A tray intelligent scheduling method, system and device

Abstract

The present application belongs to the technical field of intelligent logistics and industrial automation, and specifically relates to a tray intelligent scheduling method, system and device, which comprises: constructing the correlation relationship of scene factors, constraint conditions and weight factors and quantifying the correlation degree, introducing a correlation constraint strength correction term to complete accurate quantization of uncertain scenes; adopting a collaborative processing mechanism of parallel correlation guidance simulation and deviation optimization, combining a lightweight digital twin to generate an adaptive multi-objective weight scheme; the system is configured with a hierarchical system architecture of a perception layer, a correlation modeling layer and a double-path synchronous layer, and the device computing unit of the device is coordinated by the edge and the cloud, and short-term and long-term double-cycle optimization is realized through three-dimensional deviation feedback cycle. The problems of insufficient dynamic adaptation capability and low scheduling efficiency of the existing tray scheduling technology are solved, the dynamic adaptability of the scheduling scheme to complex scenes is effectively improved, and the efficiency of tray intelligent scheduling is greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent logistics and industrial automation technology, specifically a method, system and device for intelligent pallet scheduling. Background Technology

[0002] Pallet intelligent scheduling technology is a core supporting link in warehouse automation and industrial material flow systems. It is widely used in e-commerce warehousing, production workshops and other scenarios. Its scheduling effect directly affects the efficiency of logistics flow and operating costs. However, existing technologies still have insurmountable technical shortcomings in the application of real complex scenarios, and cannot meet the needs of efficient and accurate intelligent scheduling.

[0003] CN110689256B discloses a pallet scheduling method, apparatus, electronic device, and warehouse management system, including acquiring orders to be processed;

[0004] Based on the pending orders, the pickable items in each pallet in the current warehouse, and the preset constraints, determine the pallets that need to be scheduled so that the number of pallets that need to be scheduled is minimized, and the pallets that need to be scheduled can meet the item requirements of the pending orders.

[0005] The constraints include: the number of pallets that need to be scheduled in any aisle of the warehouse is less than or equal to the preset maximum number of pallets; the number of pallets that need to be scheduled is determined after filtering based on the constraints, after identifying multiple alternative pallet scheduling schemes based on the product requirements of the orders to be processed and the pickable products in each pallet.

[0006] Control the forklift to transport the pallets that need to be dispatched.

[0007] The main shortcomings of the relevant technologies are: insufficient dynamic adaptability and low scheduling efficiency. Existing tray-based intelligent scheduling technology cannot effectively cope with uncertain signal disturbances and changes in the correlation between factors in the scenario. The scheduling scheme is difficult to dynamically adjust with the scenario, resulting in poor adaptability. Moreover, existing technologies mostly adopt a serial optimization mode, which results in long time consumption for weight scheme generation, simulation calculation and scheduling instruction output, which cannot meet the real-time scheduling requirements.

[0008] Therefore, a method, system, and device for intelligent pallet scheduling are proposed to address the above problems. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies by providing a method, system, and apparatus for intelligent tray scheduling, thereby solving the technical problems mentioned in the background art.

[0010] To achieve the above objectives, this application proposes a method for intelligent pallet scheduling, comprising:

[0011] Based on historical data of pallet scheduling, the correlation between scenario factors, constraints and weight factors of pallet scheduling scenarios is constructed, and the correlation between scenario factors and weight factors is quantified as the degree of correlation.

[0012] Collect deterministic data and uncertain signals from the pallet scheduling scenario, perform quantitative modeling of uncertain scenarios under correlation constraints based on correlation relationships, and generate correlation-constrained scenario disturbance factors;

[0013] Based on the scenario disturbance factor with correlation constraints and the tray-specific multi-objective weight factor system, multiple weight schemes are generated by a multi-objective genetic algorithm that introduces correlation constraint mechanism. The weight schemes must meet the correlation constraint screening conditions and the scheduling logic self-consistency verification conditions.

[0014] Multiple weighting schemes are input into a lightweight digital twin, and digital twin simulation and weight optimization under associated constraints are performed simultaneously and collaboratively. The comprehensive score and logical rationality score of each scheme are output. The optimal weighting scheme is selected according to the selection logic of prioritizing logical rationality, prioritizing constraint adaptation, adapting to scenario risks, and balancing multiple objectives, and then tray scheduling is performed.

[0015] The simulation deviation, constraint adaptation deviation, and correlation matching deviation during the data collection, scheduling, and execution process are collected to construct a three-dimensional feedback loop for continuous short-term and long-term dual-cycle optimization of the weight factor system.

[0016] Preferably, the degree of correlation is within a preset degree of correlation threshold range, which is determined by calibration using historical operation data of tray scheduling.

[0017] The degree of correlation is the quantitative result when constructing the correlation relationship, and the correlation relationship includes at least the influence relationship between AGV current mutation and pallet load adaptation safety factor weight, and the influence relationship between path bumpiness degree and pallet loss avoidance cost weight.

[0018] The multi-objective weight factor system for the tray-specific system includes scheduling efficiency, load security, loss cost, and path adaptation.

[0019] Preferably, the quantitative modeling of the uncertain scenario is achieved by adding a correlation constraint strength correction term, and the formula for calculating the uncertainty of the scenario after correction is as follows:

[0020] Corrected scenario uncertainty = Initial scenario uncertainty + δ × C;

[0021] Wherein, δ is the correlation correction coefficient, and its value is within the preset correlation correction coefficient threshold range; C is the correlation degree of the core influence relationship corresponding to the current scenario, and the value of C is consistent with the quantification result when constructing the correlation relationship.

[0022] Preferably, the parallel collaborative processing includes an associated guided simulation channel and an associated matching deviation optimization channel;

[0023] The associated guidance simulation channel loads the influence relationship constraints of the current scenario and simultaneously simulates the scheduling effect of deterministic scenarios, multiple uncertain scenarios, and influence relationship triggered scenarios;

[0024] The association matching deviation optimization channel uses logical rationality score as the core reward signal and association deviation as the penalty signal, and optimizes the weight model through reinforcement learning algorithm that incorporates association logic.

[0025] The parallel collaborative processing refers to the synchronous execution process when multiple weight schemes are input into the lightweight digital twin.

[0026] Furthermore, to achieve the above objectives, this application also proposes a pallet intelligent scheduling system, comprising:

[0027] The perception layer is used to collect deterministic data and uncertain signals in the pallet scheduling scenario. The deterministic data includes at least the pallet position and AGV operating status, and the uncertain signals include at least the AGV current abnormal signal and the pallet vibration abnormal signal.

[0028] The association modeling layer, based on the data collected by the perception layer, constructs and quantifies the association relationship between scene factors, constraints, and weight factors, and dynamically updates the association degree parameters.

[0029] The scene modeling layer, based on the data collected by the perception layer and the correlation relationship with the correlation modeling layer, completes the quantitative modeling of uncertain scenes under correlation constraints and generates correlation-constrained scene perturbation factors.

[0030] The dual-path synchronization layer, based on the scene modeling layer's associated constraint-type scene disturbance factor, generates multiple weight schemes through a multi-objective genetic algorithm that introduces an associated constraint mechanism. The weight schemes are then input into a lightweight digital twin to perform parallel collaborative processing of simulation and optimization, and output the optimal weight scheme.

[0031] The constraint verification layer, based on the optimal weight scheme output by the dual-path synchronization layer, verifies whether the corresponding scheduling instructions meet the tray-specific physical constraints, ensuring the feasibility of the scheduling instructions.

[0032] The execution layer, based on the scheduling instructions that have passed the constraint verification layer, drives the AGV and the pallet to complete the scheduling action and implement the optimal scheduling solution;

[0033] The feedback layer, based on the scheduling process data of the execution layer and the verification results of the constraint verification layer, collects simulation deviation, constraint adaptation deviation and correlation matching deviation, constructs a three-dimensional feedback loop and outputs it to the correlation modeling layer and the dual-path synchronization layer to update the correlation relationship and optimize the weight model.

[0034] Preferably, the lightweight digital twin in the dual-path synchronization layer adopts an edge node and lightweight engine architecture. The lightweight twin engine is deployed on the edge node, the model size meets the preset lightweight threshold, and only the pallet, AGV, critical path and other elements that directly support the pallet scheduling execution are loaded. Moreover, its simulation calculation efficiency must meet the performance requirements of real-time scheduling.

[0035] Preferably, in the three-dimensional feedback loop of the feedback layer, the correlation matching deviation is the absolute difference between the quantified value in the correlation relationship and the measured value of the factor correlation degree in the actual scenario. When the difference exceeds the preset correlation matching deviation threshold, the weight model is triggered for emergency optimization.

[0036] Preferably, the association modeling layer is further provided with an abnormal signal and association relationship identification module, which is used to automatically match the abnormal signals collected by the perception layer to the corresponding influence relationships in the association relationship, and provide data input for modeling the uncertain scenario;

[0037] The pallet-specific physical constraints include at least the pallet load-bearing limit, the AGV load limit, and the path width adaptation threshold.

