an order feasible edge set is determined according to cost parameters for generating a transportation section e and a node v; constructing a multi-objective optimization model, and solving an optimal path enabling an objective function to be minimum; and performing interpretability analysis on the optimal path, and outputting an interpretable result including cost decomposition, risk decomposition and a time efficiency confidence index. According to the invention, the multi-modal evidence packet is updated in real time and the abnormal intensity is quantitatively calculated in the transportation execution process, so that the dynamic optimization adjustment of the cross-border logistics path is realized; and the adaptation capability of path planning to cross-border logistics complex and changeable scenes is improved.

Description

Technical Field

[0001] This invention relates to the field of cross-border logistics technology, and in particular to a method and system for optimizing cross-border logistics routes based on multimodal logistics information. Background Technology

[0002] As a core link in international trade, cross-border logistics' route planning directly determines the efficiency, cost, and risk control level of logistics transportation. However, cross-border logistics scenarios are characterized by wide node distribution, diverse transportation modes, complex regulatory rules, and fragmented multi-source information. This necessitates a comprehensive consideration of multiple dimensions, including compliance, timeliness, economy, and risk, in route optimization. Current cross-border logistics route optimization methods are mostly based on spatiotemporal transportation networks. While they collect basic logistics information to build optimization models and solve for the optimal route, they fall short in utilizing and integrating multimodal logistics information. This makes it difficult to adapt to the actual scenario of multi-source information interaction across the entire cross-border logistics chain, and fails to provide comprehensive and accurate data source support for route optimization, resulting in a low degree of matching between optimization results and actual transportation needs.

[0003] In addition, cross-border logistics route optimization methods suffer from inherent defects of static planning, and the interpretability and feasibility of optimization results are insufficient. On the one hand, route planning is mostly a one-time static solution, which cannot respond in real time to sudden abnormal events such as trajectory deviation, sensor anomalies, and port congestion during transportation execution. It lacks a dynamic replanning mechanism, causing the original planned route to lose its optimality or even feasibility under abnormal scenarios. On the other hand, the optimization results mostly only output the combination of route nodes and transportation segments, without fine-grained decomposition of indicators such as cost, risk, and timeliness, nor quantifying the marginal contribution of each transportation segment to the optimization objective, resulting in "black box" decision results. Logistics operators find it difficult to accurately identify cost high points and risk points in the route, and cannot quickly complete route adjustments and operational coordination, which greatly reduces the actual value of the route optimization results. At present, there is a need for a cross-border logistics route optimization method and system based on multimodal logistics information. Summary of the Invention

[0004] To address the problems of insufficient utilization of multimodal information and poor interpretability and applicability of optimization results in existing cross-border logistics route optimization methods, this invention provides a cross-border logistics route optimization method and system based on multimodal logistics information.

[0005] Firstly, the present invention provides a cross-border logistics route optimization method based on multimodal logistics information, which adopts the following technical solution: A method for optimizing cross-border logistics routes based on multimodal logistics information includes: Collect multimodal logistics information associated with orders to form a multimodal evidence package. And based on multimodal evidence packages Generate consistency metrics ; Multimodal evidence packages Feature encoding and fusion are performed to obtain a unified representation vector. and according to Generate the cost parameters for transport segment e and node v; Based on compliance rules, shipping restrictions, order cut-off windows, and consistency metrics. Construct a set of feasible constraints And determine the feasible edge set for the order; Construct a multi-objective optimization model and use edge selection variables. The selection of transport segment e is characterized by satisfying network connectivity and the set of constraints. Under the given conditions, find the optimal path that minimizes the objective function; The optimal path is analyzed for interpretability, and the output is an interpretable result including cost decomposition, risk decomposition, and timeliness confidence index. The timeliness confidence index includes the quantile of total delay. Conditional Value at Risk ; Anomaly intensity based on multimodal information during transportation execution Triggering event-driven local replanning fixes the executed path segments and resolves them on the remaining subnetworks to obtain the updated path.

[0006] Furthermore, the multimodal evidence package Generate consistency metrics This includes performing unified spatiotemporal alignment processing on the collected multimodal logistics information, and constructing a system indexed by time t and based on order... Multimodal evidence package as the main body The multimodal logistics information includes structured fields, document text information, image information, spatiotemporal trajectory information, and IoT sensor information, based on the multimodal evidence package. Consistency constraints between document fields and observation fields are used to calculate and generate orders. Consistency measure at time t This is used to quantify the matching degree and data credibility of multi-source, multi-modal information. The expression for the consistency measure is: , in, It is the Sigmoid activation function. The consistency weight of the k-th core attribute field of the goods. This is the document field value for the k-th core attribute of the goods. For conflict event indication functions, Let m be the observed field value of the k-th core attribute of the goods, and m be the index of the multimodal information conflict event. This is the negative correction weight for the m-th multimodal information conflict event.

[0007] Furthermore, the multimodal evidence package Feature encoding and fusion, including... Different types and dimensions of raw multimodal information are converted into high-dimensional feature vectors with uniform dimensions, thus obtaining the feature vectors encoded by each modality. ,against Calculate the modal quality index separately The modal quality index At least including the missing rate of corresponding modal information, time lag, and noise intensity, based on The adaptive fusion weights for each modality are calculated using a soft maximization function. ,Will With the corresponding adaptive fusion weights By performing a weighted summation, a unified representation vector is obtained. The formula for calculating the unified representation vector is: , , in, , These are all modality indexes of multimodal evidence packages. Let m be the bias term for the m-th mode. For the adaptive fusion weights of the m-th modality, Let be the weight matrix for the m-th modal quality index.

[0008] Furthermore, the aforementioned according to The cost parameters for transport segment e and node v are generated, including collecting the real-time operating status of the spatiotemporal transport network G=(V,E), which consists of a cross-border node set V and a transport segment set E, and using a unified representation vector. Consistency measures The system associates features with real-time operational status, uses a nonlinear mapping model to model the association between multiple features and cost parameters, and finally models the probability distribution of various cost parameters for transport segment e and node v, representing the cost parameters as a parameterized form of a non-negative distribution family, thus generating a model related to the order. Dynamic cost parameters for strong correlations.

[0009] Furthermore, the construction of the feasible constraint set It also determines the feasible edge set of orders, including converting compliance rules, prohibited or restricted shipping constraints, and order cut-off time windows into constraint expressions, and constructing a feasible constraint set based on these constraint expressions. Consistency measures With feasible constraint set Perform correlation adaptation and evaluate the feasible constraint set according to preset threshold judgment rules. Dynamic adjustments are made by introducing a comprehensive feasibility indicator function to quantify the feasibility of each transportation segment e in the spatiotemporal transportation network G=(V,E), thereby selecting all sets that satisfy the feasible constraints. For all transport segments with all constraints, a set integration algorithm is used to aggregate all feasible transport segments to form an order. Dedicated order feasible edge set.

