A unified scheduling method and system for enterprise-level multi-agent platforms

Abstract

This invention discloses a unified scheduling method and system for an enterprise-level multi-agent platform, comprising: constructing a unified scheduling center including all heterogeneous agents and their corresponding capability description data; semantically normalizing the capability description data of each heterogeneous agent to construct a capability knowledge graph; receiving task requests; decomposing the task requests into atomic task sequences based on the capability knowledge graph, wherein each atomic task sequence contains capability constraint tags; adaptively matching subtasks based on the collected real-time running status and capability constraint tags to obtain a global task allocation scheme; and driving heterogeneous agents to collaboratively execute the task requests based on the global task allocation scheme. Compared with existing technologies, this invention can significantly improve the semantic normalization level of heterogeneous resources, the logical accuracy of task decomposition, and the adaptive matching efficiency of global scheduling, overcoming the technical shortcomings of incompatible interface standards, rigid logical decomposition, and disconnected scheduling decisions.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence computing technology, and more specifically, to a unified scheduling method and system for an enterprise-level multi-agent platform. Background Technology

[0002] With the deep integration of enterprise digitalization and artificial intelligence technologies, multi-agent systems, as the core architecture for reconstructing complex business processes, have become key indicators of an enterprise's intelligence level, particularly in terms of resource integration capabilities and task orchestration efficiency. Modern enterprise-level application scenarios require scheduling platforms to possess semantic understanding across heterogeneous systems, automated decomposition of abstract business chains, and precise scheduling capabilities under dynamic load environments. Traditional hard-coded scheduling based on rule engines or independent operation of individual agents, limited by fragmented interface standards, lack of contextual logic, and static allocation mechanisms, can no longer meet the efficient collaboration needs of large-scale agent clusters. Unified scheduling platforms, with their global perspective on resource management, are gradually becoming the central carrier for heterogeneous agent collaboration. Their standardization compatibility with heterogeneous resources, the depth of planning for complex tasks, and the flexibility of scheduling strategies directly determine the continuity of business execution and the efficiency of computing resource utilization.

[0003] While some existing technologies attempt to achieve agent collaboration through API gateway aggregation or fixed workflow orchestration, they generally face technical challenges such as incompatibility of heterogeneous interface data standards, inability to automatically decompose abstract business logic, and severe disconnect between scheduling decisions and real-time node load. Furthermore, they frequently result in cross-system collaboration failures and imbalances in computing resource allocation.

[0004] Therefore, how to provide a unified scheduling method and system for enterprise-level multi-agent platforms that can significantly improve the semantic normalization level of heterogeneous resources, the logical accuracy of task decomposition, and the adaptive matching efficiency of global scheduling, and overcome the technical shortcomings of traditional solutions such as incompatible interface standards, rigid logical decomposition, and disconnected scheduling decisions, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a unified scheduling method for an enterprise-level multi-agent platform, which can significantly improve the semantic normalization level of heterogeneous resources, the logical accuracy of task decomposition, and the adaptive matching efficiency of global scheduling, overcoming the technical shortcomings of incompatible interface standards, rigid logical decomposition, and disconnected scheduling decisions.

[0006] The first technical solution provided by this invention is as follows: This invention provides a unified scheduling method for an enterprise-level multi-agent platform, comprising the following steps: S1 Constructing a unified scheduling center, which communicatively connects to all heterogeneous agents and receives capability description data corresponding to each heterogeneous agent; S2 Semantically normalizing the capability description data of each heterogeneous agent and constructing a capability knowledge graph; S3 Decomposing the received task request according to the capability knowledge graph based on the received task request to obtain an atomic task sequence, wherein the atomic task sequence contains capability constraint tags for each sub-task; S4 Collecting the real-time running status of each heterogeneous agent and adaptively matching the sub-tasks according to the real-time running status and the capability constraint tags to obtain a global task allocation scheme; S5 Driving the heterogeneous agents to collaboratively execute the task request based on the global task allocation scheme.

[0007] Furthermore, in a preferred embodiment of the present invention, the step of constructing a capability knowledge graph includes: Receive the capability description data of each heterogeneous intelligent agent, perform semantic normalization on the capability description data of each heterogeneous intelligent agent, and generate a standard capability representation; Generate capability entity nodes based on the aforementioned standard capability representation; Extract the logical collaboration relationships between the heterogeneous intelligent agents and map these relationships as topological connection edges connecting the capability entity nodes to construct a capability knowledge graph.

[0008] Furthermore, in a preferred embodiment of the present invention, the step of semantically normalizing the capability description data of each heterogeneous intelligent agent includes: The adapter interface according to the communication protocol receives the interface definition file and quality of service metrics for each heterogeneous intelligent agent. A pre-defined capability ontology is constructed based on ontology modeling technology. The pre-defined capability ontology includes standard concept classes and attribute constraints. The interface definition file is parsed, the heterogeneous input and output parameters are extracted, and the heterogeneous input and output parameters are defined as instance data, mapped and populated into the corresponding standard concept class in the preset capability ontology; The service quality indicators are quantified based on the attribute constraints to generate standard capability representations.

[0009] Furthermore, in a preferred embodiment of the present invention, the step of decomposing the task request based on the capability knowledge graph includes: Construct a prompting engineering input context, which includes a natural language description of the task request and the standard concept class associated with the task request; The prompting process is input into the context and passed into the large language model, which then drives the large language model to generate logical analysis text. Parse the logical analysis text to obtain the action derivation matching the standard concept class and the execution path determination based on action dependency; The atomic task sequence is generated based on the action derivation and the execution path determination, and the sub-tasks in the sequence are labeled with capability constraint tags containing execution priority and deadline, thus completing the task request decomposition.

[0010] Furthermore, in a preferred embodiment of the present invention, step S4 includes: Real-time collection of current load rate, health status, and network latency data for each of the heterogeneous intelligent agents; A multi-objective optimization model is constructed based on deep learning technology, and the semantic matching degree between the capability constraint label of the sub-task and the heterogeneous agent is calculated based on the multi-objective optimization model. Based on the semantic matching degree and the current load rate and execution cost factors, the multi-objective optimization model is solved globally through iteration to obtain a global task allocation scheme.

[0011] Further, in a preferred embodiment of the present invention, the step of calculating the semantic matching degree between the capability constraint label of the sub-task and the heterogeneous agent according to the multi-objective optimization model includes: The multi-objective optimization model includes an input representation layer, a vector embedding layer, and a semantic interaction layer; The input representation layer receives the capability constraint labels of the subtasks and the standard capability representations of the heterogeneous agents, respectively. Based on the vector embedding layer, the descriptive text in the capability constraint label and the standard capability representation are mapped to task feature vectors and capability feature vectors, respectively. Based on the semantic interaction layer, the feature cross-result between the task feature vector and the capability feature vector is calculated and obtained, and the semantic matching degree is output based on the feature cross-result.

