Semiconductor factory affair ai digital employee system and method based on multi-agent cooperation

Abstract

This invention provides a semiconductor plant AI digital employee system based on multi-agent collaboration. The system adopts a hierarchical multi-agent architecture, including: an anomaly perception and task generation module; a task dynamic allocation module; a collaborative scheduling and execution guidance module; and a closed-loop feedback and profile update module. This module dynamically updates the digital profile parameters of the corresponding AI digital employee agent based on the actual execution results and response indicators after task execution, and structurally stores the new processing cases in the semiconductor plant domain knowledge graph. This invention also provides a semiconductor plant digital employee task allocation method based on multi-agent collaboration, including the following steps: S1) plant status perception and task generation; S2) task dynamic broadcasting and autonomous bidding; S3) optimal task allocation and execution guidance; and S4) operation and maintenance feedback and continuous learning of the digital profile.

Description

Technical Field

[0001] This invention relates to the field of semiconductor intelligent manufacturing technology, specifically to an AI digital employee system for semiconductor plants based on multi-agent collaboration. Background Technology

[0002] Semiconductor plant management systems refer to the auxiliary systems in a wafer fabrication plant that support the production process environment, including HVAC systems, ultrapure water supply systems, special gas supply systems, and power distribution systems. These subsystems are strongly coupled, and a single anomaly can easily trigger a cascading failure, placing high demands on the timeliness of maintenance response and the ability to coordinate cross-subsystem issues.

[0003] Currently, semiconductor plant maintenance mainly adopts the following methods: First, there's the plant monitoring solution based on digital twins. While it achieves status visualization and fault early warning by constructing digital twin models of equipment, task execution still relies on manual dispatching or pre-set scripts, lacking the ability to dynamically optimize scheduling based on real-time on-site conditions.

[0004] Second, a multi-agent monitoring scheme based on centralized scheduling. Each agent is responsible for collecting subsystem operation data and reporting it to the central scheduling node for task assignment. In this architecture, the central node is prone to becoming a performance bottleneck, and no autonomous negotiation mechanism between agents is introduced. Task allocation does not comprehensively consider the dynamic balance of skill matching, personnel location, and task urgency.

[0005] Third, the anomaly diagnosis and work order generation solution using a single AI model. While AI models can identify anomalies in sensor data and generate maintenance work orders, they can only handle independent anomalies in a single subsystem. The handling strategies are fixed in the model parameters and cannot be continuously learned online. Furthermore, manual intervention is still required in the task allocation process after task generation.

[0006] In summary, existing semiconductor plant operation and maintenance solutions suffer from problems such as limited architectural scalability, lack of collaborative negotiation mechanisms between intelligent agents, lack of multi-dimensional optimization in task allocation, and rigid knowledge that is difficult to evolve. Summary of the Invention

[0007] To address the following issues in existing semiconductor plant operation and maintenance solutions: the centralized scheduling architecture has limited scalability, and the central node is prone to becoming a performance bottleneck as the equipment scales up; there is a lack of autonomous collaborative negotiation mechanisms between intelligent agents, making it difficult to achieve dynamic task alliances across subsystems; task allocation does not comprehensively consider multiple factors such as skill matching, response time and load status, and resource allocation efficiency needs to be improved; and anomaly handling knowledge is fixed in static rules or model parameters, lacking continuous evolution capabilities.

[0008] This invention provides an AI-powered digital employee system for semiconductor plant operations based on multi-agent collaboration. The system adopts a hierarchical multi-agent architecture and includes: An anomaly perception and task generation module is deployed at the device layer intelligent agent. It is used to collect sensor data of plant physical equipment in real time and identify the operating status through an anomaly detection model. When it is determined that manual intervention is required, it submits a task request to the superior coordinating intelligent agent. The task request includes at least task attribute information. A task dynamic allocation module is deployed on the coordinating agent and multiple AI digital employee agents. The coordinating agent is used to encapsulate the task request into a task bidding message and broadcast it. Each AI digital employee agent is used to receive the bidding message and autonomously calculate the bid value based on its own digital profile to participate in the bidding. The bid value comprehensively represents at least the relevant parameters of task execution cost, including capability matching degree, response timeliness and current occupancy status. The collaborative scheduling and execution guidance module is used by the coordinating agent to select the target AI digital employee agent based on the feedback bid value and issue execution instructions; the winning target AI digital employee agent calls the semiconductor plant domain knowledge graph to perform reasoning, generate standard operating procedures, or generate standard operating procedures and execution work orders to output operation and maintenance guidance information; The closed-loop feedback and profile update module is used to dynamically update the digital profile parameters of the corresponding AI digital employee agent based on the actual execution results and response indicators after the task is completed, and to structurally store the new processing cases into the semiconductor plant domain knowledge graph.

[0009] This invention also provides a method for task allocation of digital employees in semiconductor plants based on multi-agent collaboration, comprising the following steps: S1) Plant status perception and task generation: The equipment layer intelligent agent collects sensor data of the plant's physical equipment in real time to detect anomalies. When it is determined that manual intervention is required, it submits a task request carrying task attribute information to the superior coordinating intelligent agent. S2) Task dynamic broadcasting and autonomous bidding: The coordinating agent encapsulates the task request into a bidding message and broadcasts it to the corresponding AI digital employee agent pool; each AI digital employee agent autonomously calculates the bid value based on its own digital profile and provides feedback. The bid value comprehensively represents at least the relevant parameters of task execution cost, including capability matching degree, response timeliness and current occupancy status. S3) Optimal allocation and execution guidance of tasks: The coordinating agent selects the target AI digital employee agent based on the bid values ​​fed back by each agent and issues instructions; the target AI digital employee agent calls the semiconductor plant domain knowledge graph to generate standard operating procedures, or generates standard operating procedures and execution work orders to output operation and maintenance guidance information; S4) Operation and Maintenance Feedback and Continuous Learning of Digital Profiling: After the task is completed, the system updates the digital profile parameters of the target AI digital employee agent based on the actual execution results and stores the new processing cases in the domain knowledge graph.

