Robot task dynamic scheduling optimization method of annular warehousing system

Abstract

The invention discloses a robot task dynamic scheduling optimization method for an annular warehousing system, and particularly relates to the technical field of intelligent warehousing system scheduling. Task basic information is collected to construct a task state table, and robot operation states are collected to construct a robot state table R; constructing a task-robot matching graph based on T and R, and calculating a matching priority score by adopting a graph neural mechanism; generating an initial scheduling scheme according to a game optimization and state driving strategy; path conflict detection is carried out in combination with a ring topology structure and traffic rules, and a task execution path is optimized through a sliding time window and a local disturbance rearrangement algorithm; a scheduling simulation engine is constructed to perform performance evaluation, if an evaluation result does not reach a preset threshold value, structure reconstruction is performed on the scheduling graph based on a residual feedback mechanism, and the method has the advantages of being high in real-time performance, low in conflict rate and high in resource utilization efficiency and is suitable for high-reliability intelligent scheduling in a multi-robot intensive task scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent warehouse system scheduling, in particular to a robot task dynamic scheduling optimization method for a ring-shaped warehouse system. BACKGROUND

[0002] With the rapid development of intelligent manufacturing and smart logistics, automated warehouse systems are widely used in e-commerce, manufacturing, pharmaceutical and other fields. Among them, the ring-shaped warehouse system has become an important development direction of the new generation of automated three-dimensional warehouse due to its compact structure, closed path, high space utilization rate and other advantages. In this system, multiple mobile robots (such as shuttles, AGVs, etc.) need to perform tasks such as carrying and picking on the ring-shaped track, and the task intensity is high and the path conflict probability is large. If the scheduling strategy is not reasonable, it is easy to cause task delay, resource congestion, and even cause the system to malfunction.

[0003] At present, traditional robot scheduling methods mostly use static or semi-static rules, such as "shortest path first", "nearest response" and the like. These methods can maintain basic scheduling efficiency when the task quantity is small, but in a dynamic, multi-task, high-density running scene, problems such as uneven load, path congestion, scheduling conflict, etc. are prone to occur, which seriously affects the overall throughput capacity and response speed. Especially in the ring-shaped structure, the path selection is low, and the bottleneck area is easy to produce task backlog, affecting the smooth running of the system. Therefore, there is an urgent need for a method that can adjust the scheduling strategy according to the dynamic changes of the task and the real-time state of the system, to realize the efficient coordination and load balancing of the robot resources, and to improve the overall running efficiency and stability of the system. SUMMARY

[0004] The purpose of the present application is to provide a robot task dynamic scheduling optimization method for a ring-shaped warehouse system to solve the problems in the background art.

[0005] In order to achieve the above purpose, the present application provides the following technical scheme: a robot task dynamic scheduling optimization method for a ring-shaped warehouse system, comprising: S100, obtaining the basic information of all tasks to be executed in the current system, including task position, task type, priority and estimated processing time, and constructing a task state table T; S200, obtaining the real-time running state of each robot in the system, including the current position, load state, power level and historical path information, and constructing a robot state table R; S300, based on T and R, constructing a task-robot matching graph G, using a weighted matching strategy combining task priority, distance cost and load balancing factor to calculate the matching priority score of each task and robot; S400, performing a preliminary allocation on the matching graph G to generate an initial scheduling scheme; S500, performing path conflict detection on the initial scheduling scheme based on the ring path topology graph and the traffic rules, and marking potential conflict nodes; S600, for the task path with conflict, performing dynamic adjustment on the path execution order and scheduling time by using a time window sliding rearrangement algorithm, and generating an optimized conflict-free scheduling scheme; S700, simulating the task execution order of each robot in the conflict-free scheduling scheme, and performing performance evaluation according to the task completion rate, path conflict rate and average response time fed back by system simulation; S800, if the evaluation index does not meet the preset optimization threshold, iteratively updating steps S300 to S700 until the system performance requirement is met, and outputting the final task scheduling optimization scheme.

[0006] Preferably, the S300 comprises: S301, taking the task nodes in the task state table T and the robot nodes in the robot state table R as graph vertices, and constructing a heterogeneous bidirectional matching graph G according to the coupling relationship of geographical distance, energy consumption overhead and task type; S302, performing feature propagation on the graph G by using an adaptive graph convolution attention network, fusing a task priority vector, a robot load margin vector and a path overlap factor, and generating a multi-dimensional matching score tensor of task-robot pairs; S303, reordering the score tensor based on a multi-objective normalized scoring mechanism, and finally determining a matching priority score set of each task and candidate robot.

[0007] Preferably, the S400 comprises: S401, based on the priority score set obtained in the task-robot matching graph G, constructing a two-sided matching game model, taking the minimization of robot energy consumption cost and the minimization of task response delay as dual objective functions, and generating a feasible matching candidate set F; S402, introducing a dynamic compression screening mechanism to forcibly filter the allocation scheme in set F with high conflict degree or in boundary state, and forming a controllable subset F'; S403, using a state-driven distributed scheduling strategy to locally select the candidate matching pairs in F', and redistributing them according to the regional congestion threshold, self-adjusting path density and task window prediction function, and finally outputting the initial scheduling scheme.

[0008] Preferably, the S500 comprises: S501, embedding all robot paths in the initial scheduling scheme into a ring path topology graph structure, constructing a multi-path synchronous scheduling encoding matrix, and the matrix elements representing the node timing and overlapping weight occupied by the robot within any two time slices; S502, based on the multi-dimensional path prediction tensor and the traffic rule modeling graph, a node conflict recognition function is established to jointly determine the node competition priority, path intersection directionality and node occupation overlap rate; S503, using a reverse conflict deduction mechanism, a scheduling coding matrix within a prediction window is analyzed in time, potential conflict sources are traced back level by level, and a potential conflict node set with a conflict risk level higher than a preset threshold is marked on a topology graph.

[0009] Preferably, the S600 comprises: S601, for the identified potential conflict node set, a scheduling interference graph based on a sliding time window is established, wherein the edges represent the path mutual exclusion relationship between the time window overlapping tasks; S602, by applying controllable perturbations to the robot task sequence where the conflict paths are located, including sequential interchanging, delay offset and task segment splitting, a plurality of local feasible path combination candidates are generated; S603, the candidate path combination is fitted to a dynamic scheduling graph, combined with the current load index and node blocking probability of the system, a multi-objective evaluation function is used to select the reconstruction scheme with the smallest conflict, and an optimized conflict-free scheduling scheme is output.

