Intelligent bulk commodity freight logistics management method and system

Abstract

The invention discloses an intelligent bulk commodity freight logistics management method and system, and belongs to the technical field of logistics management, and the method comprises the steps: obtaining the multi-dimensional information of a transportation task, and generating a task semantic tag set; collecting state data and performance behaviors of the transportation units to construct a resource feature vector set; constructing a multi-model scheduling system including transportation risks, node passing trends and loading and unloading waiting prediction, and generating candidate scheduling strategies; performing game optimization on the strategy set in a dynamic environment through a reinforcement learning optimizer, and outputting an optimal scheduling strategy; deploying a continuous monitoring mechanism at edge nodes, and realizing adaptive reconfiguration of paths and resources according to abnormal scores; the intelligent level and the response capability of transportation scheduling can be remarkably improved, the delay risk is reduced, the resource utilization rate is improved, and the method is suitable for a bulk commodity multi-path multi-node combined transportation scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of logistics management, and in particular to an intelligent bulk commodity freight logistics management method and system. BACKGROUND

[0002] In the current bulk commodity transportation system, the transportation of materials such as coal, steel, and chemical raw materials mainly relies on manual route planning, fixed scheduling rules, and static time estimation. Although some systems have introduced path planning and vehicle monitoring methods, they mostly remain at the stage of static information processing and single-dimensional optimization, and have not yet built an intelligent management framework that can dynamically optimize multi-dimensions according to transportation task characteristics, changes in operating state, and path environmental risks.

[0003] In addition, existing systems cannot automatically adjust transportation resources and scheduling strategies according to changes in different commodity types, loading and unloading methods, transportation frequencies, and path traffic capacity, and lack real-time closed-loop feedback mechanisms. For example, in the case of multiple vehicle fleets and multiple tasks running simultaneously, resource conflicts, task delays, and even path congestion transmission effects may occur, making it difficult to adapt to highly dynamic and uncertain demands in actual operation.

[0004] Therefore, there is an urgent need for an intelligent logistics management method that has data-driven decision-making capabilities, dynamic environmental adaptation capabilities, and multi-dimensional collaborative optimization capabilities. In particular, it addresses core issues such as dynamic cargo identification, path adaptive adjustment, risk trend prediction, and intelligent resource matching in the bulk commodity transportation process, and builds an intelligent scheduling system with full-process, low-intervention, and high-response capabilities to achieve truly intelligent management of bulk logistics. SUMMARY

[0005] The purpose of the present application is to provide an intelligent bulk commodity freight logistics management method and system to address the shortcomings in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solution: an intelligent bulk commodity freight logistics management method, comprising: S100, obtaining multi-dimensional information of bulk commodity transportation tasks and generating a task semantic label set T; S200, obtaining real-time operating state data, historical task performance data, and behavior patterns of all transportation units in the transportation network, and generating a transportation resource feature vector set V; S300, based on the task semantic label set T and the resource feature vector set V, constructing a multi-model integrated scheduling system including a transportation risk prediction sub-model, a node traffic trend sub-model, and a loading and unloading waiting prediction sub-model, for generating multiple candidate scheduling strategies; S400, input the reinforcement learning optimizer for each candidate scheduling strategy, combine the path network accessibility outliers, node predicted congestion degree, weather change trend and real-time feedback data, iteratively game optimize the strategy set, and output the global optimal scheduling strategy S* under the current conditions; S500, issue the strategy S* to the scheduling execution system, which is distributed in multiple edge computing nodes, continuously monitor the transportation process, and adaptively reconfigure the current execution path and resource allocation according to the monitoring and analysis results.

[0007] Preferably, the construction of the multi-model integrated scheduling system comprises: S301, based on the historical transportation task sample set, for each type of goods and its transportation scene, extract the high-frequency combination mode in the task semantic label set T, and generate the embeddable task semantic vector T' through the high-order semantic nested coding method, wherein the coding method uses a graph convolution network with attention mechanism to extract task correlation features; S302, for the resource feature vector set V, use a time series residual convolution network to perform feature denoising and multi-scale feature extraction, and obtain a dynamically updated resource capability expression vector V'; S303, taking the task semantic vector T' and the resource capability vector V' as inputs, respectively construct a transportation risk prediction sub-model, a node passing trend sub-model and a loading and unloading waiting prediction sub-model, wherein the sub-models all adopt a multi-channel attention gate recurrent unit structure, and can output confidence indicators and abnormal early warning indicators in each prediction dimension respectively; S304, according to the prediction outputs of the three sub-models, the scheduling preferences in different dimensions are fused and evolved through a strategy clustering evolution mechanism, forming a set of candidate scheduling strategies {S1, S2, …, Sn} with evaluable confidence factors, wherein each strategy includes scheduling path, execution window and interference tolerance configuration parameters, and n represents the total number of generated scheduling strategies.