[0012] Furthermore, in a preferred embodiment of the present invention, the method further includes performing data compatibility verification on the atomic task sequence, specifically: Traverse the adjacent upstream and downstream subtasks in the atomic task sequence that have data flow relationships; Based on the capability knowledge graph, the expected output data pattern corresponding to the upstream sub-task and the required input data pattern corresponding to the downstream sub-task are extracted respectively. Based on field name similarity, data type compatibility, and data structure hierarchy differences, the semantic structure distance between the expected output data pattern and the required input data pattern is quantitatively calculated. When the semantic structure distance exceeds a preset threshold, dynamic data adaptation processing is performed on the upstream subtask and the downstream subtask.

[0013] Furthermore, in a preferred embodiment of the present invention, the step of performing dynamic data adaptation processing on the upstream subtask and the downstream subtask includes: Call the preset code generation model, input the difference features between the expected output data pattern and the required input data pattern into the code generation model, and obtain the data conversion code; Configure the data conversion runtime environment, load the data conversion code based on the data conversion runtime environment, and encapsulate the data conversion code as an intermediate processing node; The intermediate processing node is inserted serially between the upstream subtask and the downstream subtask to complete the dynamic data adaptation process.

[0014] Furthermore, in a preferred embodiment of the present invention, the method further includes optimizing the execution process of the task request and the multi-objective optimization model, specifically as follows: A task instruction message queue is established based on the unified scheduling center, and scheduling instructions are distributed to the heterogeneous intelligent agents according to the capability knowledge graph and the task instruction message queue. When multiple heterogeneous intelligent agents are detected to be contending for the same resource, the conflict is resolved according to the execution priority in the capability constraint label. When an abnormality is detected in the execution of a certain heterogeneous intelligent agent, a candidate heterogeneous intelligent agent with the capability matching the subtask is retrieved based on the capability constraint tag, and the execution instructions of the subtask are rescheduled to the candidate heterogeneous intelligent agent. The unified scheduling center records the entire process execution log and uses the entire process execution log and the result of the task request as feedback data to update and iterate the multi-objective optimization model.

[0015] The present invention provides a second technical solution as follows: This invention also provides a unified scheduling system for an enterprise-level multi-agent platform, comprising: The central construction module builds a unified scheduling center. The unified scheduling center's communication connections include all heterogeneous intelligent agents and the capability description data corresponding to each heterogeneous intelligent agent. The graph construction module performs semantic normalization on the capability description data of each heterogeneous intelligent agent and constructs a capability knowledge graph. The task decomposition module, based on the received task request, decomposes the task request according to the capability knowledge graph to obtain an atomic task sequence, wherein the atomic task sequence contains capability constraint tags for each subtask. The task allocation module collects the real-time running status of each heterogeneous intelligent agent, and adaptively matches the sub-tasks based on the real-time running status and the capability constraint labels to obtain a global task allocation scheme. The collaborative execution module, based on the global task allocation scheme, drives the heterogeneous intelligent agents to collaboratively execute the task request.

[0016] This invention provides a unified scheduling method for an enterprise-level multi-agent platform, which significantly improves the semantic normalization level of heterogeneous resources, the logical accuracy of task decomposition, and the adaptive matching efficiency of global scheduling, overcoming the technical shortcomings of incompatible interface standards, rigid logical decomposition, and disconnected scheduling decisions. The unified scheduling method for the enterprise-level multi-agent platform includes: S1 constructing a unified scheduling center, which communicatively connects to all heterogeneous agents and receives capability description data corresponding to each heterogeneous agent; S2 performing semantic normalization on the capability description data of each heterogeneous agent and constructing a capability knowledge graph; S3 decomposing the received task request according to the capability knowledge graph to obtain an atomic task sequence, the atomic task sequence containing capability constraint tags for each sub-task; S4 collecting the real-time running status of each heterogeneous agent and adaptively matching the sub-tasks based on the real-time running status and the capability constraint tags to obtain a global task allocation scheme; S5 driving the heterogeneous agents to collaboratively execute the task request based on the global task allocation scheme. Specifically, by constructing a unified scheduling center to aggregate heterogeneous intelligent agent resources, and using semantic normalization technology to standardize and map multi-source heterogeneous capability description data and construct a capability knowledge graph, a unified interaction standard is established at the semantic level to shield the data differences of underlying heterogeneous interfaces, thereby solving the technical problem of incompatible heterogeneous interface data standards. On this basis, the system uses this capability knowledge graph as the logical deduction base to deeply parse and decompose the received abstract task requests into atomic task sequences containing precise capability constraint labels. By transforming abstract business logic into machine-recognizable structured instructions with capability definitions, the system overcomes the defect that complex business logic cannot be automatically decomposed. Furthermore, the system collects the running status of each intelligent agent in real time and performs joint analysis and adaptive matching of this dynamic load data with the capability constraint labels of the aforementioned sub-tasks to generate a global task allocation scheme. Through this strategy of deeply integrating static capability requirements with dynamic running status, the system ensures that scheduling decisions respond to node load changes in real time, effectively solving the problem of cross-system collaboration failure and resource allocation imbalance caused by the disconnect between scheduling decisions and real-time load in existing technologies. Compared with existing technologies, this invention can significantly improve the semantic normalization level of heterogeneous resources, the logical accuracy of task decomposition, and the adaptive matching efficiency of global scheduling, overcoming the technical shortcomings of incompatible interface standards, rigid logical decomposition, and disconnected scheduling decisions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the steps of a unified scheduling method for an enterprise-level multi-agent platform provided in an embodiment of the present invention; Figure 2 A flowchart illustrating the steps involved in constructing a capability knowledge graph, as provided in this embodiment of the invention. Figure 3 A logical framework diagram of the data compatibility verification steps provided in this embodiment of the invention; Figure 4 A schematic diagram of the modules of the unified scheduling system of the enterprise-level multi-agent platform provided in the embodiments of the present invention; Figure 5 This is a statistical chart comparing the performance of the map construction of the embodiments of the present invention with that of the prior art. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0020] It should be noted that when a component is referred to as being "fixed to" or "set on" another component, it can be directly on or indirectly set on the other component; when a component is referred to as being "connected to" another component, it can be directly connected to or indirectly connected to the other component.

[0021] It should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "first", "second", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" or "several" means two or more, unless otherwise explicitly specified.

[0023] It should be noted that the structures, proportions, sizes, etc., shown in the accompanying drawings of this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0024] like Figures 1 to 5 As shown in the embodiments of the present invention, a unified scheduling method for an enterprise-level multi-agent platform is provided, which can significantly improve the semantic normalization level of heterogeneous resources, the logical accuracy of task decomposition, and the adaptive matching efficiency of global scheduling, overcoming the technical shortcomings of incompatible interface standards, rigid logical decomposition, and disconnected scheduling decisions.