[0010] Preferably, the S700 comprises: S701, based on the task execution timing and the path mapping relationship, a virtual scheduling simulation engine is constructed, the simulation engine takes the conflict-free scheduling scheme as input, generates a robot behavior trajectory sequence and a task execution shadow record under the condition of not interfering with the real system operation; S702, a task behavior dynamic modeler is introduced, the path travel state, task response time and execution efficiency of each robot in the simulation are periodically sampled, and a multi-dimensional performance feedback vector is generated; S703, the performance feedback vector is input into a multi-index evaluation module, cross-weighted analysis is performed on the task completion rate, path conflict rate and average response time, a performance evaluation tensor is constructed, and a comparison judgment is performed with a preset system performance threshold.

[0011] Preferably, the S800 comprises: S801, the generated performance evaluation tensor is normalized, a key performance deviation vector is extracted, and input into a scheduling scheme correction network based on residual driving, to identify the current scheduling bottleneck type and the corresponding responsible path segment; S802, based on the bottleneck feedback result, the task-robot matching graph G is locally reconstructed in terms of graph structure, including edge weight adjustment, path feasibility re-estimation and node selection mask update, to generate a new round of iteration scheduling graph G'; S803, re-executing steps S300 to S700 on the updated graph G' to dynamically adjust the task allocation weight, path execution order and time window configuration through multiple rounds of adaptive iteration until all evaluation indicators meet the preset optimization threshold, and output the final task scheduling optimization scheme.

[0012] In the above technical solution, the present application provides technical effects and advantages: 1. The present application introduces a task state table and a robot state table to construct a heterogeneous matching graph, and combines a graph attention network, an adaptive sliding time window and a bottleneck identification mechanism to achieve high-precision matching, conflict prediction and path rearrangement of multi-robot tasks. This method has end-to-end closed-loop optimization capability, can dynamically perceive system state changes, and adjust the scheduling strategy in real time, thereby significantly improving the task completion rate, reducing the path conflict rate, and ensuring the system operation stability and scheduling response efficiency.

[0013] 2. Compared with the existing scheduling method based on static priority or heuristic rules, the present application uses deep graph structure modeling and multi-round performance feedback driving mechanism to break through the technical bottleneck that traditional methods cannot cope with high-density task dynamic conflicts. Experiments show that the present application can reduce the path conflict rate to less than 3% in complex task scenarios, and reduce the average response time by more than 30%, effectively improving the system throughput and resource usage balance, and has wide industrial application prospects. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings described below are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0015] Figure 1 The method mind map of the present application. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0017] Embodiment 1, please refer to Figure 1 As shown in the figure, the robot task dynamic scheduling optimization method of the ring warehouse system described in this embodiment comprises: S100, acquire basic information of all tasks to be executed in the current system, including task position, task type, priority and expected processing time, and construct a task state table T; S200, acquire real-time running state of each robot in the system, including current position, load state, power level and historical path information, and construct a robot state table R; S300, based on T and R, construct a task-robot matching graph G, adopt a weighted matching strategy to combine task priority, distance cost and load balancing factor, and calculate matching priority score of each task and robot; S400, perform initial allocation on the matching graph G to generate an initial scheduling scheme; S500, based on the ring path topology graph and traffic rules, perform path conflict detection on the initial scheduling scheme, and mark potential conflict nodes; S600, for the task path with conflict, adopt a time window sliding rearrangement algorithm to dynamically adjust the execution order and scheduling time of the path, and generate an optimized conflict-free scheduling scheme; S700, simulate the task execution order of each robot in the conflict-free scheduling scheme, and perform performance evaluation according to the task completion rate, path conflict rate and average response time fed back by the system simulation; S800, if the evaluation index does not meet the preset optimization threshold, iteratively update steps S300 to S700 until the system performance requirement is met, and output the final task scheduling optimization scheme.

[0018] To solve the problems of insufficient task and state modeling and poor dynamic adaptability of robot scheduling in the prior art, the present application proposes a dynamic scheduling optimization method based on task state table and robot state table, and the technical solutions of S100 and S200 are further described below in combination with specific steps.

[0019] In step S100, in order to ensure the comprehensiveness and real-time of task information, first of all, the periodic state acquisition of each work node in the ring warehouse system is implemented through the fusion perception unit and the task generation module to form structured raw task data. The fusion perception unit integrates visual sensor, RFID reader and industrial Ethernet communication module, which can perceive the real-time instruction number of the buffer queue in the station, the current trigger source (such as manual input, order system request, inventory replenishment event, etc.) and its priority weight. The task generation module standardizes the collected data in format to ensure consistent subsequent processing interface.

[0020] Subsequently, the original task data is embedded by using a task feature extraction network constructed by a graph attention mechanism. The network adopts a node-edge joint attention mechanism, automatically learns the context dependency relationship between tasks, extracts task embedding features, forms a timestamped task feature vector sequence, and supports the modeling needs of time series tasks. To further optimize the explainability and scheduling relevance of the features, a time-space decoupling encoder is used to partition and reconstruct the above vector sequence, to identify the time sequence priority relationship and spatial scheduling accessibility of the task. Finally, the processing result is constructed as a task state atlas and mapped to generate a task state table T, which is used to support subsequent scheduling graph construction and allocation algorithm execution.

[0021] In step S200, a lightweight perception agent is deployed on the edge side in the system to obtain high-precision perception data of the running state of each robot. The perception agent adopts a combination of edge AI units and cache nodes, asynchronously collects the current position, load mass, battery state and current task sequence of the path cache area of each robot. The collected information is uploaded in real time through the internal CAN communication bus.

[0022] After collection is completed, a state recognition model based on Bayesian dynamic clustering is used to label the robot state data. The model can determine the robot running state level (such as empty, high load, critical battery level, etc.) and its load variation trend according to the historical state transition mode and current parameter distribution, and provide a basis for subsequent scheduling priority setting.