[0008] Preferably, the iteratively game optimizing the strategy set comprises: S401, input the candidate scheduling strategy set {S1, S2, …, Sn} as the state space, and construct a reinforcement learning optimizer based on a double-adversarial structure; S402, collect path network accessibility outliers, node predicted congestion degree, real-time weather heat map and vehicle operation feedback data, construct a multi-source fusion environment tensor E, and simulate and evaluate the response trajectory of each strategy in the E environment; S403, introduce a Bayesian game dynamic incentive mechanism in the simulation process, score the utility and compensate the risk of each round of strategy confrontation result, and update the strategy network weight; S404, output the optimal scheduling strategy S* based on the converged main strategy network.

[0009] Preferably, the step of calculating the path network accessibility anomaly value for the reinforcement learning optimizer input comprises: S411, obtaining a historical passing time delay sequence of nodes on the target transportation path and real-time time delay data in the current period , constructing a comparison matrix within a time window Δt , wherein adopting 5% truncated mean processing to eliminate abnormal outliers; S412, based on the topological correlation of node traffic state, using spectral clustering method to structure classify the accessibility change trend of adjacent nodes on the path, and identifying the local fluctuation critical region corresponding to the cluster center point; S413, constructing accessibility disturbance function for each node , wherein ΔT is the delay offset, is the historical time delay variance, and α and β are robust coefficients adjusted according to the task type, and the disturbance intensity value of each node is obtained; S414, if the disturbance intensity value exceeds the reference accessibility tolerance threshold ε, the corresponding node is marked as an abnormal node, and the topological importance in the path structure is taken as the weight to comprehensively output the path network accessibility anomaly value δ.

[0010] Preferably, the continuous monitoring and adaptive reconfiguration of the transportation process comprises: S501, deploying edge perception units at key nodes of the transportation path, collecting vehicle running trajectory, cargo environment state and path micro traffic data in real time, generating a heterogeneous data stream set D, and performing multi-scale time sequence segmentation processing on D through a hierarchical sliding window algorithm; S502, using an integrated anomaly detection model to quickly distinguish the segmented data, the model integrates isolation forest and probabilistic graph model, outputs a risk score vector R of multi-dimensional monitoring indicators, and determines whether there is a path anomaly or resource imbalance event based on a time-varying threshold mechanism; S503, if a certain dimension indicator in the risk score vector R exceeds its adaptive threshold interval for N consecutive times, the reconfiguration module is triggered, the resource-path bidirectional mapping table is called, and replaceable candidate transportation units and feasible alternative path set are identified; S504, through the resource reconfiguration engine based on graph neural network, the current task execution state and candidate resources are matched and sorted, and the execution path and transportation unit allocation are dynamically updated.

[0011] Preferably, the matching and sorting of the current task execution state and candidate resources and the dynamic updating of the execution path and transportation unit allocation comprise: S511, encode the state information of the current transportation task into a state vector including the remaining time limit of the task, the current path congestion coefficient and the remaining energy consumption of the transportation unit, as input features; S512, generate a resource evaluation vector for all candidate transportation units The vector is obtained by fusing historical task completion rate, path adaptability score and fault prediction probability, and a multi-objective orthogonal projection analysis algorithm is used to sort it; S513, bidirectional conditional matching coding of the state vector and the sorted resource evaluation vector Input into the path reconstruction module based on the graph attention mechanism to generate the transportation unit-path pair with the optimal adaptation degree; S514, call the scheduling control system to seamlessly reconfigure the current transportation plan.

[0012] The application also provides an intelligent bulk commodity freight logistics management system, comprising: A task information processing module acquires multi-dimensional information of bulk commodity transportation tasks and generates a task semantic label set T; An ability modeling module acquires real-time running state data, historical task performance data and behavior patterns of all transportation units in the transportation network and generates a transportation resource feature vector set V; A multi-model scheduling prediction module, based on the task semantic label set T and the resource feature vector set V, constructs a multi-model integrated scheduling system including a transportation risk prediction sub-model, a node passing trend sub-model and a loading and unloading waiting prediction sub-model, for generating multiple candidate scheduling strategies; A reinforcement game optimization module inputs a reinforcement learning optimizer to each candidate scheduling strategy, combines path network accessibility outliers, node predicted congestion degree, weather change trend and real-time feedback data, iteratively game optimizes the strategy set, and outputs a globally optimal scheduling strategy S* under the current conditions; A dynamic reconfiguration module, which distributes the strategy S* to the scheduling execution system, is distributed in multiple edge computing nodes, continuously monitors the transportation process, and adaptively reconfigures the current execution path and resource allocation according to the monitoring and analysis results.