[0025] A unified scheduling method for an enterprise-level multi-agent platform includes the following steps: S1 Constructing a unified scheduling center, which communicates with all heterogeneous agents and receives capability description data corresponding to each heterogeneous agent; S2 Semantically normalizing the capability description data of each heterogeneous agent and constructing a capability knowledge graph; S3 Decomposing the received task request according to the capability knowledge graph based on the received task request to obtain an atomic task sequence, which contains capability constraint labels for each sub-task; S4 Collecting the real-time running status of each heterogeneous agent and adaptively matching the sub-tasks based on the real-time running status and capability constraint labels to obtain a global task allocation scheme; S5 Driving the heterogeneous agents to collaboratively execute the task request based on the global task allocation scheme. Specifically, by constructing a unified scheduling center to aggregate heterogeneous intelligent agent resources, and using semantic normalization technology to standardize and map multi-source heterogeneous capability description data and construct a capability knowledge graph, a unified interaction standard is established at the semantic level to shield the data differences of underlying heterogeneous interfaces, thereby solving the technical problem of incompatible heterogeneous interface data standards. On this basis, the system uses this capability knowledge graph as the logical deduction base to deeply parse and decompose the received abstract task requests into atomic task sequences containing precise capability constraint labels. By transforming abstract business logic into machine-recognizable structured instructions with capability definitions, the system overcomes the defect that complex business logic cannot be automatically decomposed. Furthermore, the system collects the running status of each intelligent agent in real time and performs joint analysis and adaptive matching of this dynamic load data with the capability constraint labels of the aforementioned sub-tasks to generate a global task allocation scheme. Through this strategy of deeply integrating static capability requirements with dynamic running status, the system ensures that scheduling decisions respond to node load changes in real time, effectively solving the problem of cross-system collaboration failure and resource allocation imbalance caused by the disconnect between scheduling decisions and real-time load in existing technologies. Compared with existing technologies, this invention can significantly improve the semantic normalization level of heterogeneous resources, the logical accuracy of task decomposition, and the adaptive matching efficiency of global scheduling, overcoming the technical shortcomings of incompatible interface standards, rigid logical decomposition, and disconnected scheduling decisions.

[0026] The following detailed explanation of the steps and flow of the unified scheduling method for an enterprise-level multi-agent platform, using specific embodiments, is provided.

[0027] Specifically, such as Figure 2 As shown, in a specific embodiment of the present invention, the step of constructing a capability knowledge graph includes: S21 receiving capability description data of each heterogeneous intelligent agent, performing semantic normalization on the capability description data of each heterogeneous intelligent agent, and generating a standard capability representation; S22 generating capability entity nodes based on the standard capability representation; S23 extracting the logical cooperation relationship between heterogeneous intelligent agents, mapping the logical cooperation relationship to the topological connection edge connecting the capability entity nodes, and constructing a capability knowledge graph.

[0028] In a specific embodiment of this invention, the first step is to establish an access channel for heterogeneous intelligent agent data. This involves receiving capability description data from heterogeneous intelligent agents of different vendors and architectures based on HTTP or MQTT communication protocols. This data is typically in unstructured text or semi-structured JSON format, containing function definitions, interface parameters, and service quality indicators. To eliminate semantic ambiguity from multiple data sources and achieve a unified, computer-readable expression, a pre-trained language model based on the Transformer architecture is used as a semantic normalization encoder, such as BERT-base-uncased. This model consists of 1... The model consists of two stacked Transformer encoders, each containing 12 self-attention heads and a 768-dimensional hidden layer. GELU activation is used, and the maximum input sequence length is set to 512 tokens. Domain-adaptive fine-tuning is performed using a pre-collected corpus of industrial IoT terminology. This fine-tuning process employs unsupervised training using a masked language model task, with AdamW as the optimizer, a learning rate of 2e-5, and a weight decay coefficient of 0.01. Based on the fine-tuned encoder, each received capability description data is vectorized, and the output vector corresponding to the [CLS] marker is extracted. As a standard capability representation, this representation maps discrete text descriptions to coordinate points in a high-dimensional continuous semantic feature space. Subsequently, based on the generated standard capability representation, corresponding capability entity nodes are created in the graph database, and the high-dimensional semantic vectors are persistently stored as node attributes. When extracting logical collaboration relationships, the feature vectors of any two capability entity nodes are concatenated, and a multilayer perceptron is used to calculate their dependency probability in the business flow. The threshold for determining this dependency probability is not a fixed value, but is dynamically optimized by plotting a precision-recall curve (PR curve) on an independent validation set, specifically with a threshold of 0.01. The probability interval from 0 to 1 is traversed by step size. The F1 score at each cutoff point is calculated. The probability value corresponding to the peak of the F1 score (preferably 0.85 in this embodiment) is selected as the preset threshold. This maximizes the preservation of potential collaborative links while effectively suppressing noise interference from false connections. When the output dependency probability value exceeds the preset threshold, it is determined that there is a logical collaborative relationship between the two agents. This relationship is instantiated as a directed topological connection edge connecting the two capability entity nodes. The edge attributes include the collaboration type and confidence weight, thereby completing the construction of the topological structure of the capability knowledge graph, which can realize semantic insight and association mining of heterogeneous resources.

[0029] To verify the effectiveness and robustness of the map construction method in this embodiment, such as Figure 5As shown, a publicly available large-scale supply chain logistics multi-agent dataset was selected for experimental verification. This dataset contains 5000 heterogeneous agent nodes (including AGVs, robotic arms, sorting machines, etc.) and 20000 historical interaction records. In the data preprocessing stage, the original descriptive text was denoised, stop word filtered, and segmented. The dataset was then randomly divided into training, validation, and test sets in an 8:1:1 ratio. The experimental environment was deployed on an Ubuntu 20.04 server equipped with two NVIDIA A100 GPUs, and the deep learning framework used was PyTorch version 1.13. In the model training stage, the triplet loss function was used. By narrowing the distance between positive sample pairs with similar functions and widening the distance between negative sample pairs with large functional differences, the semantic distribution of the vector space was optimized. The training epochs were set to 100, the batch size to 64, and the Early... The stopping mechanism prevents overfitting. Experimental results show that the capability knowledge graph constructed based on this scheme achieves an entity alignment accuracy of 94.5% and a relation extraction F1 score of 91.2%. Compared with traditional graph construction methods based on keyword matching or rule templates, the entity recognition error rate is reduced by 18%, the graph construction speed is increased by 40%, and it exhibits good generalization ability when faced with novel intelligent agent descriptions that have never been seen before. This demonstrates that this scheme has extremely high semantic understanding accuracy and topology construction efficiency when processing large-scale heterogeneous intelligent agent data, and can provide a solid data foundation for subsequent task decomposition and scheduling.

[0030] Specifically, in a specific embodiment of the present invention, the step of semantically normalizing the capability description data of each heterogeneous intelligent agent includes: receiving the interface definition file and service quality indicators of each heterogeneous intelligent agent according to the adapter interface of the communication protocol; constructing a pre-set capability ontology based on ontology modeling technology, the pre-set capability ontology including standard concept classes and attribute constraints; parsing the interface definition file, extracting the heterogeneous input and output parameters therein, defining the heterogeneous input and output parameters as instance data, mapping and filling them into the corresponding standard concept classes in the pre-set capability ontology; and quantifying the service quality indicators based on the attribute constraints to generate standard capability representations.

