An agent cooperation method and system for a production field artificial intelligence system

Abstract

The application discloses an agent cooperation method and system for a production field artificial intelligence system, and belongs to the field of artificial intelligence. The method comprises the following steps: constructing a production field ontology library and a process rule library covering four dimensions of a production process, process parameters, equipment capacity and quality standards, and designing a semantic embedding layer to convert heterogeneous input into a unified semantic vector fused with field ontology knowledge; executing semantic disambiguation, completion and consistency checking through a semantic enhancement analysis engine to generate a structured analysis result; automatically generating a multi-level cooperation strategy of a strategic layer task allocation, a tactical layer resource scheduling and an execution layer execution sequence based on the analysis result, and deciding an optimal cooperation scheme through a strategy selection network; and simultaneously deploying a runtime monitoring module to collect execution indexes and trigger dynamic reconstruction. The application realizes accurate mapping from task semantics to cooperation strategies, and improves the adaptability of the system to a dynamic production environment.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence, specifically relating to an intelligent agent collaboration method and system for artificial intelligence systems in the production field. Background Technology

[0002] In artificial intelligence systems in the production field, intelligent agents need to generate collaborative strategies based on task descriptions and system states. However, there is a semantic gap between the textual form of task descriptions and the collaborative strategies that the system can execute. How to bridge this gap and achieve accurate mapping from task semantics to collaborative strategies is a core problem that urgently needs to be solved.

[0003] Common methods for generating intelligent agent collaboration strategies typically employ static task allocation or fixed mapping based on predefined rules. These methods use manually written rule bases or templates to simply convert task descriptions into collaboration strategies. However, this rule- and template-dependent mapping approach lacks a deep understanding of the complex business semantics of the production domain. It fails to accurately identify process constraints and resource dependencies in task descriptions, leading to a disconnect between the generated collaboration strategies and the actual production scenario. Strategy selection becomes rigid and unable to adapt to dynamically changing production environments, hindering reliable collaboration decisions and becoming a key bottleneck restricting the improvement of the intelligence level of production systems. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an agent collaboration method for artificial intelligence systems in the production field, which can effectively solve the problems mentioned in the background.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides an agent collaboration method for artificial intelligence systems in the production field, the method comprising the following steps: A production domain ontology library and a process rule library are constructed. The production domain ontology library adopts a hierarchical directed acyclic graph as the underlying storage structure, covering four dimensions: production process, process parameters, equipment capabilities, and quality standards. The process rule library covers three types of rule systems: process constraint rules, resource dependencies, and execution conditions. The three heterogeneous inputs—task text description, product process requirements, and resource conditions—are transformed into a unified semantic vector representation that integrates domain ontology knowledge through a designed semantic embedding layer. The unified semantic vector representation is input into the semantic enhancement parsing engine. The semantic enhancement parsing engine performs semantic disambiguation and semantic completion based on the production domain ontology library and the process rule library. After incorporating the process constraint rules as constraint vectors into the semantic vectors, semantic consistency verification is performed to output a structured parsing result containing task objectives, constraints, resource requirements, and execution sequences. Based on the structured parsing results, a multi-level collaborative strategy is automatically generated, which includes strategic task allocation, tactical resource scheduling, and execution order. The strategy layers are decoupled through a loosely coupled architecture and coordinated through constraint propagation and conflict arbitration mechanisms. The unified semantic vector representation is combined with the current system state input strategy selection network to determine the optimal cooperation scheme; Deploy a runtime monitoring module to collect execution metrics in real time and trigger adjustments to the collaboration mode based on feedback signals to support dynamic refactoring of strategies.

[0006] Furthermore, the production domain ontology library adopts a hierarchical directed acyclic graph as its underlying storage structure, covering the production process dimension, specifically including: The production process is defined as a node in the hierarchical directed acyclic graph, and the logical relationship between processes is defined as an edge in the hierarchical directed acyclic graph. Process time constraints are stored as node attributes, which include four parameters: baseline execution time, longest execution time, shortest execution time, and flexible time window.

[0007] Furthermore, the production domain ontology library covers equipment capabilities and quality standards, specifically including: The device state description adopts a finite state machine model. The state set includes six states: idle, running, debugging, fault, maintenance, and standby. The state transition rules are defined by a state transition matrix. The matrix element takes the value 0 or 1. A value of 1 indicates that the transition from a row state to a column state is allowed, and a value of 0 indicates that the transition from a row state to a column state is prohibited. When integrating the production domain ontology library and the process rule library, a cross-reference index is established, and a graph database is used as the underlying storage engine of the knowledge base after the integration of the production domain ontology library and the process rule library. Ontology nodes and edges are mapped to vertices and edges in the graph database, and rules and conditions are attached to the corresponding ontology nodes as vertex attributes, forming a unified knowledge graph structure that can be traversed and queried.

[0008] Furthermore, the semantic embedding layer converts the three heterogeneous inputs—the task text description, the product process requirements, and the resource conditions—into a unified semantic vector representation, specifically including: The initial feature vector, process feature vector, and resource feature vector are each mapped to a unified semantic space dimension through their respective linear projection layers; Through the cross-modal attention module, the query vector obtained by linear projection of the initial feature vector and the key vector obtained by linear projection of the process feature vector are used to calculate the dot product similarity and then scaled and normalized by softmax to obtain the scalar attention weights. The calculated scalar attention weights are applied to the corresponding value vectors, and then all weighted value vectors are added element by element to obtain a weighted fusion vector.

[0009] Furthermore, the semantic enhancement parsing engine performs semantic disambiguation and semantic completion, specifically including: The semantic disambiguation uses the concept hierarchy in the production domain ontology to query the equipment type and quality standards associated with the current process where the polysemous word is located, and determines the accurate meaning of the polysemous word in the current task context. The semantic completion is achieved by calculating the unified semantic vector representation and the first semantic vector in the production domain ontology. The dot product similarity of the embedding vectors corresponding to each conceptual entity is used to infer and complete the implicit entities, implicit relationships, and implicit constraints missing in the task description.

[0010] Furthermore, the integration of process constraint rules as constraint vectors into semantic vectors specifically includes: The structured process constraint rules constructed in step S1 are mapped into constraint vectors using a rule encoder. Based on the degree of enforcement of the process constraint rules, the degree of enforcement coefficient of the process constraint rules is determined; wherein, the degree of enforcement coefficient of hard constraints is 1.0, and the degree of enforcement coefficient of soft constraints is a value that is dynamically adjusted according to the task type, resource availability and historical execution data; The coverage coefficient of the process constraint rule is determined based on the ratio of the number of processes involved in the process constraint rule to the total number of processes. No. Semantic fusion weights of constraints The calculation formula is ,in For the first The degree of enforceability of the constraint. For the first Coverage coefficient of the constraint; Based on the semantic fusion weights, the constraint vector is integrated into the semantic vector through a constraint-aware attention module.

[0011] Furthermore, the semantic consistency verification includes three stages: syntactic consistency verification, domain consistency verification, and process consistency verification. The process consistency check checks whether the process parameter configuration meets the boundary conditions of the process constraint rules and whether the process sequence meets the process sequence constraint rules. For soft conflicts found during the check, a constraint relaxation algorithm is used for fault tolerance processing. While maintaining the hard constraints, the soft constraints are appropriately relaxed, and the relaxation range does not exceed 10% of the constraint boundary value.

[0012] Furthermore, the unified semantic vector representation is input to the current system state policy selection network to determine the optimal cooperation scheme, specifically including: The policy selection network includes a state encoder, a semantic encoder, a fusion layer, a decoder, and a scoring output layer; The state encoder adopts a graph neural network architecture, models the current system state as a resource node graph, and aggregates the features of resource nodes and their neighboring nodes through a message passing mechanism to generate a state feature vector. The semantic encoder employs a multilayer perceptron architecture to encode the unified semantic vector representation and generate a semantic feature vector. The fusion layer uses the semantic feature vector as a query and the state feature vector as a key and value to calculate the attention weight distribution and generate a joint feature vector. The decoder generates candidate cooperative schemes based on the joint feature vector; The scoring output layer calculates the comprehensive score of each candidate collaboration scheme through a comprehensive scoring function, and decides on the candidate collaboration scheme with the highest comprehensive score as the optimal collaboration scheme.

[0013] Furthermore, the various strategy levels are coordinated through constraint propagation and conflict arbitration mechanisms, and the runtime monitoring module triggers adjustments to the cooperation mode based on feedback signals, specifically including: When inter-layer strategies conflict, an arbitration priority formula is used. Determine the order in which to handle each conflict, among which Arbitration priority for conflicts, These are the weighting coefficients for the strategy level: 3 for the strategic level, 2 for the tactical level, and 1 for the execution level. The coefficient represents the degree of enforceability of the constraint; arbitration decisions prioritize conflicts with higher arbitration priority values. The runtime monitoring module uses a dual triggering mechanism of threshold monitoring and trend prediction to trigger the adjustment of the collaboration mode; when dynamically reconstructing the strategy, hot switching technology is used to complete the dynamic adjustment of the strategy structure while retaining the state and execution progress of the unchanged strategy parts.