[0010] Furthermore, the construction of the multi-objective optimization model includes defining binary decision variables for edge selection, constructing the objective function of the multi-objective optimization model hierarchically based on cost parameters, wherein the objective function adopts a weighted summation form, transforming the three optimization objectives of cost, risk, and robustness into a single-objective optimization form, and integrating the feasible constraint set. The network connectivity constraints are then transformed into formal constraints suitable for a multi-objective optimization model, forming a complete multi-objective optimization model: , in, The comprehensive objective function of the multi-objective optimization model is... , and These are the weighting coefficients for the cost item, risk item, and robust timeliness item, respectively. For the cost term in the objective function, For the risk term in the objective function, This is the robust time-dependent term in the objective function.

[0011] Furthermore, finding the optimal path that minimizes the objective function includes pre-validating the constraints of the multi-objective optimization model, eliminating invalid constraints, and solving the multi-objective optimization model based on an improved genetic algorithm to obtain the decision variables that minimize the comprehensive objective function J. Combination, The transport segments are spliced ​​together in spatiotemporal order, and the optimal path from the origin node to the destination node is obtained through analysis. The network connectivity constraint expression of the multi-objective optimization model is: , , in, We select binary decision variables for the edges to characterize the selection of transport segment e. Let v be the set of transport segments from node v. Let V be the set of transport segments flowing into node v. For the set of feasible nodes for the order, For the order's origin and shipment point, The destination node for the order.

[0012] Furthermore, the interpretability analysis of the optimal path includes analyzing the optimal path... A hierarchical analysis is performed, and the cost of each transportation segment and node, as well as the total route cost, are quantitatively calculated based on cost parameters to obtain cost decomposition results. Based on comprehensive risk probability parameters, the single-type and comprehensive risk probabilities of each transportation segment are quantitatively calculated, and the total route risk level is summarized to form risk decomposition results. Finally, the timeliness robustness of the optimal route is calculated, generating a timeliness confidence index. The formula for calculating the conditional value of risk in the timeliness confidence index is as follows: , in, For the confidence level of timeliness, As an auxiliary variable, For mathematical expectation operators, Let be the random variable representing the total delay of the optimal path.

[0013] Furthermore, fixing the executed path segments and resolving them on the remaining subnetworks includes: Optimal path During execution, the multimodal evidence package is updated synchronously. Based on the updated multimodal evidence package Quantitative calculation of the anomaly intensity at time t ; Preset abnormal intensity threshold The anomaly intensity will be calculated in real time. With threshold When a comparison is performed, When this happens, event-driven local replanning is triggered. When this happens, the optimal path is maintained. Continue execution; After triggering local replanning, fix the completed transportation segments in the optimal path, remove the executed nodes and transportation segments, and construct the remaining subnetwork containing unexecuted nodes and unexecuted transportation segments based on the feasible edge set of the order; The solution is re-solved on the remaining subnetworks, and the updated path that adapts to the remaining transportation links is obtained by combining the real-time updated multimodal information and cost parameters.

[0014] Secondly, a cross-border logistics route optimization system based on multimodal logistics information includes: The data acquisition module is configured to collect multimodal logistics information associated with orders and form a multimodal evidence package. And based on multimodal evidence packages Generate consistency metrics ; The encoding fusion module is configured to: process multimodal evidence packages Feature encoding and fusion are performed to obtain a unified representation vector. and according to Generate the cost parameters for transport segment e and node v; The feasible edge set module is configured to be based on compliance rules, restricted / prohibited shipping constraints, order cut-off time windows, and consistency measures. Construct a set of feasible constraints And determine the feasible edge set for the order; The model module is configured to: build a multi-objective optimization model, using edge-selected variables. The selection of transport segment e is characterized by satisfying network connectivity and the set of constraints. Under the given conditions, find the optimal path that minimizes the objective function; The decomposition module is configured to: perform interpretability analysis on the optimal path and output interpretable results including cost decomposition, risk decomposition, and timeliness confidence indicators, wherein the timeliness confidence indicators include the quantile of total delay. Conditional Value at Risk ; The output module is configured to: measure the anomaly intensity based on multimodal information during the transportation process. Triggering event-driven local replanning fixes the executed path segments and resolves them on the remaining subnetworks to obtain the updated path.

[0015] In summary, the present invention has the following beneficial technical effects: 1. This invention collects multimodal logistics information associated with cross-border logistics orders and constructs a multimodal evidence package. At the same time, it generates a consistency measure that represents the credibility of the information. This invention achieves the standardized integration and credible quantification of multi-source heterogeneous information in cross-border logistics, providing comprehensive, accurate and verifiable data source support for route optimization. It effectively solves the problems of insufficient utilization of logistics information and lack of quantitative basis for data credibility in the prior art.

[0016] 2. This invention obtains a unified representation vector by encoding modality-specific features of multimodal evidence packages and adaptively fusing them based on quality indicators. It also dynamically generates cost parameters for transportation segments and nodes by combining the real-time operating status and consistency measurement of the spatiotemporal transportation network. This achieves a strong correlation between cost parameters and order attributes and real-time network status, solving the problems of using fixed empirical values ​​for cost parameters and insufficient dynamism and accuracy in existing technologies, and improving the scientific nature of path optimization parameter modeling.

[0017] 3. This invention constructs a feasible constraint set by integrating compliance rules, restrictions on transportation, order cut-off time windows, and consistency metrics. It then selects and determines the order-specific feasible edge set through a comprehensive feasibility indicator function. Furthermore, it can dynamically adjust the constraint set based on consistency metrics, thereby realizing formal modeling and dynamic adaptation of multi-dimensional constraints in cross-border logistics. This effectively avoids the path execution failure problem caused by information contradictions and data distortion, and significantly improves the compliance and practical feasibility of path planning.

[0018] 4. This invention constructs a multi-objective optimization model that integrates cost, risk, and robust timeliness, uses binary edge selection variables to represent the selection of transportation segments, and combines network connectivity constraints and feasible constraint sets to solve for the optimal path. This achieves a synergistic balance of multiple optimization objectives in cross-border logistics, taking into account the economy, risk, and timeliness robustness of the path. It solves the problem of single or simple weighted optimization objectives in existing technologies and improves the comprehensive benefits of path optimization results.

[0019] 5. This invention performs hierarchical interpretability analysis on the optimal path, outputting interpretable results including cost decomposition results, risk decomposition results, and timeliness confidence indicators. This enables the visualization and traceability of the path optimization decision-making logic, allowing logistics operators to accurately identify cost high points and risk points in the path. It solves the problem of black box optimization results in existing technologies and greatly improves the practical value and operational convenience of path optimization results.

[0020] 6. This invention achieves dynamic optimization and adjustment of cross-border logistics routes by updating multimodal evidence packages in real time and quantifying anomaly intensity during transportation execution, triggering event-driven local replanning based on anomaly intensity thresholds, fixing executed path segments, and resolving and updating paths on the remaining sub-networks. This ensures that the routes always remain optimal and feasible during transportation, solves the defect of static planning in existing technologies that cannot respond to sudden abnormal events, and improves the adaptability of route planning to complex and ever-changing cross-border logistics scenarios. Attached Figure Description

[0021] Figure 1 This is an overall schematic diagram of a cross-border logistics route optimization method based on multimodal logistics information according to an embodiment of the present invention.

