Intelligent agent cooperative structure dynamic generation system based on directed acyclic graph search

Abstract

The application relates to the technical field of intelligent scheduling, and discloses an agent collaborative structure dynamic generation system based on a directed acyclic graph search, which comprises the following steps: after a system receives a shelf inspection task, according to a strain time sequence of a measuring point, an acceleration time sequence of the measuring point and spatial coordinates of the measuring point, a first significant response time, a spatial distance, a response similarity and a bottleneck degree are determined, a response directed acyclic graph and a shelf load response sequence bottleneck skeleton are formed; the skeleton is mapped into a sequence constraint label and a bottleneck node label, which are used for pruning a candidate directed acyclic collaborative graph. A candidate graph scoring module combines general scoring, a same direction reachability degree and an opposite direction reachability degree to determine a consistency score, a reverse violation proportion and an enhancement score. A target collaborative graph solidification module determines a target collaborative directed acyclic graph according to the consistency score, the reverse violation proportion and the enhancement score, which is used for automatic warehouse shelf inspection and risk assessment.

Description

Technical Field

[0001] This invention belongs to the field of intelligent scheduling technology, specifically relating to a dynamic generation system for agent cooperative structure based on directed acyclic graph search. Background Technology

[0002] In automated warehouses, steel pallet racks withstand dynamic loads from inbound and outbound operations, handling collisions, and periodic inspections. The field system typically records rack status using strain sensors, accelerometers, installation logs, and equipment positioning data, and the inspection platform generates risk assessment tasks. In engineering implementation, data reading, signal analysis, vibration assessment, risk summarization, and report generation are often handled by different agents or service units, requiring the scheduling platform to coordinate permissions, capabilities, latency, and resource usage.

[0003] Current intelligent agent workflow orchestration generally focuses on task dependencies and execution costs, generating directed acyclic collaborative graphs and achieving parallel scheduling. However, in shelf inspection scenarios, if the collaborative graph only reflects task textual relationships or service capabilities, it easily overlooks the differences in the order of responses, proximity relationships, and bottleneck points of different measurement points within the same loading event. While the generated candidate graph may be executable, it might schedule the summary results of later response regions before the analysis of earlier response regions, or place tasks related to key measurement points in later processing positions, leading to an increase in the number of candidates and a decrease in screening efficiency.

[0004] In engineering deployments, shelf sensor data has a field source, while agent collaboration graphs have execution constraints; these are typically processed by different modules. Without a mechanism to convert load response evidence into collaboration graph search conditions and scoring criteria, the scheduling system struggles to constrain candidate graphs using current event data without altering shelf equipment and sensor layout. Summary of the Invention

[0005] The present invention aims to provide a dynamic generation system for agent cooperative structures based on directed acyclic graph search, which solves the technical problems mentioned in the background art.

[0006] To achieve the above-mentioned objectives, the present invention provides the following technical solution.

[0007] The intelligent agent cooperative structure dynamic generation system based on directed acyclic graph search includes a task receiving and input normalization module, an atomic subtask decomposition module, a candidate directed acyclic cooperative structure search module, a candidate graph scoring module, and a target cooperative structure solidification module. It also includes a shelf load response sequence bottleneck skeleton construction module and a structure partial order label generation module. The bottleneck skeleton construction module for the sequential response of the rack under load is used to determine the first significant response time, spatial distance, response similarity and bottleneck degree based on the strain time series, acceleration time series and spatial coordinates of the measuring points in the same rack under load event, and to form a directed acyclic graph of response and a bottleneck skeleton for the sequential response of the rack under load. The structural partial order label generation module is used to map the bottleneck skeleton of the shelf load response sequence into structural partial order labels and bottleneck node labels. The candidate directed acyclic cooperative structure search module is used to prune candidate directed acyclic cooperative structures based on the partial order label of the structure and the bottleneck node label. The candidate graph scoring module is used to obtain a general score based on quality, latency, communication and resource conflicts. Based on the degree to which the directed edges in the bottleneck skeleton of the shelf load response remain in the same direction and the degree to which opposite directions of reachability appear in the candidate directed acyclic cooperative structure, the physical consistency score and the reverse violation ratio are determined respectively, and the enhancement score is determined. The target collaborative structure solidification module is used to determine the target collaborative directed acyclic graph based on the enhancement score.

[0008] Compared with the prior art, the present invention has the following substantial features and significant progress: In shelf inspection tasks, this invention transforms the strain time series, acceleration time series, and spatial coordinates of measuring points into sequential constraints and bottleneck criteria that can be used for collaborative scheduling. These criteria are then embedded into the search, pruning, and scoring process of candidate directed acyclic collaborative graphs. This enables the system to reduce candidate branches inconsistent with the current shelf load response, prioritize the analysis of key measuring points in the risk summary preprocessing, and improve the matching degree between inspection task scheduling and on-site load response while maintaining the original agent registration, assignment, execution, and replanning processes. It also reduces the search overhead caused by irrelevant candidates and enhances the traceability of the risk report formation process. Attached Figure Description

[0009] Figure 1 This is a diagram showing the combination of modules and mechanisms of the present invention. Detailed Implementation

[0010] The following description is provided in conjunction with the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the invention; those skilled in the art can make equivalent substitutions or combinations for the specific implementations.

[0011] This implementation is applied to the scenario of rack inspection and risk assessment in automated warehouses, and the processing objects are rack load-bearing units including columns, beams, diagonal braces and existing connections. The system does not change the physical structure of the rack load-bearing units, the sensor installation structure, or the existing basic processes of registration, assignment, execution, monitoring and replanning in the intelligent agent execution platform. Instead, it introduces structured response evidence generated by rack load events into the existing intelligent agent collaborative structure generation process.

[0012] like Figure 1 As shown, the intelligent agent collaborative structure dynamic generation system 100 receives a shelf inspection task 300, a measurement point strain time series 310, a measurement point acceleration time series 320, a measurement point spatial coordinate 330, and an intelligent agent registry 340. The shelf inspection task 300 is input into the task receiving and input normalization module 110; the measurement point strain time series 310, the measurement point acceleration time series 320, and the measurement point spatial coordinate 330 are input into the shelf load response sequence bottleneck skeleton construction module 130; and the intelligent agent registry 340 is input into the intelligent agent registration information reading module 150. The modules within the system 100 collaborate sequentially according to data flow and processing flow, and finally, the result summary and audit log module 230 outputs a risk report 400 and an audit log 410.

[0013] Figure 1 The task chain consists of a task receiving and input normalization module 110, an atomic subtask decomposition module 120, a candidate directed acyclic cooperative structure search module 160, an agent assignment module 170, a candidate graph scoring module 180, a target cooperative structure solidification module 190, and a topology scheduling execution module 200. The shelf response constraint chain consists of a shelf load response priority bottleneck skeleton construction module 130 and a structural partial order label generation module 140. The failure feedback chain consists of an execution monitoring and intermediate result verification module 210 and an incomplete subgraph replanning module 220.

[0014] In this embodiment, the strain time series at the measurement points is directly acquired by strain sensors installed on the load-bearing components of the rack, and the acceleration time series at the measurement points is directly acquired by acceleration sensors installed on the same load-bearing unit of the rack. The spatial coordinates of the measurement points are obtained directly through conventional measurement, installation records, or equipment positioning. The system uses these three types of directly acquired data to form a directed acyclic graph of response and a bottleneck skeleton of rack load response sequence. The bottleneck skeleton of rack load response sequence is then transformed into structural partial order labels and bottleneck node labels, thereby ensuring that the search, pruning, assignment, scoring, solidification, execution, and replanning of candidate directed acyclic collaborative structures are consistent with the response sequence relationship in the current rack load event.