[0014] On the other hand, the present invention provides an intelligent agent collaboration system for artificial intelligence systems in the production field, comprising: The ontology library construction module is used to perform the steps of constructing the production domain ontology library and the process rule library as described in claim 1. The ontology library construction module uses a graph database as the underlying storage engine. The semantic embedding module is used to perform the step in claim 1 of converting the three types of heterogeneous inputs—task text description, product process requirements, and resource conditions—into a unified semantic vector representation that integrates domain ontology knowledge; The semantic parsing module is used to perform the step of outputting structured parsing results through the semantic enhancement parsing engine as described in claim 1; The strategy generation module is used to perform the step of automatically generating a multi-level collaborative strategy as described in claim 1, which includes strategic-level task allocation, tactical-level resource scheduling, and execution-level execution order. The strategy decision module is used to execute the step in claim 1 of inputting the unified semantic vector representation and the current system state into the strategy selection network to determine the optimal cooperation scheme; The runtime monitoring module is used to execute the steps of the deployment runtime monitoring module as described in claim 1, which collects execution metrics in real time and triggers the adjustment of the collaboration mode based on feedback signals.

[0015] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention constructs a domain ontology library covering four dimensions: production process, process parameters, equipment capabilities, and quality standards. It also designs a multimodal semantic embedding layer to uniformly map three types of heterogeneous inputs—task text descriptions, product process requirements, and resource conditions—into semantic vectors. Combined with a semantic enhancement parsing engine, it performs disambiguation, completion, and consistency checks, effectively bridging the semantic gap between natural language task descriptions and system-executable collaborative strategies. This solves the problem that traditional methods rely on fixed rule templates and lack deep semantic understanding.

[0016] 2. This invention automatically generates three-layer collaborative strategies—strategy, tactical, and execution—based on structured analysis results. The layers are decoupled using a loosely coupled architecture and coordinated through constraint propagation and conflict arbitration mechanisms. This allows each layer to evolve independently to cope with local changes while ensuring global consistency. At the same time, it utilizes a strategy selection network to dynamically decide the optimal solution based on the real-time system state, overcoming the shortcomings of static allocation methods in terms of rigid strategy selection and difficulty in adapting to dynamic production environments.

[0017] 3. The deployment runtime monitoring module of this invention uses a dual triggering mechanism of threshold monitoring and trend prediction to perceive system performance deviations in real time, supports the adjustment of collaborative modes based on feedback signals, and adopts hot switching technology to complete incremental or full updates of the strategy structure without interrupting execution. This enables the system to respond quickly to emergencies such as resource anomalies and task changes, and significantly improves the robustness and continuous optimization capability of the production system in dynamic and complex environments. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall technical solution for an intelligent agent collaboration method in an artificial intelligence system for the production field.

[0019] Figure 2 This is a schematic diagram illustrating the core principles of semantic enhancement parsing and vectorized representation.

[0020] Figure 3 This is a logical flowchart of the semantic understanding-based adaptive hierarchical collaboration strategy generation mechanism.

[0021] Figure 4 A schematic diagram illustrating the core principles of network selection and optimal collaboration scheme decision-making for strategy selection.

[0022] Figure 5 This diagram illustrates the multi-level interaction relationships and data flow of the runtime adaptive refactoring module.

[0023] Figure 6 A logical flow framework diagram for the automatic generation mechanism of hierarchical collaboration strategies. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1 To be continued Figure 6 The present invention will be further described in detail below with reference to specific embodiments.

[0025] This invention provides an agent collaboration method for artificial intelligence systems in the production domain. This method, based on a semantically enhanced domain knowledge understanding architecture, achieves end-to-end intelligent mapping from task description to collaboration strategy. The method constructs a production domain ontology library and a process rule library as its knowledge foundation. It designs a multimodal semantic embedding layer to transform heterogeneous inputs into a unified vector representation. A semantically enhanced parsing engine performs deep semantic understanding and constraint injection. Based on the parsing results, it automatically generates three layers of collaboration strategies: strategy, tactics, and execution. A strategy selection network dynamically determines the optimal collaboration scheme based on the real-time system state, and a runtime monitoring module enables adaptive adjustment and dynamic reconstruction of the strategies.

[0026] The first step, S1, involves constructing a production domain ontology library and a process rule library. The ontology library covers four dimensions: production process, process parameters, equipment capabilities, and quality standards. The process rule library encompasses three rule systems: process constraint rules, resource dependencies, and execution conditions. This will be implemented according to the following steps.

[0027] Step S101: Construct the production process ontology structure, using a hierarchical directed acyclic graph as the underlying storage structure. Production processes are defined as nodes in the graph, and logical relationships between processes are defined as edges. Node attributes include a unique identifier for the process, the process name, the process type, the baseline value of the process execution time, and a process parallelism identifier.

[0028] Directed edges are assigned relation type attributes, categorized into four types: serial dependency, parallel equivalence, conditional branching, and cyclic execution. Process time constraints are stored as node attributes, including four parameters: baseline execution time, longest execution time, shortest execution time, and flexible time window.

[0029] Step S102: Construct the process parameter ontology structure. An attribute graph structure is used, with process parameter entities defined as nodes and coupling relationships between parameters defined as edges. Each node includes the process parameter name, parameter value range, coupling relationship between parameters, and parameter optimization objective.

[0030] Parameter values ​​are stored as closed intervals, including lower limit, upper limit, and recommended value. Coupling relationships between parameters are represented by weighted directed edges, where weights indicate coupling strength and direction indicates the direction of causality. Coupling types are categorized into positive coupling, negative coupling, and nonlinear coupling. Parameter optimization objectives support both single-objective and multi-objective optimization, with optimization directions including maximization, minimization, and approaching the objective value.

[0031] Step S103: Construct the equipment capability ontology structure. A multi-dimensional attribute tag system is used for organization, with equipment entities defined as nodes and relationships between devices defined as edges. Each node includes a device type definition, device processing capacity, and device status description. The device status description uses a finite state machine model, with a set of states including six types: idle, running, debugging, fault, maintenance, and standby.

[0032] State transition rules are defined by a state transition matrix. The matrix elements take the values ​​0 or 1. A value of 1 indicates that the transition from a row state to a column state is allowed, while a value of 0 indicates that the transition is prohibited.

[0033] Step S104: Construct the quality standard ontology structure. A hierarchical decision tree organization is adopted, with quality indicator definitions and quality judgment rules defined as nodes, and the relationships between indicators defined as edges. Each node contains a quality indicator definition, a quality testing method, and a quality judgment rule. The quality judgment rule is associated with a judgment condition expression and a judgment level mapping table. The judgment condition expression is a set of logical expressions that take quality indicator measurement values ​​as input and level values ​​as output.

[0034] Step S105 defines process constraint rules, which cover four subcategories: process sequence constraint rules, process parameter constraint rules, resource usage constraint rules, and quality requirement constraint rules. Process sequence constraint rules include source process identifier, target process identifier, constraint type, and enforcement level. The constraint type is divided into four types: must be executed sequentially, cannot be executed simultaneously, takes precedence over, and immediately follows. The enforcement level is divided into two categories: hard constraints and soft constraints.

[0035] The process parameter constraint rules support three forms: linear constraints, nonlinear constraints, and discrete constraints.

[0036] The linear constraint is calculated as follows: for constraints involving Linear constraints on process parameters, with parameter variables as follows: The corresponding weighting coefficient is The threshold is The constraint expression is:

[0037] or

[0038] in This represents the total number of process parameters. For the summation index, For the first The current value of each process parameter For the first The weighting coefficients of each process parameter in this constraint This represents the constraint threshold constant. If it is an equality constraint, it is expressed as:

[0039] Nonlinear constraints are expressed as multiplication and division operations or transcendental function relationships between parameters, while discrete constraints are expressed as the requirement that parameter values ​​must belong to a predefined set of discrete values.

[0040] Resource usage constraints are stored in the form of resource quotas, defining the upper limit for each resource type and the principle of fairness in usage. The principle of fairness in usage is based on the upper limit of the variance in the amount of resources obtained by each user. Assume there are a total of... When multiple users acquire the same type of resource, the first... The amount of resources obtained by each user is The arithmetic mean of the amount of resources acquired by all users is:

[0041] in The total number of users, For the summation index, For the first The amount of resources allocated to each user This represents the arithmetic mean of the resource quantities. The variance of the resource allocation is... The calculation method is as follows:

[0042] when When it is determined to satisfy the principle of fairness, among which This represents the upper limit threshold for variance. Quality requirement constraints are stored in the form of quality objectives, defining the required quality indicators and the upper limit for the nonconforming rate for each process.

[0043] Step S106 defines resource dependencies, covering four types: task-equipment dependencies, task-personnel dependencies, task-material dependencies, and task-tool dependencies. Task-equipment dependencies are stored in a resource allocation matrix, where rows correspond to task identifiers, columns to equipment identifiers, and elements are allocation weight values ​​ranging from 0 to 1. A value of 0 indicates no dependency, and a value greater than 0 indicates a dependency. Each task is marked with either an exclusive usage mode or a shared usage mode.

[0044] Task-person dependency relationships are stored using a skill matching matrix, with rows corresponding to task identifiers, columns corresponding to personnel skill identifiers, and elements representing matching scores, ranging from 0 to 100 integers. Task-material dependency relationships are stored using a bill of materials, with each item containing a task identifier, material type identifier, and material quantity requirement. Task-tool dependency relationships are stored using a tool list, with each item containing a task identifier, tool type identifier, and tool quantity limit.

[0045] Step S107 defines the execution conditions, covering four trigger condition types: prerequisite task completion conditions, resource availability conditions, environmental status conditions, and security compliance conditions. Prerequisite task completion conditions are stored in a task dependency graph, using a directed acyclic graph structure. Nodes are task identifiers, and directed edges are labeled with dependency types, which are divided into three types: full completion dependency, any completion dependency, and proportional completion dependency.