[0010] Furthermore, a long short-term memory network model is used for anomaly detection; the coordinating agent and the AI ​​digital employee agent use a contract network protocol or auction mechanism to broadcast and bid for tasks.

[0011] Furthermore, the specific formula for calculating the bid value by the AI ​​digital employee intelligent agent is as follows: Bid=αSkillMatch+β(1 / TravelTime)-γCurrentLoad; Wherein, Bid is the bid value, SkillMatch is the skill matching degree of the AI ​​digital employee, TravelTime is the estimated travel time of the AI ​​digital employee from its current physical location to the location of the abnormal device, CurrentLoad is the current load of the AI ​​digital employee, and α, β, and γ are dynamic weighting coefficients.

[0012] Furthermore, the dynamic weighting coefficients α, β, and γ are adaptively adjusted according to the urgency of the task request; when the urgency exceeds the preset high-risk threshold, the system automatically increases the proportion of the weighting coefficient β corresponding to the estimated travel time.

[0013] Furthermore, the winning AI digital employee agent will provide visual operational guidance to on-site maintenance personnel by driving robotic process automation tools or augmented reality terminals.

[0014] Furthermore, it also includes a global scheduling agent; when a detected abnormal event affects the plant's physical equipment in multiple different subsystems, the global scheduling agent identifies the cascading fault and issues cross-subsystem collaborative handling instructions to the coordinating agents of each relevant subsystem.

[0015] Furthermore, a reinforcement learning algorithm is employed, with the reward function being the minimization of the overall cost of task allocation, to iteratively optimize the dynamic weight coefficients of the AI ​​digital employee agent when calculating the bid value online.

[0016] Furthermore, a fault-tolerant backup mechanism is also configured: if a bid value that meets the minimum requirements of the task is not received within the preset time, or if the winning AI digital employee agent does not confirm acceptance of the task within the predetermined time, the system will automatically start a high-priority second bidding process; if bidding fails continuously, the task will be forcibly attached to the manual control console and an alarm will be triggered.

[0017] Furthermore, when the type of the task request is determined to be a pure software operation task, the TravelTime parameter in the bid value calculation formula is replaced with a constant minimum value or a network response latency indicator; the method also includes: when calling the plant management underlying software control interface, performing security verification on the operation command based on a dynamic permission token mechanism, wherein the dynamic permission token is generated by the global scheduling agent for a specific task and is only valid within a preset effective time window and a specified device address range.

[0018] The present invention has the following beneficial effects: 1. It adopts a hierarchical multi-agent architecture and a distributed bidding mechanism to avoid centralized scheduling bottlenecks and has strong system scalability.

[0019] 2. Digital employees bid independently based on their own profiles, taking into account factors such as ability matching, response time and workload, resulting in fast and accurate task allocation.

[0020] 3. By combining domain knowledge graphs with global scheduling agents, collaborative handling of cascading faults across subsystems can be supported.

[0021] 4. The task execution results are fed back to update the digital profile and knowledge graph, enabling continuous learning and accumulation of operation and maintenance knowledge. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the module structure of a semiconductor plant digital employee dynamic task allocation system based on multi-agent collaboration, provided as an embodiment of this application.

[0023] Figure 2 This is a flowchart illustrating a method for dynamic task allocation of digital employees in a semiconductor plant based on multi-agent collaboration, provided as an embodiment of this application.

[0024] Figure 3 This is a schematic diagram of data flow and control flow in a hierarchical multi-agent architecture provided in an embodiment of this application.

[0025] Figure 4 This is a schematic diagram illustrating the timing of task bidding and tendering interactions between a coordinating agent and a digital employee agent, provided as an embodiment of this application. Detailed Implementation

[0026] This invention provides an AI-powered digital employee system for semiconductor plant operations based on multi-agent collaboration. The system adopts a hierarchical multi-agent architecture and includes: An anomaly perception and task generation module is deployed at the device layer intelligent agent. It is used to collect sensor data of plant physical equipment in real time and identify the operating status through an anomaly detection model. When it is determined that manual intervention is required, it submits a task request to the superior coordinating intelligent agent. The task request includes at least task attribute information. A task dynamic allocation module is deployed on the coordinating agent and multiple AI digital employee agents. The coordinating agent is used to encapsulate the task request into a task bidding message and broadcast it. Each AI digital employee agent is used to receive the bidding message and autonomously calculate the bid value based on its own digital profile to participate in the bidding. The bid value comprehensively represents at least the relevant parameters of task execution cost, including capability matching degree, response timeliness and current occupancy status. The collaborative scheduling and execution guidance module is used by the coordinating agent to select the target AI digital employee agent based on the feedback bid value and issue execution instructions; the winning target AI digital employee agent calls the semiconductor plant domain knowledge graph to perform reasoning, generate standard operating procedures, or generate standard operating procedures and execution work orders to output operation and maintenance guidance information; The closed-loop feedback and profile update module is used to dynamically update the digital profile parameters of the corresponding AI digital employee agent based on the actual execution results and response indicators after the task is completed, and to structurally store the new processing cases into the semiconductor plant domain knowledge graph.

[0027] This invention also provides a method for task allocation of digital employees in semiconductor plants based on multi-agent collaboration, comprising the following steps: S1) Plant status perception and task generation: The equipment layer intelligent agent collects sensor data of the plant's physical equipment in real time to detect anomalies. When it is determined that manual intervention is required, it submits a task request carrying task attribute information to the superior coordinating intelligent agent. S2) Task dynamic broadcasting and autonomous bidding: The coordinating agent encapsulates the task request into a bidding message and broadcasts it to the corresponding AI digital employee agent pool; each AI digital employee agent autonomously calculates the bid value based on its own digital profile and provides feedback. The bid value comprehensively represents at least the relevant parameters of task execution cost, including capability matching degree, response timeliness and current occupancy status. S3) Optimal allocation and execution guidance of tasks: The coordinating agent selects the target AI digital employee agent based on the bid values ​​fed back by each agent and issues instructions; the target AI digital employee agent calls the semiconductor plant domain knowledge graph to generate standard operating procedures, or generates standard operating procedures and execution work orders to output operation and maintenance guidance information; S4) Operation and Maintenance Feedback and Continuous Learning of Digital Profiling: After the task is completed, the system updates the digital profile parameters of the target AI digital employee agent based on the actual execution results and stores the new processing cases in the domain knowledge graph.

