A warehouse prediction scheduling method and system based on machine learning

Abstract

The application discloses a warehouse prediction scheduling method and system based on machine learning, relates to the technical field of logistics warehouse scheduling, and is used for in-plant logistics warehouse scheduling of RCS docking MES / WMS, receives task instruction data, generates scheduling scheme data in combination with warehouse state data, predicts task completion time distribution and congestion risk through machine learning, and solves according to a task time limit window to realize dynamic adjustment of task release, robot allocation and task priority, so that the scheduling scheme can be quickly repaired and minimum change is maintained when a task is inserted or a robot, an elevator or a channel is abnormal. Compared with fixed rules or average time scheduling, the timely completion rate can be improved, congestion chain delay and execution shock can be reduced, and manual intervention can be reduced.

Description

Technical Field

[0001] This invention relates to the field of logistics and warehousing scheduling technology, and in particular to a warehousing prediction and scheduling method and system based on machine learning. Background Technology

[0002] In in-plant logistics and warehousing scenarios where RCS integrates with MES / WMS, the control end needs to uniformly schedule AGV / AMRs to perform handling and transfer tasks across multiple work areas, aisle nodes, and cross-floor resources. Due to the time-sensitive nature and diverse sources of task instruction data, and the presence of disturbances such as task insertion, robot unavailability, aisle blockage, and elevator unavailability, task completion times fluctuate significantly, leading to localized congestion and cascading delays. Existing technologies typically employ fixed priorities, static rules, or deterministic scheduling based solely on average time. When faced with the coupling of multiple parallel tasks and traffic conflicts, it is difficult to simultaneously ensure on-time completion within the task time window and control traffic congestion. Furthermore, it is challenging to quickly and stably restore the established plan after disturbances occur, easily resulting in frequent rescheduling, execution oscillations, and a decrease in overall throughput.

[0003] Currently, Chinese invention patent application number CN202411702871.8 discloses an efficient logistics warehousing scheduling system. Its technical solution includes a warehousing scheduling management module, a data analysis and optimization module connected to the warehousing scheduling management module, an automatic scheduling module, and a warehousing safety and stability assurance module. The automatic scheduling module includes an inventory management module and an order management module. The beneficial effects of this invention are: by reducing manual operation steps through automated warehousing operation modules, it achieves fast and accurate cargo processing, significantly improving overall logistics efficiency. It can intelligently optimize transportation routes based on real-time order volume and transportation demand, reducing waiting time and empty-running rates, further improving logistics efficiency. Automated and intelligent warehousing scheduling reduces manpower requirements and lowers labor costs. Through real-time monitoring of inventory status and accurate demand forecasting, it can reduce inventory backlog and excessive replenishment, lowering inventory costs.

[0004] The aforementioned technologies are insufficient in scenarios involving multiple work areas and cross-floor access resources in RCS scheduling. When faced with disturbances such as task insertion and equipment unavailability, they are unable to generate and quickly resolve scheduling schemes based on predictable completion time risks and congestion risks, thereby stably meeting task time limit window requirements and avoiding congestion-induced delays. Summary of the Invention

[0005] The technical problem solved by this invention is that existing technologies are difficult to develop and quickly repair scheduling schemes based on predictable completion time risks and congestion risks in scenarios with coupled multi-operation areas and cross-floor access resources in RCS scheduling, when faced with disturbances such as task insertion and equipment unavailability, so as to stably meet task time limit window requirements and avoid congestion chain delays.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A machine learning-based warehouse forecasting and scheduling method includes the following steps:

[0008] Step S1: Collect task instruction data and warehouse status data, and generate task set data based on the task time limit window data in the task instruction data;

[0009] Step S2: Call historical execution data and combine it with warehouse status data to construct task feature data, train and update the task completion time distribution model online, and output task completion time distribution data;

[0010] Step S3: Construct the warehouse interaction graph data based on the warehouse status data and task set data, input the warehouse interaction graph data and task feature data into the congestion prediction model, and output the congestion prediction data.

[0011] Step S4: Generate risk budget data based on task time limit window data, and solve scheduling scheme data within a scrolling window using task completion time distribution data and congestion prediction data as constraints. The scheduling scheme data includes task release data, robot allocation data, and task priority data.

[0012] Step S5: Collect disturbance event data, and based on the scheduling scheme data and risk budget data, call the planned repair model to output repair action data and update the scheduling scheme data;

[0013] Step S6: Send the updated scheduling scheme data to the control terminal for execution, collect execution feedback data, and use the execution feedback data to update the task completion time distribution model and congestion prediction model.

