Central and eastern european block train container intelligent stowage method and system for international supply chain

Abstract

The present application relates to the technical cross field of cross-border railway transportation, and provides a method and system for intelligent loading of Central and Eastern Europe block train containers for international supply chains, comprising responding to a block train container loading task request; obtaining loading basic data corresponding to the target block train from a multi-source data interface based on the loading task request; constructing a continuous space model for the target container, introducing a heavy convex polygon stability algorithm to establish stability constraints, and using a deep reinforcement learning model based on Transformer to determine a target loading scheme; when it is detected that the load distribution state does not satisfy the preset constraint condition, a new container loading scheme is calculated by performing local adjustment. By constructing a continuous loading state space, the space rounding error caused by gridding is eliminated, the container space utilization rate and the overall loading rate are improved; at the same time, the loading efficiency is improved under the premise of ensuring the loading quality, and the safety risk in the transportation process is reduced by introducing a physical perception mask mechanism in the model.

Description

Intelligent loading methods and systems for China-Europe freight train containers for international supply chains Technical Field

[0001] This invention relates to the field of cross-border railway transportation technology, and in particular to a method and system for intelligent container loading of China-Europe freight trains for international supply chains. Background Technology

[0002] As the scale of China-Europe freight train operations continues to expand in international land ports, railway stations, and other railway logistics hubs, the collection, loading, and dispatch of cross-border goods at railway operation sites are becoming increasingly frequent. The rationality of container loading schemes directly affects the efficiency and operational safety of cross-border transportation. Especially in the context of international supply chains, improving container loading efficiency and ensuring loading and operational safety have become significant challenges in China-Europe freight train transportation. Currently, container loading and allocation operations still rely heavily on manual experience for decision-making, typically arranging goods based on principles such as placing heavy goods at the bottom and light goods on top, and concentrating loading at the same destination. However, these experience-based methods cannot effectively handle the various complex factors involved in cross-border transportation, such as track gauge conversions between different railway systems, customs declaration status, and the flow of goods between different countries.

[0003] Currently, information such as cargo dimensions, weight, category, and destination is mostly managed through manual data entry systems or paper documents. This leads to problems such as delayed data updates and susceptibility to human factors, making it difficult to formulate accurate and quantifiable loading layout plans before loading. Furthermore, in cross-border transportation, especially in loading operations at railway logistics hubs, loading is constrained not only by physical conditions such as container size, load capacity, and center of gravity limitations, but also by dynamic constraints such as customs declaration status, station operating conditions, availability of loading and unloading equipment, and external environmental factors. Existing technologies lack a unified spatial modeling and constraint expression method for the aforementioned multi-source information and multiple constraints. Assessments of center of gravity shift, cargo stability, and the feasibility of loading and unloading sequences during loading often rely on manual judgment or simple rules. This makes it difficult to quickly and locally adjust loading operations when cargo is temporarily changed, detained during inspection, or operating conditions change. Often, the entire loading plan needs to be replanned, resulting in low loading efficiency, low space utilization, and certain operational safety hazards. This fails to meet the actual needs of China-Europe freight trains for intelligent and stable loading operations in railway logistics hubs under the demands of the international supply chain. Summary of the Invention

[0004] To address the issues of reliance on manual decision-making, and the difficulty in ensuring loading efficiency and safety in China-Europe freight train container loading for international supply chains, this invention proposes an intelligent loading method and system for China-Europe freight train containers to achieve intelligent and efficient loading operations:

[0005] To achieve the above objectives, the present invention utilizes the following techniques:

[0006] This invention provides an intelligent container stowage method for China-Europe freight trains oriented towards international supply chains, including:

[0007] Step S1: Respond to the China-Europe freight train container loading task request. The loading task request is used to calculate the loading of containers for the target train.

[0008] Step S2: Based on the loading task request, obtain the loading basic data corresponding to the target train from the multi-source data interface. The loading basic data includes the order data of the goods to be loaded, container specifications and load constraint parameters, and customs clearance status data.

[0009] Step S3: Based on the loading data, establish a three-dimensional spatial constraint for the target container, including the geometric boundary of the container and the size limit of the cargo. Based on the three-dimensional spatial constraint, divide the internal space of the target container into a continuous space to form a continuous available space model representing the loadable area inside the container. Then, organize the continuous available space model through the packing configuration tree to construct a continuous space model inside the container. Introduce the load-bearing convex polygon stability algorithm to judge the center of gravity position of the cargo to be loaded in the candidate loading space and establish stability constraints that conform to the mechanical principles.

[0010] Step S4: Using a Transformer-based deep reinforcement learning model, attention encoding is performed on the cargo features and spatial features of the cargo to be loaded, and corresponding container loading decisions are generated under the constraints of stability constraints and physical perception masks. The container loading decisions evaluate the candidate loading results by using a multi-dimensional composite reward function that includes volume utilization, loading balance, loading compactness and loading and unloading order, and the target loading scheme is determined based on the evaluation results.

[0011] Step S5: Use the target loading scheme as the container loading scheme and monitor the load distribution status during the loading process. When the load distribution status is detected to be inconsistent with the preset constraints, perform local adjustments to the affected loading units and recalculate the new container loading scheme by calling the deep reinforcement learning model through the packaged configuration tree.

[0012] Further, step S1 includes:

[0013] Step S11: Receive a loading task request from the railway logistics hub dispatch center and obtain the train identification information of the target train and the corresponding planned departure time window;

[0014] Step S12: Verify the validity of the load allocation task request, determine that the target train is in a preset operating state that allows load allocation calculation, and generate a load allocation calculation trigger command when the validity verification is passed to start the load allocation calculation of the target train.

[0015] Further, step S2 includes:

[0016] Step S21: Obtain the cargo order data associated with the target train from the cargo order data interface, and perform parameterization processing on the geometric dimension parameters, weight parameters and category information of the cargo to be loaded to generate a cargo parameter set;

[0017] Step S22: Obtain the container specifications and load constraint parameters corresponding to the target train from the container resource management interface, and extract the corresponding container capacity parameters and maximum load parameters;

[0018] Step S23: Obtain the customs clearance status data corresponding to the order data of the goods to be loaded from the customs clearance status interface, and filter out the set of goods parameters that meet the loading calculation conditions;

[0019] Step S24: Associate the filtered cargo parameter set, container capacity parameters, and maximum load-bearing parameters to form a basic parameter set for the construction of a three-dimensional spatial constraint structure, which serves as the basic data for loading.

[0020] Further, step S3 includes:

[0021] Step S31: Using the inner corner of the target container as a spatial reference point, construct a three-dimensional rectangular coordinate system with the length, width and height directions of the container corresponding to the X-axis, Y-axis and Z-axis respectively. Then, update the internal space of the target container in the three-dimensional rectangular coordinate system through geometric difference operation to generate the maximum remaining space set representing the loadable area inside the container.

[0022] Step S32: Based on the maximum remaining space set, organize the current loading status of the target container into a hierarchical packing configuration tree to form a continuous space model. Each leaf node of the packing configuration tree corresponds to a candidate loading space that can be used for loading in the maximum remaining space set, and the spatial feature parameters of the candidate loading space include spatial size and spatial coordinates.

[0023] Step S33: For the candidate loading space corresponding to the leaf node of the packing configuration tree in the continuous space model, construct the load-bearing convex polygon of the support surface based on the support surface formed by the loaded goods below the candidate loading space, and determine whether the center of gravity projection of the goods to be loaded in the candidate loading space is located inside the load-bearing convex polygon. If the center of gravity projection is located inside the load-bearing convex polygon, the candidate loading space meets the stability constraint requirements; otherwise, the candidate loading space does not meet the stability constraint, and the candidate loading space is marked as a candidate loading space that cannot be used for container loading decision.

