A container export-oriented robot scheduling method and system

Abstract

The application discloses a kind of container export-oriented robot scheduling method and system, it is related to intelligent scheduling technical field, and the multi-source data is collected by Internet of Things sensor network, and after time-space alignment, input double-depth Q network completes task pre-allocation;Rolling time domain path optimization is carried out based on three-dimensional space-time coordinate system, and trajectory conflict is solved by combining virtual guide point strategy;The weight of multi-objective cost function is dynamically adjusted using long short-term memory network, and adaptive optimization of transportation efficiency, energy consumption and other indicators is realized;For abnormal events, local conflict-free path is generated using bee colony optimization algorithm, and energy consumption is reduced by integrating container door orientation optimization strategy, which improves the collaborative scheduling efficiency of AGV cluster in container terminal and enhances the dynamic environmental adaptability.

Description

Technical Field

[0001] This invention relates to the field of intelligent scheduling technology, specifically to a robot scheduling method and system for container export. Background Technology

[0002] The core of robot scheduling for container exports is to achieve "seamless transfer" of containers from the storage area to the quay crane at the terminal by efficiently coordinating various types of robots, so as to ultimately meet the core objectives of on-time delivery, equipment utilization, and operational efficiency.

[0003] Existing technologies for scheduling container export robots have significant drawbacks: the lack of effective spatiotemporal fusion of multi-source data (container location, AGV status, task queue) leads to insufficient task allocation accuracy; static path planning makes it difficult to predict trajectory conflicts, and the delayed response to conflict resolution can easily cause congestion; the fixed weight of scheduling targets makes it impossible to dynamically balance efficiency and energy consumption based on working conditions; the rigid abnormal event handling mechanism results in excessively long local rescheduling times; and the impact of container physical properties on energy consumption is not considered, resulting in poor overall adaptability and economy. Summary of the Invention

[0004] To address the aforementioned technical shortcomings, the present invention aims to provide a robot scheduling method for container export.

[0005] To solve the above technical problems, the present invention adopts the following technical solution: The present invention provides a robot scheduling method for container export, including: S1, multi-source data acquisition: real-time acquisition of multi-source data through an Internet of Things sensor network deployed at the terminal site.

[0006] S2, Task Pre-allocation: After spatiotemporally aligning the collected data, it is input into a dual-depth Q-network. The parallel computing channels generate AGV feature vectors and task feature vectors respectively. The value score of each AGV-task pairing is calculated using the two-stream interaction mechanism of Hadamard product and feature concatenation, and the optimal initial allocation scheme is output.

[0007] S3. Rolling Time Domain Path Optimization: Establish a three-dimensional spatiotemporal coordinate system to predict the movement trajectory of all AGVs in the future time period, determine whether there is a spatiotemporal conflict between any two trajectories, and activate the virtual guide point conflict resolution strategy when a spatiotemporal conflict between any two trajectories is detected.

[0008] S4. Design a dynamic weighted multi-objective cost function: The function includes four indicators: total transportation time, total cluster energy consumption, predicted number of conflict points, and emergency task non-completion rate. The historical working condition data stream is analyzed in real time through a long short-term memory network, and the weight coefficients of the four indicators are dynamically adjusted.

[0009] S5. Local rescheduling in response to abnormal events: When an abnormal event is detected, the affected area is delineated with the event occurrence point as the center. The bee colony optimization algorithm is used to generate a new conflict-free path for the AGVs in this area, and the door orientation optimization strategy is executed.

[0010] Preferably, the multi-source data includes export container location distribution data, AGV cluster status data, and dynamic task queue data.

[0011] The export container location distribution data includes container location coordinates, dimensions, target vessels, and cut-off times.

[0012] The AGV cluster status data includes the real-time location, remaining battery power, load status, and operating speed of each AGV.

[0013] The dynamic task queue data includes a list of boxes to be transferred, task generation time, and priority identifier.