[0031] In this embodiment of the invention, a data receiving channel is first established by configuring a multi-protocol communication adapter. This adapter integrates HTTP, MQTT, and gRPC communication protocol stacks and can automatically identify and parse OpenAPI specifications from different vendors' intelligent agents, such as Swagger JSON interface definition files and XML-formatted service quality indicator documents. A pre-built capability ontology is constructed based on the OWL Web ontology language. Specifically, this pre-built capability ontology is constructed using a top-down three-layer architecture: the first layer is the basic capability domain, the second layer is the specific functional subclasses, and the third layer is the attribute constraint layer. Precise data attribute constraints are defined for each specific functional subclass; for example, the maximum load attribute of the transport class is constrained to be numerical and the unit is uniformly kilograms. During the parsing of the interface definition file and parameter mapping, to solve the problem of inconsistent naming styles for heterogeneous parameters, such as load_capacity and max_weight, a Siamese-based ontology is constructed. The BERT dual-tower architecture semantic alignment neural network model consists of two BERT-base encoders sharing weights. Each encoder contains 12 Transformer coding layers, 12 self-attention heads, and a 768-dimensional hidden layer. GELU activation is used, and the input layer is truncated to 128 tokens. Interface parameter names and standard attribute names from a pre-defined ontology are input to the two encoders respectively, generating instantaneous CLS bit vectors as semantic representations. The cosine similarity between the two vectors is calculated to measure the degree of semantic matching. The similarity threshold for determining a successful match is not fixed but determined through hyperparameter search on a labeled industrial semantic alignment validation set, specifically iterating through the interval from 0.5 to 0.99 with a step size of 0.01. The F1 score is calculated at different thresholds. The similarity value corresponding to the peak F1 score of 0.82 in this embodiment is selected as the final preset threshold. When the calculated similarity exceeds the preset threshold, the heterogeneous parameter is defined as instance data and automatically filled into the corresponding standard concept class. For the quantitative processing of service quality indicators, the attribute constraints in the preset ontology, such as the response time needing to be less than 100ms, are read. The Min-Max normalization method is used to map the original QoS data to a dimensionless interval of 0 to 1. Combined with the preset Sigmoid utility function, the physical value is converted into a standard capability score, thereby generating a standard capability representation containing structured semantics and quantitative performance data. This can effectively eliminate the semantic gap between heterogeneous devices and achieve accurate alignment and unification of multi-source data.

[0032] To verify the effectiveness of the aforementioned semantic normalization and ontology mapping scheme, a simulation dataset containing 2000 heterogeneous interface definitions across 10 different types, including AGVs, six-axis robotic arms, and industrial cameras, was constructed for experimental verification. The dataset covered various parameter naming conventions, including camelCase and underscore naming, and was divided into training, validation, and test sets in a 7:2:1 ratio. The experiments were run on a compute node operating on Ubuntu 20.04, with hardware consisting of two NVIDIA A100 Tensor Core GPUs and 256GB of RAM. The software environment was based on the PyTorch 1.10 deep learning framework. During model training, a contrastive loss function was used, optimizing the embedding space by minimizing the Euclidean distance between positive sample pairs (synonyms) and maximizing the Euclidean distance between negative sample pairs (non-synonyms). The optimizer used was AdamW, with an initial learning rate of 2e-5, and a linear warm-up was employed. The warp strategy, with a batch size of 32 and a training period of 50 epochs, uses a convergence criterion of no further decrease in validation set loss for 5 consecutive epochs. Experimental results show that the parameter mapping accuracy of this scheme on the test set reaches 96.8%, which is 24.5 percentage points higher than the traditional edit distance-based string matching method. Furthermore, the quantization error of QoS indicators is controlled within 1.2%, demonstrating that this scheme has the ability to accurately transform messy and heterogeneous data into high-quality standard capability representations, providing a reliable data foundation for subsequent scheduling decisions.

[0033] Specifically, in a specific embodiment of the present invention, the step of decomposing the task request based on the capability knowledge graph includes: constructing a prompting engineering input context, which contains a natural language description of the task request and standard concept classes associated with the task request; passing the prompting engineering input context into a large language model to drive the large language model to generate logical analysis text; parsing the logical analysis text to obtain action derivations matching the standard concept classes and execution path determinations based on action dependencies; generating an atomic task sequence based on the action derivations and execution path determinations, and labeling the subtasks in the sequence with capability constraint tags containing execution priority and deadlines, thereby completing the task request decomposition.

[0034] In this embodiment of the invention, the first step is to construct a prompting engineering input context capable of stimulating the logical reasoning ability of a large language model. The system employs a vector-based retrieval enhancement generation technique. Specifically, it uses a pre-trained Sentence-BERT model to encode the received natural language task description, such as producing 500 customized speed reducers, into a 768-dimensional query vector. The system then calculates the cosine similarity between this query vector and the attribute vectors of all standard concept class nodes in a graph database. The similarity threshold for determining successful association is determined through receiver operating characteristic (ROC) analysis on a constructed semantic relevance validation set. Specifically, a similarity of 0.75, corresponding to the maximum Youden index, is selected as the preset threshold. This allows for the selection of standard concept classes highly relevant to the task request, such as gear processing and bearing assembly, as background knowledge injected into the prompt template. The constructed prompting engineering context is then fed into a large language model fine-tuned by industrial domain instructions. This model preferably adopts the LLaMA-3-8B-Instruct architecture, containing 32 Transformer decoding layers, each equipped with 32 self-defined functions. The model employs an attention head and a 4096-dimensional hidden layer. The intermediate layers utilize the SwiGLU activation function to enhance non-linear expressiveness. Rotated Position Encoding (RoPE) is used for positional encoding to improve long sequence processing. During training, the model uses cross-entropy loss to calculate the difference between the predicted probability distribution and the true label. The optimizer is AdamW, and the learning rate decays from 2e-5 to 1e-6 using cosine annealing. The convergence criterion is set to ensure that the validation set perplexity does not decrease for three consecutive evaluation epochs. The model is driven to perform CoT reasoning on the input context, generating structured JSON-formatted logical analysis text containing step breakdowns and logical dependencies. Subsequently, a word segmentation algorithm based on a Bi-LSTM feature extractor is used to serialize the reasoning fields in the logical analysis text. A syntactic dependency tree is constructed based on the dependency parsing algorithm of the Arc-Standard transfer system. Candidate actions are extracted by traversing the core predicate nodes in the syntactic dependency tree, and the Euclidean distance between the word vector of the candidate action and the preset standard concept class vector is calculated. When the Euclidean distance is less than the semantic matching threshold 0, the result is considered valid.At 3 PM, the candidate action is mapped to a standard action node. Simultaneously, a keyword matching algorithm scans the text for logical connectors such as "premise is," "immediately afterwards," and "execute in parallel." Based on the dominant paths of these logical connectors in the syntactic dependency tree, directed edges are constructed between action nodes, generating an adjacency matrix representing the execution path determination. For example, when "refining must be performed after coarse processing is completed" is detected, the corresponding element in the adjacency matrix pointing from the coarse processing node to the fine processing node is set to 1. Finally, an atomic task sequence is generated based on the parsed action nodes and the directed acyclic graph dependencies determined by the adjacency matrix. The model's internal reward model scoring mechanism is then invoked to output an urgency score for each subtask. When the score exceeds the threshold of 0.8 determined through F1 score optimization, the subtask is marked as high priority, and the deadline is automatically calculated, thus generating an atomic task sequence with complete capability constraint labels. This enables intelligent transformation from fuzzy instructions to precise execution sequences.