[0022] Figure 2 This is an overall architecture diagram of a cross-border logistics route optimization method based on multimodal logistics information according to an embodiment of the present invention.

[0023] Figure 3 This is a diagram illustrating the implementation effect of the intelligent optimization algorithm in this embodiment of the invention. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to the accompanying drawings. Example

[0025] Reference Figure 1 This embodiment of a cross-border logistics route optimization method based on multimodal logistics information includes: S1. Collect multimodal logistics information associated with orders to form a multimodal evidence package. And based on multimodal evidence packages Generate consistency metrics ; S2, regarding multimodal evidence packages Feature encoding and fusion are performed to obtain a unified representation vector. and according to Generate the cost parameters for transport segment e and node v; S3, based on compliance rules, shipping restrictions, order cut-off windows, and consistency measurements. Construct a set of feasible constraints And determine the feasible edge set for the order; S4. Construct a multi-objective optimization model and use edge selection variables. The selection of transport segment e is characterized by satisfying network connectivity and the set of constraints. Under the given conditions, find the optimal path that minimizes the objective function; S5. Perform interpretability analysis on the optimal path, and output interpretable results including cost decomposition, risk decomposition, and timeliness confidence indicators. The timeliness confidence indicators include the quantiles of total delay. Conditional Value at Risk ; S6. Anomaly intensity based on multimodal information during transportation execution. Triggering event-driven local replanning fixes the executed path segments and resolves them on the remaining subnetworks to obtain the updated path.

[0026] Specifically, a cross-border logistics route optimization method based on multimodal logistics information includes the following steps: like Figure 1 , Figure 2 As shown, this invention is geared towards cross-border logistics orders. In this embodiment, path optimization is performed on the spatiotemporal transportation network G=(V,E), which consists of a set of cross-border nodes V and a set of transportation segments E. First, multimodal logistics information associated with orders is collected to form a multimodal evidence package. And based on multimodal evidence packages Generate consistency metrics ; Comprehensive data collection and cross-border logistics orders Strongly correlated multimodal logistics information, covering five core data types: structured, text, visual, spatiotemporal trajectory, and IoT sensing. Specifically, it includes structured fields, document text information, image information, spatiotemporal trajectory information, and IoT sensor information, where the structured field is the order... The basic attribute data includes at least the core information such as cargo category, weight, volume, departure point, destination point, and transportation timeliness requirements; document text information serves as the legal basis and operational voucher data for cross-border logistics operations, including at least the text information in documents such as customs declarations, bills of lading, and packing lists; image information consists of visual data of the actual cargo and operational scenarios, including at least images of cargo labels, shipping marks, packing status, and logistics node operational scenarios; spatiotemporal trajectory information is the spatiotemporal movement data of cargo and transport vehicles, including at least the real-time location, trajectory, and corresponding timestamp data of transport vehicles based on satellite positioning and base station positioning; IoT sensor information is real-time sensing data of cargo status and transportation environment, including at least sensor data collected by IoT sensing devices such as temperature, humidity, vibration, pressure, and location. The collection of the above multimodal logistics information covers cross-border logistics orders. The entire operational chain from shipment to delivery.

[0027] Secondly, the collected multi-source, multi-modal logistics information undergoes unified spatiotemporal alignment processing, based on orders. Using the logistics operation timeline as a unified time reference and the WGS-84 geographic coordinate system as a unified spatial reference, all types of multimodal logistics information are mapped to this unified spatiotemporal coordinate system. For modal information with missing timestamps during the collection process, a linear interpolation completion algorithm and cross-border logistics operation logic reasoning are used to accurately complete the timestamps, ensuring the consistency of the time dimension of each modal information. For modal information with missing spatial locations during the collection process, spatial location is completed by matching the spatial attributes of logistics nodes and transportation segments, ensuring the consistency of the spatial dimension of each modal information. After spatiotemporal unified alignment processing, all information related to orders... The associated multimodal logistics information is integrated into a structured data set indexed by time t, i.e., a multimodal evidence package. ,in, For structured field modal information, For document text information, For image information, For spatiotemporal trajectory information, For IoT sensing information.

[0028] Finally, based on the multimodal evidence package after spatiotemporal uniform alignment processing With consistency constraints between document fields and observation fields as the core, a consistency measure of cross-border logistics order O at time t is generated through quantitative calculation. This consistency metric Used to quantify the matching degree and data reliability of multi-source, multi-modal logistics information, its value ranges from [0,1]. The closer the value is to 1, the stronger the multi-modal evidence package. The higher the degree of matching of modal information and the stronger the data credibility, the closer the value is to 0, indicating a multimodal evidence package. The more inconsistencies and distortions exist in the data, the lower the credibility of the information; consistency measurement The specific calculation is achieved through the following formula: , in, It is the Sigmoid activation function. The consistency weight of the k-th core attribute field of the goods. This is the document field value for the k-th core attribute of the goods, and this value is directly taken from the multimodal evidence package. The document text information in the document serves as legally binding documentation for cross-border logistics operations. This is an indicator function, a binary function. The indicator function evaluates to 1 when the condition within the parentheses is true, and evaluates to 0 when the condition within the parentheses is false. This is the observed field value of the k-th core attribute of the goods, taken from the multimodal evidence package. The actual observation data obtained through image recognition and analysis, equipment scanning, and IoT sensor data inference serves as a true representation of the physical state of the goods. This is a matching indicator function for document fields and observed fields. When the value of a document field for a core attribute of a commodity matches the value of an observed field, the indicator function takes a value of 1, indicating that the information in that attribute field matches. When they do not match, the indicator function takes a value of 0, indicating that the information in that attribute field is contradictory. m is the index of the multimodal information conflict event, which takes a positive integer value. The multimodal evidence package is traversed. All multi-source information conflict events identified in the data are considered as various information conflicts, including spatiotemporal feature contradictions between different modal information, mismatches between sensor information and physical state, and mismatches between trajectory information and operation nodes, in addition to mismatches between document fields and observation fields. This is the negative correction weight for the m-th multimodal information conflict event, and its value is a positive real number. This is a conflict event indicator function. When the m-th multimodal information conflict event actually occurs, the indicator function takes a value of 1, and when the conflict event does not occur, the indicator function takes a value of 0.

[0029] In the specific calculation process, the first step is to calculate the weighted sum of the document and observation matching results for all core cargo attribute fields, i.e. Then calculate the negative-corrected weighted sum of all multimodal information conflict events, i.e. Then, the weighted sum of the matching results is subtracted from the weighted sum of the negative correction to obtain the original calculation result of the consistency measure. Finally, this original calculation result is input into the Sigmoid activation function for nonlinear mapping to obtain the final consistency measure. The value is used to quantitatively characterize the credibility of the multimodal evidence package xo,t.

