Industrial agent collaborative decision-making system and method

Abstract

The invention discloses an industrial agent collaborative decision-making system and method, and relates to the crossing field of industrial automation and artificial intelligence, and the system comprises a data collection module which collects dynamic working condition signals in real time; the local agent module receives the dynamic working condition signal and generates an action proposal signal representing own intention and resource demand based on a built-in decision model; the collaborative decision-making platform module receives the action proposal signal, and forms and sends a negotiation instruction signal for initiating negotiation to a related local agent module; when the global optimization module detects a conflict which cannot be solved autonomously, a coordination strategy signal is generated and fed back to the collaborative decision-making platform module; the knowledge base module stores an industrial field knowledge graph and historical cooperation cases, provides optimization knowledge signals for the global optimization module, and provides decision knowledge signals for the local agent module. According to the method, the problems of low industrial multi-agent collaboration efficiency and difficulty in global optimization can be solved.

Description

Technical Field

[0001] This invention relates to the intersection of industrial automation and artificial intelligence, specifically to an industrial intelligent agent collaborative decision-making system and method. Background Technology

[0002] Currently, with the deepening development of intelligent manufacturing and the Industrial Internet, industrial production systems are becoming increasingly complex, and their decision-making processes are evolving from traditional centralized control to distributed intelligence. Against this backdrop, multi-agent system technology has been introduced into the industrial field. By abstracting physical devices or logical units into intelligent agents with perception, computation, and decision-making capabilities, it aims to improve the flexibility and autonomy of the system. However, in practical industrial applications, existing multi-agent-based solutions still face significant challenges. First, at the architectural level, many systems have not completely broken free from centralized thinking, still relying on a powerful central controller for task allocation and scheduling. Limited agent autonomy leads to poor system scalability, and the central node becomes a performance bottleneck and a single point of failure. Second, at the level of collaboration mechanisms, agents often lack efficient and standardized interaction protocols and negotiation mechanisms, relying instead on simple information transmission or responses based on fixed rules. This makes it difficult to handle complex resource competition and goal conflicts, easily leading to local optima or decision-making deadlocks. Secondly, regarding decision-making quality, intelligent agents rely heavily on real-time data and pre-set models, lacking a systematic utilization of deep industrial knowledge (such as process principles and equipment relationships) and historical collaborative experience. This results in decisions that are technically feasible but not necessarily economically or globally optimal, and are ill-suited to handling rare operating conditions. Finally, in terms of dynamic adaptability, existing systems typically require interrupting existing plans and global rescheduling when faced with sudden disturbances such as order changes or equipment failures, leading to slow responses and impacting production continuity. Although some research has attempted to improve collaboration by introducing algorithms such as contract networks and game theory, these methods are often difficult to implement due to excessive computational complexity or detachment from actual industrial constraints. Therefore, the industry urgently needs a new collaborative decision-making system architecture and method that can ensure high autonomy and rapid response of each agent, achieve effective global coordination and optimization, and integrate industrial knowledge and data. Summary of the Invention

[0003] In view of the shortcomings of the prior art, the purpose of this invention is to provide an industrial intelligent agent collaborative decision-making system and method to solve the problems of low efficiency and difficulty in global optimization of multi-agent collaboration in industry. This invention solves this problem by constructing a hierarchical distributed system. The bottom-level data acquisition module and local intelligent agent module are responsible for perception and rapid response, while the upper-level collaborative decision-making platform module provides a virtual collaboration space, aggregating action proposals from various intelligent agents and performing conflict detection to drive autonomous negotiation among agents. For complex conflicts, the global optimization module initiates a distributed algorithm for coordination. A knowledge base module is present throughout, providing domain knowledge and historical experience support for decision-making and optimization. Through this mechanism of "autonomous negotiation as the main method, global optimization as a supplement, and continuous knowledge-driven approach," the system approaches the global optimum while maintaining flexibility, achieving efficient, adaptive, and evolvable industrial collaborative decision-making.

[0004] This invention provides an industrial intelligent agent collaborative decision-making system, comprising: The data acquisition module collects industrial field operating condition data in real time, including at least equipment status, production progress, and material information, and generates dynamic operating condition signals. The local intelligent agent module receives dynamic operating condition signals and generates action proposal signals representing its own intentions and resource needs based on the built-in decision model. The collaborative decision-making platform module receives action proposal signals, aggregates and detects conflicts in signals from multiple local intelligent agent modules, and forms and sends negotiation instruction signals to relevant local intelligent agent modules to initiate negotiation. When the global optimization module detects a conflict that it cannot resolve on its own, it generates an optimization trigger signal and transmits it to the global optimization module. The global optimization module then runs a distributed optimization algorithm based on the optimization trigger signal, generates a coordination strategy signal, and feeds it back to the collaborative decision-making platform module. The knowledge base module stores industrial knowledge graphs and historical collaborative cases, and provides optimization knowledge signals to the global optimization module and decision-making knowledge signals to the local intelligent agent module.

[0005] In one embodiment of the present invention, the data acquisition module includes a sensor unit deployed on a physical device and an edge computing unit located at the network edge; the sensor unit senses multimodal raw physical signals such as vibration, temperature, current and visual images during device operation in real time; the edge computing unit connects to the sensor unit, receives the multimodal raw physical signals, and performs signal filtering, feature extraction and format standardization processing to fuse heterogeneous raw data into a unified dynamic operating condition signal with timestamp and data source identification; the data acquisition module periodically or event-triggeredly pushes the dynamic operating condition signal to the local intelligent agent module through industrial network protocols, thereby providing a real-time, clean and structured data foundation for subsequent decision-making.

[0006] In one embodiment of the present invention, the local intelligent agent module integrates a reaction layer, a collaborative planning layer, and a communication interface layer. The reaction layer responds to received emergency alarm-type dynamic operating condition signals at the millisecond level based on a preset rule base, generating direct control commands. The collaborative planning layer receives regular dynamic operating condition signals and decision knowledge signals from the knowledge base module, and uses a model-based reinforcement learning algorithm to simulate the interaction consequences of its own action proposals with the possible behaviors of other intelligent agents, thereby generating optimized action proposal signals. The communication interface layer is responsible for encapsulating the action proposal signals into messages conforming to predetermined communication standards and sending them to the collaborative decision-making platform module. It is also responsible for receiving negotiation command signals and coordination strategy signals from external sources. The model-based reinforcement learning algorithm is as follows:

[0007] in, Given the mathematical expectation, traverse the state space S and the agent's policy. Environmental Model ; / The industrial operating condition state of the intelligent agent at time t / t+1 is mapped by dynamic operating condition signals; / Propose actions for the agent at time t / t+1; For instant reward functions; The discount factor balances immediate returns with future returns; For action value networks, the value of the optimal future action is estimated; The interaction penalty coefficient balances the autonomous decision-making of agents with the constraints of multi-agent cooperation; To account for the multi-agent interaction loss, quantify the current agent action proposal. Other agent candidate action sets The degree of resource and time conflicts; the greater the conflict, the higher the loss. For an agent policy network, output the probability distribution of action proposals under a given state; : An environmental model that simulates the state transition patterns of industrial operating conditions and multi-agent interactions.

[0008] In one embodiment of the present invention, the collaborative decision-making platform module internally implements a virtual collaboration space and a conflict detection engine. The virtual collaboration space, as a shared information exchange medium, continuously receives and broadcasts action proposal signals from various local intelligent agent modules to form a global situation view. The conflict detection engine analyzes the global situation view in real time and identifies competitive and mutually exclusive conflicts in equipment, materials, or production capacity by comparing resource requirements, space occupancy, and time windows in different action proposal signals. When the identified conflict conforms to the preset simple conflict rules, the conflict detection engine directly generates a negotiation instruction signal containing the identifiers of the conflicting parties and suggested solutions. When complex or circularly dependent conflicts are identified, an optimization trigger signal is generated.