[0014] Preferably, the method for calculating the value score of each AGV-task pairing is as follows: the container pose data is encoded into a three-dimensional matrix containing Cartesian coordinate values ​​and length values, with the number of rows in the matrix equal to the number of containers to be transferred; the AGV status data is encoded into a three-dimensional matrix containing Cartesian coordinate values ​​and speed values, with the number of rows in the matrix equal to the total number of available AGVs; the AGV battery power data is converted into a normalized vector; the task deadline data is converted into a vector of remaining seconds based on the current system time; and the above four types of data are structurally combined into the input state space of the deep reinforcement learning model.

[0015] The dual-depth Q-network adopts a dual-channel fully connected architecture. The first channel extracts AGV feature vectors through a fully connected layer, and the second channel extracts task feature vectors through a fully connected layer. During the feature extraction process, the ReLU activation function is used to perform element-wise multiplication on the two types of feature vectors. The result of concatenating the vectors is then linearly weighted and fused to obtain the AGV-task pairing value score matrix, and thus obtains the value score of each AGV-task pairing.

[0016] Preferably, the method for determining whether any two trajectories have a spatiotemporal conflict is as follows: based on the dock plane rectangular coordinate system, the time dimension is extended to form a three-dimensional spatiotemporal coordinate system.

[0017] The spline curve interpolation algorithm is used to calculate the continuous motion trajectory of each AGV in the future time period. The interpolation control points are determined by the real-time position, speed and acceleration data of the AGV.

[0018] Calculate the minimum Euclidean distance between any two trajectories and compare it with a preset Euclidean distance threshold in the database. If the minimum Euclidean distance between two trajectories is greater than the preset Euclidean distance threshold, it is determined that there is no spatiotemporal conflict between the two trajectories. If the minimum Euclidean distance between two trajectories is less than the preset Euclidean distance threshold, it is initially determined that there is a spatiotemporal conflict between the two trajectories. Further calculate the time difference between the trajectory points, take the absolute value, and compare the absolute value of the time difference between the two trajectory points with a preset time difference threshold in the database. If the absolute value of the time difference between the two trajectory points is less than the preset time difference threshold, it is finally determined that there is a spatiotemporal conflict between the two trajectories.

[0019] Preferably, the activation of the virtual guide point conflict resolution strategy is specifically implemented as follows: an RFID tag is set up as a virtual guide point upstream of the path intersection where a conflict is detected; the passage time window is dynamically calculated based on the real-time speed of the AGV; and a mutually exclusive passage time window sequence is allocated to the conflicting AGVs through a time division multiple access mechanism.

[0020] Preferably, the dynamic adjustment of the weight coefficients of the four indicators is specifically implemented as follows: the input layer of the Long Short-Term Memory network receives a four-dimensional temporal feature vector, including the average number of path conflicts per minute within the rolling window, the proportion of currently incomplete tasks to the maximum processing capacity, the average remaining battery power percentage of the AGV cluster, and the proportion of tasks with insufficient time to reach the port; the hidden layer of the network learns the temporal dependencies of the working condition features through forget gate, input gate, and output gate mechanisms; the output layer is normalized by the Softmax function to generate adjustment coefficients for the four parameters: transportation time weight, energy consumption weight, conflict weight, and emergency task weight.

[0021] Preferably, the method for generating conflict-free new paths for AGVs within the region using the bee colony optimization algorithm is as follows: multiple candidate paths are generated between the current position and the target position of the affected AGV to form an initial solution set. Each candidate path is scored using a multi-dimensional evaluation function that includes path execution efficiency, conflict risk, and path smoothness. Then, a search strategy that simulates bee colony foraging behavior is used to perform neighborhood search, crossover, and mutation operations on the candidate paths to gradually optimize the solution set. When the preset number of iterations is met, the optimal path solution is output.

[0022] Preferably, the container door orientation optimization strategy specifically includes: adding a container door orientation angle parameter to the task features and quantifying the angle value into a finite number of discrete directions; calculating the difference in container door orientation angles between adjacent tasks when the AGV continuously executes multiple tasks; introducing a steering energy consumption penalty term related to load status and steering angle into the reward function of deep reinforcement learning; adding a steering angle change rate during path planning; statistically calculating the steering energy consumption savings based on historical operating data; and dynamically adjusting the priority parameters of the door orientation matching algorithm when the steering energy consumption savings reach a preset threshold.

