WMS warehouse task dynamic scheduling method and system based on multi-objective optimization

Abstract

The invention discloses a WMS warehouse task dynamic scheduling method and system based on multi-objective optimization, and the method comprises the steps: collecting scheduling parameters, such as task priority, equipment load and goods location distance, through a RidgeOS intelligent warehouse scheduling platform, and inputting a dynamic scheduling task uncertainty perception model to generate an uncertainty feature vector; constructing a scheduling decision space by adopting an elastic time slice-reinforcement learning hybrid algorithm, dividing an elastic time slice interval, and iteratively updating a scheduling strategy in the interval; sorting the to-be-executed tasks of the warehouse according to the updated strategy, wherein the uncertainty feature vectors are used as constraints during sorting; a scheduling instruction is sent to the equipment through the platform, and meanwhile parameter change data in the equipment execution process is collected and fed back to the sensing model. The system comprises six units which are connected. According to the method, the task execution uncertainty can be accurately perceived, the scheduling strategy flexibility and iteration efficiency are improved, and the intelligent storage multi-target optimization scheduling requirement is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of warehouse task dynamic scheduling, and particularly relates to a WMS warehouse task dynamic scheduling method and system based on multi-objective optimization. BACKGROUND

[0002] In the modern warehouse management scenario, the WMS (full name: Warehouse Management System) warehouse task scheduling needs to cope with complex situations such as task quantity fluctuation, device state change, and difference in storage location distribution. The traditional scheduling method relies on fixed rules and is difficult to adapt to the dynamically changing warehouse environment. With the development of intelligent warehousing technology, the RidgeOS intelligent warehousing scheduling platform is gradually applied to parameter collection and instruction issuing. However, how to combine task priority, device load, storage location distance and other multi-dimensional parameters to achieve accurate scheduling has become a key to improving warehouse efficiency. At the same time, there are uncertain factors such as device failure and task delay in the task execution process, which need to be perceived by a special model. The fusion of elastic time slice division and reinforcement learning algorithm can further optimize the scheduling strategy iteration efficiency. Therefore, the research and development of the WMS warehouse task dynamic scheduling method and system based on multi-objective optimization has become an important direction to meet the efficient operation needs of modern warehouses.

[0003] The existing technology has two significant shortcomings in WMS warehouse task dynamic scheduling: on the one hand, the processing capability of task execution uncertainty is insufficient. Most scheduling technologies do not build a special uncertainty perception model, but rely on a single parameter to determine the task state, which cannot generate comprehensive uncertainty features by combining task priority, device load, storage location distance and other multi-parameters, resulting in scheduling strategies being easily affected by unexpected factors and being difficult to adapt to dynamic changes in complex warehouse environments. On the other hand, the flexibility and iteration optimization capability of the scheduling strategy is lacking. Traditional technologies do not use a hybrid algorithm of elastic time slice-reinforcement learning. The time slice division is fixed and the strategy update relies on manual adjustment, which cannot dynamically adjust the time slice interval according to real-time collected parameter data, and is also difficult to achieve autonomous optimization of the scheduling strategy through algorithm iteration, resulting in low scheduling efficiency and being unable to meet the multi-objective optimization of warehouse scheduling needs. SUMMARY

[0004] In order to overcome the shortcomings and deficiencies of the existing technology, the present application provides a WMS warehouse task dynamic scheduling method and system based on multi-objective optimization.

[0005] The technical scheme adopted by the present application is a multi-objective optimization-based WMS warehouse task dynamic scheduling method, comprising the following steps: step S1, collecting various parameters in the WMS warehouse task dynamic scheduling process through the RidgeOS (full name RidgeIntelligent Warehouse Operating System) intelligent warehouse scheduling platform, the parameters including task priority coefficient, device running load value, storage location distance weight, task execution time length threshold and device failure probability factor; step S2, inputting the collected various parameters into a dynamic scheduling task uncertainty perception model, identifying the task execution uncertainty characteristics corresponding to each parameter through the model, generating an uncertainty feature vector, and the feature vector dimension is consistent with the number of parameter types collected; step S3, based on the uncertainty feature vector, using an elastic time slice-reinforcement learning hybrid algorithm to construct a task scheduling decision space, dividing different elastic time slice intervals in the decision space, and each interval corresponds to a different task scheduling priority range; step S4, in each elastic time slice interval, the task scheduling strategy is iteratively updated through the elastic time slice-reinforcement learning hybrid algorithm, and the task priority coefficient, the device running load value and the storage location distance weight are used as the calibration input variables of the algorithm training in the updating process; step S5, according to the scheduling strategy updated by iteration, the tasks to be executed in the WMS warehouse are sequenced and sorted, and the uncertainty feature vector output by the dynamic scheduling task uncertainty perception model is called as a sorting constraint condition in the sorting process; step S6, sending the sorted task scheduling instruction to the corresponding warehouse device through the RidgeOS intelligent warehouse scheduling platform, and simultaneously collecting the parameter change data in the device execution scheduling instruction process in real time, and feeding back to the dynamic scheduling task uncertainty perception model.

[0006] Further, when the dynamic scheduling task uncertainty perception model generates the uncertainty feature vector in step S2, the task execution uncertainty index is calculated using the following model formula: U i = α·P i + β·L j + γ·D k - δ·T m , wherein U i represents the execution uncertainty index of the i-th task, α, β, γ, δ are weight coefficients of the task priority coefficient, the device running load value, the storage location distance weight and the task execution time length threshold respectively, P i represents the priority coefficient of the i-th task, L j represents the running load value of the j-th device, D k represents the distance weight of the k-th storage location, and T m represents the execution time length threshold of the m-th task.

[0007] Further, the elastic time slice-reinforcement learning hybrid algorithm divides the elastic time slice interval in step S3, and adopts the following model formula to determine the time slice length: Wherein, T s represents the elastic time slice length, μ, v, λ, σ are respectively the adjustment coefficients of the maximum value of task priority, the average load value of equipment, the average distance weight of warehouse, and the average fault probability factor of equipment, P max represents the maximum value of priority coefficient in all tasks, L avg represents the average running load value of warehouse equipment, D avg represents the average distance weight of warehouse, F avg represents the average fault probability factor of warehouse equipment.

[0008] Further, the elastic time slice-reinforcement learning hybrid algorithm iteratively updates the scheduling strategy in step S4, and adopts the following model formula to calculate the strategy update reward value: R t = η·(P i -P avg )- θ·(L j -L avg )+ ζ·(D std -D k ), wherein R t represents the strategy update reward value of the tthiteration, η, θ, ζ are respectively the reward coefficients of task priority deviation, equipment load deviation, and distance weight deviation of warehouse, P avg represents the average priority coefficient of all tasks, D std represents the standard deviation of distance weight of warehouse.

[0009] Further, the dynamic scheduling task uncertainty perception model adopts the following model formula to calculate the constraint strength when it is used as a sorting constraint condition in step S5: Wherein, C s represents the constraint strength value, ε, are respectively the constraint coefficients of uncertainty index and time slice length deviation, T avg represents the average length of all elastic time slice intervals.

[0010] Further, the RidgeOS intelligent warehouse scheduling platform adopts the following model formula to calculate the data feedback weight W f = κ·ΔL j + τ·ΔD k + ω·ΔF j when it feeds back the parameter change data in step S6, wherein W frepresents the data response weight, and κ, τ, ω are weight coefficients of the device load variation, the distance weight variation of the storage location, and the fault probability factor variation of the device, respectively, ΔL j represents the load variation of the jth device, ΔD k represents the distance weight variation of the kth storage location, ΔF j represents the fault probability factor variation of the jth device.