[0009] In one embodiment of the present invention, after receiving an optimization trigger signal, the global optimization module initiates a distributed optimization algorithm based on game theory Nash equilibrium search. The distributed optimization algorithm is as follows:

[0010] in, Adjust parameters (such as resource allocation, execution time window, and capacity allocation ratio) for the action proposals of n conflicting agents. These are the adjustment parameters for the i-th agent; Let be the global objective function, and let be the minimum objective function that integrates core industrial indicators. These are weighting coefficients, dynamically adjusted according to industrial production needs. The total energy consumption of the system is calculated by adding up the energy consumption of each agent's action proposal; It is a comprehensive quantitative indicator of the overall production efficiency of the system, including equipment utilization rate, task completion rate, and material flow efficiency. For a set of conflicting intelligent agents, Let be the feasible domain of the action parameters of the i-th agent; Let i be the revenue function of the i-th agent (such as equipment utilization rate, task completion revenue). Let be the parameters of all agents except i. This constraint is the core condition of Nash equilibrium: any agent cannot improve its own reward by adjusting its parameters alone, that is, all agents accept the current solution. To address the hard constraints of industrial processes, this algorithm takes the action proposal signals from conflicting parties gathered in the collaborative decision-making platform module as input, and uses the highest overall system production efficiency or lowest total energy consumption as the global objective function. Simultaneously, it incorporates the individual benefit constraints of each local agent module into the computational framework. The global optimization module calls upon optimization knowledge signals provided by the knowledge base module as prior information for the algorithm, and seeks a Pareto improvement solution acceptable to all conflicting parties through multiple rounds of iterative computation. Finally, the calculated resource reallocation scheme or execution sequence adjustment scheme is encapsulated as a coordination strategy signal and sent back to the collaborative decision-making platform module for distribution. The formula for determining the Pareto improvement solution is as follows:

[0011] in, The set of changes in payoffs for all conflicting agents. The change in the reward of the i-th agent; This represents the industrial tolerance threshold. This represents the change in the global objective function. <0 indicates that the overall cost has decreased compared to before optimization, meaning that the global objective has been optimized.

[0012] In one embodiment of the present invention, the knowledge base module includes an industrial knowledge graph sub-base and a collaborative case base. The industrial knowledge graph sub-base stores equipment process associations, material conversion relationships, and production constraint rules in a graph structure. The decision knowledge signals it provides are used to guide the local intelligent agent module to generate proposals that conform to the process route, and the optimization knowledge signals it provides are used to limit the feasible strategy search space for the global optimization module. The collaborative case base records historical collaborative processes and results in a structured log format, including the characteristics of successful cases and the root causes of failed cases. The cases are clustered and labeled through a machine learning model, so that the provided knowledge signals can be matched and pushed based on the similarity of the current dynamic working conditions, thereby realizing the reuse and inheritance of experience knowledge.

[0013] In one embodiment of the present invention, the action proposal signal generated by the local intelligent agent module has a standardized data structure, which includes at least the identity of the proposing intelligent agent, a description of the proposal target, a list of required resources and their quantity and time requirements, expected start and end timestamps, proposal priority weight, and proposal validity period. This structured signal design enables the collaborative decision-making platform module to perform mechanical parsing and comparison of proposals, and also provides clear optimization variables and constraints for the global optimization module, thereby ensuring the unambiguity of information transmission and the efficiency of processing between different modules.

[0014] In one embodiment of the present invention, the system further includes a digital twin simulation sandbox module, which is connected to the collaborative decision-making platform module and the local intelligent agent module respectively. The digital twin simulation sandbox module receives the global situation view signal from the collaborative decision-making platform module and the decision model copies of each local intelligent agent module, performs accelerated simulation and stress testing on various collaborative strategies in a virtual environment, and feeds back the performance evaluation signal and potential risk warning signal obtained from the simulation to the collaborative decision-making platform module and the relevant local intelligent agent module for verification and optimization of decisions before actual execution, forming a closed loop of "decision-simulation-optimization".

[0015] In one embodiment of the present invention, the system further includes a security and authorization management module. This module is responsible for the identity authentication, access control, and operation auditing of all intelligent agents within the system. It supports a three-tiered control architecture of "global-regional-local," enabling dynamic permission allocation based on three dimensions of "resource-operation-scenario," and is linked to the credit rating of intelligent agents to construct a closed-loop authorization mechanism encompassing the entire process of "application-approval-execution-monitoring-revocation." The security and authorization management module includes an identity authentication unit, an access control unit, and an audit trail unit. The identity authentication unit implements two-way authentication of intelligent agents based on digital certificates and dynamic tokens. The access control unit supports a three-tiered control architecture of "global-regional-local." The system is structured with a global level responsible for system policy formulation and cross-regional coordination, a regional level responsible for resource and task scheduling within its own region, and a local level responsible for device-level operation permission management. The permission control unit also implements a permission model based on a three-dimensional framework of "resource-operation-scenario." The resource dimension includes device, data, and network resources; the operation dimension includes read / write, control, negotiation, and optimization; and the scenario dimension includes normal production, emergency response, and maintenance / debugging. Permission allocation is dynamically linked to the agent's historical credit rating, with agents with higher credit ratings enjoying more lenient permissions and higher negotiation weight. The audit and tracing unit records the operation logs, permission change records, and abnormal behaviors of all agents, supporting real-time monitoring and post-event traceability.

[0016] This invention also includes an industrial intelligent agent collaborative decision-making method, comprising: S1: Real-time acquisition of industrial site operating condition data, including at least equipment status, production progress and material information, and formation of dynamic operating condition signals; S2: Receives dynamic operating condition signals and generates action proposal signals representing its own intentions and resource needs based on the built-in decision model; S3: Receive action proposal signals, aggregate and detect conflicts in signals from multiple local agent modules, form and send negotiation instruction signals to relevant local agent modules to initiate negotiation; S4: When a conflict that cannot be resolved autonomously is detected, an optimization trigger signal is generated and transmitted to the global optimization module. The global optimization module runs a distributed optimization algorithm based on the optimization trigger signal, generates a coordination strategy signal, and feeds it back to the collaborative decision-making platform module. S5: Stores industrial knowledge graphs and historical collaborative cases, and provides optimization knowledge signals to the global optimization module and decision-making knowledge signals to the local intelligent agent module.

[0017] This invention provides an industrial intelligent agent collaborative decision-making system and method, which solves this problem by constructing a hierarchical distributed system. The bottom-level data acquisition module and local intelligent agent module are responsible for perception and rapid response, while the upper-level collaborative decision-making platform module provides a virtual collaboration space, aggregating action proposals from various intelligent agents and performing conflict detection to drive autonomous negotiation among them. For complex conflicts, a global optimization module initiates a distributed algorithm for coordination. A knowledge base module is integrated throughout, providing domain knowledge and historical experience support for decision-making and optimization. Through this mechanism of "autonomous negotiation as the primary method, global optimization as a secondary method, and continuous knowledge-driven approach," the system approaches the global optimum while maintaining flexibility, achieving efficient, adaptive, and evolvable industrial collaborative decision-making. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a system architecture diagram of an industrial intelligent agent collaborative decision-making system. Figure 2 This is a flowchart of a collaborative decision-making method for industrial intelligent agents. Detailed Implementation

[0020] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0021] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0022] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0023] Please see Figure 1-2 The diagram illustrates an industrial intelligent agent collaborative decision-making system and method according to the present invention. The industrial intelligent agent collaborative decision-making system of the present invention includes a data acquisition module, which collects industrial field operating condition data in real time, including at least equipment status, production progress, and material information, and forms dynamic operating condition signals; a local intelligent agent module, which receives the dynamic operating condition signals and generates action proposal signals representing its own intentions and resource needs based on a built-in decision model; a collaborative decision-making platform module, which receives the action proposal signals, aggregates and detects conflicts in signals from multiple local intelligent agent modules, forms and sends negotiation instruction signals to relevant local intelligent agent modules to initiate negotiation; a global optimization module, which generates an optimization trigger signal and transmits it to the global optimization module when it detects a conflict that cannot be resolved autonomously, and runs a distributed optimization algorithm based on the optimization trigger signal to generate a coordination strategy signal and feeds it back to the collaborative decision-making platform module; and a knowledge base module, which stores industrial domain knowledge graphs and historical collaborative cases, and provides optimization knowledge signals to the global optimization module and decision knowledge signals to the local intelligent agent modules.

