Multi-agent collaborative calling method and system based on knowledge graph association

Abstract

The invention discloses a multi-agent collaborative calling method and system based on knowledge graph association, and relates to the technical field of data processing, and the method comprises the steps: carrying out semantic analysis on a task original demand text to extract a task feature vector; then a multi-agent semantic knowledge graph is called for matching, a target agent set is recognized, and a local calling sub-graph is constructed; then, the initial association weights of the sub-graphs are dynamically corrected based on the task feature vectors, and a real-time association matrix is obtained; then quantizing cooperation priority and dependence intensity, executing redundancy elimination based on semantic projection gradient, and generating a calling sequence in combination with a bottom layer physical load state; and finally, acquiring real-time physical load and resource occupation conditions during execution, executing conflict scheduling based on the dynamic access quota model, and issuing an adjustment instruction. The method has the advantages of accurate collaboration, dynamic self-adaption and good calling effect.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a multi-agent collaborative invocation method and system based on knowledge graph association. Background Technology

[0002] In complex industrial automation and distributed computing fields, multi-agent collaborative invocation systems are the core infrastructure for realizing high-dimensional business process automation. However, existing multi-agent collaborative technologies, when faced with unstructured natural language instructions, mainly rely on shallow keyword matching and lack the ability to deeply deconstruct the semantics of task actions, objects, and complex constraints, as well as the ability to fuse multi-dimensional vectors, resulting in insufficient accuracy in initial task distribution and node matching.

[0003] In terms of collaborative topology construction, traditional knowledge graphs generally use static association weights, which cannot dynamically calculate projection strength and adaptively adjust the association gain between nodes based on the semantic context of the current specific task. This results in rigid call chains and slow responses. At the same time, in large-scale intelligent agent clusters, existing redundancy removal mechanisms are limited to comparing the similarity of single node attributes. This makes them prone to misjudging complementary intelligent agents with different actual efforts as redundant nodes and removing them in complex task contexts, causing serious functional errors and wasted computing power.

[0004] More critically, existing collaborative scheduling sequence generation and queuing mechanisms are typically limited to purely software-level logical deduction, completely severing the causal mapping relationship between upper-layer semantic priorities and the real-time physical state of the underlying hardware. When the underlying physical nodes are nearing full load, the system still forcibly dispatches tasks according to static logical timing. Furthermore, when multiple agents compete for the same common logical resources, there is a lack of a dynamic underlying access quota control mechanism based on core business weights. This can easily lead to severe exponential delays in hardware response, execution deadlocks, and data throughput collapses in core business links, ultimately making it difficult for the entire multi-agent collaborative architecture to guarantee physical feasibility and overall stability in dynamic high-concurrency scenarios. Summary of the Invention

[0005] In response to the technical problems described in the background section, this invention provides a multi-agent collaborative invocation method and system based on knowledge graph association.

[0006] A multi-agent collaborative invocation method based on knowledge graph association includes: obtaining the original requirement text of the task to be executed, and performing semantic parsing on the original requirement text to extract task feature vectors; retrieving a preset multi-agent semantic knowledge graph, matching the task feature vectors with each agent node in the multi-agent semantic knowledge graph, and identifying a target agent set; constructing a local invocation subgraph based on preset association edges between agent nodes in the target agent set, and dynamically correcting the initial association weights in the local invocation subgraph based on the task feature vectors to obtain a real-time association matrix; quantifying the collaboration priority and dependency strength between agent nodes based on the real-time association matrix, performing redundancy removal based on semantic projection gradients, and generating a multi-agent collaborative invocation sequence by combining the underlying physical load state and the dependency strength; obtaining the real-time physical load state and logical resource occupancy of each agent node when executing the multi-agent collaborative invocation sequence, performing conflict scheduling based on a dynamic access quota model, and issuing collaborative adjustment instructions.

[0007] Optionally, the original requirement text of the task to be executed is obtained, and semantic parsing is performed on the original requirement text to extract the task feature vector, including: identifying the task verbs, operation objects, and constraints in the original requirement text, and mapping them to a high-dimensional semantic space respectively to obtain the corresponding verb component vector, object component vector, and constraint component vector; obtaining verb weight coefficients, object weight coefficients, and constraint weight coefficients based on a preset semantic importance and summing to one; and using the verb weight coefficients, object weight coefficients, and constraint weight coefficients, performing a weighted fusion operation on the verb component vectors, object component vectors, and constraint component vectors to obtain the task feature vector representing the global semantic features of the task to be executed.

[0008] Optionally, dynamically correcting the initial association weights in the local call subgraph based on the task feature vector to obtain a real-time association matrix includes: calculating the cosine alignment of the task feature vector with the sum of the attribute feature vectors of two connected agent nodes in the local call subgraph, as the semantic projection strength of the task feature vector between the two connected agent nodes; constructing a positive gain ratio by multiplying the semantic projection strength by a preset adjustment coefficient; multiplying the initial association weights between the two agent nodes by one and the sum of the positive gain ratio to obtain the corrected real-time association weights; and arranging all the corrected real-time association weights according to the node index to construct the real-time association matrix.

[0009] Optionally, redundancy removal is performed based on semantic projection gradient, including: calculating the cosine similarity of the attribute feature vectors of any two agent nodes in the target agent set; calculating the absolute value of the difference between the projection scalars of the attribute feature vectors of the two agent nodes on the task feature vector, as the projection gradient difference; when the cosine similarity is greater than a preset redundancy threshold and the projection gradient difference is less than a preset gradient interference threshold, it is determined that the two agent nodes have functional redundancy, and the agent node with the higher corrected real-time association weight in the real-time association matrix is ​​retained, and redundancy removal is performed.

[0010] Optionally, a multi-agent collaborative invocation sequence is generated by combining the underlying physical load state and the dependency strength, including: obtaining the data interaction handle release time of the upstream agent node (the time point when the task is completed and resources are released), and the real-time physical comprehensive load rate of the downstream agent node; multiplying a preset baseline system response constant by an exponential term with the product of a preset hardware penalty coefficient and the real-time physical comprehensive load rate as the exponent, with the natural constant as the base, to obtain the load penalty response time; dividing the load penalty response time by the corrected real-time association weight between the upstream agent node and the downstream agent node to obtain the additional delay time; adding the data interaction handle release time to the additional delay time to determine the start timestamp of the downstream agent node, and generating the multi-agent collaborative invocation sequence according to the chronological order of the start timestamps.

[0011] Optionally, performing conflict scheduling based on a dynamic access quota model and issuing collaborative adjustment instructions includes: when it is predicted that multiple agent nodes are competing for the same logical resource within the same predetermined time period, retrieving the corresponding modified real-time association weights in the real-time association matrix as heuristic factors to calculate the competition priority; adding one to the modified real-time association weights and taking the natural logarithm to obtain a logarithmic increment; multiplying the preset system basic access quota distribution rate by one and the sum of the logarithmic increment to calculate the dynamic quota replenishment rate specific to each competing agent node; opening the underlying physical communication channel with the corresponding specific dynamic quota replenishment rate, and allocating logical resource access permissions to the agent nodes according to the competition priority.