[0046] A full completion dependency requires all source tasks to be completed before the target task can be triggered; any completion dependency requires at least one source task to be completed before triggering. A proportional completion dependency requires the completion percentage of source tasks to reach a preset threshold before triggering. Let the source task set contain... There are [number] tasks, and the number of completed tasks is [number]. Completion rate for:

[0047] in The total number of tasks in the source task set. This is used to count the number of source tasks that have been completed at any given time. The target task is triggered at the time. The preset completion rate trigger threshold has a value range of [value range missing]. arrive Between. Resource availability conditions define the availability thresholds for each resource type. Environmental condition conditions define the target ranges for temperature, humidity, cleanliness, and light. Safety compliance conditions include personnel qualification requirements and equipment safety operating procedures.

[0048] Step S108 integrates the ontology structure, constraint rules, dependencies, and execution conditions constructed in steps S101 to S107 into a unified production domain knowledge base. The integration process establishes a cross-reference index: using a unique process identifier to associate the process parameter constraint rule identifier, resource dependency identifier, and execution condition identifier involved in that process; using a unique equipment identifier to associate the process identifier that the equipment can execute and the related safety and compliance condition identifier; and using the quality indicator name to associate the process identifier involved in that indicator and the corresponding quality judgment rule identifier.

[0049] The knowledge base uses a graph database as its underlying storage engine. Ontology nodes and edges are mapped to vertices and edges in the graph database. Rules and conditions are attached to the corresponding ontology nodes as vertex attributes, forming a unified knowledge graph structure that can be traversed and queried.

[0050] In summary, step S1 completes the construction of a domain knowledge graph covering four dimensions: production process, process parameters, equipment capabilities, and quality standards. This knowledge graph provides domain background knowledge for semantic disambiguation in step S2, constraint rules for semantic consistency verification in step S3, and process logic and resource capability information for automatic planning of collaboration strategies in step S4.

[0051] For step S2 in the method of this embodiment, a semantic embedding layer is designed. The semantic embedding layer is used to convert three types of heterogeneous inputs—task text description, product process requirements, and resource conditions—into a unified semantic vector representation that integrates domain ontology knowledge, including the following steps.

[0052] Step S201: Establish a multi-source input processing framework for the semantic embedding layer. The semantic embedding layer receives three types of input data. The first type of input data is a task text description, including production task instructions and operation instructions in natural language. The second type of input data is product process requirements, including processing technology type identifiers, process parameter configuration data, and process quality requirement labels.

[0053] The third type of input data consists of resource conditions, including equipment resource status information, personnel resource status information, material resource status information, and tool resource status information. These three types of input data are processed through three separate feature extraction channels.

[0054] Step S202: Construct a semantic encoder for production tasks. The semantic encoder adopts a bidirectional encoder representation architecture based on a transformer. The bidirectional encoder representation architecture includes a multi-head self-attention layer, a feedforward neural network layer, a residual connection layer, and a layer normalization layer.

[0055] The input layer of the semantic encoder performs word segmentation on the task text description, converting the input text into a sequence of word identifiers. The multi-head self-attention layer captures long-distance dependencies in the text sequence, identifying semantic associations between process names, process parameters, and quality requirements scattered in different locations in the task description.

[0056] The hidden layer dimension of the semantic encoder is set to a configurable parameter, which is adjusted according to the complexity of the task description. The output layer of the semantic encoder aggregates the variable-length sequence representation into an initial feature vector of fixed dimensions through pooling operations.

[0057] Step S203: Construct the process feature extraction network. The process feature extraction network adopts a one-dimensional convolutional neural network architecture, which includes convolutional layers, batch normalization layers, activation layers, and pooling layers.

[0058] The process feature extraction network processes the three types of sub-information in the product process requirements separately. The processing technology type is represented in one-hot encoding form and mapped to a dense vector through an embedding layer. The process parameter configuration is represented in real number vector form, with each dimension of the process parameter configuration corresponding to the current set value of a process parameter.

[0059] Process quality requirements are represented by a structured tag set, where each tag corresponds to the compliance status or grade requirement of a quality indicator. The process feature extraction network extracts local correlation patterns between three types of sub-information through convolutional operations, generating a process feature vector. This vector contains key information about the processing technology type, process parameter configuration, and process quality requirements.

[0060] Unstructured task text descriptions are converted into fixed-dimensional initial feature vectors, and semi-structured product process requirements are converted into fixed-dimensional process feature vectors. Together, they solve the problem of inconsistent heterogeneous input data formats, enabling subsequent multimodal fusion to be performed in a unified vector space.

[0061] Step S204: Construct a resource feature extraction network. The resource feature extraction network adopts a graph neural network architecture. The resource feature extraction network defines the equipment, personnel, materials, and tools in the production site as nodes in the resource topology graph, and defines the supply relationship, call relationship, and cooperation relationship between resources as edges in the resource topology graph, thus forming a resource topology graph structure.

[0062] The resource feature extraction network aggregates the feature information of neighboring nodes of each resource node in the resource topology graph through the message passing mechanism of graph neural networks, generating resource node embeddings that include local context information. The resource feature extraction network integrates the embeddings of various types of resource nodes to generate resource feature vectors, which contain four key aspects of information: equipment resource status, personnel resource status, material resource status, and tool resource status.

[0063] Equipment resource status is represented by equipment identifier and current equipment status code. Personnel resource status is represented by personnel skill vector and workload index. Material resource status is represented by material inventory vector and material consumption rate. Tool resource status is represented by tool availability identifier and tool status code.

[0064] Resource condition data with graph structure characteristics are converted into fixed-dimensional resource feature vectors, and the resource topological dependencies in the resource condition data participate in subsequent attention fusion calculations in vector form.

[0065] Step S205: Perform vector fusion operation. The vector fusion operation is implemented using a multimodal attention fusion mechanism.

[0066] The initial feature vector, process feature vector, and resource feature vector are first mapped to a unified semantic space dimension through their respective linear projection layers, so that the three vectors have the same dimension and are under the same semantic metric scale.

[0067] Then, the pairwise attention weights between the three input vectors are calculated using a cross-modal attention module. Let the initial feature vector be projected to obtain the query vector, the process feature vector to obtain the key vector, and the resource feature vector to obtain the value vector.

[0068] For the attention weights of the first modality relative to the second modality, the projection vector of the first modality is used as the query and the projection vector of the second modality is used as the key. The dot product similarity of the two projection vectors is calculated, and then scaled and normalized using softmax to obtain the scalar attention weights. The scalar weights reflect the semantic relevance between the two modalities.

[0069] The cross-modal attention module calculates pairwise attention weights among the three input vectors. Taking the attention weights between the initial feature vector and the process feature vector as an example, the calculation formula is as follows:

[0070] in The query vector is obtained by linear projection of the initial feature vector. The key vector is obtained by linear projection of the process feature vector. To query the dimensions of the vector and key vector, The attention weight is the initial feature vector relative to the process feature vector. The attention weight is a scalar with a value between 0 and 1, reflecting the semantic relevance between the task text description and the process requirements.

[0071] Attention weights of the initial feature vector relative to the resource feature vector The calculation formula is:

[0072] in The key vector is obtained by linear projection of the resource feature vector. The attention weights of the initial feature vector relative to the resource feature vector reflect the degree of matching between the task text description and the current resource state.

[0073] Attention weights of process feature vectors relative to resource feature vectors The calculation formula is:

[0074] in The query vector is obtained by linear projection of the process feature vector. This represents the attention weight of the process feature vector relative to the resource feature vector, reflecting the degree of fit between process requirements and resource capabilities.

[0075] The cross-modal attention module outputs a weighted fusion representation. Three input vectors are used as value vectors, denoted as follows: , , The calculated attention weights are applied to the corresponding value vectors to generate three weighted vectors. These three weighted vectors are then summed element-wise to obtain the weighted fusion representation. The calculation formula is:

[0076] in The weighted fusion vector has the same dimensions as the weighted fusion vector. , , same. The value vector is obtained by linear projection of the initial feature vector. The value vector is obtained by linear projection of the process feature vector. This is the value vector obtained by linear projection of the resource feature vector. Each dimension of the weighted fusion vector is a weighted sum of the corresponding dimensions of the three input vectors, and each dimension is an independent scalar operation.

[0077] The fusion weights are automatically learned through a learnable attention network, dynamically adjusting the contribution of each modal input based on different task types and scenarios. Finally, a unified semantic vector representation is output through residual connections and layer normalization operations.

[0078] Heterogeneous vectors from three different feature extraction channels are fused into a unified semantic vector, which simultaneously encodes task semantic information, process requirement information, and resource status information.

[0079] Step S206: Perform a knowledge enhancement process based on the domain ontology library. The knowledge enhancement process is implemented using knowledge graph embedding technology.

[0080] Entities and relations in the domain ontology are first vectorized and encoded using a knowledge representation learning model, generating entity embedding matrices and relation embedding matrices. Each row of the entity embedding matrix corresponds to the embedding vector of the entity, and each row of the relation embedding matrix corresponds to the embedding vector of the relation.

[0081] The knowledge enhancement process extracts key entity references from a unified semantic vector. Let the unified semantic vector be... The unified semantic vector dimension is The entity embedding matrix is ,in For the total number of entities, Let be the dimension of the entity embedding vector, which is consistent with the dimension of the unified semantic vector. For the _th ... Entity embedding vectors , No. Individual entity embedding vectors and unified semantic vectors Similarity score Calculated as the first The dot product of an entity embedding vector and a unified semantic vector, i.e. All The similarity scores of each entity form a score vector. Applying a softmax operation to the score vector yields the attention weights for each entity. The calculation formula is:

[0082] in For the first The attention weight of each entity is a real number between 0 and 1, and the sum of the attention weights of all entities is 1. For the first The dot product similarity score between an entity embedding vector and a unified semantic vector. For the first The dot product similarity score between an entity embedding vector and a unified semantic vector. For the total number of entities, It is a natural exponential function.