[0023] A second aspect of the present invention provides a system for a robot scheduling method for container export, comprising: a multi-source data acquisition module for real-time acquisition of multi-source data through an Internet of Things sensor network deployed at the terminal site.

[0024] Task pre-assignment module: After spatiotemporally aligning the collected data, it inputs it into a dual-depth Q-network, generates AGV feature vectors and task feature vectors through parallel computing channels, calculates the value score of each AGV-task pairing using the Hadamard product and feature concatenation dual-stream interaction mechanism, and outputs the optimal initial assignment scheme.

[0025] The rolling time-domain path optimization module is used to establish a three-dimensional spatiotemporal coordinate system to predict the motion trajectory of all AGVs in the future time period, determine whether there is a spatiotemporal conflict between any two trajectories, and activate the virtual guide point conflict resolution strategy when a spatiotemporal conflict between any two trajectories is detected.

[0026] The design module is a dynamic weighted multi-objective cost function. The function includes four indicators: total transportation time, total cluster energy consumption, predicted number of conflict points, and emergency task non-completion rate. It uses a long short-term memory network to analyze historical operating data streams in real time and dynamically adjusts the weight coefficients of the four indicators.

[0027] Local rescheduling module for responding to abnormal events: When an abnormal event is detected, it delineates the affected area centered on the event occurrence point, uses the bee colony optimization algorithm to generate a conflict-free new path for the AGVs in this area, and executes the door orientation optimization strategy.

[0028] The beneficial effects of this invention are as follows: by optimizing task allocation accuracy through multi-source data spatiotemporal alignment and dual-depth Q-network, improving conflict control capabilities based on three-dimensional spatiotemporal coordinate system and virtual guide point strategy, realizing dynamic weight adaptive adjustment by utilizing long short-term memory network, and rapidly responding to abnormal events by combining bee colony optimization algorithm, while integrating container door orientation optimization to reduce energy consumption, the efficiency, adaptability and economy of the scheduling system are significantly improved, realizing intelligent and efficient collaborative scheduling of robots in container terminals. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention;

[0031] Figure 2 This is a schematic diagram of the system structure connection of the present invention. Detailed Implementation

[0032] 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, and 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.

[0033] according to Figure 1 As shown, the present invention provides a robot scheduling method for container export, including: S1, multi-source data acquisition: real-time acquisition of multi-source data through an Internet of Things sensor network deployed at the terminal site.

[0034] In one specific embodiment, the multi-source data includes export container location distribution data, AGV cluster status data, and dynamic task queue data.

[0035] The export container location distribution data includes container location coordinates, dimensions, target vessels, and cut-off times.

[0036] The AGV cluster status data includes the real-time location, remaining battery power, load status, and operating speed of each AGV.

[0037] The dynamic task queue data includes a list of boxes to be transferred, task generation time, and priority identifier.

[0038] S2, Task Pre-allocation: After spatiotemporally aligning the collected data, it is input into a dual-depth Q-network. The parallel computing channels generate AGV feature vectors and task feature vectors respectively. The value score of each AGV-task pairing is calculated using the two-stream interaction mechanism of Hadamard product and feature concatenation, and the optimal initial allocation scheme is output.

[0039] It should be noted that the dual deep Q network is an improved model in deep reinforcement learning, optimized from the traditional deep Q network (DQN). It sets up two neural networks with the same structure but independent parameters (the main network and the target network), which are responsible for action value estimation and target value calculation, respectively. This can effectively alleviate the Q-value overestimation problem in the traditional DQN and improve decision stability. In this scenario, the dual deep Q network is used to process multi-source data after spatiotemporal alignment. By learning the correlation between AGV and task, it outputs a more accurate AGV-task pairing value score.

