AI-based multimodal transport resource collaborative dynamic configuration method

Abstract

The application discloses a multimodal transport resource collaborative dynamic configuration method based on AI, which comprises the following steps: constructing a real-time data layer for collecting multidimensional data; preprocessing the data of the real-time data layer; dynamically constructing a digital twin platform based on the preprocessed data information; training multiple intelligent agents; and collaboratively training the multiple intelligent agents under a federal learning cluster framework; jointly training a global AI model; an AI decision hub generating an optimal or near-optimal dynamic resource configuration decision; a dynamic configuration engine dynamically scheduling resources and generating instructions; issuing the generated detailed instructions to the physical system of the execution layer; and an IoT device continuously monitoring the execution state and physical environment changes and feeding new data back to the real-time data layer. The application breaks through the data island and response delay bottleneck, constructs an intelligent decision system balancing the cost-time-carbon emission multi-objective, and realizes cross-domain collaborative optimization through the collaborative training of multiple intelligent agents under the federal learning framework.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cargo transportation, in particular to an AI-based multimodal transport resource collaborative dynamic configuration method. BACKGROUND

[0002] Current multimodal transport resource scheduling mainly relies on manual experience and static planning systems, lacks cross-modal real-time collaboration capabilities, and cannot achieve dynamic game optimization of resources under the premise of data security, resulting in increased transportation costs and decreased time efficiency and reliability. In addition, there are problems of insufficient data collaboration, such as isolated data among road, railway, and port links, lack of real-time fusion mechanism, resulting in delayed resource state visibility; traditional optimization models are based on historical rule libraries and are difficult to respond to sudden weather events (such as typhoons) and traffic congestion, etc. Dynamic disturbances, and scheduling adjustments will increase the time consumption. SUMMARY

[0003] Therefore, the present application provides an AI-based multimodal transport resource collaborative dynamic configuration method to solve the problems in the prior art.

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

[0005] The AI-based multimodal transport resource collaborative dynamic configuration method comprises the following steps:

[0006] Step 1: Build a real-time data layer for collecting multi-dimensional data;

[0007] Step 2: Preprocess the data of the real-time data layer through the edge computing nodes deployed on the road side and the port area;

[0008] Step 3: Based on the preprocessed data information, dynamically build a digital twin platform, and update the high-fidelity virtual mapping of the entire multimodal transport network; the digital twin platform reflects the state of the physical system in real time;

[0009] Step 4: Train agents representing different transport resources or links; and collaboratively train the above agents under a federated learning cluster framework; train a global AI model together;

[0010] Step 5: The AI decision hub receives real-time state information from the digital twin platform, combines the trained global AI model with preset rules and optimization objectives, performs global optimization calculation, and generates optimal or near-optimal dynamic resource configuration decisions;

[0011] Step 6: After receiving the configuration decision scheme from the AI decision hub, the dynamic configuration engine dynamically schedules resources and generates instructions;

[0012] Step 7: The detailed instructions generated by the dynamic configuration engine are issued to the physical systems of the execution layer; the IoT devices continuously monitor the execution status and changes in the physical environment, and feed back new data to the real-time data layer.

[0013] Further: the multi-dimensional data collection mainly includes transportation main body data collection, environment perception data collection, and demand data collection.

[0014] Further: the transportation main body data collection is mainly collected by the deployed IoT devices in real time to collect physical world data of different transportation links;

[0015] The environment perception data mainly includes meteorological data and data collected by sudden situation detection;

[0016] The demand data includes supply chain order data and historical demand prediction data; the supply chain order system extracts the delivery priority and dangerous goods transportation label; the historical demand prediction data is analyzed by a mathematical calculation model to generate seasonal fluctuations.

[0017] Further: the real-time processing of data preprocessing is as follows:

[0018] (1) Sensor data denoising, eliminate GPS multipath effect by Kalman filter;

[0019] (2) Video stream target detection, identify freight vehicle type, and synchronize semantic segmentation lane line;

[0020] (3) Time series alignment, railway dispatch clock and UTC time synchronization.