[0024] Further, step S4 includes:

[0025] Step S41: Using a multi-head self-attention mechanism and a graph neural network, feature models are performed on the dependencies of the cargo sequence and the spatial topology of the packing configuration tree, respectively. This constructs cargo feature vectors and spatial topology feature vectors reflecting the current loading status of the container, including:

[0026] The cargo to be loaded is arranged into a sequence of cargo to be loaded according to a predetermined loading and unloading order. Based on the sequence of cargo to be loaded, a multi-head self-attention mechanism is used to encode the dependencies between each cargo to be loaded, and to obtain cargo feature vectors that reflect the loading correlation between the cargo to be loaded. Among them, the dependencies include loading combination correlations formed based on the destination attributes of the cargo to be loaded and / or loading order priority correlations formed based on the weight parameters of the cargo to be loaded.

[0027] Based on the continuous space model, for each leaf node in the packaging configuration tree, the spatial features of the corresponding candidate loading space are obtained, and the leaf nodes are used as graph nodes as the processing objects of the graph neural network to generate the corresponding spatial topology feature vectors.

[0028] Step S42: Based on the cargo feature vector and spatial topology feature vector, perform loading matching between the cargo to be loaded and the candidate loading spaces in the continuous spatial model. Combine the judgment results of whether the candidate loading spaces meet the stability constraints in step S33 and the three-dimensional spatial constraints, filter the candidate loading spaces. Mark the candidate loading spaces that do not meet the stability constraints or the three-dimensional spatial constraints as unselectable loading spaces and remove them from the maximum remaining space set for shielding during the container loading decision calculation process.

[0029] Step S43: Based on the maximum remaining space set after filtering, the cross-attention layer of the Transformer decoder is used to perform correlation calculation on the cargo feature vector and the spatial topology feature vector to obtain the matching degree score corresponding to each selectable candidate loading space, and the probability distribution of loading action is generated based on the matching degree score to determine the candidate loading result. The candidate loading result includes the target placement position and corresponding placement posture of the cargo to be loaded;

[0030] Step S44: Construct a multi-dimensional composite reward function that includes volume utilization, load balance, loading compactness, and loading / unloading sequence to comprehensively evaluate the candidate load allocation results. Based on the evaluation results, optimize and select the loading actions output by the deep reinforcement learning model to generate the corresponding target load allocation scheme.

[0031] Further, step S42 includes:

[0032] Determine whether the current cargo to be loaded satisfies the stability constraints or three-dimensional space constraints when placed in the candidate loading space corresponding to any leaf node.

[0033] When stability constraints or three-dimensional space constraints are not met, the corresponding loading action is marked as an unexecutable loading action, and a first mask identifier -∞ is assigned to the unexecutable loading action to mask the unexecutable loading action during the container loading decision calculation process.

[0034] When the stability constraints and three-dimensional space constraints are met, the corresponding loading action is marked as an executable loading action, and a second mask identifier of 0 is assigned to the executable loading action to allow the executable loading action to participate in the container loading decision calculation.

[0035] Further, step S43 includes:

[0036] Using cargo feature vectors as query features and spatial topology feature vectors as key and value features, a cross-attention layer is used to perform correlation calculations on query features, key features, and value features to obtain the matching score corresponding to each candidate loading space.

[0037] Based on the matching score, the loading actions corresponding to each candidate loading space are probabilistically calculated to generate a probability distribution of the loading actions. The loading action with the highest probability is selected according to the probability distribution, and the corresponding candidate loading space and placement posture are determined as the target placement location for the goods to be loaded.

[0038] The formula for probabilistic calculation is as follows: S t Given the current load allocation decision state, a t For loading action, For loading action a t In the current load decision state S t The probability distribution below, For a given loading action a t Mask markings, For loading action a t Match score, The function is used for loading action a tThe matching score is normalized to generate a probability distribution of loading actions.

[0039] Furthermore, in step S44, the calculation formula for the multidimensional composite reward function is as follows:

[0040] ;

[0041] Where i is the index number of the cargo to be loaded, and S t Let S be the current load planning decision state, and R be the current load planning decision state. t The following loading action a t The comprehensive reward value is the evaluation result of the candidate loading results. w1, w2, w3, and w4 are the weights of four indicators: volume utilization rate, loading balance, loading compactness, and loading / unloading sequence, respectively. i v represents the volume of the cargo currently awaiting loading. bin This represents the total volume of the candidate loading space used for loading. To determine the current load allocation decision state S t The spatial position of each loaded and unloaded cargo is used to calculate the ratio of the overall center of gravity height to the corresponding support area, which is used as a stability index to reflect the stability of the current loading structure. To determine the current load allocation decision state S t The contact relationship between the cargo to be loaded and the already loaded cargo and the inner wall of the container after the cargo is placed, and the compactness index is constructed based on the distribution of contact surfaces between cargo and between cargo and container; To determine the current load allocation decision state S t The correspondence between the spatial placement order of each cargo and the predetermined loading and unloading order is established, and a sequence penalty index is constructed based on the reverse placement situation in the loading results.

[0042] Further, step S5 includes:

[0043] Step S51: Convert the target loading scheme into a loading execution instruction set and send the loading execution instruction set to the corresponding loading unit. The loading execution instruction set shall at least include the loading unit identifier, spatial placement parameters and loading sequence parameters corresponding to each cargo to be loaded.

[0044] Step S52: During the loading process, collect the actual load data of each loading unit in real time, and construct the current load distribution status of the container based on the collected actual load data, as the current loading decision status;

[0045] Step S53: Calculate the difference between the current load allocation decision state and the current load distribution state. If the difference exceeds the preset threshold or if the stability constraint is not met, pause the loading execution process and trigger the reconfiguration mechanism; otherwise, continue to execute the corresponding loading action according to the loading execution instruction set.

[0046] Step S54: When the reconstruction mechanism is triggered, identify the loading units whose contribution to the load distribution deviation or center of gravity offset exceeds the corresponding threshold, determine them as the affected loading units, and extract the corresponding adjustable loading units from the maximum remaining space set to form a local adjustment area;

[0047] Step S55: Remove the goods in the local adjustment area from the current loading state and put them back into the waiting queue. Use the local adjustment area as a new leaf node, and call the deep reinforcement learning model in S4 again to perform container loading decision calculation on the local adjustment area to generate an updated container loading plan.

[0048] Based on the same inventive concept, this invention also provides an intelligent container stowage system for China-Europe freight trains oriented towards the international supply chain, employing the aforementioned intelligent container stowage method for China-Europe freight trains oriented towards the international supply chain, including:

[0049] The response module is used to respond to container loading task requests from China-Europe freight trains. The loading task request is used to perform loading calculations on the containers of the target train.

[0050] The data acquisition module is used to obtain the basic loading data corresponding to the target train from multiple data interfaces based on the loading task request. The basic loading data includes the order data of the goods to be loaded, container specifications and load constraint parameters, and customs clearance status data.