[0038] Furthermore, to achieve the above objectives, this application also proposes a pallet intelligent scheduling device, comprising:

[0039] The data acquisition module is used to collect deterministic data, uncertain signals, and deviation data during the scheduling process in the pallet scheduling scenario.

[0040] The storage module is used to store the relationships, weight factor system parameters, digital twin model data, and scheduling history data;

[0041] The processing module is connected to the data acquisition module and the storage module respectively, and is used to execute the tray intelligent scheduling method, or any possible implementation of the above method, to realize the association modeling, weight scheme generation, simulation and optimization parallel processing and feedback optimization.

[0042] The execution module, connected to the processing module, is used to receive the optimal scheduling instructions output by the processing module and drive the execution mechanism to complete the tray scheduling action.

[0043] Preferably, the processing module includes an edge computing unit and a cloud computing unit;

[0044] The edge computing unit is used to perform digital twin simulation, short-term weight optimization, and real-time scheduling instruction output;

[0045] The cloud computing unit is used to receive data synchronized by the edge computing unit, and to perform long-term weight optimization and update the association relationship.

[0046] The beneficial effects of this invention are:

[0047] Significantly improves dynamic adaptability: Through quantitative modeling of correlation relationships, quantitative modeling of correlation constraints in uncertain scenarios and three-dimensional deviation and dual-cycle optimization, the scheduling scheme can accurately respond to dynamic changes and correlation disturbances in the scenario, and the adaptability is greatly improved.

[0048] Significantly improve scheduling efficiency: Relying on parallel collaborative processing mechanism, lightweight digital twin architecture and hierarchical collaborative system design, it realizes rapid generation of weight schemes, efficient simulation and real-time instruction output, and significantly optimizes scheduling efficiency. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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.

[0050] In the attached diagram:

[0051] Figure 1 This is a flowchart illustrating the intelligent pallet scheduling method of the present invention;

[0052] Figure 2 This is a schematic diagram of the architecture of the intelligent pallet scheduling system in this invention;

[0053] Figure 3 This is a schematic diagram of the intelligent pallet scheduling device in this invention. Detailed Implementation

[0054] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0055] Specific implementation examples are given below.

[0056] Please see Figures 1-3This invention provides a method, system, and device for intelligent tray scheduling. By constructing and quantifying the correlation between scene factors, constraints, and weight factors, and combining the correlation constraint mechanism to optimize the modeling of uncertain scenarios and the generation of multi-objective weight schemes, and relying on parallel collaborative processing and dual-cycle feedback loops, the scheduling scheme is dynamically adapted and efficiently output. At the same time, through a layered system architecture and edge and cloud collaborative computing units, the real-time performance and long-term stability of the system operation are ensured.

[0057] To avoid technical ambiguity, the core technical terms of this invention are clearly defined below:

[0058] Scenario factors: refer to dynamic environmental or equipment status parameters that affect the pallet scheduling process, including but not limited to AGV operating status parameters (current, speed, position), pallet status parameters (load, vibration, wear level), path status parameters (congestion level, flatness), and environmental parameters (temperature, humidity).

[0059] Constraints: These refer to the physical or logical limitations that pallet scheduling must comply with, including but not limited to pallet load limits, AGV load capacities, path width limits, scheduling task priorities, and equipment operation safety thresholds.

[0060] Weighting factor: refers to the quantitative index of the importance of each objective in multi-objective optimization. The multi-objectives of this invention include four core dimensions: scheduling efficiency, load security, loss cost, and path adaptation.

[0061] Association constraint strength: refers to the quantitative value of the degree of mutual influence among scenario factors, constraint conditions, and weight factors, and is used to correct modeling biases in uncertain scenarios.

[0062] Three-dimensional deviations include simulation deviation, constraint adaptation deviation, and correlation matching deviation, which respectively reflect the differences between the digital twin simulation results and the actual scenario, the fit between the scheduling scheme and the constraints, and the matching accuracy of the factor correlation relationship.

[0063] Dual-cycle optimization: refers to the collaborative mechanism of short-term optimization and long-term optimization. Short-term optimization is used to quickly adapt to sudden changes in the scenario, while long-term optimization is used to accumulate data to achieve in-depth model iteration.

[0064] See Figure 2 The diagram shown is an architectural schematic of the intelligent pallet scheduling system provided in this embodiment. The system of this invention adopts a layered architecture design, specifically including a perception layer, an association modeling layer, a scene modeling layer, a dual-path synchronization layer, a constraint verification layer, an execution layer, and a feedback layer. The core functional modules of the device work in coordination with the layered architecture of the system, as specifically implemented as follows:

[0065] The perception layer is used to collect various dynamic data in the pallet scheduling scenario, providing basic input for subsequent modeling and optimization. Its hardware configuration includes, for example, LiDAR, a vision camera, an RFID reader, a current sensor, a vibration sensor, and a position sensor. These sensors are used to collect scenario factor data such as AGV position data, pallet tag information, AGV operating current, pallet vibration signals, and path congestion status. The data collected by the sensors is transmitted to the association modeling layer after noise reduction and format standardization. The data noise reduction uses a median filtering algorithm to remove extreme outliers, and the format standardization is uniformly converted to JSON format to ensure data compatibility. The acquisition frequency is set to no less than 1Hz based on the dynamic characteristics of the scheduling scenario to ensure real-time capture of dynamic changes in the scenario.

[0066] Noise reduction: A median filtering algorithm is used, and the filter window size is dynamically selected based on the scene disturbance frequency—when the scene disturbance frequency is ≤1Hz (such as a typical warehouse static path scenario), a 5th-order window is selected; when the scene disturbance frequency is >1Hz (such as a dynamic path scenario during e-commerce peak periods), a 3rd-order window is selected; the fluctuation error of the filtered data must be ≤10% of the fluctuation of the original data.

[0067] Outlier removal: Outliers are determined based on the equipment's rated parameters. For example, an abnormal AGV current value is more than 20% of its rated operating current (duration ≥ 0.5s), and an abnormal pallet vibration value is an acceleration ≥ 0.5g (g is the acceleration due to gravity). After outlier removal, missing data is supplemented using linear interpolation of data from adjacent time points (the supplemented data percentage is ≤ 5% of the total data).

[0068] Standardized format: All collected data is uniformly converted into JSON format, with fields including "sensor ID, collection timestamp, data type, raw value, preprocessed value, and data status (normal / abnormal)" to ensure compatibility of different types of sensor data.

[0069] Accuracy of deterministic data acquisition: Pallet position acquisition error ≤ ±5cm (LiDAR + RFID fusion positioning), AGV operating status parameter acquisition accuracy ≤ ±2% (rated accuracy of current sensor and speed sensor);

[0070] Uncertain signal thresholds: AGV current abnormality is when it exceeds the rated current by 20% (duration ≥ 0.5s), pallet vibration abnormality is when the acceleration is ≥ 0.5g (g is the acceleration due to gravity), and path congestion abnormality is when the path passage efficiency is ≤ 50% (actual passage time / theoretical passage time). The thresholds are determined by calibration using the equipment's factory parameters and 100 sets of fault scenario data.

[0071] The correlation modeling layer is used to construct and store the correlation relationships between scene factors, constraints, and weight factors, enabling accurate matching and quantification of the correlation degree of abnormal signals. The correlation relationship adopts a two-dimensional matrix structure, with the row dimension representing scene factors and weight factors, the column dimension representing constraints, and the matrix elements representing the quantified correlation degree C between the three (the value ranges from 0 to 1, where C=0 indicates no correlation and C=1 indicates complete correlation).

[0072] The correlation degree is quantified using the mutual information method. The specific steps are as follows: First, the collected scene factors, constraints, and weight factors data are discretized, for example, dividing the data into 10 intervals; then, the information entropy H(X) and H(Y) of individual factors are calculated using the following formula:

[0073] ;

[0074] in Let X be the probability distribution of factor X. To minimize (for example, 1e-8), avoid (0) calculation anomaly;

[0075] Next, calculate the joint entropy of the two factors. The formula is:

[0076] ;

[0077] in for and The joint probability distribution; the final degree of correlation. That is, the degree of correlation is quantified by mutual information value, and then normalized to the range of 0 to 1.

[0078] The correlation modeling layer also includes an abnormal signal and correlation identification module. This module pre-stores feature vectors of the correlations. The feature vector dimensions include the type of abnormal signal (current surge / vibration exceedance) and amplitude (such as current surge value). ), duration (e.g.) The system considers the duration t and associated factors (such as the current tray load value L). When an abnormal signal transmitted from the perception layer is received, it compares the feature vector with a pre-stored feature vector library containing combinations of abnormal signals and corresponding factors. If the cosine similarity between the newly acquired abnormal signal feature vector and the vector in the library is ≥0.9, the match is considered successful and mapped to the corresponding factor combination, providing accurate input for modeling uncertain scenarios. The cosine similarity threshold of 0.9 was determined through testing with 500 sets of abnormal signals to ensure a matching accuracy of ≥95%.