[0024] like Figure 1As shown, the technical solution protected by this invention relates to an intelligent agent collaborative decision-making system for complex industrial environments. Its core architecture consists of multiple interconnected and functionally distinct modules. These modules interact through specific signal flows to jointly achieve a complete decision-making closed loop from data perception to global coordination. The following will provide a detailed technical description of the first three fundamental and core modules of this system. First, the system's data acquisition module constitutes the sensory nerve endings of the entire intelligent decision-making system, serving as a bridge between the physical world and the digital decision-making system. This module is not a simple data aggregation point, but rather a collection of distributed edge computing units with preliminary data processing capabilities. Its core function is to continuously capture raw, multi-source, and heterogeneous operational data from the vast industrial field. This data widely originates from various sensors installed on production lines, machine tools, conveyor belts, warehousing equipment, and in the environment. Examples include vibration and temperature sensors for monitoring equipment health, visual recognition systems or RFID readers for tracking material consumption and product output, current and voltage sensors for acquiring precise equipment operating parameters, and production order progress and planning information obtained from manufacturing execution systems or enterprise resource planning systems. These raw data streams are often chaotic, unstructured, and mixed with noise in their initial state.

[0025] Furthermore, the data acquisition module is designed with integrated signal conditioning and preprocessing capabilities. Once the sensor unit captures the raw physical signal, the connected edge computing unit immediately intervenes. These edge computing units, deployed at the network edge close to the data source, first filter the raw signal to eliminate noise such as electromagnetic interference common in industrial environments. Next, they perform feature extraction operations, such as calculating RMS, peak value, and kurtosis—features characterizing equipment status—from a vibration waveform, or identifying the type, quantity, and location of materials from an image stream. Finally, these processed multimodal data are standardized and encapsulated according to a predefined, uniform format, with each data packet bearing a precise timestamp and a unique data source identifier, thus forming a dynamic operating condition signal that can be unambiguously understood by all upper-layer modules within the system. This process achieves the transformation from raw bitstream to a structured signal rich in semantic information. The module actively pushes the encapsulated dynamic operating condition signal to the corresponding local intelligent agent module in the system at fixed time periods or when a specific event is detected, via reliable industrial network protocols such as Industrial Ethernet or Time-Sensitive Networking (TSN). This design ensures that the data foundation upon which the decision-making system is based is real-time, clean, structured, and traceable, providing a solid and reliable data source for all subsequent advanced intelligent decisions. Secondly, the local intelligent agent module is the carrier of the system's distributed intelligence. Each module represents a physical or logical industrial entity, such as a machine tool, a robot, a warehouse shelf, or a production line. This module is the core unit for receiving and processing dynamic operating condition signals and generating initial decision intentions. Its internal structure is carefully designed, employing a biomimetic hierarchical decision-making architecture to balance the agility of response with the intelligence of decision-making. This architecture mainly includes a reaction layer, a collaborative planning layer, and a communication interface layer. The reaction layer is at the forefront of decision-making, and it incorporates a high-speed rule engine and a predefined library of key event patterns. Its main responsibility is to handle emergencies with high timeliness requirements. When the dynamic operating condition signals transmitted from the data acquisition module contain emergency alarm information such as the triggering of an emergency stop button, the obstruction of a safety light curtain, or the exceeding of key parameters, the reaction layer is directly triggered with a millisecond delay. Instead of complex calculations, it immediately generates a direct control command signal based on a pre-set, security-verified rule base. This signal may instruct the equipment to immediately shut down, activate a backup unit, or issue an audible and visual alarm, thereby ensuring the fastest and most reliable instinctive response in emergencies and protecting personal and equipment safety.

[0026] Specifically, the collaborative planning layer acts as the "brain" for the local agent module to make complex decisions and think strategically. It receives and processes routine dynamic operating condition signals, such as the current equipment's operating mode, progress, queue length, and decision-making knowledge signals from the knowledge base module—signals that may include relevant process parameter schemes, historical optimal operating ranges, and other experiential knowledge. The core of the collaborative planning layer is one or more built-in decision models. These models can be rule-based expert systems, but more advanced ones are agent models trained using model-based reinforcement learning algorithms. This model enables the agent to "think": when faced with a task, it can perform multiple simulations and deductions within its internally constructed, simplified model of itself and its environment. It can not only evaluate the effectiveness of its different action plans but also predict the possible reactions of other related agents, thereby assessing the consequences of interactions. Through this internal simulation, the collaborative planning layer generates an optimal solution for future actions, i.e., an action proposal signal. This signal is a structured information package that clearly explains the agent's intention, such as "Requesting the use of machine tool A for process X during the time period T1 to T2, with an estimated time H and a required amount of material M." This proposal is an agent's "offer" or "declaration" to participate in system collaboration. The communication interface layer is the unified portal for interaction between local agent modules and the outside world. It is responsible for encapsulating the action proposal signals generated by the collaborative planning layer according to the system-defined, standardized communication protocol, forming standard message frames, and reliably sending them to the collaborative decision-making platform module. Simultaneously, it is also responsible for monitoring the network, receiving negotiation instruction signals or final coordination strategy signals from the collaborative decision-making platform module, and accurately distributing them to the internal reaction layer or collaborative planning layer for processing. This design decouples the complex decision-making logic within the agent from external communication interactions, improving the module's independence and maintainability. Finally, the collaborative decision-making platform module is the "collaboration hub" and "air traffic control tower" of the entire system. It does not directly control any devices but provides an open, transparent, and orderly interaction and negotiation environment for all local agent modules. Its core function is to realize the globalization of local information and the explicitization of implicit conflicts. This module internally implements two key components: a virtual collaboration space and a conflict detection engine. The model-based reinforcement learning algorithm is as follows:

[0027] in, Given the mathematical expectation, traverse the state space S and the agent's policy. Environmental Model ; / The industrial operating condition state of the intelligent agent at time t / t+1 is mapped by dynamic operating condition signals; / Propose actions for the agent at time t / t+1; For instant reward functions; The discount factor balances immediate returns with future returns; For action value networks, the value of the optimal future action is estimated; The interaction penalty coefficient balances the autonomous decision-making of agents with the constraints of multi-agent cooperation; To account for the multi-agent interaction loss, quantify the current agent action proposal. Other agent candidate action sets The degree of resource and time conflicts; the greater the conflict, the higher the loss. For an agent policy network, output the probability distribution of action proposals under a given state; The environment model simulates the state transition patterns of industrial operating conditions and multi-agent interactions. The model-based reinforcement learning algorithm used in the collaborative planning layer of the local agent module optimizes the generation of action proposals by constructing a target value function. This algorithm uses the target value function shown in the formula... With the core optimization objective, this method integrates industrial operating conditions, agent actions, multi-agent interaction constraints, and reward quantification through mathematical modeling. This solves the problem that traditional reinforcement learning only focuses on the agent's own decisions and is detached from the actual constraints of multi-agent collaboration in industry. In the formula, E represents the mathematical expectation traversal of the industrial operating conditions, agent action strategies, and environmental interactions. and These correspond to the dynamic operating conditions at time t and time t+1 in the industrial setting, respectively. These conditions are obtained by mapping dynamic operating condition signals such as equipment status, production progress, and material information output by the data acquisition module, covering core industrial elements such as equipment operating parameters, material inventory, and production task nodes. and This refers to the action proposal generated by the agent at a given moment, including core elements such as resource requirements, time window, and execution objective. Instant reward function. Based on the actual needs of industrial production, the design quantifies key industrial indicators such as improved production efficiency, reduced energy consumption, task completion rate, and equipment utilization rate, transforming them into numerical rewards. γ serves as a discount factor to balance the immediate benefits of the agent's current actions with long-term future benefits, adapting to the long-cycle and continuous characteristics of industrial production. Action Value Network Used to estimate the value of the agent's optimal future action, providing a basis for forward-looking planning of action proposals; λ is the multi-agent interaction penalty coefficient. This is a multi-agent interaction loss function. By quantifying the degree of resource competition, time conflict, and space exclusion between the current agent's action proposal and other agent candidate action sets, this function incorporates the core constraints of multi-agent collaboration in industrial scenarios into the algorithm, preventing agents from generating independent proposals that are detached from collaborative requirements. Policy Network With environmental model These modules are responsible for outputting the probability distribution of action proposals and simulating the state transition of industrial conditions. By iteratively optimizing the policy parameter θ, the objective value function is maximized, ultimately generating optimized action proposals that both conform to the interests of the agents and adapt to the collaborative constraints of multiple agents. Those skilled in the art can complete the code implementation and parameter tuning of the algorithm based on the mathematical logic of the formula and the definition of industrial parameters. By combining the equipment characteristics, production processes, and collaborative rules of the industrial site, key parameters such as reward functions, penalty coefficients, and discount factors can be set to ensure that the local agent module can stably generate compliant, efficient, and collaborative action proposal signals that adapt to collaborative needs. This solves the problem of traditional multi-agent decision-making being divorced from actual industrial constraints from the algorithmic level.

[0028] In one embodiment of the invention, the virtual collaborative space is a shared data area that is logically centralized but physically distributed. It continuously receives action proposal signals broadcast from all local agent modules, stores, indexes, and publishes these signals in real time. Therefore, any agent accessing this space gains a constantly updated global situational view that transcends its local perspective. This view clearly shows the intentions, resource requirements, and planned timelines of all participants at any given moment, forming the foundation for collaborative cognition within the system. The conflict detection engine is an intelligent analyzer that continuously analyzes the global situational view. Based on a built-in conflict rule base, it automatically compares and correlates a large number of converged action proposal signals. Typical conflict patterns include: two or more proposals requesting the use of the same non-shareable equipment within the same time period; a downstream process's proposal starting earlier than the expected end time of its upstream process's proposal; multiple material handling proposals making mutually exclusive requests for the same batch of materials, etc. The conflict detection engine can accurately identify these competitive and mutually exclusive relationships in terms of resources, time, and space. Its decision-making logic is hierarchical: For detected simple conflicts, such as those involving only two parties with clear priorities or preset resolution rules, the engine directly generates a negotiation instruction signal. This signal includes the identifiers of the conflicting parties, the specific content of the conflict, and a preliminary rule-based solution. This signal is then precisely sent to the relevant local intelligent agent modules, triggering them to conduct one-on-one autonomous negotiation. For detected complex conflicts, such as those involving circular dependencies of three or more parties, contradictory objective functions, or situations where there is no solution in the preset rule base, the conflict detection engine determines that this is a conflict that cannot be resolved autonomously. It then generates an optimization trigger signal and transmits it to the higher-level global optimization module, requesting intervention for more complex global coordination calculations. In summary, the data acquisition module, the local intelligent agent module, and the collaborative decision-making platform module together constitute a complete decision-making loop from perception to intent expression to preliminary conflict coordination. The data acquisition module ensures the real-time and authenticity of the decision-making basis, the local intelligent agent module endows the system's underlying units with autonomous response and intelligent planning capabilities, and the collaborative decision-making platform module, by providing a shared space and conflict detection mechanism, initially connects the dispersed individual intelligences, laying the foundation for effective overall collaboration. The collaborative work of these three modules enables the system to handle most decision-making and coordination tasks in a highly distributed and autonomous manner during daily operation, thereby significantly reducing the burden and bottlenecks that centralized computing may bring while maintaining the system's flexibility and robustness.

[0029] In one embodiment of the present invention, the global optimization module is the "ultimate decision-making arsenal" for the system to cope with complex collaborative challenges. It is normally in a low-power standby state, and is only activated when it receives an optimization trigger signal from the collaborative decision-making platform module. This ensures that the system operates efficiently in a distributed manner in most cases, only utilizing more expensive centralized computing resources when necessary. The core of this module is an efficient and interpretable distributed optimization algorithm library, one typical implementation of which is a search algorithm based on the Nash equilibrium concept in game theory. The distributed optimization algorithm is as follows:

[0030] in, Adjust parameters (such as resource allocation, execution time window, and capacity allocation ratio) for the action proposals of n conflicting agents. These are the adjustment parameters for the i-th agent; Let be the global objective function, and let be the minimum objective function that integrates core industrial indicators. These are weighting coefficients, dynamically adjusted according to industrial production needs. The total energy consumption of the system is calculated by adding up the energy consumption of each agent's action proposal; It is a comprehensive quantitative indicator of the overall production efficiency of the system, including equipment utilization rate, task completion rate, and material flow efficiency. For a set of conflicting intelligent agents, Let be the feasible domain of the action parameters of the i-th agent; Let i be the revenue function of the i-th agent (such as equipment utilization rate, task completion revenue). Let be the parameters of all agents except i. This constraint is the core condition of Nash equilibrium: any agent cannot improve its own reward by adjusting its parameters alone, that is, all agents accept the current solution. As a hard constraint for industrial processes, the algorithm takes the action proposal signals of the conflicting parties gathered in the collaborative decision-making platform module as input, and takes the highest overall system production efficiency or the lowest total energy consumption as the global objective function, while incorporating the individual benefit constraints of each local intelligent agent module into the calculation framework.

[0031] When the algorithm starts, it first mathematically models the conflict scenario encapsulated in the optimization trigger signal: each involved local agent is treated as a rational game participant, and its action proposal signal corresponds to its respective strategy choice. The payoff of the strategy is related to the system's global objective function and its own individual objective. The task of the global optimization module is not to forcibly enforce a certain solution, but to find an equilibrium point that is acceptable to all participants, i.e., a Pareto improvement solution or an approximate Nash equilibrium point, under the constraint of respecting the basic interests of each agent. In this process, the optimization knowledge signals provided by the knowledge base module play a crucial guiding role. This knowledge may limit the search space of the strategy in the form of constraints, such as specifying the order in which certain process steps must be followed, or accelerate the convergence of the algorithm in the form of heuristic information, such as providing successful resolution patterns of similar conflicts in history as initial solutions. The optimization algorithm simulates the policy interaction and adjustment between agents through multiple rounds of iterative calculation, and finally outputs a coordinated strategy signal that is globally optimal and reasonable for each individual agent. This signal may include fine-tuning of time windows in the original action proposals of multiple agents, redefining resource allocation ratios, or reconstructing the task execution sequence. This signal is fed back to the collaborative decision-making platform module, which authoritatively distributes and supervises its execution, thereby resolving complex global conflicts with minimal perturbation. The formula for determining the Pareto improved solution is as follows:

[0032] in, The set of changes in payoffs for all conflicting agents. The change in the reward of the i-th agent; This represents the industrial tolerance threshold. This represents the change in the global objective function. <0 indicates that the overall cost has decreased compared to before optimization, meaning that the global objective has been optimized.