[0023] Further, by analyzing the historical path sequence of the robot and the congestion probability of different areas in the ring-shaped warehouse system, a time-state joint representation model is constructed. The model considers the historical behavior pattern of the robot and the environmental interference factor to quantify the scheduling availability, remaining energy margin and task bearing potential of the robot in the current period. Finally, the above multi-source state information is integrated to generate a robot state table R, which provides a structured decision basis for task allocation and path control strategy.

[0024] In the scheduling optimization method of the application, the core of step S300 is to construct a task-robot matching graph G and calculate a matching priority score to support subsequent scheduling allocation and conflict avoidance. Since the path resources in the ring-shaped warehouse system are limited, the robot scheduling needs to consider the urgency of task demand, robot running ability, path resource state and scheduling cost, therefore, this step comprehensively introduces a heterogeneous graph modeling, an adaptive graph neural processing mechanism and a multi-objective normalization evaluation strategy to realize an efficient matching method with global perception ability and real-time adjustment ability.

[0025] First, in S301, a task-robot matching graph G is constructed based on the task state table T and the robot state table R. Specifically, each to-be-assigned task in the task state table T is abstracted as a task node, each task node carrying its location coordinates, task type code, priority label, and time sensitivity weight; each robot in the robot state table R is abstracted as a robot node, carrying its current spatial position, load state, energy margin, and congestion area exposure frequency. Subsequently, according to the following three association rules, a heterogeneous bidirectional edge relationship is established between the task nodes and the robot nodes to form an initial graph structure: Geographical distance association: the Euclidean distance or the number of hops on the path graph between the current position of the robot and the task occurrence position is taken as the distance weight; Energy consumption overhead factor: the energy consumption overhead required for executing the task is estimated with reference to the current load mass and the remaining power level of the robot; Task type coupling: specific matching preference rules are set for specific task types (such as urgent single delivery, cold chain transportation, etc.), and robots with corresponding adaptive capabilities are preferentially selected.

[0026] The matching graph G after construction is a heterogeneous bidirectional weighted graph, and the edge weights reflect the matching cost and coupling degree between tasks and robots. This graph structure can provide a relationship basis for subsequent deep feature extraction.

[0027] In S302, to improve the expression ability and robustness of the matching calculation, an adaptive graph convolution attention network (AGCAN) is introduced to perform feature propagation on the graph G. The AGCAN model is based on the structure design of graph neural network, and introduces an attention mechanism to dynamically adjust the weight of information transmission between task and robot nodes on the basis of preserving the basic unit of graph convolution.

[0028] Specifically, the input of the AGCAN model is composed of a task node feature matrix and a robot node feature matrix, which respectively include: Task priority vector: reflecting the urgency of the task; Robot load margin vector: representing the assignable ability of the robot before executing a new task; Path overlap factor matrix: calculated from the path layout in the current scheduling graph, used to reflect the possible shared resource conflict risk between two paths.

[0029] The graph convolution unit performs multi-layer feature update on the nodes in the graph G. Each layer realizes feature aggregation and transfer through edge weight adjustment and node embedding merging strategy, while the attention mechanism dynamically allocates attention coefficients between adjacent nodes at each hop to enhance the representation ability of key task-robot pairs.

[0030] After multiple rounds of convolution propagation, the model outputs a task-robot multi-dimensional matching score tensor, each element of which represents the matching degree of a certain task and a certain robot under the current context. The tensor structure can be two-dimensional (task x robot) or three-dimensional (task x robot x index dimension), adjusted according to system configuration.

[0031] In S303, in order to convert the above high-dimensional score information into a structured result that can be used for scheduling decision, a multi-objective normalized scoring mechanism is introduced to reprocess the score tensor. The scoring mechanism is completed through the following process: First, the scores of all matching pairs are normalized to make the indicators of different dimensions (such as priority, path distance, energy consumption, etc.) in a unified comparison scale; Then, a weighted combination strategy is used to make a comprehensive score of each dimension. The weighted combination strategy supports adaptive learning, that is, the weight distribution can be dynamically adjusted according to the historical scheduling feedback of the system (such as penalizing energy consumption exceeding the limit or path conflict); Finally, the normalized scores are sorted to construct a matching priority score set between tasks and robots. The set takes the task as the primary key, lists its top N candidate robots and the corresponding matching scores, which are used as the input basis for subsequent initial allocation (S400).

[0032] The core technical advantage of this step is the integration of heterogeneous graph modeling and the dynamic adaptive ability of deep neural processing, which has the following significant technical effects compared to traditional methods based on static rules (such as shortest path first, idle first): It can dynamically generate priority matching according to the real-time state of the task and the running ability of the robot. By considering path overlap and load margin, it realizes multi-dimensional balanced matching and reduces scheduling conflicts. The task-robot matching tensor can be quickly evaluated in parallel under high task load, ensuring scheduling efficiency.

[0033] In the ring-shaped warehouse system described in the present application, the robot resource scheduling faces typical challenges such as high density, multiple tasks, and serious path overlap. In particular, in the initial matching stage of tasks and robots, how to ensure global optimal performance while considering system executability and path conflict avoidance ability is a core problem that affects the quality of the scheduling scheme. Therefore, in step S400, the present application proposes a dynamic initial scheduling generation method based on the combination of game model, compression screening mechanism and state-driven redistribution strategy, which constructs a more flexible and operable scheduling scheme.

[0034] Firstly, in step S401, a two-sided matching game model for initial task allocation is constructed based on the task-robot matching graph G and its corresponding priority score set constructed in step S300. The model takes the task set and the robot set as the game participants, and the modeling goal is to achieve adaptive equilibrium of the scheduling strategy by balancing the interests of both sides under the many-to-many matching relationship.

[0035] In this model, two dual objective functions are defined: Objective 1: Minimization of robot energy consumption cost, which includes the path cost required for the robot to move from the current location to the task starting point, the energy consumption estimation of the return path after task completion, and the task load-related execution energy overhead; Objective 2: Minimization of task response delay, which evaluates the total waiting time from task generation to actual scheduled execution, taking into account task priority, deadline, and task window urgency.