[0083] Pay attention weights to each entity With the corresponding entity embedding vector We perform a weighted summation to obtain the knowledge enhancement vector. The calculation formula is:

[0084] in For knowledge augmentation vectors, the dimension of the knowledge augmentation vector is... Same, for . For the first Attention weights for each entity For the first Embedding vectors of entities, This represents the total number of entities. The summation operation is a scalar operation independent of each dimension. The values ​​of each entity embedding vector in the same dimension are weighted and summed to obtain the corresponding value of the knowledge enhancement vector in the same dimension.

[0085] Knowledge Augmentation Vector With Unified Semantic Vector Perform a fusion update. Unify the semantic vector. With knowledge augmentation vectors By adding elements one by one, we obtain the enhanced semantic vector. The calculation formula is:

[0086] in The enhanced semantic vector has a dimension of . . To unify semantic vectors, This is a knowledge augmentation vector. Element-wise addition is a scalar operation independent of each dimension; for the ... One dimension, ,in To enhance the semantic vector of the first generation Values ​​of each dimension To unify semantic vectors Values ​​of each dimension For knowledge augmentation vectors Values ​​of each dimension The value range is 1 to The enhanced semantic vector simultaneously encodes the original semantics of the task and the domain knowledge context.

[0087] Step S2 completes the transformation from three types of heterogeneous inputs to a unified semantic space. The task text description is converted into an initial feature vector by a semantic encoder, the product process requirements are converted into a process feature vector by a process feature extraction network, and the resource conditions are converted into a resource feature vector by a resource feature extraction network.

[0088] The three feature vectors are fused into a unified semantic vector through a multimodal attention fusion mechanism, and then domain ontology knowledge is injected into the unified semantic vector through knowledge graph embedding technology. The unified semantic vector integrates information from four dimensions: task semantics, process constraints, resource status, and domain knowledge, providing a vectorized input with rich contextual information for the semantic enhancement parsing in the subsequent step S3.

[0089] In step S3, the unified semantic vector is input into the semantic enhancement parsing engine. The semantic enhancement parsing engine performs semantic expansion and constraint injection on the task vector based on the domain ontology library, integrates process rules for semantic consistency verification, and outputs parsing results containing task objectives, constraints, resource requirements, and execution sequences, implemented according to the following steps.

[0090] Step S301: Establish the overall architecture of the semantic enhancement parsing engine. The semantic enhancement parsing engine adopts an encoder-decoder framework, which includes four functional components: a semantic understanding encoder, a constraint injection module, a consistency verification module, and a decoding output module.

[0091] The semantic understanding encoder performs deep semantic encoding on the input unified semantic vector based on a transformer architecture. The constraint injection module injects process rules and domain ontology knowledge into the semantic understanding result based on knowledge graph reasoning technology. The consistency verification module checks the consistency between the elements in the semantic representation and identifies and handles conflict information. The decoding output module converts the structured semantic representation into four types of structured output: task objectives, constraints, resource requirements, and execution sequences.

[0092] Step S302: Perform a semantic extension process based on the domain ontology library. The semantic extension process employs a dual mechanism of semantic disambiguation and semantic completion.

[0093] The semantic disambiguation mechanism performs context-aware semantic inference on polysemous words in task descriptions, determining the precise meaning of polysemous words in the current context through the concept hierarchy in the domain ontology. The same process name may correspond to different process parameter configurations in different production lines or equipment contexts. The semantic disambiguation mechanism determines the meaning of the process in the current task context by querying the equipment type and quality standards associated with that process in the domain ontology.

[0094] The semantic completion mechanism intelligently infers the missing implicit information in the task description, making implicit entities, implicit relationships, and implicit constraints explicit. The inference of implicit entities is based on the co-occurrence statistics of concepts in the domain ontology library and semantic similarity calculation.

[0095] The semantic similarity is calculated as follows: for a uniform semantic vector and the first in the domain ontology library Embedding vectors corresponding to each conceptual entity The semantic similarity between the two is calculated using the dot product operation, that is:

[0096] in To unify semantic vectors, the dimension is... ; For the first in the domain ontology library The embedding vector of each conceptual entity, with dimensions and Consistency, for ; The result of the dot product operation is a scalar; a higher value indicates a stronger semantic association between the unified semantic vector and the concept entity. Dot product similarity is the standard operation for measuring vector similarity in linear algebra.

[0097] When a task description mentions a process name but does not specify its preceding processes, the missing preceding process information is supplemented using process sequence constraint rules in the ontology library. Implicit relationship inference is implemented based on relational pattern matching and causal reasoning in the domain ontology library. When the task description mentions adjusting a certain process parameter, the parameter coupling relationships in the ontology library are used to infer whether other parameters need to be adjusted in conjunction.

[0098] The inference of implicit constraints is achieved through rule reasoning and constraint propagation based on the process rule base. When a certain quality requirement is implicit in the task description, the corresponding detection method and judgment standard are completed by the quality requirement constraint rules.

[0099] Step S303: Perform the process rule injection process. Process rule injection converts process constraint rules into constraint vector representations in a semantic vector space, and then integrates the constraint vectors into the semantic vectors. Rule injection employs a dual mechanism of constraint encoding and semantic fusion.

[0100] The constraint encoding mechanism achieves the mapping transformation from rule description to rule vector through a rule encoder. The rule encoder adopts a graph neural network-based architecture, taking the structured constraint rules constructed in step S1 as input, extracting attribute information such as source process identifier, target process identifier, constraint type, and enforcement degree from the rules, and aggregating the association patterns between rules through the message passing mechanism of the graph neural network to generate constraint vectors as output.

[0101] The semantic fusion mechanism incorporates the constraint vector into the expanded semantic vector output from step S302 through a constraint-aware attention module. The constraint fusion weights are calculated as follows: for the first... Constraints, their degree of enforceability coefficient It is determined by the degree of enforceability of the constraint.

[0102] Hard constraints The value is 1.0, representing a soft constraint. The value is a parameter that is dynamically adjusted based on the task context. , Calculated based on the current task type, resource availability, and historical execution data, the value ranges from 0 to 1. (Constraint coverage coefficient) This represents the proportion of the number of processes involved in this constraint to the total number of processes in the current task. Fusion weights of constraints The calculation formula is:

[0103] in For the first The semantic fusion weights of the constraints are dimensionless scalars, with values ​​ranging from 0 to 1; For the first The coercivity coefficient of the constraint is set to 1.0 for hard constraints and 1.0 for soft constraints. ; For the first The coverage coefficient of a constraint is the ratio of the number of processes involved in the constraint to the total number of processes. When the value is greater than or equal to 0.8, the constraint dominates the semantic fusion process, and the weight of the constraint vector on the expanded semantic vector is set to a higher value. When the value is less than 0.3, the constraint is incorporated into the semantic vector as secondary reference information.

[0104] The fusion weights for hard constraints are fixed at a relatively high value to ensure that the hard constraints are strictly enforced in subsequent policy generation. The fusion weights for soft constraints are dynamically adjusted according to the task context and are considered as optimization objectives.

[0105] Step S304: Perform semantic consistency verification. Semantic consistency verification includes three stages: syntactic consistency verification, domain consistency verification, and process consistency verification.

[0106] The syntactic consistency check phase examines whether entity references in semantic vectors are consistent, whether relation descriptions are contradictory, and whether temporal information conflicts. The domain consistency check phase examines whether entity types conform to domain ontology constraints, whether attribute values ​​are within allowed ranges, and whether relationships between entities conform to domain knowledge. The process consistency check phase examines whether process parameter configurations meet the boundary conditions of process constraint rules, whether the process sequence meets process sequence constraint rules, and whether resource requirements are within resource availability limits.

[0107] During the consistency verification process, the location and type of conflicting elements are identified through the constraint propagation algorithm. First, all conflicting elements discovered during the verification are extracted and categorized into syntactic conflicts, domain conflicts, and process conflicts according to the type of constraints involved. Conflicts arising from hard constraints are processed first.

[0108] Soft conflicts are handled using a constraint relaxation algorithm for fault tolerance, appropriately relaxing soft constraints while maintaining hard constraints, with the relaxation not exceeding 10% of the constraint boundary value. Hard conflicts are handled using a constraint backtracking algorithm for root cause tracing and correction, locating the root constraint or semantic parsing error causing the conflict by analyzing the conflict dependency chain. If a semantic parsing error is located, the process returns to step S302 to re-execute semantic completion. If an irreconcilable contradiction is found between the root constraint and the current task description, the process returns to step S303 to re-execute rule injection, reducing the fusion weight of that constraint.

[0109] Step S305: Generate structured parsing results. The decoding output module converts the enhanced semantic representation, after semantic expansion, rule injection, and consistency verification, into four types of structured output.

[0110] The parsed results of the task objectives are represented as structured objects, including attribute fields such as a unique task identifier, task name, task type, task priority, task output specifications, and task quality requirements. The parsed results of the constraints are represented as constraint sets, with each constraint containing attribute fields such as constraint type, constraint expression, constraint strength, and violation penalty. The constraint types cover four dimensions: time constraints, resource constraints, quality constraints, and security constraints.