[0014] Preferably, step S1 includes the following sub-steps:

[0015] Step S101: Collect task instruction data, which includes order identifier data, order detail data, service time limit data, and carrier pickup window data, and generate task time limit window data from the service time limit data and the carrier pickup window data.

[0016] Step S102: Collect warehouse status data, which includes work-in-process data, resource availability data, and channel occupancy data.

[0017] Step S103: Based on the task time limit window data, the order details data is split into tasks to generate task set data, and the task set data is written with the work area identifier corresponding to the work area quantity data.

[0018] Preferably, the generation of the risk budget data includes:

[0019] The remaining available time data is calculated based on the task time limit window data, and the remaining available time data is mapped to the timeout probability upper limit parameter and the congestion exceeding the threshold probability upper limit parameter.

[0020] The upper limit parameter of timeout probability and the upper limit parameter of congestion exceeding threshold probability are associated with the corresponding task time limit window data, and the output is risk budget data.

[0021] Preferably, step S2 includes the following sub-steps:

[0022] Step S201: Extract historical task completion time data from historical execution data and construct task feature data together with warehouse status data;

[0023] Step S202: Input the task feature data into the task completion time distribution model for training or updating, and output the task completion time distribution data, wherein the task completion time distribution data includes at least multiple quantile data of the task completion time.

[0024] Step S203: Establish a correlation between the task completion time distribution data and the corresponding task set data.

[0025] Preferably, step S3 includes the following sub-steps:

[0026] Step S301: Construct warehouse interaction graph data based on warehouse status data. The warehouse interaction graph data includes work area node data and edge relationship data formed by resource sharing and access conflicts.

[0027] Step S302: Input the in-warehouse interaction graph data and task feature data into the congestion prediction model and output congestion prediction data. The congestion prediction data includes congestion level sequence data for future time periods and corresponding over-threshold probability data.

[0028] Step S303: Establish a correlation between the congestion prediction data and the corresponding work area node data.

[0029] Preferably, step S4 includes the following sub-steps:

[0030] Step S401: Determine the rolling window boundary data based on the task time limit window data and risk budget data;

[0031] Step S402: Generate candidate task release data and candidate robot allocation data in the scrolling window, and call the task completion time distribution data corresponding to the task set data to calculate the candidate timeout risk data;

[0032] Step S403: Calculate candidate congestion risk data by calling congestion prediction data;

[0033] Step S404: With the constraints that the candidate timeout risk data does not exceed the upper limit parameter of timeout probability and the candidate congestion risk data does not exceed the upper limit parameter of congestion threshold probability, solve and output scheduling scheme data.

[0034] Preferably, step S5 includes the following sub-steps:

[0035] Step S501: Collect disturbance event data, including order surge event data, resource unavailability event data, and stockout / replenishment event data;

[0036] Step S502: Calculate scheduling deviation data based on disturbance event data, and use the scheduling deviation data and risk budget data together as input to the planned repair model;

[0037] Step S503: The planned repair model outputs repair action data, which is used to perform minimal changes to update the task release data, robot allocation data, and task priority data to generate updated scheduling scheme data.

[0038] Preferably, step S6 includes the following sub-steps:

[0039] Step S601: The control terminal executes the scheduling scheme data and collects execution feedback data.

[0040] Step S602: Compare the execution feedback data with the task completion time distribution data to generate distribution error data, and use the distribution error data to update the task completion time distribution model;

[0041] Step S603: Compare the execution feedback data with the congestion prediction data to generate congestion error data, and use the congestion error data to update the congestion prediction model.

[0042] Preferably, the execution feedback data includes task start time data, task end time data, work area inventory observation data, and channel occupancy observation data;

[0043] Task start time data and task end time data are used to update the task completion time distribution model, while work area volume observation data and channel occupancy observation data are used to update the congestion prediction model.

[0044] A machine learning-based warehouse forecasting and scheduling system includes a data acquisition and set construction module, a distributed modeling and updating module, an interactive graphing and congestion control module, a risk constraint solving module, a disturbance repair and updating module, and an execution feedback module.

[0045] The data acquisition and collection module is used to collect task instruction data and warehouse status data, and generate task set data based on the task time limit window data in the task instruction data.

[0046] The distribution modeling and update module is used to call historical execution data and combine it with warehouse status data to construct task feature data, train and update the task completion time distribution model online, and output task completion time distribution data.

