Underground logistics network planning method and system based on artificial intelligence

Abstract

The application relates to the technical field of logistics network planning, and discloses an underground logistics network planning method and system based on artificial intelligence. The method comprises the following steps: each node predicts a local future state through a space-time graph convolutional neural network; an offline trained multi-agent reinforcement learning model is used to generate an optimal local action sequence; and space-time conflicts are eliminated through a communication-free deterministic protocol. The system comprises a distributed state sensing and prediction module, a multi-agent reinforcement learning decision module, a space-time conflict elimination module and a hierarchical task allocation module. The application realizes decentralized and high-robustness autonomous collaborative planning, and significantly improves the operation efficiency and safety of the underground logistics network in a weak communication environment.

Description

Technical Field

[0001] This invention belongs to the field of logistics network planning technology, specifically relating to an artificial intelligence-based method and system for planning underground logistics networks. Background Technology

[0002] With the accelerated development of urban underground space, underground logistics networks, as a crucial infrastructure for alleviating surface traffic pressure and improving the efficiency of goods distribution, are gradually becoming a key component of smart city construction. This network typically consists of multiple distributed nodes, encompassing warehousing and transfer stations, automated guided vehicles (AGVs), and environmental sensing equipment. Its efficient operation relies on precise task planning and real-time collaborative control. In complex underground environments, nodes need to frequently exchange status information to achieve dynamic task allocation and path optimization, which places extremely high demands on communication reliability and timeliness.

[0003] Artificial intelligence-based underground logistics network planning methods focus on using machine learning models to intelligently schedule and coordinate resources for multi-node tasks, aiming to improve overall system throughput and response speed through a data-driven approach. These methods typically rely on a centralized decision-making architecture, processing status data reported by each node centrally in the cloud and generating globally optimal or near-optimal task allocation schemes for execution. However, this model heavily depends on stable and low-latency communication links, making it difficult to adapt to scenarios in underground channels where communication quality is poor or intermittently interrupted due to geological structures, electromagnetic interference, and other factors.

[0004] In existing technologies, centralized planning architectures exhibit significant drawbacks when facing unstable communication in underground environments. On the one hand, the cloud-based decision-making cycle is constrained by the round-trip delay between uplink data acquisition and downlink command transmission, easily leading to delayed task allocation. On the other hand, when some nodes cannot receive commands in a timely manner due to communication interruptions, they often become idle and wait, wasting energy and potentially causing local congestion or even task conflicts. Furthermore, traditional methods lack the ability for autonomous negotiation and local optimization at the node level, making it impossible to maintain basic collaborative order under communication constraints. Therefore, reducing reliance on central nodes and stable communication while ensuring efficient task allocation has become a pressing technical challenge for the intelligent upgrading of underground logistics networks. Summary of the Invention

[0005] The technical problem to be solved by this invention is that existing underground logistics network planning systems rely heavily on centralized architecture for multi-node collaborative planning. When the quality of underground channel communication is poor and there are intermittent communication situations, the real-time and consistency of decision information cannot be guaranteed, which leads to task allocation conflicts, path planning deadlocks and idle node resources, significantly reducing the operating efficiency and robustness of the entire logistics network.

[0006] To address the technical problems existing in the background art, the present invention provides an artificial intelligence-based underground logistics network planning method, comprising: at each planning node of the underground logistics network, through its distributed state perception and prediction module, collecting and integrating its own node state data in real time, wherein the node state data constitutes a multi-dimensional state vector.

[0007] The distributed state perception and prediction module uses its built-in spatiotemporal graph convolutional neural network model to take the collected state data of its own node and the historical state data of neighboring nodes obtained through intermittent communication as input to predict the state sequence of the planned node and its neighboring nodes within the communication range in the future within a preset time window, and generate future state time series data containing multiple time steps.

[0008] At each planning node, the multi-agent reinforcement learning decision module it carries takes the future state time series data generated by the distributed state perception and prediction module as the decision input. The multi-agent reinforcement learning decision module has a built-in pre-trained deep policy network model. The deep policy network model outputs the optimal local action sequence that maximizes the predefined global reward function based on the input future state time series data.

[0009] At each planning node, the optimal local action sequence output by the multi-agent reinforcement learning decision-making module is converted into the spatiotemporal trajectory of the node within a future preset time window through the spatiotemporal conflict resolution and negotiation module on board, and compared with the predicted spatiotemporal trajectory generated based on the predicted state of adjacent nodes to detect potential spatiotemporal occupancy conflicts.