[0079] Matching failure handling process: If the cosine similarity between the feature vector of the abnormal signal and the vector in the library is <0.9, the data preprocessing submodule is used to reduce noise and complete the data (if the data is missing ≤10%), and then a second matching is performed; if the second matching still fails, it is determined to be an "unknown abnormal signal" and will not be included in the uncertain scene modeling for the time being. At the same time, the signal and the corresponding scene data are uploaded to the cloud storage submodule for subsequent feature vector library updates (the feature vector library is retrained every 50 sets of unknown signals).

[0080] The scene modeling layer, serving as the connecting link between the correlation modeling layer and the dual-path synchronization layer, primarily combines the raw data from the perception layer with the quantized correlation relationships from the correlation modeling layer to achieve accurate quantized modeling of uncertain scenarios. Ultimately, it outputs correlation-constrained scene perturbation factors, providing a scenario adaptation basis for the weighting scheme generated by the dual-path synchronization layer. The core logic of this layer is constraint-guided based on correlation relationships. It corrects modeling deviations by adding a correlation constraint strength correction term (see step S2 of the method implementation for specific calculation logic). The generated scene perturbation factor includes key parameters such as uncertain signal amplitude, core correlation degree, and corrected uncertainty, realizing the transformation from raw data and scene feature quantization to adaptation requirement output. This ensures that the weighting scheme generated by the dual-path synchronization layer accurately matches the dynamic changes and correlation constraints of the current scenario.

[0081] The dual-path synchronization layer includes an association-guided simulation channel and a deviation optimization channel, enabling parallel collaborative processing of simulation and optimization to improve scheduling efficiency. The association-guided simulation channel is equipped with a lightweight digital twin, loading only the pallet, AGV, critical path, and other elements that directly support pallet scheduling execution, eliminating redundant environment models and ensuring simulation computation efficiency; the deviation optimization channel incorporates a multi-objective genetic algorithm with an association constraint mechanism to optimize the weight scheme based on deviation data.

[0082] The parallel synchronization mechanism adopts a collaborative approach of hardware triggering and software synchronization signals. Specifically, the data acquisition cycle of the perception layer is used as the synchronization benchmark (exemplarily set to 1Hz). The synchronization signal is generated by the timer of the edge computing unit. After the associated guidance simulation channel receives the synchronization signal, it performs scheduling effect deduction based on the latest scene data and generates simulation results and simulation deviations. The deviation optimization channel synchronously receives the synchronization signal and simulation deviation data, and performs weight scheme optimization in parallel to ensure that the data of the two channels are aligned in real time and avoid timing deviations.

[0083] The constraint verification layer is used to specifically verify the weighting scheme and generated scheduling instructions output by the dual-path synchronization layer to ensure compliance. The verification content includes pallet-specific physical constraints and scheduling logic constraints. The physical constraints include pallet load limits, AGV load limits, and path access restrictions, while the scheduling logic constraints include task priority order and equipment operation safety thresholds.

[0084] The initial threshold for the pallet's specific physical constraints is obtained from the pallet's product manual and factory inspection report, and the pallet's load-bearing limit is preset for example. AGV current safety threshold Path width adaptation threshold Real-time threshold updates are achieved by collecting pallet status data in real time through vibration and load sensors in the sensing layer. If the pallet load L ≥ 0.9 is detected 10 times consecutively, the threshold will be updated. When the vibration value V ≥ 0.8 This triggers the constraint verification layer to be downgraded. (Example: 5% reduction) to adapt to the actual load-bearing capacity after pallet wear; the constraint threshold is stored in the edge storage submodule of storage module 202 and synchronized to the cloud storage submodule for version management.

[0085] The verification process is as follows: First, extract the scheduling parameters corresponding to the weight scheme (such as pallet load value L, AGV running current I, path width W, and task completion time T); then compare the scheduling parameters with the constraint thresholds one by one. If any parameter exceeds the threshold, it is determined that the verification fails and the scheme is fed back to the deviation optimization channel for re-optimization; the scheme that passes the verification generates a formal scheduling instruction and transmits it to the execution layer.

[0086] The execution layer communicates with the AGV on-board controller and pallet drive device to receive scheduling instructions output by the constraint verification layer, drive the execution of pallet transfer tasks, and provide real-time feedback of execution status data to support subsequent deviation calculation and optimization.

[0087] The feedback layer, a crucial component of the system's closed-loop optimization, is used to collect three key data types—simulation deviation, constraint adaptation deviation, and correlation matching deviation—based on the scheduling process status data from the execution layer (such as AGV position deviation and pallet load stability) and the verification results from the constraint verification layer. This data forms the basis for a three-dimensional feedback loop. The three types of deviation reflect the differences between the digital twin simulation and the actual scenario, the fit between the scheduling scheme and the constraints, and the matching accuracy of factor correlation relationships (see step S5 of the method implementation for specific deviation calculation logic). The feedback layer outputs the three-dimensional deviation data in real time to the correlation modeling layer and the dual-path synchronization layer: providing data support to the correlation modeling layer to dynamically update the correlation degree parameters, and feeding back the optimization direction to the dual-path synchronization layer to iterate the weight model. When the correlation matching deviation exceeds a preset threshold, it triggers emergency optimization of the weight model, providing crucial data input for the short-term (edge ​​computing) and long-term (cloud computing) dual-cycle optimization mechanism, ensuring continuous iteration of the system's scheduling performance.

[0088] Threshold calibration process: Based on 800 sets of correlation matching data, the optimized trigger accuracy (avoiding missed triggers) and false trigger rate under different thresholds are calculated, and 0.1 is determined as the optimal threshold, at which the accuracy is ≥95% and the false trigger rate is ≤3%;

[0089] False trigger protection: Emergency optimization is triggered only if the correlation matching deviation is ≥0.1 for three consecutive collections. If the deviation exceeds the standard in a single instance, it will be marked as an anomaly and optimization will not be triggered to avoid invalid adjustments caused by momentary interference.

[0090] See Figure 3 The diagram shown is a schematic of the intelligent pallet scheduling device 200 provided in this embodiment. The device 200 includes a data acquisition module 201, a storage module 202, a processing module 203, and an execution module 204. The functions of each module are as follows:

[0091] The data acquisition module 201 is used to collect deterministic data, uncertain signals, and deviation data during the scheduling process in the pallet scheduling scenario. The data acquisition module 201 includes a sensor group (LiDAR, current sensor, vibration sensor, RFID reader, etc.) and a data preprocessing submodule. The sensor group collects real-time scenario factor data such as AGV operating parameters (current, position), pallet status parameters (load, vibration), and path status parameters (congestion level). The data preprocessing submodule performs noise reduction (median filtering) and format standardization (converting to JSON format) on the collected raw data, removing invalid data. The preprocessed data is synchronously transmitted to the storage module 202 for temporary storage and then pushed to the processing module 203 for subsequent modeling and optimization.

[0092] The installation requirements for the above sensors are as follows:

[0093] LiDAR (for pallet positioning): The selected model must meet the requirements of ranging range of 0.1-50m and positioning accuracy ≤±5cm; it should be installed in an unobstructed area at the front of the AGV, at a height of 1.2m above the ground, parallel to the AGV's direction of travel, and avoid obstacles such as shelves and columns from obstructing the scanning area.

[0094] Current sensor (used for AGV operation status monitoring): The selected model must meet the requirements of a range of 0-50A and a measurement accuracy of ≤±2%; it should be installed in the power supply circuit of the AGV, connected in series with the circuit wire, and the installation position should be far away from strong electromagnetic interference sources (such as motors), with a distance of ≥20cm;

[0095] Vibration sensor (for pallet load status monitoring): The selected sensor should meet the following requirements: measurement range 0-10g, frequency response 10-1000Hz. It should be installed at the center of the bottom of the pallet and fixed with bolts to ensure a rigid connection between the sensor and the pallet and avoid vibration signal attenuation.

[0096] Storage module 202 is used to store association relationships, weight factor system parameters, digital twin model data, and scheduling history data. Storage module 202 is divided into an edge storage submodule and a cloud storage submodule. The edge storage submodule is deployed in the edge computing unit and uses embedded storage media (exemplarily flash memory). It mainly stores real-time scene data collected by the perception layer (storage period not exceeding 1 hour to avoid occupying redundant space), currently effective association relationships, weight model parameters, and recent (exemplarily within 24 hours) scheduling instructions and execution status data.

[0097] The cloud storage submodule is deployed in the cloud computing unit and adopts a distributed storage architecture. It mainly stores long-term accumulated scene data, scheduling task history, updated version of the relationship, version of the weight model parameters, and historical statistical data of 3D deviation. The edge storage submodule periodically synchronizes long-term data to the cloud storage submodule, and the cloud storage submodule sends updated model parameters to the edge storage submodule.