[0036] By solving this two-objective game model, the system can construct a feasible matching candidate set F, which contains several task-robot matching combinations that have execution feasibility, reasonable priority weight, and controllable resource scheduling conflict risk under the current state, each combination meeting or approaching the optimal energy-delay balance.

[0037] To improve computational efficiency, the matching model uses linear relaxation and Lagrange multiplier method to reconstruct the objective functions during construction, and uses an alternating optimization strategy to obtain a stable solution within a limited number of rounds.

[0038] However, due to potential path conflicts and resource deadlocks caused by graph structure overlap, node mutual exclusion, or area congestion in the candidate set F, a dynamic compression filtering mechanism is introduced in S402 to further filter the set.

[0039] The compression filtering mechanism includes the following core operations: Conflict identification algorithm: based on the path overlap graph and the resource mutual exclusion matrix, a conflict score factor is introduced for each matching pair in F, which is calculated according to the path pre-occupation length, the shared segment length with the allocated task, and the conflict window overlap, quantifying the severity of the conflict; Boundary state detection: identify robots in critical state (such as low battery, close to task saturation, path cache full, etc.), and mark the matching pairs they participate in; Forced filtering logic: hard constraint elimination is performed on matching schemes with conflict degree exceeding the threshold or obvious boundary state, to ensure that the final output scheme meets the system safety and execution stability requirements.

[0040] After screening, the generated subset F' is a candidate scheduling scheme with stable structure and controllable conflict risk, providing a good foundation for the next step of fine-grained allocation.

[0041] In the S403 step, for the subset F' formed after screening, the application further introduces a state-driven distributed scheduling strategy to improve the adaptability of the scheduling strategy in the actual complex operating environment.

[0042] The strategy takes the system operating state as the driving variable and combines the following three types of indicators for distributed reallocation: Regional congestion threshold awareness: the system divides the ring path into several logical blocks, sets congestion awareness thresholds for each block, and dynamically adjusts the start time and robot intervention time window of the task in the scheduling scheme according to the paragraph distribution density of the candidate path; Self-adjusting path density analyzer: dynamically scan the path resources, establish a path resource density map, implement a weight penalty mechanism for hot path segments, guide the task scheduling to tilt towards the idle resource area, and achieve load balancing effect; Task window prediction function: based on historical task response data and current task trend, predict the optimal execution window of each task, and reorder or postpone the matching pairs in the scheduling scheme that deviate from the predicted window.

[0043] Finally, the task-robot matching relationship optimized by the above adaptive distributed scheduling strategy is converged to the initial scheduling scheme The scheme not only avoids high-risk path conflicts in the early stage of scheduling, but also has good energy distribution balance and task response consistency.

[0044] The application introduces a three-stage processing logic of game optimization + compression screening + state-driven reallocation in the initial scheduling generation process, constructs a dynamic scheduling method for ring path topology and high-density task environment, realizes multi-objective game optimization allocation of robot resources, and comprehensively considers energy, time and path load; effectively filter potential conflict matching schemes, ensure the stability and continuity of the scheduling system; support path density dynamic adjustment and regional load prediction, significantly improve task throughput efficiency and resource use fairness.

[0045] In the robot task dynamic scheduling optimization method described in the application, in order to ensure that the scheduling scheme does not produce resource conflicts, path congestion or scheduling deadlocks and other problems in the actual deployment process, the preliminary generated scheduling scheme needs to be subjected to path conflict detection and risk assessment. In view of the characteristics of high path sharing and high-frequency task insertion in the ring warehouse system, the S500 step designs a conflict detection mechanism combining path topology structure and traffic rule modeling, ensuring the continuity and stability of the system scheduling.

[0046] In step S501, to realize the systematic representation of the path timing and space occupation relationship of each robot, first, the initial scheduling scheme is embedded in a ring path topology graph The path of each robot is embedded in the ring path topology graph. The topology graph takes each intersection node, inflection point and task point of the warehouse space as a vertex, and the track segment as an edge, forming a directed graph structure with direction.

[0047] Then, a multi-path synchronous scheduling coding matrix is constructed, the rows of the matrix represent all robots in the system, the columns represent the discretized time segments (such as dividing the time window of T seconds in the future by 1 second), and the matrix element M[i][t] represents the path node number set occupied by the robot Ri at time t. At the same time, an overlap weight parameter ω(i, j, t) is introduced, which represents the intersection overlap degree of the nodes occupied by the robots Ri and Rj at the time slice t, and the weight is dynamically weighted according to the path overlap degree, node type (such as inflection point, intersection) and task urgency.

[0048] The matrix can be used to intuitively observe the path resource distribution of the robots in the time-space domain, and provides a structured data basis for conflict detection.

[0049] In step S502, to accurately determine the potential conflict of multiple robots in the shared path area, a node conflict identification function needs to be constructed based on the path prediction tensor and the traffic rule modeling graph, which is used to analyze the node competition and identify high-risk nodes.

[0050] Specifically: Path prediction tensor generation: according to the scheduling coding matrix and the path travel speed, turning strategy and other information of each robot, a three-dimensional path prediction tensor P[i][n][t] is generated, where i is the robot number, n is the node number, and t is the future time; the tensor predicts the nodes that may be occupied by the robot and the confidence probability in the future period.

[0051] Traffic rule modeling graph construction: to avoid conflicts, a number of path traffic rules are preset in the system, including: Same direction occupation priority rule; Corner yielding rule (based on turning direction to determine intersection priority); Node backoff limit (to avoid deadlock); The above rules are modeled as priority constraint edges between nodes in the graph, which are used to judge the legal competition relationship of multiple paths on the same node.

[0052] Conflict identification function definition: combining the prediction tensor P and the traffic rule graph, the conflict function C(n, t) is defined, which is used to calculate the conflict intensity of node n at time t, considering three elements: The node competition priority difference Δπ(n, t); path crossing directionality coefficient φ(n, t); occupancy overlap probability ρ(n, t); The function is: C(n, t) = f(Δπ(n, t), φ(n, t), ρ(n, t)); if C(n, t) exceeds the set threshold, it is marked as a potential conflict node.

[0053] Through the above method, the conflict intensity at any space-time point in the system can be quantified, providing a basis for subsequent conflict tracking.