[0010] When a spatiotemporal conflict is detected, the spatiotemporal conflict resolution and negotiation module activates a deterministic conflict resolution protocol. The protocol calculates dynamic priorities based on the task priority, remaining running time, and battery power of each conflicting node. After each conflicting node synchronously completes the priority value calculation, the node with the smallest value re-plans the action sequence to avoid the spatiotemporal conflict, thereby achieving autonomous collaboration without communication.

[0011] Preferably, the distributed state perception and prediction module utilizes its built-in spatiotemporal graph convolutional neural network model, taking the collected state data of its own node and the historical state data of neighboring nodes obtained through intermittent communication as input, to predict the state sequence of the planned node and its neighboring nodes within its communication range within a preset future time window, generating future state time series data containing multiple time steps, including:

[0012] An adjacency matrix is ​​constructed based on the physical topology of the underground logistics network, where the connection weights between nodes are determined by normalizing the inverse of the Euclidean distance.

[0013] After organizing the multidimensional state vector and the historical state data of neighboring nodes into a node state sequence according to the time dimension, the sequence is input into the graph convolutional layer to extract spatial dependency features, thus obtaining the spatially enhanced feature vector of each node.

[0014] The spatially augmented feature vector sequence is input into the gated recurrent unit layer to model the temporal evolution law and obtain the temporal hidden state of each node;

[0015] The temporal hidden state is input to the fully connected output layer to be mapped to the predicted value of the future state, thereby generating the temporal data of the future state.

[0016] Preferably, the multi-agent reinforcement learning decision-making module incorporates a pre-trained deep policy network model. Based on the input future state time-series data, the deep policy network model outputs the optimal local action sequence that maximizes the predefined global reward function, including:

[0017] Future state time series data is used as input to a multilayer perceptron with 5 hidden layers, each containing 1024 neurons, and the activation function is a modified linear unit.

[0018] The selection probability distribution of five predefined action primitives is output through a flexible maximum value function. The action primitives include: moving to a specified topology node, performing a cargo picking operation, performing a cargo placement operation, entering a standby state, and autonomously navigating to a preset charging facility.

[0019] Based on the sampling of the selection probability distribution, the specific action primitives to be executed within the current decision cycle are determined, and a local action sequence that is time-aligned with the future state time series data is generated.

[0020] Preferably, the predefined global reward function is defined as a linear weighted sum, and its specific form is: the global reward function equals the task completion reward weight multiplied by the task completion status value, minus the time consumption penalty weight multiplied by the task execution time, minus the energy consumption penalty weight multiplied by the node energy consumption, and finally minus the collision and conflict penalty weight multiplied by the collision and conflict event indication value; where all weight coefficients are non-negative real numbers, and their sum is 1.

[0021] Preferably, the spatiotemporal conflict resolution and negotiation module converts the optimal local action sequence into the spatiotemporal trajectory of the node within a preset future time window, and compares it with the predicted spatiotemporal trajectory generated based on the predicted states of neighboring nodes to detect potential spatiotemporal occupancy conflicts, including:

[0022] Based on the semantics of the action primitives, the local action sequence is converted into a four-dimensional spatiotemporal trajectory consisting of a series of discrete spatiotemporal points. Each spatiotemporal point is represented as a quadruple containing three-dimensional spatial coordinates and a timestamp.

[0023] The predicted spatiotemporal trajectory of each neighboring node is generated based on the predicted state of the neighboring nodes.

[0024] The local spatiotemporal trajectory is cross-compared with the predicted spatiotemporal trajectories of all adjacent nodes at each time step. If the Euclidean distance between the spatial coordinates of the local node and any adjacent node is less than the preset safety threshold at any time step, a spatiotemporal occupancy conflict is determined to exist.

[0025] Preferably, the safety threshold is 1.5 times the physical size of the planned node.

[0026] Preferred, deterministic conflict resolution protocols calculate dynamic priorities based on the task priority, remaining runtime, and battery level of each conflicting node, including:

[0027] Calculate a priority score for each conflict node. The priority score is equal to the first weight coefficient multiplied by the task priority code, minus the second weight coefficient multiplied by the estimated task completion time, minus the third weight coefficient multiplied by the normalized battery power value.