[0012] A multi-agent collaborative invocation system based on knowledge graph association is also provided. The system includes: a task parsing module, used to obtain the original requirement text of the task to be executed and perform semantic parsing on the original requirement text to extract task feature vectors; a graph management module, used to retrieve a preset multi-agent semantic knowledge graph, the multi-agent semantic knowledge graph containing multiple agent nodes, functional attribute labels corresponding to each agent node, and preset association edges between nodes; and an association quantization module, used to match the task feature vectors with each agent node in the multi-agent semantic knowledge graph, identify the target agent set, and, based on... The target agent set constructs a local call subgraph, dynamically corrects the initial association weights based on the task feature vector, and obtains a real-time association matrix; the sequence generation module is used to quantify the cooperation priority and dependency strength between each agent node according to the real-time association matrix, perform redundancy removal based on semantic projection gradient, and generate a multi-agent collaborative call sequence by combining the underlying physical load state and the dependency strength; the conflict scheduling module is used to obtain the real-time physical load state and logical resource occupancy of each agent node when executing the multi-agent collaborative call sequence, perform conflict scheduling based on a dynamic access quota model, and issue collaborative adjustment instructions.

[0013] Optionally, the task parsing module is further configured to: identify the task verbs, operation objects, and constraints in the original requirement text, and map them to a high-dimensional semantic space respectively to obtain the corresponding verb component vectors, object component vectors, and constraint component vectors; obtain verb weight coefficients, object weight coefficients, and constraint weight coefficients that are preset based on semantic importance and sum to one; and perform a weighted fusion operation on the verb component vectors, object component vectors, and constraint component vectors using the verb weight coefficients, object weight coefficients, and constraint weight coefficients to obtain the task feature vector that represents the global semantic features of the task to be executed.

[0014] Optionally, the association quantization module is further configured to: calculate the cosine alignment of the task feature vector with the sum of the attribute feature vectors of two connected agent nodes in the local call subgraph, as the semantic projection strength of the task feature vector between the two connected agent nodes; construct a positive gain ratio by multiplying the semantic projection strength by a preset adjustment coefficient; multiply the initial association weight between the two agent nodes by one and the sum of the positive gain ratio to obtain the corrected real-time association weight; and arrange all the corrected real-time association weights according to the node index to construct the real-time association matrix.

[0015] Optionally, the sequence generation module is further configured to: calculate the cosine similarity of the attribute feature vectors of any two agent nodes in the target agent set; calculate the absolute value of the difference between the projection scalars of the attribute feature vectors of the two agent nodes onto the task feature vector, as the projection gradient difference; when the cosine similarity is greater than a preset redundancy threshold and the projection gradient difference is less than a preset gradient interference threshold, determine that the two agent nodes have functional redundancy, retain the agent node with the higher corrected real-time association weight in the real-time association matrix, and perform a redundancy removal operation.

[0016] The beneficial effects of this invention are reflected in: In the entire knowledge graph-based multi-agent collaborative invocation method, firstly, at the task parsing end, weighted fusion operations with clear mathematical boundaries are performed on multi-dimensional semantic components, effectively reducing the ambiguity and vague interference of unstructured natural language, and anchoring a clear and reproducible digital benchmark for subsequent accurate matching of execution entities. Secondly, a multiplicative gain model based on semantic projection intensity is introduced, endowing the static knowledge graph with scene awareness capabilities, enabling the real-time association weights between core collaborative nodes to adaptively amplify as the task logic center shifts, ensuring the responsiveness of complex business paths. More importantly, a physical scale of spatial projection gradient difference is introduced, constructing a strict three-dimensional redundancy elimination mechanism, accurately identifying and eliminating seemingly redundant but highly complementary key nodes in the current task's direction of effort, simplifying the invocation sequence while avoiding false positives. Furthermore, at the physical execution level, by constructing an exponential hardware penalty term that includes real-time physical comprehensive load rate, the dimensionless semantic association weights are mapped to specific physical start timestamps, realizing dynamic adaptation of upper-layer business logic to the underlying hardware capacity. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0018] Figure 1 This is a schematic diagram illustrating the steps of the multi-agent collaborative invocation method based on knowledge graph association in this invention; Figure 2 This is a partial flowchart of S1 and S2 in the knowledge graph-based multi-agent collaborative invocation method of the present invention; Figure 3 This is a partial flowchart of S2, S3, and S4 in the knowledge graph-based multi-agent collaborative invocation method of the present invention; Figure 4 This is a schematic diagram of part of the process of S4 and S5 in the multi-agent collaborative invocation method based on knowledge graph association of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0020] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0021] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0022] like Figures 1 to 4 As shown, a multi-agent collaborative invocation method based on knowledge graph association is provided. In one embodiment, the method includes: S1. Obtain the original requirement text of the task to be executed, and perform semantic parsing on the original requirement text to extract the task feature vector; S2. Retrieve a preset multi-agent semantic knowledge graph, match the task feature vector with each agent node in the multi-agent semantic knowledge graph, and identify the target agent set. S3. Construct a local call subgraph based on the preset association edges between each agent node in the target agent set, and dynamically correct the initial association weights in the local call subgraph based on the task feature vector to obtain the real-time association matrix; S4. Quantify the collaboration priority and dependency strength between each agent node according to the real-time association matrix, perform redundancy removal based on semantic projection gradient, and generate a multi-agent collaborative call sequence by combining the underlying physical load state and the dependency strength. S5. Obtain the real-time physical load status and logical resource occupancy of each agent node when executing the multi-agent collaborative call sequence, execute conflict scheduling based on the dynamic access quota model, and issue collaborative adjustment instructions.

[0023] In this embodiment, it should be noted that the agent node indices involved in S3, S4, and S5 are all variables. and It represents any two execution unit nodes in a multi-agent semantic knowledge graph that have a pre-defined relationship.

[0024] like Figure 2 As shown, in S1, a semantic baseline for task execution and a specific vector fusion mapping mechanism are established. This step first receives unstructured raw requirement text through an input interface, and then performs deep semantic parsing using natural language processing techniques. In specific implementation, three key dimensions are extracted: task verbs, operation objects, and constraints. These three dimensions are then mapped to a high-dimensional semantic space to obtain the corresponding verb component vectors, object component vectors, and constraint component vectors.

[0025] To achieve accurate weighted fusion, the system pre-defines fusion weight coefficients based on semantic importance. The specific logical formula for performing weighted fusion operations on each component vector is as follows:

[0026] in: : Represents the task feature vector that characterizes the global semantic features of the task to be executed; : Indicates the preset verb weight coefficient; : Represents the verb component vectors mapped from the task verbs to the corresponding high-dimensional semantic space; : Represents the preset object weight coefficient; : Represents the object component vector mapped to the higher-dimensional semantic space of the operation object; : Indicates the preset constraint weight coefficient; : Represents the constraint component vector mapped to the high-dimensional semantic space.