[0047] The interactive graphing congestion module is used to construct in-warehouse interactive graph data based on in-warehouse status data and task set data, input the in-warehouse interactive graph data and task feature data into the congestion prediction model, and output congestion prediction data.

[0048] The risk constraint solving module is used to generate risk budget data based on task time limit window data, and solve scheduling scheme data within a scrolling window using task completion time distribution data and congestion prediction data as constraints. The scheduling scheme data includes task release data, robot allocation data, and task priority data.

[0049] The disturbance repair and update module is used to collect disturbance event data, call the planned repair model based on the scheduling scheme data and risk budget data to output repair action data and update the scheduling scheme data;

[0050] The execution feedback module is used to send the updated scheduling scheme data to the control terminal for execution, collect execution feedback data, and use the execution feedback data to update the task completion time distribution model and congestion prediction model.

[0051] The beneficial effects of this invention are as follows: This invention is used for in-plant logistics and warehousing scheduling through RCS integration with MES / WMS. It receives task instruction data and combines it with in-warehouse status data to generate scheduling scheme data. Through machine learning, it predicts the task completion time distribution and congestion risk, and performs rolling constraint solving based on task time limits. This enables dynamic adjustment of task release, robot allocation, and task priority. When tasks are inserted or robots, elevators, or passageways malfunction, the scheduling scheme can be quickly repaired with minimal changes. Compared to fixed rules or average time scheduling, this invention can improve on-time completion rates, reduce congestion-induced delays and execution fluctuations, and minimize manual intervention. Attached Figure Description

[0052] Figure 1 A flowchart illustrating the steps of a machine learning-based warehouse prediction and scheduling method according to an embodiment of the present invention;

[0053] Figure 2 This is a basic flowchart of a machine learning-based warehouse prediction and scheduling system provided in one embodiment of the present invention. Detailed Implementation

[0054] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0055] Example 1, referring to Figure 1 This embodiment is applied to an in-plant logistics and warehousing scenario where RCS is integrated with MES / WMS. The control terminal is used to uniformly schedule AGVs / AMRs to perform handling and cross-floor transfer tasks between multiple work areas and aisle nodes within the warehouse. It receives task instruction data from MES / WMS, generates task time limit window data based on task time limit data, and forms task set data. Simultaneously, it collects warehouse status data to reflect real-time status such as robot availability, work area inventory, and passage occupancy. Based on this, it outputs scheduling scheme data and sends it to the control terminal for execution. With the goal of completing tasks within the task time limit window, it dynamically adjusts task release, robot allocation, and task priority. When disturbances such as task insertion, robot unavailability, aisle blockage, or elevator unavailability occur, the scheduling scheme data is repaired and updated to ensure that scheduling continuously adapts to changes in the field and guarantees the continuity and stability of overall operations. Therefore, a machine learning-based warehouse predictive scheduling method is provided, including the following steps:

[0056] Step S1: Collect task instruction data and warehouse status data, and generate task set data based on the task time limit window data in the task instruction data.

[0057] Step S2: Call historical execution data and combine it with warehouse status data to construct task feature data, train and update the task completion time distribution model online, and output task completion time distribution data.

[0058] Step S3: Construct the warehouse interaction graph data based on the warehouse status data and task set data, input the warehouse interaction graph data and task feature data into the congestion prediction model, and output the congestion prediction data.

[0059] Step S4: Generate risk budget data based on task time limit window data. Solve the scheduling scheme data within the scrolling window using task completion time distribution data and congestion prediction data as constraints. The scheduling scheme data includes task release data, robot allocation data, and task priority data.

[0060] Step S5: Collect disturbance event data, and based on the scheduling scheme data and risk budget data, call the planned repair model to output repair action data and update the scheduling scheme data.

[0061] Step S6: Send the updated scheduling scheme data to the control terminal for execution, collect execution feedback data, and use the execution feedback data to update the task completion time distribution model and congestion prediction model.

[0062] Step S1 includes the following sub-steps:

[0063] Step S101: Collect task instruction data, which includes order identifier data, order detail data, service time limit data, and carrier pickup window data. Then, generate task time limit window data from the service time limit data and the carrier pickup window data.

[0064] In step S101, task instruction data is collected, including order identifier data, order detail data, service time limit data, and carrier pickup window data. The service time limit data and carrier pickup window data are used to generate task time limit window data. At the same time, warehouse status data is collected, including work-in-process data, resource availability data, and channel occupancy data. Resource availability data includes available personnel data, available equipment data, and equipment fault identifier data. The above task instruction data is used to generate task set data and determine the rolling window constraints. The above warehouse status data is used to construct task feature data and warehouse interaction diagram data.