[0098] The processing module 203, connected to both the data acquisition module 201 and the storage module 202, is used to implement relational modeling, weight scheme generation, parallel simulation and optimization processing, and feedback optimization. The processing module 203 is divided into an edge processing submodule and a cloud processing submodule. The edge processing submodule is deployed in the edge computing unit and is responsible for real-time processing tasks, including receiving preprocessed data from the data acquisition module 201, calling the abnormal signal identification logic of the relational modeling layer, matching relational relationships, running the lightweight digital twin of the relational guidance simulation channel, performing scheduling effect simulation, performing short-term weight optimization, and generating real-time scheduling instructions.

[0099] Computing power allocation rules: Tasks with high real-time requirements (response time ≤ 100ms) (digital twin simulation, real-time scheduling instruction output, short-term optimization) are executed by edge computing units. The computing power configuration of edge units must meet the requirement that the processing time of a single task is ≤ 50ms. The thresholds for response time and processing time are determined based on 100 sets of real-time scheduling experiments: when the response time > 100ms, the delay in scheduling instructions will cause the probability of AGV path conflict to increase by 20%; when the processing time of a single task > 50ms, it is impossible to match the real-time running rhythm of the AGV.

[0100] Complex iterative tasks (long-term optimization, relationship update, feature vector library training) are executed by cloud computing units. The cloud computing power needs to support the simultaneous processing of synchronous data from 1000+ edge units, and the time taken for a single round of long-term optimization should be ≤30 minutes. The basis for this cloud computing power requirement is that the amount of synchronous data from 1000+ edge units is approximately 500MB / time, and a single round of long-term optimization needs to complete 500 sets of weight model iterations. Completion within 30 minutes can ensure that the optimization results can be synchronized to the edge units in a timely manner without affecting the next round of scheduling.

[0101] Dynamic adjustment: If the edge unit's computing power load is ≥80% (for 30 consecutive seconds), some non-real-time tasks (such as historical data statistics for short-term optimization) can be temporarily migrated to the cloud and migrated back when the load is ≤60%. The calibration basis for this load threshold is: when the edge unit load is ≥80% (for 30 consecutive seconds), the processing time of a single task will increase to more than 80ms, triggering a real-time risk; when the load is ≤60%, it can be restored to a stable state within 50ms.

[0102] The cloud processing submodule is deployed in the cloud computing unit and is responsible for complex iterative processing tasks. This includes receiving historical data synchronized from the edge processing submodule and updating the correlation degree quantification value; running long-term weight optimization algorithms and iteratively adjusting the parameters of multi-objective genetic algorithms; analyzing historical data of three-dimensional deviations and optimizing the trigger threshold of double-cycle optimization; the edge processing submodule synchronizes data and status with the cloud processing submodule, and the cloud processing submodule sends updated models and parameters to the edge processing submodule to achieve collaborative computing.

[0103] Data synchronization rules: Each time the edge computing unit completes a scheduling task, it synchronizes scene data, deviation data, and weight adjustment records to the cloud (synchronization delay ≤ 1s); each time the cloud computing unit completes a long-term optimization, it synchronizes the updated association relationships and weight model parameters to the edge (incremental synchronization, only transmitting the changed parts).

[0104] Conflict resolution mechanism: If the edge computing unit is performing a short-term optimization and the cloud sends an updated model, the current optimization task is paused, the cloud model is loaded first, and the optimization is restarted based on the new model, thus avoiding scheduling instruction conflicts caused by the parallel operation of the old and new models.

[0105] The execution module 204, connected to the processing module 203, receives the optimal scheduling instructions output by the processing module 203 and drives the execution mechanism to complete the pallet scheduling action. The execution module 204 includes an AGV onboard control submodule, a pallet drive control submodule, and a status feedback submodule. It receives the scheduling instructions output by the constraint verification layer and parses them into AGV path instructions and pallet pick-up / unload instructions. The AGV onboard control submodule drives the AGV to move along the execution path, and the pallet drive control submodule drives the pallet to complete the pick-up and unload actions. The status feedback submodule collects the execution status of the AGV and pallet (such as position and task completion progress) in real time and feeds it back to the processing module 203 and the storage module 202, providing data support for three-dimensional deviation calculation.

[0106] See Figure 1 The diagram shown is a flowchart of the intelligent pallet scheduling method provided in this embodiment. The method includes steps S1 to S5, as detailed below:

[0107] S1: Construct the correlation and quantify the degree of correlation. Based on the historical operation data of tray scheduling, construct the correlation between scene factors, constraints and weight factors of tray scheduling scenario, and quantify the correlation between scene factors and weight factors as the degree of correlation.

[0108] In practical implementation, historical operational data originates from the cloud storage submodule of storage module 202, containing at least 1000 complete sets of scheduling scenario data (including scenario factors, constraints, weight factors, and execution results) to ensure the statistical significance of correlations. The correlations are stored using a two-dimensional matrix. The row dimension covers core scenario factors such as AGV current, pallet load, and path bumpiness, as well as weight factors such as scheduling efficiency and load safety. The column dimension covers constraints such as pallet load limits and AGV load capacities. The matrix elements represent the quantified values ​​of the correlation degree.

[0109] The degree of correlation is calculated using the mutual information method: first, the historical data is discretized into 10 intervals, and the information entropy H(X), H(Y) and joint entropy H(X,Y) of individual factors are calculated. Then, the correlation is calculated using the mutual information value. The correlation degree is quantified and then normalized to the range of 0 to 1. For example, the correlation degree between AGV current mutation and pallet load adaptation safety factor weight is calculated to be 0.72, and the correlation degree between path bumpiness degree and pallet loss avoidance cost weight is 0.68. Both are within the preset correlation degree threshold of 0.2 to 1. This threshold is determined by calibration through historical data. Weak correlation factor combinations below 0.2 will not be included in the corresponding relationship.

[0110] S2: Collect and preprocess scene data, collect deterministic data and uncertain signals of tray scheduling scene to provide basic input for subsequent modeling, perform quantitative modeling of uncertain scene under correlation constraints, generate correlation constraint-type scene disturbance factor based on correlation relationship.

[0111] In practical implementation, deterministic data includes precisely measurable parameters such as pallet position, AGV operating status (speed, current, position), and path width, with a collection frequency of no less than 1Hz. Uncertain signals include abnormal AGV current signals, abnormal pallet vibration signals, and sudden path congestion signals, which are captured in real time by a sensor array. The preprocessing process is completed by the preprocessing submodule of the data acquisition module 201, using median filtering to remove outliers and converting all data into JSON format to ensure data compatibility. At the same time, a preliminary amplitude determination is performed on uncertain signals. For example, when the AGV current fluctuation exceeds 20% of the rated value or the pallet vibration acceleration exceeds 0.5g, it is marked as a valid abnormal signal, and subsequent correlation matching is prioritized.

[0112] In practice, the quantitative modeling of uncertain scenarios is achieved by adding a correction term for the strength of associated constraints. The formula for calculating the uncertainty of the scenario after correction is as follows:

[0113] ;

[0114] in, To correct the uncertainty of the scenario, representing the actual risk level of the current pallet scheduling scenario (such as the comprehensive risk quantification value after the superposition of AGV current fluctuations and path bumps).

[0115] The initial scenario uncertainty is the raw fluctuation value of the sensor data collected in step S2 (such as the standard uncertainty of the AGV current, obtained by calculating the experimental standard deviation through 10 repeated measurements). It is calculated using the GUM method (Guide to Measurement Uncertainty). Taking the AGV current signal as an example, the combined standard uncertainty is determined through Type A evaluation (calculating the experimental standard deviation through 10 repeated measurements) and Type B evaluation (calculating based on the sensor's maximum permissible error according to a uniform distribution). ≈0.123A;

[0116] This is the correlation correction coefficient, corresponding to the correlation density of the scene factors (0.3 for complex scenarios such as e-commerce peak periods, and 0.1 for regular static scenarios), with a value range from 0.1 to 0.5. The denser the correlation of the scene factors, the higher the correlation density. The larger the value, the better in complex e-commerce warehousing scenarios. Set it to 0.3;

[0117] This represents the degree of correlation of the core impact relationship corresponding to the current scenario, and is the quantitative result in the correlation relationship constructed in step S1 (such as the correlation value between "AGV current mutation" and "pallet load safety weight"), which is consistent with the quantitative result in the correlation relationship.

[0118] The threshold calibration process: Based on 500 sets of scheduling data from different scenarios (regular warehousing, complex workshops, peak-hour circulation), the least squares method was used to fit the corrected uncertainty with the actual scenario fluctuations, and the optimal threshold range of 0.1-0.5 was determined, with a goodness of fit ≥92%.

