A seeding wall scheduling method and system

Abstract

This invention belongs to the field of intelligent logistics and warehouse automation technology, specifically relating to a seeding wall scheduling method and system. By constructing a real-time synchronized three-dimensional digital twin environment, it achieves closed-loop intelligent scheduling through perception, prediction, decision-making, and correction. The method includes decoupling wave tasks into atomic units; dynamically and optimally allocating AGVs based on a multi-factor cost function that integrates travel distance, network load, and order urgency; conducting advanced simulation in the digital twin environment to accurately predict time overlap conflicts at path nodes and queuing conflicts at the seeding wall entrance, and employing differentiated strategies to resolve these conflicts; and simultaneously, making real-time corrections to AGV paths based on real-time communication, trajectory, and order anomaly detection results. This invention achieves a transformation from passive response to proactive prediction, and from local optimization to global collaboration, significantly improving the operational efficiency, response speed, and scheduling robustness of AGV clusters.

Description

Technical Field

[0001] This invention belongs to the field of intelligent logistics and warehouse automation technology, specifically relating to a seeding wall scheduling method and system. Background Technology

[0002] In large e-commerce warehousing and logistics distribution centers, the pre-shipment sorting process is a crucial step affecting overall fulfillment timeliness and accuracy. With the development of automation technology, this process has gradually been replaced by automated equipment such as Automated Guided Vehicles (AGVs) and mobile robots, replacing traditional manual labor. AGVs, following instructions from a central dispatch system, autonomously transport goods to designated sorting walls for precise delivery.

[0003] Current mainstream AGV dispatching methods typically plan a static, fixed sequence of work paths for each AGV when a task is assigned, such as: first go to the pickup point, then to a specific dispatch wall, and finally return to the standby point. The core problem with this method lies in the fundamental contradiction between its static planning and the dynamic environment. The actual warehousing operation environment is full of high uncertainty and dynamic changes, mainly reflected in: orders continuously arriving dynamically, with urgent order insertion needs, and dispatch wall slots dynamically closing or opening due to order boxes being full, malfunctions, etc. In addition, multiple AGVs operating in a limited channel network are prone to space contention at path intersections and convergence points, and abnormal states such as network communication quality fluctuations and occasional AGV equipment failures occur frequently.

[0004] Existing methods are mostly one-time static task allocation and path planning, or simple collision avoidance in local areas. They cannot dynamically reschedule and reallocate resources based on real-time status such as the position of each AGV, task progress, channel congestion, and order urgency. As a result, the overall system efficiency cannot reach the optimal level, and it cannot respond to high-priority needs such as urgent order insertion in a timely manner.

[0005] Secondly, existing methods often focus on simple spatial collision avoidance, failing to proactively predict and differentiate between different types of conflicts. This easily leads to frequent AGV start-stops, increased energy consumption, and can cause order delays and the spread of localized congestion, severely impacting system throughput. When anomalies such as communication interruptions, AGV trajectory deviations, or misdelivered goods occur, existing systems often only issue alarms or simple task resets at the business layer. Their path planning and task allocation struggle to perform real-time, accurate closed-loop corrections based on anomaly information. This results in poor system robustness under abnormal conditions and reduced system operating efficiency.

[0006] In conclusion, existing seeding wall scheduling technology is insufficient to meet the efficiency and robustness requirements of modern intelligent warehousing. Summary of the Invention

[0007] To address the aforementioned problems in the existing technology, this invention provides a seeding wall scheduling method and system, which solves the problems of low system efficiency, slow response, and poor robustness caused by static planning, passive response conflicts, and disconnection from anomaly handling in existing AGV-based seeding wall scheduling technology.

[0008] The objective of this invention can be achieved through the following technical solution: a seeding wall scheduling method, the method comprising the following steps:

[0009] S1: Construct a three-dimensional digital twin environment that maps the instantaneous states of moving and stationary entities in the physical world. The three-dimensional digital twin environment synchronizes the state of the physical world and synchronizes abnormal states in real time.

[0010] S2: Receive seeding wave tasks, decouple and decompose the wave tasks into multiple logically independent atomic task units, and publish them to the global dynamic task pool;

[0011] S3: Based on a preset multi-factor cost function, dynamically select the optimal AGV for each atomic task unit;

[0012] S4: In a three-dimensional digital twin environment, based on the assignment results of each atomic task unit and the preset motion trajectory of the AGV, a spatiotemporal conflict detection algorithm is used to predict the time overlap conflict of the AGV at the path node and the queuing conflict at the entrance of the seeding wall.

[0013] S5: For time overlap conflicts, perform local path replanning or speed coordination adjustment; for queuing conflicts, perform arrival time diversion adjustment or atomic task unit reallocation adjustment, and correct the adjusted path based on real-time detected abnormal status to determine the optimal path for each AGV.

[0014] S6: Publish the optimal path to the corresponding AGV for execution, and synchronously simulate and monitor the execution process in the three-dimensional digital twin environment.

[0015] Preferably, the abnormal states include communication link abnormalities, AGV motion trajectory abnormalities, and order execution abnormalities, and the collection of each abnormality includes the following process:

[0016] Based on real-time communication data, a communication quality heatmap of the warehouse area is generated using the GMM algorithm, and abnormal communication link areas are identified based on the communication quality heatmap.

[0017] The DTW algorithm is used to compare the actual trajectory of each AGV with the predicted trajectory in the 3D digital twin environment in real time to identify the abnormal trajectory status of each AGV.

[0018] The system compares the actual delivery events of goods at the seeding wall with the expected delivery events of the task instructions in real time to identify order execution anomalies.

[0019] Preferably, in S1, the three-dimensional digital twin model is a static three-dimensional environment model based on the warehouse. The instantaneous pose and status data of all operating vehicles, the occupancy status data of all planting wall grids, and the traffic status data of the path are injected through real-time network telemetry to synchronize and simulate the physical environment, thereby obtaining a digital twin environment that maps the instantaneous states of moving and stationary entities in the physical world.

[0020] Preferably, in step S3, the multi-factor cost function is a weighted summation function of the estimated travel distance cost, the future network load penalty cost of the AGV path, and the urgency gain of the order, and the weight coefficients of the multi-factor cost function are optimized using the GG-GMW algorithm.

[0021] The system employs a multi-factor cost function to dynamically calculate and rank the cost values ​​of each AGV through bidding, and selects the AGV with the lowest cost value as the optimal AGV.