[0033] The global optimization module employs a distributed optimization algorithm based on game theory Nash equilibrium search. Its core computational framework is a two-layer optimization mathematical model, combined with the Pareto improvement solution criterion. This addresses the issues of excessive computational complexity and difficulty in implementation outside of industrial process constraints inherent in traditional game theory algorithms. The distributed optimization algorithm deeply integrates the global optimization objective with individual agent constraints and industrial process constraints. The optimization variable x represents adjustment parameters such as resource allocation, time windows, and production capacity ratios from the action proposals of each conflicting agent, covering core conflict elements in industrial collaboration, including equipment, materials, and production capacity. The global objective function... Total energy consumption of the integrated system With production efficiency By dynamically adapting the weighting coefficient α to different industrial production needs such as energy saving priority, capacity priority, and cost optimization, it meets the optimization objectives of diverse industrial scenarios such as discrete manufacturing and process industries. (Constraints are listed below.) The feasible domain of each agent's action parameters is limited by actual production conditions such as the rated parameters of industrial equipment, the processing range, and material supply capacity, to avoid optimization results exceeding the carrying capacity of industrial hardware and processes. This is achieved through Nash equilibrium constraints. ≥ Ensure that individual adjustments to parameters by any agent cannot increase its own gains, making the optimization results cooperatively acceptable, and preventing the coordination strategy from failing to execute due to the harm to the interests of a single agent, thus addressing the hard constraints of industrial processes. An industrial knowledge graph, incorporating a knowledge base module, transforms core industrial logic such as equipment process relationships, material transformation relationships, process sequence, and production safety rules into mathematical constraints. This fundamentally prevents optimization results from deviating from industrial realities, such as process violations and safety hazards. The Pareto improvement solution determination formula serves as the quantitative basis for determining Pareto improvement solutions, calculating the change in agent revenue before and after optimization. Combined with the change in the global objective function ΔF, and the industrial tolerance threshold Accurately determine whether the optimized solution meets the core requirements of "no significant loss of individual gains, increased gains for some individuals, and optimization of the global objective". By clustering and labeling historical collaborative cases from the knowledge base using machine learning, this algorithm aligns with the acceptable range of profit fluctuations for agents in industrial production, avoiding situations where excessive pursuit of global optimum leads to a significant drop in individual profits and rejection of coordination strategies. The algorithm employs a distributed computing architecture, decomposing the global optimization problem into local parameter optimization for each agent and Nash equilibrium constraint verification, significantly reducing the complexity of centralized computation. Those skilled in the art can build distributed computing nodes based on the optimization model of the distributed optimization algorithm, set iterative convergence conditions using the criteria of the Pareto improved solution judgment formula, and adjust optimization parameters according to the conflict type, production goals, and process rules in the industrial setting. This rapidly generates coordination strategy signals that conform to industrial realities and are acceptable to all conflicting parties, achieving efficient global optimization for multi-agent collaboration in industry and making game theory optimization algorithms truly adaptable to practical applications in industrial scenarios.

[0034] like Figure 1As shown, the knowledge base module is the soul that runs through the entire system, endowing it with "wisdom" and "experience." It is not a static data warehouse, but a dynamically growing and continuously learning industrial knowledge hub. This module mainly comprises two complementary sub-repositories: the industrial knowledge graph sub-repository and the collaborative case repository. The industrial knowledge graph sub-repository is constructed using graph database technology. Its nodes represent industrial entities such as equipment, materials, processes, and products, while edges represent the rich relationships between them, such as "Equipment A can process material B," "Process X must follow Process Y," and "Using parameter P will improve yield but increase energy consumption." This structured knowledge network provides deep domain logic support for the system. When the local intelligent agent module generates an action proposal, it can query or receive decision knowledge signals from the knowledge base. These signals can verify the feasibility of its proposal in terms of process logic or recommend a better parameter range, thereby avoiding invalid proposals that violate physical laws or common sense in processes. For the global optimization module, optimization knowledge signals can be integrated into the optimization model in the form of hard or soft constraints, ensuring that the calculated coordination strategy is not only mathematically optimal but also feasible and safe in engineering practice. The collaborative case library focuses on accumulating procedural experience. Like a detailed "collaborative decision-making medical record," it documents the complete data chain from the occurrence of conflict, negotiation or optimization process to the final execution result. Through machine learning techniques, these cases are cleaned, labeled, and characterized to form searchable patterns. When the system encounters similar situations or early conflict scenarios again, the knowledge base module can proactively push historically successful solutions for reference or warn of potential risks that may lead to failure. This ability to "remember" and "learn" enables the system to avoid repeating mistakes and continuously solidify effective collaborative patterns into reusable knowledge, driving the continuous evolution of the entire system's decision-making level.

[0035] Furthermore, the standardized data structure of the action proposal signals generated by the local intelligent agent module serves as the "common language" and "basic protocol" for seamless collaboration across the entire massive system. This design embodies profound systems engineering principles, ensuring high-fidelity and unambiguous information flow within complex systems. This data structure is a carefully defined set of fields, each carrying clear semantics. Among them, the "Proposal Intelligent Agent Identity" is the source of the signal, ensuring that any decision and feedback can be accurately traced. The "Proposal Target Description" clearly describes the intended specific action using standardized work codes or keywords processed from natural language, such as "drilling," "spraying," or "transporting to Area A." The "List of Required Resources and Their Quantity and Time Requirements" is the primary basis for conflict detection, requiring a detailed list of equipment numbers, mold codes, material batches, and precise quantities and durations. The "Expected Start and End Timestamps" define the proposal's position on the timeline, crucial for timing coordination and conflict resolution. The "Proposal Priority Weight" is an important flexible coordination field, dynamically calculated based on order urgency, customer level, or process criticality, serving as a key trade-off factor in negotiation and optimization. The "proposal validity period" imbues decisions with a time attribute, preventing erroneous coordination due to expired information. This highly structured signal design enables the conflict detection engine of the collaborative decision-making platform module to mechanically and efficiently analyze and compare proposals, much like a computer parsing program, quickly locating conflict points. Simultaneously, it provides a direct mathematical foundation for accurately defining optimization variables, objective functions, and constraints in the global optimization module, greatly simplifying the modeling process for complex optimization problems. In short, this standardized signal design integrates heterogeneous intelligent agents distributed in different physical locations and performing different functions into a community capable of rational dialogue and collaboration, serving as an indispensable technological cornerstone for the system's efficient collaboration.

[0036] In one embodiment of the invention, the enhancement module is a digital twin simulation sandbox, which provides the entire decision-making system with a safe, controllable, and efficient "pre-launch battlefield" and "decision laboratory." This module is not independent of the real-time production control flow but is deeply integrated with it, forming a dynamic, bidirectional, interactive verification and optimization closed loop. The digital twin simulation sandbox module physically connects the collaborative decision-making platform module and each local agent module. It continuously subscribes to global situational view signals from the collaborative decision-making platform module. These signals, like snapshots of the real world, are captured by the sandbox module and used to initialize or update its internally constructed high-fidelity virtual factory environment. Simultaneously, it obtains or calls copies or simplified versions of the core decision models of each local agent module through specific interfaces. These model copies can simulate the decision-making logic of the agents in the virtual environment. When the collaborative decision-making platform module is about to trigger a complex round of multi-agent negotiation, or when the global optimization module generates a new set of coordination strategy signals, these collaborative schemes to be executed can be sent to the digital twin simulation sandbox module first. The module then leverages its powerful parallel computing capabilities to run hundreds or thousands of simulations of the solution in a virtual environment at an accelerated timescale. These simulations not only model normal processes but also inject various pre-set or randomly generated disturbances, such as sudden equipment slowdowns, delayed material delivery, or urgent order requests. Through this large-scale "stress test" and "Monte Carlo simulation," the sandbox module can objectively evaluate the performance of different collaborative solutions under multiple possible future paths, quantifying their robustness, efficiency bottlenecks, and potential risks. Ultimately, it generates two key signals to feed back to the system: first, a performance evaluation signal, which compares the advantages and disadvantages of different solutions with detailed indicator data, providing data for the final decision of the collaborative decision-making platform module or the parameter tuning of the global optimization module; second, a potential risk warning signal, which identifies vulnerable links or conflict points in the solution that may collapse under specific disturbances, enabling relevant local intelligent agent modules to prepare contingency plans in advance. This mechanism fundamentally changes the decision-making model, shifting from ex-post remediation of "execution-error detection-correction" to ex-ante verification of "simulation-optimization-re-execution", greatly improving the probability of successful decision-making and the system's ability to cope with uncertainty.