[0040] The parallel computing refers to setting up two independent computing channels in the dual-depth Q network to extract features from AGV status data and task data simultaneously: one channel focuses on extracting AGV feature vectors from data such as AGV position and power, while the other channel simultaneously extracts task feature vectors from data such as task priority and port closing time. This parallel processing method avoids the waiting delay of serial computing, improves feature extraction efficiency, and adapts to the real-time requirements of terminal scheduling.

[0041] The Hadamard product is a vector / matrix operation that involves element-wise multiplication of two vectors (or matrices) of the same dimension to obtain a new vector (or matrix) of the same dimension. In the dual-stream interaction mechanism of this invention, the Hadamard product is used for element-level interaction between AGV feature vectors and task feature vectors, capturing the correlation strength between the two in corresponding dimensions (such as the matching degree between AGV battery power and task urgency); simultaneously, combined with feature concatenation (connecting the two vectors end-to-end to form a higher-dimensional vector), it comprehensively integrates element-level correlations and global features, and finally calculates a more accurate AGV-task pairing value score through linear weighting.

[0042] In one specific embodiment, the method for calculating the value score of each AGV-task pairing is as follows: the container pose data is encoded into a three-dimensional matrix containing Cartesian coordinate values ​​and length values, with the number of rows in the matrix equal to the number of containers to be transferred; the AGV status data is encoded into a three-dimensional matrix containing Cartesian coordinate values ​​and speed values, with the number of rows in the matrix equal to the total number of available AGVs; the AGV battery power data is converted into a normalized vector; the task deadline data is converted into a vector of remaining seconds based on the current system time; and the above four types of data are structurally combined into the input state space of the deep reinforcement learning model.

[0043] The dual-depth Q-network adopts a dual-channel fully connected architecture. The first channel extracts AGV feature vectors through a fully connected layer, and the second channel extracts task feature vectors through a fully connected layer. During the feature extraction process, the ReLU activation function is used to perform element-wise multiplication on the two types of feature vectors. The result of concatenating the vectors is then linearly weighted and fused to obtain the AGV-task pairing value score matrix, and thus obtains the value score of each AGV-task pairing.

[0044] S3. Rolling Time Domain Path Optimization: Establish a three-dimensional spatiotemporal coordinate system to predict the movement trajectory of all AGVs in the future time period, determine whether there is a spatiotemporal conflict between any two trajectories, and activate the virtual guide point conflict resolution strategy when a spatiotemporal conflict between any two trajectories is detected.

[0045] In one specific embodiment, the method for determining whether any two trajectories have a spatiotemporal conflict is as follows: based on the dock plane rectangular coordinate system, the time dimension is extended to form a three-dimensional spatiotemporal coordinate system.

[0046] The spline curve interpolation algorithm is used to calculate the continuous motion trajectory of each AGV in the future time period. The interpolation control points are determined by the real-time position, speed and acceleration data of the AGV.

[0047] It should be noted that spline curve interpolation is a mathematical method for constructing smooth curves using piecewise polynomial functions. In AGV trajectory prediction, cubic B-spline curves are commonly used. The method for determining the control points is as follows: using the AGV's current coordinates as the current position point P0, calculate the velocity influence point P1 based on the current speed (P1 = P0 + Δt * current speed, where Δt is the time step), and combine this with acceleration calculations to obtain the acceleration influence point P2 (P2 = P1 + 1 / 2Δt). 2 *Acceleration), and take the task target point or the next key node in the path planning as the target position point P3. Construct a cubic B-spline curve through these four points to describe the motion trajectory of the AGV in the short term. This method integrates physical kinematic parameters into the mathematical model, ensuring the physical rationality and mathematical computability of the trajectory.

[0048] Calculate the minimum Euclidean distance between any two trajectories and compare it with a preset Euclidean distance threshold in the database. If the minimum Euclidean distance between two trajectories is greater than the preset Euclidean distance threshold, it is determined that there is no spatiotemporal conflict between the two trajectories. If the minimum Euclidean distance between two trajectories is less than the preset Euclidean distance threshold, it is initially determined that there is a spatiotemporal conflict between the two trajectories. Further calculate the time difference between the trajectory points, take the absolute value, and compare the absolute value of the time difference between the two trajectory points with a preset time difference threshold in the database. If the absolute value of the time difference between the two trajectory points is less than the preset time difference threshold, it is finally determined that there is a spatiotemporal conflict between the two trajectories.