[0015] Figure 1 Number 170 indicates the same agent assignment module. Figure 1The two numbers 170 appearing in the diagram represent the processing relationship of the same agent assignment module at two input convergence points. One plotting location represents the input of the candidate directed acyclic cooperative structure set, and the other plotting location represents the input of the agent capability matrix. These are not two different agent assignment modules. The number 170 below only indicates that the agent capability matrix and the candidate directed acyclic cooperative structure set converge at the agent assignment module. The bottleneck skeleton construction module 130 does not directly output control or assignment data to the agent assignment module 170. The output of the agent assignment module 170 is always a candidate directed acyclic cooperative structure with agent assignment, which is then input into the candidate graph scoring module 180.

[0016] In this implementation, the candidate directed acyclic cooperative structure represents the dependencies and execution order between task nodes in the inspection task to be executed, and the target cooperative directed acyclic graph represents the cooperative structure selected by enhanced scoring and used for actual execution. The response directed acyclic graph represents the response relationship formed by each measuring point in the same shelf loading event according to the strictly increasing first significant response time, and the shelf loading response sequence bottleneck skeleton represents the structural object obtained by retaining the directed response path passing through the bottleneck node from the response directed acyclic graph.

[0017] The task receiving and input standardization module 110 receives the shelf inspection task 300 and converts it into a standardized task object that can be processed by the atomic subtask decomposition module 120. The shelf inspection task 300 can be a natural language task or a structured inspection task; its content includes the target shelf area, task objective, deadline, permission scope, available resources, output requirements, prohibited operations, and completion criteria.

[0018] The task receiving and input standardization module 110 performs field recognition, time format standardization, target shelf area identification, and output requirement organization for the shelf inspection task 300. If the shelf inspection task 300 contains multiple target shelf areas, the task receiving and input standardization module 110 retains each target shelf area in the standardized task object, enabling the subsequent atomic subtask splitting module 120 to generate corresponding atomic subtask sets according to the target shelf areas. If the shelf inspection task 300 lacks completion criteria, the task receiving and input standardization module 110 generates completion criteria corresponding to the risk report 400 based on the inspection task template.

[0019] A standardized task object must include at least the task instance number, target shelf area, output requirements, prohibited operations, access permissions, and completion criteria. The task instance number is used to associate the strain time series of the measuring point (310), the acceleration time series of the measuring point (320), the spatial coordinates of the measuring point (330), the candidate directed acyclic collaborative structure, the target collaborative directed acyclic graph, the risk report (400), and the audit log (410). The target shelf area is used to restrict the bottleneck skeleton construction module (130) to only use the measuring point data corresponding to this target shelf area, limiting the sequence of shelf load responses.

[0020] The atomic subtask decomposition module 120 is used to generate a set of atomic subtasks and general dependencies based on standardized task objects and existing task templates. Task nodes in the atomic subtask set can include data reading tasks, strain analysis tasks, vibration analysis tasks, structural risk assessment tasks, risk summary nodes, and report generation tasks. General dependencies are used to represent the logical relationships between tasks that are independent of the bottleneck frame and the order of the rack's load response.

[0021] The atomic subtask decomposition module 120 selects a task template based on the target shelf area, output requirements, and completion criteria, and expands the processing items in the task template into task nodes. Each task node must have at least a task node number, task type, required input and output fields, permission requirements, resource requirements, and completion criteria. General dependencies are recorded as directed relationships between task node numbers; for example, a data reading task should precede a strain analysis task and a vibration analysis task, and a structural risk assessment task should precede a report generation task.

[0022] The atomic subtask decomposition module 120 outputs a set of atomic subtasks and general dependencies to the candidate directed acyclic cooperative structure search module 160. The general dependencies serve as the search basis for the candidate directed acyclic cooperative structure search module 160, and together with the structural partial order labels and bottleneck node labels output by the structural partial order label generation module 140, they constrain the candidate branches.

[0023] When the shelf inspection task 300 involves multiple measurement points within the same target shelf area, the atomic subtask splitting module 120 generates analysis tasks related to the measurement points based on the existing correspondence between the measurement point number, the structural analysis object number, and the task node number. This correspondence is used for subsequent node task mapping. If multiple measurement points have already been assigned to the same structural analysis object in the task template, the atomic subtask splitting module 120 retains this attribution relationship and outputs it to the structural partial order label generation module 140.

[0024] The shelf load response sequence bottleneck skeleton construction module 130 is used to determine the first significant response time, spatial distance, response similarity, and bottleneck degree based on the strain time series, acceleration time series, and spatial coordinates of the measuring points in the same shelf load event, forming a directed acyclic graph of the response and a shelf load response sequence bottleneck skeleton. For example... Figure 1 As shown, the bottleneck skeleton construction module 130 for the sequential load response of the shelving includes a data dimensionless processing unit 131, a joint response intensity forming unit 132, a directed acyclic graph generation unit 133, and a bottleneck skeleton extraction unit 134 for the sequential load response of the shelving.

[0025] like Figure 1 As shown, the dimensionless data processing unit 131 receives the strain time series 310 and acceleration time series 320 of the measuring points, and outputs the normalized strain change and normalized acceleration. The joint response intensity forming unit 132 receives the normalized strain change and normalized acceleration, and outputs the joint response intensity. The directed acyclic graph (DAG) generation unit 133 receives the joint response intensity and the spatial coordinates 330 of the measuring points, and outputs the DAG. The bottleneck skeleton extraction unit 134 for the shelf load response sequence receives the DAG and the joint response intensity, and outputs the shelf load response sequence bottleneck skeleton to structural partial order label generation module 140 and the candidate image scoring module 180.

[0026] Within the same shelf loading event, the system captures the strain and acceleration time series of each measuring point using a unified event window. The unified event window can correspond to pallet placement, pallet removal, minor forklift collisions, or periodic inspection excitations. The discrete sampling times within the unified event window are denoted as... ,in Numbering the discrete sampling points, , To standardize the number of sampling points in the event window.

[0027] If different sensors have different sampling frequencies, the system uses conventional linear interpolation or resampling methods that maintain the time order of sampling points to convert the strain time series and acceleration time series of the measurement points to a unified time axis. The sampling interval of the unified time axis is taken as the larger value among the original sampling intervals of each sensor or the sampling interval pre-given by the task template. The time alignment error after resampling does not exceed half of the unified time axis sampling interval; if it exceeds this error, the measurement point data corresponding to the current rack loading event is marked as time alignment failure and is not used to form the bottleneck skeleton of the rack loading response sequence.

[0028] The data dimensionless processing unit 131 first starts with the unified event window. The strain is used as the local baseline for this shelf loading event, and the strain change at each measuring point is calculated.

[0029] ; in, For the first Each measurement point at discrete sampling time The change in strain. For the first Each measurement point at discrete sampling time The strain time series at the measuring points For the first At the start of the unified event window, each measurement point... The strain time series at the measuring points. The strain change has the same dimensions as the strain time series at the measuring points.

[0030] The dimensionless data processing unit 131 determines the maximum absolute value of strain change and the maximum absolute value of acceleration, respectively.

[0031] ; ; in, It is the maximum absolute value of the strain change, and its dimensions are the same as those of the strain change. It represents the maximum absolute value of acceleration, and its dimensions are the same as the acceleration time series at the measuring point; For the first Each measurement point at discrete sampling time The acceleration time series at the measurement points.

[0032] The dimensionless data processing unit 131 performs dimensionless processing using the strain normalization constant and the acceleration normalization constant to obtain the normalized strain change and normalized acceleration.