[0013] In the above technical solution, the application provides technical effects and advantages: 1. This invention provides an intelligent bulk commodity freight logistics management method. By constructing a task semantic label set and a transportation resource feature vector set, and combining multi-model prediction and reinforcement learning optimization, it can achieve refined matching and scheduling based on task attributes and resource status, significantly improving the accuracy and personalization of scheduling decisions. Compared with traditional rule-based path planning and static resource allocation methods, this invention can dynamically generate and optimize candidate strategies under multiple paths, multiple resources, and multiple constraints, exhibiting strong adaptability and high scheduling efficiency, making it particularly suitable for the frequently changing scheduling environment in bulk commodity transportation.

[0014] 2. This invention employs a distributed edge node deployment approach to achieve continuous monitoring and adaptive reconfiguration of the transportation process, exhibiting strong robustness and responsiveness to environmental disturbances. By integrating accessibility anomaly identification, resource status anomaly detection, and path self-reconfiguration mechanisms, the system can automatically and rapidly reconfigure resources and paths in the event of sudden congestion, vehicle malfunctions, or node congestion, ensuring uninterrupted transportation tasks and significantly reducing transportation delay rates and resource idle rates. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the method of the present invention.

[0017] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1, please refer to Figure 1 As shown in this embodiment, an intelligent bulk commodity freight logistics management method includes: S100: Obtain multi-dimensional information on bulk commodity transportation tasks and generate a task semantic tag set T; S200, acquire real-time running state data, historical task fulfillment data and behavior patterns of all transportation units in the transportation network, and generate a transportation resource feature vector set V; S300, based on the task semantic label set T and the resource feature vector set V, construct a multi-model integrated scheduling system including a transportation risk prediction sub-model, a node traffic trend sub-model and a loading and unloading waiting prediction sub-model, for generating multiple candidate scheduling strategies; S400, input each candidate scheduling strategy into a reinforcement learning optimizer, combine the path network accessibility outliers, node predicted congestion degree, weather change trend and real-time feedback data, and perform iterative game optimization on the strategy set to output a global optimal scheduling strategy S* under the current conditions; S500, issue the strategy S* to a scheduling execution system, which is distributed on multiple edge computing nodes, continuously monitors the transportation process, and adaptively reconfigures the current execution path and resource allocation according to the monitoring and analysis results.

[0020] The embodiment provides a task semantic label construction method for bulk commodity transportation scheduling, which is suitable for the task perception and classification stage in an intelligent logistics management system, and the specific process is as follows: The system receives a transportation task instruction from a business platform, which contains the following original task information fields: cargo type (such as coal, ore, steel); starting point and destination code (GIS code or railway / port node); cargo volume (unit: m³), mass (unit: ton); required arrival time and time window width (unit: hour); unloading method (bucket wheel machine, grab type, belt conveyor, etc.); danger level (such as whether it involves chemical flammable goods); transportation restriction conditions (prohibition of night operation, closure of specific sections, etc.).

[0021] After the system structurally processes the above information, it enters the task label construction process. First, a coding strategy based on rules and models is used to map the original fields to semantic dimensions, and the dimensions are shown as follows: : cargo attribute label (for example: fragile / incompressible / high dust); : time constraint level (loose, medium, urgent); : path adaptability requirement (whether it is suitable for port-rail joint transportation, long-distance high-speed transportation, etc.); : operation coordination complexity (whether it involves multi-node interactive loading and unloading); : safety control level (ordinary, key supervision, special protection); In the process of generating labels, the system introduces a multi-task classification model, which is trained based on historical scheduling data, uses BERT variants for text field embedding, and combines multi-channel fully connected layers to realize label prediction and correction, and finally generates a structured label set T.

[0022] To improve the expressiveness of labels, the system also introduces a graph embedding module to project task labels into a task graph, so that labels not only have ontology description capability, but also support subsequent inter-label relationship calculation (such as conflictiveness, coupling degree, etc.), to serve scheduling strategy matching.

[0023] The final generated task semantic label set T will be input into the subsequent resource matching and scheduling path evaluation module to form the pre-cognition layer of the intelligent scheduling process.

[0024] Through the above embodiments, the present application can perform multi-dimensional feature structured modeling on bulk commodity transportation tasks, realize the mapping process from raw data to semantic abstraction, provide high expressiveness input for subsequent intelligent matching scheduling strategy, and has stronger generalization ability and scheduling adaptation ability compared with traditional task allocation methods based on field screening.

[0025] The embodiment provides a method for generating a transportation resource feature vector set V, which is suitable for state modeling and scheduling adaptability evaluation stage of multi-source heterogeneous transportation units, and the specific process is as follows: The system collects real-time running related data of each transportation unit (including trucks, train sets, port transfer vehicles, etc.) in the transportation network before the transportation task starts. The data includes: Real-time running state data: including current position (latitude and longitude), current load (tons), running speed (km / h), remaining oil / electricity capacity (%), traffic area congestion index; Historical task performance data: number of tasks executed in the past 30 days, average punctuality rate, longest / shortest task execution time, historical path type coverage rate; Behavioral pattern data: including driver operation stability indicators (frequent braking, frequent acceleration), task interruption frequency, maintenance record density, abnormal report frequency, etc.