[0021] Further: when training agents representing different transportation resources or links, agent i learns the strategy independently through reinforcement learning , the goal is to maximize the long-term cumulative reward; defined as follows:

[0022] The state space is , representing the local state of agent i at time t;

[0023] The action space is , representing the decision of agent i;

[0024] The reward function is , which measures the immediate return of the action, and optimizes the overall goal of cost, time and carbon emissions;

[0025] The discount factor is , used to weigh future rewards;

[0026] The calculation formula of the reward function is as follows:

[0027] ;

[0028] where α1, α2, α3 are weight coefficients, satisfying ;

[0029] The agent i updates the local policy parameters by a policy gradient algorithm (such as PPO), whose formula is

[0030] ;

[0031] where is the learning rate;

[0032] is the policy gradient estimate, whose formula is

[0033] ;

[0034] where is the state value function; is the baseline value, used to reduce variance; γ is the discount factor.

[0035] Further: the server of federated learning updates the global model parameters by aggregating the local training results of each agent , whose specific steps are as follows:

[0036] (1) Local training: each agent i trains on the local data set to obtain the updated parameters ;

[0037] (2) Parameter upload: the agent uploads the local parameters or gradient to the server;

[0038] (3) Global aggregation: the server aggregates the parameters by weighted average:

[0039] ;

[0040] where wi is the weight of agent i, usually allocated according to data volume or contribution; N represents the number of agents participating in federated learning;

[0041] (4) Parameter distribution: the server distributes the global parameters to each agent as the initial parameters for the next round of training.

[0042] Further: the optimization target includes the lowest cost, the shortest time, and the least carbon emission.

[0043] Further, the dynamic configuration engine converts the collaborative decision scheme of the AI decision hub into specific and executable detailed operation instructions; the dynamic configuration engine is responsible for ensuring the feasibility of the instructions under the constraints of time, space and resources, and processing possible real-time conflicts.

[0044] Further, the physical system includes a railway resource pool, a highway fleet, and a port system, and the physical system executes specific operations according to the received instructions.

[0045] Further, after the new data feedbacks back to the real-time data layer, the feedback data is used to update the digital twin and as input of the subsequent AI decision hub, so as to realize continuous optimization of dynamic configuration.

[0046] The present application has the following advantages: breaking through the data island and response delay bottleneck, constructing an intelligent decision system balancing cost, time efficiency and carbon emission; realizing state mapping through the digital twin platform by pre-processing data through the edge node; realizing cross-domain collaborative optimization through multi-agent collaborative training under the federated learning framework; improving the generation speed of scheduling schemes in emergency situations through online updating of the model driven by IoT feedback data.

[0047] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more intuitively illustrate the prior art and the present application, exemplary drawings are given below. It should be understood that the specific shapes, structures shown in the drawings should not be regarded as limiting conditions in the implementation of the present application; for example, based on the technical concepts disclosed in the present application and the exemplary drawings, those skilled in the art can easily make routine adjustments or further optimizations to the increase / decrease / assignment of certain units (components), specific shapes, positional relationships, connection methods, size ratio relationships, etc.

[0049] Figure 1 An implementation flowchart of the AI-based multimodal transport resource collaborative dynamic configuration method provided for an embodiment of the present application is shown in the following figure. DETAILED DESCRIPTION

[0050] The following embodiments of the present application are explained by way of specific examples, and other advantages and effects of the present application will be readily appreciated by skilled in the art from the following description. It is obvious that the described embodiments are a part of the embodiments of the present application, but not all of the embodiments. It is understood that the embodiments are only for further explanation of the present application, and should not be construed as limiting the scope of the present application. The skilled in the art can make some non-essential improvements and adjustments to the present application according to the above-mentioned embodiments of the present application; based on the embodiments of the present application, all other embodiments obtained by the skilled in the art without creative labor are within the scope of the present application.