[0037] like Figure 2As shown, the present invention also includes an industrial intelligent agent collaborative decision-making method, comprising: S1: real-time acquisition of industrial site working condition data including at least equipment status, production progress, and material information, and forming a dynamic working condition signal; S2: receiving the dynamic working condition signal and generating an action proposal signal representing its own intention and resource needs based on a built-in decision model; S3: receiving the action proposal signal, aggregating and conflict detecting signals from multiple local intelligent agent modules, forming and sending negotiation instruction signals to relevant local intelligent agent modules for initiating negotiation; S4: when a conflict that cannot be resolved autonomously is detected, generating an optimization trigger signal and transmitting it to the global optimization module, the global optimization module running a distributed optimization algorithm based on the optimization trigger signal, generating a coordination strategy signal and feeding it back to the collaborative decision-making platform module; S5: storing an industrial domain knowledge graph and historical collaborative cases, and providing optimization knowledge signals to the global optimization module and decision knowledge signals to the local intelligent agent modules.

[0038] Specifically, in the implementation of the negotiation mechanism, the system adopts and deeply improves the classic Contract Network protocol, forming an intelligent collaborative contract establishment process that integrates technical assessment and trust management. This process is mediated and coordinated by the collaborative decision-making platform module, enabling cooperation between local intelligent agent modules to transcend simple task distribution and become a partner selection based on comprehensive value. The entire interaction process begins with a clear collaborative need: when a local intelligent agent module (called the manager) needs the services or resources of other intelligent agents for its task, it does not directly send instructions to specific objects, but instead issues a structured bidding signal through the collaborative decision-making platform module. This signal fully discloses the technical specifications, quality requirements, expected time window, and any necessary constraints of the task, equivalent to a public "bid." The collaborative decision-making platform module broadcasts this bidding signal to all local intelligent agent modules that may have the relevant capabilities. The relevant intelligent agents that receive the bidding signal (called bidders) immediately initiate internal evaluation. They do not respond blindly, but conduct rigorous feasibility analysis based on their current dynamic operating conditions, capability models, and load status. Only bidders who confirm that they can meet the technical requirements and that the expected benefits are reasonable will generate a rigorous bidding signal and submit it. This bidding signal is also highly structured, including the committed technical solution, estimated costs (such as time and energy consumption), the possible start time, and a summary of its own historical performance data. Subsequently, managers collect all bidding signals through the collaborative decision-making platform module, and the evaluation phase officially begins. The core improvement here lies in the multidimensionality of the evaluation criteria. Managers not only compare the merits and costs of the technical solutions in each bidding signal, but also introduce a crucial dimension: historical reputation. The system queries the knowledge base module or a dedicated trust management unit to check each bidder's historical records, including success rate in fulfilling contracts, task completion quality, and cooperative spirit in the face of unforeseen circumstances. Technical feasibility is the entry ticket to cooperation, while overall reputation is the cornerstone of establishing a long-term, reliable collaborative relationship. Managers comprehensively evaluate and rank all bids using a preset, potentially dynamically weighted evaluation function. Finally, managers generate an award signal, which explicitly designates the successful bidder and includes the final confirmed task details, the conditions promised by both parties, and metaphors for breach of contract clauses, which are formally sent to the successful bidder through the collaborative decision-making platform module. Once the successful bidder receives the award signal, it must return a confirmation signal. At this point, a binding virtual contract is established, guiding the actions of both parties and subject to system monitoring. This mechanism organically combines market bidding principles with trust management, guiding agents not only to focus on short-term gains but also to maintain their long-term reputation, thereby spontaneously cultivating a healthy and reliable collaborative ecosystem within the system.

[0039] In one embodiment of this invention, the management and control responsibilities in an industrial scenario are inherently hierarchical, typically manifested as a three-tiered control model: factory-level (global), workshop or production line-level (regional), and equipment or workstation-level (local). This system innovatively maps this real-world management architecture to a digital collaborative decision-making system, constructing a corresponding "global-regional-local" three-tiered collaborative security control architecture. The global control layer, as the central hub for formulating and deciding on system security strategies, is responsible for defining a unified security baseline, authentication standards, core resource access policies, and cross-regional collaborative security rules for the entire system. Like the system's brain, it examines the security situation from a holistic perspective, handles major security events and permission conflicts involving multiple regions or affecting the entire production objective, and maintains the credit evaluation system for all intelligent agents in the system. The regional control layer, acting as a bridge between the upper and lower layers, corresponds to specific production units (such as a welding workshop or an assembly line) and has regional security management nodes. These nodes are responsible for localizing and refining the global strategy, approving resource collaboration and task execution permissions between intelligent agents within their region, coordinating and resolving permission competition within the region, and performing preliminary detection and response to abnormal operational behaviors within their region. The local control layer is the final execution endpoint of the security policy, and its logic is directly embedded within each local agent module. It is responsible for performing final permission verification on every control command and data access request issued by the agent within millisecond response time, executing allow or block operations, and collecting the most basic operation audit logs. This three-tier architecture achieves a unified approach to centralized and decentralized security management: the global level grasps the direction and rules, the regional level achieves flexible adaptation and efficient coordination, and the local level ensures the accurate implementation and rapid response of policies. It not only aligns with the physical and organizational realities of industrial production but also enables security control to expand synchronously with production, providing flexibility for each region to address specific local security needs while ensuring unified security standards. To achieve granular control over complex entities and operations in the industrial field, this system abandons the coarse-grained permission management approach and innovatively adopts a three-dimensional permission subdivision model based on "resource-operation-scenario". The first dimension is "resources," which encompasses all entities in an industrial system that require protection. This includes not only physical equipment (such as CNC machine tools, robotic arms, and AGVs), sensors, and actuators, but also digital assets such as production process parameters, real-time production data, historical case libraries, optimization algorithm models, and network bandwidth, computing and storage resources. The second dimension is "operations," which defines the specific actions an entity can perform on resources, including but not limited to reading status, writing parameters, start / stop control, task negotiation initiation, optimization algorithm invocation, debugging command issuance, and log auditing. Different operations correspond to different risk levels; for example, the risk of "start / stop control" is far higher than that of "reading status." The third dimension is "scenario," which is crucial for granting dynamic and context-aware permissions.The system predefines several typical operating scenarios, such as "normal continuous production," "planned equipment maintenance," "emergency fault handling," "process parameter debugging," and "system upgrade and switchover." In different scenarios, the permission status of the same agent performing the same operation on the same resource may differ. For example, in the "normal production" scenario, a regular handling agent is not authorized to modify machine tool parameters; however, in the "equipment maintenance" scenario, an authorized maintenance agent can perform parameter debugging operations. The three-dimensional permission model, through the combination of (resource, operation, scenario), can precisely define under what circumstances, who, can, what, and how to process, achieving minimal permission allocation and contextualized adaptation.