[0026] After the above data is processed by format standardization and time sequence adjustment, the system introduces a time sequence feature convolution network to compress and model the behavior sequence, so as to extract the running rule and behavior pattern trend of the transportation unit, and generate a one-dimensional feature embedding vector V_b.

[0027] At the same time, a statistical feature extraction module is introduced to normalize the performance and state data, including mean, variance, skewness coefficient, and outlier density, to generate a static capability vector V_s of the transportation unit in performance.

[0028] The dynamic behavior embedding vector V_b is spliced with the static capability vector V_s, and an attention weighted fusion layer is introduced to reconstruct the features, so as to eliminate dimensional differences and time sequence fluctuation interference, and finally output a uniform dimension transport resource feature vector .

[0029] The above process is repeatedly performed on all transport units in the transport network, and finally a resource feature vector set V is obtained, which is used for subsequent matching with the task label set T and the generation process of the scheduling strategy.

[0030] By fusing the static fulfillment capability and the dynamic behavior pattern into a unified feature vector, the present application can more accurately depict the real schedulable capability and risk level of the transport unit, solve the problem of estimating the resource capability only according to the GPS position and the load in the traditional method, and improve the reliability and system adaptability of the scheduling strategy matching.

[0031] The embodiment provides a construction method of a multi-model integrated scheduling system, which is used for multi-dimensional correlation analysis of a task semantic label set and a transport resource capability vector, and generates a structured candidate scheduling strategy based on a prediction result, and the specific steps are as follows: Before constructing the scheduling system, the system first calls a historical transport task sample set, including task labels, scheduling results, execution feedback and the like. For different types of goods (such as coal, steel and liquid chemical products) and their typical transport paths (such as port-rail intermodal transport, plateau route and multi-node transfer), the highest frequency label combination mode in the task semantic label set T is counted, such as “high time efficiency + high dust + long distance path”.

[0032] The label combination is vectorized by using a high-order semantic nested coding method, a graph convolution network (GAT-GCN) with an attention mechanism is used to construct a graph structure between the task labels, the implicit coupling features between the tasks are extracted through self-learning of the weights between the nodes, and an embeddable task semantic vector T' is generated, which is a high-order expression of the task semantics.

[0033] The transport resource feature vector set V is input into a time sequence residual convolution network, which retains the continuity of the historical behavior sequence, uses a residual path to perform feature noise reduction and nonlinear conversion, extracts multi-scale time sequence features, and obtains a resource capability expression vector V', which reflects the stability, path adaptability and behavior credibility of the transport unit.

[0034] T' and V' are taken as inputs to construct three sub-models respectively: A transport risk prediction sub-model: outputting the probability of existing time delay, interruption or deviation from the trajectory in the future operation; A node traffic trend sub-model: predicting potential blocking risks based on historical traffic of the node and real-time feedback; Loading and unloading waiting prediction sub-model: predict the service capacity and queuing time of the loading and unloading point within the target time window.

[0035] The above sub-models all adopt a multi-channel attention gate recurrent unit structure (MC-Attn-GRU), each channel independently models a monitoring index dimension, and focuses on the mutation trend area through a channel attention mechanism, and finally outputs the confidence score of each dimension prediction and the abnormal early warning index.

[0036] After obtaining the prediction results, the system introduces a strategy clustering evolution mechanism, and inputs the prediction index set to a strategy clusterer. Based on the K-means++ and density dynamic adaptive evolution mechanism, the cluster center is formed in the strategy space, and a set of candidate scheduling strategies {S1, S2, …, Sn} with scheduling paths, execution window configurations, and interference tolerance levels are generated, where n represents the total number of strategies.

[0037] Each candidate strategy is accompanied by a confidence index, which represents the robustness and response ability of the strategy under the current state, for subsequent optimizer iteration selection.

[0038] The multi-model integrated scheduling system constructed by the embodiment can not only realize high-order expression of task and resource characteristics, but also can predict and analyze the risk dimension in the transportation process, and generate a confident multi-strategy output on this basis, which provides a solid foundation for subsequent reinforcement learning optimization and dynamic scheduling, and significantly improves the scheduling decision quality and the ability to adapt to complex environments.