[0027] Furthermore, the aforementioned weighting coefficients satisfy... The mathematical constraints are defined as follows: Specifically, their values ​​can be determined based on the relative importance of the task verb, the object of operation, and the constraints to the agent's capabilities in a specific application scenario. These values ​​are typically obtained through expert experience or statistical learning from historical task data. For example, in an industrial automated warehousing scenario, the object of operation usually determines the agent's core domain capabilities, hence its highest weight; the task verb defines the specific execution interface and has the next highest weight; and the constraints, as environmental fine-tuning factors, have the lowest weight. The implementation example sets... , , This set of values ​​has been experimentally verified to effectively balance the semantic components, ensuring that the generated task feature vector accurately reflects the business focus. This fusion formula physically clarifies the specific contribution of each semantic component to the final task feature vector, guaranteeing the uniqueness and reproducibility of the mapping from natural language to digital features.

[0028] Furthermore, in this embodiment, it should be noted that after receiving the original requirement text, the system no longer performs simple word frequency statistics, but instead utilizes natural language processing technology to perform deep semantic parsing, deconstructing the unstructured text into three key dimensions: task verbs, operation objects, and constraints. This three-dimensional decomposition logic can accurately cover the core elements of instructions in complex industrial or business scenarios. For example, when the system receives the text "execute the task of emergency outbound and refined quality inspection of high-value fragile electronic components," the parsing module extracts "outbound" and "quality inspection" as task verbs, the core entity "electronic components" as operation objects, and "emergency" and "fragile" as constraints, representing the environment or state boundary. Subsequently, the system maps these three types of features to a preset high-dimensional semantic space, transforming them into digital vectors that can be processed by the underlying logic. Assuming a simplified three-dimensional semantic space, the verb component vectors obtained by the system mapping... for object component vector for Constraint component vector for This conversion process solves the problem that traditional instruction distribution systems cannot convert fuzzy language into machine-measurable indicators. Its beneficial effect is that it transforms human semantics into directional components with clear mathematical and physical meaning, providing a solid data foundation for subsequent quantitative calculations and scheduling.

[0029] After obtaining the three independent component vectors, the system needs to fuse them into a unified task feature vector to represent the global semantics of the task to be executed. To address the technical problem of traditional bag-of-words or average pooling models where features have equal weight, leading to the core business object being easily diluted or masked by minor modifiers, step S1 introduces a weighted fusion formula based on semantic importance. And set The mathematical constraints are as follows. The reason for adopting this computational logic is that different semantic components objectively have different proportions of the agent's execution capabilities required. In collaborative scheduling scenarios, the operation object usually determines the basic domain capabilities that the agent must possess, and its importance is relatively high; the task verb defines the specific execution interface; and the constraints are fine-tuned as surrounding environmental variables. If the system presets verb weight coefficients... Object weight coefficient Constraint weight coefficients Substituting the aforementioned spatial data into the weighted fusion operation yields: ,Right now The final generated task feature vector is This computational logic mitigates the semantic drift risk during feature space mapping. Its beneficial effect is that the fused task feature vector can more accurately reflect the actual business focus distribution, thus providing an objective digital comparison benchmark for cosine similarity matching with multi-agent knowledge graph nodes in step S2, and improving the accuracy of initial node selection.

[0030] like Figure 2 As shown, in S2, the most matching execution entity is selected from the multi-agent semantic knowledge graph. This process first retrieves the preset multi-agent semantic knowledge graph, which consists of multiple agent nodes. Each agent node is bound to a corresponding functional attribute label, and the system obtains the attribute feature vector corresponding to the functional attribute label of each agent node.

[0031] Subsequently, the cosine similarity between the task feature vector and each attribute feature vector is calculated. By setting a preset matching threshold, agent nodes with a cosine similarity greater than the preset matching threshold are selected and included in the target agent set. The preset matching threshold can be set according to the node size of the knowledge graph, the task accuracy requirements, and the desired balance between recall and precision, and generally ranges from 0.8 to 0.95. The higher the threshold, the higher the semantic alignment between the selected nodes and the task, but the lower the recall; conversely, the recall increases but may introduce weakly relevant nodes. In this embodiment, it is set to 0.85, which is an empirical value obtained through cross-validation in a typical warehousing and logistics scenario, which can effectively filter out interfering nodes while ensuring that core nodes are not missed.

[0032] Furthermore, in this embodiment, it should be noted that, following the task feature vector representing the global semantic features of the task to be executed generated in step S1, the system first retrieves a preset multi-agent semantic knowledge graph. This graph pre-decomposes the functions of various agent nodes in the cluster and uniformly maps attribute labels representing their processing speed, specialized domain, etc., into attribute feature vectors in the same high-dimensional space. Based on this, the system uses the cosine similarity between the task feature vector and the attribute feature vectors of each agent node as the underlying evaluation index. The reason for adopting the cosine similarity calculation logic is that, in the semantic mapping space, the magnitude of a vector is easily affected by the length of the basic functional description or the number of labels, while cosine similarity removes the influence of the magnitude and specifically measures the cosine value of the angle between two vectors in the spatial geometric distribution direction. When the angle is small and the cosine value is close to one, it indicates that the basic functional distribution of the agent is consistent with the business focus direction of the current task. This logic solves the problem that simply relying on metrics such as Euclidean distance can easily lead to misjudgment of similarity due to differences in the order of magnitude of attribute labels. By transforming functional matching into angular measurements in geometric space, the system can quantify entity capabilities based on objective vector directional associations. The beneficial effect of this process is that it converts natural language-level business requirements into computable mathematical spatial angles, enabling the system to quantify the capability alignment of candidate agents before task allocation. This provides a logically consistent data filtering foundation for the subsequent construction of local call subgraphs.

[0033] Based on the specific data obtained in step S1, we can extrapolate and assume that the fused task feature vector is... At this point, the graph management module traverses each node in the multi-agent semantic knowledge graph. Taking a visual detection agent node as an example, assume its preset attribute feature vector, after mapping, is... The system substitutes these two vectors into the cosine similarity formula, first calculating their dot product as follows: ,Right now The magnitude of the task feature vector is then calculated to be approximately... The magnitude of the feature vector of a visual detection agent node is approximately Dividing the dot product by the product of the moduli of both nodes yields the cosine similarity between the node and the current task, which is approximately [value missing]. After obtaining the cosine similarity of all nodes, the system sets a preset matching threshold (e.g., set to 1). This is used to perform threshold filtering. Due to the similarity of the nodes in the aforementioned visual detection agent... Greater than The system filters them out and categorizes them into the target agent set; while for those whose calculation results are only... Ordinary transport agent nodes are intercepted at this stage. The reason for setting a preset matching threshold for hard blocking is that there are many weak or irrelevant potential connections between nodes in complex knowledge graphs. Including all nodes in the subsequent dynamic correction of the association matrix would lead to ineffective consumption of system computing resources and increase the convergence time of the pathfinding algorithm. This threshold filtering step solves the problem of graph computation redundancy caused by full graph traversal. Its beneficial effect is that, through this quantitative screening process, the system can converge the entire cluster architecture into a core business subset consisting only of relevant entities in the initial stage, control the overall data processing scale of the system, and eliminate interfering nodes with mismatched functional directions before executing physical scheduling, maintaining the efficiency and consistency of system response.