[0111] The parsed results of resource requirements are represented as a resource list, with each resource requirement including attribute fields such as resource type, resource specifications, resource quantity, and resource time window. The parsed results of the execution sequence are represented as a directed acyclic graph, where nodes represent execution steps, edges represent execution order and dependencies, and each edge is labeled with dependency type and synchronization requirements. Dependency types are divided into three categories: sequential dependency, parallel dependency, and conditional dependency.

[0112] Step S3 completes the transformation from the unified semantic vector to the structured task parsing result. The semantic enhancement parsing engine performs deep encoding on the input unified semantic vector through a semantic understanding encoder, supplements the missing implicit information in the task description through semantic disambiguation and semantic completion, injects process rules into the semantic vector through constraint encoding and semantic fusion, and ensures the accuracy of the parsing result through a three-stage consistency check. Finally, the decoding output module generates four types of structured parsing results: task objective, constraints, resource requirements, and execution sequence. This parsing result provides task structure information and constraint boundary conditions for the subsequent multi-level collaboration strategy generation in step S4.

[0113] In step S4, a multi-level collaboration strategy is automatically generated based on the parsing results output in step S3. The multi-level collaboration strategy includes three levels: strategic task allocation, tactical resource scheduling, and execution order. The strategy levels are automatically decoupled and coordinated, and implemented according to the following steps.

[0114] Step S401: Establish the overall framework for generating multi-level collaboration strategies. Step S4: Receive the task objective parsing results, constraint parsing results, resource requirement parsing results, and execution sequence parsing results output from Step S3 as input.

[0115] The multi-level collaboration strategy generation process is conducted in three layers: strategic, tactical, and execution. Each layer corresponds to a different decision-making timescale and optimization objective. The strategic layer is responsible for breaking down complex tasks into independently executable sub-task units and assigning them to responsible entities. The tactical layer is responsible for allocating appropriate production resources to each sub-task and determining the timing of resource usage. The execution layer is responsible for determining the execution order and synchronization control methods for the internal steps of each sub-task.

[0116] Step S402: Generate the strategic-level task allocation strategy. The generation of the strategic-level task allocation strategy uses the task objective analysis results output in step S3 as the core input, and is achieved collaboratively through the task decomposition module and the responsibility subject matching module.

[0117] The task decomposition module hierarchically breaks down complex tasks according to process flow and functional modules, generating a multi-level task tree structure containing root tasks, subtasks, and atomic tasks. The root task corresponds to the original production task request submitted by the user, the subtasks correspond to the functional decomposition units of the original task, and the atomic task corresponds to the smallest indivisible execution unit.

[0118] During the task decomposition process, the task decomposition module queries the process sequence constraint rules in the domain ontology library constructed in step S1, determines the splitting boundary according to the serial dependency relationship and parallel equivalence relationship between processes, and ensures that the split task structure conforms to the production process logic.

[0119] The responsible entity matching module matches the most suitable responsible entity from the domain ontology library based on the attribute characteristics of the atomic task. Attribute characteristics include process type, required equipment type, required personnel skill level, and quality requirement level. The responsible entity matching process comprehensively considers three factors: the responsible entity's capability matching degree, load balancing degree, and collaboration history evaluation.

[0120] The overall score is based on the matching between the a-th atomic task and the b-th responsible entity. To optimize the objective, The calculation method is a weighted sum of three factors:

[0121] in This is the atomic task index, with values ​​ranging from 1 to... , Total number of atomic tasks; For the index of the responsible entity, the value ranges from 1 to... , The total number of responsible entities; For the first The atomic task is assigned to the first The overall matching score for each responsible entity is a dimensionless scalar. For the first The atomic task and the first The capability matching degree of each responsible entity is a dimensionless scalar with a value range of 0 to 1, and is determined by the matching relationship between process type and equipment type in the domain ontology library. For the first The load balancing factor of each responsible entity is a dimensionless scalar with a value range of 0 to 1. It is calculated based on the proportion of the currently assigned tasks of the responsible entity to the total capacity. The lower the proportion, the higher the load balancing factor. For the first The responsible entity and the first The historical collaboration score of each atomic task is a dimensionless scalar with a value range of 0 to 1, calculated from the collaboration success rate in historical execution data. , , These are the weighting coefficients for the three factors, all of which are dimensionless scalars, ranging from 0 to 1, and their sum is 1. They are dynamically adjusted based on the current production scenario.

[0122] Overall Matching Score A higher value indicates that the responsible entity is more suitable to perform atomic tasks.

[0123] For each atomic task, compare the overall matching scores of all responsible parties. The responsible entity corresponding to the highest score is selected as the assignment object for the atomic task, forming a one-to-one correspondence between each atomic task and the responsible entity. If multiple atomic tasks are matched with the same responsible entity, the priority is determined according to the relative high and low of the overall matching score, ensuring that the atomic task with the higher score is assigned a responsible entity first, and the atomic task with the lower score is re-matched with the next best responsible entity.

[0124] Step S403: Generate tactical-level resource scheduling strategy. The generation of the tactical-level resource scheduling strategy uses the resource demand analysis results and constraint analysis results output in step S3 as core inputs, and is achieved collaboratively by the resource planning module and the scheduling optimization module.

[0125] The resource planning module calculates the total and peak demand for each type of production resource based on the resource demand list. The module then matches and verifies the calculated resource demand against the resource availability information in the domain ontology library constructed in step S1 to determine whether available resources meet the total demand.

[0126] Resource planning employs an inventory buffer mechanism to reserve safety stock levels for critical resources to cope with demand fluctuations and abnormal consumption. Safety stock levels The calculation method is as follows:

[0127] in This is a safety stock level, with dimensions consistent with the quantity of resources. The preset safety factor is a dimensionless scalar value set according to production reliability requirements, with a value range of 1.0 to 3.0. It represents the standard deviation of historical demand fluctuations, with dimensions consistent with resource quantity, and is derived from statistical data on resource consumption over several past production cycles. The lead time for replenishment is measured in units of time and is determined by the material procurement or resource allocation process.

[0128] The scheduling optimization module determines the resource allocation scheme and scheduling sequence based on resource constraints and task timing constraints. The module aims to minimize task completion time, using resource capacity constraints, process sequence constraints, and quality stability constraints as conditions, and establishes a mixed-integer programming model for solution. The objective function for minimizing task completion time is:

[0129] in The total completion time for all tasks, measured in units of time. Total number of tasks; For the first The completion time of each task is measured in units of time. Resource capacity constraints require that the total amount of resources allocated within each time period does not exceed the available upper limit of the resource type. Process sequence constraints require that tasks with process sequence dependencies use resources in the order they occur. Quality stability constraints require that the resource status remain consistent when the same process is executed consecutively.

[0130] Solving the mixed-integer programming model yields the optimal allocation scheme and scheduling sequence for each resource in each time period. The tactical-level resource scheduling strategy is represented in the form of a scheduling plan table, which includes four types of time-dimensional information: resource allocation period, resource occupation period, resource release period, and resource reservation period.

[0131] Step S404: Generate the execution layer execution order strategy. The generation of the execution layer execution order strategy uses the execution sequence parsing result output in step S3 as the core input, and is achieved through the collaborative efforts of the sequence orchestration module and the synchronization control module.

[0132] The sequence orchestration module generates a linearized execution order sequence based on the directed acyclic graph (DAG) representation in the execution sequence parsing result. Nodes in the DAG represent execution steps, and edges represent dependencies between steps. The linearization process employs a topological sorting algorithm, transforming the graph structure into a sequence structure according to the order of dependencies. During topological sorting, the sequence orchestration module simultaneously evaluates the parallelism opportunities of each node, marking nodes without dependencies as parallelizable nodes, thus generating an execution order sequence that balances execution efficiency and resource utilization.

[0133] The synchronization control module establishes a synchronization mechanism for parallel task nodes, defining synchronization points, synchronization conditions, and synchronization timeout handling strategies. Synchronization points are set at the merging points of parallel branches. Synchronization conditions are defined as the merging step only starting after all parallel branches have completed. The synchronization timeout handling strategy specifies the waiting time limit and alternative execution paths after a timeout. The execution layer execution order strategy is represented in the form of an execution flowchart. Flowchart nodes represent execution steps, edges represent execution order and synchronization relationships, and node attributes include step name, step type, step execution duration, and step input / output definitions.

[0134] Step S405: Decoupling the strategy layers. The three layers—strategy, tactics, and execution—are decoupled using a loosely coupled architecture. Each layer maintains its own independent strategy state storage and strategy execution engine, and information is exchanged between layers through standardized interfaces.

[0135] After the strategy layer outputs task allocation decisions, the tactical layer receives these decisions and independently makes resource scheduling decisions without waiting for further confirmation from the strategy layer. Similarly, after the tactical layer outputs resource scheduling decisions, the execution layer receives these decisions and independently makes execution order decisions without waiting for further confirmation from the tactical layer. Each layer's strategy evolves independently based on changes in the production environment; adjustments to a single layer's strategy do not trigger coordinated adjustments across the entire system. This loosely coupled architecture supports incremental policy updates; when a strategy at one layer changes, the execution of strategies at other layers remains unaffected.

[0136] Step S406: Implement the coordination mechanism between policy layers. The coordination mechanism employs a dual mechanism of constraint propagation and conflict arbitration to prevent policy inconsistencies caused by decoupling between layers.

[0137] The constraint propagation mechanism passes key constraints from upper-level strategies to lower-level strategies, ensuring that the results generated by lower-level strategies meet the constraint requirements of upper-level strategies. During constraint propagation, task boundary constraints generated at the strategy layer are passed to the tactical layer as range constraints for resource scheduling, and resource capacity constraints generated at the tactical layer are passed to the execution layer as upper limit constraints for step parallelism.