[0028] Among them, the task priority code is pre-specified by the global task publishing system, the estimated task completion time is dynamically estimated by the node based on the distance to the task target point and the current speed, and all weight coefficients are positive real numbers pre-set by the system;

[0029] Compare the priority scores of all conflicting nodes. The node with the lowest priority score treats the predicted spatiotemporal trajectories of other conflicting nodes as dynamic obstacles and calls its local path planner to regenerate a conflict-free sequence of actions.

[0030] Preferably, the method further includes processing global logistics tasks through a hierarchical task decomposition and allocation module, including:

[0031] The macro-logistics objectives are decomposed into a series of atomic tasks with clear start points, end points, and load requirements using a hierarchical task network planning algorithm.

[0032] Atomic task information is disseminated throughout the underground logistics network via a neighbor-based broadcast protocol;

[0033] Each planning node autonomously calculates the expected cost and efficiency of executing each atomic task based on its current status, location, and prediction of the network status, and generates a bid value.

[0034] Through a distributed contract network protocol, the node with the best bid obtains the right to execute atomic tasks and announces the task acceptance confirmation information to the entire network through a broadcast protocol.

[0035] This invention also provides an artificial intelligence-based underground logistics network planning system for implementing the aforementioned artificial intelligence-based underground logistics network planning method, comprising:

[0036] The distributed state perception and prediction module is used to collect and integrate the node state data of each planned node in the underground logistics network in real time. The node state data constitutes a multi-dimensional state vector, which includes the node's real-time three-dimensional spatial coordinates, three-axis velocity components, three-axis acceleration components, remaining battery power, current load weight, and equipment operating status code. The module is also used to use its built-in spatiotemporal graph convolutional neural network model to predict the state sequence of the planned node and its neighboring nodes within the communication range within a future preset time window, using the collected node state data and the historical state data of neighboring nodes obtained through intermittent communication as input, and generate future state time series data containing multiple time steps.

[0037] A multi-agent reinforcement learning decision module is used to take the future state time series data as decision input. The module has a built-in pre-trained deep policy network model. The deep policy network model outputs the optimal local action sequence that maximizes the predefined global reward function based on the input future state time series data.

[0038] The spatiotemporal conflict resolution and negotiation module is used to convert the optimal local action sequence into the spatiotemporal trajectory of the node within a future preset time window, and compare it with the predicted spatiotemporal trajectory generated based on the predicted state of adjacent nodes to detect potential spatiotemporal occupancy conflicts. When a spatiotemporal occupancy conflict is detected, the module activates a deterministic conflict resolution protocol. The protocol calculates dynamic priorities based on the task priority, remaining running time, and battery power of each conflicting node. After each conflicting node synchronously completes the priority value calculation, the node with the smallest value replans the action sequence to avoid the spatiotemporal occupancy conflict.

[0039] The hierarchical task decomposition and allocation module is deployed on several aggregation nodes or global management servers in the network. It is used to decompose high-level logistics instructions into executable atomic tasks and publish them to the entire network through asynchronous broadcast. At the same time, this module is also used to monitor task bidding and acceptance information in the network to maintain a macro view of the global task status.

[0040] Preferably, the embedded computing unit in the distributed state perception and prediction module has a hardware architecture including a graphics processing unit with at least 1024 processing cores and no less than 32 gigabytes of unified addressing memory, to ensure that the spatiotemporal graph convolutional neural network model can complete a complete prediction inference at a frequency of no less than 10 times per second.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] 1. By deploying a complete perception, prediction, and decision-making system at each node, this invention constructs a completely decentralized planning architecture, fundamentally eliminating the dependence on central servers and real-time communication. This enables the underground logistics network to maintain efficient and stable operation even in harsh environments such as communication delays, bandwidth limitations, or even complete outages, demonstrating extremely high system robustness.

[0043] 2. This invention introduces a state prediction mechanism based on spatiotemporal graph neural networks, enabling each node to foresee the future evolution trend of its local network environment, thereby changing the decision-making basis from "current state" to "future state", realizing a paradigm shift from passive response to active prediction, and significantly reducing the probability of decision-making errors due to outdated information.

[0044] 3. The multi-agent reinforcement learning decision-making model adopted enables nodes to autonomously learn complex and non-intuitive collaborative strategies through massive training in a simulation environment. Its efficiency far exceeds that of traditional methods based on manual rules, and it can adaptively cope with various complex logistics scenarios and network congestion.