[0033] ; ; in, To normalize strain changes, For normalized acceleration, For strain normalization constant, This is the acceleration normalization constant. The strain normalization constant has the same dimensions as the strain change, and the acceleration normalization constant has the same dimensions as the acceleration time series at the measuring point. The strain normalization constant is taken as the effective resolution of the strain sensor or... of to The larger of the multiples; when When the strain normalization constant is taken as the effective resolution of the strain sensor, the acceleration normalization constant is taken as the effective resolution of the acceleration sensor or... of to The larger of the multiples; when When the acceleration normalization constant is taken as the effective resolution of the accelerometer.

[0034] The joint response intensity forming unit 132 synthesizes the normalized strain change and normalized acceleration into a joint response intensity using the Euclidean norm.

[0035] ; in, For the first Each measurement point at discrete sampling time The combined response intensity. Since both normalized strain change and normalized acceleration are dimensionless, the combined response intensity is also dimensionless. The combined response intensity simultaneously reflects the local force change corresponding to the strain time series at the measuring point and the dynamic response corresponding to the acceleration time series at the measuring point.

[0036] The directed acyclic graph generation unit 133 determines the first significant response time based on the joint response intensity. For each measurement point, the system uses the peak value of the joint response intensity of that measurement point within a unified event window multiplied by a trigger threshold as the significant response criterion, and selects the discrete sampling time at which the criterion is first reached as the first significant response time.

[0037] ; in, For the first The first significant response time of each measuring point This is the trigger threshold. The trigger threshold is a dimensionless ratio, and its value ranges from [value range missing]. Default value When noise at measurement points causes multiple measurement points to be triggered near the starting position of a unified event window, the trigger threshold can be increased within the above range; when the shelf load event is weak but the order of responses still needs to be preserved, the trigger threshold can be decreased within the above range. If a measurement point does not meet the significant response criterion determined by the trigger threshold within the unified event window, then that measurement point will not be used to establish a directed edge in the directed acyclic graph of the response.

[0038] The directed acyclic graph generation unit 133 calculates the spatial distance between any two measuring points based on the spatial coordinates of the measuring points.

[0039] ; in, For spatial distance, For the first The spatial coordinates of each measuring point For the first The spatial coordinates of each measuring point. The dimension of the spatial coordinates of the measuring points is length, and the dimension of the spatial distance is length.

[0040] Preset distance This represents the upper limit of spatial distance, measured in units of length. The preset distance is determined by the installation records of adjacent measuring points within the target shelving area; when the installation records provide the design distance between adjacent measuring points, the preset distance is taken as that design distance; when the installation records do not provide the design distance between adjacent measuring points, the preset distance is taken as the upper quartile of the nearest neighbor spatial distance for each measuring point within the same shelving load unit. The preset distance must be greater than zero and must not exceed the maximum spatial distance between any two measuring points within the target shelving area.

[0041] The directed acyclic graph generation unit 133 calculates the response similarity based on the joint response intensity sequence of the two measurement points. The response similarity calculation uses a response similarity normalization constant.

[0042] ; in, In response to similarity, This is the response similarity normalization constant. The response similarity normalization constant is dimensionless and its value ranges from [value missing]. Default value Response similarity is used to preserve the relationship between measurement points with similar response patterns in the same shelf loading event.

[0043] Response similarity threshold To establish a lower bound for response similarity when there are strictly sequential directed edges, the response similarity threshold is a dimensionless quantity with a value range of [value range missing]. Default value When the noise level at the measurement points within the target shelf area is high, the response similarity threshold can be increased within the above range; when the number of measurement points is small and more candidate directed edges need to be retained, the response similarity threshold can be decreased within the above range.

[0044] The response directed acyclic graph generation unit 133 establishes strictly sequential directed edges only when three conditions are met: the first condition is the measurement point. The first significant response time was earlier than that at the measuring point. The first significant response time, the second condition is that the spatial distance does not exceed the preset distance, and the third condition is that the response similarity reaches the response similarity threshold.

[0045] ; ; in, In response to a directed acyclic graph, In response to the set of graph nodes, For the directed edge set of the response graph, For the preset distance, The response similarity threshold is used. The set of nodes in the response graph consists of measurement points that meet the significant response criterion, and the set of directed edges in the response graph consists of pairs of measurement points that satisfy the strict sequential directed edge condition.

[0046] The acyclicity of the directed acyclic graph of the response is guaranteed by the strictly increasing relationship of the first significant response time.

[0047] ; For any directed edge in the response graph, the starting response time is strictly less than the ending response time. Therefore, along any directed path, the first significant response time continuously increases and it is impossible to return to the starting point of the path. The resulting graph is a directed acyclic graph of responses.

[0048] The average joint response intensity is calculated on the directed acyclic graph of the response of the rack under load, based on the bottleneck skeleton extraction unit 134.

[0049] ; in, For the first The average joint response intensity of each measuring point To standardize the number of sampling points within the event window, the average joint response intensity is a dimensionless quantity.

[0050] The bottleneck degree is used to reflect the extent to which a measurement point is traversed by multiple shortest directed response paths in the directed acyclic graph of the response. Combined with the average joint response intensity, it represents the response penetration and actual response magnitude of the measurement point in the current shelf loading event.

[0051] ; in, As a bottleneck level, For the node To the node The number of shortest directed response paths, For the nodes The number of shortest directed response paths. (Summation range limited) , and All of them are nodes in the response graph node set, and , , .like This node is not included in the bottleneck degree calculation. The bottleneck degree is a dimensionless quantity.

[0052] The bottleneck skeleton extraction unit 134 extracts the bottleneck degree from the directed acyclic graph of the response according to the preset high quantile screening conditions, and retains the directed response path that passes through the measurement point that meets the preset high quantile screening conditions, so as to form the bottleneck skeleton of the load response sequence of the shelf.

[0053] ; ; in, For the set of bottleneck nodes, For a set of path nodes, For the bottleneck skeleton, the set of directed edges. The bottleneck framework is designed to address the load response sequence of the shelving. Preset high-quantile screening criteria are based on quantile points. It is confirmed that quantiles are dimensionless quantities, and their range of values ​​is [range missing]. Default value When the number of measuring points in the target shelf area is large, the quantile level can be increased within the above range; when the number of measuring points is small, the quantile level can be decreased within the above range.

[0054] If the set of directed edges in the response graph is empty or the set of directed edges in the bottleneck skeleton is empty, the candidate directed acyclic cooperative structure search module 160 reverts to a candidate directed acyclic cooperative structure search based solely on general dependencies. At this point, the candidate graph scoring module 180 records the physical consistency score as... The reverse violation ratio is denoted as And in audit log 410, it was recorded that the bottleneck skeleton did not form a set of directed edges for the load response of the shelf.

[0055] The structural partial order label generation module 140 is used to map the nodes in the bottleneck skeleton of the shelf load response sequence to structural analysis objects, map the directed edges in the bottleneck skeleton of the shelf load response sequence to the structural partial order labels used for the earlier response analysis results before the later response summary, and set bottleneck node labels for the analysis tasks corresponding to the bottleneck nodes.

[0056] like Figure 1 As shown, the bottleneck skeleton construction module 130 for the sequential load response of the shelving outputs the bottleneck skeleton to the structural partial order label generation module 140. The structural partial order label generation module 140 then outputs the structural partial order labels and bottleneck node labels to the candidate directed acyclic collaborative structure search module 160. This positional relationship ensures that the bottleneck skeleton for the sequential load response of the shelving is not simply a record result in the risk report 400, but rather serves as a search constraint for the candidate directed acyclic collaborative structure search module 160.