[0022] Preferably, the GG-GMW algorithm is used to optimize the weight coefficients of the multi-factor cost function, including the following process:

[0023] The initial weights are determined using the EMW algorithm;

[0024] The initial weights are optimized using the GA algorithm. The fitness function F of the GA algorithm is calculated as follows:

[0025] ;

[0026] In the formula, These are all set balance coefficients. These are the average task completion time, order delay rate, and total system cost, respectively.

[0027] Preferably, the formula for calculating the multi-factor cost function C is as follows:

[0028] ;

[0029] In the formula, The weight coefficients after GA-EMW algorithm optimization are summed to 1, D is the estimated travel distance cost, L is the future network load penalty cost, and E is the order urgency gain.

[0030] Preferably, the calculation of the future network load penalty cost L includes the following process:

[0031] First, calculate the contribution L1 of any planned path to the future network load. The formula for calculating L1 is:

[0032] ;

[0033] Where e represents any element of the sequence set of the planned path. The current real-time load of edge e, This represents the number of AGVs that will traverse edge e in the planned path within the predicted future time window ΔT. The time it takes for the AGV to reach edge e. It is a time decay function;

[0034] After normalizing L1, we get L.

[0035] Preferably, the spatiotemporal conflict detection algorithm performs the following process:

[0036] Generate precise spatiotemporal trajectories for each AGV in future time periods;

[0037] For path nodes, detect whether the time windows in which different AGVs are expected to occupy the safe area of ​​that node overlap;

[0038] For the seeding wall entrance, simulate the queuing process based on the first-come, first-served rule to detect whether there will be waiting area capacity overflow or service position time conflict.

[0039] Preferably, in step S5, correcting the path based on the abnormal state includes the following process:

[0040] When the path crosses a communication anomaly area, replan the detour route or issue a fault-tolerant instruction sequence;

[0041] When an abnormal trajectory of an AGV is detected, its task is marked as pending rescue and a high-priority rescue task is generated;

[0042] When an order execution anomaly is detected, a high-priority error correction and order replacement task is immediately generated and rescheduled.

[0043] A seeding wall scheduling system, the system comprising:

[0044] A digital twin construction module is used to build and maintain the three-dimensional digital twin environment, and to achieve state synchronization with the physical world and anomaly detection;

[0045] The task management decomposition module is used to receive the seeding wave tasks issued by the upper-level warehouse management system and decompose them into multiple logically independent atomic task units.

[0046] The AGV dynamic optimization module is used to select the optimal AGV for the atomic task unit based on the multi-factor cost function.

[0047] The conflict prediction and resolution module is used to perform the spatiotemporal conflict detection, conflict classification, and corresponding adjustment strategies.

[0048] The anomaly response and path correction module is used to correct the AGV path in real time based on the detected anomaly status.

[0049] The route publishing and monitoring module is used to distribute the final determined optimal route to the AGV and perform synchronous monitoring in the digital twin environment.

[0050] The beneficial effects of this invention are as follows:

[0051] By using multi-factor dynamic optimization (distance, network load, order urgency) and predictive conflict resolution, the system achieves a globally optimal match between resources and demands, moving from local optimization to system collaboration and improving the overall coordination capability of the scheduling process. By distinguishing and proactively handling spatiotemporal conflicts and queuing conflicts, the system replaces brute-force emergency stops with refined time / space adjustments, transforming passive response into proactive predictive adjustment. The adjusted paths are then corrected, creating a second-level closed loop from perception to correction, enhancing the overall robustness of the system. This reduces AGV empty runs and waiting, increases throughput to improve operational efficiency, and enhances response agility and system stability. Attached Figure Description

[0052] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0053] Figure 1 This is a schematic diagram of the scheduling method process of the present invention;

[0054] Figure 2 This is a schematic diagram of the process of correcting the path based on the abnormal state in step S5 of the scheduling method of the present invention;

[0055] Figure 3 This is a schematic diagram of the scheduling system structure of the present invention. Detailed Implementation

[0056] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0057] Please see Figures 1-2 This embodiment provides a seeding wall scheduling method, which includes the following steps:

[0058] S1: Construct a three-dimensional digital twin environment that maps the instantaneous states of moving and stationary entities in the physical world, and the three-dimensional digital twin environment synchronizes the state of the physical world and detects abnormal states in real time. The abnormal states include communication link abnormalities, AGV motion trajectory abnormalities, and order execution abnormalities.

[0059] S2: Receive seeding wave tasks, decouple and decompose the wave tasks into multiple logically independent atomic task units, and publish them to the global dynamic task pool;

[0060] S3: Select the optimal AGV for each atomic task unit based on the preset optimization rules;

[0061] S4: In a three-dimensional digital twin environment, based on the optimal AGV task assignment results, preset motion trajectories and work area maps corresponding to each atomic task unit, a spatiotemporal collision detection algorithm is used to predict the time overlap conflict at the path intersection and the queuing conflict at the seeding wall entrance of each AGV.

[0062] S5: For time overlap conflicts, perform local path replanning or speed coordination adjustment; for queuing conflicts, perform arrival time diversion adjustment or atomic task unit reallocation adjustment to determine the AGV path after adjustment.

[0063] S6: Publish the determined AGV paths to each AGV, and synchronously simulate and monitor the execution process of each AGV path in the 3D digital twin environment.

[0064] The detailed execution sub-steps of each step include:

[0065] S1: Construct a 3D digital twin environment that maps the instantaneous states of moving and stationary entities in the physical world, and synchronize the 3D digital twin environment with the physical world's state and abnormal states:

[0066] A 3D digital twin model is a high-precision static 3D environment model built based on the actual physical space of the warehouse. Its construction process requires the integration of core geographic information such as warehouse building CAD drawings, installation location parameters of shelves / planting walls, aisle width dimensions, and ground marking lines. The physical environment is replicated through laser scanning modeling technology to ensure that the position and size of static entities in the model (including shelves, planting walls, columns, safety aisle boundaries, etc.) are completely consistent with the physical world.

[0067] The physical synchronization function of the digital twin environment is achieved through real-time network telemetry technology. It adopts a 5G+edge computing architecture to collect and inject multi-dimensional real-time data at a frequency of 100ms / time. Specifically, this includes: instantaneous pose data of all AGVs (including X / Y / Z three-dimensional coordinates, heading angle, and angular velocity), equipment status data (battery power, load weight, drive module operating status, sensor working status, and fault codes); occupancy status data of all seeding wall grids (idle, occupied, awaiting delivery, delivery completed, fault locked), order number and product information corresponding to the grid; and traffic status data of the path, including the traffic congestion level of each channel segment, temporary construction closure signs, and priority passage permissions.