[0049] It should be noted that the calculation of the minimum Euclidean distance between any two trajectories is existing technology and will not be elaborated upon here. The preset Euclidean distance threshold in the database needs to comprehensively consider the physical properties of the AGV and the characteristics of the operating environment: based on the maximum external dimensions of the AGV, the error tolerance of positioning and path tracking is superimposed, and additional safety margins are added with reference to port automation safety specifications to form a distance critical value that can both avoid mechanical collisions and adapt to actual operational errors, ensuring that the AGV maintains a physical safety distance during movement. The calculation of the time difference value of trajectory points is existing technology and will not be elaborated upon here. The preset time difference threshold in the database is based on the motion performance of the AGV and the system response characteristics: the time required for safe braking is calculated in combination with the maximum operating speed of the AGV, and the response delay of the system from collision detection to instruction execution is included to form a time critical value that can cover the braking process and system reaction, ensuring that the AGV has enough time to avoid the detected risk of close-range trajectory intersection.

[0050] In one specific embodiment, the activation of the virtual guide point conflict resolution strategy is implemented as follows: an RFID tag is set up as a virtual guide point upstream of the path intersection where a conflict is detected; the passage time window is dynamically calculated based on the real-time speed of the AGV; and a mutually exclusive passage time window sequence is allocated to the conflicting AGVs through a time division multiple access mechanism.

[0051] It should be noted that the time-division multiple access mechanism divides the passage time of the virtual guide point into multiple non-overlapping time slots, and dynamically allocates time slots according to the AGV's expected arrival time, task priority, and real-time speed. This allows each AGV to exclusively occupy the path intersection within its dedicated time slot, avoiding spatiotemporal conflicts through orderly scheduling in the time sequence. The time slot length is calculated based on the AGV size, speed, and safety margin, and can be dynamically adjusted according to the working conditions. This ensures the determinism of AGV passage, efficiently utilizes resources, and adapts to fluctuations in task priority and changes in AGV status during dock operations. It is the core technology for resolving virtual guide point conflicts.

[0052] S4. Design a dynamic weighted multi-objective cost function: The function includes four indicators: total transportation time, total cluster energy consumption, predicted number of conflict points, and emergency task non-completion rate. The historical working condition data stream is analyzed in real time through a long short-term memory network, and the weight coefficients of the four indicators are dynamically adjusted.

[0053] In one specific embodiment, the dynamic adjustment of the weight coefficients of the four indicators is implemented as follows: the input layer of the Long Short-Term Memory network receives a four-dimensional temporal feature vector, including the average number of path conflicts per minute within the rolling window, the proportion of currently incomplete tasks to the maximum processing capacity, the average remaining battery percentage of the AGV cluster, and the proportion of tasks with insufficient time to reach the port; the hidden layer of the network learns the temporal dependencies of the working condition features through forget gate, input gate, and output gate mechanisms; the output layer is normalized by the Softmax function to generate adjustment coefficients for the four parameters: transportation time weight, energy consumption weight, conflict weight, and emergency task weight.

[0054] It should be noted that the forget gate, input gate, and output gate mechanisms are the core gating mechanisms of Long Short-Term Memory (LSTM) networks, used to solve the gradient vanishing problem in traditional recurrent neural networks and to handle long-term dependencies in temporal data. Specifically, the forget gate filters and discards irrelevant information from the previous hidden state and the current input; the input gate selects key information from the current input and updates the cell state; and the output gate controls the transmission of relevant information from the cell state to the next step. These three mechanisms work together to enable LSTM to effectively retain important historical information, absorb key real-time data, and transmit effective information, thereby accurately learning the long-term dependencies in temporal data.

[0055] S5. Local rescheduling in response to abnormal events: When an abnormal event is detected, the affected area is delineated with the event occurrence point as the center. The bee colony optimization algorithm is used to generate a new conflict-free path for the AGVs in this area, and the door orientation optimization strategy is executed.