[0028] Furthermore, a long short-term memory network model is used for anomaly detection; the coordinating agent and the AI ​​digital employee agent use a contract network protocol or auction mechanism to broadcast and bid for tasks.

[0029] Furthermore, the specific formula for calculating the bid value by the AI ​​digital employee intelligent agent is as follows: Bid=αSkillMatch+β(1 / TravelTime)-γCurrentLoad; Wherein, Bid is the bid value, SkillMatch is the skill matching degree of the AI ​​digital employee, TravelTime is the estimated travel time of the AI ​​digital employee from its current physical location to the location of the abnormal device, CurrentLoad is the current load of the AI ​​digital employee, and α, β, and γ are dynamic weighting coefficients.

[0030] Furthermore, the dynamic weighting coefficients α, β, and γ are adaptively adjusted according to the urgency of the task request; when the urgency exceeds the preset high-risk threshold, the system automatically increases the proportion of the weighting coefficient β corresponding to the estimated travel time.

[0031] Furthermore, the winning AI digital employee agent will provide visual operational guidance to on-site maintenance personnel by driving robotic process automation tools or augmented reality terminals.

[0032] Furthermore, it also includes a global scheduling agent; when a detected abnormal event affects the plant's physical equipment in multiple different subsystems, the global scheduling agent identifies the cascading fault and issues cross-subsystem collaborative handling instructions to the coordinating agents of each relevant subsystem.

[0033] Furthermore, a reinforcement learning algorithm is employed, with the reward function being the minimization of the overall cost of task allocation, to iteratively optimize the dynamic weight coefficients of the AI ​​digital employee agent when calculating the bid value online.

[0034] Furthermore, a fault-tolerant backup mechanism is also configured: if a bid value that meets the minimum requirements of the task is not received within the preset time, or if the winning AI digital employee agent does not confirm acceptance of the task within the predetermined time, the system will automatically start a high-priority second bidding process; if bidding fails continuously, the task will be forcibly attached to the manual control console and an alarm will be triggered.

[0035] Furthermore, when the type of the task request is determined to be a pure software operation task, the TravelTime parameter in the bid value calculation formula is replaced with a constant minimum value or a network response latency indicator; the method also includes: when calling the plant management underlying software control interface, performing security verification on the operation command based on a dynamic permission token mechanism, wherein the dynamic permission token is generated by the global scheduling agent for a specific task and is only valid within a preset effective time window and a specified device address range.

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0037] Figure 1 This illustration shows a schematic diagram of the module structure of a dynamic task allocation system for digital employees in a semiconductor plant based on multi-agent collaboration, according to an embodiment of this application. Figure 1 As shown, the system includes: an anomaly detection and task generation module, a dynamic task allocation module, a collaborative scheduling and execution guidance module, and a closed-loop feedback and profile update module. These modules are deployed in a hierarchical multi-agent architecture, with agents at each layer interacting via industrial Ethernet or an internal plant communication bus.

[0038] Figure 3 This illustration shows a schematic diagram of data flow and control flow in a hierarchical multi-agent architecture provided in an embodiment of this application. Combined with... Figure 3 This hierarchical multi-agent architecture, from bottom to top, includes device-level agents, subsystem coordination agents, and optional global scheduling agents. Device-level agents are located on the physical equipment side of each plant, responsible for collecting sensor data and reporting anomalies and task requests (such as...). Figure 3 (As shown by the solid arrows from bottom to top). Subsystem coordinating agents are deployed according to plant subsystems, such as HVAC (Heating, Ventilation and Air Conditioning) subsystem coordinating agents, power subsystem coordinating agents, ultrapure water subsystem coordinating agents, and special gas subsystem coordinating agents, etc., responsible for issuing execution instructions to equipment-level agents (such as...). Figure 3 (As shown by the solid arrows from top to bottom). When a cascading failure occurs across subsystems, the subsystem coordinating agent reports a coordination request to the global scheduling agent, which then issues coordination instructions to the relevant subsystem coordinating agents. The right side of the system also displays a semiconductor plant domain knowledge graph and a digital employee profile database for use by agents at each layer.

[0039] The following combination Figure 1 The module structure shown provides a detailed explanation of the specific implementation methods of each module.

[0040] I. Anomaly Detection and Task Generation Module

[0041] The anomaly detection and task generation module is deployed at the device-level intelligent agent. These agents are located on the physical equipment side of the plant, such as cooling water circulation pumps in the HVAC system, reverse osmosis membrane modules in the ultrapure water system, and gas holder valves in the special gas supply system. Each device-level intelligent agent establishes data connections with the sensors and actuators of its corresponding equipment, collecting real-time sensor data such as vibration, temperature, pressure, flow rate, current, and valve opening. The sampling frequency for critical equipment can be set to once per second, while for general equipment it is once every five seconds.

[0042] This module incorporates a lightweight anomaly detection model. In one embodiment, this model is implemented using a Long Short-Term Memory (LSTM) network, a variant of recurrent neural networks suitable for time-series data, which can effectively capture the time-dimensional dependencies of sensor data. Before deployment, the LSTM model is trained using historical normal operation data of the corresponding device. The training data includes sensor time-series data from at least one month of continuous operation under different load rates and ambient temperatures. The model learns the distribution characteristics of the mean, variance, and periodic fluctuation patterns of the normal operation state. During operation, the module inputs the real-time collected sensor data into the LSTM model in the form of a sliding time window, with a window length of, for example, 60 sampling points. The model outputs a judgment result indicating whether the operating state deviates from the normal pattern and an anomaly confidence score between 0 and 1, with a higher score indicating a greater probability of an anomaly.