[0065] Step S102: Collect warehouse status data, which includes work-in-process data, resource availability data, and channel occupancy data.

[0066] Based on the task time limit window data, order details data is split into tasks to generate task set data. The task set data includes at least picking task data, review and packing task data, and replenishment task data. For each task, work area identifier data and resource type identifier data are written; the work area identifier data is mapped from the storage location area of ​​the SKU in the order details data, and the resource type identifier data is used to generate robot allocation data later. To make order cutoff calculable, the remaining available time data is calculated for each order, and the remaining available time data satisfies the following:

[0067] ;

[0068] in, This is the time limit window data for the task corresponding to the order. The current moment;

[0069] The remaining available time data will be used to generate risk budget data and verify scheduling timeout risks.

[0070] The task set data is obtained by splitting the task instruction data, and may include cross-floor transport sub-tasks, with pre-order and post-order relationships between sub-tasks.

[0071] Step S103: Based on the task time limit window data, the order details data is split into tasks to generate task set data, and the task set data is written with the work area identifier corresponding to the work area quantity data.

[0072] The generation of risk budget data includes:

[0073] The remaining available time data is calculated based on the task time limit window data, and the remaining available time data is mapped to the timeout probability upper limit parameter and the congestion exceeding the threshold probability upper limit parameter.

[0074] The upper limit parameter of timeout probability and the upper limit parameter of congestion exceeding threshold probability are associated with the corresponding task time limit window data, and the output is risk budget data.

[0075] To characterize the current pressure within the warehouse, non-existent state-derived data is generated based on the warehouse status data and written into the task set data as task attributes. This includes work area pressure data and channel congestion data; the work area pressure data satisfies:

[0076] ;

[0077] in, The number of tasks to be executed on the site. Number of available robots;

[0078] The channel congestion data meets the following requirements:

[0079] ;

[0080] in, For channel occupancy data, This is a constant for the channel length. The aforementioned state-derived data is used for predicting the completion time distribution data of subsequent tasks and calculating congestion prediction data, thereby ensuring that the collected data is fully utilized in subsequent processing.

[0081] Step S2 includes the following sub-steps:

[0082] Step S201: Extract historical task completion time data from historical execution data and construct task feature data together with warehouse status data.

[0083] Historical execution data is retrieved from the database. This data includes historical task start time data, historical task end time data, historical work area quantity observation data, and historical channel occupancy observation data. The historical task completion time data is calculated from the historical task start time data and end time data, satisfying the following conditions:

[0084] ;

[0085] The historical task completion time data, along with the work area pressure data, channel congestion data, and order line count and SKU quantity data that can be directly statistically analyzed in the order details data, are used to construct task feature data. The task feature data includes at least work area pressure data, channel congestion data, order line count, SKU quantity, and available resource quantity.

[0086] Step S202: Input the task feature data into the task completion time distribution model for training or updating, and output the task completion time distribution data. The task completion time distribution data includes at least multiple quantile data of the task completion time.

[0087] A task completion time distribution model is established to output task completion time distribution data. The task completion time distribution model adopts a quantile learning approach, enabling it to output multiple quantile completion times for the same task under different warehouse pressures, thereby representing tail risk;

[0088] by For quantile sets, the model output is... And training is performed using Pinball loss:

[0089] ;

[0090] in .

[0091] The resulting task completion time distribution data includes at least the following: Quantile data quantile data and The data is divided into quantiles and associated with the corresponding task set data for subsequent scheduling constraints.

[0092] Step S203: Establish a correlation between the task completion time distribution data and the corresponding task set data.

[0093] After execution at the control end, the execution feedback data is obtained, and the task completion time distribution model is updated with the task start time data and task end time data in the execution feedback data. The update method is online incremental training, so that the quantile output of the same task feature data under different warehouse states is adaptively calibrated over time, thereby ensuring that the model can still predict the tail risk of completion time under sudden congestion and resource fluctuations.

[0094] Step S3 includes the following sub-steps:

[0095] Step S301: Construct warehouse interaction graph data based on warehouse status data. The warehouse interaction graph data includes work area node data and edge relationship data formed by resource sharing and access conflicts.