[0011] Further, step S3 comprises the following sub-steps: step S31, obtaining the uncertainty feature vector output from the dynamic scheduling task uncertainty perception model, extracting the parameter values corresponding to each dimension in the feature vector, arranging the values in ascending order to form a parameter sequence; step S32, based on the parameter sequence, setting an initial threshold for the flexible time slice interval division, the initial threshold being determined according to the five-equal-division result of the difference between the maximum value and the minimum value of the parameter sequence, each equal-division interval corresponding to an initial time slice interval; step S33, inputting the initial time slice interval into the flexible time slice-reinforcement learning hybrid algorithm, the algorithm adjusting the initial interval boundary according to the distribution of the task priority coefficient, expanding or reducing the interval range, so that the variance of the task priority coefficient in each interval is within the preset range; step S34, recording the boundary values of the adjusted flexible time slice interval and the corresponding task quantity, generating an interval-task quantity mapping table as basic data for subsequent scheduling strategy iteration and update.

[0012] Further, step S4 comprises the following sub-steps: step S41, calling historical scheduling data from the RidgeOS intelligent warehouse scheduling platform, and screening out historical data groups similar to the current flexible time slice interval parameters, the similarity being determined according to whether the parameter value deviation is less than a preset deviation threshold; step S42, inputting the screened historical data groups into the training module of the flexible time slice-reinforcement learning hybrid algorithm, taking the task completion order and device running state in the historical data as the training samples of the algorithm, and initializing the reward function parameters of the algorithm; step S43, selecting part of the to-be-executed tasks in the current flexible time slice interval as test samples, inputting them into the trained algorithm model, outputting a preliminary scheduling strategy, and calculating the deviation value of the preliminary strategy from the historical optimal strategy; step S44, adjusting the learning rate parameter of the algorithm according to the deviation value, inputting the test samples again for strategy iteration, and repeating the iteration process until the deviation value is less than a preset upper deviation limit, to determine the final scheduling strategy update result.

[0013] Further, step S5 includes the following sub-steps: step S51, obtaining the scheduling strategy update result output by step S4, extracting the task priority sorting rule and device allocation rule included in the strategy, and arranging to form a sorting rule list; step S52, obtaining the uncertainty feature vector of all tasks to be executed from the dynamic scheduling task uncertainty perception model, and calculating the similarity of each task vector and the standard vector in the sorting rule list; step S53, according to the similarity, the tasks are preliminarily sorted, the higher the similarity, the higher the task is sorted, and the real-time load data of the devices in the WMS warehouse are called to perform device load adaptability verification on the preliminary sorting result; step S54, adjusting the position of the task that does not meet the device load requirement after verification, and referring to the location distance weight parameter in the adjustment process to ensure that the adjusted task order meets the device load limit and the total sum of the location distance weight is minimum.

[0014] The WMS warehouse task dynamic scheduling system based on multi-objective optimization comprises: a warehouse parameter acquisition and instruction sending unit, which is connected with each device in the WMS warehouse to acquire various parameters of the WMS warehouse task dynamic scheduling and send task scheduling instructions; a scheduling task feature recognition and vector generation unit, which is connected with the warehouse parameter acquisition and instruction sending unit through a data bus, receives the acquired parameters and generates an uncertainty feature vector; a scheduling decision space construction and interval division unit, which is connected with the scheduling task feature recognition and vector generation unit through a communication interface, constructs a decision space based on the feature vector and divides time slice intervals; a scheduling strategy iteration update and optimization unit, which is connected with the scheduling decision space construction and interval division unit through a data link, performs iteration update of the scheduling strategy; a WMS warehouse task to be executed scheduling order sorting and constraint verification unit, which is connected with the scheduling strategy iteration update and optimization unit and the scheduling task feature recognition and vector generation unit through signal lines, and completes task sorting and constraint verification; a scheduling data real-time feedback and model parameter adjustment unit, which is connected with the warehouse parameter acquisition and instruction sending unit through a communication protocol at one end and with the scheduling task feature recognition and vector generation unit through a data interface at the other end, and performs data feedback and model adjustment.

[0015] Beneficial effects: The application proposes a WMS warehouse task dynamic scheduling method and system based on multi-objective optimization. By constructing a dynamic scheduling task uncertainty perception model, multi-dimensional parameters such as task priority, device running load, storage location distance, task execution time threshold, and device failure probability factor can be comprehensively considered to generate a comprehensive uncertainty feature vector. Instead of relying on a single parameter to determine the task state, the uncertainty factors in task execution can be accurately perceived, so that the scheduling strategy can adapt to the dynamic changes of the complex warehouse environment, avoid being affected by sudden factors, and greatly improve the stability of scheduling. Through the elastic time slice-reinforcement learning hybrid algorithm, the elastic time slice interval can be dynamically divided, the interval range can be adjusted based on real-time collected parameter data, and the scheduling strategy can be updated autonomously through algorithm iteration without manual adjustment. The problems of lack of scheduling flexibility and iterative optimization ability in traditional technologies are solved, and the scheduling efficiency is significantly improved. In addition, through the RidgeOS intelligent warehouse scheduling platform, real-time parameter collection, instruction sending, and data feedback are realized to form a closed loop of "collection-analysis-scheduling-feedback". It can not only accurately integrate multi-dimensional parameters to realize multi-objective optimization scheduling, but also adjust the model and strategy in real time according to the device execution data to further improve the warehouse task execution efficiency and device utilization, meeting the needs of modern intelligent warehouse efficient operation. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The method steps flowchart of the application;

[0017] Figure 2 The system unit composition diagram of the application. DETAILED DESCRIPTION

[0018] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict, and the present application will be further described in detail below in combination with the drawings and specific embodiments.

[0019] As Figure 1 shown, the WMS warehouse task dynamic scheduling method based on multi-objective optimization includes the following steps:

[0020] Step S1, collect various parameters in the WMS warehouse task dynamic scheduling process through the RidgeOS intelligent warehouse scheduling platform, the parameters including task priority coefficient, device running load value, storage location distance weight, task execution time threshold, and device failure probability factor;

[0021] Specifically, step S1 is the basic data collection link of the whole WMS warehouse task dynamic scheduling process, providing comprehensive and accurate original parameter support for subsequent model analysis and algorithm operation. This step realizes data collection through the RidgeOS intelligent warehouse scheduling platform. The collected parameters are all key indicators that affect the efficiency and accuracy of warehouse task scheduling, covering multiple dimensions such as task attributes, device status, storage location characteristics, and risk factors. The completeness and real-time of these parameters directly determine the recognition accuracy of the subsequent uncertainty perception model and the optimization effect of the elastic time slice-reinforcement learning hybrid algorithm, and are the prerequisite for ensuring the scientific and reasonable scheduling strategy. In the specific implementation process, the RidgeOS intelligent warehouse scheduling platform establishes real-time data connection with the WMS system, warehouse equipment sensors, and storage location management module, collects task priority coefficient, device running load value, storage location distance weight, task execution time threshold, and device failure probability factor. Among them, the task priority coefficient is set to an integer from 1 to 5 according to the task urgency, with the emergency outbound task coefficient set to 5 and the regular replenishment task coefficient set to 2; the device running load value is calculated by the ratio of the number of tasks processed by the device in real time to the maximum processing capacity, with the value ranging from 0 to 1, and the forklift load value reaching 0.8 is determined as high load; the storage location distance weight is set according to the straight-line distance between the storage location and the outbound port, with the weight of the storage location within 10 meters set to 0.3 and the weight of the storage location over 50 meters set to 0.9; the task execution time threshold is set according to the task type, with the sorting task threshold set to 30 minutes and the inventory task threshold set to 60 minutes; the device failure probability factor is calculated by referring to the device failure frequency in the past 30 days, with the factor of the device with more than 3 failures per month set to 0.7 and the factor of the fault-free device set to 0.1, and the collection frequency is set to every 10 seconds to ensure real-time parameter update.