[0043] When an anomaly is detected, this module determines whether to generate a task request to seek human or digital employee intervention based on the following criteria, using a comprehensive judgment based on both handling confidence level and physical constraints: First, physical operation constraint determination. This module queries the attribute tags of plant equipment, which predefine the operation mode type of each equipment component. If an abnormally associated equipment component is marked as a physical interaction component in the knowledge graph, it is automatically determined that manual intervention is required. Physical interaction components include physical consumables that need to be manually replaced, such as air filters and purification resin columns; pipes or chambers that need to be manually cleaned; mechanical valves or circuit breakers that need to be manually reset on-site; and mechanical wear parts that need to be visually inspected. These tasks cannot be completed by pure software commands.

[0044] Second, the logical confidence threshold is determined. The inference unit within this module attempts to generate an automated handling plan based on a preset rule base. The rule base includes automated response strategies for common anomaly types, such as outputting instructions to increase the cooling tower fan speed or activate the backup cooling circuit when a high cooling water temperature is detected. The inference unit outputs a success confidence score P for the generated automated plan. This score comprehensively considers the rule matching degree, the success rate of similar historical scenarios, and the degree of consistency between the current equipment status and the applicable conditions of the plan. If P is lower than a preset threshold, such as 0.85, the anomaly is determined to be beyond the scope of automated handling capabilities, requiring manual intervention.

[0045] Third, the system performs feedback anomaly detection. If the module has triggered automated handling, such as executing parameter adjustment commands by calling an RPA script, the system will continuously monitor relevant sensor values. If, within a preset observation time window (e.g., 30 seconds to 2 minutes after parameter adjustment), the relevant sensor values ​​do not return to the normal operating range, the automated handling is deemed a failure and requires manual intervention. For example, if, after executing a command to increase the cooling tower fan speed, the cooling water temperature does not drop below the set value within two minutes, the self-healing process is deemed a failure.

[0046] Once intervention is deemed necessary, the module generates a task request and submits it to the coordinating agent of its superior subsystem. The task request includes task attribute information, specifically: task type (e.g., repair, replacement, calibration, inspection, cleaning); physical location identifier of the abnormal equipment (e.g., plant area number, floor number, equipment tag number, area grid code); task urgency rating (e.g., categorized by impact on production processes, from 1 to 5, with 5 being the highest urgency indicating potential production line shutdown); and required skill tags for handling the task (e.g., HVAC - inverter commissioning, special gas - gas holder valve replacement, ultrapure water - reverse osmosis membrane cleaning, power - low-pressure cabinet operation). The task attribute information may also include a brief description of the abnormal phenomenon and relevant sensor reading snapshots as supplementary information.

[0047] II. Dynamic Task Allocation Module

[0048] The task dynamic allocation module is deployed in the subsystem coordinating agent and multiple AI digital employee agents.

[0049] The subsystem coordination intelligent agent is deployed according to the plant management subsystem, manages the equipment layer intelligent agents and callable AI digital employee intelligent agents in its subsystem, and maintains the digital employee registry to record online status, capability tags and load information.

[0050] Figure 4 This diagram illustrates the timing sequence of task bidding and tendering interactions between the coordinating agent and the digital employee agent. Combined with... Figure 4 The specific interaction process is as follows: Upon receiving a task request, the subsystem coordinating agent encapsulates it into a task bidding message conforming to the Contract Network Protocol (CNP) format. The Contract Network Protocol is a distributed task allocation protocol where task publishers broadcast their requirements, potential executors assess their capabilities and submit bids, and the publisher selects the best bidders to award the task contract. The bidding message includes a bidding identifier, a summary of task attribute information, a bid deadline (e.g., within 30 seconds), and minimum bid requirements such as a list of mandatory skill tags. In an alternative embodiment, the CNP protocol can be replaced by an auction mechanism.

[0051] like Figure 4As shown in step 1, the subsystem coordinating agent broadcasts the tender message to eligible digital employees. The broadcast range is filtered based on task attributes; for example, if the task skill tag is HVAC-inverter commissioning, it will only be sent to digital employees with that tag; low-priority inspection tasks can be broadcast to all available digital employees.

[0052] The AI-powered digital employee agent maintains its own digital profile, which is dynamically updated. This profile includes at least a set of capability tags such as RPA_Control, Human_Binding, HVAC-Expert, and Power-Junior; its current physical location or network access node; its current load status (the ratio of completed tasks to queue capacity); and historical task completion quality scores such as success rate and average response time. Digital employees are not limited to static classifications and can possess multiple capability tags simultaneously. For example, a digital employee bound to a plant engineer can have both remote software operation capabilities and on-site AR guidance capabilities, with additional professional domain tags.

[0053] During task broadcasting, the subsystem coordinating agent filters objects based on the Boolean field requiring physical operations in the task attributes. When this field is true, the tender message is sent only to online digital employees who are bound to real personnel; when this field is false, the message is broadcast to digital employees with RPA capabilities.

[0054] like Figure 4 As shown in step 2, after receiving the bidding message, the digital employee independently calculates the bid value (Bid) based on their own profile. A higher bid value indicates greater suitability for the task. The formula for calculating the bid value is: Bid = α·SkillMatch + β·(1 / TravelTime) - γ·CurrentLoad SkillMatch represents the degree of match between ability tags and task skill tags, with a value ranging from 0 to 1. If the mandatory skill tag is fully included, the basic match degree is 1. Partial matches are proportionally reduced and multiplied by a discount factor, such as 0.8. Additional points are awarded for having extended skill tags.

[0055] TravelTime represents the estimated travel time, in minutes, from the current location to the location of the malfunctioning device. For digital employees linked to real personnel, the real-time location of the person is obtained based on indoor positioning systems such as UWB, Wi-Fi fingerprinting, or Bluetooth beacons. The shortest path distance to the target device is calculated and divided by a preset walking speed, such as 80 meters per minute. If located outdoors in a cleanroom, the air shower passage time of approximately 30 seconds needs to be added. The TravelTime value for purely software-based digital employees is explained below.

[0056] CurrentLoad represents the current load, quantified as the ratio of the number of accepted but unfinished tasks to the maximum capacity of the task queue, with a value ranging from 0 to 1.