[0138] The conflict arbitration mechanism is triggered when inter-layer strategies conflict. The conflict detection module monitors the consistency of inter-layer strategies in real time, identifying the type and severity of conflicts. Conflict types include three categories: constraint violation conflicts, resource contention conflicts, and timing conflicts. Constraint violation conflicts occur when a lower-level strategy violates constraints passed down from the upper layer; resource contention conflicts occur when multiple tasks compete for the same type of scarce resource at the same time; and timing conflicts occur when the synchronization conditions for parallel execution steps cannot be met.

[0139] The conflict arbitration process uses an arbitration priority formula to determine the order in which conflicts are processed. Arbitration Priority The calculation method is as follows:

[0140] in This represents the arbitration priority for conflicts. It is a dimensionless scalar, and a higher value indicates that the conflict is processed with higher priority. The weight coefficients at the strategy level are dimensionless scalars. The value is 3, the tactical layer is 2, and the execution layer is 1. The constraints of the higher-level strategy take precedence over the constraints of the lower-level strategy. , where is the constraint enforcement coefficient, a dimensionless scalar; a value of 1.0 is used for hard constraints, and a value of 0.5 is used for soft constraints. By comparing the various conflicts... The value determines the arbitration order. Arbitration will be conducted with priority given to the party with the highest value. Arbitration decisions will prioritize high-level strategic constraints and constraints with a high degree of enforcement, followed by low-level strategic constraints and constraints with a low degree of enforcement.

[0141] Step S4 completes the transformation from task parsing results to executable collaboration strategies. The generated multi-level collaboration strategies cover the entire chain from task allocation to resource scheduling and execution ordering. The three strategy levels evolve independently while maintaining global consistency through constraint propagation and conflict arbitration. The multi-level collaboration strategies output by step S4 provide a set of candidate solutions and decision-making basis for the strategy selection network in the subsequent step S5.

[0142] For step S5, a policy selection network is constructed. The policy selection network takes the unified semantic vector output from step S2 and the current system state as input, outputs the optimal cooperation scheme, and continuously optimizes the decision model using a reinforcement learning mechanism, as follows.

[0143] Step S501: Establish the overall model architecture of the policy selection network. The policy selection network adopts an encoder-decoder architecture, which includes five functional modules: state encoder, semantic encoder, fusion layer, decoder, and scoring output layer.

[0144] Step S502: Encode the current system state using a state encoder. The state encoder employs a graph neural network architecture, modeling the current system state as a resource node graph. Nodes in the resource node graph represent production resources, and edges represent the relationships between resources. The state encoder aggregates the feature information of resource nodes and their neighboring nodes through the message passing mechanism of the graph neural network to generate a state feature vector.

[0145] The input to the status encoder includes three types of system status data: resource load information, task queue information, and device status information. Resource load information is input to the status encoder in the form of resource node attributes, task queue information is input in the form of task node attributes, and device status information is input in the form of device node attributes and device topology relationships.

[0146] Step S503: The unified semantic vector is encoded using a semantic encoder. The semantic encoder employs a multilayer perceptron architecture, taking the unified semantic vector output from step S2 as input, extracting semantic features through multilayer nonlinear transformations, and outputting a semantic feature vector.

[0147] System state data and semantic data are converted into state feature vectors and semantic feature vectors respectively through their respective independent encoding channels.

[0148] Step S504: The state feature vector and semantic feature vector are fused through a fusion layer. The fusion layer employs an attention weighting mechanism, using the semantic feature vector as the query and the state feature vector as the key and value, to calculate the attention weight distribution of the semantic features relative to the state features, thereby generating a joint feature vector.

[0149] Step S505: Candidate cooperation schemes are generated through the decoder. The decoder employs a transformer-decoder architecture, using the joint feature vector output in step S504 as the initial hidden state. The conditional probability distributions of the candidate cooperation schemes are gradually generated through the multi-head self-attention layer and cross-attention layer in the transformer-decoder. At each time step, the decoder outputs the action probability distribution for the current step, and the entire sequence constitutes a complete candidate cooperation scheme.

[0150] Step S506: Output a comprehensive score for each candidate collaboration scheme through the scoring output layer. The scoring output layer adopts a fully connected network architecture, mapping the decoder hidden state vector corresponding to each candidate collaboration scheme generated in step S505 to a scalar score value. The dimension of the score vector is the same as the number of candidate collaboration schemes, and each dimension corresponds to the comprehensive score of a candidate collaboration scheme.

[0151] The comprehensive scoring process considers three evaluation dimensions: task completion efficiency, resource utilization efficiency, and quality compliance rate. Task completion efficiency is evaluated using task completion time, task delay rate, and task throughput as indicators. Resource utilization efficiency is evaluated using equipment utilization rate, personnel hour utilization rate, and material turnover rate. Quality compliance rate is evaluated using first-pass yield rate, rework rate, and scrap rate.

[0152] For the c-th candidate collaboration scheme, the comprehensive score is... The calculation method is a weighted sum of the normalized evaluation indicators:

[0153] in For candidate collaboration scheme indexes, values ​​range from 1 to... , The total number of candidate collaboration solutions; To evaluate the dimensional index, This represents the dimension of task completion efficiency. This represents the dimension of resource utilization efficiency. This represents the quality compliance rate dimension; This provides an index of evaluation metrics for each dimension, with each dimension containing three evaluation metrics. The value can be between 1 and 3; For the first The weight coefficients of each evaluation dimension are dimensionless scalars, which are dynamically adjusted according to the priority of the current production scenario. The sum of the weight coefficients of the three dimensions is 1. For the first The first evaluation dimension The weight coefficients of each evaluation indicator are dimensionless scalars, and the sum of the weight coefficients of the three indicators within each dimension is 1. For the first The candidate solution is in the... The first dimension The original measured values ​​of each indicator; The normalization function maps the original measurements to the interval between 0 and 1 using linear interpolation, with the historical minimum and maximum values ​​of the corresponding indicator as boundaries. (Comprehensive score) It is a dimensionless scalar; the higher the value, the better the overall performance of the candidate collaboration scheme.

[0154] Weighting coefficients of each evaluation indicator and The algorithm is learned offline and dynamically adjusted based on task type and business scenario. The scoring output layer outputs the candidate collaboration solution with the highest comprehensive score as the optimal collaboration solution in the current system state. The optimal collaboration solution is determined by taking the maximum comprehensive score.

[0155] The strategy selection network generates a set of candidate cooperation schemes based on the fused joint feature vectors, and then selects the candidate cooperation scheme with the highest comprehensive score from them through a multi-objective weighted scoring mechanism.

[0156] Step S507: Train the policy selection network using a reinforcement learning mechanism. The reinforcement learning mechanism employs an actor-critic framework. The actor network, or policy selection network, is responsible for generating cooperative decision-making based on the current input. The critic network is responsible for evaluating the long-term value of the current decision and outputting a state-action value function estimate.

[0157] Training data comes from historical production execution records. Each training sample includes current state features, selected action, immediate reward obtained after execution, and state features at the next moment. During training, the system performance index after the collaborative scheme is executed is used as the reward signal, and the gradient is calculated and the actor network parameters are updated using the policy gradient algorithm.

[0158] The loss function of the critic network is the mean squared error of the temporal difference error:

[0159] in Let be the loss value of the commentator network, which is a dimensionless scalar. The training batch size is the number of samples taken from the experience pool for each training iteration. This is the sample index within the batch, with a value ranging from 1 to... ; For the first The instantaneous reward value obtained after each sample performs an action has the same dimensions as the reward signal design. This is a discount factor, a dimensionless scalar with a value ranging from 0 to 1, used to balance the importance of current rewards and future rewards; For the critics' network on the next state Value estimation, dimensions and Consistent; For the network of critics on the current state Value estimation, dimensions and Consistent. The critic network is trained by minimizing a loss function, allowing the value estimate to gradually approach the true return.

[0160] The actor network is updated using a policy gradient algorithm, with the gradient direction increasing the expected cumulative reward. The update amount is proportional to the advantage function value output by the critic network. The advantage function value is the difference between the sum of the immediate reward and the discounted future value, minus the estimated value of the current state. A positive advantage function strengthens the probability of choosing the corresponding action, while a negative advantage function reduces the probability of choosing the corresponding action.

[0161] The reward function design comprehensively considers four types of rewards: task completion reward, resource efficiency reward, quality reward, and collaboration reward. The task completion reward provides a positive reward for successful task completion and a negative reward for task failure. The resource efficiency reward is based on the resource utilization rate. The quality reward is based on the achievement of quality standards. The collaboration reward is based on the effectiveness of cooperation between agents.

[0162] Step S508: Perform the continuous optimization process of the decision model. The continuous optimization process is implemented using an experience replay mechanism and a target network mechanism.

[0163] The experience replay mechanism stores historical decision-making experience in an experience pool. During training, it randomly samples experience samples from the pool for learning, breaking the temporal correlation between samples. The target network mechanism maintains an independent set of network parameters as the training objective, periodically copying parameters from the main network to update the target network, thus improving training stability.

[0164] The decision model supports two optimization modes: online learning and offline learning. In online learning mode, model parameters are continuously updated based on real-time execution feedback. In offline learning mode, batches of execution data are collected, and model parameters are updated periodically through batch training.