[0068] After data injection, a data fusion algorithm is used to verify and complete the multi-source heterogeneous data, removing abnormal and noisy data, and interpolating missing data temporarily to ensure data consistency. Based on the completed data, the digital twin environment updates the motion trajectory of moving entities and the state parameters of stationary entities in real time, achieving synchronization between the physical environment and the digital twin environment. At the same time, a simulation engine simulates dynamic behaviors such as AGV motion inertia and the physical process of goods delivery, ultimately constructing a digital twin environment that can accurately map the instantaneous states of moving and stationary entities in the physical world, providing visualization and data support for subsequent task scheduling, conflict prediction, and anomaly response.

[0069] Furthermore, this includes using the Kalman filter (KF) algorithm to correct biases in multi-source detection data, and dynamically compensating for data transmission delays and sensor noise through state equations and observation equations to ensure consistency between the digital twin environment and the physical world. The formula for calculating the state equation is as follows:

[0070] ;

[0071] In the formula, X is the entity state vector, including the AGV's position, velocity, and attitude. The state prediction value at time k is obtained based on the information at time k-1. Let A be the state estimate at time k-1, A be the state transition matrix, and B be the control matrix, which describes the effect of the control input on the state. Let be the control input vector at time k-1. for The process noise at time t is subject to a Gaussian distribution, and the calculation formula for the observation equation is as follows:

[0072] ,in, Let H be the observation vector at time k, and let H be the observation matrix, which describes how the state vector is mapped to the observation vector. The set observation noise;

[0073] Furthermore, abnormal states include communication link anomalies, AGV motion trajectory anomalies, and order execution anomalies. The detection process for each anomaly includes:

[0074] Communication link anomaly detection: Based on real-time acquisition of AGV communication signal strength, data packet loss rate, and transmission delay, a Gaussian Mixture Model (GMM) algorithm is used to generate and dynamically update a communication quality heatmap of the warehouse area. Anomaly areas in the communication link are identified according to preset thresholds. The specific process includes:

[0075] I1: By deploying sensors on the AGV vehicle terminal, seeding wall control node, and fixed base station in the work area, communication data of the communication link is collected in real time. The communication data includes the communication signal strength RSSI, data packet loss rate LPR, and transmission round-trip time RTD. The collection location coordinates, collection timestamp, and unique code of the AGV are added to the data frame of each collected communication data. The raw communication data is normalized to map all data to the 0-1 range to eliminate the influence of units and ensure the effectiveness of subsequent model training.

[0076] I2: Modeling normal communication behavior based on GMM:

[0077] Initialize the GMMM model and set the number K of Gaussian components in the GMM. The value of K can be determined by the Bayesian Information Criterion (BIC) combined with cross-validation. Typically, for a medium-sized warehouse environment, K=3 to 5 can effectively model various normal sub-states such as signal coverage from strong to weak and load from light to heavy. Randomly initialize the weights, mean vector, and covariance matrix for each Gaussian component. The initial covariance matrix is ​​used to reflect the correlation between different communication indicators, such as the negative correlation between signal strength and packet loss rate.

[0078] The EM algorithm is used to iteratively optimize the parameters: the expectation step (E-Step) calculates the posterior probability of each training sample belonging to each Gaussian component; the maximization step (M-Step) updates the mean vector, covariance matrix and initial weights of each Gaussian component based on the posterior probability to maximize the log-likelihood function of the GMM model, and iteratively executes the expectation step and the maximization step until the GMM model parameters converge or the maximum number of iterations is reached to obtain the final GMM model;

[0079] After training, the probability density values ​​of all training data are calculated using the GMM model. The probability density values ​​are sorted in ascending order, and the density value corresponding to a lower quantile (such as 1% or 5%) is taken as the anomaly threshold 'a'. The anomaly threshold is used to indicate that the probability of normal data falling into a region with a density lower than this anomaly threshold is extremely low. Therefore, when the probability density of new data is less than the anomaly threshold, it is determined that the communication link is abnormal.

[0080] I3: Input the real-time communication data and the coordinate data of the AGV carried by the communication data into the GMM model, calculate the probability density value of the communication data. If the probability density value is greater than or equal to a, the communication data is determined to be a normal communication sample. If the probability density value is less than a, it is determined to be a communication anomaly. Record the abnormal GAV number, location coordinates, timestamp, specific abnormal index value (RSSI too low, LPR too high, etc.) and its probability density value corresponding to the abnormal data.

[0081] Furthermore, it also includes incremental learning and bias protection of the GMM model using real-time communication data. To adapt to slow environmental changes, the model is incrementally updated using the following methods:

[0082] Only high-confidence samples that are judged as normal and have good communication quality (e.g., normalized RSSI>0.7 and LPR<0.05) multiple times consecutively (e.g. within a time window) are added to an incremental cache pool.

[0083] Periodically use new samples from the cache pool to fine-tune the GMM model with a small learning rate (Mini-batch EM). Simultaneously, retain a copy of the original model as a baseline.

[0084] The performance of the fine-tuned model and the baseline model is periodically compared on an independent validation set. If the detection performance deteriorates, the model is rolled back to the baseline model, and an environment check alert is triggered. This prevents the model from failing due to learning continuous communication degradation.

[0085] I4: Generates a communication quality heatmap. To overcome the limitations of single-point detection, the system spatially visualizes the discrete detection results to distinguish between equipment faults and regional interference.

[0086] Heatmaps are generated using spatial interpolation: All communication data reported by AGVs, including both normal and abnormal data, are aggregated within a fixed time window (e.g., 5 minutes). For each communication indicator (e.g., RSSI or comprehensive probability density value), spatial interpolation is performed on the warehouse two-dimensional plane, using the two-dimensional coordinates of each data point as support points and the indicator value as height. Inverse distance weighted (IDW) interpolation or Kriging interpolation is used to generate a continuous communication quality raster (heatmap) covering the entire area.

[0087] On the heatmap generated based on probability density, areas with density values ​​below the abnormal threshold 'a' are colored as abnormal areas. At the same time, heatmaps based on original indicators (such as RSSI) can be overlaid for auxiliary analysis.

[0088] I5: Perform anomaly location:

[0089] If an AGV continuously reports abnormal data, but its location is not an abnormal area on the heat map, and other AGVs in the vicinity communicate normally, the communication failure is determined to be from the AGV's own communication module.