[0040] Furthermore, the allocation of permissions is not static but dynamically linked to each agent's "credit rating." The credit rating is a comprehensive quantitative indicator automatically calculated through continuous analysis of the agent's historical behavior records. Calculation factors include, but are not limited to: task fulfillment success rate, compliance rate with collaborative commitments, proposal execution quality, frequency of anomalies or alarm triggers, historical security violation records, and collaborative evaluations in feedback from other agents. An agent that consistently collaborates reliably and efficiently completes tasks will gradually increase its credit rating; conversely, an agent that frequently causes conflicts, fails to execute tasks, or exhibits suspicious behavior will see its credit rating decrease. Credit rating directly affects permissions: agents with high credit ratings enjoy higher priority and shorter approval paths when initiating resource requests, and may even gain access to a wider range of potential resource access schemes; while in high-risk or critical scenarios, certain operational permissions for agents with low credit ratings may be temporarily restricted or subject to stronger approval constraints. This linkage mechanism transforms security control from passive rule enforcement to proactive trust management, encouraging agents to accumulate credit capital through standardized, reliable, and efficient behavior, thereby gaining greater autonomy and collaborative opportunities within the system, forming a virtuous cycle of mutual promotion between security and efficiency. Based on the aforementioned 3D model and credit system, the system constructs a complete, closed-loop, and traceable dynamic authorization management mechanism encompassing the entire process of "application-approval-execution-monitoring-revocation," ensuring that every use of permissions is within a controlled lifecycle. The process begins with "application": When an agent (or its representative human operator) needs to perform an operation exceeding its current default permissions, it must submit a structured permission request signal to the security and authorization management module through a standardized interface. This signal must clearly include key information such as the target resource, the requested operation, the applicable scenario, the reason for the application, and the expected validity period. Next is the "approval" stage: the approval process intelligently decides based on the 3D permission strategy, the real-time credit rating of the requesting agent, the current system operating scenario, and resource load status. For routine and low-risk applications, the system can automatically approve within seconds according to preset rules; for applications involving critical resources, high-risk operations, or those with average credit ratings, the approval may be escalated to regional or global human review nodes for confirmation. Approval decisions and their basis are fully recorded. After approval, the process enters the "execution" phase: the corresponding permissions are temporarily activated, and a time-limited digital token is generated and simultaneously distributed to the requesting agent and related resource management nodes. The agent can only execute authorized operations with this token, and all operation requests must include the token for verification. "Monitoring" is integrated throughout the entire execution process: the audit trail unit monitors and analyzes the operation flow with the permission token in real time, comparing the actual operation with the authorized content to detect any unauthorized access, abuse, or abnormal patterns. Once a violation is detected, the system can issue an alert and intervene in real time, while simultaneously recording all operation logs to create an unalterable audit trail.Finally, there's "revocation": when a task is completed, the permission token expires, the credit rating is dynamically downgraded, or a scenario change renders the authorization inappropriate, the system will automatically or be triggered to revoke the permission, ensuring its timeliness and minimizing its holding. This closed-loop process achieves refined management of permissions throughout their entire lifecycle, from generation and use to expiration. This allows the system to dynamically adjust permission configurations based on the actual progress of production tasks, the real-time performance of agents, and environmental changes, ensuring smooth operation of necessary tasks while minimizing the security attack surface and internal operational risks. In summary, the security and authorization management system deeply integrated into this system achieves precise alignment of control responsibilities with industrial realities through a three-tiered "global-regional-local" architecture; it achieves extremely fine-grained and context-aware permission definitions through a three-dimensional "resource-operation-scenario" model; it introduces dynamic trust and behavioral incentive mechanisms through the linkage between permissions and credit ratings; and finally, through a closed-loop "application-approval-execution-monitoring-revocation" process, it achieves manageable, controllable, and auditable authorization lifecycles. This mechanism does not operate in isolation, but rather interacts deeply with core modules such as the collaborative decision-making platform, local intelligent agents, and knowledge base: the conflict detection engine invokes permission status for pre-judgment when identifying resource conflicts; the generation and execution of negotiation instructions are constrained by the permissions of the participating parties; the knowledge base provides historical cases and risk knowledge for permission approval; and the digital twin simulation sandbox can also deduce and verify authorization strategies. This makes security no longer a rigid external barrier, but rather an adaptive immune capability flowing in the system's veins, comprehensively supporting the industrial intelligent agent collaborative decision-making system to achieve highly autonomous and flexible yet absolutely safe and reliable efficient operation and continuous optimization in complex, dynamic, and high-security industrial environments.

[0041] This invention discloses an industrial intelligent agent collaborative decision-making system and method, which solves this problem by constructing a hierarchical distributed system. The bottom-level data acquisition module and local intelligent agent module are responsible for perception and rapid response, while the upper-level collaborative decision-making platform module provides a virtual collaborative space that aggregates action proposals from various intelligent agents, performs conflict detection, and drives autonomous negotiation among the agents. For complex conflicts, a global optimization module initiates a distributed algorithm for coordination. A knowledge base module is integrated throughout the system, providing domain knowledge and historical experience to support decision-making and optimization. Through this mechanism of "autonomous negotiation as the primary method, global optimization as a secondary method, and continuous knowledge-driven approach," the system approaches the global optimum while maintaining flexibility, achieving efficient, adaptive, and evolvable industrial collaborative decision-making.

[0042] Therefore, the industrial intelligent agent collaborative decision-making system and method of the present invention solves the problems of low collaborative efficiency and difficulty in global optimization of industrial multi-agent collaboration.

[0043] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. An industrial intelligent agent collaborative decision-making system, characterized in that, include: The data acquisition module collects industrial field operating condition data in real time, including at least equipment status, production progress and material information, and generates dynamic operating condition signals. The local intelligent agent module receives the dynamic operating condition signal and generates an action proposal signal representing its own intention and resource needs based on the built-in decision model. The collaborative decision-making platform module receives the action proposal signal, aggregates and detects conflicts in the signals of multiple local intelligent agent modules, and forms and sends a negotiation instruction signal to the relevant local intelligent agent modules to initiate negotiation. A global optimization module, when it detects a conflict that it cannot resolve on its own, generates an optimization trigger signal and transmits it to the global optimization module. The global optimization module runs a distributed optimization algorithm based on the optimization trigger signal, generates a coordination strategy signal, and feeds it back to the collaborative decision-making platform module. The knowledge base module stores industrial knowledge graphs and historical collaborative cases, and provides optimization knowledge signals to the global optimization module and decision-making knowledge signals to the local intelligent agent module.

2. The industrial intelligent agent collaborative decision-making system according to claim 1, characterized in that, The data acquisition module includes sensor units deployed on physical devices and edge computing units located at the network edge. The sensor units sense multimodal raw physical signals such as vibration, temperature, current, and visual images during device operation in real time. The edge computing unit connects to the sensor units, receives the multimodal raw physical signals, and performs signal filtering, feature extraction, and format standardization processing to fuse heterogeneous raw data into a unified dynamic operating condition signal with timestamps and data source identifiers. The data acquisition module periodically or event-triggeredly pushes the dynamic operating condition signal to the local intelligent agent module through industrial network protocols, thereby providing a real-time, clean, and structured data foundation for subsequent decision-making.

3. The industrial intelligent agent collaborative decision-making system according to claim 1, characterized in that, The local intelligent agent module integrates a reaction layer, a collaborative planning layer, and a communication interface layer. The reaction layer, based on a pre-set rule base, provides millisecond-level responses to received emergency alarm-type dynamic condition signals, generating direct control commands. The collaborative planning layer receives regular dynamic condition signals and decision knowledge signals from the knowledge base module, and uses a model-based reinforcement learning algorithm to simulate the interaction consequences of its own action proposals with the possible behaviors of other intelligent agents, thereby generating optimized action proposal signals. The communication interface layer is responsible for encapsulating the action proposal signals into messages conforming to predetermined communication standards and sending them to the collaborative decision-making platform module. It is also responsible for receiving negotiation command signals and coordination strategy signals from external sources. The model-based reinforcement learning algorithm is as follows: ; in, Given the mathematical expectation, traverse the state space S and the agent's policy. Environmental Model ; / The industrial operating condition state of the intelligent agent at time t / t+1 is mapped by dynamic operating condition signals; / Propose actions for the agent at time t / t+1; For instant reward functions; The discount factor balances immediate returns with future returns; For action value networks, the value of the optimal future action is estimated; The interaction penalty coefficient balances the autonomous decision-making of agents with the constraints of multi-agent cooperation; To account for the multi-agent interaction loss, quantify the current agent action proposal. Other agent candidate action sets The degree of resource and time conflicts; the greater the conflict, the higher the loss value. For an agent policy network, output the probability distribution of action proposals under a given state; : An environmental model that simulates the state transition patterns of industrial operating conditions and multi-agent interactions.