[0057] The structural partial order label generation module 140 establishes a node task mapping. The node task mapping takes the node numbers in the bottleneck skeleton of the shelf load response sequence as input and the task node numbers in the candidate directed acyclic cooperative structure as output, saving the correspondence between the measurement point numbers, structural analysis object numbers, and atomic subtask numbers. The node task mapping preferentially uses the measurement point numbers and structural analysis object numbers from the installation record; if the installation record lacks structural analysis object numbers, it uses the correspondence between the target shelf area and the atomic subtask set in the task template.

[0058] When a node in the bottleneck skeleton corresponding to a load response sequence of a shelf corresponds to multiple analysis tasks, the node task mapping selects the analysis task that first produces the structural analysis result for that node as the mapping result, and requires other related analysis tasks to be located before or in parallel with this mapping result. If multiple nodes correspond to the same task node, then the same task node is used as the mapping result for multiple nodes; when both ends of a directed edge in the bottleneck skeleton's directed edge set are mapped to the same task node, this directed edge is treated as maintaining reachability in the same direction in the physical consistency score, but not as reachingable in the opposite direction in the reverse violation ratio.

[0059] When the node task mapping cannot determine the task node for the node in the bottleneck skeleton of the shelf load response sequence, the structural partial order label generation module 140 marks the corresponding node as a mapping failure node. If the mapping failure node belongs to the bottleneck node set, the structural partial order label generation module 140 retains the corresponding bottleneck node label and mapping failure record. The candidate directed acyclic collaborative structure search module 160 judges any candidate branch that needs to use the bottleneck node label but cannot determine the corresponding task node as not satisfying the bottleneck node label constraint and discards it until the node task mapping is completed by the task template or installation record before allowing it to re-enter the search. The candidate graph scoring module 180 records the reason for the mapping failure in the audit log 410. If the mapping failure node does not belong to the bottleneck node set, the candidate graph scoring module 180 removes the directed edge with the mapping failure node as the endpoint from the bottleneck skeleton directed edge set, and uses the remaining directed edges that can simultaneously determine the task mapping of the nodes at both ends as the set of valid directed edges for physical scoring. The physical consistency score and the reverse violation ratio are calculated only based on the set of valid directed edges for physical scoring.

[0060] For any directed edge in the bottleneck skeleton of the shelf load response sequence bottleneck skeleton, the structural partial order label generation module 140 generates a corresponding structural partial order label. The constraint expressed by this structural partial order label is: the analysis task output corresponding to the earlier response measurement point should be used before the summary or judgment process corresponding to the later response measurement point. For nodes in the bottleneck node set, the structural partial order label generation module 140 sets bottleneck node labels. The bottleneck node labels are used to require that the analysis task corresponding to the bottleneck node be completed before the risk summary node in the candidate directed acyclic collaborative structure.

[0061] The agent registration information reading module 150 reads the agent registry 340 and forms an agent capability matrix. The agent registry 340 includes agent ID, capability tag, interface address, permission flag, historical success rate, average latency, resource usage, and current status. The current status includes online status, offline status, interface failure status, resource usage status, and permission status.

[0062] The agent capability matrix indicates whether an agent in the set of available agents can execute task nodes in the set of atomic subtasks. Each element in the agent capability matrix is ​​determined by the capability label matching result, permission status, and current status. If the agent's capability label contains the capability label required by the task node, the permission status allows execution of the corresponding task node, and the current status is online, then the agent is available for that task node; otherwise, the agent is not available for that task node.

[0063] Historical success rate is determined by the proportion of times the agent successfully completes similar tasks within a recent fixed number of tasks or a fixed time window. Average latency is determined by the average execution time of the agent in completing similar tasks within the same statistical range. If the agent lacks historical records, the historical success rate uses the initial success rate in the agent registry 340, and the average latency uses the expected latency in the task template. The above statistical range is configured by the agent registry 340 and remains unchanged within the same shelf inspection task 300.

[0064] The agent assignment module 170 is used to map task nodes to executable agents based on the agent capability matrix. For example... Figure 1 As shown, the agent registration information reading module 150 outputs the agent capability matrix to the agent assignment module 170, the candidate directed acyclic cooperative structure search module 160 outputs the candidate directed acyclic cooperative structure set to the agent assignment module 170, and the agent assignment module 170 then outputs the candidate directed acyclic cooperative structure with agent assignment to the candidate graph scoring module 180.

[0065] For each task node in the candidate directed acyclic cooperative structure, the agent assignment module 170 first filters executable agents, then sorts them according to their historical success rate from high to low; when the historical success rates are the same, they are sorted according to their average latency from low to high; when the average latency is the same, they are sorted according to their resource consumption from low to high. If multiple task nodes in the same topology layer select the same non-concurrent resource, the agent assignment module 170 records this as a resource conflict and prioritizes the agent assignment of the analysis task corresponding to the bottleneck node label.

[0066] The candidate directed acyclic cooperative structure search module 160 is used to search based on the general dependency relationship formed by the atomic subtask splitting module 120. When expanding the task edges, verification nodes or summary nodes of the candidate directed acyclic cooperative structure and forming parallel branches, it checks whether the candidate directed acyclic cooperative structure reverses the partial order label and prunes the candidate branches with reversed partial order labels.

[0067] The candidate directed acyclic cooperative structure search module 160 receives the set of atomic subtasks, general dependencies, partial order labels of structures, bottleneck node labels, node-task mappings, and graph search width. The graph search width represents the upper limit of the number of candidate directed acyclic cooperative structures retained after each round of search; its value is a positive integer, with a default value of [value missing]. When the number of atomic subtasks is small, the graph search width can be set to... to When the number of atomic subtasks is large, the graph search width can be taken as... to .

[0068] The candidate directed acyclic cooperative structure search module 160 generates an initial candidate directed acyclic cooperative structure based on the general dependency relationship. Subsequently, it generates candidate extensions according to the task edges, verification nodes, summary nodes, and parallel branches allowed by the task template. For each candidate extension generated, the candidate directed acyclic cooperative structure search module 160 calculates the transitive closure after the addition of the task edge and checks whether the addition of the task edge forms a directed cycle.

[0069] The transitive closure can be obtained by performing a depth-first search on each task node, or by iteratively updating the Boolean reachability matrix. For the Boolean reachability matrix implementation, if a new task edge is... Then Reachable Record it as true, and record all reachable... Task nodes and all those made by The reachability relationship between reachable task nodes is recorded as true. If a new task edge is added, it makes... Reachable If both conditions are met, then the candidate directed acyclic cooperative structure contains a directed cycle, and the candidate branch is discarded.

[0070] The candidate directed acyclic cooperative structure search module 160 performs pruning based on the partial order labels of the structures. This includes pruning the directed edges in the bottleneck skeleton for the order of rack load response. If the node task mapping is obtained and Furthermore, the transitive closure of the candidate directed acyclic cooperative structure contains elements from... arrive If there is an reachability relation, then the candidate branch reverses the partial order label and is pruned. If there exists a reachability relation in the transitive closure from... arrive If a candidate branch has an reachability relationship with a given structure, it retains the partial order label of that structure. If neither of these relationships exists, the candidate branch can be retained and further differentiated by physical consistency score in the candidate graph scoring module 180.

[0071] The candidate directed acyclic collaborative structure search module 160 forms a constraint based on the bottleneck node label that the analysis task corresponding to the bottleneck node must be completed before the risk summary node. If there is no reachable relationship between the task node corresponding to the bottleneck node label and the risk summary node, the candidate directed acyclic collaborative structure search module 160 attempts to add a task edge from the task node corresponding to the bottleneck node label to the risk summary node; if adding the task edge results in a directed cycle, the corresponding candidate branch is discarded.