[0045] 4. The deterministic, communication-free spatiotemporal conflict resolution protocol designed in this invention provides the ultimate security guarantee for autonomous collaboration between nodes, ensuring that the network will not deadlock or collide even under the worst communication conditions, thus guaranteeing the absolute safety of logistics operations.

[0046] 5. The hierarchical and distributed task allocation mechanism effectively decouples the macro-planning of global tasks from the micro-decision-making of node execution, taking into account both global logistics efficiency and the autonomy of individual nodes, and realizing the optimal allocation of network resources. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall technical solution architecture of the artificial intelligence-based underground logistics network planning method and system proposed in this invention;

[0048] Figure 2 This is a schematic diagram of the core principle framework of decentralized decision-making based on spatiotemporal graph neural network and multi-agent reinforcement learning in this invention;

[0049] Figure 3 This is a logical flowchart of the distributed state perception and prediction module in this invention.

[0050] Figure 4 This is a logical flowchart of the multi-agent reinforcement learning decision-making module in this invention;

[0051] Figure 5 This is a logical flowchart of the spatiotemporal conflict resolution and negotiation module in this invention. Detailed Implementation

[0052] Please refer to Figures 1 to 5 In the operating environment of underground logistics networks, communication channels are highly unstable due to factors such as geological structure, electromagnetic shielding, and dense node deployment. This leads to frequent problems such as task allocation conflicts, path deadlocks, and node idleness in traditional planning methods that rely on centralized cloud decision-making. To solve this technical problem, this invention provides an artificial intelligence-based underground logistics network planning method and system. Its core lies in constructing a fully decentralized local intelligent decision-making architecture, enabling each logistics node to autonomously complete the entire process of state perception, future prediction, action decision-making, and conflict resolution without the support of real-time central communication.

[0053] The method is executed independently at each planned node of the underground logistics network, and its process strictly follows the following sequence: S1, S2, S3, S4, S5. There are strict temporal dependencies between each step, and each step contains multiple mandatory sub-operations to ensure the integrity and feasibility of the technical solution.

[0054] Step S1 involves distributed state awareness and data integration. In this step, each planning node acquires its real-time 3D spatial coordinates through its onboard 3D positioning sensor; obtains three-axis velocity and acceleration components through its inertial measurement unit; reads the remaining battery power through its battery management unit; measures the current load weight through its load sensor; and obtains the device operating status code through its device operating status monitoring interface. These five types of data together constitute a multi-dimensional state vector, with a fixed nine-dimensional dimension, arranged as follows: x-coordinate, y-coordinate, z-coordinate, vx velocity, vy velocity, vz velocity, ax acceleration, ay acceleration, az acceleration, remaining battery power, current load weight, and device operating status code. This multi-dimensional state vector is fully collected and cached in local memory at the beginning of each decision cycle. Simultaneously, the node receives historical state data packets from all neighboring nodes within its communication range through an intermittent wireless communication interface. Each historical state data packet contains a complete multi-dimensional state vector sequence of neighboring nodes over several past time steps, with a time step interval of 100 milliseconds and a maximum backtracking length of 30 seconds. After the received data packets undergo timestamp and integrity verification, they are stored in the local neighborhood state buffer for use as input to the subsequent prediction model.

[0055] Step S2 involves future state prediction based on a spatiotemporal graph convolutional neural network. In this step, the distributed state perception and prediction module calls its built-in spatiotemporal graph convolutional neural network model, using the local multidimensional state vector generated in step S1 and the historical state data of neighboring nodes in the neighborhood state buffer as joint input. The structure of this spatiotemporal graph convolutional neural network model, from input to output, includes a graph convolutional layer, a gated recurrent unit layer, and a fully connected output layer. The graph convolutional layer first constructs an adjacency matrix based on the physical topology of the underground logistics network, where the connection weights between nodes are determined by normalizing the inverse of the Euclidean distance. This layer aggregates spatial features from the input multidimensional state vector, outputting a spatially enhanced feature vector for each node.

[0056] Subsequently, the gated recurrent unit layer performs temporal modeling on the spatial augmented feature vector sequence of each node, using time steps as the unit, to capture long-term dependencies in state evolution. Finally, the fully connected output layer maps the hidden states of the gated recurrent unit layer to predicted values ​​of future states. The output of this model is future state temporal data, with a time span of a preset time window of 150 seconds and a time resolution of 500 milliseconds, containing a total of 300 time steps. Each time step corresponds to a complete nine-dimensional state vector prediction value, covering coordinates, velocity, acceleration, battery level, load, and status code. The prediction results are encapsulated into two independent data structures: a local future state sequence and a neighborhood future state sequence, and then passed to the next module.