[0056] It should be noted that the aforementioned swarm optimization algorithm is a swarm intelligence optimization algorithm inspired by the foraging behavior of bee colonies. It solves complex optimization problems by simulating the process of bees working together to find the optimal food source. Its core is to abstract the solution to the optimization problem as a "food source" and set up three types of "artificial bees": hired bees are responsible for mining known food sources (local search) and optimizing solutions through neighborhood perturbation; observer bees choose to follow based on the food source quality information (such as fitness) transmitted by hired bees, reinforcing the search for high-quality solutions; and scout bees randomly explore new areas when food sources are depleted (global search), avoiding getting trapped in local optima. Through information sharing and division of labor, these three types of bees achieve efficient searching of the solution space, exhibiting characteristics such as fewer parameters, strong robustness, and outstanding global optimization capabilities.

[0057] In one specific embodiment, the method of generating conflict-free new paths for AGVs in the region using the bee colony optimization algorithm is as follows: multiple candidate paths are generated between the current position and the target position of the affected AGV to form an initial solution set. Each candidate path is scored by a multi-dimensional evaluation function that includes path execution efficiency, conflict risk and path smoothness. Then, a search strategy that simulates bee colony foraging behavior is used to perform neighborhood search, crossover and mutation operations on the candidate paths to gradually optimize the solution set. When the preset number of iterations is met, the optimal path solution is output.

[0058] It should be noted that the generation of multiple candidate paths to form an initial solution set involves first rasterizing the affected area and extracting navigation key points, then using the A* algorithm to generate basic paths, randomly perturbing the basic paths, generating tangent paths based on obstacle distribution, and reusing historical successful paths with adaptive adjustments. At the same time, parameter perturbations are introduced into each strategy to enhance diversity. Finally, paths with obvious conflicts are filtered out to form an initial solution set containing multiple path characteristics.

[0059] In one specific embodiment, the container door orientation optimization strategy is as follows: adding a container door orientation angle parameter to the task features and quantifying the angle value into a finite number of discrete directions; calculating the difference in container door orientation angles between adjacent tasks when the AGV continuously executes multiple tasks; introducing a steering energy consumption penalty term related to load status and steering angle into the reward function of deep reinforcement learning; adding a steering angle change rate during path planning; statistically calculating the steering energy consumption savings based on historical operating data; and dynamically adjusting the priority parameters of the door orientation matching algorithm when the steering energy consumption savings reach a preset threshold.

[0060] It should be noted that the calculation of the difference in the orientation angle of the boxes between adjacent tasks is existing technology and will not be elaborated on further here.

[0061] according to Figure 2 As shown, a system for a robot scheduling method for container exports includes: a multi-source data acquisition module for real-time acquisition of multi-source data through an Internet of Things (IoT) sensor network deployed at the terminal site.

[0062] Task pre-assignment module: After spatiotemporally aligning the collected data, it inputs it into a dual-depth Q-network, generates AGV feature vectors and task feature vectors through parallel computing channels, calculates the value score of each AGV-task pairing using the Hadamard product and feature concatenation dual-stream interaction mechanism, and outputs the optimal initial assignment scheme.

[0063] The rolling time-domain path optimization module is used to establish a three-dimensional spatiotemporal coordinate system to predict the motion trajectory of all AGVs in the future time period, determine whether there is a spatiotemporal conflict between any two trajectories, and activate the virtual guide point conflict resolution strategy when a spatiotemporal conflict between any two trajectories is detected.

[0064] The design module is a dynamic weighted multi-objective cost function. The function includes four indicators: total transportation time, total cluster energy consumption, predicted number of conflict points, and emergency task non-completion rate. It uses a long short-term memory network to analyze historical operating data streams in real time and dynamically adjusts the weight coefficients of the four indicators.

[0065] Local rescheduling module for responding to abnormal events: When an abnormal event is detected, it delineates the affected area centered on the event occurrence point, uses the bee colony optimization algorithm to generate a conflict-free new path for the AGVs in this area, and executes the door orientation optimization strategy.