[0072] For measurement points with the same initial significant response time, the candidate directed acyclic cooperative structure search module 160 allows the corresponding atomic subtasks to form parallel branches when there is no general dependency. The method for determining parallel branches is as follows: if there are no directional constraints between two task nodes in terms of general dependency and structural partial order labels, and their required resources do not conflict, they can be placed in the same topology layer. If resource conflicts exist, the agent assignment module 170 serializes their processing.

[0073] After each round of candidate expansion, the candidate directed acyclic cooperative structure search module 160 first deletes candidate branches containing directed cycles, reversed structure partial order labels, or those that cannot satisfy bottleneck node label constraints. Then, based on the temporary score that can be calculated by the candidate graph scoring module 180, it retains candidate directed acyclic cooperative structures that do not exceed the graph search width. The temporary score is determined by a fixed priority based on the number of satisfied structure partial order labels, the number of satisfied bottleneck node labels, the number of covered atomic subtasks, and the number of candidate directed edges. The comparison rules are as follows: from most to least satisfied structure partial order labels, from most to least satisfied bottleneck node labels, from most to least covered atomic subtasks, from least to most candidate directed edges, and from smallest to largest candidate structure generation sequence number. The search ends when there are no more task edges, verification nodes, or summary nodes to be added, or when the number of candidates reaches the graph search width and all candidates have covered the set of atomic subtasks.

[0074] The candidate graph scoring module 180 is used to calculate the transitive closure of the candidate directed acyclic cooperative structure. It determines the physical consistency score based on the degree to which the directed edges in the bottleneck skeleton remain in the same direction in the transitive closure after the shelf load response. It also determines the reverse violation ratio based on the degree to which the directed edges in the bottleneck skeleton appear in opposite directions in the transitive closure after the shelf load response.

[0075] The candidate graph scoring module 180 receives candidate directed acyclic cooperative structures with agent assignments output by the agent assignment module 170, and receives the bottleneck skeleton of the shelf load response sequence output by the shelf load response sequence construction module 130. For example... Figure 1 As shown, the candidate graph scoring module 180 is located between the agent assignment module 170 and the target cooperative structure solidification module 190, and simultaneously receives candidate directed acyclic cooperative structures and the bottleneck skeleton of the shelf load response sequence, so that the enhanced score reflects both the general score and the structural constraints corresponding to the current shelf load event.

[0076] Let the candidate directed acyclic cooperative structure be... The bottleneck skeleton of the rack is the order of load response. Node task mapping is The transitive closure of the candidate directed acyclic cooperative structure is .in, Let be the set of task nodes in the candidate directed acyclic cooperative structure. This represents the set of task edges in the candidate directed acyclic cooperative structure. (Transitive closure) It is a non-reflexive transitive closure; when At that time, the candidate graph scoring module 180 has already processed the results from both ends simultaneously according to the same task node.

[0077] The set of valid directed edges for physical scoring represents directed edges in the bottleneck skeleton where both ends of the directed edge set have had their task nodes determined through node-task mapping. If either end of a bottleneck skeleton directed edge is a mapping failure node, that bottleneck skeleton directed edge will not be included in the set of valid directed edges for physical scoring.

[0078] ; in, This is the set of valid directed edges for physics scoring.

[0079] The physical consistency score is used to represent the proportion of directed edges of the bottleneck skeleton in the bottleneck skeleton that remain in the same direction and reachable in the candidate directed acyclic cooperative structure, according to the order of the load response of the shelf.

[0080] ; in, For physical consistency score, For the directed edges in the set of valid directed edges for physics scoring, For nodes The task nodes are mapped to the corresponding task nodes. For nodes The task nodes are mapped to the corresponding task nodes. The physical consistency score is a dimensionless proportion.

[0081] The reverse violation ratio is used to represent the proportion of bottleneck skeleton directed edges with opposite reachable directions in a candidate directed acyclic cooperative structure.

[0082] ; in, This is the reverse violation ratio. The reverse violation ratio is a dimensionless ratio. When the bottleneck skeleton directed edge set is empty or the physical score valid directed edge set is empty, the candidate graph scoring module 180 records the physical consistency score as... The reverse violation ratio is denoted as .

[0083] The candidate graph scoring module 180 obtains a general score based on quality, latency, communication, and resource conflicts. The quality score is obtained by averaging the historical success rates of the agents corresponding to each task node in the agent assignment results; the latency score is obtained by the estimated completion latency of the candidate directed acyclic cooperative structure relative to the deadline of shelf inspection task 300; the communication score is obtained by the number of task edges in the candidate directed acyclic cooperative structure relative to the maximum number of task edges allowed by the task template; and the resource conflict score is obtained by the number of task node pairs with resource conflicts in the same topology layer. All of the above scores are normalized to... to The dimensionless quantity between them, the larger the value, the better the corresponding index.

[0084] ; ; ; ; in, To rate the quality, For latency scoring, Scoring for communication Score resource conflicts. Assign results to the agent. For task nodes The historical success rate of the corresponding agent in executing this task node. The expected completion time of candidate directed acyclic cooperative structures under agent assignment results. The maximum allowable delay corresponding to the deadline of shelf inspection task 300. This represents the maximum number of task edges allowed in the task template. This represents the number of resource conflicts. If the value is positive, the delay scoring formula described above will be used; if the shelf inspection task 300 does not have a deadline configured, the default maximum allowable delay in the task template will be used; if Then, candidate graph scoring module 180 uses the default maximum allowable delay instead. Perform the calculation. If Communication score is recorded as .

[0085] ; in, For general rating, , , and All are dimensionless fixed weights and satisfy the following conditions: Without further configuration, , , , The above default values ​​prioritize quality and latency over communication and resource conflicts and can be pre-configured by task templates.

[0086] The candidate graph scoring module 180 adds physical consistency score and reverse violation ratio to the general scoring to form an enhanced score.

[0087] ; in, To enhance the rating, and Fixed weights. and All are dimensionless non-negative weights, with the default value being... , When the bottleneck skeleton is used in a strongly constrained scenario, the response sequence of the shelving under load is determined by the load. Not less than This is to increase the penalty intensity for the opposite reachable direction.

[0088] The candidate graph scoring module 180 is also used to mark candidate directed acyclic cooperative structures containing directed cycles or missing permissions as invalid. The criterion for missing permissions is: at least one task node in the candidate directed acyclic cooperative structure lacks an executable agent, or the agent's permission state in the agent assignment result does not allow execution of the corresponding task node. The criterion for the presence of a directed cycle is: any task node in the transitive closure of the candidate directed acyclic cooperative structure is reachable from itself.

[0089] The target collaborative structure solidification module 190 is used to determine the target collaborative directed acyclic graph based on the enhancement score. For example... Figure 1 As shown, the candidate graph scoring module 180 outputs the enhanced score to the target collaborative structure solidification module 190, and the target collaborative structure solidification module 190 then outputs the target collaborative directed acyclic graph to the topology scheduling execution module 200.

[0090] The target collaborative structure solidification module 190 selects the candidate directed acyclic collaborative structure with the highest enhancement score and the most effective result from the candidate results output by the candidate graph scoring module 180 as the target collaborative directed acyclic graph. If multiple candidate directed acyclic collaborative structures have the same enhancement score, the target collaborative structure solidification module 190 selects the candidate directed acyclic collaborative structure with fewer task nodes, fewer task edges, and shorter expected completion delay in sequence. This sorting rule remains unchanged within the same shelf inspection task 300.