[0030] S2, regarding multimodal evidence packages Feature encoding and fusion are performed to obtain a unified representation vector. and according to Generate the cost parameters for transport segment e and node v; First, execute the multimodal evidence package. The feature encoding operation employs a dedicated encoding technique for the information characteristics of each modality, transforming the original multimodal information of different types and dimensions into high-dimensional feature vectors with unified dimensions. , For modal indexing, all encoded feature vectors are uniformly dimensional to d, with d ranging from 512 to 2048. This dimensionality can be configured according to actual computing power and accuracy requirements. For structured field modal... First, continuous fields are mapped to the [0,1] interval using the min-max normalization method. Discrete fields (such as product category, node type, and transportation mode) are converted into binary vectors using one-hot encoding. Then, the normalized continuous features are concatenated with the one-hot encoded discrete features to obtain the structured encoded feature vector. ; For document text information modality A pre-trained BERT-based model is used as the encoding backbone. The extracted document text is segmented into 128-character segments. Semantic features are extracted through the model's embedding and encoder layers. Finally, mean pooling is applied to the hidden states output by the encoder to obtain a fixed-dimensional text encoding feature vector. For image information modalities A pre-trained ResNet50 model with the last fully connected layer removed was used as the feature extractor. The acquired images were first normalized in size (uniformly adjusted to 224×224 pixels) and pixel values ​​were standardized before being input into the model for forward propagation. The output of the average pooling layer was used as the image encoding feature vector. ; For spatiotemporal trajectory information modes The method employs a spatiotemporal graph neural network (ST-GNN) for encoding, constructing a spatiotemporal graph of trajectory points according to time series. Nodes represent the location and time information of trajectory points, while edges represent the temporal relationships between adjacent trajectory points. Spatial features are captured through graph convolutional layers, and temporal features are captured through temporal convolutional layers. Finally, global pooling is used to obtain the spatiotemporal trajectory encoding feature vector. For IoT sensing information modes The system employs a Temporal Convolutional Network (TCN) for encoding, dividing the sensor data into fixed-length temporal sequences according to time windows. Temporal features are extracted through stacked 1D convolutional layers, batch normalization layers, and the ReLU activation function. Finally, global average pooling is used to obtain the sensor encoding feature vector. .

[0031] After completing the feature encoding for each modality, the encoded feature vector for each modality is... Calculate modal quality index , The vector is a three-dimensional vector, corresponding to three core dimensions: missing rate, time lag, and noise intensity of modality information. The calculation methods for each dimension are as follows: The missing rate is calculated by the complement of the ratio of the effective data volume of the modality within the current time window to the total amount of data to be collected, i.e., missing rate = 1 − (effective data volume / total amount of data to be collected). The effective data volume is the number of data entries after data cleaning to remove outliers and null values. The total amount of data to be collected is the theoretical number of data entries to be collected for the modality within the time window. The missing rate ranges from [0,1], and the closer it is to 0, the better the integrity of the modality data; the time lag is calculated by the latest effective data entry of the modality. The difference between the timestamp of the data and the current time t is calculated in minutes and then normalized to the [0,1] interval using a linear mapping. The closer the time lag value is to 0, the better the real-time performance of the modal data. For numerical sensor data (such as temperature, humidity, and vibration values), noise intensity is calculated using the coefficient of variation (standard deviation / mean). For non-numerical data such as text and images, it is calculated using the data distribution entropy value. The results are then normalized to the [0,1] interval. The closer the noise intensity value is to 0, the higher the purity of the modal data. Finally, the normalized missing rate, time lag, and noise intensity are concatenated in sequence to obtain the modal quality index. .

[0032] Based on modal quality indicators The adaptive fusion weights for each modality are calculated using a softmax function. The calculation formula is: , where j is the modal index (which has the same range as m). This is the weight matrix for the m-th modal quality index, with dimensions 3×1, used to... The three dimensions of the metrics are weighted, and the initial value of wm is generated by random normal distribution and subsequently optimized through model training. Let be the bias term of the m-th mode, which is a scalar and is initially set to 0. It is also optimized through training. exp(⋅) is the natural exponential function, which is used to convert the result of the linear combination into a non-negative value.

[0033] Obtain the adaptive fusion weights for each modality Then, the encoded feature vector of each modality is... With the corresponding By performing a weighted summation, a unified representation vector is obtained. The calculation formula is: ,in, The dimension of the vector is the same as that of the encoded feature vector, which is d-dimensional. This vector fully integrates the effective information of the five modalities, and at the same time, it suppresses the interference of low-quality modalities through weight adjustment, thereby achieving the complementarity and enhancement of multimodal information.

[0034] Collection of spatiotemporal transportation networks Real-time running status, It consists of a cross-border node set V (including ports of origin, transit hubs, clearance points, ports of destination, etc.) and a transportation segment set E (including transportation links such as sea, air, and land transportation). The real-time operating status includes the congestion level and equipment operating status of node v, and the vehicle status of transportation segment e. All real-time status data are collected at a frequency of minutes through the API interface of the logistics network monitoring system and converted into standardized numerical feature vectors after collection.

[0035] Next, the unified representation vector will be used. Consistency measures The real-time running state feature vector is associated with the above-mentioned feature vector. Specifically, the three are integrated into a joint feature vector of dimension d+1+k through a concatenation operation, where k is the dimension of the real-time running state feature. This joint feature vector is then input into a nonlinear mapping model to realize the association model between multiple features and cost parameters. The nonlinear mapping model is constructed using a three-layer fully connected neural network. The input layer dimension is d+1+k, the first hidden layer has a height of 2(d+1+k), the second hidden layer has a height of (d+1+k), and the output layer dimension is determined according to the type of cost parameter. The activation function is the ReLU function, and no activation function is used in the output layer to retain continuous value output.

[0036] Finally, probability distribution models are performed on various cost parameters for transportation segment e and node v, representing the cost parameters as parameterized forms of a non-negative distribution family. Dynamic cost parameters strongly correlated with order O are generated, including the transportation cost of transportation segment e. Transportation delay distribution Comprehensive risk probability and the operation cost of node v Customs clearance / inspection time delay distribution Regarding transportation costs Operating costs The output value of the nonlinear mapping model is directly used as the deterministic dynamic parameter, retaining two decimal places; for the transportation delay distribution Distribution of customs clearance / inspection delays The gamma distribution is used for parameterization, i.e. , where shape parameters with rate parameter All outputs are obtained through a nonlinear mapping model and satisfy the following conditions: The Gamma distribution can accurately characterize the nonnegativity and right skewness of time delay; for comprehensive risk probability The Sigmoid function is used to map the output value of the nonlinear mapping model to the interval [0,1], and the risk probability is... It includes at least the combined probability of delay risk, temperature control failure risk, and loss risk, and sets configurable weights for different risk types to form a comprehensive risk probability.

[0037] , in, For risk type indexing, To avoid the risk of delay, To mitigate the risk of temperature exceeding limits, To mitigate the risk of loss, Configurable weights for the r-th type of risk, Let be the probability of a single type of risk occurring in order O on transportation segment e, representing type r risk.