[0054] In step S503, to identify and locate the conflict source, the system uses a reverse conflict deduction mechanism. This mechanism simulates a time-reversing path replay process, gradually backtracking from the end of the current prediction window, identifying the conflict propagation path and its causes.

[0055] The specific process is as follows: time window setting and conflict window selection: set the future time window Select high conflict degree nodes in this interval to form the conflict window set Wc = {(n, t) | C(n, t) > θ}; Conflict path tracking graph construction: starting from each (n, t) node, trace all possible robot path segments that may cause the current conflict in reverse, and construct a conflict causal chain graph. Each edge represents the scheduling dependency relationship between time-adjacent path nodes.

[0056] Path backtracking and conflict aggregation analysis: backtracking and reasoning on the causal chain graph, marking the earliest path segment and robot operation that causes the conflict to increase, forming the conflict source path set.

[0057] Risk level marking: according to the conflict frequency, the number of robots involved, the length of the affected path, and other factors, calculate the risk level of the node, and mark the potential conflict node set C_critical whose conflict risk level is higher than the preset value on the topology graph as the core input for subsequent path adjustment (S600) and scheduling repair.

[0058] The S500 step has the following key technical values: Through synchronous scheduling encoding and path prediction tensor, the space-time coverage modeling of robot behavior is realized; traffic rule graph constraints improve the physical reasonableness and actual feasibility of conflict judgment; compared with traditional methods based on static conflict point filtering, reverse deduction can identify deep potential bottlenecks and locate the root cause; through risk quantification evaluation, the subsequent rearrangement module realizes more accurate path adjustment.

[0059] To effectively solve the task path conflict problem caused by overlapping of scheduling resources or node mutual exclusion, in step S600, the application proposes a path rearrangement method based on a sliding time window, which combines local controllable disturbance and dynamic scheduling diagram fitting mechanism to realize efficient correction of the conflict path in the initial scheduling scheme. This step not only has the sensing ability to high-risk conflict nodes, but also can generate an optimized conflict-free scheduling scheme without destroying the original scheduling structure, thereby improving the overall throughput performance and stability of the system.

[0060] In S601, the system first constructs a scheduling interference relationship graph oriented to the time dimension according to the potential conflict node set C_critical marked in step S500.

[0061] Specifically, each task path segment involving conflict is regarded as a node in the graph, and each node is associated with attributes including task number, associated robot number, path occupation segment, and execution time window [t_start, t_end]. Then, the system constructs a graph edge under the following conditions to represent the scheduling conflict interference relationship: Time window overlap condition: if the execution time windows of two task paths have an intersection, that is, t_overlap > 0 is satisfied; Path node mutual exclusion condition: if two task paths occupy the same topological node or adjacent path segments; Robot scheduling dependency condition: if two tasks are assigned to the same robot and are adjacent in order, there is an order restriction.

[0062] Through the above three types of logic, the system constructs a scheduling interference graph D = (V, E), where each edge e_ij ∈ E represents the possibility of potential conflict or scheduling interference between task Ti and task Tj. This graph can be used to identify rearrangeable local areas as a constraint structure for subsequent path disturbance adjustment.

[0063] In S602, to solve the high conflict degree area in the scheduling interference graph, the application proposes a controllable perturbation path rearrangement mechanism, which constructs multiple feasible candidate scheduling schemes by applying fine disturbance to the local task sequence.

[0064] The disturbance operation includes but is not limited to the following three types: Swap: under the premise of not violating the robot task execution dependency, the positions of two adjacent tasks in the same robot task queue are exchanged to reduce the concentration of high conflict nodes; Shift: the start time window of the conflict task is slid backward by several time slices to release its competition for path resources, which is suitable for scenarios with high tolerance for task response delay; Task segment split: divide the task path into multiple segments, and re-time or replace the robot in the middle segment. It is especially suitable for dynamic intervention in long path and high overlap task segments.

[0065] Each perturbation operation is bounded by the schedule interference graph, and the disturbance range is limited to the influence domain of the conflict propagation chain, so as to ensure that the global scheduling structure is not destroyed, and the controllability and effectiveness of local adjustment are improved.

[0066] Through the combination of multiple disturbance strategies, the system will generate a set of local feasible path combination candidates S candidate, each candidate corresponds to a set of task scheduling adjustment configurations, and is accompanied by preliminary evaluation results of its modification range and conflict resolution degree.

[0067] In S603, in order to select the optimal scheduling reconstruction version from multiple candidate schemes, the system constructs a dynamic scheduling graph fitting mechanism, and combines real-time system state quantities and multi-objective evaluation functions to sort and filter the candidate schemes.

[0068] Scheduling graph fitting process: embed the candidate scheduling path set into the current global scheduling graph of the system, recalculate the arrangement of each robot path segment on the time-node graph, and form a new scheduling graph G'; Key state index fusion: System load index (Load Index): represents the ratio of total scheduling task amount to robot resource per unit time, used to reflect the current saturation degree of the system; Node block probability (Node Block Probability): calculated according to the historical conflict frequency and task density statistics of the node, representing the risk of path segment being blocked during operation; Scheduling variation cost function: evaluate the modification degree of the original scheduling structure by the current disturbance operation (such as task postponement amount, path change amplitude); Multi-objective evaluation function design: construct a unified scoring function F(Si), which comprehensively considers the above indexes, takes the maximum conflict resolution ability, the minimum scheduling variation, and the optimal resource balance as the goal, and selects the optimal reconstruction scheme.

[0069] Finally, the system outputs the scheme with the highest score as the optimized conflict-free scheduling scheme , and uses it as the input for subsequent simulation evaluation and deployment execution.

[0070] The S600 step of the present application embodies the following key technical innovations: For the first time, the time window sliding mechanism is combined with the task scheduling graph structure to accurately identify the conflict concentration area and the scheduling coupling node; By constructing a local solution space through perturbation operation library, the structural solution of high conflict path is realized quickly while maintaining the stability of the scheduling system; Unlike traditional greedy optimization or static weight scoring mechanism, a system state driven fitting feedback method is adopted, so that the rearrangement result is highly matched with the system running environment.

[0071] In the actual measurement of the ring warehouse system scene, the above method can effectively reduce the path conflict rate to less than 10% of the original scheduling scheme, significantly improve the scheduling throughput and task completion time, and has good scalability and fault tolerance.