[0051] The data processing module is used to establish three-dimensional spatial constraints on the target container based on the loading basic data, including the geometric boundaries of the container and the size restrictions of the cargo. Based on the three-dimensional spatial constraints, the internal space of the target container is continuously divided into spaces to form a continuous available space model representing the loadable area inside the container. Then, the continuous available space model is organized through the packing configuration tree to construct a continuous space model inside the container. The load-bearing convex polygon stability algorithm is introduced to determine the center of gravity position of the cargo to be loaded in the candidate loading space and establish stability constraints that conform to the mechanical principles.

[0052] The output module utilizes a Transformer-based deep reinforcement learning model to encode the cargo features and spatial features of the cargo to be loaded with attention. Under the constraints of stability constraints and physical perception masks, it generates corresponding container loading decisions. The container loading decision evaluates candidate loading results using a multi-dimensional composite reward function that includes volume utilization, loading balance, loading compactness, and loading / unloading sequence. Based on the evaluation results, a target loading scheme is determined. The target loading scheme is used as the container loading scheme, and the load distribution status is monitored during the loading process. When the load distribution status is found to be inconsistent with the preset constraints, local adjustments are made to the affected loading units. The deep reinforcement learning model is then called again through the packaged configuration tree to calculate a new container loading scheme.

[0053] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0054] This invention abandons the static planning approach based on discrete grids in traditional technologies. Instead, it continuously models the internal space of containers using maximum remaining space and a packing configuration tree, constructing a continuous loading state space. This eliminates spatial rounding errors caused by gridding, improving container space utilization and overall loading rate. Simultaneously, it uses a pre-trained Transformer model for loading decision reasoning, resulting in high computational efficiency. This ensures loading quality while improving loading efficiency. Furthermore, by introducing a physical perception mask mechanism into the model, loading actions that do not meet stability or spatial constraints are masked during the decision-making stage, reducing safety risks during transportation from the outset. Moreover, during actual loading, this invention only performs local optimization and reconstruction on affected loading units—those impacted by load distribution or center of gravity—without requiring global calculation of the entire loading scheme. This enhances the adaptability of the loading scheme to actual loading deviations and temporary adjustment needs, reducing manual intervention and repetitive calculation costs. It is suitable for efficient and stable container loading applications in China-Europe freight trains within international supply chain logistics hubs. Attached Figure Description

[0055] Figure 1 is a flowchart of the steps of the intelligent loading method for China-Europe freight train containers for international supply chains according to the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0057] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0058] First Embodiment

[0059] In the context of international supply chains within railway logistics hubs, the loading operations of China-Europe freight train containers are typically affected by multiple factors, including train departure times, differences in cargo attributes, container structural limitations, and on-site loading and unloading conditions. In cross-border rail transport, container loading involves not only cargo arrangement but is also closely related to external factors such as coordination between different national railway systems, customs procedures, and the availability of loading and unloading equipment.

[0060] As the scale of China-Europe freight train operations continues to expand, the efficiency and safety of container loading operations have become core challenges in cross-border logistics. In actual loading, these challenges also relate to load distribution, overall center of gravity stability, and the feasibility and safety of the loading process, making it a typical multi-constraint coupled optimization problem. Especially in the complex environment of international transportation, the impact of different spatial locations within a container on the overall load distribution and center of gravity varies significantly. For example, loading areas located at the bottom of the container or near its geometric center have a relatively small impact on the overall center of gravity; while loading areas located at the top or ends of the container, even small load changes, can cause a significant shift in the center of gravity, increasing safety risks during transportation. However, existing loading technologies typically treat the internal space of the container as a homogeneous loading area, failing to differentiate the stability sensitivity of different loading positions. When loading deviations or adjustments are needed, the entire loading scheme often requires recalculation, resulting in high computational costs and difficulty in meeting the practical needs of rapid on-site adjustments.

[0061] Furthermore, during the loading execution phase, the actual load distribution will inevitably deviate from the initial load allocation calculation results due to factors such as weighing accuracy, differences in cargo packaging conditions, and temporary changes in loading and unloading sequences. Without a continuous perception and feedback mechanism for the loading process, the load allocation plan cannot be promptly corrected based on real-time conditions, easily introducing new load imbalances or center of gravity shifts during local adjustments.

[0062] Based on the above problems, this embodiment proposes a new container loading technology approach. Instead of treating the container's internal space as a homogeneous whole, it constructs a set of loading units with spatial location, load constraints, and center of gravity influence characteristics. The load contribution and center of gravity sensitivity of different loading units under the current loading state are quantitatively evaluated, thus distinguishing between hold-type and adjustable loading units. During loading optimization and adjustment, reconstruction and optimization calculations are performed only on the local loading areas composed of adjustable loading units. This improves the flexibility and computational efficiency of loading adjustments while ensuring overall stability. Furthermore, this embodiment combines multi-source data acquisition, three-dimensional spatial constraint modeling, and a load distribution monitoring mechanism during loading execution to realize an intelligent container loading method for China-Europe freight trains suitable for railway logistics hub operations. The specific implementation method is as follows.

[0063] As shown in Figure 1, this invention provides an intelligent container stowage method for China-Europe freight trains oriented towards the international supply chain, including:

[0064] Step S1: Respond to the China-Europe freight train container loading task request. The loading task request is used to calculate the loading of containers for the target train.

[0065] Step S2: Based on the loading task request, obtain the loading basic data corresponding to the target train from the multi-source data interface. The loading basic data includes the order data of the goods to be loaded, container specifications and load constraint parameters, and customs clearance status data.

[0066] Step S3: Based on the loading data, establish a three-dimensional spatial constraint for the target container, including the container's geometric boundaries and cargo size limitations. Utilize the maximum remaining space and the packing configuration tree to construct a high-precision continuous spatial model. Introduce a load-bearing convex polygon to establish stability constraints that conform to mechanical principles. Specifically, based on the three-dimensional spatial constraints, continuously divide the internal space of the target container to form a continuous available space model representing the loadable area inside the container. Then, organize the continuous available space model through the packing configuration tree to construct a continuous spatial model inside the container. Introduce a load-bearing convex polygon stability algorithm to determine the center of gravity position of the cargo to be loaded in the candidate loading space, and establish stability constraints that conform to mechanical principles.

[0067] Step S4: Utilizing a Transformer-based deep reinforcement learning model, cargo and spatial features are encoded through an attention mechanism, and a physically feasible millisecond-level loading decision is directly generated by combining a physical perception mask. The optimization objective is driven by a multi-dimensional composite reward function, comprehensively considering volume utilization, loading stability, loading compactness, and loading / unloading sequence to achieve collaborative processing of multiple objectives. Specifically, by encoding the cargo and spatial features of the cargo to be loaded with attention, and generating corresponding container loading decisions under stability constraints and physical perception mask constraints, the container loading decision evaluates candidate loading results using a multi-dimensional composite reward function that includes volume utilization, loading balance, loading compactness, and loading / unloading sequence, and determines the target loading scheme based on the evaluation results.

[0068] Step S5: Use the target loading scheme as the container loading scheme and monitor the load distribution status during the loading process. When the load distribution status is detected to be inconsistent with the preset constraints, perform local adjustments to the affected loading units and recalculate the new container loading scheme by calling the deep reinforcement learning model through the packaged configuration tree.

[0069] It's important to clarify that the primary goal of ship container stowage schemes is to optimize the discrete distribution of containers across ship holds or decks, ensuring that cargo is effectively allocated to each hold according to the designated loading and unloading sequence. For ship container stowage, stability constraints mainly focus on the overall stability of the hull, particularly how to cope with waves, tilting, and the hull's inherent stability requirements during navigation. These constraints primarily concern the arrangement of containers' relative positions and weight distribution to ensure safe navigation. However, the technical solutions for ship container stowage differ fundamentally from those for China-Europe freight trains in rail transport scenarios. In rail transport, especially in cross-border rail transport and international supply chain environments, the challenges of container stowage extend beyond simply distributing cargo rationally within a limited space; they also include the stability of cargo within containers, particularly controlling the center of gravity and optimizing space utilization.