[0057] Step S3 is the local action decision driven by multi-agent reinforcement learning. In this step, the multi-agent reinforcement learning decision module receives the future state time-series data output from step S2 and uses it as the sole input to the deep policy network model. This deep policy network model is a multilayer perceptron with 5 hidden layers, each containing 1024 neurons, and the activation function uses modified linear units. The input layer dimension is 2700 (i.e., 300 time steps multiplied by nine-dimensional states), and the output layer dimension is 5, corresponding to 5 predefined action primitives: move to a specified topology node, perform a cargo pickup operation, perform a cargo placement operation, enter a standby state, and autonomously navigate to a preset charging facility. The network outputs the selection probability distribution of each action primitive through a flexible maximum function. The policy network is trained offline using a centralized training and decentralized execution framework: in a high-fidelity simulation environment, the observation, action, and reward data of all nodes are collected centrally to train the global evaluation network; this evaluation network guides the parameter updates of the policy network of each node through a policy gradient algorithm. After training, only the policy network is embedded into the embedded computing unit of each node. In actual operation, the policy network calculates the optimal local action sequence that maximizes the predefined global reward function based on the input future state time series data. This global reward function is defined as:

[0058] ;

[0059] in, This represents the task completion status value (1 for completion, 0 for incomplete). The task execution time (in seconds). This represents the node's energy consumption (in watt-hours). This is the collision event indicator value (1 if a collision occurs, 0 otherwise); , , , Non-negative real number weight coefficients, and satisfying The system default configuration is as follows: The probability distribution output by the policy network is sampled to determine the specific action primitives to be executed within the current decision period, and a local action sequence with a length of 300 time steps is generated. This sequence is strictly aligned in time with the future state time series data.

[0060] Step S4 is for spatiotemporal trajectory generation and conflict detection. In this step, the spatiotemporal conflict resolution and negotiation module receives the local action sequence output from step S3 and converts it into a precise four-dimensional spatiotemporal occupancy trajectory. The conversion process is based on the semantics of the action primitives: if the action is "move to a specified topology node", the local path planner is invoked to generate a smooth trajectory from the current position to the target node, and the timestamps of reaching each intermediate point are estimated based on the current speed and acceleration; if the action is "perform a cargo pick-up / placement operation", the trajectory remains stationary at the operation point for a preset operation duration (5 seconds for pick-up, 4 seconds for placement); if the action is "standby", the trajectory remains unchanged at the current position for the entire time window; if the action is "navigate to charging facility", the trajectory points to the nearest available charging pile and includes charging waiting and connection time. The final generated spatiotemporal trajectory consists of a series of discrete spatiotemporal points, each represented as a quadruple (x, y, z, t), where t is the absolute timestamp. Simultaneously, based on the future state sequence of the neighborhood in step S2, the module generates the predicted spatiotemporal trajectory for each neighboring node. Subsequently, the module performs a time-step cross-comparison between the local trajectory and all predicted trajectories from the neighborhood. The conflict determination criterion is: within any time step, the Euclidean distance between the spatial coordinates of the local node and any neighboring node is less than a preset safety threshold. This safety threshold is set to 1.5 times the physical size of the planned node, with a typical value of 2.25 meters (assuming a node diameter of 1.5 meters). If a distance less than this threshold is detected at any time step, a spatiotemporal occupancy conflict is determined, and the time window of the conflict and the node identifiers involved are recorded.

[0061] Step S5 is a deterministic conflict resolution method based on dynamic priority. When step S4 detects a spatiotemporal occupancy conflict, it triggers a simplified, low-power short-frame broadcast, transmitting only the "priority score" data (≤8 bytes), without any additional state data, a significant difference from conventional high-bandwidth continuous communication; moreover, the broadcast only starts when a conflict is triggered, without continuously occupying communication resources, adapting to low-bandwidth requirements in weak communication environments. The formula for calculating the priority score is:

[0062] ;

[0063] in, Encode the task priority (pre-specified by the global task publishing system, with a value range of 1 to 10, the larger the value, the higher the priority). The estimated task completion time (in seconds, dynamically estimated by the node based on the distance to the task target point and the current average speed). This is the normalized value of battery power. Compatibility with temporary no-communication scenarios: If there is a short period of complete no-communication (unable to broadcast), the priority scores of adjacent nodes derived from historical data are used directly, combined with the user's own priority score, to determine the nodes to be actively avoided, without real-time broadcast interaction. Backup solution for temporary no-communication scenarios: When there is no intermittent communication, no broadcasting, and no historical data, the "task urgency priority" fallback rule is implemented: If the user's own task priority code is ≥8 (urgent task), the original action sequence is maintained, and the "exclusive trajectory" strategy is executed; if it is a non-urgent task (priority code <8), it immediately enters a standby state until the conflict time window ends (maximum 300 seconds) or communication is restored, ensuring no collision risk.