[0072] After completing the conflict resolution and generating the optimized conflict-free scheduling scheme, in order to avoid scheduling failure, path congestion or task delay in actual operation, system-level scheduling scheme simulation verification and performance evaluation need to be carried out before deployment. Therefore, the present application proposes a multi-dimensional performance evaluation mechanism integrating simulation engine, behavior modeler and evaluation module in step S700, which can predict the executability, robustness and timeliness of the scheduling scheme under non-invasive conditions, ensuring that the final output scheduling scheme has engineering practicability and running stability.

[0073] In step S701, the system first constructs a virtual scheduling simulation engine based on the task execution time sequence and path mapping relationship. The engine is used to reproduce the complete behavior trajectory of the robot executing the task according to the conflict-free scheduling scheme without affecting the actual warehouse system operation.

[0074] The construction method is as follows: Input data structure: each robot task in the conflict-free scheduling scheme contains path sequence, expected start time, execution duration, task type, node occupation plan and other fields; Topological space mapper: the engine simulates the actual warehouse topology inside, establishes the connectivity and reachable path table between nodes, and maps the robot movement path to the simulation space coordinates; Time sequence behavior executor: drive the robot to simulate the movement behavior according to the time slice promotion model, record the path segment, occupied node and execution task state of the robot at each time; Shadow record generator: generate the corresponding task execution "shadow record" for each robot, including path timestamp, occupation history, response starting point and completion time, and store it in the log structure in a backtracking format.

[0075] The simulation engine can quickly simulate large-scale task scheduling process under simulation time acceleration conditions, avoiding the cost and risk of relying on real robot testing, providing data basis for behavior modeling and performance evaluation.

[0076] In S702, in order to more accurately quantify the behavior state and scheduling performance of the robot in the simulation execution, the application introduces a task behavior dynamic modeler to periodically sample and analyze the robot behavior trajectory and extract its running characteristics.

[0077] The modeler implementation steps are as follows: Path travel state extraction: analyze the moving speed, path occupation continuity and waiting time of the robot in each time slice, and identify whether there are blocking, reverse movement or frequent start-stop behavior patterns in the scheduling graph; Task response time calculation: calculate the real response time of the task from generation to completion, and analyze the difference with the expected time in the scheduling scheme to identify the scheduling delay causes; Execution efficiency estimation: calculate the number of tasks completed by the robot per unit time and the average task execution time to form an execution efficiency index vector.

[0078] The above indexes are sampled in units of period Δt, and the system automatically merges the behavior data of different time periods into a unified format multi-dimensional vector group to form a performance feedback vector set F of each robot in the simulation process; The vector set has time sequence and measurability, supporting the subsequent comprehensive evaluation and quantitative judgment of the overall quality of the scheduling scheme by the system.

[0079] In S703, the system inputs the performance feedback vector into a multi-index evaluation module to build a complete scheduling performance evaluation system.

[0080] The specific process is as follows: Core index extraction: extract the following three core performance indexes from each feedback vector: Task completion rate (TCR, Task Completion Rate): the number of completed tasks divided by the total number of tasks, reflecting the integrity of the scheme execution; Path conflict rate (PCR, Path Conflict Rate): the number of path overlaps or waiting times in the simulation process divided by the total number of scheduling times, measuring the effectiveness of path scheduling; Average response time (ART, Average Response Time): the average of all task response times, measuring the timeliness of the scheduling scheme; Cross-weighting analysis mechanism: to achieve uniform quantification and comparison of indexes, a cross-weighting function is introduced: ; Where the weight coefficient can be dynamically set according to different business scenarios, for example, increase for time-sensitive businesses and increase for businesses with high stability requirements; Performance evaluation tensor construction: tensor splicing is performed on the Score values of all robots and time periods to form a multi-dimensional performance tensor P_tensor of the scheduling scheme, which can be used to represent the scheduling performance of the entire system in different time intervals, different regions, and different robot dimensions; Comparison judgment mechanism: P_tensor is compared with the preset system performance threshold P_threshold dimension by dimension, and if any dimension is lower than the threshold, it is determined that the current scheduling scheme does not meet the standard, and the scheme iteration update needs to be performed back to step S800.

[0081] Through the multi-dimensional performance simulation mechanism of this step, the system can comprehensively identify potential performance problems, path bottlenecks or resource allocation problems before the scheduling scheme is formally executed, effectively avoiding the risk of running interruption or resource imbalance.

[0082] Step S700 has the following significant technical advantages: high-fidelity simulation verification of the scheduling scheme can be realized without actually running the robot; feedback modeling can be performed at the task behavior level rather than only at the logic level, enabling in-depth analysis of behavior deviation; the enterprise is allowed to adjust the evaluation strategy according to its own focus (efficiency, cost, safety); the simulation output result can be used as a feedback source for the scheduling system, realizing end-to-end scheduling strategy closed-loop self-optimization.

[0083] After completing the simulation evaluation of the conflict-free scheduling scheme (S700), if the evaluation index does not meet the preset optimization threshold, a scheduling iteration optimization mechanism with performance feedback driving capability needs to be introduced to improve the system running efficiency and resource allocation rationality. The present application proposes a scheduling adaptive optimization strategy driven by performance evaluation tensor, combined with bottleneck identification and graph structure evolution in step S800, to realize the structure-level reconstruction of the task allocation graph G and the path timing readjustment, and through cyclic iteration to ensure that the final output task scheduling optimization scheme meets the global performance target.

[0084] In step S801, the system first normalizes the performance evaluation tensor P_tensor generated in step S700 to improve the comparability and feedback sensitivity of different dimension indicators. The normalization method adopts the Max-Min scaling strategy or Z-score standardization, which uniformly maps the task completion rate, path conflict rate and response time to the [0, 1] interval.

[0085] Subsequently, a key performance deviation vector is extracted, which represents the performance difference of the current scheduling scheme relative to the target threshold, and its calculation method is: ; in the formula, represents the target performance vector, and represents the normalized performance evaluation tensor.

[0086] The residual vector is input into a residual-guided optimization network (RONet) as an input. The RONet is one of the core innovative structures of the application, and its core function is to associate the performance deviation with the scheduling graph structure, and output specific bottleneck types and corresponding responsibility path segments.