[0039] The embodiment provides a method for iterative game optimization of a candidate scheduling strategy set, which is used for dynamically selecting a current global optimal transportation scheduling strategy under a variable traffic environment and a transportation resource state, and the specific steps are as follows: After the system completes the construction of the candidate scheduling strategy set {S1, S2, …, Sn}, the system inputs the set as a state space to construct a reinforcement learning optimizer based on a double-adversarial structure. The optimizer is composed of two cooperative training networks: a main strategy network adopts a variable structure attention type Actor-Critic model to evaluate the value of the scheduling strategy and output actions; and an auxiliary strategy network is used to simulate the change of the strategy robustness under path disturbance to form an adversarial training.

[0040] To realize dynamic environment modeling, the system obtains the following indexes from the real-time data platform: ① path network accessibility abnormal value; ② node predicted congestion degree (based on historical + short-term trend prediction); ③ weather heat map in the target area (such as rainfall, wind speed level, etc.); and ④ current vehicle running state feedback (such as fuel consumption rate, off-track frequency, etc.). The system encodes the above heterogeneous data into a unified format to construct a high-dimensional tensor E.

[0041] The calculation step of the path network reachability anomaly value for the reinforcement learning optimizer input comprises: Obtaining a historical passing time delay sequence of nodes on a target transportation path And real-time time delay data in the current period Constructing a comparison matrix in a time window Δt Wherein Adopting 5% tail mean processing to eliminate abnormal outliers; Based on the topological correlation of node traffic states, the spectral clustering method is used to classify the structural trend of the reachability of adjacent nodes on the path, and the local fluctuation critical region corresponding to the cluster center point is identified; A reachability disturbance function is constructed for each node Where ΔT is a delay offset, is a historical time delay variance, and α and β are robust coefficients adjusted according to the task type, and the disturbance intensity value of each node is obtained; If the disturbance intensity value exceeds the reference reachability tolerance threshold ε, the corresponding node is marked as an abnormal node, and the topological importance of the node in the path structure is used as a weight to comprehensively output the path network reachability anomaly value δ.

[0042] Under the driving of the multi-source fusion environment tensor E, the simulation trajectory of each candidate strategy Si is evaluated, and the task completion rate, average delay time, path switching frequency and resource occupation redundancy under the disturbance condition are recorded to form a strategy response trajectory set R = {r1, r2,..., rn}.

[0043] In the strategy training process, a Bayesian game dynamic incentive mechanism is introduced to convert the inter-strategy antagonistic game into a benefit-risk trade-off problem. In each simulation round, the system calculates the expected utility value U(Si) of the strategy response trajectory, and evaluates its disturbance sensitivity σ(Si). Through the Bayesian inference model, a risk adjustment coefficient γ(Si) is given to the strategy, so as to dynamically adjust the learning weight parameters of the main strategy network.

[0044] When the main strategy network converges to a stable utility distribution in continuous training rounds, the system outputs the optimal scheduling strategy S* under the current environmental conditions, and records its confidence score and adaptive environmental feature vector for subsequent dynamic reconfiguration module calling.

[0045] Through the iterative game optimization mechanism constructed by the embodiment, the system can actively perceive the strategy execution risk under variable environmental conditions, and dynamically update the optimization direction with the help of antagonistic learning and Bayesian game theory. Compared with the traditional reinforcement learning optimization method, the method can simultaneously consider the strategy utility and risk redundancy, and has stronger strategy generalization ability and environmental adaptability.

[0046] The embodiment provides a method for realizing continuous monitoring and adaptive reconfiguration in the process of bulk commodity transportation, which is used for perceiving real-time changes of transportation paths, resource states and external environment during the execution of transportation tasks, and automatically performing dynamic reconfiguration of scheduling resources and execution paths after identifying abnormal states, so as to guarantee the reliability and continuity of the transportation tasks. The method contains two nested structures: one is a multi-source state monitoring and reconfiguration triggering mechanism at the path level and the resource level; the other is a resource matching and path adaptation updating mechanism based on the current task state.

[0047] At the key nodes of the transportation path (such as toll stations, unloading ports, port truck zones and the like), the system deploys an edge perception unit, which has local operation and low-delay data transmission capability. The perception unit collects the following three types of data in real time: vehicle running trajectory data, including vehicle GPS position, speed vector, trajectory deviation rate; cargo environment state data, such as temperature and humidity, vibration frequency, and box gas concentration; path microscopic traffic data, such as instantaneous accessibility, average travel time, and micro-jam duration of road segments.

[0048] The collected multi-type heterogeneous data is uniformly packaged into a data stream set D, and then input into a layered sliding window algorithm for timing processing. The algorithm uses a short window (such as 10s) on the high-frequency data dimension (such as vehicle trajectory), and uses a long window (such as 5min) on the low-frequency change dimension (such as node traffic rate), so as to guarantee the dual perception ability of sudden change and trend fluctuation, and to segment and archive the data for subsequent calling by the abnormal detection model.