[0035] Specifically, in this embodiment of the invention, step S4 includes: real-time collection of the current load rate, health status, and network latency data of each heterogeneous intelligent agent; construction of a multi-objective optimization model based on deep learning technology, and calculation of the semantic matching degree between the capability constraint labels of sub-tasks and heterogeneous intelligent agents based on the multi-objective optimization model; and global iterative solution of the multi-objective optimization model based on the semantic matching degree combined with the current load rate and execution cost factors to obtain a global task allocation scheme.

[0036] In a specific embodiment of this invention, Prometheus monitoring probes deployed on edge computing nodes first collect real-time operational status data for each heterogeneous agent. The collection frequency is set to once per second, and the data dimensions include CPU load rate, memory usage rate, health score calculated based on packet loss rate, and network round-trip time (RTT) between nodes. A sliding window algorithm with a time span of ten seconds is used to smooth the original time-series data to eliminate instantaneous jitter interference. A multi-objective optimization model is constructed based on deep learning technology. Specifically, this model is constructed using a dual-tower semantic matching neural network architecture combined with a multi-objective evolutionary algorithm: the left tower of the dual-tower network is the task encoder, which uses a 12-layer BERT architecture to process the capability constraint label text of atomic tasks; the right tower is the agent encoder, which uses a 3-layer fully connected layer (DenseLayer) to process the static capability representation of heterogeneous agents. Each layer contains 256 neurons, and the activation function is Swish to avoid gradient vanishing. The feature vectors output by the two towers are fused in the interaction layer through a dot product operation, and then connected to a regression prediction head composed of a 3-layer perceptron, outputting a semantic matching degree value between 0 and 1. During global iterative solution... The non-dominated sorting genetic algorithm, NSGA-II, is used as the solver. The semantic matching degree output by the deep learning model is used as one of the maximizing objective functions, while the variance of the real-time collected current load rate and the execution cost based on the energy pricing model are used as minimizing objective functions. A population containing 100 random assignment schemes is initialized. In each generation iteration, new individuals are generated by simulating binary crossover and polynomial mutation operators. The preset threshold for the mutation probability is not empirically specified but determined through hyperparameter optimization using a Bayesian optimization algorithm in a historical scheduling log replay environment. Using the hypervolume index as the evaluation criterion, the optimal value is searched within a search space of 0.01 to 0.2. Finally, the value preceding the inflection point of the hypervolume index decrease is selected. In this embodiment, 0.05 is preferred as the threshold. The iteration stops when the number of iterations reaches 200 or the displacement distance of the Pareto front of the population within 5 consecutive generations is less than the convergence threshold of 0.001 determined by grid search. Finally, the solution with the largest crowding distance is selected from the Pareto optimal solution set as the global task allocation scheme, which can ensure that the task allocation in high-concurrency scenarios is both accurate in matching capabilities and balanced in cluster load.

[0037] To verify the performance advantages of the task allocation method based on deep learning multi-objective optimization in this embodiment, a simulation experimental environment adapted from the publicly available Google Borg cluster dataset was constructed. This simulated a high-pressure scenario involving 500 heterogeneous nodes and 1000 burst task requests per second. The dataset was divided into training and test sets according to time series. The experiment was run on a workstation equipped with four NVIDIA RTX 4090 GPUs. The deep learning framework used was TensorFlow 2.10, and the training of the dual-tower matching model employed the contrastive loss function. The Loss algorithm aims to increase the Euclidean distance between mismatched tasks and agent pairs. The optimizer chosen is Adam, with a learning rate of 1e-4 and a batch size of 128. In the multi-objective optimization phase, this scheme is compared with traditional weighted round-robin and greedy algorithms. Evaluation metrics include average task completion time, cluster load balancing variance, and task assignment success rate. Experimental results show that this scheme reduces the average task completion time by 35% and the cluster load variance by 42% on the test set. Furthermore, even with severely limited heterogeneous resources, the task assignment success rate remains above 98.5%, demonstrating that this scheme effectively solves the resource scheduling problem under complex constraints by combining deep semantic matching with a multi-objective evolutionary algorithm, thus optimizing the overall system performance.

[0038] Specifically, in a specific embodiment of the present invention, the step of calculating the semantic matching degree between the capability constraint label of the sub-task and the heterogeneous intelligent agent according to the multi-objective optimization model includes: the multi-objective optimization model includes an input representation layer, a vector embedding layer, and a semantic interaction layer; the input representation layer receives the capability constraint label of the sub-task and the standard capability representation of the heterogeneous intelligent agent respectively; based on the vector embedding layer, the descriptive text in the capability constraint label and the standard capability representation are mapped to the task feature vector and the capability feature vector respectively; based on the semantic interaction layer, the feature cross result between the task feature vector and the capability feature vector is calculated and obtained, and the semantic matching degree is output based on the feature cross result.

[0039] First, a neural network model based on the Deep Structured Semantic Model (DSSM) architecture is constructed. This model logically consists of three cascaded stages: an input representation layer, a vector embedding layer, and a semantic interaction layer. Specifically, the input representation layer is mainly responsible for data preprocessing and formatting. It converts the capability constraint labels of subtasks into token sequences using a token segmenter and transforms the standard capability representations of heterogeneous agents (such as positioning accuracy of 0.01 mm) into numerical feature vectors. Based on the vector embedding layer, a dual-tower network structure is used to perform feature mapping on the heterogeneous data. The left tower processes the task text, using a 12-layer BERT-based pre-trained model with a fixed hidden layer dimension of 768. It captures long-distance dependencies in the text through a multi-head self-attention mechanism. The right tower processes capability representations, using a multilayer perceptron consisting of three fully connected layers. The number of neurons in each layer decreases sequentially to 512, 256, and 128, with the Leaky activation function selected. ReLU addresses the problem of neuron death in the negative interval. Both towers ultimately output 128-dimensional normalized feature vectors, namely the task feature vector and the ability feature vector. Based on the semantic interaction layer, to capture the fine-grained matching relationship between the two, simple cosine similarity calculation is abandoned. Instead, a feature cross-network is constructed. Specific operations include element-wise dot product, element-wise subtraction and absolute value taking, and direct concatenation of the task and ability feature vectors to form a 384-dimensional mixed feature vector. This mixed feature vector is then input into a three-layer fully connected layer for nonlinear transformation. Finally, the output is compressed to a closed interval of 0 to 1 using the Sigmoid activation function, yielding a semantic matching degree value representing the degree of matching. Regarding the positive and negative sample judgment threshold during model training, it is not randomly set but determined by plotting an ROC curve on the validation set and calculating the Youden exponent. Specifically, it iterates through the threshold interval from 0 to 1, finding the maximum value of the sum of sensitivity and specificity minus 1. In this embodiment, the calculated threshold is 0.72. During training, a sample is considered correctly classified only when the error between the predicted value and the true label is less than this threshold range.