[0070] This application addresses the problem of continuous spatial loading and center of gravity stability of cargo within containers in cross-border rail transport scenarios, employing continuous spatial modeling and container stability assessment methods. Unlike the constraints on center of gravity stability and overall hull stability in ship container loading, the technical solution in this application focuses more on the loading sequence, load distribution, and center of gravity stability of cargo within individual containers. Particularly in cross-border rail transport environments like the China-Europe Railway Express, it addresses how to handle the impact of different loading positions on the overall stability of the container, and how to make local adjustments to improve loading efficiency and flexibility.

[0071] Therefore, the technical solution of this application mainly addresses the issues of loading stability and center of gravity control of cargo inside containers. It adopts a loading method based on a continuous space model, ensuring cargo stability through precise modeling of the loading space and center of gravity analysis. Compared with ship container loading methods, the technical object, spatial model, and stability constraint hierarchy of this application are significantly different. It focuses on the optimization and dynamic adjustment of the internal space of containers in cross-border railway transportation environments, offering greater flexibility and accuracy.

[0072] Therefore, to describe the method of this embodiment in another way, in the actual process of container loading of China-Europe freight trains, considering the railway gauge conversion, loading deviation and special requirements of loading and unloading operations in cross-border transportation, the process is as follows: (1) Obtain information on the train's cargo orders, container specifications, loading constraints and customs clearance status from the logistics hub dispatch center. After these data are uniformly formatted and parameterized, a standardized dataset for loading calculation is formed. (2) Obtain information on the train's cargo orders, container specifications, loading constraints and customs clearance status from the logistics hub dispatch center. After being uniformly formatted and parameterized, these data form a standardized dataset for loading calculation. (3) By constructing a continuous available space model based on three-dimensional spatial constraints, the available space inside the container is organized using the maximum remaining space and packing configuration tree method to ensure the spatial accuracy of each loading position and avoid spatial fragmentation. (4) When deviations or changes occur during the actual loading process, this technical solution can perform local optimization for the affected loading units to avoid global recalculation, thereby improving the flexibility and efficiency of loading adjustment. (5) When deviations or changes occur during the actual loading process, this technical solution can perform local optimization for the affected loading units to avoid global recalculation, thereby improving the flexibility and efficiency of loading adjustment. By collecting load distribution data inside the container in real time, the system can adjust the loading scheme in real time during the loading process to ensure the physical feasibility and safety of each loading action.

[0073] Further, step S1 includes:

[0074] Step S11: Receive a loading task request from the railway logistics hub dispatch center and obtain the train identification information of the target train and the corresponding planned departure time window;

[0075] Step S12: Verify the validity of the load allocation task request, determine that the target train is in a preset operating state that allows load allocation calculation, and generate a load allocation calculation trigger command when the validity verification is passed to start the load allocation calculation of the target train.

[0076] In this embodiment, the loading task request is automatically generated or manually triggered by the railway logistics hub dispatch center based on the train organization plan. It includes at least the unique train identifier of the target train, the corresponding shipping route information, and the planned shipping time window. The planned shipping time window is used to define the time boundary for loading calculations and subsequent loading operations, avoiding unnecessary or unexecutable loading calculations on trains nearing departure or already in the loading lock-in stage, thereby improving the overall timeliness and stability of the system. Simultaneously, loading calculations are not directly performed on the received loading task request; instead, the current operating status of the target train is first validated based on the train identifier information. This validation includes at least: determining whether the target train has completed train formation confirmation, whether it has not yet entered the loading freeze state, whether it has not been marked as a temporary adjustment or abnormal handling state, and whether it meets the preset loading calculation triggering conditions. Only when all the above validation conditions are met is a corresponding loading calculation triggering instruction generated to initiate the container loading calculation process for the target train.

[0077] Further, step S2 includes:

[0078] Step S21: Obtain the cargo order data associated with the target train from the cargo order data interface, and perform parameterization processing on the geometric dimension parameters, weight parameters and category information of the cargo to be loaded to generate a cargo parameter set;

[0079] Step S22: Obtain the container specifications and load constraint parameters corresponding to the target train from the container resource management interface, and extract the corresponding container capacity parameters and maximum load parameters;

[0080] Step S23: Obtain the customs clearance status data corresponding to the order data of the goods to be loaded from the customs clearance status interface, and filter out the set of goods parameters that meet the loading calculation conditions;

[0081] Step S24: Associate the filtered cargo parameter set, container capacity parameters, and maximum load-bearing parameters to form a basic parameter set for the construction of a three-dimensional spatial constraint structure, which serves as the basic data for loading.

[0082] Specifically, data interaction is performed with the cargo order data interface within the railway logistics hub to obtain cargo order data associated with the target train. This cargo order data includes at least the cargo's dimensions, weight information, and cargo category identifier. This information undergoes unified data formatting and parameterization, converting cargo information from different sources and with different representations into a standardized set of cargo parameters usable for subsequent calculations. The container specification information and load constraint parameters corresponding to the target train are obtained through the container resource management interface, and container capacity parameters and maximum load-bearing parameters for loading calculations are extracted. The container capacity parameter limits the available loading space of a single container, and the maximum load-bearing parameter constrains the total weight limit of the container during loading to avoid overloading risks. Customs clearance status information corresponding to the cargo order data is obtained through the customs clearance status interface, and the cargo parameter set is filtered based on this information. Specifically, cargo that has not yet cleared customs, is under inspection, or is marked as temporarily unreleased is excluded from the current loading calculation scope, thereby avoiding including cargo that does not meet actual loading conditions in the loading optimization process and reducing subsequent adjustment costs.

[0083] Simultaneously, the filtered cargo parameter set is correlated with the container capacity parameters and maximum load-bearing parameters to form a basic parameter set for constructing the three-dimensional spatial constraint structure. This basic parameter set comprehensively reflects the cargo characteristics and the container's physical constraints, providing a unified data foundation for subsequently establishing the container's internal three-dimensional spatial constraint structure and performing multi-objective loading optimization calculations.

[0084] Further, step S3 includes:

[0085] Step S31: Using the inner corner of the target container as a spatial reference point, construct a three-dimensional rectangular coordinate system with the length, width and height directions of the container corresponding to the X-axis, Y-axis and Z-axis respectively. Then, update the internal space of the target container in the three-dimensional rectangular coordinate system through geometric difference operation to generate the maximum remaining space set representing the loadable area inside the container.

[0086] Step S32: Based on the maximum remaining space set, organize the current loading status of the target container into a hierarchical packing configuration tree to form a continuous space model. Each leaf node of the packing configuration tree corresponds to a candidate loading space that can be used for loading in the maximum remaining space set, and the spatial feature parameters of the candidate loading space include spatial size and spatial coordinates.

[0087] Step S33: For the candidate loading space corresponding to the leaf node of the packing configuration tree in the continuous space model, construct the load-bearing convex polygon of the support surface based on the support surface formed by the loaded goods below the candidate loading space, and determine whether the center of gravity projection of the goods to be loaded in the candidate loading space is located inside the load-bearing convex polygon. If the center of gravity projection is located inside the load-bearing convex polygon, the candidate loading space meets the stability constraint requirements; otherwise, the candidate loading space does not meet the stability constraint, and the candidate loading space is marked as a candidate loading space that cannot be used for container loading decision.