[0091] The target collaboration structure solidification module 190 stores the target collaboration directed acyclic graph and its version information. Version information includes the task instance number corresponding to shelf inspection task 300, a summary of the bottleneck skeleton indicating the shelf load response sequence, node task mapping, agent assignment results, enhancement score, and generation time. This version information is used for the subsequent generation of audit log 410.

[0092] The topology scheduling and execution module 200 is used to schedule agents to execute target cooperative directed acyclic graphs according to topology layers. The topology scheduling and execution module 200 calculates the topology layer of each task node in the target cooperative directed acyclic graph, prioritizing the execution of task nodes that are ready for input and have no resource conflicts. After a task node completes its task, the topology scheduling and execution module 200 releases its successor task nodes, allowing the successor task nodes to enter the executable state when the input is ready.

[0093] The topology scheduling execution module 200 outputs execution status and intermediate results to the execution monitoring and intermediate result verification module 210. Execution status includes task node startup, running, completion, timeout, interface failure, and output error states. Intermediate results include data reading results, strain analysis results, vibration analysis results, structural risk assessment results, and intermediate outputs from the risk summary node.

[0094] The execution monitoring and intermediate result verification module 210 is used to monitor the execution status of task nodes in the target collaborative directed acyclic graph and to verify intermediate results. For example... Figure 1 As shown, the topology scheduling execution module 200 outputs the execution status and intermediate results to the execution monitoring and intermediate result verification module 210. When the verification passes, the execution monitoring and intermediate result verification module 210 outputs the valid intermediate results to the result summary and audit log module 230. When a task node fails to execute, it outputs a failure flag to the incomplete subgraph replanning module 220.

[0095] The execution monitoring and intermediate result verification module 210 verifies the following: field integrity, data type, target shelf area consistency, completion criterion fulfillment, and interface return status. Field integrity is determined according to the output field list of the task node; data type is determined according to the field types specified in the task template; target shelf area consistency is determined according to the task instance number and target shelf area; and completion criterion fulfillment is determined according to the completion criteria generated by the task receiving and input standardization module 110.

[0096] The criteria for determining a timeout state are: the actual execution time of a task node exceeds twice the expected delay of that task node, or exceeds the deadline of shelf inspection task 300. The criteria for determining an interface failure state are: the agent interface returns an error code, the connection is interrupted, or there is no response after two consecutive calls. The criteria for determining an output error state are: the intermediate result lacks necessary fields, the field type is incorrect, or the target shelf area is inconsistent. For recoverable interface failure states, the execution monitoring and intermediate result verification module 210 will retry a maximum of two times; if it still fails after two retries, a failure flag will be output.

[0097] The incomplete subgraph replanning module 220 is used to freeze completed task nodes, delete unusable agent assignments, and re-call the candidate directed acyclic cooperative graph search module 160, agent assignment module 170, candidate graph scoring module 180, and target cooperative structure solidification module 190 for incomplete task nodes, so that the regenerated alternative cooperative subgraph is still subject to structural partial order labels, bottleneck node labels, and enhanced scoring constraints.

[0098] like Figure 1 As shown, the incomplete subgraph replanning module 220 sends back the incomplete task nodes and the updated agent state to the candidate directed acyclic cooperative structure search module 160 via the dashed arrow. The updated agent state is derived from the failure flag output by the execution monitoring and intermediate result verification module 210 and the current state read by the agent registration information reading module 150, including at least the offline state, interface failure state, resource occupancy state, and permission state. The candidate directed acyclic cooperative structure search module 160 uses the incomplete task nodes to regenerate the candidate directed acyclic cooperative structure, and the agent assignment module 170 uses the updated agent state to re-form the agent assignment result.

[0099] The incomplete subgraph replanning module 220 only replans incomplete task nodes and does not repeat the execution of task nodes that have already produced valid intermediate results. If multiple anomalies occur simultaneously, the incomplete subgraph replanning module 220 processes them in the following order: missing permissions, interface failure status, resource conflict, output error status, and timeout status. The execution monitoring and intermediate result verification module 210 records the task node, failure type, failing agent, and failure time corresponding to the failure marker as anomaly task features; the incomplete subgraph replanning module 220 sends the anomaly task features back to the candidate directed acyclic cooperative structure search module 160. The candidate directed acyclic cooperative structure search module 160 increases the failure priority of the corresponding incomplete task node and does not select task node and agent combinations with exactly the same anomaly task features within the same replanning cycle. Missing permissions have the highest priority because missing permissions will directly cause the corresponding task node to be unexecutable.

[0100] The results aggregation and audit log module 230 is used to merge all valid intermediate results, generate a risk report 400, and save the audit log 410. The risk report 400 may include the target shelf area, a summary of the bottleneck skeleton corresponding to the shelf load event's response sequence, structural risk assessment results, risk aggregation node output, and report generation task output. The audit log 410 may include task input, a summary of the bottleneck skeleton for the shelf load response sequence, a target collaborative directed acyclic graph, agent assignment results, enhancement scores, failure markers, replanning records, and operation logs.

[0101] For example, in a complete workflow embodiment, the automated warehouse system receives a shelf inspection task 300 to inspect the 12th row of shelves in area A and output the risk level. The task receiving and input standardization module 110 first identifies the target shelf area, output requirements, deadline, permission scope and completion criteria, forms a standardized task object, and outputs the standardized task object to the atomic subtask splitting module 120.

[0102] The atomic subtask decomposition module 120 generates a set of atomic subtasks and general dependencies based on standardized task objects and existing task templates. The set of atomic subtasks includes data reading tasks, strain analysis tasks, vibration analysis tasks, structural risk assessment tasks, risk summary nodes, and report generation tasks. The general dependencies require that the data reading task precedes the strain analysis and vibration analysis tasks, and that the structural risk assessment task precedes the report generation task.

[0103] When the same rack loading event occurs, strain sensors collect strain time series 310 at the measuring points, acceleration sensors collect acceleration time series 320 at the measuring points, and installation records or equipment positioning provide spatial coordinates of the measuring points 330. The bottleneck skeleton construction module 130 receives the above three types of directly collected data in the rack loading response sequence, and the dimensionless data processing unit 131 calculates the strain change, normalized strain change, and normalized acceleration within a unified event window.

[0104] The joint response intensity forming unit 132 generates the joint response intensity, and the response directed acyclic graph generation unit 133 determines the first significant response time based on the joint response intensity, and establishes a response directed acyclic graph from the earlier significant response measurement point to the later significant response measurement point by combining the spatial distance obtained from the spatial coordinates of the measurement points and the response similarity obtained from the joint response intensity. The shelf load response priority bottleneck skeleton extraction unit 134 calculates the bottleneck degree and forms the shelf load response priority bottleneck skeleton according to the preset high quantile screening conditions.

[0105] The structural partial order label generation module 140 maps nodes in the bottleneck skeleton of the shelf load response sequence to structural analysis objects, establishes a node task mapping, maps directed edges in the bottleneck skeleton of the shelf load response sequence to structural partial order labels, and sets bottleneck node labels for the analysis tasks corresponding to bottleneck nodes. The candidate directed acyclic collaborative structure search module 160 extends the candidate directed acyclic collaborative structure based on the general dependency relationship, and uses structural partial order label pruning to reverse the candidate branches of the structural partial order labels. At the same time, it uses bottleneck node labels to ensure that the analysis tasks corresponding to bottleneck nodes are completed before the risk summary nodes.