[0090] If a heatmap shows that a certain physical area is in a continuous abnormal state, and multiple different AGVs in that area report abnormalities, it can be determined that there is communication interference or infrastructure (such as base stations or mobile hotspots) failure in that area.

[0091] After locating the problem, the corresponding strategy is triggered: for the faulty AGV, an instruction is issued to restrict its entry into complex areas and prioritize its return for maintenance; for areas with communication abnormalities, the path can be avoided during path planning, and maintenance personnel are notified to inspect and repair the infrastructure in the designated area.

[0092] The AGV motion trajectory anomaly detection process includes:

[0093] Based on the preset AGV motion trajectory (including path nodes, driving speed, and turning timing) in the 3D digital twin environment, the actual AGV trajectory and the predicted trajectory are compared in real time through the Dynamic Time Warping (DTW) algorithm. During the comparison process, the motion inertia error of the AGV is considered, and the allowable range of trajectory deviation is set. The allowable range of deviation includes the allowable range of planar coordinate deviation and the allowable range of heading angle deviation.

[0094] The DTW algorithm uses dynamic programming to find the optimal matching path between two trajectories and calculates the similarity distance between the trajectories. When the similarity distance exceeds a preset threshold, or when the trajectory deviation exceeds the allowable range for n consecutive sets of periods (n can be 5), trajectory anomaly identification is triggered. At the same time, the anomaly type is further determined, including: path offset anomaly (actual position deviates from the preset path), speed anomaly (actual speed exceeds the preset speed range ±20%), and stagnation anomaly (AGV stagnation time exceeds the preset threshold of 5s and there is no task pause instruction).

[0095] Detection of order execution anomalies: Deploy barcode sensors and weight sensors in each grid of the planting wall to collect real-time data on product delivery events, including: product barcode information, delivery timestamp, delivery grid number, and product weight. At the same time, extract expected delivery event information from the task instructions, including: product barcode to be delivered, target grid number, planned delivery time window, and theoretical product weight.

[0096] The system performs multi-dimensional comparisons between actual and expected delivery events to identify three types of order execution anomalies: 1) Misdelivery anomaly, where the product barcode matches but the delivery slot does not match the target slot, or the slot matches but the product barcode does not match; 2) Missed delivery anomaly, where the corresponding slot does not detect the delivery signal of the target product after the planned delivery time window has ended; 3) Information mismatch anomaly, where the product barcode and slot match, but the actual weight deviates from the theoretical weight by more than ±5%, or the product batch, specifications, and order requirements do not match. After anomaly identification, anomaly details are recorded in real time.

[0097] S2: Receive seeding wave tasks, decouple and decompose the wave tasks into multiple logically independent atomic task units, and publish them to the global dynamic task pool. This process includes the following steps:

[0098] The system receives planting wave tasks from the upper-level system. These tasks include planting requirements for multiple orders, covering core parameters such as product information, target planting grid size, work time limit, and priority. An improved K-means clustering algorithm is used to decouple and decompose the wave tasks. With the goal of maximizing order similarity and maximizing task independence, a clustering objective function is constructed. The formula for calculating the clustering objective function is as follows:

[0099] ;

[0100] In the formula, For the i-th atomic task unit, The task similarity function is calculated based on the target grid distance, product type, and priority. As cluster center, The similarity penalty function, ranging from 0.3 to 0.5, uses an improved K-means clustering algorithm to decouple and decompose the wave task into multiple logically independent atomic task units. Each unit corresponds to part or all of the product seeding requirements of a single order, possessing a complete pickup-delivery closed loop without dependencies. After decomposition, each unit is labeled with parameters (pickup location, target grid, product details, etc.) and published to a global dynamic task pool. The global dynamic task pool updates the task status (pending assignment, assigned, etc.) in real time based on a hash table structure, supporting rapid querying and adjustment in subsequent scheduling processes.

[0101] S3: Select the optimal AGV for each atomic task unit based on the preset optimization rules;

[0102] The optimization rules include using a preset multi-factor cost function to calculate the cost value of each AGV through dynamic bidding and sorting them, and selecting the AGV with the lowest cost value to win the bid;

[0103] The multi-factor cost function includes a weighted summation function of the estimated travel distance cost required to execute the atomic task unit, the future network load penalty cost of the AGV path, the processing efficiency coefficient corresponding to the target seeding wall grid, and the urgency gain of the order associated with the atomic task unit. The weight coefficients are optimized using a genetic algorithm and entropy weight method (GA-EMW). The specific implementation process includes:

[0104] The formula for calculating the multi-factor cost function is as follows:

[0105] ;

[0106] In the formula, The weight coefficients after the GA-EMW algorithm optimization are summed to 1. D is the estimated travel distance cost, L is the future network load penalty cost, and E is the order urgency gain. All four are normalized to the range of 0-1.

[0107] The first step is to calculate each cost factor, including the following process:

[0108] Estimated travel distance cost D: based on pre-calculated Dijkstra's or A The distance matrix is ​​used to calculate the shortest path length of the AGV from its current location to the pickup point and then to the target planting wall grid.

[0109] Based on the energy consumption per unit distance of the AGV specified in the design parameters The calculation formula is: The denominator uses the maximum value function to take the maximum energy consumption value of similar historical tasks to ensure that D is normalized. The higher the energy consumption, the larger the value of D.

[0110] The future network load penalty cost L is used to predict the impact of AGV task execution on the future congestion level of the path network;

[0111] The planned path for any GAV vehicle to perform any atomic task is a sequence of edges. The contribution L1 of this planned path to the future network load is calculated using the following formula:

[0112] ;

[0113] In the formula, e represents any element of the sequence set of the planned path. This represents the current real-time load of edge e, i.e., the number of AGVs currently passing through. This represents the number of AGVs that will traverse edge e in the planned path within the predicted future time window ΔT. The time it takes for the AGV to reach edge e. The time decay function represents the degree of decay of the load effect. The Min-Max normalization formula is used to normalize L1, and the original costs of different orders of magnitude and different units (such as distance, load count, queue length, etc.) are uniformly mapped to the [0,1] interval, so that they can be directly weighted and summed to avoid the problem of inconsistent dimensions.

[0114] The order urgency gain E reflects the timeliness requirement of an order; the higher the urgency, the lower the cost.

[0115] Calculate the initial urgency level E1:

[0116] ;

[0117] In the formula, The deadline for this order as specified for the atomic task (the promised deadline for the order). For the predicted estimated completion time of the order, The current system time. For urgency reward function, To prevent the elimination of the zero minimum constant, the calculated E1 value is normalized.