[0088] Specifically, step S31 uses the inner corner of the target container as a spatial reference point to construct a three-dimensional Cartesian coordinate system with the length, width, and height directions corresponding to the X, Y, and Z axes, respectively. A maximum remaining space (EMS) set is dynamically generated and maintained through geometric difference operations, accurately describing all remaining usable cuboid regions within the container with floating-point precision, eliminating spatial fragmentation and accuracy loss. Step S32 organizes the real-time state of the container into a hierarchical Packing Configuration Tree (PCT) structure, where leaf nodes directly correspond to the currently available EMS set and its spatial features (such as dimensions and coordinates), providing efficient and structured state input for subsequent deep reinforcement learning models. Step S33 introduces the Loaded Convex Polygon (LBCP) algorithm, which conforms to Newtonian mechanics. By calculating the joint convex hull of the support surfaces, it verifies whether the center of gravity projection of the cargo to be loaded falls within the effective support area, thus establishing a more stringent and secure physical stability hard constraint than simply calculating the center of gravity distance.

[0089] Further, step S31 includes:

[0090] S311: Before container loading begins, based on the container's geometric dimensions, define the initial loadable space inside the container as follows: ,in , , Let the length, width, and height of the container be the values, and initialize the maximum remaining space set using the initial loadable space as the unique element. ;

[0091] S312: When any cargo is to be loaded Space inside the container and occupy size After the three-dimensional space, based on the current Each space in the set Perform spatial updates based on geometric difference:

[0092] when Candidate space With goods If the occupied spaces do not overlap, then the candidate space is retained. ;

[0093] when Candidate space With goods If the occupied spaces overlap, then the candidate spaces... Goods Perform a segmentation process and generate up to three new rectangular subspaces based on the segmentation surfaces, located respectively within the cargo. Above, in front of, and to the side; the generated subspaces are aggregated to form a new set of subspaces. The subspace set and the candidate spaces that do not overlap are combined into a temporary maximum remaining space set.

[0094] S313 performs a redundant space filtering operation on the temporary maximum remaining space set, deleting subspaces that are completely contained by other candidate spaces, and retaining only the maximal rectangular spaces that cannot be completely covered in at least one dimension, to obtain the updated set. This maximum remaining space set is used to accurately describe all contiguous space areas inside the container that are available for subsequent loading at the current moment.

[0095] Further, step S32 includes:

[0096] Step S321: Based on the container's geometric boundaries, construct the root node of the packing configuration tree. The root node represents the overall loadable space state of the container and serves as the initial node of the packing configuration tree.

[0097] Step S322: Based on the loading order of the loaded goods and their space occupancy status inside the container, construct internal nodes level by level in the packing configuration tree. Each internal node is used to represent the division status of the container's internal space and the constraint relationship between the loaded goods and the subsequent loading space at the corresponding loading stage.

[0098] Step S323: Based on the maximum set of remaining space at the current time. In the package configuration tree, construct the corresponding set of leaf nodes, where each leaf node... This corresponds to a maximum remaining space in the maximum remaining space set and is used to represent the candidate load space currently available for loading; for each leaf node Construct the corresponding spatial feature vectors. ,in These are the coordinates of the corner points in space. For spatial ruler, This describes the characteristics of the supporting surface at the bottom of the space, indicating the space's load-bearing conditions.

[0099] Further, step S33 includes:

[0100] Step S331: For any candidate leaf node in the packaged configuration tree Based on ray detection or spatial projection, identify the height of the bottom surface of the candidate loading space. Below, the set of top surfaces of loaded goods that can provide physical support for the candidate loading space, i.e., the set of support surfaces. ;

[0101] Step S332: Based on the set of support surfaces Perform a joint convex hull operation on the geometric projection regions of each support surface to obtain the load-bearing convex polygons corresponding to the candidate loading spaces. Among them, the load-bearing convex polygon This indicates the area of ​​physical support that can effectively support the cargo above at the current altitude level;

[0102] Step S333: Determine the bottom area of ​​the cargo to be loaded based on its geometric dimensions. And calculate the projection point of the physical center of gravity of the cargo to be loaded on the horizontal plane. Based on load-bearing convex polygons Construct the following stability constraints:

[0103] (1) Centroid envelope constraint: This means that the projection of the center of gravity of the cargo to be loaded must be located inside the load-bearing convex polygon to prevent the cargo from tipping over after loading.

[0104] (2) Effective contact area constraint: ;in, The preset threshold is used to avoid large areas of the bottom of the cargo to be loaded being suspended, thereby reducing the risk of localized stress concentration or slippage.

[0105] Step S334: When a candidate loading space simultaneously satisfies the center of gravity envelope constraint and the effective contact area constraint, the candidate loading space is determined to satisfy the stability constraint; otherwise, the candidate loading space is determined not to satisfy the stability constraint and is marked as a candidate loading space that cannot be used for subsequent container loading decisions.

[0106] It should be further explained that, in order to transform the loading problem of the target container into a computable and constrained spatial optimization problem, a three-dimensional Cartesian coordinate system is first constructed using an interior corner point inside the target container as a unified spatial reference point. In this three-dimensional Cartesian coordinate system, the length direction of the container corresponds to the X-axis, the width direction corresponds to the Y-axis, and the height direction corresponds to the Z-axis. The spatial boundaries of the three-dimensional Cartesian coordinate system in each coordinate axis direction are determined according to the specifications of the container, thereby limiting the loadable geometric range of the target container. In this way, any spatial position inside the container can be represented by a unique three-dimensional coordinate. Whenever a cargo to be loaded is planned to be placed in a specific position inside the current container, a spatial update mechanism is triggered, traversing every remaining space in the current EMS set to perform intersection detection, spatial cutting, and validity filtering. To prevent excessive fragmentation caused by spatial cutting from leading to computational explosion, the system performs maximization filtering on all newly generated subspaces, and the final set retained is the maximum remaining space, which constitutes the accurate candidate placement point set at the current moment.

[0107] The algorithm utilizes a packaging configuration tree structure to organize and manage the generated continuous spatial data. It constructs a hierarchical tree-like data structure: the root node represents the initial empty state of the container; as the loading process progresses, the intermediate nodes record the historical information of each placed item and its occupied space; and the leaf nodes are the core of the structure, specifically used to store all the maximum remaining space after filtering at the current moment. Each leaf node records the specific length, width, and absolute coordinates of the corresponding available space. In this way, the algorithm transforms complex spatial geometry into a list of decision candidates that is easily processed by a computer, enabling the subsequent intelligent decision-making model to directly select the optimal loading position from these leaf nodes, thus forming the candidate space for loading actions.

[0108] The algorithm first identifies the support surfaces by scanning the top surfaces of all loaded goods below the current placement space to determine effective support sources. Next, it calculates the joint geometric convex hull of these dispersed support surfaces, generating a minimum polygonal region that covers all support points. This is called the load-bearing convex polygon, representing the actual physically stable support range at that height level. Finally, based on this, the algorithm generates two strict constraints: first, the projection point of the newly placed goods' center of gravity must fall precisely within this convex polygon region to prevent the goods from tipping over; second, the contact area ratio between the bottom surface of the goods and the supporting surface below must exceed a preset safety threshold to prevent excessive local stress leading to crushing. Only placement actions that simultaneously meet both conditions are considered physically feasible.