[0022] Step S2, input each parameter collected into the dynamic scheduling task uncertainty perception model, identify the task execution uncertainty characteristics corresponding to each parameter through the model, generate an uncertainty feature vector, the dimension of the feature vector is consistent with the number of parameter types collected;

[0023] Specifically, step S2 performs deep analysis on the collected multi-dimensional parameters by the dynamic scheduling task uncertainty perception model, converts the dispersed parameter information into a unified uncertainty feature vector, and realizes accurate identification of potential risks and fluctuations in the task execution process. The core value of this step lies in breaking the limitations of single parameter judgment in traditional scheduling, considering the influence of multiple factors such as tasks, equipment, and storage locations on scheduling, providing standardized features for subsequent construction of scheduling decision space, and avoiding scheduling strategy deviation caused by isolated parameter analysis. In specific implementation, first, the task priority coefficient, equipment running load value, storage location distance weight, task execution time threshold, and equipment failure probability factor collected in step S1 are arranged in a parameter matrix in a predetermined order. The number of rows of the matrix is consistent with the number of tasks collected in real time, and the number of columns is the same as the number of parameter types. Then, the parameter matrix is input into the dynamic scheduling task uncertainty perception model. The model maps the features of each parameter through a multi-layer perception network. The number of network nodes in each layer is determined according to the parameter dimension. The number of input layer nodes is set to 5 (corresponding to 5 types of parameters), the number of hidden layer nodes is set to 12, and the number of output layer nodes is set to 1 (corresponding to a single task uncertainty feature value). The model performs feature weighting processing on each type of parameter. The weight proportion of the task priority coefficient is set to 30%, the weight proportion of the equipment running load value is set to 25%, the weight proportion of the storage location distance weight is set to 20%, the weight proportion of the task execution time threshold is set to 15%, and the weight proportion of the equipment failure probability factor is set to 10%. Finally, an uncertainty feature vector corresponding to the number of tasks is generated. The value range of each element in the vector is 0 to 1. The larger the value, the higher the uncertainty of the task execution.

[0024] Step S3, based on the uncertainty feature vector, a flexible time slice-reinforcement learning hybrid algorithm is used to construct a task scheduling decision space. Different flexible time slice intervals are divided in the decision space, and each interval corresponds to a different task scheduling priority range.

[0025] Specifically, step S3 builds a task scheduling decision space through the elastic time slice-reinforcement learning hybrid algorithm based on the uncertainty feature vector generated in step S2, and divides different elastic time slice intervals to realize the classified management and precise optimization of the scheduling tasks. This step decomposes the complex multi-task scheduling problem into sub-problems in multiple intervals, reduces the difficulty of algorithm optimization, and through the design of elastic time slices, adapts to tasks with different uncertainty levels, provides a clear optimization range for subsequent policy iteration updates, and improves the pertinence and efficiency of scheduling decisions. In the specific implementation process, first, according to the value range (0 to 1) of the elements in the uncertainty feature vector, the scheduling decision space is divided into three primary intervals of low uncertainty (0 to 0.3), medium uncertainty (0.3 to 0.7), and high uncertainty (0.7 to 1). Then, in each primary interval, the elastic time slice-reinforcement learning hybrid algorithm is used to further divide the elastic time slice intervals, and the time slice length is dynamically adjusted according to the number of tasks and the uncertainty feature value in the interval. In the low uncertainty interval, when the number of tasks is less than 20, the time slice length is set to 5 minutes; when the number of tasks is between 20 and 50, the time slice length is set to 8 minutes; and when the number of tasks exceeds 50, the time slice length is set to 12 minutes. In the medium uncertainty interval, when the number of tasks is less than 15, the time slice length is set to 6 minutes; when the number of tasks is between 15 and 40, the time slice length is set to 10 minutes; and when the number of tasks exceeds 40, the time slice length is set to 15 minutes. In the high uncertainty interval, when the number of tasks is less than 10, the time slice length is set to 8 minutes; when the number of tasks is between 10 and 30, the time slice length is set to 12 minutes; and when the number of tasks exceeds 30, the time slice length is set to 18 minutes. Each elastic time slice interval corresponds to a specific task scheduling priority range, and the time slice in the low uncertainty interval corresponds to the priority range 1 to 2, the time slice in the medium uncertainty interval corresponds to the priority range 3 to 4, and the time slice in the high uncertainty interval corresponds to the priority range 5.

[0026] Step S4, in each elastic time slice interval, the task scheduling strategy is iteratively updated through the elastic time slice-reinforcement learning hybrid algorithm, and the task priority coefficient, device running load value, and location distance weight are used as the calibration input variables for algorithm training during the update process.

[0027] Specifically, step S4 continuously iteratively updates the task scheduling strategy in each elastic time slice interval divided in step S3 through the elastic time slice-reinforcement learning hybrid algorithm, and constantly optimizes the matching scheme of tasks and devices. The core value of this step lies in using the autonomous learning ability of reinforcement learning combined with the dynamic adjustment characteristics of elastic time slices, so that the scheduling strategy can be adaptively optimized according to real-time parameter changes, avoiding the problem that traditional fixed strategies cannot adapt to dynamic changes in the warehouse environment, and ensuring that the scheduling strategy is always in the optimal state. In specific implementation, the scheduling strategy is initialized for each elastic time slice interval, and the initial strategy allocates devices according to task priority coefficients from high to low, and the device selection preferentially matches devices with a load value below 0.6. Then the iteration process of the elastic time slice-reinforcement learning hybrid algorithm is started, and the task priority coefficient, device running load value and distance weight of the goods location in the interval are taken as input variables each time. In the iteration process, the algorithm calculates the strategy optimization degree every 10 seconds, and the optimization degree is calculated by three indexes including task completion time deviation (the difference between the actual completion time and the threshold value), device load balancing degree (the standard deviation of each device load value) and goods location distance sum. When the optimization degree improvement amplitude is less than 0.5% for 3 times in a row, the iteration of the time slice interval is stopped; if the optimization degree does not reach the preset threshold value (set to 90%), the algorithm learning rate is adjusted (the initial learning rate is set to 0.01, and it is reduced by 0.001 every 10 iterations, and the minimum is 0.001), and the iteration is continued. The upper limit of the iteration number of the low uncertainty interval is set to 50 times, the medium uncertainty interval is set to 80 times, and the high uncertainty interval is set to 120 times, so as to ensure that the scheduling strategy of each interval can be fully optimized, and the updated strategy is stored in the strategy database of the RidgeOS intelligent warehouse scheduling platform.