[0165] Step S5 completes the construction of the policy selection network and the decision-making of cooperative policies based on real-time state. The state encoder and semantic encoder encode the system state data and unified semantic vector into state feature vectors and semantic feature vectors, respectively. The fusion layer generates a joint feature vector, the decoder generates a set of candidate cooperative schemes, and the scoring output layer selects the optimal scheme from them through multi-objective weighted scoring. The reinforcement learning mechanism continuously optimizes the model using historical execution data. The optimal cooperative scheme output in step S5 serves as the decision-making basis for runtime execution and monitoring in the subsequent step S6.

[0166] Finally, in step S6, deploy the runtime monitoring module. The runtime monitoring module collects execution metrics in real time, triggers adjustments to the collaboration mode based on feedback signals, and supports online policy updates and dynamic refactoring. This is implemented according to the following steps.

[0167] Step S601: Establish the overall framework of the runtime monitoring module. The runtime monitoring module distributes the optimal collaboration scheme output in step S5 to each execution unit for execution, and continuously collects execution metrics during the execution process.

[0168] Step S602: Deploy the task progress monitoring unit. The task progress monitoring unit interfaces with the task execution system to retrieve real-time execution status data for each task. The collected data items include the task start timestamp, current task status, task progress percentage, and reason for task blocking. The collected raw data is cleaned, transformed, and aggregated to generate standardized task progress metrics, which are stored in a time-series database to support time-range queries.

[0169] The task progress monitoring unit compares the actual task progress percentage with the planned progress percentage. An alarm is triggered when the absolute value of the deviation between the actual and planned progress percentages exceeds a preset deviation threshold. The deviation threshold is set according to the task type and ranges from 5% to 20%.

[0170] Step S603: Deploy the resource utilization monitoring unit. The resource utilization monitoring unit adopts a distributed data acquisition architecture. Monitoring agents for each resource type are deployed locally on the resource, collecting data such as equipment uptime, equipment load rate, effective working hours of personnel, material consumption rate, and tool usage frequency, and reporting them to the central monitoring service.

[0171] Real-time utilization rate of equipment resources The calculation method is the proportion of the actual operating time of the equipment to the total time within the statistical period:

[0172] in This represents the real-time utilization rate of equipment resources, a dimensionless scalar with a value range of 0 to 1. The actual operating time of the equipment within the statistical period is expressed in units of time. The total time of the statistical period is expressed in units of time.

[0173] Real-time utilization rate of human resources The calculation method is the proportion of an employee's effective working hours to the total working hours during the statistical period:

[0174] in The real-time utilization rate of personnel resources is a dimensionless scalar with a value range of 0 to 1. The effective working hours of personnel within the statistical period are measured in units of time. The total working hours of personnel during the statistical period are measured in units of time.

[0175] The average utilization rate of each type of resource over a statistical period is the arithmetic mean of the real-time utilization rate at each moment within that period. It supports outputting analysis results by multiple time dimensions such as hour, shift, day, and week.

[0176] Step S604: Deploy the abnormal event monitoring unit. The abnormal event monitoring unit receives abnormal event reports in real time through interfaces with various monitoring systems. The abnormal event types include four categories: equipment failure events, material shortage events, quality abnormality events, and personnel absence events.

[0177] The abnormal event monitoring unit classifies, correlates, and assesses the impact of collected events. The event classification process assigns events to corresponding categories based on event type tags. The event correlation process performs correlation analysis on related abnormal events to identify causal chains and propagation paths. The event impact assessment process analyzes the potential impact of abnormal events on task progress, resource utilization, and quality compliance. The event alerting layer generates alert information based on the assessment results and pushes it to relevant responsible personnel.

[0178] Step S605: Calculate system performance indicators. System performance indicators include four items: task on-time completion rate, overall resource utilization rate, equipment failure rate, and quality pass rate.

[0179] On-time completion rate of tasks The calculation method is as follows:

[0180] in The on-time completion rate of the task is a dimensionless scalar with a value range of 0 to 1. The number of tasks completed on time within the statistical period is a dimensionless integer. This represents the total number of tasks planned to be completed within the statistical period, and is a dimensionless integer.

[0181] Comprehensive resource utilization rate The calculation uses a weighted average of the utilization rates of different types of resources:

[0182] in The comprehensive resource utilization rate is a dimensionless scalar with a value range of 0 to 1. Total number of resource types; Index for resource types; For the first The weight coefficient of a resource type is a dimensionless scalar, and the sum of the weight coefficients of all resource types is 1. For the first The average utilization rate of a type of resource within a statistical period is a dimensionless scalar.

[0183] Equipment failure rate The calculation method is as follows:

[0184] in The equipment failure rate is a dimensionless scalar with a value range of 0 to 1. The total downtime of equipment due to malfunctions within the statistical period is measured in units of time. This represents the total planned operating time of the equipment within the statistical period, measured in units of time.

[0185] Quality pass rate The calculation method is as follows:

[0186] in The quality pass rate is a dimensionless scalar quantity with a value range from 0 to 1. This represents the number of batches that passed inspection within the statistical period, and is a dimensionless integer. This represents the total number of batches inspected within the statistical period, and is a dimensionless integer.

[0187] The system performance metrics support two modes: real-time calculation and historical statistical calculation. The real-time calculation mode calculates the instantaneous metric value based on the data at the current moment, while the historical statistical calculation mode calculates the statistical metric value based on historical data for a specified time period.

[0188] Step S606: Execute the triggering process for collaboration mode adjustment. The collaboration mode adjustment adopts a dual triggering mechanism of threshold monitoring and trend prediction.

[0189] The threshold monitoring method continuously monitors performance metrics and triggers adjustments when a performance metric falls below a preset threshold. Threshold settings support both static and dynamic threshold modes. Static thresholds are manually set and used permanently. Dynamic thresholds... Automatically calculated based on the mean and standard deviation of historical statistical data:

[0190] in The threshold is dynamic, and its dimensions are consistent with the corresponding performance index. It is the average of the corresponding performance index within the historical statistical period, and its dimensions are consistent with those of the corresponding performance index. The adjustment coefficient is a dimensionless scalar with a value range of 1.0 to 3.0. This represents the standard deviation of the corresponding performance index within the historical statistical period, with dimensions consistent with the corresponding performance index. When the calculated result is negative, the dynamic threshold is set to 0.

[0191] The trend prediction method is based on time series data of performance indicators. It uses a time series prediction model to predict the performance indicator values ​​at future moments and triggers adjustments in advance when the predicted value is lower than the performance threshold.

[0192] After an adjustment is triggered, it proceeds through three phases in sequence: adjustment decision generation, adjustment plan verification, and adjustment execution. In the adjustment decision generation phase, an adjustment strategy is matched from the strategy adjustment knowledge base based on the type and scope of the abnormal event. In the adjustment plan verification phase, the adjustment plan is simulated in a sandbox environment to assess its feasibility and expected effects. In the adjustment execution phase, the verified adjustment plan is distributed to each execution unit, triggering real-time strategy updates.

[0193] Step S607: Perform an online policy update. Online policy updates support both incremental and full update modes.

[0194] Incremental update mode only updates the changed policy components, retaining the state and execution progress of the unchanged policy components. The incremental update process first identifies the changed policy elements, calculates the impact of the policy element changes on the system state, generates incremental update instructions, and issues them for execution.

[0195] The full update mode is suitable for scenarios where the system architecture has undergone significant changes or where incremental updates cannot solve the problem. The full update process first pauses the execution of the current policy, saves a snapshot of the current system state, regenerates the complete policy based on the current state, restores the system state, and starts the execution of the new policy.

[0196] Step S608: Perform dynamic refactoring. Dynamic refactoring automatically adjusts the topology of the collaboration strategy based on the type and scope of the abnormal event.

[0197] Resource unavailability events trigger resource replacement refactoring, reallocating suitable alternative resources to affected tasks. Task priority change events trigger task scheduling refactoring, adjusting task execution order and resource allocation schemes. Process flow change events trigger process adaptation refactoring, adjusting task decomposition structure and execution order. System capacity change events trigger capacity adaptation refactoring, adjusting resource scheduling and task allocation strategies. Dynamic refactoring employs hot-swapping technology to dynamically adjust the strategy structure without affecting currently executing tasks.

[0198] Step S6 completes the deployment of the runtime monitoring module and the adaptive adjustment of the collaboration strategy. The task progress monitoring unit and resource utilization monitoring unit collect execution indicators in real time, the abnormal event monitoring unit identifies and evaluates various abnormalities, the system performance indicators quantify the system operating status, the collaboration mode adjustment trigger mechanism initiates adjustments when performance deviates, the online strategy updates and corrects the strategy content, and the adjustment strategy structure is dynamically reconstructed.

[0199] This invention also provides an intelligent agent collaboration system for artificial intelligence systems in the production field. The intelligent agent collaboration system comprises six components: an ontology library construction module, a semantic embedding module, a semantic parsing module, a policy generation module, a policy decision-making module, and an operation monitoring module.

[0200] The ontology building module connects to the semantic embedding module. The ontology building module uses a graph database as the underlying storage engine to execute the production domain ontology and process rule base building functions in step S1, and provides data access interfaces for domain knowledge for subsequent modules.

[0201] The semantic embedding module connects the ontology library construction module and the semantic parsing module. The semantic embedding module includes a production-task-oriented semantic encoder, a process feature extraction network, a resource feature extraction network, and a cross-modal attention fusion layer. It is used to perform the multimodal semantic embedding function in step S2, converting heterogeneous inputs into a unified semantic vector.