[0051] Please refer to Figure 1 , the AI-based multimodal transport resource collaborative dynamic configuration method includes the following steps:

[0052] Step 1: Build a real-time data layer for collecting multi-dimensional data.

[0053] Multi-dimensional data collection mainly includes transportation subject data collection, environmental perception data collection, and demand data collection.

[0054] Transportation subject data collection mainly collects physical world data (such as vehicle location, state, cargo information, environmental conditions, port berth occupancy, railway line state, etc.) of different transportation links such as railways, highways, and ports through deployed IoT devices (sensors, GPS, RFID, etc.).

[0055] Environmental perception data mainly includes meteorological data and data collected by sudden situation detection.

[0056] Among them, meteorological data is collected through the construction of a centimeter-level weather-road coupling field, which integrates numerical prediction models and roadside laser radars; sudden situation detection is based on unmanned aerial vehicle visible light / infrared dual-spectrum imaging to identify road obstacles.

[0057] Demand data includes supply chain order data and historical demand prediction data.

[0058] Among them, the supply chain order system extracts JIT distribution priority (regular expression matching "urgent" keywords) and dangerous goods transportation labels (UN number identification) through the NLP engine.

[0059] Historical demand prediction data is generated by analyzing seasonal fluctuations through an LSTM model, such as a surge in demand for transportation capacity during the harvest season for agricultural products.

[0060] Step 2: Preprocess data on the edge.

[0061] The preprocessing of real-time data layer data is mainly achieved by deploying MEC nodes on the roadside and in the port area. The roadside is equipped with an edge server every 50 kilometers, and the port area is equipped with a 5G base station integrated with an MEC module.

[0062] The real-time processing of data preprocessing is as follows:

[0063] (1) Sensor data denoising, eliminate GPS multipath effect through Kalman filtering;

[0064] (2) Video stream target detection, identify freight vehicle types, and synchronize semantic segmentation lane lines. In this embodiment, the YOLOv5 model is used to identify freight vehicle types;

[0065] (3) Time series alignment, synchronize railway dispatch clock with UTC time.

[0066] Step 3: Build and update the digital twin environment;

[0067] Based on the preprocessed data information, a digital twin platform is dynamically built, and a high-fidelity virtual mapping of the entire multimodal transport network (including railway resource pool, truck fleet, port system, etc.) is updated.

[0068] The digital twin platform reflects the state of the physical system in real time (such as resource location, availability, congestion, plan execution progress, etc.), providing accurate environmental models for subsequent simulation and decision-making.

[0069] In this embodiment, the hub three-dimensional structure can be reconstructed through oblique photography and BIM technology, and the LiDAR scanning data is fused to achieve centimeter-level spatial accuracy. Based on vibration spectrum analysis, the diesel engine piston wear is mapped in real time (error <0.02mm), and the bearing health is monitored using voiceprint recognition (accuracy 99.1%). Business process simulation is achieved, Petri net modeling is used to simulate the timing of transshipment operations, and AnyLogic is used to simulate the throughput under different scheduling strategies (error <3%).

[0070] Step 4: AI model training and collaborative learning;

[0071] Using the multi-agent reinforcement learning (MARL) algorithm, agents representing different transportation resources or links (such as trains, trucks, port cranes, dispatchers) are trained.

[0072] Each agent i learns the strategy independently through reinforcement learning , the goal is to maximize the long-term cumulative reward. Defined as follows:

[0073] State space (State): , represents the local state of agent i at time t, such as resource location, task queue.

[0074] Action (Action): , representing the decision of agent i, such as scheduling instructions, path planning.

[0075] Reward (Reward): , measuring the immediate income of action, optimizing the target of cost, time and carbon emission, the calculation formula of reward function is as follows:

[0076] ;

[0077] Where, α1, α2, α3 are weight coefficients, satisfying α1+α2+α3=1.

[0078] Discount Factor (Discount Factor): , used to weigh future rewards.