[0119] Scenario-based adjustment rules: For normal scenarios (disturbance factor amplitude ≤ 0.3), δ = 0.1-0.2 is used; for complex scenarios (disturbance factor amplitude 0.3-0.7), δ = 0.2-0.4 is used; for extreme scenarios (disturbance factor amplitude ≥ 0.7), δ = 0.4-0.5 is used. The adjustment trigger condition is that the average of three consecutive collections of the scenario's disturbance factor exceeds the corresponding interval threshold.

[0120] The generated correlation-constrained scenario disturbance factor includes parameters such as the amplitude of the uncertain signal, the degree of correlation, and the corrected uncertainty, which are used for scenario adaptation considerations when generating subsequent weighting schemes.

[0121] The core of this correlation constraint-type scenario disturbance factor is the preset correlation between the scenario factors corresponding to uncertain signals in the current scenario (such as abnormal AGV current, excessive pallet vibration, path congestion, etc.) and the constraints of pallet scheduling (such as pallet load limit, AGV load limit, path width adaptation threshold, etc.) and weight factors (such as load safety weight, loss cost weight, scheduling efficiency weight, etc.).

[0122] S3: Based on the perturbation factor of the associated constraint scenario, combined with the tray-specific multi-objective weight factor system, multiple weight schemes are generated by a multi-objective genetic algorithm that introduces an associated constraint mechanism. The weight schemes must meet the associated constraint screening conditions and the scheduling logic self-consistency verification conditions.

[0123] In practical implementation, the multi-objective weight factor system for pallet-specific systems includes scheduling efficiency, load security, loss cost, and path adaptation.

[0124] Scheduling efficiency: refers to the time efficiency and turnover capacity of pallet scheduling. Specific metrics include the average time from order placement to completion, the number of pallet transfers per unit time, and the on-time delivery rate of urgent orders.

[0125] Load safety: refers to the safety and stability of pallet load and AGV operation. Specific indicators include the ratio of the actual load of the pallet to the load-bearing limit, the difference between the AGV operating current and the safety threshold, and the peak value of pallet vibration acceleration.

[0126] Loss cost: refers to the cost of equipment wear and energy consumption during the scheduling process. Specific indicators include the energy consumption per unit transfer distance of AGV, pallet wear rate (times / unit time), and the proportion of equipment maintenance cycle reduction.

[0127] Path adaptation: refers to the fit between the scheduling path and the scenario. Specific metrics include the ratio of path length to the shortest path, traffic efficiency during congested periods in alleyways, and the adaptability of the path to the turning radius of the equipment.

[0128] The specific measurement criteria of the tray-specific multi-objective weighting factor system are as follows:

[0129] Dispatch efficiency: Order completion time ≤ preset threshold (2 hours for urgent orders, 4 hours for regular orders), pallet transfer times per unit time ≥ 10 times / hour;

[0130] Load safety: Pallet load ≤ 95% of the load limit, AGV operating current ≤ 90% of the safety threshold, vibration acceleration ≤ 0.3g;

[0131] Loss costs: AGV unit transfer energy consumption ≤0.5kWh / km, pallet wear frequency ≤5 times per day, equipment maintenance cycle ≥3 months;

[0132] Path adaptation: Actual path length ≤ 1.2 times the shortest path, traffic efficiency during congested periods ≥ 60%, and path turning radius ≥ 1.1 times the minimum turning radius of the AGV.

[0133] The weighting ratios were determined through 100 pallet transfer experiments, and were exemplarily set to 0.3 (scheduling efficiency), 0.35 (load safety), 0.2 (loss cost), and 0.15 (path adaptation), which can be adjusted according to actual scheduling needs. The multi-objective genetic algorithm is based on the NSGA-II framework, uses real-number encoding, and each chromosome corresponds to a set of weighting schemes. The chromosome length is 4, the value of each gene position ranges from 0 to 1, and the sum of the values ​​of all gene positions is 1.

[0134] The chromosome corresponds to "a set of weight schemes": which perfectly matches the requirement that "each level of confidence corresponds to a unique weight combination". The four gene loci can directly correspond to four core optimization objectives (such as gene 1 = space utilization weight, gene 2 = equipment load balancing weight, gene 3 = access efficiency weight, gene 4 = order response speed weight).

[0135] Value selection rules (0-1 range, sum of 1): Ensure the rationality of the weight scheme (normalization of the total weight, which conforms to the mathematical logic of multi-objective optimization) and avoid weight conflicts or invalid combinations;

[0136] The NSGA-II framework has efficient multi-objective Pareto optimization capabilities, which can find the optimal trade-off between multiple conflicting objectives (such as space utilization and order response speed) and support the generation of "differentiated weight combinations" (high confidence scenarios focus on one type of objective, and low confidence scenarios focus on another type of objective).

[0137] The fitness function includes the objective optimization term and the constraint violation term, and its expression is:

[0138] ;

[0139] in, The fitness function value represents the overall qualification of the weighted scheme (schemes with F ≥ 80 points will proceed to the next screening stage).

[0140] The objective function value for scheduling efficiency is... The objective function value for load safety. The objective function value of loss cost. The objective function value for path adaptation is obtained by normalizing all the above function values ​​to the range of 0 to 1.

[0141] , , , The weight coefficients for the corresponding objectives correspond to the priority of the pallet-specific multi-objective weight factor system (exemplary). , , , );

[0142] For constraint violations, this represents the degree to which the weighting scheme violates physical constraints (such as when the pallet load exceeds the limit). (Value increases), specifically ( This represents the actual value of the factor correlation corresponding to the current weighting scheme. (The threshold for the association constraint is 0.2, for example).

[0143] To constrain the penalty coefficient for violations, its value is set much larger than the sum of the objective function weights; for example, it is set to 10, ensuring that the algorithm prioritizes minimizing constraint violations. In the genetic operation, a tournament selection method (randomly selecting 3 individuals) is used, with a crossover probability of 0.8 and a mutation probability of 0.05. The association constraint screening requires that the degree of association between the proposed solutions be no less than 0.2. The scheduling logic self-consistency verification must satisfy physical constraint self-consistency (pallet load ≤ load limit, AGV operating current ≤ safety threshold, etc.), multi-objective logic self-consistency (ratio of scheduling efficiency improvement rate to load safety coefficient ≤ 2), and association constraint self-consistency. Solutions that pass all three criteria proceed to the subsequent screening stage.

[0144] The general standard for the correlation constraint screening conditions is as follows: the correlation degree threshold (0.2) is determined by calibration using 1000 sets of historical scheduling data. Factor combinations below this threshold can be ignored due to their weak influence. If the correlation between scenario factors is dense (such as during peak e-commerce warehousing periods), it can be dynamically adjusted to 0.3 through the cloud computing unit. The basis for the adjustment is that the average amplitude of the scenario disturbance factor is ≥0.5.

[0145] The criteria for verifying the self-consistency of scheduling logic are as follows: physical constraint self-consistency requires that all scheduling parameters (load, current, etc.) be ≤ 95% of the constraint threshold (with reserved safety redundancy); multi-objective logic self-consistency requires that the variance of the score of each objective be ≤ 10; and correlation constraint self-consistency requires that the correlation matching deviation of the core influence relationship be ≤ 0.1. If all three conditions are met, the algorithm is deemed to have passed; otherwise, it is fed back to the algorithm for re-optimization.

[0146] S4: Lightweight digital twin simulation and optimization parallel collaborative processing. Multiple weight schemes are input into the lightweight digital twin, and digital twin simulation and weight optimization under associated constraints are executed simultaneously in parallel collaborative processing. The comprehensive score and logical rationality score of each scheme are output. The optimal weight scheme is selected by prioritizing logical rationality, constraint adaptation, scenario risk adaptation, and multi-objective balance. The optimal weight scheme is then checked for constraints. Scheduling instructions are generated based on the optimal weight scheme. The scheduling instructions are checked to ensure that they meet the tray-specific physical constraints, thus ensuring the feasibility of the scheduling instructions.

[0147] The overall score is calculated as follows: Scheduling efficiency score × 0.3 + Load security score × 0.35 + Loss cost score × 0.2 + Path adaptation score × 0.15; where each objective score is calculated as a function value × 100 (out of 100). The logical rationality score is determined based on the following rules:

[0148] Scheduling efficiency score ≥ 60 points and load safety score ≥ 70 points (basic logic).

[0149] The difference between the loss cost score and the path adaptation score is ≤30 points (balance logic).

[0150] No target score is less than 40 points (no serious logical defects); if all 3 conditions are met, the logical rationality score is ≥80 points; if 2 conditions are met, the score is 60-79 points; if only 1 condition is met, the score is <60 points.