[0096] Based on the warehouse status data and task set data, an internal interaction graph is constructed. The internal interaction graph data includes work area node data and edge relationship data. The work area node data is determined by the work area identification data, and the node characteristics include the work area's production volume data, work area pressure data, and resource availability data. The edge relationship data includes resource sharing edge relationship data and passage conflict edge relationship data. The resource sharing edge relationship data is formed by work area nodes with the same resource type identification data, and the passage conflict edge relationship data is formed by adjacent work areas corresponding to the passage occupancy data.

[0097] The work area node can correspond to a station, road segment, or elevator entrance. The work area quantity data is calculated from the waiting queue length.

[0098] ;

[0099] in, For a moment At the work area node The number of tasks in a waiting state is obtained from the task queue record on the control end.

[0100] Step S302: Input the in-warehouse interaction graph data and task feature data into the congestion prediction model and output congestion prediction data. The congestion prediction data includes the congestion level sequence data for future time periods and the corresponding over-threshold probability data.

[0101] A congestion prediction model is established, and incorporating in-warehouse interaction graph data and task characteristic data, it outputs congestion prediction data. The congestion prediction model employs a graph time-series prediction structure, outputting future congestion level sequence data for each work area node; node state updates satisfy the following:

[0102] ;

[0103] And Mapped to congestion level:

[0104] ;

[0105] Congestion prediction data includes at least future... Congestion level sequence data for each time period and congestion exceeding the threshold probability data ,in The congestion threshold is a constant, and the probability data of congestion exceeding the threshold is used for subsequent risk budget constraints.

[0106] Step S303: Establish a correlation between the congestion prediction data and the corresponding work area node data.

[0107] The congestion prediction data is correlated with the node data of the work area, and the impact of candidate wave releases on congestion is further calculated as non-existent data for subsequent scheduling scheme selection; the impact is calculated as follows:

[0108] ;

[0109] in, Release tasks for candidates. It is obtained by weighting the number of order lines and the number of SKUs in the task feature data, so that how many tasks are released into which area can be quantified and participate in congestion constraints.

[0110] Step S4 includes the following sub-steps:

[0111] Step S401: Determine the rolling window boundary data based on the task time limit window data and risk budget data.

[0112] Risk budget data is generated based on task time limit window data. This risk budget data includes at least the upper limit parameters for timeout probability and congestion exceeding the threshold probability. The average remaining available time is calculated for each cut-off window. The average remaining available time of the window is mapped to a timeout probability upper limit parameter, which in this embodiment satisfies the following:

[0113] ;

[0114] in, It is a scale constant;

[0115] At the same time, a maximum parameter for the probability of congestion exceeding the threshold is set for each work area. and will , Establish a link with the corresponding task time limit window data and work area node data, and output risk budget data.

[0116] Step S402: Generate candidate task release data and candidate robot allocation data in the scrolling window, and call the task completion time distribution data corresponding to the task set data to calculate candidate timeout risk data.

[0117] Within a scrolling window, candidate task release data, candidate robot allocation data, and candidate task priority data are generated. Candidate timeout risk data is calculated based on task completion time distribution data. For a set of tasks belonging to the same order, the order completion time quantile is conservatively estimated using the 95th percentile completion time of the critical path, satisfying the following conditions:

[0118] ;

[0119] in, This is the critical path of the task chain, and the timeout is calculated based on the remaining available time data:

[0120] like Then the candidate solution is for the order This creates a timeout risk; the percentage of timeout orders within the scrolling window is used as candidate timeout risk data, and compared with the upper limit parameter of timeout probability. Compare them.

[0121] Step S403: Calculate candidate congestion risk data by calling congestion prediction data.

[0122] Candidate congestion risk data is calculated based on congestion prediction data. This candidate congestion risk data includes the aggregated probability of congestion exceeding a threshold for each work area within a rolling window, satisfying the following:

[0123] ;

[0124] When candidate task releases data, it causes When it increases, it is superimposed on The probability is calculated again in the prediction, so that the candidate wave release plan and the congestion risk are constrained by the same indicator.

[0125] Step S404: With the constraints that the candidate timeout risk data does not exceed the upper limit parameter of timeout probability and the candidate congestion risk data does not exceed the upper limit parameter of congestion threshold probability, solve and output scheduling scheme data.

[0126] Solve for the output scheduling scheme data under constraints; the constraints should include at least the following: the number of candidate timeout risk data within the rolling window does not exceed the upper limit parameter of timeout probability. Furthermore, the probability of congestion exceeding the threshold in each work area of ​​the candidate congestion risk data does not exceed the upper limit parameter of the probability of congestion exceeding the threshold. .