[0079] Agent i updates local policy parameters by policy gradient algorithm (such as PPO), whose formula is:

[0080] ;

[0081] Where, is the learning rate.

[0082] is the policy gradient estimate value, and the calculation formula is:

[0083] ;

[0084] Where, is the state value function; is the baseline value, used to reduce variance; γ is the discount factor.

[0085] These agents are collaboratively trained under the framework of federated learning cluster; federated learning allows each participant (such as different transport companies, different ports) to share model parameters or gradients while protecting local data security, and collaboratively train a more powerful global AI model to improve resource coordination capability.

[0086] The federated learning server updates the global model parameters θ global by aggregating the local training results of each agent, and the specific steps are as follows:

[0087] (1) Local training: each agent i trains on local data set Di to get updated parameters ;

[0088] (2) Parameter upload: the agent uploads the local parameters or gradient Uploaded to the server.

[0089] (3) Global aggregation: The server aggregates parameters by weighted average:

[0090] ;

[0091] Where wi is the weight of agent i, which is usually assigned based on the amount of data or contribution; N represents the number of agents participating in federated learning.

[0092] For example: Suppose a multimodal transport network includes the following participants: railway company A (agent 1), road logistics company B (agent 2), and port C (agent 3).

[0093] Then N=3, and the global model parameters are:

[0094] ;

[0095] w1, w2, and w3 are allocated based on the amount of data or negotiated weights for each party.

[0096] (4) Parameter distribution: The server distributes global parameters. The parameters are distributed to each agent as initial parameters for the next round of training.

[0097] In addition, to capture spatiotemporal correlations, a graph neural network (GNN) and a spatiotemporal attention mechanism are introduced, and the weights are dynamically adjusted.

[0098] The graph neural network construction includes determining node attributes and designing edge weights. When determining node attributes, it is necessary to determine the available time window for port berth tidal flow (predicted by harmonic analysis model) and the probability matrix of track occupancy at marshalling yards (modeled by Markov chain). Edge weights are designed by taking into account transportation costs, carbon emission coefficients, and timeliness reliability.

[0099] Spatiotemporal attention mechanism; using Transformer model to capture long-distance dependencies, such as the impact of typhoon path prediction on the joint scheduling of maritime and land transportation; and dynamically adjusting weights; for example, automatically reducing the weight of coastal highways and increasing the priority of inland waterway shipping when a typhoon warning is issued.

[0100] Step 5: The AI ​​decision center receives real-time status information from the digital twin platform, combines it with the trained global AI model and possible preset rules and optimization objectives; performs global optimization calculations; and generates optimal or near-optimal dynamic resource allocation decisions.

[0101] Among them, the optimization target includes the lowest cost, the shortest time, the least carbon emission, etc.; the dynamic resource configuration decision includes the collaborative scheduling scheme across transportation modes, such as when a batch of goods is transferred from the railway to the highway, which truck is connected at which port, which berth is idle for loading and unloading, etc.

[0102] Step 6: The dynamic configuration engine receives the configuration decision scheme from the AI decision hub, dynamically schedules resources, and generates instructions.

[0103] The dynamic configuration engine converts the collaborative decision scheme of the AI decision hub into specific and executable motion instructions (to vehicles and equipment), path instructions (to vehicles and ships), berth plans (to port terminals), and other detailed operation instructions.

[0104] The dynamic configuration engine is responsible for ensuring the feasibility of the instructions under the constraints of time, space, and resources, and handling possible real-time conflicts.

[0105] Step 7: The generated detailed instructions are issued to the physical systems of the execution layer; IoT devices continuously monitor the execution status and changes in the physical environment, and feed new data back to the real-time data layer.

[0106] The physical system includes the railway resource pool (train, locomotive, station), highway fleet (truck, driver), and port system (berth, quay crane, yard equipment). The physical system performs specific transportation, loading and unloading, and transfer operations according to the received instructions.

[0107] After the new data is fed back to the real-time data layer, these feedback data are used to update the digital twin (step 3) and as input for subsequent AI decisions (step 5), achieving continuous optimization of dynamic configuration.