[0057] α, β, and γ are dynamic weighting coefficients that control the relative importance of skill matching, response timeliness, and load status, respectively. Initial values ​​are obtained through multi-objective optimization training on historical maintenance work order data. Task attributes, assigned personnel, completion time, and quality evaluations of work orders from the past 6 months to 1 year are collected. The goal is to minimize response time, minimize travel distance, and maximize the first-time success rate. Genetic algorithms or particle swarm optimization are used to find the optimal combination, resulting in values ​​such as α=0.5, β=0.3, and γ=0.2.

[0058] During system operation, the weighting coefficients are adaptively adjusted based on task urgency and system operating conditions. In high-load mode, when the backlog of pending tasks exceeds a threshold, such as when the number of waiting tasks exceeds twice the total number of digital employees, the system automatically increases the γ weight by multiplying it by an amplification factor of 1.5 to 2 to strengthen load balancing and give priority to digital employees with lighter loads.

[0059] In emergency fault mode, when the urgency score reaches level 4 or 5, the system automatically increases the β weight by multiplying it by a magnification factor of 2 to 3, ensuring that the qualified digital employee closest to the fault point wins the bid, achieving rapid loss mitigation at the nearest location. For level 5 tasks such as special gas leaks and severe power fluctuations, the β weight can temporarily rise to a dominant position.

[0060] During periods of low manpower, such as night shifts, weekends, or holidays, the system automatically increases the alpha weight to ensure that highly skilled matching digital employees are selected, reducing the probability of rework. This adjustment can be dynamically triggered based on the time schedule or the number of online digital employees.

[0061] like Figure 4 As shown in step 3, digital employees provide their bid values, identifiers, and signatures before the bid deadline. In the example, digital employee A submits 0.9, and digital employee B submits 0.5.

[0062] For purely software operation tasks, since there is no physical movement, when the field requiring physical operation is false, the system replaces TravelTime with a constant minimum value ε, such as 0.001, rendering this item ineffective in the calculation. In an alternative embodiment, TravelTime is redefined as a network response latency metric, i.e., the round-trip time (RTT) between the compute node and the target controller, in milliseconds, prioritizing the software instance with the lowest network latency.

[0063] This module is configured with a fault-tolerant fallback mechanism, including three levels: Task timeout triggers reselection. If no bid meeting the minimum requirements is received within the bidding waiting period, or the winning employee fails to confirm acceptance of the task within the scheduled time, or the bound personnel's location does not move towards the target, the system automatically revoks the winning bid, initiates a high-priority second bidding process, shortens the deadline to 15 seconds, and marks it as priority-upgraded. The winning bid qualification is frozen, for example, for 5 minutes.

[0064] Forced task preemption. For tasks with the highest urgency, the system grants the right to forcibly interrupt them. If all eligible digital employees are busy, their current task priorities can be assessed, low-priority tasks can be forcibly withdrawn, digital employees can be reset to idle status and assigned high-priority tasks, and withdrawn tasks will enter a re-bidding process. The preemption operation will notify the preempted party and the central control center.

[0065] Manual takeover mode. If two consecutive rounds of bidding fail, the task is marked as a failed bid, attached to the on-duty engineer's panel in the plant control center, highlighted in red, and triggers an audible and visual alarm, prompting manual intervention. The on-duty engineer can manually assign the task or handle it themselves.

[0066] III. Collaborative Scheduling and Execution Guidance Module

[0067] The collaborative scheduling and execution guidance module is deployed in the subsystem to coordinate the intelligent agent and the winning AI digital employee intelligent agent.

[0068] like Figure 4 As shown in step 4, after the bidding deadline, the subsystem coordinates the intelligent agent to collect all bid values ​​and select the digital employee with the highest bid value as the target AI digital employee. If multiple digital employees have the same highest bid value, a preset secondary rule is used for selection, such as prioritizing those with higher historical task success rates or random selection. Figure 4 As shown in step 5, after selecting the target digital employee, the module issues a winning bid instruction and complete task information to it, including all fields of the task attributes and a snapshot of sensor readings attached to the anomaly detection module.

[0069] like Figure 4 As shown in step 6, the winning digital employee uses a semiconductor plant management knowledge graph to generate a solution. This knowledge graph is a graph database that structurally represents plant management expertise. It stores information such as Standard Operating Procedures (SOPs), equipment operation manuals, historical failure cases and handling experience, and Bill of Materials (BOM) for equipment components in the form of entity-relationship-attribute triples. For example, entity "cooling water circulation pump A" is connected to entity "bearing overheat" through the relationship "failure mode," and entity "bearing overheat" is connected to entity "bearing replacement SOP" through the relationship "recommended handling steps." The knowledge graph can use graph databases such as Neo4j or JanusGraph.

[0070] Digital employees input task type, abnormal equipment model, and description of the abnormal phenomenon, and use a graph inference engine to retrieve matching sequences of handling steps. The inference engine can employ rule-based reasoning, such as retrieving the process with SOP number Z if the equipment model is X and the abnormal phenomenon is Y, based on predefined rules; or embedding-based reasoning, mapping entities and relationships to a low-dimensional vector space and discovering associated handling solutions through vector similarity. The inference result is an ordered sequence of operation steps, each step including an action description, the object being operated on, the expected result, and safety precautions. Based on this, digital employees generate standard operating procedures and can generate execution work orders. Work orders are structured data for plant management systems such as Computerized Maintenance Management Systems (CMMS), containing fields such as work order number, task description, executor, and planned time.

[0071] After generating the operation steps, this module selects different technical paths based on the nature of the task and outputs operation and maintenance guidance information: First, purely software operation tasks. These include remotely adjusting inverter parameters, resetting the controller, switching backup circuits, modifying setpoints, and downloading configuration files. The digital employee maps logical actions to executable instructions. For upper-level plant software that provides an Application Programming Interface (API), this is mapped to an API call sequence, including endpoint addresses, request methods, parameters, and authentication tokens. For legacy systems that do not provide APIs, Robotic Process Automation (RPA) technology is used, mapping to graphical user interface (GUI) operation paths. The digital employee maintains a GUI element identifier library, recording the XPath paths, Automation IDs, or image templates of controls such as buttons, input boxes, and drop-down menus in each software interface. This drives RPA tools such as UiPath and Automation Anywhere to sequentially locate elements and simulate clicks, inputs, and other operations, automatically completing software-level control.