[0127] When the constraints are satisfied, the scheduling scheme data is selected based on the objective function, which includes at least the expected total completion time and congestion penalty, and satisfies the following:

[0128] ;

[0129] This outputs task release data, robot allocation data, and task priority data.

[0130] Step S5 includes the following sub-steps:

[0131] Step S501: Collect disturbance event data, which includes order surge event data, resource unavailability event data, and stockout / replenishment event data.

[0132] Collect disturbance event data, including order surge event data, resource unavailability event data, and stockout / replenishment event data, and calculate scheduling deviation data based on the disturbance event data. The scheduling deviation data includes at least new task volume data, resource capacity loss data, and replenishment queue data, and is used to recalculate work area pressure data and channel congestion data, thereby driving changes in task characteristic data and congestion prediction data.

[0133] Step S502: Calculate scheduling deviation data based on disturbance event data, and use the scheduling deviation data and risk budget data together as input to the planned repair model.

[0134] The planned repair model outputs repair action data. The planned repair model, adhering to the principle of minimizing changes to restore feasible behavior, performs partial updates to the scheduling scheme data. The system defines scheduling change cost data.

[0135] ;

[0136] in, Release data for the task. Assign data to the robot. This is task priority data. , and For the repaired data, , and These are the weighting coefficients.

[0137] The planned repair model minimizes the change cost data while meeting the risk budget data constraints, thereby outputting repair action data and updating the scheduling plan data; the repair action data includes at least wave delay action data, task reassignment action data, and replenishment priority adjustment action data.

[0138] Step S503: The planned repair model outputs repair action data. The repair action data is used to perform minimal changes to update the task release data, robot allocation data, and task priority data, generating updated scheduling scheme data.

[0139] The updated scheduling scheme data is used to recalculate candidate timeout risk data and candidate congestion risk data as repair and verification data; if the repair and verification data still exceed the risk budget data, then the risk level is increased. The corresponding wave delay weights are applied and the planned repair model is called back, so that the system prioritizes reducing the release intensity to reduce congestion and timeout risks, thereby ensuring the stability of the repair action data.

[0140] Step S6 includes the following sub-steps:

[0141] Step S601: The control terminal executes the scheduling scheme data and collects execution feedback data.

[0142] The updated scheduling plan data is sent to the control terminal for execution. The control terminal releases tasks according to the task release data and assigns personnel or equipment to perform tasks according to the robot allocation data. Execution feedback data is collected synchronously, including task start time data, task end time data, work area inventory observation data, and channel occupancy observation data. Among them, the work area inventory observation data and channel occupancy observation data are used to supervise and update the congestion prediction model.

[0143] Step S602: Compare the execution feedback data with the task completion time distribution data to generate distribution error data, and use the distribution error data to update the task completion time distribution model.

[0144] The task completion time distribution model is updated based on execution feedback data, with the actual completion time and model percentiles used in the update process. Output the calculated distribution error data:

[0145] ;

[0146] The model parameters are updated online according to the gradient of the Pinball loss, making the output of each quantile more sensitive to tail timeouts, thereby calculating candidate timeout risk data more accurately in the next rolling window.

[0147] Step S603: Compare the execution feedback data with the congestion prediction data to generate congestion error data, and use the congestion error data to update the congestion prediction model.

[0148] The execution feedback data includes task start time data, task end time data, work area quantity observation data, and channel occupancy observation data.

[0149] Task start time data and task end time data are used to update the task completion time distribution model, while work area volume observation data and channel occupancy observation data are used to update the congestion prediction model.

[0150] The congestion prediction model is updated based on execution feedback data. During the update, actual congestion observation data is formed by combining the operational area's volume observation data and the lane occupancy observation data. and with The difference between the data and the predicted congestion level sequence data forms the congestion error data. The congestion error data is used to update the model parameters in reverse, so that the congestion exceeding the threshold probability data is closer to the actual congestion evolution.

[0151] Example 2, refer to Figure 2 This paper presents a machine learning-based warehouse forecasting and scheduling system, which includes a data acquisition and set construction module, a distributed modeling and updating module, an interactive graphing and congestion module, a risk constraint solving module, a disturbance repair and updating module, and an execution feedback module.

[0152] The data acquisition and collection module is used to collect task instruction data and warehouse status data, and generate task collection data based on the task time limit window data in the task instruction data.