[0038] S3, based on compliance rules, shipping restrictions, order cut-off windows, and consistency measurements. Construct a set of feasible constraints And determine the feasible edge set for the order; The compliance rules, restrictions on transport, and order cut-off time windows are each converted into quantifiable constraint expressions, forming a set of feasible constraints. The foundation for its construction lies in the following: Compliance rules and regulations cover import and export tariff rules, trade control rules, customs clearance documentation requirements, and cargo inspection and quarantine rules. Regarding tariff rules, they are based on order... Based on the categories of goods, declared value, and tariff rates of the country of origin / destination, construct a tariff compliance constraint expression. ,in, For orders The amount of customs duty payable when transported via transport segment e is calculated by multiplying the declared value by the corresponding customs duty rate. For orders The upper limit of tariffs payable is based on the maximum tariff reductions available under trade agreements; constraints are established for determining controlled goods in accordance with trade control rules. ,in, For the list of prohibited / controlled goods published by the country of origin / destination, For indicator functions, when the cargo attribute The function value is 1 if the item belongs to the manifest, and 0 otherwise; the constraint requires that the function value must be 0. A document completeness constraint is constructed based on customs clearance document requirements. ,in This is a set of legally required documents for the corresponding type of goods. The set of documents actually provided for order O.

[0039] For restrictions on transport, a feasibility constraint expression for such restrictions is constructed based on the attributes of the goods and the regulatory requirements of the transport segment. ,in It includes core attributes such as cargo category, cargo type, hazard class, and temperature control requirements. The regulatory requirements for transport segment e include the types of goods that can be transported, restrictions on the types of goods, and transport condition requirements. For logical satisfaction symbols, when the cargo attributes meet the regulatory requirements of the transport segment, the indicator function... The value is 1 if it is not 1, and 0 otherwise. For the order cut-off time window constraint, a feasibility constraint expression for the time window is constructed by combining the transport segment's shift schedule and order operation plan. ,in Let t be the time window formed by the cut-off time of transportation segment e and the departure time of the scheduled transport. Let t be the planned operation time of order O in transportation segment e. When the planned operation time falls within the time window, the indicator function... The value is 1 if it is not 0 otherwise.

[0040] After completing the formal transformation of constraints in each dimension, all constraint expressions are integrated to construct a basic set of feasible constraints. Then, the consistency metric will be used. It is associated and adapted with the set of feasible constraints, and the set of constraints is dynamically adjusted according to preset threshold judgment rules to form the final set of feasible constraints. Preset consistency metric threshold (Value range is 0.6-0.8), when This indicates high consistency of multimodal information, strong data reliability, and maintenance of the basic feasible constraint set. Unchanged; when When this occurs, it indicates that there are contradictions in the information and low data reliability, requiring stronger constraints to reduce the risk of path execution failure. In this case, the regulatory requirements for prohibited or restricted transport constraints need to be upgraded, narrowing the range of permissible cargo types and improving the accuracy of transport condition matching. The corresponding constraint expressions need to be adjusted. ,in To meet the upgraded regulatory requirements At the same time, the constraints on the order cut-off time window are tightened, and the time window is... Adjusted to ,in, To tighten the time window duration, the corresponding constraint expression is adjusted as follows: The adjusted constraint set is .

[0041] To achieve segment-by-segment feasibility quantification for transportation sections, a comprehensive feasibility indicator function is introduced: ; in The product of indicator functions for all compliance rule constraints is given; the product term is 1 if all compliance rule constraints are satisfied, and 0 otherwise; (Comprehensive feasibility indicator function) The value of is determined by the satisfaction of constraints in each dimension. The function takes the value of 1 only when the restrictions on transport, time window constraints and all compliance rule constraints are satisfied, otherwise it takes the value of 0. This implements the logical AND operation of multi-dimensional constraints and ensures the strictness of constraint judgment.

[0042] Based on the comprehensive feasibility indicator function, the feasibility of each transportation segment e in the spatiotemporal transportation network G=(V,E) is quantitatively determined segment by segment: traversing all transportation segments in the set of transportation segments E, for each transportation segment e, it is sequentially verified whether it satisfies the set of feasible constraints. Calculate the comprehensive feasibility indicator function based on all constraints. The possible values; filter out all The feasible transport segments are those that satisfy all constraints. Then, a set integration algorithm is used to aggregate all feasible transport segments. This algorithm employs a hash table-based fast deduplication and classification strategy, constructing a hash table using the transport segment code as the key. Each selected feasible transport segment is stored in the hash table, automatically removing duplicates. Finally, the set of transport segments stored in the hash table represents the feasible edge set specific to order O. Meanwhile, based on feasible edge sets Extract all nodes connected to feasible transport segments to form the feasible node set of order O. .

[0043] S4. Construct a multi-objective optimization model and use edge selection variables. The selection of transport segment e is characterized by satisfying network connectivity and the set of constraints. Under the given conditions, find the optimal path that minimizes the objective function; Define edge selection binary decision variables This variable only applies to the feasible edge set of orders. The inner value is used to accurately represent the selection of transport segment e, and it is defined as follows: , among which when When, it indicates that the optimization model selects transportation segment e as a component of the path of order O; when When the value is 'e', ​​it means that the optimization model does not select transportation segment e as a component of the path of order O. By defining this decision variable, it is ensured that the subsequent model solution is only performed within the range of feasible transportation segments that meet the constraints, thus avoiding the generation of infeasible paths from the source.

[0044] Next, based on the hierarchical construction of transportation network cost parameters, a multi-objective optimization model objective function is constructed. The objective function adopts a weighted summation form, transforming the three optimization objectives of cost, risk, and robust timeliness into a single-objective optimization form to achieve a synergistic balance among multiple objectives. Its expression is: , in, The comprehensive objective function of the multi-objective optimization model is... , and These are the weighting coefficients for the cost item, risk item, and robust timeliness item, respectively. The cost term in the objective function represents the total logistics cost of the route. It is calculated as the sum of the costs of all selected transportation segments and node operation costs in the route. This is the risk term in the objective function, representing the overall risk probability of the route. It is calculated as the weighted sum of the overall risk probabilities of all selected transportation segments within the route. For the robust time-dependent term in the objective function, conditional value of risk is used. Characterization, used to quantify the extreme delay risk of a path. ,in, For the confidence level of timeliness, As an auxiliary variable, For mathematical expectation operators, Let be the random variable representing the total delay of the optimal path.

[0045] After constructing the objective function, integrate the set of feasible constraints. The network connectivity constraints are then transformed into formal constraints suitable for a multi-objective optimization model, forming a complete multi-objective optimization model. The network connectivity constraints include flow conservation constraints and origin / destination node constraints. The expression for the flow conservation constraint is: , , in, We select binary decision variables for the edges to characterize the selection of transport segment e. Let v be the set of transport segments from node v. Let V be the set of transport segments flowing into node v. For the set of feasible nodes for the order, For the order's origin and shipment point, For the destination node of the order, this constraint ensures that for any intermediate node in the path, the number of inbound transport segments is equal to the number of outbound transport segments, guaranteeing the connectivity of the path; the expression for the origin and destination node constraints is as follows: and Origin node The number of outbound transport segments is 1, and the destination node is... The number of inflow transport segments is 1, ensuring that the path starts from the origin node and eventually reaches the destination node. Simultaneously, the feasible constraint set is... The compliance rules constraints, prohibited and restricted transport constraints, and cut-off time window constraints are transformed into formal constraints and incorporated into the model. The compliance rules constraints are implemented by verifying whether transport segment e meets requirements such as tariff compliance, trade control, and document completeness. The prohibited and restricted transport constraints and the cut-off time window constraints are implemented through a comprehensive feasibility indicator function. Constraints are applied to ensure that the model selects transport segments only within the feasible edge set EO. ​​Additionally, the model includes probabilistic service level constraints, expressed as follows: Where P(⋅) is a probability operator, For order O, the service time threshold. For the timeliness confidence level, this constraint guarantees that the total delay of the optimal path is within the confidence level. The timeliness must not exceed the service timeliness threshold to meet the timeliness requirements of the order.