[0064] Furthermore, the method includes a global task decomposition and allocation mechanism. While this mechanism does not directly participate in the periodic decision-making loops of each node, it provides task input. A hierarchical task decomposition and allocation module is deployed at several aggregation nodes in the network, responsible for receiving macro-level logistics instructions (such as "transport goods A from warehouse X to site Y"). This module first decomposes the macro-level instructions into a series of atomic tasks using a hierarchical task network planning algorithm. Each atomic task contains clearly defined start and end coordinates, goods identifier, load requirements, and deadline. Subsequently, the atomic task information is packaged into a task announcement message and disseminated throughout the network via a neighbor-based broadcast protocol. After receiving the task announcement, each planning node calculates the expected cost of executing the task, including time cost, energy cost, and conflict risk cost, based on its current state, location, power consumption, and predictions of the future state of the local network. Based on this, it generates a bid value (the lower the value, the better the bid). All nodes send their bid values ​​to the task publishing source via a reverse broadcast path. After collecting all bid values, the publishing source selects the node with the lowest value as the task executor and broadcasts a task allocation confirmation message to the entire network. The node that receives the task adds it to its local task queue, which serves as the basis for calculating the reward function in the subsequent S3 step.

[0065] At the system level, the above method is implemented by an embedded system deployed on each planning node. This system comprises four core functional modules: a distributed state perception and prediction module, a multi-agent reinforcement learning decision-making module, a spatiotemporal conflict resolution and negotiation module, and a task management interface that works in conjunction with them. The distributed state perception and prediction module integrates a high-precision 3D positioning sensor, a six-degree-of-freedom inertial measurement unit, an intelligent battery management unit, and a weighing load sensor. All sensor data achieves microsecond-level time alignment through a hardware synchronization mechanism. The embedded computing unit of this module uses a dedicated AI acceleration chip, containing no fewer than 1024 parallel processing cores and 32 gigabytes of unified addressing memory, ensuring that the spatiotemporal graph convolutional neural network model can complete a full inference within 100 milliseconds, meeting the prediction frequency requirement of 10 times per second. The multi-agent reinforcement learning decision-making module stores the trained deep policy network model in read-only memory with fixed weights. The inference process is completed entirely locally, without relying on external computing resources. The spatiotemporal conflict resolution and negotiation module is implemented using a deterministic algorithm with linear computational complexity, capable of completing trajectory comparison and conflict handling for 300 time steps within 10 milliseconds. The entire system schedules the execution sequence of each module through a real-time operating system, ensuring that the entire process from state acquisition to action output is completed within a 200-millisecond decision cycle.

[0066] This method and system design completely eliminates the dependence on a central server. Even in the event of a complete communication interruption, each node can still make predictions within a limited range based on the last received neighborhood state data and make independent decisions in conjunction with its local task queue. When communication is restored, the node automatically corrects prediction biases by receiving new neighborhood state data, achieving seamless continuity. Experiments show that in a simulated underground tunnel environment, when the communication packet loss rate reaches 70%, the task completion rate of the logistics network planned by this invention remains above 92%, while the completion rate of traditional centralized methods drops to below 35%. The node idle rate is reduced from 28% in traditional methods to 4%, and the number of path conflicts is reduced by 96%. This fully verifies the robustness and efficiency of this technical solution under extreme communication conditions.