[0028] Step S5, according to the iteratively updated scheduling strategy, the scheduling order of the tasks to be executed in the WMS warehouse is sequenced, and the uncertainty feature vector output by the dynamic scheduling task uncertainty perception model is called as a sequencing constraint condition in the sequencing process;

[0029] Specifically, step S5 updates the scheduling strategy according to step S4, and sorts the scheduling sequence of all tasks to be executed in the WMS warehouse, while introducing the uncertainty feature vector generated in step S2 as a constraint condition to ensure that the sorting result meets the optimization goal and can cope with the uncertainty risk in task execution. This step converts the optimized scheduling strategy into a specific task execution sequence, reduces the scheduling risk through constraint conditions, avoids task delays or equipment overload caused by ignoring uncertainty factors, and improves the stability and reliability of scheduling execution. In the specific implementation process, first, the corresponding iterative scheduling strategy of each flexible time slice interval is retrieved from the strategy database of the RidgeOS intelligent warehouse scheduling platform, and the task-device matching rules, priority sorting logic and load limit conditions in the strategy are extracted. Then, the tasks to be executed in the WMS warehouse are classified according to the corresponding flexible time slice interval, and the tasks in the same interval are sorted according to the priority sorting logic in the scheduling strategy first, and the tasks with the same priority are sorted according to the weight of the location distance from small to large. In the sorting process, the uncertainty feature vector output by the dynamic scheduling task uncertainty perception model is called, and the constraint conditions are set: when the task uncertainty feature value is greater than 0.8, the task needs to be allocated to the device with a load value less than 0.5, and the execution interval with the previous task needs to be greater than 5 minutes; when the feature value is between 0.5 and 0.8, the device load value needs to be less than 0.6, and the execution interval needs to be greater than 3 minutes; when the feature value is less than 0.5, the device load value needs to be less than 0.7, and the execution interval needs to be greater than 1 minute. After sorting, a task scheduling sequence table is generated, which contains task number, interval, device number, planned start time, planned completion time and constraint condition remarks, and the scheduling sequence table is synchronized to the task management module of the RidgeOS intelligent warehouse scheduling platform in real time.

[0030] Step S6, through the RidgeOS intelligent warehouse scheduling platform, the sorted task scheduling instruction is sent to the corresponding warehouse equipment, and the parameter change data in the process of executing the scheduling instruction is collected in real time and fed back to the dynamic scheduling task uncertainty perception model.

[0031] Specifically, step S6 converts the task scheduling sequence table generated in step S5 into device executable scheduling instructions through the RidgeOS intelligent warehouse scheduling platform, and sends them to the corresponding warehouse devices, while collecting parameter change data in the device execution process in real time, feeding back to the dynamic scheduling task uncertainty perception model, forming a "scheduling-execution-feedback-optimization" closed loop. This step realizes the landing execution of the scheduling strategy, while providing the basis for model parameter adjustment and subsequent scheduling optimization through real-time data feedback, ensuring that the entire scheduling system can continuously adapt to changes in the warehouse environment and improve long-term scheduling results. In specific implementation, the instruction generation module of the RidgeOS intelligent warehouse scheduling platform first disassembles each task in the task scheduling sequence table into device action instructions, such as forklift pickup location, travel path, and unloading location, etc. The instruction format conforms to the communication protocol of the warehouse device (such as Modbus protocol). Then the instructions are sent to the corresponding device through the wireless communication module (using 4G or WiFi communication method, communication rate not less than 1 Mbps), and the instruction sending time and device receiving confirmation information are recorded after successful sending. At the same time, the real-time data acquisition module of the platform collects parameter change data in the execution process through device sensors, including device real-time load value (collected every 5 seconds), task actual execution time (time difference from device starting action to completing action), actual distance of goods location (actual path length of device travel), and device fault warning information (triggered when device temperature exceeds 60℃ or current exceeds rated value 110%). The collected data are classified and arranged according to task number, and a data feedback report is generated every 30 seconds, which is transmitted to the parameter update module of the dynamic scheduling task uncertainty perception model through the data interface. The model adjusts the weight proportion of each parameter according to the feedback data, such as when the device fault warning information increases, the weight proportion of the device fault probability factor is increased by 5% to 15%, ensuring that the model can accurately reflect the current warehouse environment state.

[0032] Preferably, the dynamic scheduling task uncertainty perception model calculates the task execution uncertainty index when generating the uncertainty feature vector in step S2 using the following model formula: i = α · P i + β · L j + γ · D k - δ · T m wherein U i represents the execution uncertainty index of the i-th task, α, β, γ, δ are the weight coefficients of the task priority coefficient, device running load value, goods location distance weight, and task execution time threshold, respectively, P i represents the priority coefficient of the i-th task, L j represents the running load value of the j-th device, D k represents the distance weight of the k-th goods location, and T mAn execution duration threshold value of the mth type of task.

[0033] Specifically, the dynamic scheduling task uncertainty perception model generates the uncertainty feature vector in step S2 by quantifying the task execution uncertainty index through a specific calculation method, providing accurate uncertainty measurement basis for subsequent scheduling decisions. In implementation, the calculation method integrates four types of key parameters: task priority coefficient, device running load value, storage location distance weight, and task execution duration threshold value, and sets corresponding weight coefficients for each type of parameter. The weight coefficient of the task priority coefficient is set by referring to the daily task urgency distribution of the warehouse, and the value of the coefficient is higher when the proportion of urgent tasks is high, usually between 0.2 and 0.4. The weight coefficient of the device running load value needs to be combined with the device overload risk threshold value, and the value of the coefficient increases when the maximum load threshold value of the device is lower, usually between 0.15 and 0.3. The weight coefficient of the storage location distance weight is adjusted according to the warehouse storage location distribution density, and the value of the coefficient increases when the dispersion degree of the storage location is high, usually between 0.1 and 0.25. The weight coefficient of the task execution duration threshold value needs to match the severity of the consequences of task overtime, and the coefficient is higher for tasks with greater overtime impact, usually between 0.05 and 0.2. Through the combination calculation of these parameters and weight coefficients, the execution uncertainty index of each task is obtained, and the index value range is controlled between 0 and 5. The higher the index, the greater the possibility of abnormality caused by interference from various factors in the task execution process. The calculation result is directly used for dimension filling of the uncertainty feature vector in step S2, ensuring that the feature vector can accurately reflect the task uncertainty level.

[0034] Preferably, the elastic time slice-reinforcement learning hybrid algorithm determines the time slice length by using the following model formula when dividing the elastic time slice interval in step S3: wherein, T s represents the length of the elastic time slice, μ, v, λ, σ are respectively the adjustment coefficients of the maximum value of the task priority, the average load value of the device, the average storage location distance weight, and the average device failure probability factor, P max represents the maximum value of the priority coefficient in all tasks, L avg represents the average running load value of the warehouse device, D avg represents the average distance weight of the warehouse storage location, F avg represents the average failure probability factor of the warehouse device.

[0035] Specifically, the process of dividing the flexible time slice interval in step S3 is to calculate the length of the flexible time slice to achieve precise adaptation of the time slice interval to the task characteristics. In the implementation process, the calculation needs to rely on four types of parameters: the maximum value of task priority, the average load value of the device, the average distance weight of the goods location, and the average device failure probability factor. At the same time, an adjustment coefficient is configured for each type of parameter. The adjustment coefficient of the maximum value of task priority needs to be set according to the priority distribution range of the warehouse. When the priority span is large (such as 1 to 5), the coefficient value is higher, generally between 0.3 and 0.5; the adjustment coefficient of the average load value of the device needs to be combined with the safe load interval of the device. When the average load of the device is close to the upper limit of the safe load, the value is increased, usually between 0.2 and 0.4; the adjustment coefficient of the average distance weight of the goods location is adjusted according to the average operation radius of the warehouse. When the operation radius is large, the value is increased, generally between 0.1 and 0.3; the adjustment coefficient of the average device failure probability factor needs to refer to the historical failure frequency of the device. When the proportion of frequently malfunctioning devices is high, the value is increased, usually between 0.05 and 0.2. The calculated length of the flexible time slice needs to be controlled within the range of 3 to 20 minutes, and needs to meet the adaptation logic that the longer the time slice length, the more the number of tasks in the same uncertainty interval and the greater the parameter fluctuation, to ensure that the tasks in each time slice interval can be completed within a reasonable time range. Optimization scheduling is avoided due to too short time slice or delay due to too long scheduling.