[0046] like Figure 3 As shown, after constraint verification, the multi-objective optimization model is solved based on the improved intelligent optimization algorithm. An appropriate solution strategy is selected for the scale of the cross-border logistics spatiotemporal transportation network: For small-scale networks with fewer than 50 nodes and fewer than 200 transportation segments, the branch and bound method is used for accurate solution. By continuously branching and dividing the feasible region and calculating the lower bound of each branch, non-optimal solution regions are gradually eliminated, and finally, the global optimal solution is obtained. For large-scale networks with more than or equal to 50 nodes and more than or equal to 200 transportation segments, an improved genetic algorithm is used for solution. This algorithm achieves optimal solution search by simulating the biological evolution process. The specific steps are as follows: First, population initialization is performed, randomly generating N initial individuals, each individual having a length equal to the feasible edge set. The number of transport segments is represented by a binary vector. The elements in the vector are either 0 or 1, which correspond to the unselected and selected states of the transport segments, respectively. Each individual segment must satisfy network connectivity constraints and origin-destination node constraints. The value of N ranges from 100 to 500 and can be adjusted according to the network size. The second step is to calculate the fitness function. The reciprocal of the comprehensive objective function J of the multi-objective optimization model is used as the fitness function value, that is, fitness value = 1 / J. The higher the fitness value, the better the path corresponding to the individual. The third step is to perform a selection operation, using the roulette wheel selection method. The probability of an individual being selected is determined based on its fitness value as a percentage of its population. Individuals with higher fitness values ​​have a greater probability of being selected. At the same time, the best individual in the population is retained (elite retention strategy) to avoid losing the optimal solution. The fourth step is to perform a crossover operation, using a single-point crossover strategy. A crossover point is randomly selected, and the gene segments after the crossover point of the two parent individuals are exchanged to generate offspring individuals. The crossover probability is set to 0.6-0.8. The fifth step involves mutation, where each gene locus in the offspring is flipped with a certain probability (0 becomes 1, 1 becomes 0), and the mutation probability is set to 0.01-0.05 to maintain population diversity. The sixth step is to determine whether the termination condition has been met. The termination condition is that the number of iterations reaches the preset maximum number of iterations (the value range is 200-500) or the fitness value has not significantly improved for 20 consecutive generations. If the termination condition is met, the iteration stops and the best individual in the population is output. Otherwise, return to the second step to continue the iteration.

[0047] Solving for the decision variables that minimize the comprehensive objective function J yields the answer. After combination, The transport segments are spliced ​​together in spatiotemporal order, and the data is analyzed to obtain the sequence from the origin node. to the destination node Optimal path During the assembly process, the starting point is used as the first reference. Starting from, find all And the starting point is The first transport segment of the route is determined; then, using the end node of that transport segment as the current node, all transport segments are searched. The starting point is the current node's transport segment, which becomes the next transport segment on the path; this process is repeated until the current node becomes the destination node. Complete the splicing of all transportation segments to form a complete optimal route. .

[0048] S5. Perform interpretability analysis on the optimal path, and output interpretable results including cost decomposition, risk decomposition, and timeliness confidence indicators. The timeliness confidence indicators include the quantiles of total delay. Conditional Value at Risk ; First traverse the optimal path All For each transportation segment e, extract the transportation cost. The transportation cost is broken down into its components, and the proportion of each transportation segment's cost to the total route transportation cost is calculated. Then, the operating costs of all nodes v along the route are extracted. Break down the costs of each activity into its components, calculate the proportion of each node's activity cost to the total cost of all nodes in the route, and then sum up the costs of all transportation segments and the activity costs of each node to obtain the total cost of the route. Furthermore, the total cost is broken down by transportation mode (sea freight, air freight, land freight, multimodal transport) and logistics links (departure, transshipment, customs clearance, delivery), and the cost proportion of different transportation modes and links is calculated to form a complete cost breakdown result. After completing the cost decomposition, risk decomposition is performed, based on the comprehensive risk probability generated by S2. and single-type risk probability This achieves a two-dimensional breakdown of risk, first traversing the optimal path. For each transport segment e in the data, extract the comprehensive risk probability of that transport segment. The process involves decomposing the risks of delays, temperature exceeding limits, and loss into individual probabilities and their corresponding weights. Then, the proportion of the overall risk probability of each transport segment to the total route risk is calculated, and the transport segments are classified into risk levels based on their overall risk probabilities. Finally, the overall risk probabilities of all transport segments are summed, and a weighted summation method is used to calculate the total route risk level. That is, by using the proportion of transportation segment costs as weights, the risks are weighted and aggregated to form a risk decomposition result.

[0049] After risk decomposition, the timeliness robustness of the optimal path is calculated, and a timeliness confidence index is generated. The timeliness confidence index includes the quantile of the total delay. Conditional Value at Risk All calculations are based on the transportation delay distribution and customs clearance / inspection delay distribution generated by S2. In addition, to quantify the marginal contribution of each transportation segment to the overall objective, the calculations for each transportation segment are performed. marginal contribution value ,in , , The weighting coefficients set in S4 The total path delay includes and excludes this transport segment. The difference, The smaller the value, the greater the positive contribution of the transportation segment to the overall objective function. Finally, the cost decomposition results, risk decomposition results, timeliness confidence index, and quantitative results of the marginal contribution of the transportation segment are integrated.

[0050] S6. Anomaly intensity based on multimodal information during transportation execution. Triggering event-driven local replanning fixes the executed path segments and resolves them on the remaining subnetworks to obtain the updated path.

[0051] In the optimal path During the transportation process, the multimodal evidence package is updated synchronously according to the dynamically adjusted collection frequency. The data collection frequency is adapted to the characteristics of the logistics process: data is collected hourly during transportation and minutely during key operations such as customs clearance and transshipment. The collection scope is consistent with S1, covering structured fields, document text information, image information, spatiotemporal trajectory information, and IoT sensor information. After collection, the spatiotemporal alignment process of S1 is repeated to uniformly map each modal of information to the WGS-84 geographic coordinate system and the order logistics operation timeline, generating a real-time updated multimodal evidence package. Meanwhile, the consistency metric is recalculated based on the updated evidence package to ensure the timeliness and consistency of the data.