[0066] It should be noted that it also includes a database, which stores preset Euclidean distance thresholds, preset time difference thresholds, and historical running data.

[0067] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in this specification, they should all fall within the protection scope of the present invention.

Claims

1. A robot scheduling method for container export, characterized in that, include: S1. Multi-source data acquisition: Real-time acquisition of multi-source data through an IoT sensor network deployed at the dock site; S2, Task Pre-allocation: After spatiotemporally aligning the collected data, input it into a dual-depth Q network. AGV feature vectors and task feature vectors are generated through parallel computing channels. The value score of each AGV-task pairing is calculated using the dual-stream interaction mechanism of Hadamard product and feature concatenation, and the optimal initial allocation scheme is output. S3. Rolling Time Domain Path Optimization: Establish a three-dimensional spatiotemporal coordinate system to predict the motion trajectory of all AGVs in the future time period, determine whether there is a spatiotemporal conflict between any two trajectories, and activate the virtual guide point conflict resolution strategy when a spatiotemporal conflict between any two trajectories is detected. S4. Design a dynamic weighted multi-objective cost function: The function includes four indicators: total transportation time, total cluster energy consumption, predicted number of conflict points, and emergency task non-completion rate. The historical working condition data stream is analyzed in real time through a long short-term memory network, and the weight coefficients of the four indicators are dynamically adjusted. S5. Local rescheduling in response to abnormal events: When an abnormal event is detected, the affected area is delineated with the event occurrence point as the center. The bee colony optimization algorithm is used to generate a new conflict-free path for the AGVs in this area, and the door orientation optimization strategy is executed.

2. The robot scheduling method for container export according to claim 1, characterized in that, The multi-source data includes export container location distribution data, AGV cluster status data, and dynamic task queue data; The export container location distribution data includes container location coordinates, dimensions, target vessels, and cut-off times; The AGV cluster status data includes the real-time location, remaining battery power, load status, and operating speed of each AGV. The dynamic task queue data includes a list of boxes to be transferred, task generation time, and priority identifier.

3. The robot scheduling method for container export according to claim 2, characterized in that, The method for calculating the value score of each AGV-task pairing is as follows: The container pose data is encoded into a three-dimensional matrix containing Cartesian coordinates and length values, with the number of rows equal to the number of containers to be transferred. The AGV status data is encoded into a three-dimensional matrix containing Cartesian coordinates and speed values, with the number of rows equal to the total number of available AGVs. The AGV battery power data is converted into a normalized vector. The task deadline data is converted into a vector of remaining seconds based on the current system time. The above four types of data are structured and combined into the input state space of the deep reinforcement learning model. The dual-depth Q-network adopts a dual-channel fully connected architecture. The first channel extracts AGV feature vectors through a fully connected layer, and the second channel extracts task feature vectors through a fully connected layer. During the feature extraction process, the ReLU activation function is used to perform element-wise multiplication on the two types of feature vectors. The result of concatenating the vectors is then linearly weighted and fused to obtain the AGV-task pairing value score matrix, and thus obtains the value score of each AGV-task pairing.

4. A robot scheduling method for container export according to claim 3, characterized in that, The specific implementation method for determining whether any two trajectories have a spatiotemporal conflict is as follows: Based on the Cartesian coordinate system of the dock, the time dimension is extended to form a three-dimensional spatiotemporal coordinate system; The spline curve interpolation algorithm is used to calculate the continuous motion trajectory of each AGV in the future time period. The interpolation control points are determined by the real-time position, speed and acceleration data of the AGV. Calculate the minimum Euclidean distance between any two trajectories and compare it with a preset Euclidean distance threshold in the database. If the minimum Euclidean distance between two trajectories is greater than the preset Euclidean distance threshold, it is determined that there is no spatiotemporal conflict between the two trajectories. If the minimum Euclidean distance between two trajectories is less than the preset Euclidean distance threshold, it is initially determined that there is a spatiotemporal conflict between the two trajectories. Further calculate the time difference between the trajectory points, take the absolute value, and compare the absolute value of the time difference between the two trajectory points with a preset time difference threshold in the database. If the absolute value of the time difference between the two trajectory points is less than the preset time difference threshold, it is finally determined that there is a spatiotemporal conflict between the two trajectories.