[0106] The agent registration information reading module 150 reads the agent registry 340 to form an agent capability matrix. The agent assignment module 170 maps task nodes in the candidate directed acyclic cooperative structure to executable agents based on the agent capability matrix. The candidate graph scoring module 180 calculates the transitive closure of the candidate directed acyclic cooperative structure, determines the physical consistency score based on the degree to which directed edges in the bottleneck skeleton remain reachable in the same direction in the transitive closure based on the order of shelf load response, determines the reverse violation ratio based on the degree to which opposite reachable directions appear, and forms an enhancement score by combining the general score, the physical consistency score, and the reverse violation ratio.

[0107] The target collaboration structure solidification module 190 determines the target collaboration directed acyclic graph based on the enhancement score. The topology scheduling and execution module 200 schedules agents to execute the target collaboration directed acyclic graph according to the topology layer. The execution monitoring and intermediate result verification module 210 verifies the execution status and intermediate results. If there are failure markers, the incomplete subgraph replanning module 220 freezes the completed task nodes and replans the incomplete task nodes. If all necessary task nodes produce valid intermediate results, the result summary and audit log module 230 outputs a risk report 400 and an audit log 410.

[0108] As an optional implementation, in the normal operating condition embodiment, the shelf loading event is a single pallet placement event. The strain time series 310 and acceleration time series 320 at the measuring points are both fully acquired within a unified event window, and the spatial coordinates 330 of the measuring points are provided by the installation record. The dimensionless data processing unit 131 obtains the normalized strain change and normalized acceleration based on the strain normalization constant and acceleration normalization constant; the joint response intensity forming unit 132 obtains the joint response intensity; and the directed acyclic graph (DAG) generation unit 133 establishes the DAG based on the first significant response time, spatial distance, and response similarity.

[0109] In this normal operating condition embodiment, all directed edges in the directed acyclic graph of the response satisfy the strict increasing first significant response time. The bottleneck skeleton extraction unit 134, based on the bottleneck degree, determines the set of bottleneck nodes and retains the directed response paths passing through the measurement points in the bottleneck node set. The structural partial order label generation module 140 generates structural partial order labels and bottleneck node labels from the bottleneck skeleton of the shelf load response sequence. The candidate directed acyclic collaborative structure search module 160 thereby reduces candidate branches that are inconsistent with the current shelf load event response sequence.

[0110] The candidate graph scoring module 180 adds physical consistency score and reverse violation ratio to the general scoring. The target collaborative structure solidification module 190 selects the candidate directed acyclic collaborative structure with the highest enhancement score and the most effective as the target collaborative directed acyclic graph. Finally, the results summary and audit log module 230 generates a risk report 400 and saves the audit log 410.

[0111] As an alternative implementation, in the feature-triggered scenario embodiment, a loading event on the same shelf causes a measurement point near a central upright to show a significant response earlier, while a measurement point near an adjacent beam shows a significant response later. If the candidate directed acyclic cooperative structure search is performed solely based on general dependencies, it may result in candidate branches used by the analysis task corresponding to the later response measurement point before those corresponding to the earlier response measurement point.

[0112] In this embodiment, the bottleneck skeleton construction module 130 constructs a directed acyclic graph of responses from earlier significant response points to later significant response points based on the time of the first significant response, spatial distance, and response similarity, and identifies bottleneck nodes by the degree of bottleneck. The structural partial order label generation module 140 converts this sequence relationship into structural partial order labels and sets bottleneck node labels for the analysis tasks corresponding to the bottleneck nodes.

[0113] When expanding candidate branches, the candidate directed acyclic cooperative structure search module 160 prunes the candidate branch once it detects a reverse structure partial order label. For candidate directed acyclic cooperative structures that are not pruned but still have opposite reachable directions in the transitive closure, the candidate graph scoring module 180 enhances the score by reducing the reverse violation ratio. Through this processing, the system reflects the constraint effect of the bottleneck skeleton of the shelf load response sequence at both the candidate search and candidate selection levels.

[0114] As an optional implementation, in the abnormal fault-tolerant operating condition embodiment, some measuring points do not reach the significant response criterion determined by the trigger threshold within a unified event window. The bottleneck skeleton construction module 130 for the shelf load response sequence does not use these measuring points to build the directed edges in the directed acyclic graph of the response, but instead continues to use other measuring points that have reached the significant response criterion to build the directed acyclic graph of the response. This process avoids the unstable impact of non-significant response measuring points on the calculation of the first significant response time and the degree of bottleneck.

[0115] As an alternative implementation, in another fault-tolerant embodiment, the set of directed edges in the response graph is empty or the set of directed edges in the bottleneck skeleton is empty. In this case, the candidate directed acyclic cooperative structure search module 160 can revert to a search based solely on general dependencies, and the candidate graph scoring module 180 records the physical consistency score as... The reverse violation ratio is denoted as This process ensures that the system can still generate the agent collaborative structure according to the original general dependencies even when there is a bottleneck skeleton for effective shelf load response.

[0116] As an optional implementation, in the abnormal operating condition embodiment, the agent interface corresponding to an analysis task in the target cooperative directed acyclic graph fails. The execution monitoring and intermediate result verification module 210 outputs a failure flag and forms abnormal task features. The incomplete subgraph replanning module 220 freezes completed task nodes, deletes unusable agent assignments, uses the abnormal task features as constraint inputs for the next round of candidate search, and re-invokes the candidate directed acyclic cooperative structure search module 160, agent assignment module 170, candidate graph scoring module 180, and target cooperative structure solidification module 190 for incomplete task nodes. The regenerated alternative cooperative subgraph is still subject to structural partial order labels, bottleneck node labels, and enhanced scoring constraints.

[0117] This implementation can be executed by a processor in a server, industrial computer, edge computing device, or warehouse inspection platform. The processor can run program instructions for implementing the task receiving and input normalization module 110, the atomic subtask splitting module 120, the shelf load response priority bottleneck skeleton construction module 130, the structural partial order label generation module 140, the agent registration information reading module 150, the candidate directed acyclic cooperative structure search module 160, the agent assignment module 170, the candidate graph scoring module 180, the target cooperative structure solidification module 190, the topology scheduling execution module 200, the execution monitoring and intermediate result verification module 210, the incomplete subgraph replanning module 220, and the result summary and audit log module 230.

[0118] The strain time series 310 at the measuring point can come from a strain sensor or its data acquisition unit; the acceleration time series 320 at the measuring point can come from an acceleration sensor or its data acquisition unit; and the spatial coordinates 330 at the measuring point can come from installation records, routine measurement records, or equipment positioning records. The agent registry 340 can be stored in a database, configuration file, or service registry. All of the above data sources are standard data sources in automated warehouse racking inspection systems.

[0119] System 100 can read the strain time series 310, acceleration time series 320 and spatial coordinates 330 of the measuring point through the data access interface, receive the shelf inspection task 300 through the task parsing interface, read the intelligent agent registry 340 through the intelligent agent registration information reading module 150, and save the audit log 410 through the log storage interface.

[0120] In this embodiment, the trigger threshold, strain normalization constant, acceleration normalization constant, response similarity normalization constant, preset distance, response similarity threshold, preset high quantile filtering conditions, quantile points, graph search width, and fixed weight can all be preset as system configuration items.

[0121] Through the shelf load response sequence bottleneck skeleton construction module 130, this embodiment converts the measurement point strain time series 310, measurement point acceleration time series 320 and measurement point spatial coordinates 330 into a response directed acyclic graph and a shelf load response sequence bottleneck skeleton, so that the intelligent agent collaborative structure dynamic generation process can use the structured response evidence in the current shelf load event.