[0151] In practical implementation, the lightweight digital twin adopts an edge node and lightweight engine architecture, loading only the tray, AGV, critical path, and other elements directly supporting tray scheduling execution. The model size meets the preset lightweight threshold (exemplarily ≤500MB), and the simulation computing efficiency is ≥10 frames / second, meeting real-time scheduling requirements. Parallel collaborative processing includes two channels: the association-guided simulation channel loads the influence relationship constraints of the current scene, synchronously deduces the scheduling effect of deterministic scenarios, multiple uncertain scenarios, and influence relationship-triggered scenarios, and outputs a comprehensive score (covering scheduling efficiency, safety factor, cost control, etc., with a maximum score of 100 points); the association matching deviation optimization channel uses the logical rationality score (maximum score of 100 points) as the core reward signal and the association deviation as the penalty signal, optimizing the weight model through a reinforcement learning algorithm incorporating association logic. The two channels use a 1Hz synchronization signal as a benchmark to ensure real-time data alignment and avoid timing deviations.

[0152] The core framework of the reinforcement learning algorithm is: adopting the DQN (Deep Q Network) architecture, the state space is the associated combination of scene factors, constraints and weight factors, and the action space is the adjustment range of weight factors (±5%-±15%).

[0153] Reward / penalty signal calculation: Reward signal = Logical rationality score × 0.6 - Association deviation × 0.4 (weights are determined based on 100 sets of optimization experiments). When the reward signal is ≥ 80 points, the model stops iterating; when the penalty signal is ≥ 30 points, the weight model is urgently adjusted.

[0154] Iteration frequency: synchronized with digital twin simulation (1Hz) to ensure real-time collaboration between optimization and simulation.

[0155] In practice, the screening process is divided into three levels: the first level sorts the solutions by logical rationality score and selects the top 30%; the second level selects solutions from this group with a constraint fit (the degree of fit between scheduling parameters and constraint thresholds) ≥ 90%; the third level combines the scenario risk level (focusing on safety and losses in complex scenarios, and efficiency and fit in conventional scenarios) and the multi-objective balance coefficient (variance of each objective score ≤ 10) to finally determine the optimal solution. For example, a solution with a logical rationality score of 92, constraint fit of 95%, multi-objective balance coefficient of 8.5, and the highest scenario risk fit score is selected as the optimal weighted solution.

[0156] In practical implementation, pallet-specific physical constraints include pallet load limits, AGV load capacities, and path width adaptation thresholds, which are retrieved from storage module 202 by the constraint verification layer. During verification, core parameters (such as pallet load value, AGV operating current, and planned path width) are extracted from the scheduling instruction and compared one by one with preset thresholds. If any parameter exceeds the threshold, the solution is fed back to the dual-path synchronization layer for re-optimization; if the verification passes, a formal scheduling instruction is generated and transmitted to the execution layer.

[0157] S5: Execute tray scheduling and feedback status data. Execute tray scheduling based on verified scheduling instructions, provide real-time feedback on status data during scheduling execution, calculate three-dimensional deviations and construct feedback loops, collect simulation deviations, constraint adaptation deviations and correlation matching deviations during scheduling execution, construct three-dimensional feedback loops, and continuously optimize the weight factor system in two cycles. Based on the three-dimensional feedback loops, perform short-term and long-term dual-cycle continuous optimization of the weight factor system.

[0158] In practical implementation, the AGV on-board control submodule of execution module 204 parses the path instructions and drives the AGV to move along the planned path. The moving speed is dynamically adjusted according to the scenario (≤1.5m / s for normal paths, ≤0.8m / s for complex paths). The pallet drive control submodule completes the picking-up (hydraulic lifting height adapts to pallet size) and unloading actions according to the instructions. The status feedback submodule collects data such as AGV position deviation, pallet load stability, and task completion progress every 500ms, and transmits them synchronously to processing module 203 and storage module 202 to provide data support for subsequent deviation calculations.

[0159] In practice, the simulated deviation is calculated using Euclidean distance, as shown in the formula:

[0160] ;

[0161] in, To simulate deviation, representing the gap between the simulated effect of the digital twin and the actual scheduling effect; Simulated scheduling efficiency, calculated in the digital twin, is the number of pallet transfers per hour. Actual scheduling efficiency refers to the number of pallet transfers completed by the AGVs per hour on site. Simulated task completion time, which is the single-pallet transfer time calculated in the digital twin; The actual task time is the actual single-pallet transfer time on site.

[0162] Constraint adaptation deviation ;

[0163] in These are the actual values ​​of the constraint parameters for the scheduling scheme. The threshold value is the constraint condition.

[0164] Association Matching Deviation ;

[0165] in, The association matching deviation represents the difference between the preset association relationship and the actual scenario association relationship (the weight model optimization is triggered when Δ3≥0.1).

[0166] The preset degree of correlation is the theoretical correlation value constructed in step S1. The actual degree of correlation refers to the actual correlation values ​​of scene factors collected on-site, and the quantified values ​​in the correlation relationship. Measurement of the correlation between factors and real-world scenarios The absolute value of the difference between them can avoid the interference of positive and negative deviation directions, and focus only on the actual magnitude of the deviation, making the evaluation of the correlation matching accuracy more objective.

[0167] The 3D deviation data is transmitted in real time to the associated modeling layer and the dual-path synchronization layer, forming a feedback closed loop.

[0168] In practice, short-term optimization is performed on a single scheduling task. When any deviation value in the three dimensions exceeds a short-term threshold (e.g., Δ1>5%, Δ2>0, Δ3>0.1), the edge computing unit is triggered to quickly adjust the weight scheme to adapt to sudden changes in the scenario. Long-term optimization is performed on a preset time period (e.g., 24 hours) or a cumulative number of scheduling tasks (e.g., 1000 times). The cloud computing unit iteratively updates the parameters of the correlation relationship and the multi-objective genetic algorithm (e.g., fitness function weights) based on historical deviation data and task completion status to achieve deep model optimization.

[0169] The update of the correlation relationship is triggered when the cumulative number of scheduled tasks reaches 1000, or when the average correlation matching deviation of the three-dimensional deviation exceeds 0.1 for three consecutive days. During the update, the mutual information value of all factor combinations is recalculated, the correlation degree quantification value is updated, weak correlation factor combinations with a correlation degree <0.05 are removed, and newly emerging scenario factor combinations are added. The updated correlation relationship is verified by substituting the new correlation relationship into the uncertain scenario quantification process. If the fit between the corrected uncertainty and the actual scenario fluctuation is improved by ≥10%, the update takes effect. The cloud processing submodule distributes the updated correlation relationship to the edge storage submodule to replace the original correlation relationship. Short-term optimization ensures real-time scenario adaptation, while long-term optimization ensures continuous model iteration. The two form a closed loop to avoid the decay of scheduling effect.

[0170] Triggering conditions: After accumulating 100 scheduling tasks in the short term, the edge computing unit will synchronize the deviation data and weight adjustment records to the cloud, triggering long-term optimization; or when the long-term optimization period (24 hours) expires, an update will be performed regardless of the number of tasks.

[0171] Feedback mechanism: After the cloud completes long-term optimization (updating the relationship and weight model), it is distributed to the edge computing unit through incremental synchronization. The edge unit loads the updated parameters before the next scheduled task starts, ensuring that short-term optimization is performed based on the latest model.

[0172] The overall workflow is as follows:

[0173] Processing module 203 calls historical data from storage module 202 and executes step S1 to construct and quantify the correlation relationship;

[0174] The data acquisition module 201 executes step S2 to acquire and preprocess scene data, and transmits it to the correlation modeling layer and the scene modeling layer; the scene modeling layer completes the quantitative modeling of uncertain scenes based on the preprocessed data in step S2 and the correlation relationship in S1, and generates correlation-constrained scene disturbance factors.

[0175] The dual-path synchronous layer executes step S3 to generate and verify multiple weight schemes, and then executes parallel collaborative processing to output scores through step S4. After determining the optimal weight scheme according to the screening logic, it passes the constraint verification of S4 and sends a scheduling instruction to the execution module 204.

[0176] Execution module 204 executes step S5 to complete the scheduling action and feeds back status data; the feedback layer calculates the three-dimensional deviation and constructs a feedback loop based on the scheduling status data of S5.

[0177] The processing module 203 performs a two-cycle optimization based on the feedback loop execution step S5, updates the weight factor system and correlation relationship, and forms a closed loop.

[0178] Pallet scheduling scenario in an automotive parts production workshop:

[0179] This scenario applies to an automotive engine production workshop (10 production lines, 15 AGVs, and 300 dedicated pallets). The core requirement is to precisely supply cylinder blocks, crankshafts, and other components at a production line cycle of 2 minutes per engine, avoiding downtime due to material shortages. Simultaneously, it avoids AGV overload (load limit 800kg), equipment wear, and path congestion issues. The specific implementation process is as follows:

[0180] Relationship Construction and Quantification: Retrieve 3 months of historical scheduling data from the workshop (8000+ sets), construct the relationship between scenario factors, constraints and weight factors, and quantify the core correlation degree: the correlation degree between production line changeover signal and scheduling efficiency weight is 0.9, the correlation degree between component weight fluctuation and load safety weight is 0.85, and the correlation degree between AGV path intersection and path adaptation weight is 0.8, clarifying the basis for weight priority under different scenarios.