[0108] Embodiment

[0109] Assumptions: Train A is located at Nanjing Hub Station, the goods are containers X / Y / Z, and the expected departure time is 14:00; Truck B is located on G42 Expressway Changzhou Section, is empty, and is expected to arrive at Nanjing Hub Station at 13:40; Berth 3 is currently occupied by ship C, which is expected to depart at 15:30; Container X is in the state of not arriving at the port.

[0110] All data is aggregated to the real-time data layer to form a unified view; the digital twin platform dynamically generates a virtual scene.

[0111] Nanjing Hub Station: Train A is delayed by 10 minutes (due to track maintenance ahead);

[0112] Highway network: Truck B is rerouted due to congestion and arrives at 13:55;

[0113] Shanghai Port: Ship C on Berth 3 is delayed in departure, and the idle time is postponed to 15:45.

[0114] After the digital twin platform simulation, resource conflicts are found:

[0115] Container X original plan: 14:30 transfer by truck B → 15:30 arrive at port for loading;

[0116] Now the prediction: truck B arrives at Nanjing hub at 13:55; container X needs 15 minutes to load; highway transportation needs 100 minutes; arrival time at port is 17:00 > ship D customs clearance time is 16:30.

[0117] Federated learning collaborative training: railway / highway / port three parties train agents locally, only share model parameters (do not expose raw data).

[0118] Among them, the railway agent learns to adjust the train schedule to compensate for delays; the highway agent learns dynamic path planning to avoid congestion; and the port agent learns flexible berth allocation strategy.

[0119] Multi-agent collaborative decision making; the global goal is to minimize the total transportation time of container X.

[0120] Multi-agent collaboration scheme is if the railway delay > set threshold; then the highway agent calls the backup truck C (closer but higher cost); the port agent negotiates with ship D to delay customs clearance for 30 minutes.

[0121] Input the conflict state of the digital twin early warning, combined with the trained MARL model output strategy;

[0122] The decision-making scheme of the AI decision-making center is as follows:

[0123] Railway, train A prioritizes unloading container X (compresses stop time to 5 minutes); highway, dispatches backup truck C (location: Wuxi service area, arrival time 13:45); port, applies for delayed customs clearance for ship D to 17:00, assigns berth 4 (originally planned for maintenance, urgently activated).

[0124] Optimization basis: the total cost increases by 800 yuan (backup truck fee + port overtime fee), but avoids a breach of contract penalty of 5000 yuan.

[0125] The dynamic configuration engine converts the decision of the AI decision-making center (high level) into atomic instructions, as shown in the following table.

[0126]

[0127] The physical system of the execution layer starts execution:

[0128] Truck C arrives at Nanjing hub at 13:45, container X completes loading at 13:50; real-time path correction of road transport (detour due to accident), arrival time at port is 16:45; port clears berth 4 in advance, ship D completes loading and departs at 16:50.

[0129] Data closed-loop feedback: IoT device reports new status, container X has been loaded on the ship, actual departure time is 16:50; data is returned to the real-time data layer, which can update the digital twin and be used for subsequent model training and optimization.