[0118] The second step is to optimize the three weight coefficients using GA-EMW. First, the EMW algorithm is used to determine the initial weights, and then the GA algorithm is used to optimize the initial weights. The specific process includes:

[0119] II1: Determine initial weights using the EMW algorithm:

[0120] II11 collects historical task allocation sample data. The sample data includes historical task execution records extracted from the warehouse management system (WMS) and AGV scheduling logs over a period of time. Each record should include: the four original cost factor values ​​of the atomic task unit (travel distance, path load, order urgency), as well as the actual completion time of the task, whether there is a delay, and other result indicators. Abnormal data is cleaned and records with obvious errors or unusable records are removed.

[0121] II12: Construct an N-row, 3-column evaluation matrix from the cleaned N valid records. Each row in the matrix represents a historical task sample, and each column corresponds to a cost factor. Since the dimensions and orders of magnitude of each cost factor are different, the data in each column is first normalized and scaled to the range of 0 to 1 to eliminate the influence of dimensions and lay the foundation for subsequent entropy calculation.

[0122] II13: Calculate the initial weights according to the EMW algorithm: For each cost factor, calculate its proportion in all historical samples, and calculate the information entropy value of each cost factor according to information theory.

[0123] Subtracting the information entropy value from 1 yields the difference coefficient. The larger the difference coefficient, the stronger the evaluation effect of the factor. Normalizing the difference coefficients of each factor yields the initial weight set based on the objective laws of historical data.

[0124] II2: Optimize the initial weights using the GA algorithm:

[0125] II21: The weight optimization problem is transformed into a genetic algorithm search problem. Each possible weight combination (satisfying that the weight sum is 1 and all weights are non-negative) is treated as an individual. Using the initial weights obtained by the entropy weight method calculated in II1 as the core, a certain number (e.g., 50) of weight combinations are randomly generated in its vicinity to form the initial population.

[0126] II22: Using the weight combinations obtained in II21, the task scheduling process is replayed or simulated in a simulation environment built with historical data. The comprehensive index scores corresponding to each weight combination are then established based on the fitness function calculation. Higher scores indicate better scheduling performance from that weight combination. The formula for calculating the fitness function F is:

[0127] ;

[0128] In the formula, These are all set balance coefficients. These are the average task completion time, order delay rate, and total system cost, respectively.

[0129] II23: Perform iterative evolution. Based on the fitness score of each individual (weight combination), select using mechanisms such as roulette wheel selection. The new individuals generated through selection, crossover, and mutation form a new generation of population, replacing the old population. Repeat steps II22 and II23 to perform iterative evolution until the maximum number of generations (e.g., 200 generations) is reached or the fitness function converges, thus completing the optimization.

[0130] II24: Based on the continuous collection of new scheduling performance data during runtime, the weights are updated using incremental learning at fixed intervals.

[0131] II3: For each newly arrived atomic task unit, perform the following steps:

[0132] The available AGV set is filtered based on the current status of the AGV (battery level, load capacity, health status);

[0133] The original values ​​of the three cost factors are calculated in parallel for each element (AGV vehicle) in the AGV set;

[0134] For each newly arrived atomic task unit, Min-Max normalization is performed on each factor of the candidate AGV;

[0135] The multi-factor cost function with optimized weight coefficients is applied to calculate the overall cost of each GAV in parallel.

[0136] The AGV with the lowest overall cost is selected as the winner and determined as the optimal AGV for executing the atomic task unit.

[0137] Based on preset optimization rules, the optimal AGV is selected for each atomic task unit. A multi-factor cost function is obtained by weighting three cost factors, and the GA-EMW algorithm is used to optimize the weight values ​​of the multi-factor cost function. By combining the objective and data-driven characteristics of the Entropy Weight Method (EMW) with the powerful global search capability of the Genetic Algorithm (GA), EMW provides a high-quality search starting point and a benchmark for online fine-tuning, avoiding GA from getting stuck in local optima or blindly searching. GA can find better weight combinations that EMW alone cannot discover, and the fitness function is directly linked to the scheduling performance index, so the weight optimization process is the process of improving scheduling performance. It continuously seeks the optimal solution, providing a solid and adaptive foundation for intelligent scheduling decisions.

[0138] S4: In a three-dimensional digital twin environment, based on the optimal AGV task assignment results, preset motion trajectories and work area maps corresponding to each atomic task unit, a spatiotemporal collision detection algorithm is used to predict the time overlap conflict at the path intersection and the queuing conflict at the seeding wall entrance of each AGV.

[0139] Since GAVs may experience spatial and temporal conflicts during operation—that is, two AGVs occupying the same spatial location or having overlapping spatial locations within the same time period—the spatiotemporal collision detection algorithm is used to predict temporal overlap conflicts at path intersections and queuing conflicts at the seeding wall entrance. This process includes the following steps:

[0140] S41: For each AGV (numbered g) of an assigned atomic task unit, according to A... The algorithm generates a preset motion trajectory, kinematic model, and current initial state, which in turn generate a precise spatiotemporal trajectory for the AGV within a future time period T in the digital twin environment. This spatiotemporal trajectory is represented as a function of time t and is used to describe the AGV's coordinates and orientation in a two-dimensional map at the current moment.

[0141] S42: Define conflict detection nodes including time conflicts at path intersections and queuing conflicts at the seeding wall entrance:

[0142] For each path intersection, which includes path crossing points, merging points, and diverging points, collect all AGVs expected to pass through the path intersection and their expected arrival times. If the expected arrival time difference between two or more AGVs is less than a set first time threshold, it is determined that there is a time overlap conflict at the path intersection.

[0143] For each seeding wall entrance, all AGVs expected to arrive at that entrance and their expected arrival times are collected. If the expected arrival time difference is less than a set second time threshold, it is determined that there is a queuing conflict at the seeding wall entrance.

[0144] S43: Execution path intersection time overlap conflict detection:

[0145] S431: For each path node, perform the following operations:

[0146] Calculate the estimated time for each AGV that plans to pass through the node to enter and leave the safe zone surrounding the node. The safe zone is the physical size of the AGV plus a preset safety margin.

[0147] Compare the time windows of all AGVs that plan to pass through the node. If the time periods of any two AGVs that are expected to occupy the safe area of ​​the node overlap, it is determined that a time overlap conflict has occurred. Even if the time windows do not directly overlap, if the interval between the two is less than the preset safe time threshold, it is marked as a potential conflict.