[0052] Based on the updated multimodal evidence package Quantitatively calculate the anomaly intensity at the current time t. The anomaly intensity comprehensively considers three core dimensions: trajectory deviation, sensor anomalies, and external events. The calculation process is as follows: First, extract the current actual spatiotemporal trajectory of the order and the optimal path planning trajectory. Calculate the straight-line distance between the actual trajectory point and the corresponding planned trajectory point using the spatial distance formula. Combine this with the time difference to calculate the trajectory deviation. The larger the deviation, the more severe the deviation from the planned path. Second, extract real-time data collected by IoT sensor devices and compare it with a preset normal threshold range. Calculate the deviation amplitude and duration exceeding the threshold to obtain the sensor anomaly quantity. The larger the deviation amplitude and the longer the duration, the more severe the sensor anomaly. Next, collect external event information within the current time window, including abnormal weather, sudden changes in road / sea / air conditions, port shutdowns, and trade policy adjustments. Assign different weights according to the impact of the event on transportation, and sum the weighted values ​​to obtain the external event intensity. Finally, multiply the trajectory deviation, sensor anomaly quantity, and external event intensity by preset weight coefficients and sum them to obtain the anomaly intensity. Its value ranges from [0,10], and the larger the value, the more severe the abnormality.

[0053] Preset abnormal intensity threshold This threshold can be flexibly configured based on the mode of transport, cargo attributes, and logistics scenario, with a value range of [3,6]. For example, due to the significant impact of sea conditions on sea transport, the threshold can be set to 5; for urgent air shipments due to high time requirements, the threshold can be set to 3. The real-time calculated anomaly intensity will be used to determine the threshold. With threshold Perform real-time comparison: when When a significant anomaly is detected during transportation that may cause the original optimal path to lose its optimality or feasibility, event-driven local replanning is immediately triggered; when If the transportation process is deemed normal, the original optimal route should be maintained. Continue execution while continuously monitoring the updates of multimodal information.

[0054] After triggering a local replanning, the first step is to fix the completed transportation segments in the optimal path. By verifying logistics operation records and real-time location data, the segment numbers, corresponding origin and destination nodes, and completion times of the completed transportation segments are identified. These transportation segments are then locked as unchangeable executed parts of the path to ensure the continuity of the transportation process. Subsequently, executed nodes and transportation segments are removed, i.e., from the set of feasible nodes for the order. Remove nodes that have completed tasks from the order's feasible edge set. Remove executed transport segments from the network, retaining only unexecuted nodes and transport segments; based on the remaining unexecuted nodes and unexecuted transport segments, and in conjunction with the constraints of the order feasible edge set, construct the remaining spatiotemporal transport subnetwork. The remaining node set Includes the current node of the order (i.e., the endpoint of the previous executed transport segment), the destination node, and all feasible nodes located between the current node and the destination node, and the set of remaining transport segments. Includes all start and end nodes. Furthermore, for feasible transport segments that have not been eliminated, this sub-network focuses on incomplete transport links and is much smaller in scale than the original spatiotemporal transport network, which can significantly improve the solution efficiency of replanning.

[0055] In the remaining subnetwork The solution is then re-solved, first by incorporating a real-time updated multimodal evidence package. With consistency metrics Repeat step S2 to update the cost parameters, including the transportation cost, transportation delay distribution, and overall risk probability of unexecuted transportation segments, as well as the operation cost and customs clearance / inspection delay distribution of unexecuted nodes, to ensure that the cost parameters match the real-time transportation status. Then, repeat the constraint construction process of S3 to adjust the feasible constraint set based on the updated consistency metric, real-time monitoring rules, and schedule changes. The remaining subnetworks are selected for their viable edges. Then, using the multi-objective optimization model framework of S4, the weights of cost, risk, and robust timeliness are dynamically adjusted based on current transportation demand. For example, if the replanning is triggered by timeliness delay risk, the weight of the robust timeliness term can be appropriately increased. Finally, the same solution strategy as S4 is adopted: a branch-and-bound method is used for precise solutions in small-scale remaining subnetworks, while an improved genetic algorithm is used for fast solutions in large-scale remaining subnetworks, yielding locally optimal paths that fit the remaining transportation links. These locally optimal paths are then concatenated with the executed transportation segments in spatiotemporal order to form a complete updated path. The data is then pushed to the logistics scheduling system and transportation execution terminal in real time to guide subsequent transportation operations. Example

[0056] The difference between this embodiment and Embodiment 1 is that this embodiment provides a cross-border logistics route optimization system based on multimodal logistics information, including: The data acquisition module is configured to collect multimodal logistics information associated with orders and form a multimodal evidence package. And based on multimodal evidence packages Generate consistency metrics ; The encoding fusion module is configured to: process multimodal evidence packages Feature encoding and fusion are performed to obtain a unified representation vector. and according to Generate the cost parameters for transport segment e and node v; The feasible edge set module is configured to be based on compliance rules, restricted / prohibited shipping constraints, order cut-off time windows, and consistency measures. Construct a set of feasible constraints And determine the feasible edge set for the order; The model module is configured to: build a multi-objective optimization model, using edge-selected variables. The selection of transport segment e is characterized by satisfying network connectivity and the set of constraints. Under the given conditions, find the optimal path that minimizes the objective function; The decomposition module is configured to: perform interpretability analysis on the optimal path and output interpretable results including cost decomposition, risk decomposition, and timeliness confidence indicators, wherein the timeliness confidence indicators include the quantile of total delay. Conditional Value at Risk ; The output module is configured to: measure the anomaly intensity based on multimodal information during the transportation process. Triggering event-driven local replanning fixes the executed path segments and resolves them on the remaining subnetworks to obtain the updated path.

[0057] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing cross-border logistics routes based on multimodal logistics information, characterized in that, include: Collect multimodal logistics information associated with orders to form a multimodal evidence package. And based on multimodal evidence packages Generate consistency metrics ; Multimodal evidence packages Feature encoding and fusion are performed to obtain a unified representation vector. and according to Generate the cost parameters for transport segment e and node v; Based on compliance rules, shipping restrictions, order cut-off windows, and consistency metrics. Construct a set of feasible constraints And determine the feasible edge set for the order; Construct a multi-objective optimization model and use edge selection variables. The selection of transport segment e is characterized by satisfying network connectivity and the set of constraints. Under the given conditions, find the optimal path that minimizes the objective function; The optimal path is analyzed for interpretability, and the output is an interpretable result including cost decomposition, risk decomposition, and timeliness confidence index. The timeliness confidence index includes the quantile of total delay. Conditional Value at Risk ; Anomaly intensity based on multimodal information during transportation execution Triggering event-driven local replanning fixes the executed path segments and resolves them on the remaining subnetworks to obtain the updated path.

2. The method for optimizing cross-border logistics routes based on multimodal logistics information according to claim 1, characterized in that, The multimodal evidence package Generate consistency metrics This includes performing unified spatiotemporal alignment processing on the collected multimodal logistics information, and constructing a system indexed by time t and based on order... Multimodal evidence package as the main body The multimodal logistics information includes structured fields, document text information, image information, spatiotemporal trajectory information, and IoT sensor information, based on the multimodal evidence package. Consistency constraints between document fields and observation fields are used to calculate and generate orders. Consistency measure at time t This is used to quantify the matching degree and data credibility of multi-source, multi-modal information. The expression for the consistency measure is: ， in, It is the Sigmoid activation function. The consistency weight of the k-th core attribute field of the goods. This is the document field value for the k-th core attribute of the goods. For conflict event indication functions, Let m be the observed field value of the k-th core attribute of the goods, and m be the index of the multimodal information conflict event. This is the negative correction weight for the m-th multimodal information conflict event.