[0072] Second, on-site maintenance personnel execute tasks. Digital employees utilize spatial anchoring technology to break down operational steps into spatial coordinates and action commands. First, they obtain the spatial coordinates of the target equipment within the factory area from a knowledge graph or equipment management database, such as 3D coordinates extracted from Building Information Modeling (BIM). For each operational action, such as rotating a valve 90 degrees clockwise, pressing a reset button for three seconds, or observing a pressure gauge reading within the range of 0.4 to 0.6 MPa, it is associated and labeled with the spatial pose of the corresponding operating part of the equipment. The pose of the operating part can be pre-entered through 3D scanning or manual calibration. The digital employee combines the action text, target pose, and parameters into a visual command data package and sends it to the rendering engine.

[0073] The rendering engine, deployed on edge computing nodes or in the cloud, generates visual guidance for augmented reality (AR) terminals based on instructions. The generation process includes: generating virtual indicators based on the operation type (e.g., an arc-shaped arrow and target angle marker for rotation, a highlighted border and countdown ring for pressing); 3D registration and alignment of the virtual indicators with the target pose; and calculating the 2D projection position of the indicators on the AR terminal screen from the current viewpoint. After rendering, the video stream or 3D data is sent via Wi-Fi 6 or a 5G private network to the AR terminal worn by maintenance personnel, such as Microsoft HoloLens, RealWear industrial AR glasses, or handheld tablets. The AR terminal overlays the virtual indicators onto the real equipment image, highlighting the operating components with highlighted borders, indicating direction with animated arrows, and indicating the sequence with numerical codes, guiding accurate execution step by step. For example, maintenance personnel using AR glasses might see a cooling water pump valve surrounded by a green highlighted border, with an overlaid clockwise rotating arrow animation and text prompting to rotate 90 degrees clockwise.

[0074] This module also includes a security verification mechanism to perform security verification when digital employees call the plant's underlying software control interface, preventing accidents caused by accidental operations or malicious commands. The mechanism includes two layers of protection: dynamic permission tokens and a blacklist for blocking. A dynamic permission token mechanism is implemented. Digital employees who win a bid do not have permanent access. The collaborative scheduling module initiates a permission request to the scheduling agent. The scheduling agent then calls the production control system's permission interface for the task, requesting the generation of a one-time, single-operation token. The token is an encrypted string with time and scope restrictions, such as a JSON Web Token. It is bound to a valid time window (e.g., valid for five minutes after generation), a specified device address range (e.g., a specific IP segment, controller ID, or register address), and a whitelist of allowed operation types (e.g., only allowing parameter writing, prohibiting firmware upgrades or configuration resets). When a digital employee calls the underlying interface, the token must be included in the request header. Execution only proceeds after the interface verifies the token's validity, timeliness, and scope. Tokens expire automatically and require a new request for subsequent operations.

[0075] Security blacklist interception. Before executing RPA scripts or API call instructions generated by digital employees, the security verification module performs a static scan. The module has a built-in blacklist rule library, defining sensitive operations that are absolutely prohibited from execution in production. Rules use regular expressions or keyword matching, such as instructions containing keywords like shutdown, power_off, kill_process, reset_factory, or write operation instructions whose target address is a critical equipment controller, such as the main cooling water pump controller or the process chamber power supply contactor. If a blacklist match is found, the instruction is intercepted and not executed, a high-level alarm is sent to the central control center, and an audit log is recorded. Instructions that are suspected of being dangerous but are not on the blacklist can be set to require secondary confirmation from the on-duty engineer before execution.

[0076] IV. Closed-loop feedback and profile update module

[0077] The closed-loop feedback and profile update module is deployed on the subsystem coordinating intelligent agent and knowledge graph storage node.

[0078] After the task is completed, this module dynamically updates the digital profile parameters of the corresponding AI digital employee agent based on the actual execution results and response metrics, including: Task completion quality is scored. Acceptance is conducted by on-site maintenance personnel or automatically by the system. Personnel acceptance uses AR terminals or mobile work order applications to rate the task on a scale of one to five stars, or to indicate success on the first attempt or the need for rework. Automatic system acceptance compares whether sensor values ​​returned to normal within a preset window after task execution, and whether the same equipment reported similar anomalies again in subsequent cycles, comprehensively determining success or failure. The score updates the statistical values ​​in the digital profile, including historical success rates and average scores.

[0079] The deviation between actual response time and estimated time. The system records the total time from tender issuance to task completion and compares it with the estimated time based on the average time of similar historical tasks. The deviation is used to adjust the TravelTime estimate for future bids by this employee. If the actual time is significantly longer than expected multiple times, the timeliness score can be lowered.

[0080] Current load status updated. The number of accepted but not yet completed tasks is decremented by one after each task is completed, and the CurrentLoad value is updated accordingly.

[0081] Simultaneously, this module structures the case handling process for this task, including extracting standardized descriptions of abnormal phenomena (e.g., a temperature rise of five degrees Celsius within ten seconds), recording the executed handling steps and actual parameters for each step, and recording the evaluation of handling effectiveness as successful, unsuccessful, or partially successful. This structured information is stored in a semiconductor plant domain knowledge graph in a triple format. For example, a new entity case number is created, and the corresponding digital employee entity is connected through the relationship "handled by [specific action]", and the corresponding fault mode entity is connected through the relationship "resolved fault mode". New cases enrich the graph coverage, improving the accuracy and recall rate of subsequent reasoning.

[0082] Furthermore, the system employs a reinforcement learning algorithm, using the minimization of the overall task allocation cost as the reward function, to iteratively optimize the initial values ​​of the bid weight coefficients α, β, and γ, as well as the adjustment strategy parameters online. The reinforcement learning environment is a plant operation and maintenance system. The state space includes the length of the task queue to be assigned, the load distribution of each digital employee, and the characteristics of the current time period. The action space includes the direction and magnitude of weight adjustment. The reward function is defined as the negative value of the weighted sum of indicators such as task response time, personnel movement distance, and task failure risk. After each task allocation cycle, the reward is calculated based on the actual overall cost. Algorithms such as Deep Q-Network (DQN) or Proximal Point Optimization (PPO) are used to update the strategy network parameters, ensuring that the task allocation strategy approaches its optimal state in the long run.