[0148] S432: For each seeding wall entrance area, simulate an ordered work queue. The specific process includes:

[0149] Collect all AGVs whose path target is the seeding wall, and sort them according to the expected arrival time of each AGV at the entrance area based on the FCFS algorithm;

[0150] Considering the limited physical capacity of the waiting area at the entrance of the seeding wall (e.g., only able to accommodate N vehicles waiting simultaneously), simulate the queuing process for each AGV:

[0151] Since an AGV must wait until there is an empty space in the waiting area before it can enter, the actual time it is allowed to enter is the later of its estimated arrival time and the time when the next space in the waiting area becomes available.

[0152] After entering the waiting area, the AGV still needs to wait for the seeding wall service position (delivery position) to become available. The time when it starts delivering services is the later of the time it enters the waiting area and the time when the previous AGV finishes its service and leaves.

[0153] During the simulated queuing process, if it is found that any two AGVs are assigned to occupy the seeding wall service position in the exact same time period, it is determined that there is a service position conflict.

[0154] If the simulation results show that the delay time from the arrival of the AGV to being allowed to enter the waiting area is too long, or the waiting time exceeds the acceptable threshold of the system, then a queuing conflict is determined to have occurred, indicating that the entrance area is about to become congested or has already become congested.

[0155] The AGV cluster conflict is precisely classified into two risks: path space-time contention and facility resource queuing. In a three-dimensional digital twin environment, based on the predicted AGV motion trajectory, a fundamental leap from static spatial collision to dynamic spatial collaboration is achieved. The physical waiting area capacity constraint is integrated with the FCFS scheduling algorithm to realize the prediction of queuing congestion and overflow risks. The paradigm shift is from passive response to active prediction and from geometric collision avoidance to system optimization scheduling. The entire process of AGV operation is monitored, and subsequent replanning is performed for predicted conflicts or potential conflicts to resolve the conflicts.

[0156] S5: For time overlap conflicts, perform local path replanning or speed coordination adjustment; for queuing conflicts, perform arrival time diversion adjustment or atomic task unit reallocation adjustment. Determine the adjusted AGV path, and correct the adjusted AGV path based on the detected abnormal status to determine the optimal path for each AGV. The specific process includes the following:

[0157] S51: For time-overlapping conflicts, perform local path replanning or speed coordination adjustment.

[0158] Identify the conflict point and define its scope: Determine the coordinates of the intersection point where the conflict occurs and the estimated time window; delineate an affected area; prioritize the AGVs involved in the conflict based on the urgency gain of the tasks performed by the AGVs involved; and select the AGVs to be adjusted in descending order of priority.

[0159] The system replans the local path for the selected adjustment object to bypass the original conflict point. During replanning, the system searches for the shortest feasible alternative path between the current AGV position and the downstream necessary point, avoiding the conflict point and minimizing new time conflicts.

[0160] If replanning the local path cannot resolve the current conflict, then the speed of all conflicting AGVs will be adjusted in a coordinated manner, appropriately reducing the speed of the following vehicle and increasing the speed of the preceding vehicle, to ensure that after the adjustment, the time interval between each AGV passing through the conflict point is greater than the preset safety threshold.

[0161] S52: Replanning for queuing conflicts: When it is predicted that the number of AGVs arriving at a certain seeding wall entrance will exceed the waiting area capacity, perform the following steps:

[0162] Identify the seed wall entrance where an overflow is about to occur, and all AGVs that will arrive at this entrance within the expected arrival time window and their order;

[0163] Without changing the route, adjust the arrival time of AGVs that are not at the head of the queue:

[0164] The AGV is instructed to temporarily wait at the last decision point before arrival (e.g., 50 meters from the entrance of the seeding wall). By controlling the waiting time, the interval between AGVs arriving at the entrance is increased, ensuring that the number of AGVs waiting at the entrance of the seeding wall at the same time does not exceed the physical capacity.

[0165] If time diversion cannot solve the problem (e.g., excessive waiting time will cause order timeout), then task reallocation will be initiated, which will reassign some of the AGV's atomic task units to other available seed walls and update their destinations and paths.

[0166] S53: Based on real-time detected communication link anomalies, AGV device behavior anomalies, and order execution anomalies, the path adjusted by S51 or S52 is corrected in real time to ensure the feasibility and robustness of the path in a dynamic environment. The correction process includes:

[0167] Correction based on abnormal communication link status:

[0168] When the communication quality heatmap shows a specific area as a communication degradation zone or a communication interruption zone, the AGV path involving that area should be corrected:

[0169] During the adjustment, the detour route for AGVs that traverse communication anomaly areas is replanned, and a larger penalty cost is imposed on the path segments in the anomaly area in the multi-factor cost function to guide the planning algorithm to generate alternative paths that completely avoid the area.

[0170] For situations that cannot be completely avoided, or where communication is downgraded to temporary or localized, a fault-tolerant design should be adopted. The design should include:

[0171] Before the AGV enters the degraded zone, a buffered instruction sequence containing multiple consecutive actions is sent to its on-board controller through a still stable link to ensure that the AGV can still drive according to the preset logic during the brief period of disconnection.

[0172] At the same time, a behavior instruction is issued after the loss of contact. For example, if no further instruction is received within a specified time, the driver should proceed slowly along the current passage to the next prominent landmark (such as an intersection sign) and then stop and wait.

[0173] Temporarily dispatch an AGV located in a good signal area or activate a fixed relay device as a mobile relay node to attempt to provide temporary signal relay for AGVs in the degraded area;

[0174] When the DTW algorithm detects abnormal behavior such as serious trajectory deviation, abnormal stagnation, or speed loss of a certain AGV, it performs correction. The correction process includes: marking all atomic task units currently carried by the faulty AGV as pending rescue status, generating a high-priority rescue task package, which includes two sub-tasks: one is to dispatch a new AGV to the fault point to transfer the goods; the other is to continue transporting the transferred goods to the original destination or safe area.

[0175] When the seeding wall sensor or verification system detects order execution anomalies such as misdelivery or omission of goods, the system performs real-time closed-loop error correction at the business level and feeds back to the scheduling path layer. This includes: the WMS (Warehouse Management System) automatically generating a highest-priority error correction and supplementary atomic task unit within seconds. This task clearly defines the required goods, the correct target grid, and an extremely short completion time limit, assigns an AGV to it, and plans a direct or fast path with minimal interference for execution.