[0130] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An AI-based dynamic configuration method for multimodal transport resource collaboration, characterized in that, The method comprises the following steps: Step 1: constructing a real-time data layer for collecting multi-dimensional data; Step 2: pre-processing the data of the real-time data layer through edge computing nodes deployed on the roadside and in the port area; the real-time processing of the data pre-processing is as follows: (1) sensor data denoising, eliminating GPS multipath effect through Kalman filtering; (2) video stream target detection, identifying freight vehicle types, and synchronously segmenting lane lines; (3) time series alignment, synchronizing the railway dispatching clock with the UTC time; Step 3: dynamically constructing a digital twin platform based on the pre-processed data information, and updating the high-fidelity virtual mapping of the entire multimodal transport network; the digital twin platform reflects the state of the physical system in real time; Step 4: training agents representing different transport resources or links; and collaboratively training the above agents under a federated learning cluster framework; training a global AI model together; When training agents representing different transport resources or links, agent i learns the strategy pi independently through reinforcement learning, with the goal of maximizing the long-term cumulative reward; the definition is as follows: The state space is si t∈Si, which represents the local state of agent i at time t; The action space is ai t∈Ai, which represents the decision of agent i; The reward function is rit, which measures the immediate return of the action, and comprehensively optimizes the objectives of cost, time, and carbon emissions; The discount factor is γ∈[0,1), which is used to weigh future rewards; The calculation formula of the reward function is as follows: ； Where α1, α2, α3 are weight coefficients, satisfying α1+α2+α3=1; The agent i updates the local policy parameters θ by a policy gradient algorithm i The formula is: ； Where ηi is the learning rate; For the policy gradient estimate, the formula is: ； where V θi is the state value function; bit is a baseline value to reduce variance; γ is a discount factor; Step 5: the AI decision hub receives real-time state information from the digital twin platform, combines the trained global AI model with preset rules and optimization objectives, performs global optimization calculation, and generates optimal or near-optimal dynamic resource allocation decisions; Step 6: after receiving the allocation decision scheme from the AI decision hub, the dynamic configuration engine dynamically schedules resources and generates instructions; Step 7: the detailed instructions generated by the dynamic configuration engine are issued to the physical system of the execution layer; IoT devices continuously monitor the execution status and changes in the physical environment, and feed new data back to the real-time data layer. 2.The AI-based method for dynamically configuring multi-modal transport resources in coordination, according to claim 1, wherein, The multi-dimensional data collection mainly includes transport subject data collection, environmental perception data collection, and demand data collection. 3.The AI-based method for dynamically configuring multimodal transport resources in coordination, according to claim 2, wherein, The transport subject data collection mainly collects physical world data of different transport links in real time through deployed IoT devices; Environmental perception data mainly includes meteorological data and data collected through sudden situation detection; Demand data includes supply chain order data and historical demand prediction data; the supply chain order system extracts distribution priorities and dangerous goods transportation labels; historical demand prediction data is analyzed through mathematical calculation models to analyze seasonal fluctuations. 4.The AI-based method for dynamically configuring multi-modal transport resources in coordination, according to claim 1, wherein, The server of the federated learning updates the global model parameters θ by aggregating the local training results of each agent global The specific steps are as follows: (1) Local training: Each agent i trains on its local dataset Di, obtaining updated parameters Δθ i； (2) Parameter upload: the agent uploads local parameters θ i or gradients ∇ θi J i (θ i ) to the server; (3) Global aggregation: the server aggregates parameters through weighted averaging: ； Where wi is the weight of agent i, usually allocated according to data volume or contribution; N represents the number of agents participating in federated learning; (4) Parameter issuing: the server issues the global parameter θ global to each agent as the initial parameter for the next round of training. 5.The AI-based method for dynamically configuring multi-modal transport resources in coordination, according to claim 1, wherein, The optimization objectives include the lowest cost, the shortest time, and the least carbon emissions. 6.The AI-based method for dynamically configuring multi-modal transport resources in coordination, according to claim 1, wherein, The dynamic configuration engine translates the collaborative decision-making scheme of the AI decision hub into specific and executable detailed operation instructions; the dynamic configuration engine is responsible for ensuring the feasibility of the instructions under the constraints of time, space and resources, and handling possible real-time conflicts.

7. The AI-based multimodal transport resource collaborative dynamic configuration method according to claim 1, wherein, The physical system includes a railway resource pool, a highway fleet, and a port system, which executes specific operations according to the received instructions. 8.The AI-based method for dynamically configuring multi-modal transport resources in coordination, according to claim 1, wherein, After the new data feedbacks back to the real-time data layer, the feedback data is used to update the digital twin and as input for the subsequent AI decision hub, realizing the continuous optimization of dynamic configuration.

Images