[0049] The system inputs the sliding segmented data into an integrated abnormal detection model. The model integrates two sub-modules: an isolation forest sub-module for detecting point mutations such as vehicle deviation and trajectory anomaly; a probabilistic graph model sub-module for modeling conditional dependency between variables and identifying potential fault chains or trend anomalies triggered by joint anomalies in multiple dimensions.

[0050] The model finally outputs a set of risk score vectors R, where each component represents the risk level of a monitoring dimension, and the numerical range is [0, 1]. To improve sensitivity, the system introduces a time-varying threshold mechanism, that is, the risk indicator threshold interval is automatically adjusted according to the historical periodic pattern and real-time distribution, so as to avoid the lack of sensitivity of static threshold to actual disturbance or false alarm.

[0051] If any component of the risk score vector R continuously exceeds its corresponding adaptive threshold interval for N times (for example, N = 3), the system will determine that the current transportation state has deviated from the normal control limit, and automatically trigger the reconfiguration module.

[0052] The reconfiguration module retrieves the following information from the current transportation network and resource pool by calling the pre-built resource-path bidirectional mapping table: An alternative set of backup transportation units R', whose screening conditions include residual carrying capacity, current accessibility, state health level, etc. A set of feasible alternative paths P', whose screening conditions include path accessibility threshold not exceeding, node service capacity not exceeding, etc.

[0053] The resource-path mapping table is organized in a graph structure, with nodes representing transportation units and path segments, and edges representing schedulable relationships, with real-time state and risk labels attached for subsequent scheduling engine calls.

[0054] After entering the reconfiguration phase, the system calls the resource reconfiguration engine based on graph neural networks. This engine takes the current task execution state and candidate resource state as input and performs the following sub-steps: The system encodes the execution state of the current transportation task into a state vector , containing the following features: Remaining transportation time limit: the number of remaining hours from the current time to the expected completion time; Current path congestion coefficient: calculated by modeling the average node accessibility and path discontinuity; Transportation unit remaining energy consumption: the product of fuel / electricity remaining ratio and unit load energy consumption ratio.

[0055] Vector as a representation of the current task constraints and risk tolerance for the matching sorting algorithm.

[0056] For each candidate transportation unit , construct a resource evaluation vector containing the following dimensions: Historical task completion rate; Path adaptability score: reflects its completion stability under the current path type; Fault prediction probability: estimated by the prediction model within the next T hours.

[0057] Use multi-objective orthogonal projection analysis algorithm to perform principal component mapping and normalized sorting on the above multi-dimensional nonlinear indicators to generate a resource priority list.

[0058] The state vector and the candidate resource evaluation vector set }input to the path reconstruction module based on the graph attention mechanism. The module adopts a bidirectional conditional matching mechanism: the task state imposes constraints on the resource priority (such as emergency tasks tend to select high-stability resources), and the resource state imposes bias on path selection (such as units with high failure probability avoid long paths). The module outputs the "transport unit-path" combination pair with the optimal adaptation degree .

[0059] The system finally calls the dispatch control interface to load the new transportation plan to the current task execution thread in an uninterrupted manner, including: The current path is switched to , and the path segment scheduling token is regenerated; The current transport unit R is released and replaced by , and the task context parameters are pushed to the new resource; The original ID and tracking information of the task are retained to realize complete closed-loop tracking.

[0060] Through the transportation process continuous monitoring and adaptive reconfiguration mechanism provided by the embodiment, the system can realize multi-dimensional dynamic perception of path state, resource state and environmental disturbance during transportation, and when potential risks are detected, intelligent rescheduling based on task state and resource characteristics is performed. Compared with the traditional scheduling system based on static rules, the method has significant real-time, robustness and task continuity guarantee capability, and is suitable for large-scale logistics scheduling tasks under multi-scene, multi-path and multi-resource constraints.

[0061] Embodiment 2: In order to verify the technical effect of the "intelligent bulk commodity freight logistics management method" in the actual transportation scene, a simulation experiment platform is constructed by integrating real scheduling data and high-fidelity path network, and a comparative test is carried out with the existing scheduling method.

[0062] Experimental area: select a typical coastal port in China as the center, covering a radius of 300 kilometers of railway-highway-port combined transport channel; Simulation task: simulate 1000 bulk commodity transportation tasks, covering four types of coal, steel, iron ore and liquid chemical products; Transport unit pool: including 300 vehicles (covering tractor groups, heavy trucks and container trains); Data source: the scheduling historical data is derived from the operation log of the region in the past three months; node traffic prediction and weather disturbance are derived from the open traffic + meteorological platform.

[0063] Control system A (traditional rule scheduling): Based on the shortest distance of the static path matching; Without considering the task semantic label; Risk-free prediction model, resource status only based on location and availability.

[0064] The system B (intelligent scheduling system) of the present application: Enable task semantic vector modeling (T'); Multi-model integrated prediction (transport risk, node trend, loading and unloading waiting); Reinforcement learning strategy optimization; Real-time monitoring + adaptive reconfiguration mechanism.