[0036] Preferably, when the flexible time slice-reinforcement learning hybrid algorithm iteratively updates the scheduling strategy in step S4, the following model formula is used to calculate the strategy update reward value: R t = η · (P i -P avg ) - θ · (L j -L avg ) + ζ · (D std -D k ), where R t represents the strategy update reward value of the tthiteration, η, θ, ζ are the reward coefficients of the task priority deviation, the device load deviation, and the goods location distance weight deviation, respectively, P avg represents the average priority coefficient of all tasks, D std represents the standard deviation of the warehouse goods location distance weight.

[0037] Specifically, the elastic time slice-reinforcement learning hybrid algorithm iteratively updates the scheduling strategy in step S4 by calculating a strategy update reward value to guide the algorithm to iterate in a better scheduling direction and improve the efficiency of strategy optimization. In practice, the calculation needs to be based on three key deviation indicators: task priority deviation, device load deviation, and location distance weight deviation, and a reward coefficient is set for each type of indicator. The reward coefficient of the task priority deviation needs to be combined with the priority execution accuracy requirement. When the warehouse has strict time efficiency requirements for high-priority task execution, the coefficient takes a higher value, generally between 0.25 and 0.45. The reward coefficient of the device load deviation needs to match the device load balancing goal. When the device load difference is too large and is likely to cause failure, the value is increased, usually between 0.2 and 0.4. The reward coefficient of the location distance weight deviation is adjusted according to the warehouse operation energy consumption control requirement. The value of this coefficient increases in energy-sensitive warehouses, generally between 0.1 and 0.3. In the calculation process, the task priority deviation is obtained by the difference between the priority coefficient of a single task and the average priority coefficient of all tasks, the device load deviation is the difference between the running load value of a single device and the average load value of the device, and the location distance weight deviation is the difference between the distance weight of a single location and the standard deviation of the location distance weight. Through the combination calculation of these deviation indicators and reward coefficients, the strategy update reward value of each iteration is obtained, the reward value ranges between -2 and 3, a positive value represents that the current strategy adjustment direction meets the optimization goal, and a negative value indicates that it needs to be adjusted in the opposite direction. This reward value is directly used as the basis for updating the strategy parameters in the algorithm iteration process, ensuring that each iteration can proceed in the direction of improving scheduling efficiency and stability.

[0038] Preferably, when the dynamic scheduling task uncertainty perception model is used as a sorting constraint condition in step S5, the following model formula is used to calculate the constraint strength: wherein C s represents the constraint strength value, ε、 are the uncertainty index and the constraint coefficient of the time slice length deviation, respectively, T avg represents the average length of all elastic time slice intervals.

[0039] Specifically, the process of using the dynamic scheduling task uncertainty perception model as a sorting constraint condition in step S5 is to calculate the constraint strength value, which explicitly indicates the influence of uncertainty factors on task sorting, and ensures that the sorting result takes into account both optimization objectives and risk control. In implementation, the calculation needs to combine the task execution uncertainty index and the elastic time slice length deviation two types of parameters, and configure a constraint coefficient for each type of parameter. The constraint coefficient of the task execution uncertainty index needs to be set according to the warehouse risk bearing capacity, and the higher the risk sensitivity of the warehouse, the higher the value of this coefficient, which is generally between 0.3 and 0.5; the constraint coefficient of the elastic time slice length deviation needs to match the time slice interval scheduling efficiency requirement, and the value is increased when the time slice length deviation is too large to cause scheduling confusion, which is usually between 0.2 and 0.4. Among them, the elastic time slice length deviation is obtained by the difference between the length of a single elastic time slice and the average length of all elastic time slice intervals, and the positive difference represents that the length of the time slice is longer than the average level, and the negative difference represents that the length of the time slice is shorter than the average level. The constraint strength value obtained by calculation is controlled in the range of 0 to 4, and the higher the strength value, the stricter the constraint condition that the task needs to follow in the sorting process, for example, when the strength value is higher than 3, the task sorting needs to meet the requirements of lower device load, longer execution interval, etc. The calculation result is directly used for the constraint rule making of task sorting in step S5, which ensures that high-uncertainty tasks can get more cautious device allocation and time arrangement in sorting.

[0040] Preferably, when the RidgeOS intelligent warehouse scheduling platform feeds back parameter change data in step S6, the following model formula is used to calculate the data feedback weight: W f = κ·ΔL j + τ·ΔD k + ω·ΔF j , wherein W f represents the data feedback weight, κ, τ, ω are weight coefficients of the device load change amount, the distance weight change amount of the storage location, and the fault probability factor change amount of the device, respectively, ΔL j represents the load change amount of the jth device, ΔD k represents the distance weight change amount of the kth storage location, and ΔF j represents the fault probability factor change amount of the jth device.

[0041] Specifically, the RidgeOS intelligent warehouse scheduling platform feedback parameter change data in step S6 is calculated by calculating data feedback weight to determine the influence priority of different parameter change data on model adjustment, to ensure that the model can accurately respond to key parameter changes. In the implementation process, three types of dynamic parameters need to be relied on, including equipment load change, storage location distance weight change and equipment failure probability factor change. And a weight coefficient is set for each type of parameter. The weight coefficient of the equipment load change needs to be set in combination with the equipment load change rate. When the load fluctuates greatly in a short time, the coefficient value is higher, generally between 0.3 and 0.5. The weight coefficient of the storage location distance weight change needs to be adjusted according to the storage location adjustment frequency. The coefficient value is improved for the warehouse with frequent storage location changes, usually between 0.15 and 0.3. The weight coefficient of the equipment failure probability factor change needs to be referred to the equipment failure influence range. The coefficient of the equipment that will cause multiple task interruptions due to failure is higher, generally between 0.2 and 0.4. The parameter change value is obtained by the parameter difference before and after the device executes the scheduling instruction, for example, the equipment load change is the difference between the load value after execution and the load value before execution. A positive value represents an increase in load, and a negative value indicates a decrease in load. The data feedback weight calculated has a value range of 0 to 3. The higher the weight, the greater the influence of the parameter change data on the model parameter adjustment after being fed back to the dynamic scheduling task uncertainty perception model. For example, the equipment failure probability factor change with a weight higher than 2 will directly trigger the adjustment of the related weight proportion of the factor in the model, and the adjustment range is usually between 5% and 15%, to ensure that the model can prioritize optimizing its recognition accuracy according to key parameter changes.

[0042] Preferably, step S3 comprises the following steps: step S31, obtaining the uncertainty feature vector output from the dynamic scheduling task uncertainty perception model, extracting the parameter values corresponding to each dimension in the feature vector, arranging the values in ascending order to form a parameter sequence; step S32, based on the parameter sequence, setting the initial threshold for the interval division of the elastic time slice, the initial threshold is determined according to the five equal division results of the difference between the maximum value and the minimum value of the parameter sequence, and each equal interval corresponds to an initial time slice interval; step S33, inputting the initial time slice interval into the elastic time slice-reinforcement learning hybrid algorithm, the algorithm adjusts the initial interval boundary according to the distribution of the task priority coefficient, expands or reduces the interval range, so that the variance of the task priority coefficient in each interval is within the preset range; step S34, recording the boundary values of the adjusted elastic time slice interval and the corresponding task quantity, generating an interval-task quantity mapping table as basic data for subsequent scheduling strategy iteration and update.