[0040] Specifically, such as Figure 3 As shown, in the embodiments of the present invention, the method further includes performing data compatibility verification on the atomic task sequence, specifically: traversing adjacent upstream and downstream subtasks in the atomic task sequence that have data flow relationships; extracting the expected output data pattern corresponding to the upstream subtask and the required input data pattern corresponding to the downstream subtask based on the capability knowledge graph; quantifying the semantic structure distance between the expected output data pattern and the required input data pattern based on field name similarity, data type compatibility, and data structure hierarchy differences; and performing dynamic data adaptation processing on the upstream and downstream subtasks when the semantic structure distance exceeds a preset threshold.

[0041] In a specific embodiment of this invention, a traversal engine based on a depth-first search algorithm is first launched to sequentially scan each node in the atomic task sequence, identifying adjacent upstream and downstream subtask pairs with clear data flow relationships. Based on the capability knowledge graph stored in the Neo4j graph database, the expected output data pattern (Output Schema) of the upstream subtask and the required input data pattern (Input Schema) of the downstream subtask are extracted using Cypher queries. These patterns are defined using the standard JSON Schema format, including field names, data types such as Integer and String, and nested hierarchical structures. To accurately quantify the differences between the two, a multi-dimensional semantic structure distance calculation model is constructed. Specifically, this calculation model includes three independent measurement dimensions: the first dimension is field name similarity, which uses a pre-trained RoBERTa-large model to extract 768-dimensional context embedding vectors of field names and calculates the cosine distance between vectors; the second dimension is data type compatibility, which is based on a pre-set type conversion cost matrix (Type). CostMatrix queries the cost of converting from upstream to downstream types. For example, the cost of converting from Float to Integer is set to 0.5, while the cost of converting from String to Image is set to 1.0. The third dimension is the difference in data structure hierarchy. The Zhang-Shasha Algorithm is used to calculate the minimum tree edit distance between two schema trees, covering the weighted operation costs of inserting, deleting, and renaming nodes. The normalized values ​​of the above three dimensions are aggregated into the final semantic structure distance using a weighted summation formula. The distance threshold for determining whether adaptation is needed is determined by maximizing the F1 score on a validation set containing 50,000 sets of historical interface connection data. Specifically, the distance interval from 0 to 1 is traversed with a step size of 0.01, and the distance value corresponding to the peak F1 score (preferably 0.35 in this embodiment) is selected as the preset threshold. When the calculated semantic structure... When the structural distance exceeds the preset threshold, the data is determined to be incompatible, and dynamic data adaptation processing is triggered. This processing is based on the CodeT5+ code generation model, which contains a 12-layer encoder and a 12-layer decoder with a hidden layer dimension of 768. The input is a description of the differences between the source schema and the target schema, and the output is a Python transformation script that can perform data cleaning and reconstruction, such as dictionary mapping or type casting functions. This script is injected into the task execution container in real time, thereby dynamically smoothing out data differences at runtime and ensuring seamless connection and logical closed loop of data flow between heterogeneous intelligent agents.

[0042] Specifically, in a specific embodiment of the present invention, the step of performing dynamic data adaptation processing on the upstream subtask and the downstream subtask includes: calling a preset code generation model, inputting the difference features between the expected output data pattern and the required input data pattern into the code generation model to obtain data conversion code; configuring a data conversion runtime environment, loading the data conversion code based on the data conversion runtime environment, and encapsulating the data conversion code into an intermediate processing node; and serially inserting the intermediate processing node between the upstream subtask and the downstream subtask to complete the dynamic data adaptation processing.

[0043] In a specific embodiment of the present invention, during the dynamic data adaptation process for upstream and downstream subtasks, the following steps are taken: First, data flow anomalies are captured based on the heterogeneous system integration bus. The actual output JSON Schema of the upstream task and the required input JSON Schema of the downstream task are extracted. The differences in field naming, data type, and nested structure between the two are calculated using a differential algorithm, and these differences are serialized into natural language prompts. Then, a pre-defined code generation model is invoked. Specifically, this model is a finely tuned version of CodeT5+ built on the Transformer architecture: the model contains a 12-layer bidirectional Transformer encoder and a 12-layer unidirectional Transformer decoder, with a hidden layer dimension of 768, a multi-head attention mechanism with 12 heads, a feedforward network dimension of 3072, and GELU activation function to improve nonlinear fitting ability. The maximum input length is 1024 tokens. Based on the received difference feature descriptions, the model generates Python data conversion code that maps the source data format to the target data format through autoregressive decoding. The confidence threshold for the generated code is determined by performing a Pascal test on the code generation validation set. The correlation analysis between s@1 (pass rate) and the model confidence score determines the confidence score. Specifically, it iterates through the confidence interval from 0 to 1 with a step size of 0.05, and selects the lowest confidence score corresponding to Pass@1 reaching 95%. In this embodiment, 0.88 is preferred as the preset threshold. The code is only adopted when the average Logits probability of the generated code exceeds this threshold. Configure the data conversion runtime environment. Based on lightweight container technology such as Docker or WebAssembly runtime, build an isolated sandbox environment, load the generated data conversion code, and encapsulate the code into an intermediate processing node with an independent HTTP call endpoint through a serverless architecture. Finally, based on graph theory algorithm, dynamically update the directed acyclic graph (DAG) of task execution, disconnect the original direct connection edges, and create two new directed edges from the upstream subtask to the intermediate processing node and from the intermediate processing node to the downstream subtask. This allows the intermediate processing node to be inserted serially into the execution chain, completing the dynamic data adaptation processing.

[0044] Specifically, in a specific embodiment of the present invention, the method further includes optimizing the execution process of the task request and the multi-objective optimization model, specifically: establishing a task instruction message queue based on a unified scheduling center, and distributing scheduling instructions to heterogeneous intelligent agents according to the capability knowledge graph and the task instruction message queue; when multiple heterogeneous intelligent agents are detected to be contending for the same resource, conflict resolution is performed according to the execution priority in the capability constraint label; when an execution anomaly of a certain heterogeneous intelligent agent is detected, a candidate heterogeneous intelligent agent with the capability matching the sub-task is retrieved based on the capability constraint label, and the execution instructions of the sub-task are rescheduled to the candidate heterogeneous intelligent agent; the unified scheduling center records the entire process execution log, and uses the entire process execution log and the result of the task request as feedback data to update the multi-objective optimization model and perform iteration.