4. The industrial intelligent agent collaborative decision-making system according to claim 1, characterized in that, The collaborative decision-making platform module internally implements a virtual collaborative space and a conflict detection engine; the virtual collaborative space, as a shared information exchange medium, continuously receives and broadcasts action proposal signals from each local intelligent agent module to form a global situational view. The conflict detection engine analyzes the global situation view in real time and identifies competitive and mutually exclusive conflicts in equipment, materials or production capacity by comparing the resource requirements, space occupancy and time windows in different action proposal signals. When the identified conflict meets the preset simple conflict rules, the conflict detection engine directly generates a negotiation instruction signal containing the identifiers of the conflicting parties and suggested solutions. When a conflict involving complex or circular dependencies is identified, the optimization trigger signal is generated.

5. The industrial intelligent agent collaborative decision-making system according to claim 1, characterized in that, Upon receiving the optimization trigger signal, the global optimization module initiates a distributed optimization algorithm based on game theory Nash equilibrium search, as follows: ； in, Adjust parameters (such as resource allocation, execution time window, and capacity allocation ratio) for the action proposals of n conflicting agents. These are the adjustment parameters for the i-th agent; Let be the global objective function, and let be the minimum objective function that integrates core industrial indicators. These are weighting coefficients, dynamically adjusted according to industrial production needs. The total energy consumption of the system is calculated by adding up the energy consumption of each agent's action proposal; It is a comprehensive quantitative indicator of the overall production efficiency of the system, including equipment utilization rate, task completion rate, and material flow efficiency. For a set of conflicting intelligent agents, Let be the feasible domain of the action parameters of the i-th agent; Let i be the revenue function of the i-th agent (such as equipment utilization rate, task completion revenue). Let be the parameters of all agents except i. This constraint is the core condition of Nash equilibrium: any agent cannot improve its own reward by adjusting its parameters alone, that is, all agents accept the current solution. As a hard constraint on industrial processes, the algorithm takes the action proposal signals from the conflicting parties gathered in the collaborative decision-making platform module as input, and uses the highest overall system production efficiency or lowest total energy consumption as the global objective function. Simultaneously, it incorporates the individual benefit constraints of each local intelligent agent module into the computational framework. The global optimization module calls the optimization knowledge signals provided by the knowledge base module as prior information for the algorithm, and seeks a Pareto improvement solution acceptable to all conflicting parties through multiple rounds of iterative calculation. Finally, the calculated resource reallocation scheme or execution sequence adjustment scheme is encapsulated as the coordination strategy signal and sent back to the collaborative decision-making platform module for distribution. The formula for determining the Pareto improvement solution is as follows: ； in, The set of changes in payoffs for all conflicting agents. The change in the reward of the i-th agent; This represents the industrial tolerance threshold. This represents the change in the global objective function. <0 indicates that the overall cost has decreased compared to before optimization, meaning that the global objective has been optimized.

6. The industrial intelligent agent collaborative decision-making system according to claim 1, characterized in that, The knowledge base module includes an industrial knowledge graph sub-base and a collaborative case base. The industrial knowledge graph sub-base stores equipment process associations, material conversion relationships, and production constraint rules in a graph structure. The decision knowledge signals it provides are used to guide the local intelligent agent module to generate proposals that conform to the process route, and the optimization knowledge signals it provides are used to limit the feasible strategy search space for the global optimization module. The collaborative case base records historical collaborative processes and results in a structured log format, including the characteristics of successful cases and the root causes of failed cases. Through machine learning models, cases are clustered and labeled, enabling the provided knowledge signals to be matched and pushed based on the similarity of the current dynamic working conditions, thereby realizing the reuse and inheritance of experience knowledge.

7. The industrial intelligent agent collaborative decision-making system according to claim 1, characterized in that, The action proposal signal generated by the local intelligent agent module has a standardized data structure, which includes at least the identity of the proposing intelligent agent, the description of the proposal target, the list of required resources and their quantity and time requirements, the expected start and end timestamps, the proposal priority weight, and the proposal validity period. This structured signal design enables the collaborative decision-making platform module to perform mechanical analysis and comparison of proposals, and also provides clear optimization variables and constraints for the global optimization module, thereby ensuring the unambiguity of information transmission and the efficiency of processing between different modules.

8. The industrial intelligent agent collaborative decision-making system according to claim 1, characterized in that, The system also includes a digital twin simulation sandbox module, which is connected to the collaborative decision-making platform module and the local intelligent agent module. The digital twin simulation sandbox module receives global situational view signals from the collaborative decision-making platform module and copies of decision models from each local intelligent agent module. It performs accelerated simulation and stress testing of various collaborative strategies in a virtual environment, and feeds back the performance evaluation signals and potential risk warning signals obtained from the simulation to the collaborative decision-making platform module and relevant local intelligent agent modules. This is used to verify and optimize decisions before actual execution, forming a closed loop of "decision-simulation-optimization".

9. The industrial intelligent agent collaborative decision-making system according to claim 1, characterized in that, The system also includes a security and authorization management module, which is responsible for the identity authentication, access control, and operation auditing of all intelligent agents within the system. It supports a three-tiered control architecture of "global-regional-local," enabling dynamic permission allocation based on three dimensions: "resource-operation-scenario," and is linked to the intelligent agent's credit rating to construct a closed-loop authorization mechanism covering the entire process of "application-approval-execution-monitoring-revocation." The security and authorization management module includes an identity authentication unit, an access control unit, and an audit trail unit. The identity authentication unit implements two-way authentication of intelligent agents based on digital certificates and dynamic tokens. The access control unit supports a three-tiered control mode of "global-regional-local," wherein... The global level is responsible for system strategy formulation and cross-regional coordination; the regional level is responsible for resource and task scheduling within its own region; and the local level is responsible for device-level operation permission management. The permission control unit also implements a permission model based on three dimensions: "resource-operation-scenario". The resource dimension includes device, data, and network resources; the operation dimension includes read / write, control, negotiation, and optimization; and the scenario dimension includes normal production, emergency response, and maintenance and debugging. Permission allocation is dynamically linked to the agent's historical credit rating, with agents with higher credit ratings enjoying more lenient permissions and higher negotiation weight. The audit and tracing unit records the operation logs, permission change records, and abnormal behaviors of all agents, supporting real-time monitoring and post-event traceability.

10. A method for an industrial intelligent agent collaborative decision-making system according to claims 1-9, comprising: S1: Real-time acquisition of industrial site operating condition data, including at least equipment status, production progress, and material information, and formation of dynamic operating condition signals; S2: Receive the dynamic operating condition signal and generate an action proposal signal representing its own intentions and resource needs based on the built-in decision model; S3: Receive the action proposal signal, aggregate and detect conflicts in the signals of multiple local intelligent agent modules, form and send a negotiation instruction signal to the relevant local intelligent agent modules to initiate negotiation; S4: When a conflict that cannot be resolved autonomously is detected, an optimization trigger signal is generated and transmitted to the global optimization module. The global optimization module runs a distributed optimization algorithm based on the optimization trigger signal, generates a coordination strategy signal, and feeds it back to the collaborative decision-making platform module. S5: Store industrial knowledge graphs and historical collaborative cases, and provide optimization knowledge signals to the global optimization module and decision-making knowledge signals to the local intelligent agent module.

Images