[0043] Specifically, step S3 includes four sub-steps to ensure the accuracy of the elastic time slice interval division. In step S31, after obtaining the uncertainty feature vector from the dynamic scheduling task uncertainty perception model, the dimension parameter values are extracted, arranged in ascending order to form a parameter sequence, and the parameter sequence length is consistent with the number of tasks to be scheduled. For example, when there are 50 tasks to be scheduled in the warehouse, the parameter sequence contains 50 values arranged in order; in step S32, based on the parameter sequence, set the initial threshold value, first calculate the difference between the maximum and minimum values of the parameter sequence, and divide the difference into five equal parts to obtain five initial threshold values, each threshold value corresponds to an initial time slice interval. If the difference is 1, the threshold interval of each equal interval is 0.2, forming five initial intervals; in step S33, the initial interval is input into the elastic time slice-reinforcement learning hybrid algorithm, and the algorithm adjusts the interval boundary according to the task priority coefficient distribution. The variance of the task priority coefficient in the interval is used as the judgment standard in the adjustment process, and the preset variance range is 0.05 to 0.1. If the variance of a certain interval exceeds 0.1, the interval range is expanded, and if it is lower than 0.05, the interval range is reduced, until the variance meets the requirements; in step S34, the adjusted interval boundary value and the number of tasks in each interval are recorded, and an interval-task quantity mapping table is generated. The mapping table needs to clearly indicate the upper and lower limits of each interval and the corresponding number of tasks. For example, the interval boundary is 0.2 to 0.4, and the corresponding number of tasks is 12. This mapping table is used as the basis data for the iteration and update of the scheduling strategy in step S4, ensuring that the iteration process is only carried out for the tasks in the interval, and improving the optimization efficiency.

[0044] Preferably, step S4 includes the following sub-steps: step S41, calling historical scheduling data from the RidgeOS intelligent warehouse scheduling platform, and screening out historical data groups similar to the current elastic time slice interval parameters. The similarity judgment is based on the parameter value deviation being less than the preset deviation threshold; step S42, inputting the screened historical data group into the training module of the elastic time slice-reinforcement learning hybrid algorithm, taking the task completion order and device running state in the historical data as the training sample of the algorithm, and initializing the reward function parameters of the algorithm; step S43, selecting part of the tasks to be executed in the current elastic time slice interval as test samples, inputting the trained algorithm model, outputting the preliminary scheduling strategy, and calculating the deviation value of the preliminary strategy and the historical optimal strategy; step S44, adjusting the learning rate parameter of the algorithm according to the deviation value, inputting the test sample again for strategy iteration, and repeating the iteration process until the deviation value is less than the preset upper deviation limit, to determine the final scheduling strategy update result.

[0045] Specifically, step S4 realizes efficient iterative updating of the scheduling strategy through four sub-steps. In step S41, historical scheduling data is called from the RidgeOS intelligent warehouse scheduling platform, and when filtering similar historical data sets, the parameter value deviation threshold is set to 0.08, that is, the difference between the task priority coefficient and the device running load value of the current elastic time slice interval and the corresponding parameters of the historical data set should be less than 0.08, so as to be identified as a similar data set, for example, the current device running load value is 0.6, and the load value in the historical data set between 0.52 and 0.68 is filtered out; in step S42, the filtered historical data set is input into the algorithm training module, and the task completion order and device running state in the historical data are used as training samples to initialize the reward function parameters, and the initial weight of the reward function is set to 40% for task completion efficiency, 35% for device load balancing, and 25% for location distance optimization; in step S43, 30% of the tasks to be executed in the current interval are selected as test samples, and the trained algorithm model is input to output a preliminary scheduling strategy, and the deviation value of the preliminary strategy and the historical optimal strategy is calculated, the deviation value is calculated by the difference between the two in task completion time, device load fluctuation and other indicators, and the preset upper limit of the deviation is 0.12; in step S44, if the deviation value is greater than 0.12, the algorithm learning rate is adjusted, the initial learning rate is 0.01, and it is reduced by 0.001 each time, and the minimum is 0.002, and the test sample is input again for iteration, and the iteration is repeated until the deviation value is less than 0.12, and the final scheduling strategy updating result is determined, which ensures that the scheduling strategy not only refers to historical experience, but also adapts to the current task characteristics, and improves the feasibility of the strategy.

[0046] Preferably, step S5 includes the following sub-steps: step S51, obtaining the scheduling strategy updating result output by step S4, extracting the task priority sorting rules and device allocation rules included in the strategy, and arranging to form a sorting rule list; step S52, obtaining the uncertainty feature vector of all tasks to be executed from the dynamic scheduling task uncertainty perception model, and calculating the similarity between each task vector and the standard vector in the sorting rule list; step S53, preliminarily sorting the tasks according to the similarity, and the higher the similarity, the higher the task sorting; at the same time, real-time load data of the devices in the WMS warehouse is called to perform device load adaptability verification on the preliminary sorting result; step S54, adjusting the position of the task that does not meet the device load requirement after verification, and referring to the location distance weight parameter in the adjustment process to ensure that the adjusted task order meets the device load limit and the total location distance weight is minimized.

[0047] Specifically, step S5 guarantees the rationality and feasibility of task sequencing through four sub-steps. In step S51, after obtaining the scheduling strategy update result of step S4, the task priority sequencing rule (such as high priority task priority sequencing) and the device allocation rule (such as low load device priority allocation) are extracted, and a sequencing rule list is formed. The execution order of the rules in the list needs to be clear, for example, first sequencing by task priority, and then adjusting by device load adaptability. In step S52, the uncertainty feature vector of the task to be executed is obtained from the dynamic scheduling task uncertainty perception model. The cosine similarity algorithm is used to calculate the vector similarity, and the similarity value ranges from 0 to 1. The closer the value is to 1, the more similar the task vector is to the standard vector. For example, if the similarity of a certain task vector to the standard vector is 0.92, it is determined to be high similarity. In step S53, the tasks are preliminarily sequenced from high to low similarity, and the real-time load data of the devices is called. The device load adaptability threshold is set to 0.7, that is, the total load of the tasks allocated to the device should be less than 0.7. If the total load of a certain device after task allocation in the preliminary sequencing reaches 0.85, which exceeds the threshold, the adaptability check fails. In step S54, the position of the task that fails the check is adjusted. When adjusting, the location distance weight parameter is referred to, and the goal is to minimize the total location distance weight of all tasks after adjustment. For example, the task originally allocated to device A is adjusted to device B, and the total weight is reduced from 15.6 to 14.2. The adjusted task order needs to meet the conditions of device load less than 0.7 and total location distance weight minimum, ensuring that the sequencing result takes into account efficiency and resource optimization, and avoiding device overload or long transportation path.

[0048] The dynamic scheduling task uncertainty perception model is a core model used to identify the uncertainty characteristics in the execution process of WMS warehouse tasks in the present application. Its essence is to convert scattered task, device, and location related data into standardized uncertainty feature vectors through multi-dimensional parameter fusion analysis. In terms of implementation, the model first receives the task priority coefficient, device running load value, location distance weight, task execution time threshold, and device failure probability factor collected by the RidgeOS intelligent warehouse scheduling platform, and arranges them into a parameter matrix in a predetermined order. Then, the parameters are mapped through a multi-layer perception network. The number of input layer nodes matches the types of parameters, the number of hidden layer nodes is set according to the dimension of parameters, and the output layer generates a single task uncertainty feature value. At the same time, different parameters are configured with different weight proportions, such as a task priority coefficient weight of 30% and a device running load value of 25%. Finally, an uncertainty feature vector corresponding to the number of tasks is formed. The role of this model is to break the limitations of traditional single parameter judgment, accurately capture potential risks in task execution, and provide constraints for subsequent scheduling decision space construction and task sequencing. It solves the problem of insufficient handling of uncertainty in existing technologies, allowing the scheduling strategy to adapt to the dynamic changes of complex warehouse environments and improving scheduling stability and reliability.