[0083] V. Global Scheduling Agent and Cross-Subsystem Collaboration

[0084] Refer again Figure 3 The system may optionally include a global scheduling agent. In the full architecture, this agent, as the highest-level scheduling node, is deployed in the plant control system or edge cloud platform and is responsible for resource coordination and cascading fault handling across subsystems.

[0085] When an abnormal event is detected that may affect multiple different subsystems, such as a voltage dip causing an HVAC inverter to trip due to undervoltage, an ultrapure water booster pump to stop, or a false alarm triggered by a special gas system, the global scheduling agent is activated. Activation conditions can be triggered by preset cross-subsystem association rules, or by the subsystem coordination agent actively reporting when it determines that an abnormality may spill over.

[0086] The global scheduling agent maintains a cross-subsystem relationship model, describing the coupling relationships of energy flow, material flow, and signal flow between subsystems. For example, the model includes relationships such as power supply from the power subsystem to the cooling water pumps and fans of the HVAC subsystem, power supply from the power subsystem to the high-pressure pumps of the ultrapure water subsystem, and cooling provided by the HVAC subsystem to the process chambers. The model is based on a directed graph representation, where nodes represent key equipment in each subsystem, and edges represent dependencies and the probability of influence propagation. The global scheduling agent performs inference based on this model, using graph traversal algorithms such as breadth-first search to propagate forward along dependent edges, calculate the list of affected equipment, estimate the probability of cascading events based on historical statistical data, and identify potential cascading failure paths that the current anomaly might trigger.

[0087] After identifying cascading faults, the global scheduling agent generates a cross-subsystem collaborative handling strategy, including a set of parallel or time-dependent subsystem-level handling instructions. For example, for a power voltage dip, the strategy includes issuing a command to the power subsystem coordinating agent to switch to the backup feeder and confirm voltage recovery (this command has the highest priority); simultaneously issuing a command to the HVAC subsystem coordinating agent to switch critical cooling equipment to backup power and await restart; and issuing a command to the ultrapure water subsystem coordinating agent to suspend water production and initiate internal circulation. Figure 3 As shown by the solid arrow, the global scheduling agent sends the collaborative processing instructions to the coordinating agents of each relevant subsystem. After receiving the instructions, each coordinating agent schedules the digital employees of its subsystem to execute them according to the aforementioned task allocation and execution process.

[0088] The global scheduling agent is also responsible for resolving resource conflicts between subsystem coordination agents. When multiple subsystem coordination agents simultaneously request the same scarce resource, the global scheduling agent acts as an arbitrator. Scarce resources include the same digital employee with advanced skills across subsystems, such as a shift engineer with advanced skills in both power and HVAC; a single critical spare part such as a specific model of frequency converter in the warehouse; and a shared workspace. The global scheduling agent makes a game-theoretic decision based on preset production priority rules, comprehensively considering factors such as the urgency of the requested task, the current operating status of related production lines (e.g., full production lines have higher priority than idle lines), and the estimated time of the task (e.g., short tasks prioritized, long tasks prioritized, configurable). Through weighted scoring or expert system decision-making, the agent allocates the resource to the task with the highest overall priority and returns suggestions to other applicants to wait in line or find alternative solutions.

[0089] In the lightweight architecture, for smaller-scale equipment such as 6-inch or 8-inch mature process wafer fabs with only one production line, and semiconductor factories with low subsystem coupling, the global scheduling agent can be omitted. Under this architecture, Figure 3 The top-level global scheduling agent layer has been removed, and subsystem coordinating agents directly resolve conflicts and coordinate simply through point-to-point communication protocols. For example, when a subsystem coordinating agent needs to occupy resources that may affect other subsystems, it broadcasts the resource occupation request to the relevant subsystems. Each relevant coordinating agent responds with agreement or rejection based on its own state, achieving lightweight distributed coordination through preset negotiation timeouts and retry mechanisms. Omitting the global scheduling agent reduces communication overhead and computational latency, improving response speed in simple scenarios.

[0090] Method Implementation Examples

[0091] Figure 2 This illustration shows a flowchart of a dynamic task allocation method for digital employees in semiconductor plants based on multi-agent collaboration, according to an embodiment of this application. (Combined with...) Figure 2 The method includes the following steps: Step S1: Plant Status Awareness and Task Generation. The equipment-level intelligent agent collects real-time sensor data from the plant's physical equipment to detect anomalies, employing the LSTM model described earlier for timing anomaly detection. When manual intervention is deemed necessary, a task request carrying task attribute information is submitted to the superior subsystem's coordinating agent. The specific criteria for determining the need for manual intervention are as described earlier, including physical operation constraint determination, logical confidence threshold determination, and execution feedback anomaly determination. The logical confidence threshold is, for example, set to 0.85.

[0092] Step S2: Task Dynamic Broadcasting and Autonomous Bidding. The subsystem coordinating agent encapsulates task requests into bidding messages conforming to the Contract Network protocol or auction mechanism and broadcasts them to the corresponding AI digital employee agent pool. Each digital employee autonomously calculates its bid value (Bid) based on its own digital profile and provides feedback before the deadline. The bid value calculation formula, the meaning and values ​​of each parameter, the initial value training method for weight coefficients, the adaptive adjustment logic under three working conditions, the TravelTime correction method for pure virtual tasks, and the fault tolerance fallback mechanism are all as described in the previous task dynamic allocation module. The fault tolerance fallback includes task timeout reselection, forced task preemption, and manual takeover modes. Detailed interaction sequences for this step can be found in [link to relevant documentation]. Figure 4 .

[0093] Step S3: Optimal task allocation and execution guidance. The subsystem coordinating agent selects the target digital employee with the highest bid value and issues execution instructions. The winning digital employee invokes the semiconductor plant domain knowledge graph for reasoning, generating standard operating procedures or simultaneously generating execution work orders, and outputting operation and maintenance guidance information. The specific technical path of execution guidance includes RPA script UI identifier mapping and API call sequence generation, AR terminal spatial anchor point registration and visualization instruction rendering, and security access control mechanisms including dynamic access token application and verification, and security blacklist interception rules, all as described in the collaborative scheduling and execution guidance module above.