[0202] The semantic parsing module connects the semantic embedding module and the policy generation module. The semantic parsing module includes a semantic understanding encoder, a constraint injection module, a consistency verification module, and a decoding output module. It is used to execute the semantic enhancement parsing function in step S3, converting the unified semantic vector into a structured parsing result.

[0203] The strategy generation module connects the semantic parsing module and the strategy decision-making module. The strategy generation module includes a task decomposition module, a resource planning module, and a sequence orchestration module. It is used to execute the multi-level collaborative strategy automatic generation function in step S4, generating strategic-level task allocation strategies, tactical-level resource scheduling strategies, and execution-level execution sequence strategies.

[0204] The strategy decision-making module connects the strategy generation module and the operation monitoring module. The strategy decision-making module includes a state encoder, a semantic encoder, a fusion layer, a decoder, and a scoring output layer. It is used to execute the strategy selection network function in step S5, determine the optimal cooperation scheme from candidate cooperation schemes, and output the result to the operation monitoring module.

[0205] The operation monitoring module connects to the strategy decision module. The operation monitoring module includes a task progress monitoring unit, a resource utilization monitoring unit, an abnormal event monitoring unit, and a collaboration mode adjustment triggering unit. It is used to execute the runtime monitoring function in step S6, continuously collect execution indicators during the execution of the plan, and trigger collaboration mode adjustments based on feedback signals.

[0206] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0207] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for intelligent agent collaboration in an artificial intelligence system for the production field, characterized in that, Includes the following steps: Construct a production domain ontology library and a process rule library, wherein the production domain ontology library adopts a hierarchical directed acyclic graph as the underlying storage structure, covering four dimensions: production process, process parameters, equipment capabilities, and quality standards. The process rule base covers three categories of rule systems: process constraint rules, resource dependencies, and execution conditions. The three heterogeneous inputs—task text description, product process requirements, and resource conditions—are transformed into a unified semantic vector representation that integrates domain ontology knowledge through a designed semantic embedding layer. The unified semantic vector representation is input into the semantic enhancement parsing engine. The semantic enhancement parsing engine performs semantic disambiguation and semantic completion based on the production domain ontology library and the process rule library. After incorporating the process constraint rules as constraint vectors into the semantic vectors, semantic consistency verification is performed to output a structured parsing result containing task objectives, constraints, resource requirements, and execution sequences. Based on the structured parsing results, a multi-level collaborative strategy is automatically generated, which includes strategic task allocation, tactical resource scheduling, and execution order. The strategy layers are decoupled through a loosely coupled architecture and coordinated through constraint propagation and conflict arbitration mechanisms. The unified semantic vector representation is combined with the current system state input strategy selection network to determine the optimal cooperation scheme; Deploy a runtime monitoring module to collect execution metrics in real time and trigger adjustments to the collaboration mode based on feedback signals to support dynamic refactoring of strategies.

2. The method according to claim 1, characterized in that, The production domain ontology library uses a hierarchical directed acyclic graph as its underlying storage structure, covering the production process dimension, specifically including: The production process is defined as a node in the hierarchical directed acyclic graph, and the logical relationship between processes is defined as an edge in the hierarchical directed acyclic graph. Process time constraints are stored as node attributes, which include four parameters: baseline execution time, longest execution time, shortest execution time, and flexible time window.

3. The method according to claim 1, characterized in that, The production domain ontology library covers equipment capabilities and quality standards dimensions, specifically including: The device state description adopts a finite state machine model. The state set includes six states: idle, running, debugging, fault, maintenance, and standby. The state transition rules are defined by a state transition matrix. The matrix element takes the value 0 or 1. A value of 1 indicates that the transition from a row state to a column state is allowed, and a value of 0 indicates that the transition from a row state to a column state is prohibited. When integrating the production domain ontology library and the process rule library, a cross-reference index is established, and a graph database is used as the underlying storage engine of the knowledge base after the integration of the production domain ontology library and the process rule library. Ontology nodes and edges are mapped to vertices and edges in the graph database, and rules and conditions are attached to the corresponding ontology nodes as vertex attributes, forming a unified knowledge graph structure that can be traversed and queried.

4. The method according to claim 1, characterized in that, The semantic embedding layer converts the three heterogeneous inputs—the task text description, the product process requirements, and the resource conditions—into a unified semantic vector representation, specifically including: The initial feature vector, process feature vector, and resource feature vector are each mapped to a unified semantic space dimension through their respective linear projection layers; Through the cross-modal attention module, the query vector obtained by linear projection of the initial feature vector and the key vector obtained by linear projection of the process feature vector are used to calculate the dot product similarity and then scaled and normalized by softmax to obtain the scalar attention weights. The calculated scalar attention weights are applied to the corresponding value vectors, and then all weighted value vectors are added element by element to obtain a weighted fusion vector.

5. The method according to claim 1, characterized in that, The semantic enhancement parsing engine performs semantic disambiguation and semantic completion, specifically including: The semantic disambiguation uses the concept hierarchy in the production domain ontology to query the equipment type and quality standards associated with the current process where the polysemous word is located, and determines the accurate meaning of the polysemous word in the current task context. The semantic completion is achieved by calculating the unified semantic vector representation and the first semantic vector in the production domain ontology. The dot product similarity of the embedding vectors corresponding to each conceptual entity is used to infer and complete the implicit entities, implicit relationships, and implicit constraints missing in the task description.

6. The method according to claim 1, characterized in that, The process constraint rules are incorporated into the semantic vector as constraint vectors, specifically including: The structured process constraint rules constructed in step S1 are mapped into constraint vectors using a rule encoder. Based on the degree of enforcement of the process constraint rules, the degree of enforcement coefficient of the process constraint rules is determined; wherein, the degree of enforcement coefficient of hard constraints is 1.0, and the degree of enforcement coefficient of soft constraints is a value that is dynamically adjusted according to the task type, resource availability and historical execution data; The coverage coefficient of the process constraint rule is determined based on the ratio of the number of processes involved in the process constraint rule to the total number of processes. No. Semantic fusion weights of constraints The calculation formula is ,in For the first The degree of enforceability of the constraint. For the first Coverage coefficient of the constraint; Based on the semantic fusion weights, the constraint vector is integrated into the semantic vector through a constraint-aware attention module.

7. The method according to claim 1, characterized in that, The semantic consistency verification includes three stages: syntactic consistency verification, domain consistency verification, and process consistency verification. The process consistency check checks whether the process parameter configuration meets the boundary conditions of the process constraint rules and whether the process sequence meets the process sequence constraint rules. For soft conflicts found during the check, a constraint relaxation algorithm is used for fault tolerance processing. While maintaining the hard constraints, the soft constraints are appropriately relaxed, and the relaxation range does not exceed 10% of the constraint boundary value.

8. The method according to claim 1, characterized in that, The unified semantic vector representation is input into the current system state input policy selection network to determine the optimal cooperation scheme, specifically including: The policy selection network includes a state encoder, a semantic encoder, a fusion layer, a decoder, and a scoring output layer; The state encoder adopts a graph neural network architecture, models the current system state as a resource node graph, and aggregates the features of resource nodes and their neighboring nodes through a message passing mechanism to generate a state feature vector. The semantic encoder employs a multilayer perceptron architecture to encode the unified semantic vector representation and generate a semantic feature vector. The fusion layer uses the semantic feature vector as a query and the state feature vector as a key and value to calculate the attention weight distribution and generate a joint feature vector. The decoder generates candidate cooperative schemes based on the joint feature vector; The scoring output layer calculates the comprehensive score of each candidate collaboration scheme through a comprehensive scoring function, and decides on the candidate collaboration scheme with the highest comprehensive score as the optimal collaboration scheme.

9. The method according to claim 1, characterized in that, The various strategy layers are coordinated through constraint propagation and conflict arbitration mechanisms, and the runtime monitoring module adjusts the cooperation mode based on feedback signals, specifically including: When inter-layer strategies conflict, an arbitration priority formula is used. Determine the order in which to handle each conflict, among which Arbitration priority for conflicts, These are the weighting coefficients for the strategy level: 3 for the strategic level, 2 for the tactical level, and 1 for the execution level. The coefficient represents the degree of enforceability of the constraint; arbitration decisions prioritize conflicts with higher arbitration priority values. The runtime monitoring module uses a dual triggering mechanism of threshold monitoring and trend prediction to trigger the adjustment of the collaboration mode; when dynamically reconstructing the strategy, hot switching technology is used to complete the dynamic adjustment of the strategy structure while retaining the state and execution progress of the unchanged strategy parts.

10. An intelligent agent collaboration system for artificial intelligence systems in the production field, characterized in that, include: The ontology library construction module is used to perform the steps of constructing the production domain ontology library and process rule library as described in claim 1. The ontology library construction module uses a graph database as the underlying storage engine. The semantic embedding module is used to perform the steps described in claim 1, which convert the three types of heterogeneous inputs—task text description, product process requirements, and resource conditions—into a unified semantic vector representation that integrates domain ontology knowledge. The semantic parsing module is used to perform the step of outputting structured parsing results through the semantic enhancement parsing engine as described in claim 1; The strategy generation module is used to perform the step of automatically generating a multi-level collaborative strategy as described in claim 1, which includes strategic-level task allocation, tactical-level resource scheduling, and execution-level execution order. The strategy decision module is used to execute the steps described in claim 1, which involve inputting a unified semantic vector representation and the current system state into a strategy selection network to determine the optimal cooperation scheme. The runtime monitoring module is used to perform the steps described in claim 1, including collecting execution metrics in real time and triggering adjustments to the collaboration mode based on feedback signals.

Images