[0181] Quantitative modeling of uncertain scenarios: Deterministic data (production line cycle time 2 minutes / unit, AGV rated current 40A) and uncertain signals (sudden model change on production line 3, requiring 20 additional crankshaft pallets, AGV path congestion rate 40%) are collected. After median filtering for noise reduction and JSON format standardization preprocessing, the data is modeled using formulas. The corrected scenario uncertainty is calculated as follows: U = 0.15A (initial uncertainty) + 0.45 (correlation correction coefficient) × 0.9 (correlation degree between model change and scheduling efficiency) = 0.555A, which accurately quantifies the risk of sudden scenarios.

[0182] Multi-objective weight scheme generation: Based on the perturbation factor of the related constraint scenario, multiple weight schemes are generated by the multi-objective genetic algorithm of the NSGA-II framework. The optimal scheme chromosome that adapts to the current transformation + congestion scenario is [0.45, 0.3, 0.1, 0.15] (scheduling efficiency 0.45, load safety 0.3, loss cost 0.1, path adaptation 0.15), which meets the related constraint screening and logical self-consistency verification conditions of scheduling efficiency score ≥80 and path adaptation score ≥70.

[0183] Parallel processing and scheme selection of digital twins: The weighted scheme is input into a lightweight digital twin (model size 480MB, simulation efficiency 12 frames / second) of 10 production lines, 15 AGVs and cross-path layout. Parallel processing of correlation guidance simulation and deviation optimization is performed simultaneously: the correlation guidance simulation channel deduces that the AGV avoids congested paths and the 20 crankshaft pallets are supplied on time without material shortage; the correlation matching deviation optimization channel outputs a logic rationality score of 93 points. Finally, the scheme is selected and constraint verification is completed (actual pallet load 650kg≤800kg×95%, AGV current 35A≤40A).

[0184] Scheduling and Dual-Cycle Optimization: Scheduling instructions are issued to drive AGVs to complete pallet transfers, with real-time status data feedback; during execution, congestion is added to the AGV path, and the path adaptation deviation increases to 0.12, triggering short-term optimization (the edge computing unit increases the path adaptation weight to 0.25 and replans the path); there are a total of 3 production line changes within 24 hours, and the cloud computing unit updates the correlation (the correlation between change and path adaptation increases to 0.88), optimizes the genetic algorithm fitness function, and improves the efficiency of subsequent change scenario solution generation by 20%.

[0185] In this scenario, the correlation provides a basis for the degree of correlation in the quantification of uncertain scenarios. The quantification results of uncertain scenarios guide the multi-objective genetic algorithm to generate targeted weight schemes. Digital twins verify the adaptability of the schemes. Dual-cycle optimization updates the correlation and algorithm parameters in reverse, forming a complete closed loop of "data-model-execution-feedback". This effectively solves the combined pain points of sudden model changes, path congestion, and load fluctuations in the production workshop.

[0186] This invention focuses on addressing the core pain points of existing pallet scheduling, namely "poor dynamic adaptability and low efficiency," and forms a closed-loop optimization through five key designs: 1. Constructing correlation relationships and quantifying the degree of factor correlation to provide data support for weight allocation and avoid blind decision-making; 2. Introducing a correlation constraint strength correction term to achieve accurate quantification of uncertain scenarios and adapt to sudden situations such as order fluctuations and equipment anomalies; 3. Using a multi-objective genetic algorithm based on the NSGA-II framework to generate dynamic weight schemes and solve the problem of rigidity in multi-objective optimization; 4. Lightweight digital twins and parallel collaborative processing to verify the feasibility of the scheme in advance and improve scheduling reliability; 5. A three-dimensional deviation feedback and dual-cycle optimization mechanism to quickly adapt to scenario changes in the short term and iteratively optimize the model in the long term to ensure continuous system evolution. The above innovative synergistic effects ultimately achieve precise, intelligent, and efficient pallet scheduling.

[0187] Based on the above solutions, a further improved and optimized scheme is designed: a two-layer decoupled optimization architecture. This architecture decouples global optimization from real-time response by separating the time dimension. The upper layer focuses on long-term global path planning, while the lower layer handles short-term dynamic adjustments. The two layers work together to ensure scheduling efficiency and real-time performance. The upper layer (global optimization layer) periodically generates a global path baseline using a rolling time window (15-30 minutes) to solve the global optimization problem of large-scale forklift scheduling (e.g., minimizing total travel distance). The lower layer (real-time response layer) handles sudden scenarios (e.g., lane congestion, high-priority order insertion) with millisecond-level response times, ensuring the dynamic adaptability of scheduling instructions through local path correction. The upper layer updates the global path baseline to the lower layer every 5 minutes; the lower layer feeds back data such as lane occupancy and order execution status to the upper layer in real time for parameter optimization within the upper layer's rolling window.

[0188] The upper layer of rolling time-domain planning, namely upper-level global path optimization, has the core objective of minimizing the total travel distance of forklifts in a dynamic environment while satisfying order priority and resource constraints.

[0189] The time window (T) is set to 20 minutes (adjustable between 15-30 minutes depending on scenario complexity), covering order demand and resource status over a future period. The update cycle (Δt) re-acquires system status (such as lane occupancy, forklift position, and incomplete orders) every 5 minutes, updating the global path baseline based on the latest data to prevent solution failure due to dynamic environmental changes. Simultaneously, a genetic algorithm is used to optimize the process, aiming to minimize the total forklift travel distance while considering order priority (higher priority orders have higher time penalty coefficients for completion).

[0190] Real-number encoding is used, with each chromosome representing a forklift-order-path matching scheme. Gene bits contain information such as forklift ID, order sequence, and path nodes. A tournament selection method is used (5 individuals are randomly selected each time, and the 2 with the highest fitness are retained); partial matching crossover (PMX) is employed with a crossover probability of 0.8; random gene bit swapping generates mutations with a mutation probability of 0.05.

[0191] By controlling the population size (50-100 individuals) and employing heuristic pruning (such as path conflict pre-detection), the computational complexity is kept at O(n log n) (where n is the number of orders), ensuring that the computation is completed within a 5-minute update cycle. Big O notation is the most crucial tool in algorithm complexity analysis; its role is to ignore secondary factors and focus on the core trends of algorithm efficiency.

[0192] The underlying layer employs real-time heuristic adjustments with dynamic response. Its core objective is to rapidly respond to unexpected scenarios such as lane congestion and order insertions, ensuring real-time optimization of local paths based on a global path baseline. Triggering conditions include: lane occupancy ≥85% (capacity safety threshold), high-priority order insertion (e.g., urgent production material needs), forklift malfunction, or temporary path blockage. Response time from event detection to path correction completion is ≤500ms, achieving low-latency processing through edge computing units (e.g., embedded chips deployed in the forklift controller).

[0193] Dynamic weighting function: w(t) = α(t)·distance cost + β(t)·capacity penalty

[0194] Distance cost is the Euclidean distance between the current path and the target point; capacity penalty is the additional cost after the lane occupancy rate exceeds the threshold (the penalty coefficient increases by 0.02 for every 1% increase in occupancy rate); adaptive coefficient: α(t) + β(t) = 1, dynamically adjusted according to the load, where α(t) is the distance cost weighting coefficient and β(t) is the capacity penalty weighting coefficient.

[0195] Low load (lane occupancy <60%): α=0.7, β=0.3 (prioritize distance optimization);

[0196] High load (cargo lane occupancy rate 60%-85%): α=0.5, β=0.5 (balancing distance and capacity);

[0197] Overload warning (freight lane occupancy rate ≥ 85%): α = 0.3, β = 0.7 (prioritize avoiding congestion).

[0198] The comprehensive cost of candidate paths is calculated based on the weighting function, and the path with the lowest cost is selected as the correction result, while ensuring that the deviation rate from the global path benchmark is less than 15% (to avoid local adjustments from destroying the global optimization objective).

[0199] Reserve 10-15% of lane capacity for high-priority orders (e.g., if a lane has a total capacity of 10 pallets, reserve 1-1.5 empty slots). These reserved resources are only available to high-priority orders. Within a rolling time window, a lightweight simulation model (sharing scenario data with the digital twin) predicts lane occupancy trends for the next 20 minutes. If the predicted demand for high-priority orders exceeds the reserved capacity for a certain period, a global path adjustment at the upper layer is triggered in advance (e.g., optimizing order allocation for other lanes). When the reserved capacity is insufficient, a "priority preemption" mechanism is used: low-priority orders automatically relinquish some resources (delayed processing or reassigned to other lanes) to ensure the resource needs of high-priority orders are met.