[0065] Evaluation index

[0066] Experimental results

[0067] In summary, through semantic label modeling and resource feature vector fusion, the system realizes more accurate task-resource-path matching, greatly shortening the scheduling time; The multi-model prediction mechanism identifies potential congestion and loading bottlenecks in advance, making the average delay rate decrease by more than 60%; The reinforcement learning optimizer actively avoids unstable path segments in high-risk environments, improving the adaptability of accessibility; Real-time monitoring and reconfiguration mechanism avoids long-term idle or incorrect occupation of resources, significantly improving resource utilization efficiency.

[0068] Through simulation experiments and data comparison, it is shown that the method described in the present application is superior to the prior art in multi-objective scheduling, dynamic response capability and resource utilization efficiency, has clear substantial technical progress and quantifiable beneficial effects, and can be widely applied to bulk commodity highway / railway / port joint transportation scenarios.

[0069] Example 3, please refer to Figure 2 The intelligent bulk commodity freight logistics management system described in the present embodiment comprises: A task information processing module acquires multi-dimensional information of bulk commodity transportation tasks and generates a task semantic label set T; An ability modeling module acquires real-time running state data, historical task performance data and behavior patterns of all transportation units in the transportation network, and generates a transportation resource feature vector set V; A multi-model scheduling prediction module, based on the task semantic label set T and the resource feature vector set V, constructs a multi-model integrated scheduling system including a transportation risk prediction sub-model, a node passing trend sub-model and a loading and unloading waiting prediction sub-model, for generating multiple candidate scheduling strategies; The reinforcement game optimization module inputs each candidate scheduling strategy into a reinforcement learning optimizer, combines path network reachability outliers, node predicted congestion degree, weather change trend and real-time feedback data, iteratively optimizes the strategy set through game optimization, and outputs a global optimal scheduling strategy S* under the current conditions; The dynamic reconfiguration module under the strategy S* is sent to the scheduling execution system, is distributed and deployed on multiple edge computing nodes, continuously monitors the transportation process, and adaptively reconfigures the current execution path and resource allocation according to the monitoring analysis result.

[0070] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. An intelligent bulk commodity freight logistics management method, characterized in that: include: S100: Obtain multi-dimensional information on bulk commodity transportation tasks and generate a task semantic tag set T; S200: Obtain real-time operational status data, historical task performance data, and behavioral patterns of all transportation units in the transportation network, and generate a set of transportation resource feature vectors V; S300, based on the task semantic label set T and the resource feature vector set V, constructs a multi-model integrated scheduling system including a transportation risk prediction sub-model, a node passage trend sub-model and a loading and unloading waiting prediction sub-model, which is used to generate multiple sets of candidate scheduling strategies. S400: Input each candidate scheduling strategy into the reinforcement learning optimizer, and combine path network accessibility outliers, node predicted congestion, weather change trends and real-time feedback data to perform iterative game optimization on the strategy set, and output the global optimal scheduling strategy S* under the current conditions. S500 distributes the strategy S* to the scheduling and execution system, which is deployed in a distributed manner across multiple edge computing nodes to continuously monitor the transportation process and adaptively reconfigure the current execution path and resource allocation based on the monitoring and analysis results.

2. The intelligent bulk commodity freight logistics management method according to claim 1, characterized in that: The construction of the multi-model integrated scheduling system includes: S301. Based on the historical transportation task sample set, for each type of goods and its transportation scenario, extract the high-frequency combination pattern in the task semantic label set T, and generate an embeddable task semantic vector T′ through a high-order semantic nesting encoding method. The encoding method uses a graph convolutional network with attention mechanism to extract task-related features. S302. For the resource feature vector set V, a temporal residual convolutional network is used to perform feature denoising and multi-scale feature extraction to obtain a dynamically updatable resource capability expression vector V′. S303. Using the task semantic vector T′ and the resource capability vector V′ as input, construct the transportation risk prediction sub-model, the node passage trend sub-model, and the loading and unloading waiting prediction sub-model respectively. The sub-models all adopt a multi-channel attention-gated cyclic unit structure, which can output the confidence index and the anomaly warning index of each prediction dimension respectively. S304. Based on the predicted outputs of the three sub-models, the scheduling preferences of different dimensions are fused and evolved through the policy clustering evolution mechanism to form multiple sets of candidate scheduling policies {S1, S2, …, Sn} with evaluable confidence factors. Each policy set includes scheduling path, execution window and interference tolerance configuration parameters, and n represents the total number of generated scheduling policies.