[0109] Further, step S4 includes:

[0110] Step S41: The dependencies of the cargo sequence and the spatial topology of the packing configuration tree are modeled by multi-head self-attention mechanism and graph neural network respectively. The complex physical state is transformed into a mathematical vector that can be processed by the neural network, and cargo feature vector and spatial topology feature vector are constructed to reflect the container loading state.

[0111] Step S42: Before model decision-making, a hard constraint filtering mechanism is introduced. Based on the LBCP stability algorithm, all candidate positions that violate mechanical equilibrium or geometric boundaries are automatically screened to ensure that every loading action output by the agent is physically inherently safe. Specifically, based on the cargo feature vector and the spatial topology feature vector, the cargo to be loaded is matched with the candidate loading spaces in the continuous spatial model. Combining the judgment results of whether the candidate loading spaces meet the stability constraints in step S33 and the three-dimensional spatial constraints, the candidate loading spaces are screened. Candidate loading spaces that do not meet the stability constraints or the three-dimensional spatial constraints are marked as unselectable loading spaces and removed from the maximum remaining space set for screening during the container loading decision calculation process.

[0112] Step S43: Based on the maximum remaining space set after filtering, the cross-attention layer of the Transformer decoder is used to perform correlation calculation on the cargo feature vector and the spatial topology feature vector to obtain the matching degree score corresponding to each selectable candidate loading space. Based on the matching degree score, the probability distribution of loading actions is generated to determine the candidate loading results. The candidate loading results include the target placement position and corresponding placement posture of the cargo to be loaded, realizing fast end-to-end decision-making.

[0113] Step S44: Construct a multi-dimensional composite reward function that includes volume utilization, load balance, loading compactness, and loading / unloading sequence to comprehensively evaluate the candidate load allocation results. Based on the evaluation results, optimize and select the loading actions output by the deep reinforcement learning model to generate the corresponding target load allocation scheme.

[0114] Further, step S41 includes:

[0115] The cargo to be loaded is arranged into a sequence of cargo to be loaded according to a predetermined loading and unloading order. Based on the sequence of cargo to be loaded, a multi-head self-attention mechanism is used to encode the dependencies between each cargo to be loaded, and to obtain cargo feature vectors that reflect the loading correlation between the cargo to be loaded. Among them, the dependencies include loading combination correlations formed based on the destination attributes of the cargo to be loaded and / or loading order priority correlations formed based on the weight parameters of the cargo to be loaded.

[0116] Based on the continuous space model, spatial features of the corresponding candidate loading space are obtained for each leaf node in the packaging configuration tree, and the leaf nodes are used as graph nodes as the processing objects of the graph neural network to generate the corresponding spatial topology feature vectors.

[0117] Specifically, cargo coding uses a multi-head self-attention mechanism to process cargo sequences. The system extracts the dependency features between goods, such as the implicit patterns of "goods with the same destination should be grouped together" and "heavy goods should be given priority"; spatial coding uses a graph neural network to process the leaf nodes of the PCT to generate spatial topological feature vectors.

[0118] Further, step S42 includes:

[0119] Determine the current cargo to be loaded. Placed on any leaf node Whether the corresponding candidate loading space satisfies stability constraints or three-dimensional space constraints;

[0120] When stability constraints or three-dimensional space constraints are not met, the corresponding loading action is marked as an unexecutable loading action, and a first mask identifier -∞ is assigned to the unexecutable loading action to mask the unexecutable loading action during the container loading decision calculation process.

[0121] When the stability constraints and three-dimensional space constraints are met, the corresponding loading action is marked as an executable loading action, and a second mask identifier 0 is assigned to the executable loading action to allow the executable loading action to participate in the container loading decision calculation, so that any action output by the agent is physically safe and feasible, without the need for subsequent verification and rejection.

[0122] Further, step S43 includes:

[0123] Using cargo feature vectors as query features, spatial topology feature vectors as key features and value features as values, a cross-attention layer is used to correlate and calculate the query features, key features, and value features to obtain the matching score for each candidate loading space. ;

[0124] Based on matching score For each candidate loading space, a probabilistic calculation is performed on the loading action corresponding to it, generating a probability distribution of the loading action. Based on the probability distribution, the loading action with the highest probability is selected, and the corresponding candidate loading space and placement posture are determined as the target placement location for the cargo to be loaded.

[0125] The formula for probabilistic calculation is as follows: S t Given the current load allocation decision state, a tFor loading action, For loading action a t In the current load decision state S t The probability distribution below, For a given loading action a t Mask markings, For loading action a t Match score, The function is used for loading action a t The matching score is normalized to generate a probability distribution for loading actions. The time complexity of this process is only [time complexity missing]. It is far superior to the exponential complexity of iterative search algorithms.

[0126] Furthermore, in step S44, in order to guide the model to learn the Pareto optimal solution, the following multidimensional composite reward function R is designed, and its calculation formula is as follows:

[0127] ;

[0128] Where i is the index number of the cargo to be loaded, and S t Let S be the current load planning decision state, and R be the current load planning decision state. t The following loading action a t The comprehensive reward value is the evaluation result of the candidate loading results. w1, w2, w3, and w4 are the weights of four indicators: volume utilization rate, loading balance, loading compactness, and loading / unloading sequence, respectively. i v represents the volume of the cargo currently awaiting loading. bin This represents the total volume of the candidate loading space used for loading. To determine the current load allocation decision state S t The spatial position of each loaded and unloaded cargo is used to calculate the ratio of the overall center of gravity height to the corresponding support area, which is used as a stability index to reflect the stability of the current loading structure. To determine the current load allocation decision state S t The contact relationship between the cargo to be loaded and the already loaded cargo and the inner wall of the container after the cargo is placed, and the compactness index is constructed based on the distribution of contact surfaces between cargo and between cargo and container; To determine the current load allocation decision state S t The correspondence between the spatial placement order of each cargo and the predetermined loading and unloading order is established, and a sequence penalty index is constructed based on the reverse placement cases in the loading results. Specifically, among them, To maximize volume utilization, we encourage the inclusion of more volume. As a stability indicator, the ratio of the center of gravity height to the support area of ​​the current stacked structure is calculated, and structures with "larger bottom and smaller top, heavier bottom and lighter top" are rewarded. As a compactness metric, rewards are given for newly loaded cargo that has more contact surface with already loaded cargo; As a sequence penalty, negative feedback is given to the number of inversion pairs in the loading and unloading order; w1, w2, w3, and w4 are the weights of the four indicators, respectively.

[0129] Further, step S5 includes:

[0130] Step S51: Convert the target loading scheme into a loading execution instruction set and send the loading execution instruction set to the corresponding loading unit. The loading execution instruction set shall at least include the loading unit identifier, spatial placement parameters and loading sequence parameters corresponding to each cargo to be loaded.

[0131] Step S52: During the loading process, collect the actual load data of each loading unit in real time, and construct the current load distribution status of the container based on the collected actual load data, as the current loading decision status;

[0132] Step S53: Calculate the difference between the current load allocation decision state and the current load distribution state. If the difference exceeds the preset threshold or if the stability constraint is not met, pause the loading execution process and trigger the reconfiguration mechanism; otherwise, continue to execute the corresponding loading action according to the loading execution instruction set.