3. The method for optimizing cross-border logistics routes based on multimodal logistics information according to claim 1, characterized in that, The multimodal evidence package Feature encoding and fusion, including... Different types and dimensions of raw multimodal information are converted into high-dimensional feature vectors with uniform dimensions, thus obtaining the feature vectors encoded by each modality. ,against Calculate the modal quality index separately The modal quality index At least including the missing rate of corresponding modal information, time lag, and noise intensity, based on The adaptive fusion weights for each modality are calculated using a soft maximization function. ,Will With the corresponding adaptive fusion weights By performing a weighted summation, a unified representation vector is obtained. The formula for calculating the unified representation vector is: ， ， in, , These are all modality indexes of multimodal evidence packages. Let m be the bias term for the m-th mode. For the adaptive fusion weights of the m-th modality, Let be the weight matrix for the m-th modal quality index.

4. The method for optimizing cross-border logistics routes based on multimodal logistics information according to claim 1, characterized in that, According to The cost parameters for transport segment e and node v are generated, including collecting the real-time operating status of the spatiotemporal transport network G=(V,E), which consists of a cross-border node set V and a transport segment set E, and using a unified representation vector. Consistency measures The system associates features with real-time operational status, uses a nonlinear mapping model to model the association between multiple features and cost parameters, and finally models the probability distribution of various cost parameters for transport segment e and node v, representing the cost parameters as a parameterized form of a non-negative distribution family, thus generating a model related to the order. Dynamic cost parameters for strong correlations.

5. The method for optimizing cross-border logistics routes based on multimodal logistics information according to claim 1, characterized in that, The construction of feasible constraint set It also determines the feasible edge set of orders, including converting compliance rules, prohibited or restricted shipping constraints, and order cut-off time windows into constraint expressions, and constructing a feasible constraint set based on these constraint expressions. Consistency measures With feasible constraint set Perform correlation adaptation and evaluate the feasible constraint set according to preset threshold judgment rules. Dynamic adjustments are made by introducing a comprehensive feasibility indicator function to quantify the feasibility of each transportation segment e in the spatiotemporal transportation network G=(V,E), thereby selecting all sets that satisfy the feasible constraints. For all transport segments with all constraints, a set integration algorithm is used to aggregate all feasible transport segments to form an order. Dedicated order feasible edge set.

6. The method for optimizing cross-border logistics routes based on multimodal logistics information according to claim 1, characterized in that, The construction of the multi-objective optimization model includes defining binary decision variables for edge selection, constructing the objective function of the multi-objective optimization model hierarchically based on cost parameters, and adopting a weighted summation form for the objective function. This transforms the three optimization objectives of cost, risk, and robustness / timeliness into a single-objective optimization form, and integrates the feasible constraint set. The network connectivity constraints are converted into formal constraints suitable for a multi-objective optimization model, forming a complete multi-objective optimization model. The expression of the multi-objective optimization model is as follows: ， in, The comprehensive objective function of the multi-objective optimization model is... , and These are the weighting coefficients for the cost item, risk item, and robust timeliness item, respectively. For the cost term in the objective function, For the risk term in the objective function, This is the robust time-dependent term in the objective function.

7. The method for optimizing cross-border logistics routes based on multimodal logistics information according to claim 6, characterized in that, Finding the optimal path that minimizes the objective function involves pre-validating the constraints of the multi-objective optimization model, eliminating invalid constraints, and solving the multi-objective optimization model using an improved genetic algorithm to obtain the decision variables that minimize the comprehensive objective function J. Combination, The transport segments are spliced ​​together in spatiotemporal order, and the optimal path from the origin node to the destination node is obtained through analysis. The network connectivity constraint expression of the multi-objective optimization model is: ， ， in, We select binary decision variables for the edges to characterize the selection of transport segment e. Let v be the set of transport segments from node v. Let V be the set of transport segments flowing into node v. For the set of feasible nodes for the order, For the order's origin and shipment point, The destination node for the order.

8. The method for optimizing cross-border logistics routes based on multimodal logistics information according to claim 1, characterized in that, The interpretability analysis of the optimal path includes the optimal path... A hierarchical analysis is performed, and the cost of each transportation segment and node, as well as the total route cost, are quantitatively calculated based on cost parameters to obtain cost decomposition results. Based on comprehensive risk probability parameters, the single-type and comprehensive risk probabilities of each transportation segment are quantitatively calculated, and the total route risk level is summarized to form risk decomposition results. Finally, the timeliness robustness of the optimal route is calculated, generating a timeliness confidence index. The formula for calculating the conditional value of risk in the timeliness confidence index is as follows: ， in, For the confidence level of timeliness, As an auxiliary variable, For mathematical expectation operators, Let be the random variable representing the total delay of the optimal path.

9. A method for optimizing cross-border logistics routes based on multimodal logistics information according to claim 1, characterized in that, The step of fixing the executed path segments and resolving them on the remaining sub-networks includes: Optimal path During execution, the multimodal evidence package is updated synchronously. Based on the updated multimodal evidence package Quantitative calculation of the anomaly intensity at time t ; Preset abnormal intensity threshold The anomaly intensity will be calculated in real time. With threshold When a comparison is performed, When this happens, event-driven local replanning is triggered. When this happens, the optimal path is maintained. Continue execution; After triggering local replanning, fix the completed transportation segments in the optimal path, remove the executed nodes and transportation segments, and construct the remaining subnetwork containing unexecuted nodes and unexecuted transportation segments based on the feasible edge set of the order; The solution is re-solved on the remaining subnetworks, and the updated path that adapts to the remaining transportation links is obtained by combining the real-time updated multimodal information and cost parameters.

10. A cross-border logistics route optimization system based on multimodal logistics information, executed according to the method of claim 1, characterized in that, include: The data acquisition module is configured to collect multimodal logistics information associated with orders and form a multimodal evidence package. And based on multimodal evidence packages Generate consistency metrics ; The encoding fusion module is configured to: process multimodal evidence packages Feature encoding and fusion are performed to obtain a unified representation vector. and according to Generate the cost parameters for transport segment e and node v; The feasible edge set module is configured to be based on compliance rules, restricted / prohibited shipping constraints, order cut-off time windows, and consistency measures. Construct a set of feasible constraints And determine the feasible edge set for the order; The model module is configured to: build a multi-objective optimization model, using edge-selected variables. The selection of transport segment e is characterized by satisfying network connectivity and the set of constraints. Under the given conditions, find the optimal path that minimizes the objective function; The decomposition module is configured to: perform interpretability analysis on the optimal path and output interpretable results including cost decomposition, risk decomposition, and timeliness confidence indicators, wherein the timeliness confidence indicators include the quantile of total delay. Conditional Value at Risk ; The output module is configured to: measure the anomaly intensity based on multimodal information during the transportation process. Triggering event-driven local replanning fixes the executed path segments and resolves them on the remaining subnetworks to obtain the updated path.

Images