3. The intelligent bulk commodity freight logistics management method according to claim 1, characterized in that: The iterative game optimization of the strategy set mentioned above includes: S401. Using the set of candidate scheduling policies {S1, S2, …, Sn} as the input to the state space, construct a reinforcement learning optimizer based on a dual adversarial structure. S402. Collect outliers in the accessibility of the path network, predicted congestion of nodes, real-time weather heat maps and vehicle operation feedback data, construct a multi-source fusion environment tensor E, and simulate and evaluate the response trajectory of each strategy in the E environment. S403. In the simulation process, a Bayesian game dynamic incentive mechanism is introduced to score the utility and risk compensation of the strategy confrontation results in each round and update the strategy network weights. S404. Output the optimal scheduling policy S* based on the converged master policy network.

4. The intelligent bulk commodity freight logistics management method according to claim 3, characterized in that: The steps for calculating the path network accessibility outliers used as input to the reinforcement learning optimizer include: S411. Obtain the historical travel delay sequence of the nodes traversed by the target transportation route. and real-time latency data within the current period Construct the contrast matrix within the time window Δt. ,in A 5% truncated mean was used to remove outliers. S412. Based on the topological correlation of node traffic status, use spectral clustering method to classify the accessibility change trend of adjacent nodes on the path and identify the local fluctuation critical region corresponding to the cluster center point. S413. Construct accessibility perturbation functions for each node. Where ΔT is the delay offset. Let α and β be the historical delay variance, and α and β be the robustness coefficients adjusted according to the task type, to obtain the disturbance intensity value of each node; S414. If the disturbance intensity value exceeds the baseline accessibility tolerance threshold ε, the corresponding node is marked as an abnormal node, and its topological importance in the path structure is used as the weight to output the path network accessibility anomaly value δ.

5. The intelligent bulk commodity freight logistics management method according to claim 1, characterized in that: This includes continuous monitoring of the transportation process and adaptive reconfiguration, which includes: S501. Deploy edge sensing units at key nodes of the transportation route to collect vehicle running trajectory, cargo environmental status and micro-traffic data in real time, generate heterogeneous data stream set D, and perform multi-scale time-series segmentation processing on D through a hierarchical sliding window algorithm. S502. Use an integrated anomaly detection model to quickly identify segmented data. The model combines isolated forest and probabilistic graphical models, outputs a risk score vector R of multi-dimensional monitoring indicators, and determines whether there are path anomalies or resource imbalance events based on a time-varying threshold mechanism. S503. If a certain dimension of the risk scoring vector R exceeds its adaptive threshold range N times consecutively, the reconfiguration module is triggered, and the resource-path bidirectional mapping table is called to identify replaceable candidate transportation units and feasible alternative path sets. S504. Through a resource reconfiguration engine based on graph neural networks, the current task execution status is matched and sorted with candidate resources, and the execution path and transportation unit allocation are dynamically updated.

6. The intelligent bulk commodity freight logistics management method according to claim 5, characterized in that: The process of matching and sorting the current task execution status with candidate resources and dynamically updating the execution path and transportation unit allocation includes: S511. Encode the current transportation task's status information into a state vector. The remaining time limit of the task, the current path congestion coefficient, and the remaining energy consumption of the transportation unit are used as input features. S512. Generate resource evaluation vectors for all candidate transport units. The vector is obtained by fusing historical task completion rate, path adaptability score and fault prediction probability, and sorted by a multi-objective orthogonal projection analysis algorithm; S513, Transfer the state vector With the sorted resource evaluation vector Bidirectional conditional matching encoding is performed and input into the path reconstruction module based on graph attention mechanism to generate the optimal transport unit-path pair; S514: Call the dispatch control system to seamlessly reconfigure the current transportation plan.

7. An intelligent bulk commodity freight logistics management system, used to implement the intelligent bulk commodity freight logistics management method according to any one of claims 1-6, characterized in that: include: The task information processing module acquires multi-dimensional information about bulk commodity transportation tasks and generates a task semantic tag set T. The capability modeling module acquires real-time operational status data, historical task performance data, and behavioral patterns of all transportation units in the transportation network, and generates a set of transportation resource feature vectors V. The multi-model scheduling prediction module, based on the task semantic label set T and the resource feature vector set V, constructs a multi-model integrated scheduling system including a transportation risk prediction sub-model, a node passage trend sub-model and a loading and unloading waiting prediction sub-model, which is used to generate multiple sets of candidate scheduling strategies. The game optimization module is enhanced. Each candidate scheduling strategy is input into the reinforcement learning optimizer. The strategy set is iteratively optimized by combining path network accessibility outliers, node predicted congestion, weather change trends and real-time feedback data, and outputs the global optimal scheduling strategy S* under the current conditions. The dynamic reconfiguration module distributes the strategy S* to the scheduling and execution system, which is deployed in a distributed manner across multiple edge computing nodes. It continuously monitors the transportation process and adaptively reconfigures the current execution path and resource allocation based on the monitoring and analysis results.

Images