[0133] Step S54: When the reconstruction mechanism is triggered, identify the loading units whose contribution to the load distribution deviation or center of gravity offset exceeds the corresponding threshold, determine them as the affected loading units, and extract the corresponding adjustable loading units from the maximum remaining space set to form a local adjustment area;

[0134] Step S55: Remove the goods in the local adjustment area from the current loading state and put them back into the waiting queue. Use the local adjustment area as a new leaf node, and call the deep reinforcement learning model in S4 again to perform container loading decision calculation on the local adjustment area to generate an updated container loading plan.

[0135] Second Embodiment

[0136] Based on the same inventive concept, this invention also provides an intelligent container stowage system for China-Europe freight trains oriented towards the international supply chain, employing the aforementioned intelligent container stowage method for China-Europe freight trains oriented towards the international supply chain, including:

[0137] The response module is used to respond to container loading task requests from China-Europe freight trains. The loading task request is used to perform loading calculations on the containers of the target train.

[0138] The data acquisition module is used to obtain the basic loading data corresponding to the target train from multiple data interfaces based on the loading task request. The basic loading data includes the order data of the goods to be loaded, container specifications and load constraint parameters, and customs clearance status data.

[0139] The data processing module is used to establish three-dimensional spatial constraints on the target container based on the loading basic data, including the geometric boundaries of the container and the size restrictions of the cargo. Based on the three-dimensional spatial constraints, the internal space of the target container is continuously divided into spaces to form a continuous available space model representing the loadable area inside the container. Then, the continuous available space model is organized through the packing configuration tree to construct a continuous space model inside the container. The load-bearing convex polygon stability algorithm is introduced to determine the center of gravity position of the cargo to be loaded in the candidate loading space and establish stability constraints that conform to the mechanical principles.

[0140] The output module utilizes a Transformer-based deep reinforcement learning model to encode the cargo features and spatial features of the cargo to be loaded with attention. Under the constraints of stability constraints and physical perception masks, it generates corresponding container loading decisions. The container loading decision evaluates candidate loading results using a multi-dimensional composite reward function that includes volume utilization, loading balance, loading compactness, and loading / unloading sequence. Based on the evaluation results, a target loading scheme is determined. The target loading scheme is used as the container loading scheme, and the load distribution status is monitored during the loading process. When the load distribution status is found to be inconsistent with the preset constraints, local adjustments are made to the affected loading units. The deep reinforcement learning model is then called again through the packaged configuration tree to calculate a new container loading scheme.

[0141] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A smart container loading method for China-Europe freight trains oriented towards international supply chains, characterized in that, include: Step S1: Respond to the China-Europe freight train container loading task request; Step S2: Obtain the basic loading data corresponding to the target train based on the loading task request. The basic loading data includes the order data of the goods to be loaded, container specifications and load constraint parameters, and customs clearance status data. Step S3: Establish a three-dimensional spatial constraint for the target container based on the basic loading data, including the container's geometric boundaries and cargo size limitations. Based on the three-dimensional spatial constraint, continuously divide the internal space of the target container to form a continuous available space model representing the loadable area inside the container. Then, organize the continuous available space model through a packing configuration tree to construct a continuous space model inside the container, introducing load capacity. The convex polygon stability algorithm determines the center of gravity position of the cargo to be loaded in the candidate loading space and establishes stability constraints that conform to mechanical principles; Step S4: Using a Transformer-based deep reinforcement learning model, attention encoding is performed on the cargo features and spatial features of the cargo to be loaded, and corresponding container loading decisions are generated under the constraints of the stability constraints and the physical perception mask. The container loading decisions evaluate the candidate loading results by using a multi-dimensional composite reward function that includes volume utilization, loading balance, loading compactness and loading and unloading order, and determine the target loading scheme based on the evaluation results. Step S5: Use the target loading scheme as the container loading scheme and monitor the load distribution status during the loading process. When it is detected that the load distribution status does not meet the preset constraints, perform local adjustments on the affected loading units and recalculate the new container loading scheme by calling the deep reinforcement learning model through the packing configuration tree.

2. The intelligent container loading method for China-Europe freight trains oriented towards international supply chains as described in claim 1, characterized in that, Step S1 includes: Step S11: Receive the loading task request from the railway logistics hub dispatch center, and obtain the train identification information and corresponding planned departure time window of the target train; Step S12: Perform validity verification on the loading task request, determine that the target train is in a preset operating state that allows the loading calculation to be performed, and when the validity verification is passed, generate a loading calculation trigger command to start the loading calculation of the target train.

3. The intelligent container loading method for China-Europe freight trains oriented towards international supply chains as described in claim 1, characterized in that, Step S2 includes: Step S21: Obtain the cargo order data associated with the target train from the cargo order data interface, and perform parameterization processing on the geometric dimension parameters, weight parameters, and category information of the cargo to be loaded to generate a cargo parameter set; Step S22: Obtain the container specifications and load constraint parameters corresponding to the target train from the container resource management interface, and extract the corresponding container capacity parameters and maximum load-bearing parameters; Step S23: Obtain the customs clearance status data corresponding to the cargo order data to be loaded from the customs clearance status interface, and filter out the cargo parameter set that meets the loading calculation conditions; Step S24: Associate the filtered cargo parameter set, the container capacity parameters, and the maximum load-bearing parameters to form a basic parameter set for the construction of the three-dimensional spatial constraint structure as the loading basic data.

4. The intelligent container loading method for China-Europe freight trains oriented towards international supply chains as described in claim 3, characterized in that, Step S3 includes: Step S31: Using the inner corner point of the target container as a spatial reference point, construct a three-dimensional Cartesian coordinate system with the length, width, and height directions corresponding to the X-axis, Y-axis, and Z-axis, respectively. Update the internal space of the target container in this three-dimensional Cartesian coordinate system using geometric difference operations to generate a maximum remaining space set representing the loadable area inside the container. Step S32: Based on the maximum remaining space set, organize the current loading state of the target container into a hierarchical packing configuration tree to form the continuous spatial model. Each leaf node of the packing configuration tree corresponds to a candidate loading space available for loading in the maximum remaining space set. The spatial characteristic parameters of the loading space include spatial dimensions and spatial coordinates; Step S33: For the candidate loading space corresponding to the leaf node of the packing configuration tree in the continuous spatial model, construct the load-bearing convex polygon of the support surface formed by the load-bearing goods below the candidate loading space, and determine whether the center of gravity projection of the goods to be loaded at the candidate loading space is located within the load-bearing convex polygon. If the center of gravity projection is located within the load-bearing convex polygon, the candidate loading space satisfies the stability constraint requirement; otherwise, the candidate loading space does not satisfy the stability constraint, and the candidate loading space is marked as a candidate loading space that cannot be used for the container loading decision.