[0034] like Figure 3 As shown, in S3, the correlation response features between agent nodes are captured and dynamically corrected. Based on the identified set of target agents, pre-defined correlation edges between agent nodes are extracted to construct a local call subgraph. The specific steps for dynamically correcting the initial correlation weights in the local call subgraph based on the task feature vector include: Get the first in the local call subgraph The first intelligent agent node and the first The initial association weights between the agent nodes.

[0035] The cosine alignment of the task feature vector with the sum of the attribute feature vectors of the two connected agent nodes in the local call subgraph is calculated as the semantic projection strength of the task feature vector between the two connected agent nodes. A positive gain ratio is constructed by multiplying the semantic projection strength by a preset adjustment coefficient; the initial association weight between the two agent nodes is multiplied by a factor equal to the sum of the positive gain ratio to obtain the corrected real-time association weight.

[0036] The specific calculation formula for the corrected real-time correlation weight is expressed as follows:

[0037] in: : Indicates the corrected real-time correlation weight; : indicates a local call to the first subgraph The first intelligent agent node and the first Initial association weights between agent nodes; : Indicates the preset adjustment coefficient; : Represents the task feature vector; : indicates the first The attribute feature vector corresponding to the functional attribute labels of each agent node; : indicates the first The attribute feature vector corresponding to the functional attribute label of each agent node.

[0038] In this formula, a multiplicative gain model is employed. When the direction of the sum of the task feature vector and the feature vectors of the two agent nodes converges, the gain term takes a maximum positive value, resulting in a substantial increase in the weight of the core association path. Subsequently, all the corrected real-time association weights are arranged according to the node index to construct the real-time association matrix.

[0039] Furthermore, in this embodiment, it should be noted that step S3, following the target agent set selected in the previous steps, establishes a dynamic weight correction mechanism based on semantic projection gradient. After the system extracts the preset initial association edges between target nodes from the underlying database to construct a local call subgraph, in order to overcome the defect that static connection frequency cannot reflect the real-time business focus, this solution introduces a mathematical model to calculate the semantic projection strength of the task feature vector on the association path of adjacent nodes. The reason for adopting the calculation logic of calculating the cosine alignment of the task feature vector with the sum of the feature vectors of two connected agent nodes in this model is that the vector sum of the feature vectors of the two nodes represents the comprehensive functional orientation when this group of collaborative units exert their combined force in geometric space; the smaller the angle between this comprehensive orientation and the current task feature vector, that is, the closer the cosine alignment is to one, the higher the relevance of the collaborative path to achieving the current specific business goal. The calculated semantic projection strength is multiplied by a preset adjustment coefficient to construct a positive gain ratio, and the multiplicative gain formula is used. The initial weights are adjusted, which addresses the problem in traditional additive models where differences in basic weights are easily masked by a single constant gain term. This ensures that the adjustment process respects the physical topological connections between nodes while sensitively capturing dynamic shifts in business semantics. The beneficial effect of this computational mechanism is that it endows the static network topology with scene awareness and dynamic evolution capabilities, allowing the association weights between core collaborative nodes to undergo objective and controlled numerical stretching as task logic switches. This provides first-principles-compliant numerical support for subsequent quantification of collaboration priorities.

[0040] To further clarify the underlying operational logic of this dynamic weight correction mechanism, a deduction and verification are performed using the specific vector data obtained in steps S1 and S2. Assume the fused task feature vector... for Its vector magnitude is calculated to be approximately In the extracted local call subgraph, there are upstream sorting agent nodes with pre-defined collaborative relationships. With downstream visual inspection intelligent agent nodes Assuming the former's attribute feature vector for The latter's attribute feature vector for The system first calculates the sum vector of the two. ,Right now The magnitude of the sum vector is calculated to be approximately Subsequently, the system calculates the dot product of the task feature vector and the sum vector, i.e. The calculation result is Divide the dot product by the product of the magnitudes of the two ( The cosine alignment, which characterizes the semantic projection intensity, is approximately... Assume the initial association weight between these two nodes. for Preset adjustment coefficient for Substituting into the aforementioned multiplicative gain formula, the system calculates the positive gain ratio as follows: Corrected real-time correlation weights Then it is This data analysis objectively demonstrates that when the joint capabilities of two agents are highly aligned with task requirements, their association weights are amplified accordingly. This process resolves the computational bottleneck of the inability to quantify and assess resource dependency intensity in a multi-task parallel environment. Its beneficial effect is that it transforms abstract business relationships into real-time association matrices with clear gradients, thus providing downstream steps S4 and S5 with verifiable and sortable numerical data when allocating physical timestamps and underlying communication quotas.

[0041] like Figure 3 As shown, in S4, a collaborative invocation quantification evaluation mechanism and timestamp mapping logic based on deep semantic and physical dimensions are established. Critical paths in the local invocation subgraph are identified based on the real-time association matrix. To reduce the false positive rate of seemingly redundant but actually complementary calls in complex tasks, a three-dimensional redundancy removal mechanism based on semantic projection gradients is executed. This is done by determining any two agent nodes (set as nodes) in the target agent set. and nodes When considering whether functional redundancy exists, in addition to calculating the cosine similarity of the attribute feature vectors of the two entities, the projected gradient difference between them on the task feature vectors is also calculated. The formula for calculating the projected gradient difference is:

[0042] in: : Represents a node and nodes The gradient difference projected onto the task feature vector; : Represents the task feature vector; : Represents a node Attribute feature vectors; : Represents a node The attribute feature vector.