[0087] The RONet network structure includes the following components: A bottleneck classification module: maps the performance deviation vector to a scheduling bottleneck type label, such as "robot overload", "path congestion", "node contention", "task delay", etc. A path tracking mapper: reflects the bottleneck type to specific path segments, task nodes or robot task sequences on the original scheduling graph G, for guiding subsequent scheduling graph reconstruction; A bottleneck impact factor scorer: assigns impact weight values to all labeled path segments, and outputs a bottleneck priority list in order.

[0088] Through this module, the system realizes the causal tracking from macro performance degradation to specific scheduling structure, laying the foundation for graph structure adaptive correction.

[0089] In the S802 step, according to the bottleneck identification result, the original task-robot matching graph G is locally reconstructed at the structure level to generate a new round of to-be-optimized scheduling graph G', which includes three types of operations: Edge weight adjustment: for the edges between tasks and robots on the bottleneck path segment, adjust the edge weight function, add conflict penalty factor, congestion penalty factor or load limit factor, so that the scheduling optimization model reduces the priority of this matching pair in subsequent iterations; Path feasibility re-estimation: for the bottleneck task path, execute the path planning module again, use the alternative paths in the alternative path set P_alt to replace, and update the path availability label in the topology structure in real time; Node selection mask update: update the "selection mask matrix" of the task node, set a cooling time or an occupancy threshold for part of the hot nodes, so as to avoid repeated allocation of bottleneck resources in the subsequent graph matching propagation process.

[0090] The reconstructed graph G' keeps the task and robot set unchanged, but the edge weight, connection relationship and node priority strategy have been adaptively mutated according to the bottleneck feedback, providing structural optimization space for the next round of scheduling calculation.

[0091] In the S803 step, the system takes the updated graph structure G' as input and re-executes the whole process of steps S300 to S700, including task matching, initial allocation, conflict detection, sliding rearrangement and performance simulation.

[0092] In each iteration, the system dynamically adjusts the following three key variables: Task assignment weight matrix W(t): Based on historical evaluation bias, gradually strengthen the performance of good task-robot combination, suppress the continuous failure of scheduling pair; Path execution order update strategy: Allow a certain range of task execution sequence rearrangement to adapt to the resource release timing of local bottleneck nodes; Time window configuration adjustment mechanism: Perform slight stretching adjustment on the task time window to avoid node resource peak period.

[0093] The adaptive iterative process performs performance evaluation tensor generation and threshold comparison after each round. If all indicators meet the set optimization target, the iteration is terminated, and the final task scheduling optimization scheme S* is output.

[0094] To prevent falling into infinite iteration or local optimal oscillation, the system sets the maximum number of iterations N_max and the minimum improvement increment threshold Δ_min. When the performance improvement of the last two rounds is less than Δ_min or N_max has been reached, the system stops iteration and enters the user intervention or parameter reset process.

[0095] S800 step constructs a full-flow scheduling feedback closed-loop optimization mechanism, and its technical advantages include: structure-level scheduling graph reconstruction capability: supports multi-level interaction reconstruction of edge weight, path, and node, breaking through the limitations of traditional scheduling optimization in graph hierarchy; Introducing feedback-enhanced dynamic scheduling strategy to achieve round-by-round evolutionary optimization under the performance bias; Combined with simulation, modeling, evaluation and reconstruction, a full-automatic scheduling quality control system is constructed.

[0096] Example 2: To verify the scheduling efficiency and conflict control capability of the "robot task dynamic scheduling optimization method of ring warehouse system" in high-density task environment, the following simulation experiment scene is designed, and two mainstream scheduling algorithms in the prior art are compared and analyzed.

[0097] Simulation platform: Python + SimPy constructs a discrete event simulation environment, combined with NetworkX to represent the path topology graph structure; Warehouse topology: typical three-ring four-zone structure, each ring path average length 20 nodes, a total of 60 path nodes, containing 8 workstations; Number of robots: 16 differential drive mobile robots (supporting bidirectional travel); Task generation model: Poisson distribution arrival model, average arrival rate λ=8 tasks / min; Task type: sorting task, replenishment task, internal transportation task, proportion 6:2:2; Task priority: automatically assigned as high, medium and low according to order timeliness; Experimental time period: the total simulation time is 60 minutes (3600 seconds); Hardware platform: Intel i7 CPU, 32GB memory environment.

[0098] To verify the performance advantages of the method of the application, the following two typical algorithms are selected as comparison: Method Brief Description A: Shortest Path First (SPF) All tasks are assigned to the nearest idle robot, path selection is based on Dijkstra algorithm, task order is static B: Time Window-based Priority Matching (TWPM) Scheduling based on task priority and path available window, no conflict prediction and structure rearrangement is introduced The method of the application (denoted as method C) introduces graph neural scheduling matching, conflict identification, adaptive disturbance rearrangement and feedback-driven graph evolution mechanism in S300 to S800.

[0099] Key evaluation indicators: Indicators Description Task Completion Rate (TCR) Number of successfully completed tasks / total number of tasks Average Response Time (ART) Average time from task generation to execution completion (unit: seconds) Path Conflict Rate (PCR) Number of path occupation conflicts occurred in simulation / total number of scheduling times Scheduling Reconstruction Times Number of times the system rearranges the scheduling plan (reflects adaptability) Comparison of experimental results (average value): Indicators Method A (SPF) Method B (TWPM) Inventive Method C Task Completion Rate (↑) 83.5% 88.2% 96.7% Average Response Time (↓) 19.6 s 15.8 s 11.2 s Path Conflict Rate (↓) 18.3% 10.7% 2.9% Scheduling Reconstruction Times (↓) 0 (static) 4 2 (adaptive) In summary, by introducing the graph attention matching and scheduling feedback iteration mechanism, the robot scheduling task is more balanced, the task delay accumulation is avoided, and the average response time is reduced by more than 30%; by using the conflict prediction + interference graph + sliding rearrangement combined mechanism, the path conflict rate is reduced to less than 3%, which is much better than the conventional strategy; due to the use of the bottleneck identification driven minimum disturbance reconstruction mechanism, the scheduling system only needs a small amount of iteration to complete the optimization, and has good engineering stability and low intervention requirement.