[0049] The elastic time slice-reinforcement learning hybrid algorithm is a key algorithm for realizing scheduling strategy optimization and iterative updating in the present application, and has the dual characteristics of elastic time slice dynamic division and reinforcement learning autonomous optimization, which is used to decompose complex scheduling problems and realize continuous optimization of strategies. In terms of implementation, the algorithm first divides a first interval according to the uncertainty degree based on the feature vector output by the dynamic scheduling task uncertainty perception model, and then dynamically adjusts the time slice length according to the number of tasks in the interval and the parameter fluctuation, such as setting the time slice to 5 minutes when the number of tasks in the low uncertainty interval is less than 20, and setting the time slice to 12 minutes when the number exceeds 50. Then, the task priority coefficient, device running load value, etc. are used as input variables to initialize the scheduling strategy, and the strategy is adjusted by iterative calculation of the optimization degree and reward value. If the optimization degree improvement amplitude is less than the threshold value for many times in succession in the iteration, the iteration is stopped, and if the preset optimization target is not reached, the learning rate is adjusted and the iteration is continued. The role of this algorithm is to build a scientific scheduling decision space, realize autonomous iterative updating of the scheduling strategy, and avoid the limitations of fixed time slices and manual adjustment. The flexibility and optimization efficiency of the scheduling strategy are improved, the strategy can adapt to changes in the warehouse environment in real time, and the problems of low efficiency and poor adaptability of traditional scheduling are solved, providing algorithm support for multi-objective optimization scheduling.

[0050] The RidgeOS intelligent warehouse scheduling platform is a core hardware and software integrated system for connecting data collection, instruction sending, and data feedback in the present application, and is the "central nervous system" of the entire scheduling process. In terms of implementation, the platform establishes real-time connection with the WMS system, device sensors, and storage location management module, collects task, device, and storage location related parameters at a fixed frequency (such as every 10 seconds), and decomposes the task sorting table into action instructions conforming to the device communication protocol during the scheduling execution phase, and sends them to the corresponding device through the 4G or WiFi module, and records the instruction transmission status. At the same time, it collects parameter change data during device execution in real time, such as real-time device load and actual task execution time, and generates a data feedback report every 30 seconds, which is transmitted to the dynamic scheduling task uncertainty perception model. The role of this platform is to realize the closed-loop process of "parameter collection-analysis-scheduling-feedback", provide data input support for the model and algorithm, and at the same time ensure the landing execution of scheduling instructions and real-time feedback of data. It integrates the data flow and instruction transmission links in warehouse management, solves the problem of data dispersion and feedback lag in the prior art, provides real-time basis for model parameter adjustment and strategy optimization, and ensures the efficient operation of the entire scheduling system.

[0051] As Figure 2As shown, the WMS warehouse task dynamic scheduling system based on multi-objective optimization includes: a warehouse parameter acquisition and instruction sending unit, which is connected with each device in the WMS warehouse to collect parameters for WMS warehouse task dynamic scheduling and send task scheduling instructions; a scheduling task feature recognition and vector generation unit, which is connected with the warehouse parameter acquisition and instruction sending unit through a data bus, receives the collected parameters and generates an uncertainty feature vector; a scheduling decision space construction and interval division unit, which is connected with the scheduling task feature recognition and vector generation unit through a communication interface, constructs a decision space based on the feature vector and divides the time slice interval; a scheduling strategy iteration update and optimization unit, which is connected with the scheduling decision space construction and interval division unit through a data link, performs iteration update of the scheduling strategy; a WMS warehouse task scheduling order sorting and constraint checking unit, which is connected with the scheduling strategy iteration update and optimization unit and the scheduling task feature recognition and vector generation unit through signal lines, completes task sorting and constraint checking; a scheduling data real-time feedback and model parameter adjustment unit, which is connected with the warehouse parameter acquisition and instruction sending unit through a communication protocol at one end and with the scheduling task feature recognition and vector generation unit through a data interface at the other end, performs data feedback and model adjustment.

[0052] The WMS warehouse task dynamic scheduling method and system based on multi-objective optimization effectively overcome the shortcomings of the prior art in terms of insufficient handling of task execution uncertainty by constructing a dynamic scheduling task uncertainty perception model. Instead of relying on a single parameter to determine the task state, the model generates a comprehensive uncertainty feature vector by integrating multiple-dimensional parameters such as task priority, device operating load, storage location distance, task execution time threshold, and device failure probability factor. This multi-parameter fusion perception method can accurately capture various uncertainty factors in task execution, making the scheduling strategy less susceptible to sudden factors and significantly improving the stability of scheduling in complex warehouse environments, meeting the dynamic warehouse scheduling requirements.

[0053] Through the elastic time slice-reinforcement learning hybrid algorithm, the system solves the problems of lack of scheduling flexibility and iteration optimization ability in traditional technology, and exhibits significant advantages. The algorithm can dynamically divide the elastic time slice interval based on the feature vector output by the dynamic scheduling task uncertainty perception model, rather than using a fixed time slice division method, and can adjust the interval range according to real-time parameter data to adapt to different task priorities. At the same time, the algorithm iteratively updates the scheduling strategy autonomously without human intervention, continuously optimizes the task scheduling order and device allocation scheme, significantly improves the scheduling efficiency, and enables the scheduling strategy to continuously adapt to the dynamic changes in the warehouse environment.

[0054] Relying on the RidgeOS intelligent warehouse scheduling platform, the system forms a complete closed loop of "parameter collection-feature analysis-strategy generation-instruction sending-data feedback", further strengthening the precision and adaptability of scheduling. The platform can not only collect various key parameters of WMS warehouse task dynamic scheduling in real time to provide accurate input for the model and algorithm, but also accurately send the sorted scheduling instructions to the corresponding equipment and collect the parameter change data during the equipment execution to feed back to the uncertainty perception model. This closed loop mechanism ensures that the model and algorithm can be continuously optimized based on real-time data, not only achieving multi-objective optimization scheduling, but also improving the efficiency of warehouse task execution and equipment utilization, and fully compensating for the shortcomings of existing technologies.

[0055] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "set", "install", "connect", "connect", "fix" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integral connection; can be mechanical connection, or electrical connection; can be directly connected, or indirectly connected through an intermediate medium, or the internal communication of two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0056] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various equivalent changes, modifications, replacements and variations of the embodiments can be made without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalent scope.

Claims

1. A WMS warehouse task dynamic scheduling method based on multi-objective optimization, characterized in that, The method comprises the following steps: Step S1, collecting various parameters in the dynamic scheduling process of the WMS warehouse task through the RidgeOS intelligent warehouse scheduling platform, the parameters including task priority coefficient, device running load value, storage location distance weight, task execution time length threshold and device failure probability factor; Step S2, inputting the collected various parameters into the dynamic scheduling task uncertainty perception model, identifying the task execution uncertainty characteristics corresponding to the various parameters through the model, generating an uncertainty feature vector, and the dimension of the feature vector is consistent with the number of parameter types collected; Step S3, based on the uncertainty feature vector, using an elastic time slice-reinforcement learning hybrid algorithm to construct a task scheduling decision space, dividing different elastic time slice intervals in the decision space, and each interval corresponds to a different task scheduling priority range; Step S4, in each elastic time slice interval, the task scheduling strategy is iteratively updated through the elastic time slice-reinforcement learning hybrid algorithm, and the task priority coefficient, the device running load value and the storage location distance weight are used as the calibration input variables of the algorithm training in the updating process; Step S5, according to the scheduling strategy updated iteratively, the scheduling order of the tasks to be executed in the WMS warehouse is sorted, and the uncertainty feature vector output by the dynamic scheduling task uncertainty perception model is called as a sorting constraint condition in the sorting process; Step S6, sending the sorted task scheduling instruction to the corresponding warehouse equipment through the RidgeOS intelligent warehouse scheduling platform, and collecting the parameter change data in the process of executing the scheduling instruction by the equipment in real time and feeding back to the dynamic scheduling task uncertainty perception model.