[0176] By distinguishing between two core types of conflicts—time overlap and queuing conflicts—and applying targeted strategies such as local replanning / speed adjustment and time diversion / task reallocation respectively, and then correcting the adjusted paths based on detected anomalies, conflicts can be quickly eliminated with minimal scheduling actions and resource consumption. This avoids the computational complexity and response latency caused by global replanning, significantly reducing AGV stalls, path congestion, and order outbound delays caused by conflicts. The targeted strategies use time differences to replace inefficient space grabbing or long-distance detours. This allows most conflicts to be resolved through fine-tuning of speed or short waiting in safe areas, avoiding unnecessary extra travel distances. Thus, while ensuring timeliness, it reduces the overall energy consumption of the AGV cluster and extends equipment runtime.

[0177] S6: Publish the determined AGV paths to each AGV and simulate and monitor them synchronously in the 3D digital twin environment.

[0178] This embodiment constructs a three-dimensional digital twin environment as a virtual sandbox, transforming traditional static scheduling into predictive dynamic scheduling. It establishes an intelligent closed-loop logic of perception-prediction-decision-execution-correction, synchronizing the state and anomalies of the physical world in real time. In the twin environment, it simulates the future movements of the AGV group in advance, accurately predicting the spatiotemporal contention conflicts at path nodes and the resource queuing conflicts at the seeding wall entrance. It also adopts targeted differentiated resolution strategies, fine-tuning the speed or replanning locally for time conflicts, diverting arrival times or redistributing tasks for queuing conflicts, and integrating abnormal states such as communication interruptions and equipment failures into the subsequent path correction process. This enables anomaly-driven instantaneous path correction and emergency task generation, thereby forming a continuous self-optimization capability in a dynamic environment.

[0179] The technical effects achieved by this application include: 1. By using multi-factor (distance, network load, order urgency) dynamic optimization and conflict prediction to resolve conflicts, the global optimal matching of resources and demands is realized, improving the global coordination capability of the scheduling process from local optimization to system coordination; 2. By distinguishing and proactively handling spatiotemporal conflicts and queuing conflicts, fine-grained time / space adjustments replace crude emergency stops and waiting, transforming passive response into proactive prediction and adjustment; 3. The adjusted path is corrected, achieving a second-level closed loop from perception to correction, enhancing the overall robustness of the system, improving operational efficiency by reducing AGV empty runs and waiting, increasing throughput, and improving response agility and system stability.

[0180] Please see Figure 3 This embodiment provides a seeding wall scheduling system. This system is used to execute the seeding wall scheduling method corresponding to any of the aforementioned method embodiments. Its hardware is deployed on an edge computing server cluster in a warehouse center and connected to on-site AGVs, seeding walls, and various sensors via a 5G private network. System modules communicate and collaborate through a unified data bus and message middleware.

[0181] The system includes:

[0182] The digital twin construction module, based on pre-input 3D warehouse model data (including building structure, shelving layout, seeding wall location, and aisle topology), builds a 1:1 mapping of the physical warehouse into a 3D digital twin environment. The module collects and injects dynamic entity data and environmental and operational status data at high frequency through a real-time data interface. The dynamic entity data comes from the real-time pose, speed, power consumption, and equipment status of all AGVs, while the environmental and operational status data comes from the seeding wall control system's grid occupancy status, order information, and path access status from the monitoring system.

[0183] The digital twin construction module includes a data fusion and filtering unit and an anomaly detection unit. The data fusion and filtering unit uses the Kalman filter algorithm to verify and complete multi-source data, ensuring that the digital twin environment remains synchronized with the physical world and providing a high-fidelity simulation foundation for the system. The anomaly detection unit is used to analyze communication data, AGV trajectory data, and order delivery data in real time, generating a communication quality heatmap and identifying trajectory deviations and order execution errors.

[0184] The task management and decomposition module receives picking wave tasks from the upper-level warehouse management system (WMS). The task decoupling unit within the module employs an intelligent algorithm based on cluster analysis. According to the target location of the order, product attributes, and urgency, it decomposes complex wave tasks into multiple logically independent, parallel-executable atomic task units. For example, an atomic task unit might retrieve 5 items of product SKU123 from picking station A and deliver them to slot number 15 on picking wall B.

[0185] All resolved atomic task units are published to a globally shared dynamic task pool. The task pool maintains the status of each task unit (pending assignment, assigned, executing, completed) using a defined hash table, allowing other modules to query and update it in real time.

[0186] The AGV dynamic optimization module, when a new task is released in the global dynamic task pool, calculates the comprehensive cost of all available AGVs executing the task for each atomic task unit to be assigned in parallel. The module includes a multi-factor cost function calculation unit and a weight optimization unit. The multi-factor cost function calculation unit comprehensively considers the estimated travel distance cost, the future network load penalty cost of the path, and the order urgency gain based on the set multi-factor cost function. The weight optimization unit, based on a mechanism that combines historical data-driven optimization algorithms, is responsible for dynamically adjusting the weights of each factor in the cost function, continuously optimizing the scheduling strategy. Finally, the AGV dynamic optimization module selects the AGV with the lowest comprehensive cost and binds it to the task unit, completing the dynamic optimization of AGV allocation.

[0187] After AGV task assignment is completed, the conflict prediction and resolution module performs high-precision spatiotemporal simulation of the future movement of all AGVs in a three-dimensional digital twin environment.

[0188] The conflict prediction and resolution module includes a spatiotemporal conflict detection unit and a conflict resolution unit. The spatiotemporal conflict detection unit detects static collisions and predicts two types of dynamic conflicts: path node time overlap conflicts (different AGVs have estimated arrival times that are too close at intersections and other nodes) and seed wall entrance queuing conflicts (the estimated number of arriving AGVs exceeds the physical capacity of the entrance). Upon detecting a conflict, the conflict resolution unit initiates differentiated strategies based on the conflict type: for time overlap conflicts, local path replanning or speed coordination adjustment is used; for queuing conflicts, arrival time diversion or task reassignment is triggered. All adjustments take effect after simulation verification in a virtual environment.

[0189] The anomaly response and path correction module continuously monitors abnormal status signals. This module includes a path avoidance unit and a task rescue unit. Upon receiving an anomaly (such as communication interruption, AGV malfunction, or order misdelivery), the module immediately initiates the corresponding closed-loop correction procedure. For example, for areas with communication anomalies, the path avoidance unit will replan the path of the affected AGVs; for equipment malfunctions, the task rescue subunit will generate high-priority rescue tasks and reschedule them. Correction instructions from the anomaly response and path correction module are fed back in real-time to the conflict prediction and resolution module and the AGV dynamic optimization module, ensuring the timely updating and global consistency of the overall scheduling plan.