[0045] In a specific embodiment of the present invention, a high-throughput distributed task instruction message queue is first built based on Apache Kafka. Independent Topic partitions are created for different categories of heterogeneous intelligent agents according to the topology in the capability knowledge graph. The scheduling center, acting as a producer, serializes atomic task instructions into Protocols. The data is formatted into buffers and written to the corresponding partition. The intelligent agent, acting as a consumer, pulls instructions via the gRPC interface. When multiple heterogeneous intelligent agents are detected contending for the same shared resource, such as Ethernet bandwidth or computing nodes containing GPUs, a Redis-based distributed lock mechanism is introduced to resolve the conflict. Specifically, this mechanism runs a Lua script to atomically compare the priority of the task currently holding the lock with the priority of the task requesting the lock. The priority value is derived from the capability constraint label, ranging from 1 to 10. When the requester's priority is higher than the holder's, preemption logic is triggered, the lock is forcibly released, and the lower-priority task is pushed back to the head of the message queue to wait for retry. When an anomaly is detected during the execution of a heterogeneous intelligent agent, the system initiates a fault migration process. The threshold for determining heartbeat timeout is determined by fitting a Gaussian distribution to the heartbeat intervals of the normal operation logs over a week. Specifically, the mean plus three times the standard deviation is used as the dynamic threshold. For example, if the calculation result is 350 milliseconds, an anomaly is identified upon timeout. Based on the "equivalent capability" relationship in the capability knowledge graph, a backup agent with the same capability is queried, and the task context is seamlessly migrated to the new node for retry. At the same time, the unified scheduling center records the entire process execution log, including task execution time, resource consumption, and final state, through the ELK technology stack. These logs are used as real feedback data. An online learning strategy is adopted to iteratively update the multi-objective optimization model. The specific update process is as follows: the weights of the bottom feature extraction layer in the dual-tower model are frozen, and only the top three-layer perceptron regression prediction head is fine-tuned. The mean squared error loss function is used to calculate the residual between the prediction matching degree and the normalized score of the actual execution effect. The optimizer is stochastic gradient descent, and the learning rate is set to 1e-5. Backpropagation is performed once every 100 feedback data points are accumulated, so that the model can continuously adapt to the performance drift caused by the dynamic physical environment and equipment aging.

[0046] To verify the stability and adaptability of the aforementioned closed-loop optimization mechanism in long-term operation, a 72-hour continuous industrial production simulation experiment was constructed. The dataset included 10,000 task scheduling logs and corresponding equipment status stream data collected from an automated factory. 5% of random equipment failures and network congestion events were artificially injected. The data from the first 24 hours of the timeline was used as the initial training set, and the data from the subsequent 48 hours was used for online validation. The experimental environment ran on a cluster consisting of three Kubernetes nodes, each configured with a 16-core CPU and 64GB of memory. The online updates of the deep learning model were deployed using TensorFlow. Within the Serving container, the experiment compared the system performance of enabling the online update mechanism with that of using a static model. The evaluation metrics selected were average task response time, fault recovery success rate, and the decay curve of model prediction accuracy. The experimental results showed that after 48 hours of operation, the model prediction accuracy of this solution remained at a high level of 96% compared to the static model, which decreased to 78%. The average task response time gradually decreased by 18% as the model iterated. Furthermore, when faced with sudden equipment failures, the priority-based conflict resolution and migration mechanism controlled the average task lag time to within 500 milliseconds. This demonstrates that the solution has strong self-evolution capabilities and robustness, ensuring long-term efficient collaboration of the multi-agent platform in complex and ever-changing environments.

[0047] This invention also provides a unified scheduling system for an enterprise-level multi-agent platform, such as... Figure 4 As shown, it includes: a central construction module, which constructs a unified scheduling center, the unified scheduling center communication connection including all heterogeneous intelligent agents and the capability description data corresponding to each heterogeneous intelligent agent; a graph construction module, which performs semantic normalization on the capability description data of each heterogeneous intelligent agent and constructs a capability knowledge graph; a task decomposition module, which decomposes the received task request according to the capability knowledge graph to obtain an atomic task sequence, the atomic task sequence containing capability constraint tags for each sub-task; a task allocation module, which collects the real-time running status of each heterogeneous intelligent agent and adaptively matches the sub-tasks according to the real-time running status and the capability constraint tags to obtain a global task allocation scheme; and a collaborative execution module, which drives the heterogeneous intelligent agents to collaboratively execute the task request based on the global task allocation scheme.

[0048] As described above, the unified scheduling method for an enterprise-level multi-agent platform disclosed in this invention significantly improves the semantic normalization level of heterogeneous resources, the logical accuracy of task decomposition, and the adaptive matching efficiency of global scheduling, overcoming the technical shortcomings of incompatible interface standards, rigid logical decomposition, and disconnected scheduling decisions. The unified scheduling method for the enterprise-level multi-agent platform includes: S1 constructing a unified scheduling center, the unified scheduling center communication connection including all heterogeneous agents and capability description data corresponding to each heterogeneous agent; S2 performing semantic normalization on the capability description data of each heterogeneous agent and constructing a capability knowledge graph; S3 decomposing the received task request according to the capability knowledge graph based on the received task request, obtaining an atomic task sequence, the atomic task sequence containing capability constraint tags for each sub-task; S4 collecting the real-time running status of each heterogeneous agent, and adaptively matching the sub-tasks according to the real-time running status and the capability constraint tags to obtain a global task allocation scheme; S5 driving the heterogeneous agents to collaboratively execute the task request based on the global task allocation scheme. Specifically, by constructing a unified scheduling center to aggregate heterogeneous intelligent agent resources, and using semantic normalization technology to standardize and map multi-source heterogeneous capability description data and construct a capability knowledge graph, a unified interaction standard is established at the semantic level to shield the data differences of underlying heterogeneous interfaces, thereby solving the technical problem of incompatible heterogeneous interface data standards. On this basis, the system uses this capability knowledge graph as the logical deduction base to deeply parse and decompose the received abstract task requests into atomic task sequences containing precise capability constraint labels. By transforming abstract business logic into machine-recognizable structured instructions with capability definitions, the system overcomes the defect that complex business logic cannot be automatically decomposed. Furthermore, the system collects the running status of each intelligent agent in real time and performs joint analysis and adaptive matching of this dynamic load data with the capability constraint labels of the aforementioned sub-tasks to generate a global task allocation scheme. Through this strategy of deeply integrating static capability requirements with dynamic running status, the system ensures that scheduling decisions respond to node load changes in real time, effectively solving the problem of cross-system collaboration failure and resource allocation imbalance caused by the disconnect between scheduling decisions and real-time load in existing technologies. Compared with existing technologies, this invention can significantly improve the semantic normalization level of heterogeneous resources, the logical accuracy of task decomposition, and the adaptive matching efficiency of global scheduling, overcoming the technical shortcomings of incompatible interface standards, rigid logical decomposition, and disconnected scheduling decisions.