[0067] In summary, this embodiment constructs an intelligent logistics planning system capable of stable operation in harsh underground communication environments through a rigorous sequence of steps, deterministic algorithmic logic, high-precision prediction models, and a decentralized decision-making mechanism. All technical features are explicitly disclosed with specific parameters, data structures, and processing flows, ensuring that those skilled in the art can directly implement the system based on this description, and that the scope of protection is clear and unambiguous.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based method for planning underground logistics networks, characterized in that, include: At each planned node of the underground logistics network, the distributed state perception and prediction module it carries collects and integrates its own node state data in real time, and the node state data constitutes a multi-dimensional state vector. The distributed state perception and prediction module uses its built-in spatiotemporal graph convolutional neural network model to take the collected state data of its own node and the historical state data of neighboring nodes obtained through intermittent communication as input to predict the state sequence of the planned node and its neighboring nodes within the communication range in the future within a preset time window, and generate future state time series data containing multiple time steps. At each planning node, the multi-agent reinforcement learning decision module it carries takes the future state time series data generated by the distributed state perception and prediction module as the decision input. The multi-agent reinforcement learning decision module has a built-in pre-trained deep policy network model. The deep policy network model outputs the optimal local action sequence that maximizes the predefined global reward function based on the input future state time series data. At each planning node, the optimal local action sequence output by the multi-agent reinforcement learning decision-making module is converted into the spatiotemporal trajectory of the node within a future preset time window through the spatiotemporal conflict resolution and negotiation module on board, and compared with the predicted spatiotemporal trajectory generated based on the predicted state of adjacent nodes to detect potential spatiotemporal occupancy conflicts. When a spatiotemporal conflict is detected, the spatiotemporal conflict resolution and negotiation module activates a deterministic conflict resolution protocol. The protocol calculates dynamic priorities based on the task priority, remaining running time, and battery power of each conflicting node. After each conflicting node synchronously completes the priority value calculation, the node with the smallest value re-plans the action sequence to avoid the spatiotemporal conflict, thereby achieving autonomous collaboration without communication. The distributed state awareness and prediction module utilizes its built-in spatiotemporal graph convolutional neural network model. Taking its own node state data and historical state data of neighboring nodes obtained through intermittent communication as input, it predicts the state sequence of the planned node and its neighboring nodes within its communication range within a preset future time window, generating future state time-series data containing multiple time steps, including: An adjacency matrix is ​​constructed based on the physical topology of the underground logistics network, where the connection weights between nodes are determined by normalizing the inverse of the Euclidean distance. After organizing the multidimensional state vector and the historical state data of neighboring nodes into a node state sequence according to the time dimension, the sequence is input into the graph convolutional layer to extract spatial dependency features, thus obtaining the spatially enhanced feature vector of each node. The spatially augmented feature vector sequence is input into the gated recurrent unit layer to model the temporal evolution law and obtain the temporal hidden state of each node; The temporal hidden state is input to the fully connected output layer to be mapped to the predicted value of the future state, thereby generating the temporal data of the future state; The spatiotemporal conflict resolution and negotiation module converts the optimal local action sequence into the spatiotemporal trajectory of the node within a preset future time window, and compares it with the predicted spatiotemporal trajectory generated based on the predicted states of neighboring nodes to detect potential spatiotemporal occupancy conflicts, including: Based on the semantics of the action primitives, the local action sequence is converted into a four-dimensional spatiotemporal trajectory consisting of a series of discrete spatiotemporal points. Each spatiotemporal point is represented as a quadruple containing three-dimensional spatial coordinates and a timestamp. The predicted spatiotemporal trajectory of each neighboring node is generated based on the predicted state of the neighboring nodes. The local spatiotemporal trajectory is cross-compared with the predicted spatiotemporal trajectories of all adjacent nodes at each time step. If the Euclidean distance between the spatial coordinates of the local node and any adjacent node is less than the preset safety threshold at any time step, it is determined that there is a spatiotemporal occupancy conflict. The deterministic conflict resolution protocol calculates dynamic priorities based on the task priorities of each conflicting node, the estimated task completion time, and the battery level, including: Calculate a priority score for each conflict node. The priority score is equal to the first weight coefficient multiplied by the task priority code, minus the second weight coefficient multiplied by the estimated task completion time, minus the third weight coefficient multiplied by the normalized battery power value. Among them, the task priority code is pre-specified by the global task publishing system, the estimated task completion time is dynamically estimated by the node based on the distance to the task target point and the current speed, and all weight coefficients are positive real numbers pre-set by the system; Compare the priority scores of all conflicting nodes. The node with the lowest priority score treats the predicted spatiotemporal trajectories of other conflicting nodes as dynamic obstacles and calls its local path planner to regenerate a conflict-free sequence of actions.