[0043] When the node and the node The cosine similarity between the two agents is greater than a preset redundancy threshold (the preset redundancy threshold is used to determine whether the attribute feature vectors of two agent nodes are highly similar in function, and is the first criterion for redundancy elimination; this threshold is generally set to a relatively high value, such as 0.9 to 0.98, to ensure that only nodes with extremely similar functions enter the next step of judgment; setting the threshold too high may lead to the omission of redundant nodes, while setting it too low may lead to the misjudgment of complementary nodes; in this embodiment, 0.95 is used, which is determined by cluster analysis based on the distribution characteristics of node attribute vectors in high-dimensional space and can reliably identify agents with overlapping functions), and the difference in the projection gradient of the two agents on the task feature vectors is also greater than the preset redundancy threshold (the preset redundancy threshold is used to determine whether the attribute feature vectors of two agent nodes are highly similar in function, and is the first criterion for redundancy elimination; this threshold is generally set to a relatively high value, such as 0.9 to 0.98, to ensure that only nodes with extremely similar functions enter the next step of judgment; setting the threshold too high may lead to the omission of redundant nodes, while setting it too low may lead to the misjudgment of complementary nodes; in this embodiment, 0.95 is used, which is based on the distribution characteristics of node attribute vectors in high-dimensional space and can reliably identify agents with overlapping functions), and the difference in the projection gradient of the two agents on the task feature vectors is also greater than the preset redundancy threshold (the preset redundancy threshold is used to determine whether the attribute feature vectors of two agent nodes are highly similar in function, and is the first criterion for redundancy elimination; this threshold is generally set to a relatively high value, such as 0.9 to 0.98, to ensure that only nodes with extremely similar functions enter the next step of judgment; setting the threshold too high may lead to the omission of redundant nodes, while setting it too low may lead to the misjudgment of complementary nodes; in this embodiment, 0.95 is used, If the difference is less than the preset gradient interference threshold (the preset gradient interference threshold is used to measure the difference between the projection scalars of two agent nodes on the current task feature vector. The smaller the difference, the closer their contributions to the current task are. Combined with cosine similarity, it can be determined whether it is a true redundancy. This threshold needs to be set according to the magnitude range of the task feature vector and the fluctuation range of the actual projection value. It is usually a small positive number, such as 0.01 to 0.1. In this example, it is 0.05. This value was determined after statistical analysis of the projection value distribution of multiple typical tasks. It can effectively distinguish between functional overlap and complementary situations and avoid false positives), then it is absolutely determined to be a functional redundancy. The system retains the agent nodes with higher correlation weights in the real-time correlation matrix and performs a redundancy removal operation.

[0044] Subsequently, based on the topological connection order between each agent node and the corrected real-time association weights, combined with the hardware load-aware dynamic response mechanism, the start timestamp and data interaction handle of each agent are determined. The reconstruction calculation formula for mapping the dimensionless real-time association weights to the start timestamps of specific physical time dimensions is as follows:

[0045] in: : Indicates a downstream node that is strictly dependent on the output of the upstream node. The startup timestamp of each agent node; : Indicates a local call to the upstream of the subgraph Release time of data interaction handles for each intelligent agent node; : Represents the preset reference system response constant; : represents the natural constant; : Indicates the preset hardware penalty coefficient; : indicates the first Real-time physical load rate of each intelligent agent node; : indicates the first The first intelligent agent node and the first The corrected real-time association weights between the intelligent agent nodes.

[0046] This calculation formula creates a hard constraint closed loop between semantic scheduling and the physical carrying capacity of the underlying hardware: even if the weights generated by the upper-layer semantic projection are extremely high, once the underlying hardware is nearing full load, the exponential term... This will drastically increase the additional latency. The system will arrange the agent nodes with start timestamps and data interaction handles according to their execution sequence, generating a multi-agent collaborative invocation sequence.

[0047] Furthermore, in this embodiment, it should be noted that after obtaining the real-time association matrix corrected in S3, step S4 establishes a three-dimensional redundancy removal mechanism based on semantic projection gradient. The system no longer relies solely on the cosine of the angle between nodes, but instead introduces the spatial projection gradient difference formula. The reason for adopting this dual-judgment calculation logic is that two nodes may exhibit high cosine similarity in their overall functional labels, but when facing a specific task, the actual projection strength of their core capabilities may be complementary. Only when two nodes in the target agent set... and The cosine similarity of the attribute features is greater than a preset redundancy threshold, and the difference between their projected gradients on the task feature vectors is greater than the redundancy threshold. Only when the gradient interference threshold is less than the preset threshold is it considered a true functional overlap. Based on specific data, this is extrapolated to assume the existence of two visual detection agent nodes. and The cosine similarity between the two is as high as It meets the requirement of being greater than the preset redundancy threshold. The initial screening conditions are then determined. The system further calculates the projection difference between the two onto the task feature vector. It is assumed that the system obtains the node... The projection scalar in this task direction is ,node The projection scalar is Then the difference in the projected gradients of the two for Because this value is less than the preset gradient interference threshold. The system determines that the two overlap in the current task's direction of effort, and therefore retains the node with the higher real-time correlation weight calculated in the previous steps. This process performs redundancy removal. It effectively mitigates the conflict between false positives and wasted computing power in complex tasks, achieving the beneficial effect of retaining nodes with potential complementary value while streamlining the call sequence, thus improving the effective utilization of system computing resources.

[0048] After simplifying the node sequence, and addressing the technical issue that traditional collaborative scheduling is often limited to software-level logical deduction and fails to establish a mapping relationship between semantic priorities and the underlying hardware physical state, leading to execution congestion under high load on underlying nodes, step S4 further combines hardware load status to generate a physical call sequence with timestamps. The system adopts a reconstructed startup timestamp calculation formula. Allocate data interaction handles to each node. This involves introducing an exponential hardware penalty. The reason for this computational logic is that abstract semantic association weights can only guide actual hardware execution when transformed into concrete physical latency, and this mapping process must be constrained by the real-time carrying capacity of the nodes. This is achieved by adjusting the real-time physical load factor. As a natural constant The exponential variable allows the system to increase additional latency through an exponential amplification effect when nodes are nearing full load, thus implementing underlying hardware protection. Based on the specific data derivation from S3 and the preceding steps, a baseline system response constant is set. milliseconds, hardware penalty coefficient Among them, the reference system response constant This represents the ideal response time of the system under no-load conditions and serves as a benchmark for mapping load penalty time. Its value depends on the underlying hardware processing capabilities and the basic latency of the communication link, and is typically obtained through no-load testing, such as measuring the average time from receiving an instruction to starting execution. In this embodiment, based on actual measurement data from the hardware platform, it is taken as... Milliseconds, this value accurately reflects the system's basic response capability, providing a reasonable benchmark for subsequent load penalty calculations. Meanwhile, the hardware penalty coefficient... Used to control real-time physical overall load rate The degree of amplification of delayed penalties, The larger the value, the more severe the impact of load on latency; this coefficient needs to be set according to the system's sensitivity and tolerance to load, and its value range is generally 1 to 5; too small a value will cause problems. Unable to effectively suppress high loads, excessive This could lead to a surge in latency, causing system jitter; in the example, [the following is taken] This value was determined through load impact experiments. It can apply a reasonable delay penalty to nodes that are close to full load while ensuring system stability, thereby achieving load balancing.