[0049] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A unified scheduling method for an enterprise-level multi-agent platform, characterized in that, include: S1 constructs a unified scheduling center, which communicates with all heterogeneous intelligent agents and receives capability description data corresponding to each heterogeneous intelligent agent. S2 performs semantic normalization on the capability description data of each heterogeneous intelligent agent and constructs a capability knowledge graph; S3, based on the received task request, decomposes the task request according to the capability knowledge graph to obtain an atomic task sequence, wherein the atomic task sequence contains capability constraint tags for each subtask; S4 collects the real-time running status of each heterogeneous intelligent agent, and performs adaptive matching of the sub-tasks based on the real-time running status and the capability constraint labels to obtain a global task allocation scheme. Based on the global task allocation scheme, S5 drives the heterogeneous intelligent agents to collaboratively execute the task request.

2. The unified scheduling method for an enterprise-level multi-agent platform according to claim 1, characterized in that, The steps involved in building a capability knowledge graph include: Receive the capability description data of each heterogeneous intelligent agent, perform semantic normalization on the capability description data of each heterogeneous intelligent agent, and generate a standard capability representation; Generate capability entity nodes based on the aforementioned standard capability representation; Extract the logical collaboration relationships between the heterogeneous intelligent agents and map these relationships as topological connection edges connecting the capability entity nodes to construct a capability knowledge graph.

3. The unified scheduling method for an enterprise-level multi-agent platform according to claim 2, characterized in that, The step of semantically normalizing the capability description data for each heterogeneous agent includes: The adapter interface according to the communication protocol receives the interface definition file and quality of service metrics for each heterogeneous intelligent agent. A pre-defined capability ontology is constructed based on ontology modeling technology. The pre-defined capability ontology includes standard concept classes and attribute constraints. The interface definition file is parsed, the heterogeneous input and output parameters are extracted, and the heterogeneous input and output parameters are defined as instance data, mapped and populated into the corresponding standard concept class in the preset capability ontology; The service quality indicators are quantified based on the attribute constraints to generate standard capability representations.

4. The unified scheduling method for an enterprise-level multi-agent platform according to claim 3, characterized in that, The step involves decomposing the task request based on the capability knowledge graph, including: Construct a prompting engineering input context, which includes a natural language description of the task request and the standard concept class associated with the task request; The prompting process is input into the context and passed into the large language model, which then drives the large language model to generate logical analysis text. Parse the logical analysis text to obtain the action derivation matching the standard concept class and the execution path determination based on action dependency; The atomic task sequence is generated based on the action derivation and the execution path determination, and the sub-tasks in the sequence are labeled with capability constraint tags containing execution priority and deadline, thus completing the task request decomposition.

5. The unified scheduling method for an enterprise-level multi-agent platform according to claim 4, characterized in that, Step S4 includes: Real-time collection of current load rate, health status, and network latency data for each of the heterogeneous intelligent agents; A multi-objective optimization model is constructed based on deep learning technology, and the semantic matching degree between the capability constraint label of the sub-task and the heterogeneous agent is calculated based on the multi-objective optimization model. Based on the semantic matching degree and the current load rate and execution cost factors, the multi-objective optimization model is solved globally through iteration to obtain a global task allocation scheme.

6. The unified scheduling method for an enterprise-level multi-agent platform according to claim 5, characterized in that, The step of calculating the semantic matching degree between the capability constraint labels of the sub-task and the heterogeneous agent based on the multi-objective optimization model includes: The multi-objective optimization model includes an input representation layer, a vector embedding layer, and a semantic interaction layer; The input representation layer receives the capability constraint labels of the subtasks and the standard capability representations of the heterogeneous agents, respectively. Based on the vector embedding layer, the descriptive text in the capability constraint label and the standard capability representation are mapped to task feature vectors and capability feature vectors, respectively. Based on the semantic interaction layer, the feature cross-result between the task feature vector and the capability feature vector is calculated and obtained, and the semantic matching degree is output based on the feature cross-result.

7. The unified scheduling method for an enterprise-level multi-agent platform according to claim 1, characterized in that, It also includes performing data compatibility verification on the atomic task sequence, specifically: Traverse the adjacent upstream and downstream subtasks in the atomic task sequence that have data flow relationships; Based on the capability knowledge graph, the expected output data pattern corresponding to the upstream sub-task and the required input data pattern corresponding to the downstream sub-task are extracted respectively. Based on field name similarity, data type compatibility, and data structure hierarchy differences, the semantic structure distance between the expected output data pattern and the required input data pattern is quantitatively calculated. When the semantic structure distance exceeds a preset threshold, dynamic data adaptation processing is performed on the upstream subtask and the downstream subtask.

8. The unified scheduling method for an enterprise-level multi-agent platform according to claim 7, characterized in that, The step of performing dynamic data adaptation processing on the upstream subtask and the downstream subtask includes: Call the preset code generation model, input the difference features between the expected output data pattern and the required input data pattern into the code generation model, and obtain the data conversion code; Configure the data conversion runtime environment, load the data conversion code based on the data conversion runtime environment, and encapsulate the data conversion code as an intermediate processing node; The intermediate processing node is inserted serially between the upstream subtask and the downstream subtask to complete the dynamic data adaptation process.

9. The unified scheduling method for an enterprise-level multi-agent platform according to claim 8, characterized in that, It also includes optimizing the execution process of the task request and the multi-objective optimization model, specifically: A task instruction message queue is established based on the unified scheduling center, and scheduling instructions are distributed to the heterogeneous intelligent agents according to the capability knowledge graph and the task instruction message queue. When multiple heterogeneous intelligent agents are detected to be contending for the same resource, the conflict is resolved according to the execution priority in the capability constraint label. When an abnormality is detected in the execution of a certain heterogeneous intelligent agent, a candidate heterogeneous intelligent agent with the capability matching the subtask is retrieved based on the capability constraint tag, and the execution instructions of the subtask are rescheduled to the candidate heterogeneous intelligent agent. The unified scheduling center records the entire process execution log and uses the entire process execution log and the result of the task request as feedback data to update and iterate the multi-objective optimization model.

10. A unified scheduling system for an enterprise-level multi-agent platform, characterized in that, The system is used to execute the unified scheduling method for an enterprise-level multi-agent platform as described in any one of claims 1 to 8; including: The central construction module builds a unified scheduling center. The unified scheduling center's communication connections include all heterogeneous intelligent agents and the capability description data corresponding to each heterogeneous intelligent agent. The graph construction module performs semantic normalization on the capability description data of each heterogeneous intelligent agent and constructs a capability knowledge graph. The task decomposition module, based on the received task request, decomposes the task request according to the capability knowledge graph to obtain an atomic task sequence, wherein the atomic task sequence contains capability constraint tags for each subtask. The task allocation module collects the real-time running status of each heterogeneous intelligent agent, and adaptively matches the sub-tasks based on the real-time running status and the capability constraint labels to obtain a global task allocation scheme. The collaborative execution module, based on the global task allocation scheme, drives the heterogeneous intelligent agents to collaboratively execute the task request.

Images