[0100] This embodiment fully verifies that the method of the application has significant scheduling robustness and resource coordination efficiency in the warehouse path environment with high task density and serious resource competition. Through actual construction of scheduling graph, behavior simulation and index evaluation, the application can greatly reduce path conflicts and improve system throughput efficiency while ensuring task completion rate, showing clear technical progress and practicality.

[0101] The above is merely a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which shall be covered within the protection scope of the present application.

Claims

1. A method for dynamic scheduling and optimization of robot tasks in a circular warehousing system, characterized by: include: S100. Obtain basic information about all tasks to be executed in the current system, including task location, task type, priority and estimated processing time, and construct task status table T; S200. Obtain the real-time operating status of each robot in the system, including current position, load status, power level and historical path information, and construct robot status table R; S300. Based on T and R, construct a task-robot matching graph G. Use a weighted matching strategy that combines task priority, distance cost and load balancing factor to calculate the matching priority score between each task and the robot. S400. Perform initial allocation on the matching graph G to generate an initial scheduling scheme; S500, based on the ring path topology map and traffic rules, performs path conflict detection on the initial scheduling scheme and marks potential conflict nodes; S600. For conflicting task paths, a time window sliding rearrangement algorithm is used to dynamically adjust the execution order and scheduling time of the paths, and generate an optimized conflict-free scheduling scheme. S700: Simulate the task execution order of each robot in the conflict-free scheduling scheme, and evaluate the performance based on the task completion rate, path conflict rate and average response time feedback from the system simulation. S800. If the evaluation index does not meet the preset optimization threshold, then steps S300 to S700 are iteratively updated until the system performance requirements are met, and the final task scheduling optimization scheme is output.

2. The method for dynamic scheduling and optimization of robot tasks in a circular warehousing system according to claim 1, characterized in that: The S300 includes: S301. Using the task nodes in the task status table T and the robot nodes in the robot status table R as graph vertices, construct a heterogeneous bidirectional matching graph G based on geographical distance, energy consumption, and task type coupling relationship. S302. Use an adaptive graph convolutional attention network to propagate features to graph G, and fuse task priority vector, robot load margin vector and path overlap factor to generate a multidimensional matching score tensor for task-robot pairs. S303. Based on the multi-objective normalized scoring mechanism, the score tensor is reordered to finally determine the matching priority score set between each task and the candidate robot.

3. The method for dynamic scheduling and optimization of robot tasks in a circular warehousing system according to claim 1, characterized in that: The S400 includes: S401. Based on the priority score set obtained in the task-robot matching graph G, construct a two-sided matching game model, with minimizing robot energy consumption cost and minimizing task response delay as the dual objective functions, and generate a feasible matching candidate set F. S402. Introduce a dynamic compression and filtering mechanism to forcibly filter the allocation schemes with high conflict or boundary states in set F, forming an adjustable subset F'. S403. Using a state-driven distributed scheduling strategy, local selection is performed on candidate matching pairs in F'. Based on the regional congestion threshold, self-adjusting path density, and task window prediction function, the pairs are redistributed, and the initial scheduling scheme is finally output.

4. The method for dynamic scheduling and optimization of robot tasks in a circular warehousing system according to claim 1, characterized in that: The S500 includes: S501. Embed all robot paths in the initial scheduling scheme into a ring path topology structure, and construct a multi-path synchronous scheduling coding matrix. The matrix elements represent the node timing and overlap weight occupied by the robot in any two time segments. S502. Based on the multidimensional path prediction tensor and traffic rule modeling graph, establish a node conflict identification function to jointly determine the node competition priority, path intersection direction and node occupancy overlap rate. S503. Using a reverse conflict inference mechanism, the scheduling coding matrix within the prediction window is analyzed back over time to trace potential conflict sources step by step, and the set of potential conflict nodes with a conflict risk level higher than a preset threshold is marked on the topology map.

5. The method for dynamic scheduling and optimization of robot tasks in a circular warehousing system according to claim 1, characterized in that: The S600 includes: S601. For the identified set of potential conflict nodes, establish a scheduling interference graph based on a sliding time window, where the edges represent the path mutual exclusion relationship between tasks with overlapping time windows. S602. By applying controllable perturbations to the robot task sequence where the conflict path is located, including sequence swapping, delay offsetting and task segmentation, multiple local feasible path combination candidates are generated. S603. Perform dynamic scheduling graph fitting on the candidate path combinations. Combining the current system load index and node blocking probability, use a multi-objective evaluation function to select the reconstruction scheme with the least conflict and output the optimized conflict-free scheduling scheme.

6. The method for dynamic scheduling and optimization of robot tasks in a circular warehousing system according to claim 1, characterized in that: The S700 includes: S701. Based on the task execution timing and path mapping relationship, a virtual scheduling simulation engine is constructed. The simulation engine takes a conflict-free scheduling scheme as input and generates a robot behavior trajectory sequence and task execution shadow record without interfering with the operation of the real system. S702. Introduce a task behavior dynamic modeler to periodically sample the path travel state, task response time and execution efficiency of each robot in the simulation and generate a multi-dimensional performance feedback vector. S703. Input the performance feedback vector into the multi-index evaluation module, perform cross-weighted analysis on task completion rate, path conflict rate and average response time, construct a performance evaluation tensor, and compare it with the preset system performance threshold.

7. The method for dynamic scheduling and optimization of robot tasks in a circular warehousing system according to claim 1, characterized in that: The S800 includes: S801. Normalize the generated performance evaluation tensor, extract the key performance deviation vector, and input it into the residual-driven scheduling scheme correction network to identify the current scheduling bottleneck type and the corresponding responsible path segment. S802. Based on the bottleneck feedback results, the task-robot matching graph G is reconstructed locally, including edge weight adjustment, path feasibility reassessment and node selection mask update, to generate a new round of iterative scheduling graph G'. S803. Re-execute steps S300 to S700 on the updated graph G'. Through multiple rounds of adaptive iteration, dynamically adjust the task allocation weight, path execution order, and time window configuration until all evaluation indicators meet the preset optimization threshold, and output the final task scheduling optimization scheme.

Images