2. The multi-objective optimization based WMS warehouse task dynamic scheduling method according to claim 1, characterized in that, The dynamic scheduling task uncertainty perception model generates an uncertainty feature vector in step S2, and calculates a task execution uncertainty index using the following model formula: i = α · P i + β · L j + γ · D k - δ · T m , where U i represents the execution uncertainty index of the i-th task, α, β, γ, δ are weight coefficients of the task priority coefficient, the equipment running load value, the distance weight of the storage location, and the execution time threshold of the task, respectively, P i represents the priority coefficient of the i-th task, L j represents the running load value of the j-th equipment, D k represents the distance weight of the k-th storage location, and T m represents the execution time threshold of the m-th type of task.

3. The multi-objective optimization based WMS warehouse task dynamic scheduling method according to claim 1, characterized in that, The elastic time slice-reinforcement learning hybrid algorithm adopts the following model formula to determine the time slice length when dividing the elastic time slice interval in step S3: wherein, T s represents the elastic time slice length, μ, v, λ, σ are respectively the adjustment coefficients of the maximum value of task priority, the average load value of equipment, the average distance weight of goods location, and the average failure probability factor of equipment, P max represents the maximum value of the priority coefficient in all tasks, L avg represents the average running load value of the warehouse equipment, D avg represents the average distance weight of the warehouse goods location, F avg represents the average failure probability factor of the warehouse equipment.

4. The multi-objective optimization based WMS warehouse task dynamic scheduling method according to claim 1, characterized in that, The elastic time slice-reinforcement learning hybrid algorithm uses the following model formula to calculate the policy update reward value when iteratively updating the scheduling policy in step S4: t = η · (P i - P avg ) - θ · (L j - L avg ) + ζ · (D std - D k ), wherein R t represents the policy update reward value of the tthiteration, η, θ, ζ are reward coefficients of task priority deviation, device load deviation, and location distance weight deviation, respectively, P avg represents the average priority coefficient of all tasks, D std represents the standard deviation of the warehouse location distance weight.

5. The multi-objective optimization based WMS warehouse task dynamic scheduling method according to claim 1, characterized in that, The dynamic scheduling task uncertainty perception model uses the following model formula to calculate the constraint strength when it is used as a sequencing constraint condition in step S5: wherein C s represents the constraint strength value, ε、 are the constraint coefficients of the uncertainty index and the time slice length deviation, respectively, T avg represents the average length of all flexible time slice intervals.

6. The multi-objective optimization based WMS warehouse task dynamic scheduling method according to claim 1, characterized in that, The RidgeOS intelligent warehouse scheduling platform adopts the following model formula to calculate the data feedback weight when feeding back the parameter change data in step S6: W f = κ·ΔL j + τ·ΔD k + ω·ΔF j , wherein W f represents the data feedback weight, κ, τ, ω are weight coefficients of the equipment load change amount, the distance weight change amount of the storage location, and the fault probability factor change amount of the equipment, respectively, ΔL j represents the load change amount of the jth equipment, ΔD k represents the distance weight change amount of the kth storage location, and ΔF j represents the fault probability factor change amount of the jth equipment.

7. The multi-objective optimization based WMS warehouse task dynamic scheduling method according to claim 1, characterized in that, Step S3 comprises the following steps: Step S31, obtaining the uncertainty feature vector output from the dynamic scheduling task uncertainty perception model, extracting the parameter values corresponding to each dimension in the feature vector, arranging the values in order from small to large to form a parameter sequence; Step S32, based on the parameter sequence, setting an initial threshold for dividing the elastic time slice intervals, and the initial threshold is determined according to the five equal division results of the difference between the maximum value and the minimum value of the parameter sequence, and each equal division interval corresponds to an initial time slice interval; Step S33, inputting the initial time slice interval into the elastic time slice-reinforcement learning hybrid algorithm, and the algorithm adjusts the boundaries of the initial interval according to the distribution of the task priority coefficient, expands or reduces the interval range, so that the variance of the task priority coefficient in each interval is within a preset range; Step S34, recording the boundary values of the adjusted elastic time slice intervals and the corresponding task quantities, generating an interval-task quantity mapping table as basic data for subsequent scheduling strategy iterative updating.

8. The multi-objective optimization based WMS warehouse task dynamic scheduling method according to claim 1, characterized in that, Step S4 includes the following sub-steps: step S41, calling historical scheduling data from the RidgeOS intelligent warehouse scheduling platform, filtering out historical data groups similar to the current flexible time slice interval parameters, and the similarity judgment basis is that the parameter value deviation is less than the preset deviation threshold; step S42, inputting the filtered historical data groups into the training module of the flexible time slice-reinforcement learning hybrid algorithm, taking the task completion order and device running state in the historical data as the training sample of the algorithm, and initializing the reward function parameters of the algorithm; step S43, selecting part of the to-be-executed tasks in the current flexible time slice interval as test samples, inputting the trained algorithm model, outputting the preliminary scheduling strategy, and calculating the deviation value of the preliminary strategy and the historical optimal strategy; step S44, adjusting the learning rate parameter of the algorithm according to the deviation value, inputting the test sample again for strategy iteration, and repeating the iteration process until the deviation value is less than the preset upper deviation limit, and determining the final scheduling strategy update result.

9. The multi-objective optimization based WMS warehouse task dynamic scheduling method according to claim 1, characterized in that, Step S5 includes the following sub-steps: step S51, obtaining the scheduling strategy update result output by step S4, extracting the task priority ordering rules and device allocation rules included in the strategy, and organizing to form an ordering rule list; step S52, obtaining the uncertainty feature vector of all to-be-executed tasks from the dynamic scheduling task uncertainty perception model, and calculating the similarity of each task vector and the standard vector in the ordering rule list; step S53, preliminarily ordering the tasks according to the similarity, and the higher the similarity, the higher the task ordering, and simultaneously calling the real-time load data of the devices in the WMS warehouse, and performing device load adaptability verification on the preliminary ordering result; step S54, adjusting the position of the tasks that do not meet the device load requirement after verification, and referring to the location distance weight parameter in the adjustment process, to ensure that the adjusted task order meets the device load limit and the total location distance weight is minimum.

10. A WMS warehouse task dynamic scheduling system based on multi-objective optimization, characterized in that, Comprise: The warehouse parameter acquisition and instruction sending unit is connected with each device in the WMS warehouse to acquire parameters of dynamic scheduling of the WMS warehouse task and send task scheduling instructions; the scheduling task feature recognition and vector generation unit is connected with the warehouse parameter acquisition and instruction sending unit through a data bus to receive the acquired parameters and generate an uncertainty feature vector; the scheduling decision space construction and interval division unit is connected with the scheduling task feature recognition and vector generation unit through a communication interface to construct a decision space based on the feature vector and divide a time slice interval; the scheduling strategy iteration update and optimization unit is connected with the scheduling decision space construction and interval division unit through a data link to perform iteration update of the scheduling strategy; the WMS warehouse to-be-executed task scheduling sequence ordering and constraint checking unit is connected with the scheduling strategy iteration update and optimization unit and the scheduling task feature recognition and vector generation unit through signal lines to complete task ordering and constraint checking; and the scheduling data real-time feedback and model parameter adjustment unit is connected with the warehouse parameter acquisition and instruction sending unit through a communication protocol at one end and connected with the scheduling task feature recognition and vector generation unit through a data interface at the other end to perform data feedback and model adjustment.

Images