5. The intelligent container loading method for China-Europe freight trains oriented towards international supply chains according to claim 4, characterized in that, Step S4 includes: Step S41: Using a multi-head self-attention mechanism and a graph neural network, feature modeling is performed on the dependency relationship of the cargo sequence and the spatial topology of the packing configuration tree, respectively, to construct a cargo feature vector and a spatial topology feature vector reflecting the current loading state. This includes: constructing a cargo sequence to be loaded according to the predetermined loading and unloading order of the cargo to be loaded; based on the cargo sequence to be loaded, using the multi-head self-attention mechanism to encode the dependency relationship between each cargo to be loaded, and obtaining the cargo feature vector reflecting the loading correlation between the cargo to be loaded; wherein, the dependency relationship includes the dependency relationship based on the cargo sequence to be loaded. The loading combination association formed by the destination attributes of the loaded goods and / or the loading order priority association formed based on the weight parameters of the goods to be loaded; based on the continuous spatial model, for each leaf node in the packing configuration tree, the spatial features of the corresponding candidate loading space are obtained, and the leaf node is used as the graph node as the processing object of the graph neural network to generate the corresponding spatial topology feature vector; Step S42: According to the cargo feature vector and the spatial topology feature vector, the loading matching of the goods to be loaded and the candidate loading space in the continuous spatial model is performed, combined with step S33. The process involves: judging whether the candidate loading space meets the stability constraint and the three-dimensional spatial constraint; filtering the candidate loading spaces; marking those that do not meet the stability constraint or the three-dimensional spatial constraint as unselectable loading spaces and removing them from the maximum remaining space set for masking during the container loading decision calculation; step S43: based on the filtered maximum remaining space set, using the cross-attention layer of the Transformer decoder to perform association calculations on the cargo feature vector and the spatial topology feature vector to obtain the matching degree score corresponding to each selectable candidate loading space; and generating the probability distribution of loading actions based on the matching degree score to determine the candidate loading result, which includes the target placement position and corresponding placement posture of the cargo to be loaded; step S44: constructing a multi-dimensional composite reward function that includes the volume utilization rate, the loading balance, the loading compactness, and the loading and unloading order to comprehensively evaluate the candidate loading result; and optimizing the loading action output by the deep reinforcement learning model based on the evaluation result to generate the corresponding target loading scheme.

6. The intelligent container loading method for China-Europe freight trains oriented towards international supply chains according to claim 5, characterized in that, Step S42 includes: determining whether the current cargo to be loaded placed in the candidate loading space corresponding to any of the leaf nodes satisfies the stability constraint or the three-dimensional space constraint; when the stability constraint or the three-dimensional space constraint is not satisfied, the corresponding loading action is marked as an unexecutable loading action, and a first mask identifier -∞ is assigned to the unexecutable loading action to shield the unexecutable loading action during the container loading decision calculation; when the stability constraint and the three-dimensional space constraint are satisfied, the corresponding loading action is marked as an executable loading action, and a second mask identifier 0 is assigned to the executable loading action to allow the executable loading action to participate in the container loading decision calculation.

7. The intelligent container loading method for China-Europe freight trains oriented towards international supply chains as described in claim 5, characterized in that, Step S43 includes: using the cargo feature vector as a query feature, and the spatial topology feature vector as a key feature and a value feature; performing association calculations on the query feature, the key feature, and the value feature through the cross-attention layer to obtain the matching score corresponding to each candidate loading space; based on the matching score, performing probabilistic calculations on the loading actions corresponding to each candidate loading space to generate the probability distribution of the loading actions; selecting the loading action with the highest probability according to the probability distribution; and determining the corresponding candidate loading space and the placement posture as the target placement position of the cargo to be loaded, wherein the formula for the probabilistic calculation is... S t Given the current load allocation decision state, a t For the loading action, For the loading action a t In the current load decision state S t The probability distribution is given below, where Mask is a given loading action a. t The mask identifier, Logits is the loading action a t Match score, The function is used for the loading action a t The matching score is normalized to generate the probability distribution of the loading action.

8. The intelligent container loading method for China-Europe freight trains oriented towards international supply chains as described in claim 7, characterized in that, In step S44, the calculation formula for the multidimensional composite reward function is as follows: Where i is the index number of the cargo to be loaded, and S t S represents the current load allocation decision state. t The following describes the loading action a t The comprehensive reward value, which is the evaluation result of the candidate loading results, where w1, w2, w3, and w4 are the weights of the four indicators: volume utilization rate, loading balance, loading compactness, and loading / unloading sequence, respectively. i v represents the volume of the cargo to be loaded. bin The total volume of the candidate loading space used for loading; Pyramid(S t ) is based on the current load allocation decision state S t The spatial positions of the loaded goods and the goods to be loaded are used to calculate the ratio of the overall center of gravity height to the corresponding support area, which is a stability index reflecting the stability of the current loading structure; Compactness (S t ) is based on the current load allocation decision state S t The contact relationship between the goods to be loaded and the already loaded goods and the inner wall of the container after placement is described below, and the compactness index is constructed based on the distribution of contact surfaces between goods and between goods and the container; SeqPenalty(S t ) is based on the current load allocation decision state S t The correspondence between the spatial placement order of each cargo and the predetermined loading and unloading order is established, and a sequence penalty index is constructed based on the reverse placement situation in the loading results.

9. The intelligent container loading method for China-Europe freight trains oriented towards international supply chains as described in claim 8, characterized in that, Step S5 includes: Step S51: Converting the target loading scheme into a loading execution instruction set and issuing the loading execution instruction set to the corresponding loading unit. The loading execution instruction set includes at least the loading unit identifier, spatial placement parameters, and loading sequence parameters corresponding to each cargo to be loaded; Step S52: During the loading execution process, collecting the actual load data of each loading unit in real time, and constructing the current load distribution state of the container based on the collected actual load data as the current loading decision state; Step S53: Calculating the difference between the planned current loading decision state and the current load distribution state. If the difference exceeds a preset threshold or the stability constraint is not met, the loading is paused. The loading execution process is described and a reconfiguration mechanism is triggered; otherwise, the corresponding loading action continues to be executed according to the loading execution instruction set; Step S54: When the reconfiguration mechanism is triggered, the loading unit whose contribution to the load distribution deviation or center of gravity offset exceeds the corresponding threshold is identified, the affected loading unit is determined, and the corresponding adjustable loading unit is extracted from the maximum remaining space set to form a local adjustment area; Step S55: The goods in the local adjustment area are removed from the current loading state and put back into the waiting queue, the local adjustment area is used as the new leaf node, and the deep reinforcement learning model in S4 is called again to perform the container loading decision calculation on the local adjustment area to generate the updated container loading scheme.

10. A smart container stowage system for China-Europe freight trains oriented towards international supply chains, employing the smart container stowage method for China-Europe freight trains oriented towards international supply chains as described in any one of claims 1 to 9, characterized in that, include: The response module is used to respond to container loading task requests from China-Europe freight trains; The data acquisition module is used to obtain the basic loading data corresponding to the target train from the multi-source data interface based on the loading task request. The basic loading data includes the order data of the goods to be loaded, container specifications and load constraint parameters, and customs clearance status data. The data processing module is used to establish three-dimensional spatial constraints on the target container based on the loading baseline data, including the container's geometric boundaries and cargo size limitations. Based on these three-dimensional spatial constraints, the internal space of the target container is continuously divided to form a continuous available space model representing the loadable area inside the container. This continuous available space model is then organized using a packing configuration tree to construct a continuous space model inside the container. A load-bearing convex polygon stability algorithm is introduced to determine the center of gravity position of the cargo to be loaded in the candidate loading space, establishing stability constraints that conform to mechanical principles. The result output module is used to utilize a Transformer-based deep reinforcement learning model to analyze the cargo... Attention encoding is performed on object features and spatial features, and corresponding container loading decisions are generated under the constraints of stability constraints and physical perception masks. The container loading decisions evaluate candidate loading results using a multi-dimensional composite reward function that includes volume utilization, loading balance, loading compactness, and loading / unloading sequence. Based on the evaluation results, a target loading scheme is determined. The target loading scheme is used as the container loading scheme, and the load distribution status is monitored during the loading process. When the load distribution status is detected to be inconsistent with the preset constraints, local adjustments are performed on the affected loading units, and the deep reinforcement learning model is called again through the packaging configuration tree to calculate a new container loading scheme.

Images