[0200] In an automotive parts warehousing scenario (15 AGVs, 300 pallets), the on-time delivery rate of high-priority orders increased to 99.5%, and equipment utilization improved by 18%. In a simulated scenario involving 100+ forklifts, 500+ aisles, and 1000+ orders, digital twin verification showed: global path planning time: ≤4 minutes / cycle (meeting the 5-minute update requirement); real-time adjustment response time: average 320ms (meeting the <500ms requirement); travel distance: reduced by 21% compared to traditional static scheduling algorithms, and aisle congestion events decreased by 95%.

[0201] This two-layer decoupled optimization architecture can be integrated into the dual-path synchronization layer of the original intelligent pallet scheduling system: the upper-layer rolling time-domain planning serves as an extension module for the associated guidance simulation channel, optimizing the global weight scheme through a genetic algorithm; the lower-layer real-time heuristic adjustment serves as the core logic for the associated matching deviation optimization channel, achieving local deviation correction based on a dynamic weight function; the capacity reservation mechanism, combined with a three-dimensional feedback loop, continuously optimizes the reservation ratio and simulation model parameters based on the lane occupancy deviation data collected by the feedback layer. Through the above design, a deep decoupling of global optimization and real-time response is achieved, balancing efficiency and dynamic adaptability in large-scale forklift scheduling scenarios.

[0202] In the description of this invention, it should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired results of the technical solutions provided in this disclosure can be achieved, and no limitation is imposed herein.

[0203] The above description is merely a preferred embodiment of the present invention and does not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for intelligent pallet scheduling, characterized in that, Includes the following steps: S1: Based on the historical operation data of pallet scheduling, construct the correlation between scenario factors, constraints and weight factors of pallet scheduling scenarios, and quantify the correlation between scenario factors and weight factors as the degree of correlation. S2: Collect deterministic data and uncertain signals of the pallet scheduling scenario. After preprocessing, based on the correlation relationship, complete the quantitative modeling of the uncertain scenario under the correlation constraint by adding a correlation constraint strength correction term, and generate the correlation constraint type scenario disturbance factor. S3: Based on the perturbation factor of the associated constraint scenario, combined with the pallet-specific multi-objective weight factor system, multiple weight schemes are generated by the multi-objective genetic algorithm that introduces the associated constraint mechanism. The weight schemes must meet the associated constraint screening conditions and the scheduling logic self-consistency verification conditions. Among them, the multi-objective weight factor system of the tray-specific multi-objective system includes at least scheduling efficiency, load security, loss cost, and path adaptation; S4: Input multiple weight schemes into the lightweight digital twin, simultaneously perform parallel collaborative processing of digital twin simulation and weight optimization under associated constraints, output the comprehensive score and logical rationality score of each scheme, and select the optimal weight scheme according to the selection logic of logical rationality priority, constraint adaptation priority, scenario risk adaptation, and multi-objective balance. S5: Based on the optimal weight scheme, perform tray scheduling, collect simulation deviation, constraint adaptation deviation and correlation matching deviation during the scheduling process, construct a three-dimensional feedback loop, and continuously optimize the weight factor system in both short and long term cycles.

2. The method for intelligent pallet scheduling according to claim 1, characterized in that, In step S1, the degree of correlation is within the preset correlation threshold range and is determined by calibration using historical pallet scheduling data. The degree of correlation is the quantitative result when constructing the correlation relationship, and the correlation relationship includes at least the influence relationship between AGV current mutation and pallet load adaptation safety factor weight, and the influence relationship between path bumpiness degree and pallet loss avoidance cost weight.

3. The method for intelligent pallet scheduling according to claim 1, characterized in that, It also includes using a two-layer decoupling optimization architecture to decouple global scheduling optimization from real-time scene response; The dual-layer decoupled optimization architecture includes an upper global optimization layer and a lower real-time response layer. The upper layer adopts a rolling time-domain planning mechanism, which periodically updates the global scheduling path benchmark based on a preset time-domain window, and optimizes the global scheduling strategy through a multi-objective optimization algorithm. The lower layer focuses on dynamic adjustment and local path correction in emergency scenarios. It completes the dynamic correction of scheduling strategy through a scenario-adaptive dynamic weight function, and the corrected strategy is adapted to the global scheduling benchmark. The upper and lower layers achieve coordinated scheduling through bidirectional synchronization of state data and baseline strategies.

4. The method for intelligent pallet scheduling according to claim 1, characterized in that, In step S2, the quantitative modeling of uncertain scenarios is achieved by adding a correction term for the strength of associated constraints, and the formula for calculating the uncertainty of the scenario after correction is as follows: Corrected scenario uncertainty = Initial scenario uncertainty + δ × C; Wherein, δ is the correlation correction coefficient, and its value is within the preset correlation correction coefficient threshold range; C is the correlation degree of the core influence relationship corresponding to the current scenario, and the value of C is consistent with the quantification result when constructing the correlation relationship.

5. A pallet intelligent scheduling system, characterized in that, include: The perception layer is used to collect deterministic data and uncertain signals in the pallet scheduling scenario. The deterministic data includes at least the pallet position and AGV operating status, and the uncertain signals include at least the AGV current abnormal signal and the pallet vibration abnormal signal. The association modeling layer, based on the data collected by the perception layer, constructs and quantifies the association relationship between scene factors, constraints, and weight factors, and dynamically updates the association degree parameters. The scene modeling layer, based on the data collected by the perception layer and the correlation relationship with the correlation modeling layer, completes the quantitative modeling of uncertain scenes under correlation constraints and generates correlation-constrained scene perturbation factors. The dual-path synchronization layer, based on the scene modeling layer's associated constraint-type scene disturbance factor, generates multiple weight schemes through a multi-objective genetic algorithm that introduces an associated constraint mechanism. The weight schemes are then input into a lightweight digital twin to perform parallel collaborative processing of simulation and optimization, outputting the optimal weight scheme. The constraint verification layer, based on the optimal weight scheme output by the dual-path synchronization layer, verifies whether the corresponding scheduling instructions meet the tray-specific physical constraints, ensuring the feasibility of the scheduling instructions. The execution layer, based on the scheduling instructions that have passed the constraint verification layer, drives the AGV and the pallet to complete the scheduling action and implement the optimal scheduling solution; The feedback layer, based on the scheduling process data of the execution layer and the verification results of the constraint verification layer, collects simulation deviation, constraint adaptation deviation and correlation matching deviation, constructs a three-dimensional feedback loop and outputs it to the correlation modeling layer and the dual-path synchronization layer to update the correlation relationship and optimize the weight model.

6. The intelligent pallet scheduling system according to claim 5, characterized in that, The association modeling layer is also equipped with an abnormal signal and association relationship identification module, which is used to automatically match the abnormal signals collected by the perception layer to the corresponding influence relationship in the association relationship, and provide data input for the modeling of the uncertain scenario; The pallet-specific physical constraints include at least the pallet load-bearing limit, the AGV load limit, and the path width adaptation threshold.

7. The intelligent pallet scheduling system according to claim 5, characterized in that, The lightweight digital twin in the dual-path synchronization layer adopts an edge node and lightweight engine architecture. The lightweight twin engine is deployed on the edge node, the model size meets the preset lightweight threshold, and only the pallet, AGV, critical path and other elements that directly support the pallet scheduling execution are loaded. Moreover, its simulation calculation efficiency must meet the performance requirements of real-time scheduling.

8. The intelligent pallet scheduling system according to claim 5, characterized in that, The system also includes a high-priority order guarantee module, which reserves dedicated scheduling resources for high-priority orders. When the reserved resources are insufficient, a priority preemption mechanism is triggered, which adjusts the scheduling sequence and resource allocation of low-priority orders to ensure the execution needs of high-priority orders.

9. A pallet intelligent scheduling device, characterized in that, include: The data acquisition module is used to collect deterministic data, uncertain signals, and deviation data during the scheduling process in the pallet scheduling scenario. The storage module is used to store the relationships, weight factor system parameters, digital twin model data, and scheduling history data; The processing module is connected to the data acquisition module and the storage module respectively, and is used to execute the tray intelligent scheduling method according to any one of claims 1 to 4, and realize the association relationship modeling, weight scheme generation, simulation and optimization parallel processing and feedback optimization. The execution module, connected to the processing module, is used to receive the optimal scheduling instructions output by the processing module and drive the execution mechanism to complete the tray scheduling action.

10. The intelligent pallet scheduling device according to claim 9, characterized in that, The processing module includes an edge computing unit and a cloud computing unit; The edge computing unit is used to perform digital twin simulation, short-term weight optimization and real-time scheduling instruction output, and also carries out the task of dynamic adjustment of sudden scenarios in the lower real-time response layer. The cloud computing unit is used to receive data synchronized by the edge computing unit, perform long-term weight optimization and update the correlation relationship, and at the same time carry out the global scheduling strategy generation and optimization tasks of the upper global optimization layer.

Images