[0190] The path publishing and monitoring module receives the finalized optimal paths for each AGV. Through a secure communication protocol, these path instructions are broken down into serialized control commands, which are then sent to the corresponding physical AGV onboard controllers. This drives the 3D digital twin environment to perform a 1:1 simulation execution synchronized with the physical world. The module also displays the real-time position, status, task progress, and system alarm information of all AGVs on a visual monitoring interface, providing maintenance personnel with a comprehensive situational awareness and decision support.

[0191] In this embodiment, the system constructs and synchronizes a virtual environment; when a new wave of tasks arrives, it decomposes and publishes the new wave of tasks to a task pool, retrieves tasks from the task pool in real time, and assigns appropriate AGVs to them; it simulates, predicts, and resolves potential conflicts in the initial allocation scheme; it processes various real-time anomalies in parallel and corrects the paths; finally, it issues safe, efficient, and reliable optimal path instructions to the AGVs for execution and monitors the entire process. The entire process forms an intelligent closed loop of perception-prediction-decision-execution-correction, realizing efficient and accurate scheduling between the AGV cluster and the seeding wall in a dynamic and complex warehousing environment.

[0192] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for scheduling seeding walls, characterized in that: The method includes the following steps: S1: Construct a three-dimensional digital twin environment that maps the instantaneous states of moving and stationary entities in the physical world. The three-dimensional digital twin environment synchronizes the state of the physical world and synchronizes abnormal states in real time. S2: Receive seeding wave tasks, decouple and decompose the wave tasks into multiple logically independent atomic task units, and publish them to the global dynamic task pool; S3: Based on a preset multi-factor cost function, dynamically select the optimal AGV for each atomic task unit; S4: In a three-dimensional digital twin environment, based on the assignment results of each atomic task unit and the preset motion trajectory of the AGV, a spatiotemporal conflict detection algorithm is used to predict the time overlap conflict of the AGV at the path node and the queuing conflict at the entrance of the seeding wall. S5: For time overlap conflicts, perform local path replanning or speed coordination adjustment; for queuing conflicts, perform arrival time diversion adjustment or atomic task unit reallocation adjustment, and correct the adjusted path based on real-time detected abnormal status to determine the optimal path for each AGV. S6: Publish the optimal path to the corresponding AGV for execution, and synchronously simulate and monitor the execution process in the three-dimensional digital twin environment; In step S3, a multi-factor cost function is used to calculate the cost value of each AGV through dynamic bidding and sort them. The AGV with the lowest cost value is selected as the optimal AGV. The GG-GMW algorithm is used to optimize the weight coefficients of the multi-factor cost function, including the following process: The initial weights are determined using the EMW algorithm; The initial weights are optimized using the GA algorithm. The fitness function F of the GA algorithm is calculated as follows: ； In the formula, These are all set balance coefficients. These are the average task completion time, order delay rate, and total system cost, respectively. The formula for calculating the multi-factor cost function C is as follows: ； In the formula, The weight coefficients after GA-EMW algorithm optimization are summed to 1, D is the estimated travel distance cost, L is the future network load penalty cost, and E is the order urgency gain. The calculation of the future network load penalty cost L includes the following process: First, calculate the contribution L1 of any planned path to the future network load. The formula for calculating L1 is: ； Where e represents any element of the sequence set of the planned path. The current real-time load of edge e, This represents the number of AGVs that will traverse edge e in the planned path within the predicted future time window ΔT. The time it takes for the AGV to reach edge e. It is a time decay function; After normalizing L1, we get L.

2. The seeding wall scheduling method according to claim 1, characterized in that: The abnormal states include communication link abnormalities, AGV motion trajectory abnormalities, and order execution abnormalities. The collection of each abnormality includes the following process: Based on real-time communication data, a communication quality heatmap of the warehouse area is generated using the GMM algorithm, and abnormal communication link areas are identified based on the communication quality heatmap. The DTW algorithm is used to compare the actual trajectory of each AGV with the predicted trajectory in the 3D digital twin environment in real time to identify the abnormal trajectory status of each AGV. The system compares the actual delivery events of goods at the seeding wall with the expected delivery events of the task instructions in real time to identify order execution anomalies.

3. The seeding wall scheduling method according to claim 1, characterized in that: In S1, the three-dimensional digital twin model is a static three-dimensional environment model based on the warehouse. Through real-time network telemetry, the instantaneous pose and status data of all operating vehicles, the occupancy status data of all planting wall grids, and the traffic status data of the path are injected to synchronize and simulate the physical environment, thereby obtaining a digital twin environment that maps the instantaneous states of moving and stationary entities in the physical world.

4. The seeding wall scheduling method according to claim 1, characterized in that: In step S4, the spatiotemporal conflict detection algorithm performs the following process: Generate precise spatiotemporal trajectories for each AGV in future time periods; For path nodes, detect whether the time windows in which different AGVs are expected to occupy the safe area of ​​that node overlap; For the seeding wall entrance, simulate the queuing process based on the first-come, first-served rule to detect whether there will be waiting area capacity overflow or service position time conflict.

5. The seeding wall scheduling method according to claim 1, characterized in that: In step S5, correcting the path based on the abnormal state includes the following process: When the path crosses a communication anomaly area, replan the detour route or issue a fault-tolerant instruction sequence; When an abnormal trajectory of an AGV is detected, its task is marked as pending rescue and a high-priority rescue task is generated; When an order execution anomaly is detected, a high-priority error correction and order replacement task is immediately generated and rescheduled.

6. A seeding wall scheduling system, applied to a seeding wall scheduling method as described in any one of claims 1-5, characterized in that: The system includes: The digital twin building module is used to build and maintain a 3D digital twin environment, and to achieve state synchronization with the physical world and anomaly detection. The task management decomposition module is used to receive seeding wave tasks issued by the upper-level warehouse management system (WMS) and decompose them into multiple logically independent atomic task units. The AGV dynamic optimization module is used to select the optimal AGV for atomic task units based on a multi-factor cost function. The conflict prediction and resolution module is used to perform spatiotemporal conflict detection, conflict classification, and corresponding adjustment strategies. The anomaly response and path correction module is used to correct the AGV path in real time based on the detected anomaly status. The route publishing and monitoring module is used to distribute the final determined optimal route to the AGV and perform synchronous monitoring in the digital twin environment.

Images