[0049] Furthermore, assume that the downstream visual detection agent nodes are retained at this point. Overall load rate Upstream sorting intelligent agent node The data interaction handle release time is The system calculates an exponential penalty term. The corrected real-time correlation weights calculated in S3 are introduced. Substituting into the time mapping formula, the downstream node The additional delay time is Milliseconds, its start timestamp is Milliseconds. This mechanism solves the problem of the disconnect between physical and logical allocation. Its beneficial effect is to establish a queuing sequence with flexible buffering capabilities, so that the compactness of upper-layer services can reasonably compromise with the physical carrying capacity of the lower layer, and alleviate system-level lag caused by hardware overload.

[0050] like Figure 4 As shown, in S5, dynamic conflict detection and underlying communication quota allocation are performed on the multi-agent collaborative call sequence based on real-time load status. The memory usage, computing unit load, and communication link bandwidth usage of each agent node during execution are acquired in real time. A resource usage state vector is constructed and compared with a preset resource threshold. A dynamic access quota model for the physical communication channel is established. When it is predicted that multiple agent nodes will compete for the same logical resource within the same predetermined time period, the system directly reshapes the quota replenishment rate of the underlying channel based on the real-time correlation matrix. Specifically, the logic is as follows: the corrected real-time correlation weights in the real-time correlation matrix are retrieved as heuristic factors to calculate the competition priority; the corrected real-time correlation weights are incremented by one and the natural logarithm is taken to obtain a logarithmic increment; the preset system basic access quota allocation rate is multiplied by one and the sum of the logarithmic increment to calculate the dynamic quota replenishment rate specific to each competing agent node. The system basic access quota allocation rate... This is the default bandwidth quota for the underlying physical channel in a contention-free state. Its value depends on the maximum bandwidth of the physical network and the system's preset resource allocation strategy; it can generally be set to a certain percentage of the maximum bandwidth (e.g., 10% to 30%) to leave a margin for handling sudden contention; in the embodiment, it is set as follows: Mbps. Assuming the maximum physical link bandwidth is 1Gbps, the base rate accounts for 10%. This value ensures both the needs of regular communication and leaves sufficient room for dynamic adjustment. The logarithmic gain model can provide multiple times the bandwidth increase for the core node, while still within the physical limit.

[0051] For the first The calculation formula for each intelligent agent node is:

[0052] in: : indicates that it is assigned to the first The dedicated dynamic quota replenishment rate for each intelligent agent node; : Indicates the preset system basic access quota allocation rate; : Represents the corrected real-time association weight in the real-time association matrix.

[0053] Based on the above formula, the corrected real-time correlation weight is directly converted into a mathematical factor for allocating bandwidth to the underlying hardware interface. The underlying layer opens the physical channel with the corresponding dedicated dynamic quota replenishment rate and allocates logical resource access permissions to the intelligent agent node according to the competition priority. While resolving the logical resource competition conflict, the data throughput of the core link is absolutely guaranteed.

[0054] When a system detects an agent node in the call chain whose response latency exceeds a preset latency threshold (the preset latency threshold is used to determine whether the agent node's response latency is too long due to a fault or overload, thereby triggering path redirection and collaborative adjustment; this threshold needs to be set according to the business's tolerance for response time, usually several times the normal response time; in this embodiment, based on the real-time requirements of the warehousing and logistics task, it is set to 200 milliseconds, that is, when the node latency exceeds 200 milliseconds, the system considers that the node can no longer meet the business requirements, and starts a backup node to take over the task to ensure the continuity of the overall process), the system searches for backup agent nodes with similar functional attribute labels based on the preset association edges in the multi-agent semantic knowledge graph, calculates the cosine similarity between the backup agent node and the current task feature vector as the semantic alignment, and generates a collaborative adjustment instruction containing path redirection information and context state transition data for execution.

[0055] Furthermore, in this embodiment, it should be noted that after generating the physical call sequence with timestamps, the system continuously acquires the memory usage, computing unit load, and communication link bandwidth usage of each agent node during execution, thereby constructing a multi-dimensional resource usage state vector and comparing it with a preset resource threshold to predict potential congestion. When it is predicted that multiple agent nodes will compete for the same logical resource (e.g., common path control) within the same predetermined time period, the system directly reshapes the quota replenishment rate of the underlying channel based on the real-time association weights corrected in step S3, and adopts a dedicated dynamic quota replenishment rate calculation formula. The reason for introducing the natural logarithmic function in this calculation is that the physical bandwidth or resource quota of the underlying hardware has a rigid upper limit. If a linear amplification model is used, high correlation weights can easily cause the allocation rate to overflow the physical limit. The logarithmic model, on the other hand, can provide reasonable gains to nodes with high weights while exhibiting a smooth control characteristic of diminishing marginal returns, keeping the increment within a safe physical boundary. Based on the data deduction from the previous steps, let the system's basic access quota allocation rate be... Mbps, the corrected real-time correlation weight of the upstream sorting agent node Substituting into the formula, its logarithmic increment is The system calculates the dedicated dynamic quota replenishment rate as follows: Mbps. This calculation mechanism improves the problem of blind allocation during logical resource contention. Its beneficial effect is that it transforms the dimensionless semantic weight into a bandwidth tilt factor of the physical communication interface, enabling the core collaborative path to obtain matching underlying channel support and maintaining the smooth flow of major business data.

[0056] In addition to implementing quota bias in the early stages of resource competition, step S5 also designs a path redirection and collaborative adjustment mechanism based on semantic alignment to address the technical problem of some nodes experiencing response delays due to sudden hardware failures in multi-agent collaborative architectures under dynamic concurrency scenarios, leading to call chain failures. During the execution of the aforementioned dynamic access quota model, the system continuously monitors the actual response performance of each node. Once an agent node in the call chain is detected with a response delay exceeding a preset latency threshold, the system considers that node to have encountered a physical execution bottleneck. At this time, the system immediately invokes the multi-agent semantic knowledge graph used in step S2, searching for candidate agent nodes with similar functional attribute labels based on preset association edges in the graph. To ensure that this last-minute replacement does not disrupt the original business intent of the call sequence, the system uses the cosine similarity between the attribute feature vector of the candidate node and the feature vector of the current task as the semantic alignment for verification. Suppose that the original visual detection agent experiences a response delay of 500 milliseconds due to memory overflow (exceeding the preset threshold of 200 milliseconds), the system graph retrieves a candidate detection node and calculates its semantic alignment as 0.91 (meeting the basic coordination requirements).

[0057] Subsequently, the system not only redirects the task to the alternative node but also generates a collaborative adjustment instruction containing context state transition data for execution. The reason for packaging and distributing the path redirection information along with the context transition data is that task relay in a distributed environment requires a slice of the preceding node's execution. If only address replacement is performed, the new node will be forced to recalculate due to the loss of historical operation states, causing secondary delays. This instruction solves the problems of node failures and logical breaks in dynamic execution environments. Its beneficial effect is that when facing unexpected situations at the underlying level, the system can quickly construct a new feasible path based on the guidance of the task feature vector, allowing the newly taking over agent to smoothly inherit the data interaction handles of the preceding node, ensuring the execution integrity of complex business chains and the robustness of the overall system architecture.