[0153] The distributed modeling and update module is used to call historical execution data and combine it with warehouse status data to construct task feature data, train and update the task completion time distribution model online, and output task completion time distribution data.

[0154] The interactive graph congestion module is used to construct in-warehouse interactive graph data based on in-warehouse status data and task set data. The in-warehouse interactive graph data and task feature data are input into the congestion prediction model, and the congestion prediction data is output.

[0155] The risk constraint solving module is used to generate risk budget data based on task time limit window data. Within the scrolling window, it solves the scheduling scheme data with task completion time distribution data and congestion prediction data as constraints. The scheduling scheme data includes task release data, robot allocation data, and task priority data.

[0156] The disturbance repair and update module is used to collect disturbance event data, and based on the scheduling scheme data and risk budget data, it calls the planned repair model to output repair action data and update the scheduling scheme data.

[0157] The execution feedback module is used to send the updated scheduling scheme data to the control terminal for execution, collect execution feedback data, and use the execution feedback data to update the task completion time distribution model and congestion prediction model.

[0158] This invention targets in-plant logistics and warehousing applications that integrate RCS with MES / WMS. It generates task set data from task instruction data and combines this with warehouse status data to obtain real-time on-site status, enabling scheduling decisions to reflect changes in robot availability, work area load, and traffic occupancy. Building upon this, the invention utilizes machine learning to output task completion time distribution data and predict congestion, freeing scheduling from reliance on average completion times or fixed rules, and allowing for quantifiable constraints on on-time completion and congestion risk. Furthermore, it generates risk budget data based on task time limit window data, and solves for scheduling scheme data including task release, robot allocation, and task priority within a rolling window, enabling the scheduling scheme to suppress time-limit violations and congestion spread before execution. For disturbances such as task insertion, robot unavailability, passageway blockage, and elevator unavailability, the invention uses a planned repair model to output repair action data, updating the scheduling scheme data with minimal modifications, avoiding execution oscillations caused by full reordering, and improving the continuity and stability of on-site scheduling. Therefore, this invention can improve the on-time completion rate of tasks in complex traffic control and multi-robot collaborative operation environments, reduce the probability of chain delays and stagnation caused by congestion, reduce the cost of manual intervention and repeated parameter tuning, and improve the interpretability and traceability of scheduling strategies. It is applicable to multi-floor, multi-task concurrent in-plant logistics and warehousing systems.

[0159] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0160] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A warehouse forecasting and scheduling method based on machine learning, characterized in that, Includes the following steps: Step S1: Collect task instruction data and warehouse status data, and generate task set data based on the task time limit window data in the task instruction data; Step S2: Call historical execution data and combine it with warehouse status data to construct task feature data, train and update the task completion time distribution model online, and output task completion time distribution data; Step S3: Construct the warehouse interaction graph data based on the warehouse status data and task set data, input the warehouse interaction graph data and task feature data into the congestion prediction model, and output the congestion prediction data. Step S4: Generate risk budget data based on task time limit window data, and solve scheduling scheme data within a scrolling window using task completion time distribution data and congestion prediction data as constraints. The scheduling scheme data includes task release data, robot allocation data, and task priority data. Step S5: Collect disturbance event data, and based on the scheduling scheme data and risk budget data, call the planned repair model to output repair action data and update the scheduling scheme data; Step S6: Send the updated scheduling scheme data to the control terminal for execution, collect execution feedback data, and use the execution feedback data to update the task completion time distribution model and congestion prediction model.

2. The warehouse forecasting and scheduling method based on machine learning as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S101: Collect task instruction data, which includes order identifier data, order detail data, service time limit data, and carrier pickup window data, and generate task time limit window data from the service time limit data and the carrier pickup window data. Step S102: Collect warehouse status data, which includes work-in-process data, resource availability data, and channel occupancy data. Step S103: Based on the task time limit window data, the order details data is split into tasks to generate task set data, and the task set data is written with the work area identifier corresponding to the work area quantity data.

3. The warehouse forecasting and scheduling method based on machine learning as described in claim 2, characterized in that, The generation of the risk budget data includes: The remaining available time data is calculated based on the task time limit window data, and the remaining available time data is mapped to the timeout probability upper limit parameter and the congestion exceeding the threshold probability upper limit parameter. The upper limit parameter of timeout probability and the upper limit parameter of congestion exceeding threshold probability are associated with the corresponding task time limit window data, and the output is risk budget data.