[0122] By using the structural partial order label generation module 140 and the candidate directed acyclic cooperative structure search module 160, this embodiment transforms the bottleneck skeleton of the shelf load response sequence into structural partial order labels and bottleneck node labels, and prunes and reverses the candidate branches of the structural partial order labels during the candidate branch expansion process, thereby reducing the number of candidate directed acyclic cooperative structures that formally meet the directed acyclic requirements but are inconsistent with the current shelf load response sequence.

[0123] Through the candidate graph scoring module 180, this implementation introduces physical consistency scores and reverse violation ratios on the basis of general scoring, and selects target cooperative directed acyclic graphs through enhanced scoring. This processing ensures that the selection of candidate cooperative directed acyclic graphs considers not only quality, latency, communication, and resource conflicts, but also the degree of same-direction and opposite-direction reachability of directed edges in the bottleneck skeleton of the shelf load response sequence in the candidate cooperative directed acyclic graphs.

[0124] Through the incomplete subgraph replanning module 220, this embodiment can freeze completed task nodes, delete unusable agent assignments, and regenerate alternative collaborative subgraphs for incomplete task nodes when task node execution fails. Since the alternative collaborative subgraph is still subject to structural partial order labels, bottleneck node labels, and enhancement scores, the system maintains consistency with the bottleneck skeleton of the shelf load response sequence during the replanning process.

[0125] Through the result summary and audit log module 230, this implementation can save the audit log 410 in addition to the risk report 400, recording task input, bottleneck skeleton summary of shelf load response sequence, target collaborative directed acyclic graph, agent assignment results, enhancement scoring and replanning records, so that the agent collaborative structure generation process of shelf inspection task 300 is traceable.

[0126] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A dynamic generation system for agent cooperative structures based on directed acyclic graph search, comprising a task receiving and input normalization module, an atomic subtask decomposition module, a candidate directed acyclic cooperative structure search module, a candidate graph scoring module, and a target cooperative structure solidification module, characterized in that, It also includes a bottleneck skeleton construction module for the order of shelf load response and a structural partial order label generation module; The bottleneck skeleton construction module for the sequential response of the rack under load is used to determine the first significant response time, spatial distance, response similarity and bottleneck degree based on the strain time series, acceleration time series and spatial coordinates of the measuring points in the same rack under load event, and to form a directed acyclic graph of response and a bottleneck skeleton for the sequential response of the rack under load. The structural partial order label generation module is used to map the bottleneck skeleton of the shelf load response sequence into structural partial order labels and bottleneck node labels. The candidate directed acyclic cooperative structure search module is used to prune candidate directed acyclic cooperative structures based on the partial order label of the structure and the bottleneck node label. The candidate graph scoring module is used to obtain a general score based on quality, latency, communication and resource conflicts. Based on the degree to which the directed edges in the bottleneck skeleton of the shelf load response remain in the same direction and the degree to which opposite directions of reachability appear in the candidate directed acyclic cooperative structure, the physical consistency score and the reverse violation ratio are determined respectively, and the enhancement score is determined. The target collaborative structure solidification module is used to determine the target collaborative directed acyclic graph based on the enhancement score.

2. The intelligent agent cooperative structure dynamic generation system based on directed acyclic graph search according to claim 1, characterized in that, The bottleneck skeleton construction module for the sequential response of the shelf under load is used to perform dimensionless processing on the strain time series and acceleration time series of the measuring points respectively within a unified event window, and to form a joint response intensity from the dimensionless strain time series and acceleration time series of the measuring points. The joint response intensity is used to determine the first significant response time of each measuring point in the same shelf under load event.

3. The intelligent agent cooperative structure dynamic generation system based on directed acyclic graph search according to claim 2, characterized in that, The bottleneck skeleton construction module for the sequential load response of the shelving is used to establish a directed acyclic graph of the response from earlier significant response measurement points to later significant response measurement points based on the sequential relationship of the first significant response time, the spatial distance obtained from the spatial coordinates of the measurement points, and the response similarity obtained from the joint response intensity.

4. The intelligent agent cooperative structure dynamic generation system based on directed acyclic graph search according to claim 3, characterized in that, The bottleneck skeleton construction module for the sequential load response of the shelving is used to determine the measurement points whose bottleneck degree meets the preset high quantile screening conditions from the directed acyclic graph of the response according to the preset high quantile screening conditions, and retain the directed response paths that pass through the measurement points that meet the preset high quantile screening conditions, so as to form the bottleneck skeleton for the sequential load response of the shelving.

5. The intelligent agent cooperative structure dynamic generation system based on directed acyclic graph search according to claim 1, characterized in that, The structural partial order label generation module is used to map the nodes in the bottleneck skeleton of the shelf load response sequence to structural analysis objects, map the directed edges in the bottleneck skeleton of the shelf load response sequence to the structural partial order labels used when the analysis results of earlier responses are summarized before those of later responses, and set the bottleneck node labels for the analysis tasks corresponding to the bottleneck nodes.

6. The intelligent agent cooperative structure dynamic generation system based on directed acyclic graph search according to claim 5, characterized in that, The candidate directed acyclic cooperative structure search module is used to search based on the general dependency relationship formed by the atomic subtask splitting module. When expanding the task edges, verification nodes, or summary nodes of the candidate directed acyclic cooperative structure and forming parallel branches, it checks whether the candidate directed acyclic cooperative structure reverses the partial order label of the structure and prunes the candidate branches that reverse the partial order label of the structure.

7. The intelligent agent cooperative structure dynamic generation system based on directed acyclic graph search according to claim 6, characterized in that, The candidate directed acyclic collaborative structure search module is also used to ensure that the analysis task corresponding to the bottleneck node is completed before the risk summary node in the candidate directed acyclic collaborative structure, based on the bottleneck node label, and to form the parallel branch, so that the atomic subtasks corresponding to the measurement points with the same first significant response time can be carried out in parallel when there is no general dependency relationship.

8. The intelligent agent cooperative structure dynamic generation system based on directed acyclic graph search according to claim 1, characterized in that, The candidate graph scoring module is used to calculate the transitive closure of the candidate directed acyclic cooperative structure. It determines the physical consistency score based on the degree to which the directed edges in the bottleneck skeleton of the shelf load response remain in the same direction in the transitive closure, and determines the reverse violation ratio based on the degree to which the directed edges in the bottleneck skeleton of the shelf load response appear in opposite directions in the transitive closure.

9. The intelligent agent cooperative structure dynamic generation system based on directed acyclic graph search according to claim 8, characterized in that, The candidate graph scoring module is used to establish a node-task mapping between the nodes in the bottleneck skeleton of the shelf load response sequence and the task nodes in the candidate directed acyclic collaborative structure. Based on the node-task mapping, it determines whether the directed edges in the bottleneck skeleton of the shelf load response sequence remain in the same direction and whether there are opposite reachable directions in the candidate directed acyclic collaborative structure. Candidate directed acyclic collaborative structures with directed cycles or missing permissions are marked as invalid.

10. The intelligent agent cooperative structure dynamic generation system based on directed acyclic graph search according to claim 1, characterized in that, It also includes a module for reading agent registration information, a module for assigning agents, and a module for replanning incomplete subgraphs; The agent registration information reading module is used to form an agent capability matrix, and the agent assignment module is used to map task nodes to executable agents according to the agent capability matrix. The incomplete subgraph replanning module is used to freeze completed task nodes, delete unusable agent assignments, and re-invoke the candidate directed acyclic cooperative graph search module, the agent assignment module, the candidate graph scoring module, and the target cooperative structure solidification module for incomplete task nodes, so that the regenerated alternative cooperative subgraph is still subject to the structural partial order label, the bottleneck node label, and the enhanced scoring constraints.

Images