[0058] In summary, the multi-agent collaborative invocation method based on knowledge graph association firstly, at the task parsing end, by performing weighted fusion operations with clear mathematical boundaries on multi-dimensional semantic components, the ambiguity and fuzzy interference of unstructured natural language are effectively reduced, providing a clear and reproducible digital benchmark for subsequent accurate matching of execution entities. Secondly, the introduction of a multiplicative gain model based on semantic projection intensity endows the static knowledge graph with scene awareness, enabling the real-time association weights between core collaborative nodes to adaptively amplify as the task logic center shifts, ensuring the responsiveness of complex business paths. More importantly, the introduction of spatial projection gradient difference as a physical scale constructs a strict three-dimensional redundancy removal mechanism, accurately identifying and removing seemingly redundant but highly complementary key nodes in the current task's direction of effort, simplifying the invocation sequence while avoiding false positives. Furthermore, at the physical execution level, by constructing an exponential hardware penalty term that includes real-time physical comprehensive load rate, the dimensionless semantic association weights are mapped to specific physical start timestamps, achieving dynamic adaptation of upper-layer business logic to the underlying hardware capacity.

[0059] A multi-agent collaborative invocation system based on knowledge graph association is also provided, the system comprising: The task parsing module is used to obtain the original requirement text of the task to be executed, and to perform semantic parsing on the original requirement text to extract the task feature vector; The graph management module is used to retrieve a preset multi-agent semantic knowledge graph, which includes multiple agent nodes, functional attribute labels corresponding to each agent node, and preset association edges between nodes. The association quantization module is used to match the task feature vector with each agent node in the multi-agent semantic knowledge graph, identify the target agent set, construct a local call subgraph based on the target agent set, dynamically correct the initial association weight based on the task feature vector, and obtain the real-time association matrix. The sequence generation module is used to quantify the collaboration priority and dependency strength between each agent node according to the real-time association matrix, perform redundancy removal based on semantic projection gradient, and generate a multi-agent collaborative calling sequence by combining the underlying physical load state and the dependency strength. The conflict scheduling module is used to obtain the real-time physical load status and logical resource occupancy of each agent node when executing the multi-agent collaborative call sequence, perform conflict scheduling based on the dynamic access quota model, and issue collaborative adjustment instructions according to the detection results.

[0060] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principle of this invention is further explained below in conjunction with a specific industrial automation warehousing and logistics scheduling application scenario: S1: The task parsing module obtains the original requirement text "Execute the emergency outbound and refined quality inspection task for high-value, fragile electronic components". Semantic parsing extracts the corresponding verb component vector, object component vector, and constraint component vector. Preset verb weight coefficients are set. Object weight coefficient Constraint weight coefficients According to the aforementioned fusion formula Calculate the task feature vector that is specifically mapped in the semantic space.

[0061] S2: The graph management module retrieves a multi-agent semantic knowledge graph containing sorting agent nodes, handling agent nodes, and visual inspection agent nodes. Calculations show that the cosine similarity of the visual inspection agent node reaches 0.92, which is greater than the preset matching threshold of 0.85. Therefore, it is filtered out by the system and included in the target agent set.

[0062] S3: The correlation quantization module constructs a local call subgraph. Assume the upstream sorting agent node in the local call subgraph (i.e., the...) The first intelligent agent node) and the downstream visual detection intelligent agent node (i.e., the first intelligent agent node) Initial association weights among intelligent agent nodes The system calculates that the semantic projection strength of the task feature vector between the two connected agent nodes is 0.80. An adjustment coefficient is set. That is, the positive gain ratio is 0.4; specifically, the adjustment coefficient γ is used to control the correction magnitude of the semantic projection intensity on the initial association weight, and its value determines the sensitivity of the dynamic adjustment of the weight. It can be set according to the system's response requirements to changes in task semantics, and is generally set to 0.1 to 1.0; If the weight is too small, the weight adjustment will be insufficient and it will be difficult to reflect the characteristics of the task; An excessively large value may cause drastic fluctuations in weight, affecting the stability of the call; in the example, it is taken as... This ensures a moderate positive gain ratio. The corrected real-time association weights fully reflect the semantic projection strength of the task while maintaining a reasonable correlation with the initial weights. Simulation tests show that this value can effectively improve the response timeliness of the critical path. Substituting the corrected real-time association weights into the calculation logic yields the corrected real-time association weights. .

[0063] S4: The sequence generation module performs 3D redundancy removal based on semantic projection gradients. Assume there are two visual detection agent nodes (nodes...). and nodes The cosine similarity between the two is as high as 0.96 (greater than the preset redundancy threshold of 0.95). The system calculates the difference in the projected gradients of the two onto the task feature vectors. Because this value is less than the preset gradient interference threshold of 0.05, the system absolutely determines that the two overlap in the current task's force direction and retains the node with the higher weight. .

[0064] When determining the startup timestamp, the reference system response constant is set. milliseconds, hardware penalty coefficient Assume that the downstream visual detection agent node is currently... Real-time physical overall load rate (i.e., 80% full load rate), upstream sorting intelligent agent nodes The data interaction handle release time is Exponential penalty items Substituting the time mapping reconstruction formula, the downstream visual detection agent nodes... The startup timestamp is: millisecond.

[0065] Through this mechanism, even if the semantic association weight is extremely high, the system will implement reasonable hardware latency protection based on the underlying physical load.

[0066] S5: During execution, if it is detected that the sorting agent node and the ordinary handling agent node are competing for the logical resource of common path control, the conflict scheduling module immediately establishes a dynamic access quota model. Assume a preset system basic access quota allocation rate. Mbps, real-time association weight of sorting agent nodes Substitute the formula for the dedicated dynamic quota replenishment rate: Mbps.

[0067] The system forces the opening of the underlying physical channel for the sorting agent nodes at an excess quota replenishment rate of 161Mbps and allocates logical resource access permissions, thus perfectly eliminating resource bottlenecks at the physical level.

[0068] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0069] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0070] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A multi-agent collaborative invocation method based on knowledge graph association, characterized in that, The methods include: Obtain the original requirement text of the task to be executed, and perform semantic parsing on the original requirement text to extract the task feature vector; Retrieve a preset multi-agent semantic knowledge graph, match the task feature vector with each agent node in the multi-agent semantic knowledge graph, and identify the target agent set; A local call subgraph is constructed based on the preset association edges between each agent node in the target agent set, and the initial association weights in the local call subgraph are dynamically corrected based on the task feature vector to obtain a real-time association matrix; The collaboration priority and dependency strength between each agent node are quantified based on the real-time association matrix, redundancy elimination is performed based on semantic projection gradient, and a multi-agent collaborative calling sequence is generated by combining the underlying physical load state and the dependency strength. The system acquires the real-time physical load status and logical resource occupancy of each agent node when executing the multi-agent collaborative invocation sequence, performs conflict scheduling based on the dynamic access quota model, and issues collaborative adjustment instructions.