4. The warehouse forecasting and scheduling method based on machine learning as described in claim 1, characterized in that, Step S2 includes the following sub-steps: Step S201: Extract historical task completion time data from historical execution data and construct task feature data together with warehouse status data; Step S202: Input the task feature data into the task completion time distribution model for training or updating, and output the task completion time distribution data, wherein the task completion time distribution data includes at least multiple quantile data of the task completion time. Step S203: Establish a correlation between the task completion time distribution data and the corresponding task set data.

5. The warehouse forecasting and scheduling method based on machine learning as described in claim 1, characterized in that, Step S3 includes the following sub-steps: Step S301: Construct warehouse interaction graph data based on warehouse status data. The warehouse interaction graph data includes work area node data and edge relationship data formed by resource sharing and access conflicts. Step S302: Input the in-warehouse interaction graph data and task feature data into the congestion prediction model and output congestion prediction data. The congestion prediction data includes congestion level sequence data for future time periods and corresponding over-threshold probability data. Step S303: Establish a correlation between the congestion prediction data and the corresponding work area node data.

6. The warehouse forecasting and scheduling method based on machine learning as described in claim 1, characterized in that, Step S4 includes the following sub-steps: Step S401: Determine the rolling window boundary data based on the task time limit window data and risk budget data; Step S402: Generate candidate task release data and candidate robot allocation data in the scrolling window, and call the task completion time distribution data corresponding to the task set data to calculate the candidate timeout risk data; Step S403: Calculate candidate congestion risk data by calling congestion prediction data; Step S404: With the constraints that the candidate timeout risk data does not exceed the upper limit parameter of timeout probability and the candidate congestion risk data does not exceed the upper limit parameter of congestion threshold probability, solve and output scheduling scheme data.

7. The warehouse forecasting and scheduling method based on machine learning as described in claim 1, characterized in that, Step S5 includes the following sub-steps: Step S501: Collect disturbance event data, including order surge event data, resource unavailability event data, and stockout / replenishment event data; Step S502: Calculate scheduling deviation data based on disturbance event data, and use the scheduling deviation data and risk budget data together as input to the planned repair model; Step S503: The planned repair model outputs repair action data, which is used to perform minimal changes to update the task release data, robot allocation data, and task priority data to generate updated scheduling scheme data.

8. The warehouse forecasting and scheduling method based on machine learning as described in claim 1, characterized in that, Step S6 includes the following sub-steps: Step S601: The control terminal executes the scheduling scheme data and collects execution feedback data. Step S602: Compare the execution feedback data with the task completion time distribution data to generate distribution error data, and use the distribution error data to update the task completion time distribution model; Step S603: Compare the execution feedback data with the congestion prediction data to generate congestion error data, and use the congestion error data to update the congestion prediction model.

9. The warehouse forecasting and scheduling method based on machine learning as described in claim 8, characterized in that, The execution feedback data includes task start time data, task end time data, work area production monitoring data, and channel occupancy monitoring data; Task start time data and task end time data are used to update the task completion time distribution model, while work area volume observation data and channel occupancy observation data are used to update the congestion prediction model.

10. A machine learning-based warehouse forecasting and scheduling system, applied in any one of the machine learning-based warehouse forecasting and scheduling methods described in claims 1-9, characterized in that, It includes a data acquisition and set building module, a distributed modeling and updating module, an interactive graphing and congestion control module, a risk constraint solving module, a disturbance repair and updating module, and a distribution, execution, and feedback module; The data acquisition and collection module is used to collect task instruction data and warehouse status data, and generate task set data based on the task time limit window data in the task instruction data. The distribution modeling and update module is used to call historical execution data and combine it with warehouse status data to construct task feature data, train and update the task completion time distribution model online, and output task completion time distribution data. The interactive graphing congestion module is used to construct in-warehouse interactive graph data based on in-warehouse status data and task set data, input the in-warehouse interactive graph data and task feature data into the congestion prediction model, and output congestion prediction data. The risk constraint solving module is used to generate risk budget data based on task time limit window data, and solve scheduling scheme data within a scrolling window using task completion time distribution data and congestion prediction data as constraints. The scheduling scheme data includes task release data, robot allocation data, and task priority data. The disturbance repair and update module is used to collect disturbance event data, call the planned repair model based on the scheduling scheme data and risk budget data to output repair action data and update the scheduling scheme data; The execution feedback module is used to send the updated scheduling scheme data to the control terminal for execution, collect execution feedback data, and use the execution feedback data to update the task completion time distribution model and congestion prediction model.

Images