5. A robot scheduling method for container export according to claim 4, characterized in that, The specific implementation method of the activation virtual guidance point conflict resolution strategy is as follows: Radio frequency identification (RFID) tags are set up as virtual guidance points upstream of the path intersection where a conflict is detected. The passage time window is dynamically calculated based on the real-time speed of the AGV, and a mutually exclusive passage time window sequence is allocated to the conflicting AGVs through a time division multiple access (TDMA) mechanism.

6. A robot scheduling method for container export according to claim 5, characterized in that, The specific method for dynamically adjusting the weight coefficients of the four indicators is as follows: The input layer of the Long Short-Term Memory (LSTM) network receives a four-dimensional temporal feature vector, including the average number of path conflicts per minute within the rolling window, the proportion of currently incomplete tasks to the maximum processing capacity, the average remaining battery percentage of the AGV cluster, and the proportion of tasks with insufficient time to reach the port. The hidden layer of the network learns the temporal dependencies of the working condition features through forget gates, input gates, and output gates. The output layer is normalized by the Softmax function to generate adjustment coefficients for four parameters: transportation time weight, energy consumption weight, conflict weight, and emergency task weight.

7. A robot scheduling method for container export according to claim 6, characterized in that, The method for generating conflict-free new paths for AGVs within the region using the bee colony optimization algorithm is as follows: Multiple candidate paths are generated between the current position and the target position of the affected AGV to form an initial solution set. Each candidate path is scored by a multi-dimensional evaluation function that includes path execution efficiency, conflict risk and path smoothness. Then, a search strategy that simulates bee colony foraging behavior is used to perform neighborhood search, crossover and mutation operations on the candidate paths to gradually optimize the solution set. When the preset number of iterations is met, the optimal path solution is output.

8. A robot scheduling method for container export according to claim 7, characterized in that, The door orientation optimization strategy is as follows: The container door orientation angle parameter is added to the task features and the angle value is quantified into a finite number of discrete directions. When the AGV executes multiple tasks consecutively, the difference in container door orientation angle between adjacent tasks is calculated. A steering energy consumption penalty term related to load status and steering angle is introduced into the reward function of deep reinforcement learning. At the same time, the steering angle change rate is added during the path planning process. The steering energy consumption savings are statistically calculated based on historical operation data. When the steering energy consumption savings reach a preset threshold, the priority parameters of the door orientation matching algorithm are dynamically adjusted.

9. A system applying the robot scheduling method for container export according to any one of claims 1-8, characterized in that, include: Multi-source data acquisition module: Used to acquire multi-source data in real time through an IoT sensor network deployed at the dock site; Task pre-allocation module: After spatiotemporally aligning the collected data, it inputs it into a dual-depth Q network, generates AGV feature vectors and task feature vectors through parallel computing channels, calculates the value score of each AGV-task pairing using the Hadamard product and feature concatenation dual-stream interaction mechanism, and outputs the optimal initial allocation scheme. The rolling time-domain path optimization module is used to establish a three-dimensional spatiotemporal coordinate system to predict the motion trajectory of all AGVs in the future time period, determine whether there is a spatiotemporal conflict between any two trajectories, and activate the virtual guide point conflict resolution strategy when a spatiotemporal conflict between any two trajectories is detected. Design a dynamic weighted multi-objective cost function module: The function includes four indicators: total transportation time, total cluster energy consumption, predicted number of conflict points, and emergency task non-completion rate. It uses a long short-term memory network to analyze historical working condition data streams in real time and dynamically adjusts the weight coefficients of the four indicators. Local rescheduling module for responding to abnormal events: When an abnormal event is detected, it delineates the affected area centered on the event occurrence point, uses the bee colony optimization algorithm to generate a conflict-free new path for the AGVs in this area, and executes the door orientation optimization strategy.

Images