2. The multi-agent collaborative invocation method based on knowledge graph association according to claim 1, characterized in that, The step of obtaining the original requirement text of the task to be executed and performing semantic parsing on the original requirement text to extract the task feature vector includes: Identify the task verbs, operation objects, and constraints in the original requirement text, and map them to a high-dimensional semantic space to obtain the corresponding verb component vectors, object component vectors, and constraint component vectors. Obtain verb weight coefficients, object weight coefficients, and constraint weight coefficients that are preset based on semantic importance and sum to one; Using the verb weight coefficient, the object weight coefficient, and the constraint weight coefficient, a weighted fusion operation is performed on the verb component vector, the object component vector, and the constraint component vector to obtain the task feature vector that represents the global semantic features of the task to be executed.

3. The multi-agent collaborative invocation method based on knowledge graph association according to claim 1, characterized in that, The step of dynamically correcting the initial association weights in the local call subgraph based on the task feature vector to obtain the real-time association matrix includes: Calculate the cosine alignment of the task feature vector with the sum of the attribute feature vectors of two connected agent nodes in the local call subgraph, and use it as the semantic projection strength of the task feature vector between the two connected agent nodes. A positive gain ratio is constructed by multiplying the semantic projection intensity by a preset adjustment coefficient; The initial association weight between the two agent nodes is multiplied by one and the sum of the positive gain ratio to obtain the corrected real-time association weight. All the corrected real-time association weights are then arranged according to the node index to construct the real-time association matrix.

4. The multi-agent collaborative invocation method based on knowledge graph association according to claim 1, characterized in that, The redundancy removal based on semantic projection gradient includes: Calculate the cosine similarity of the attribute feature vectors of any two agent nodes in the target agent set; Calculate the absolute value of the difference between the projection scalars of the attribute feature vectors of the two agent nodes onto the task feature vector, and use it as the projection gradient difference; When the cosine similarity is greater than a preset redundancy threshold and the projected gradient difference is less than a preset gradient interference threshold, it is determined that the two agent nodes have functional redundancy, and the agent node with the higher corrected real-time association weight in the real-time association matrix is ​​retained, and a redundancy removal operation is performed.

5. The multi-agent collaborative invocation method based on knowledge graph association according to claim 1, characterized in that, The process of generating a multi-agent collaborative invocation sequence by combining the underlying physical load state with the dependency strength includes: Obtain the release time of the data interaction handle of the upstream intelligent agent node, and the real-time physical comprehensive load rate of the downstream intelligent agent node; The load penalty response time is obtained by multiplying the preset baseline system response constant by an exponential term that has a base of natural constant, an exponential term that is the product of the preset hardware penalty coefficient and the real-time physical comprehensive load rate. The additional delay time is obtained by dividing the load penalty response time by the corrected real-time association weight between the upstream agent node and the downstream agent node. The data interaction handle release time plus the additional delay time is used to determine the start timestamp of the downstream agent node, and the multi-agent collaborative call sequence is generated according to the time sequence of the start timestamps.

6. The multi-agent collaborative invocation method based on knowledge graph association according to claim 1, characterized in that, The execution of conflict scheduling based on the dynamic access quota model and the issuance of collaborative adjustment instructions include: When it is predicted that multiple intelligent agent nodes will compete for the same logical resource within the same predetermined time period, the corresponding corrected real-time association weights in the real-time association matrix are retrieved as heuristic factors to calculate the competition priority. Add one to the corrected real-time correlation weight and take the natural logarithm to obtain the logarithmic increment; The preset system basic access quota distribution rate is multiplied by one and the sum of the logarithmic increments to calculate the dynamic quota replenishment rate specific to each competing intelligent agent node. The underlying physical communication channel is opened at the corresponding dedicated dynamic quota replenishment rate, and logical resource access permissions are allocated to the intelligent agent node according to the competition priority.

7. A multi-agent collaborative invocation system based on knowledge graph association, characterized in that, The system includes: The task parsing module is used to obtain the original requirement text of the task to be executed, and to perform semantic parsing on the original requirement text to extract the task feature vector; The graph management module is used to retrieve a preset multi-agent semantic knowledge graph, which includes multiple agent nodes, functional attribute labels corresponding to each agent node, and preset association edges between nodes. The association quantization module is used to match the task feature vector with each agent node in the multi-agent semantic knowledge graph, identify the target agent set, construct a local call subgraph based on the target agent set, dynamically correct the initial association weight based on the task feature vector, and obtain the real-time association matrix. The sequence generation module is used to quantify the collaboration priority and dependency strength between each agent node according to the real-time association matrix, perform redundancy removal based on semantic projection gradient, and generate a multi-agent collaborative calling sequence by combining the underlying physical load state and the dependency strength. The conflict scheduling module is used to obtain the real-time physical load status and logical resource occupancy of each agent node when executing the multi-agent collaborative call sequence, perform conflict scheduling based on the dynamic access quota model, and issue collaborative adjustment instructions.

8. The multi-agent collaborative invocation system based on knowledge graph association according to claim 7, characterized in that, The task parsing module is also used for: Identify the task verbs, operation objects, and constraints in the original requirement text, and map them to a high-dimensional semantic space to obtain the corresponding verb component vectors, object component vectors, and constraint component vectors. Obtain verb weight coefficients, object weight coefficients, and constraint weight coefficients that are preset based on semantic importance and sum to one; Using the verb weight coefficient, the object weight coefficient, and the constraint weight coefficient, a weighted fusion operation is performed on the verb component vector, the object component vector, and the constraint component vector to obtain the task feature vector that represents the global semantic features of the task to be executed.

9. The multi-agent collaborative invocation system based on knowledge graph association according to claim 7, characterized in that, The correlation quantization module is also used for: Calculate the cosine alignment of the task feature vector with the sum of the attribute feature vectors of two connected agent nodes in the local call subgraph, and use it as the semantic projection strength of the task feature vector between the two connected agent nodes. A positive gain ratio is constructed by multiplying the semantic projection intensity by a preset adjustment coefficient; The initial association weight between the two agent nodes is multiplied by one and the sum of the positive gain ratio to obtain the corrected real-time association weight. All the corrected real-time association weights are then arranged according to the node index to construct the real-time association matrix.

10. The multi-agent collaborative invocation system based on knowledge graph association according to claim 7, characterized in that, The sequence generation module is also used for: Calculate the cosine similarity of the attribute feature vectors of any two agent nodes in the target agent set; Calculate the absolute value of the difference between the projection scalars of the attribute feature vectors of the two agent nodes onto the task feature vector, and use it as the projection gradient difference; When the cosine similarity is greater than a preset redundancy threshold and the projected gradient difference is less than a preset gradient interference threshold, it is determined that the two agent nodes have functional redundancy, and the agent node with the higher corrected real-time association weight in the real-time association matrix is ​​retained, and a redundancy removal operation is performed.

Images