2. The artificial intelligence-based underground logistics network planning method according to claim 1, characterized in that, The multi-agent reinforcement learning decision-making module incorporates a pre-trained deep policy network model. Based on the input future state time-series data, the deep policy network model outputs the optimal local action sequence that maximizes the predefined global reward function, including: Future state time series data is used as input to a multilayer perceptron with 5 hidden layers, each containing 1024 neurons, and the activation function is a modified linear unit. The selection probability distribution of five predefined action primitives is output through a flexible maximum value function. The action primitives include: moving to a specified topology node, performing a cargo picking operation, performing a cargo placement operation, entering a standby state, and autonomously navigating to a preset charging facility. Based on the sampling of the selection probability distribution, the specific action primitives to be executed within the current decision cycle are determined, and a local action sequence that is time-aligned with the future state time series data is generated.

3. The artificial intelligence-based underground logistics network planning method according to claim 2, characterized in that, The predefined global reward function is defined as a linear weighted sum, and its specific form is: the global reward function equals the task completion reward weight multiplied by the task completion status value, minus the time consumption penalty weight multiplied by the task execution time, minus the energy consumption penalty weight multiplied by the node energy consumption, and finally minus the collision and conflict penalty weight multiplied by the collision and conflict event indication value; where all weight coefficients are non-negative real numbers, and their sum is 1.

4. The artificial intelligence-based underground logistics network planning method according to claim 1, characterized in that, The safety threshold is 1.5 times the physical size of the planned node.

5. The artificial intelligence-based underground logistics network planning method according to claim 1, characterized in that, The method also includes processing global logistics tasks through a hierarchical task decomposition and allocation module, including: The macro-logistics objectives are decomposed into a series of atomic tasks with clear start points, end points, and load requirements using a hierarchical task network planning algorithm. Atomic task information is disseminated throughout the underground logistics network via a neighbor-based broadcast protocol; Each planning node autonomously calculates the expected cost and efficiency of executing each atomic task based on its current status, location, and prediction of the network status, and generates a bid value. Through a distributed contract network protocol, the node with the best bid obtains the right to execute atomic tasks and announces the task acceptance confirmation information to the entire network through a broadcast protocol.

6. An artificial intelligence-based underground logistics network planning system, characterized in that, The method for implementing the artificial intelligence-based underground logistics network planning method according to any one of claims 1-5 includes: The distributed state perception and prediction module is used to collect and integrate the node state data of each planned node in the underground logistics network in real time. The node state data constitutes a multi-dimensional state vector, which includes the node's real-time three-dimensional spatial coordinates, three-axis velocity components, three-axis acceleration components, remaining battery power, current load weight, and equipment operating status code. The module is also used to use its built-in spatiotemporal graph convolutional neural network model to predict the state sequence of the planned node and its neighboring nodes within the communication range within a future preset time window, using the collected node state data and the historical state data of neighboring nodes obtained through intermittent communication as input, and generate future state time series data containing multiple time steps. A multi-agent reinforcement learning decision module is used to take the future state time series data as decision input. The module has a built-in pre-trained deep policy network model. The deep policy network model outputs the optimal local action sequence that maximizes the predefined global reward function based on the input future state time series data. The spatiotemporal conflict resolution and negotiation module is used to convert the optimal local action sequence into the spatiotemporal trajectory of the node within a future preset time window, and compare it with the predicted spatiotemporal trajectory generated based on the predicted state of adjacent nodes to detect potential spatiotemporal occupancy conflicts. When a spatiotemporal occupancy conflict is detected, the module activates a deterministic conflict resolution protocol. The protocol calculates dynamic priorities based on the task priority, remaining running time, and battery power of each conflicting node. After each conflicting node synchronously completes the priority value calculation, the node with the smallest value replans the action sequence to avoid the spatiotemporal occupancy conflict. The hierarchical task decomposition and allocation module is deployed on several aggregation nodes or global management servers in the network. It is used to decompose high-level logistics instructions into executable atomic tasks and publish them to the entire network through asynchronous broadcast. At the same time, this module is also used to monitor task bidding and acceptance information in the network to maintain a macro view of the global task status.

7. The artificial intelligence-based underground logistics network planning system according to claim 6, characterized in that, The embedded computing unit in the distributed state perception and prediction module has a hardware architecture including a graphics processing unit with at least 1024 processing cores and no less than 32 gigabytes of unified addressing memory to ensure that the spatiotemporal graph convolutional neural network model can complete a complete prediction inference at a frequency of no less than 10 times per second.

Images