[0094] Step S4: Operational Feedback and Continuous Learning of Digital Profiles. Upon task completion, the system dynamically updates the digital profile parameters of the target digital employee based on the actual execution results, including task completion quality score, deviation between actual and estimated response time, and current load status. The system structurally stores this handling case in the semiconductor plant domain knowledge graph. Simultaneously, a reinforcement learning algorithm is used, with the minimum overall cost of task allocation as the reward function, to iteratively optimize the bid value weight coefficients online. The above closed-loop feedback and learning mechanism is as described in the previous section on closed-loop feedback and profile update.

[0095] In an embodiment including a global scheduling agent, when the anomaly detected in step S1 affects multiple subsystems, the global scheduling agent is activated before step S2. Based on a cross-subsystem correlation model, it identifies cascading fault paths, generates a collaborative handling strategy, and distributes it to the coordinating agents of each relevant subsystem. Each coordinating agent then executes steps S2 to S4 accordingly. This cross-subsystem coordination process can be referred to... Figure 3 In the lightweight architecture, the global scheduling agent is omitted, and the subsystem coordinating agents collaborate through point-to-point communication.

[0096] The embodiments described above are merely further illustrations of the present invention and are not intended to limit the present invention in any other way. The present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding modifications and changes based on the present invention, but all such modifications and changes should fall within the protection scope of the present invention.

Claims

1. A semiconductor plant AI digital employee system based on multi-agent collaboration, characterized in that, The system adopts a hierarchical multi-agent architecture, including: An anomaly perception and task generation module is deployed at the device layer intelligent agent. It is used to collect sensor data of plant physical equipment in real time and identify the operating status through an anomaly detection model. When it is determined that manual intervention is required, it submits a task request to the superior coordinating intelligent agent. The task request includes at least task attribute information. A task dynamic allocation module is deployed on the coordinating agent and multiple AI digital employee agents. The coordinating agent is used to encapsulate the task request into a task bidding message and broadcast it. Each AI digital employee agent is used to receive the bidding message and autonomously calculate the bid value based on its own digital profile to participate in the bidding. The bid value comprehensively represents at least the relevant parameters of task execution cost, including capability matching degree, response timeliness and current occupancy status. The collaborative scheduling and execution guidance module is used by the coordinating agent to select the target AI digital employee agent based on the feedback bid value and issue execution instructions; the winning target AI digital employee agent calls the semiconductor plant domain knowledge graph to perform reasoning, generate standard operating procedures, or generate standard operating procedures and execution work orders to output operation and maintenance guidance information; The closed-loop feedback and profile update module is used to dynamically update the digital profile parameters of the corresponding AI digital employee agent based on the actual execution results and response indicators after the task is completed, and to structurally store the new processing cases in the semiconductor plant domain knowledge graph.

2. A method for task allocation of digital employees in semiconductor plants based on multi-agent collaboration, characterized in that, Includes the following steps: S1) Plant status perception and task generation: The equipment layer intelligent agent collects sensor data of the plant's physical equipment in real time to detect anomalies. When it is determined that manual intervention is required, it submits a task request carrying task attribute information to the superior coordinating intelligent agent. S2) Task dynamic broadcasting and autonomous bidding: The coordinating agent encapsulates task requests into bidding messages and broadcasts them to the corresponding AI digital employee agent pool; Each AI digital employee intelligent agent autonomously calculates its bid value based on its own digital profile and provides feedback. The bid value comprehensively represents at least the relevant parameters including capability matching degree, response timeliness, and task execution cost of the current occupancy status. S3) Optimal allocation and execution guidance of tasks: The coordinating agent selects the target AI digital employee agent based on the bid values ​​fed back by each agent and issues instructions; the target AI digital employee agent calls the semiconductor plant domain knowledge graph to generate standard operating procedures, or generates standard operating procedures and execution work orders to output operation and maintenance guidance information; S4) Operation and Maintenance Feedback and Continuous Learning of Digital Profiling: After the task is completed, the system updates the digital profile parameters of the target AI digital employee agent based on the actual execution results and stores the new processing cases in the domain knowledge graph.

3. The method according to claim 2, characterized in that, Anomaly detection is performed using a long short-term memory network model; the coordinating agent and the AI ​​digital employee agent use a contract network protocol or auction mechanism to broadcast and bid for tasks.

4. The method according to claim 2, characterized in that, The specific formula for calculating the bid value by the AI ​​digital employee agent is as follows: Bid=αSkillMatch+β(1 / TravelTime)-γCurrentLoad; Wherein, Bid is the bid value, SkillMatch is the skill matching degree of the AI ​​digital employee, TravelTime is the estimated travel time of the AI ​​digital employee from its current physical location to the location of the abnormal device, CurrentLoad is the current load of the AI ​​digital employee, and α, β, and γ are dynamic weighting coefficients.

5. The method according to claim 4, characterized in that, The dynamic weighting coefficients α, β, and γ are adaptively adjusted according to the urgency of the task request; when the urgency exceeds the preset high-risk threshold, the system automatically increases the proportion of the weighting coefficient β corresponding to the estimated travel time.

6. The method according to claim 2, characterized in that, The winning AI digital employee agent will provide visual operation guidance to on-site maintenance personnel by driving robotic process automation tools or augmented reality terminals.

7. The method according to claim 2, characterized in that, It also includes a global scheduling agent; when a detected abnormal event affects the plant physical equipment of multiple different subsystems, the global scheduling agent identifies the cascading fault and issues cross-subsystem collaborative handling instructions to the coordinating agents of each relevant subsystem.

8. The method according to claim 2, characterized in that, A reinforcement learning algorithm is used, with the reward function being the minimum overall cost of task allocation, to iteratively optimize the dynamic weight coefficients of the AI ​​digital employee agent when calculating the bid value online.

9. The method according to claim 2, characterized in that, It is also equipped with a fault tolerance and backup mechanism: if a bid value that meets the minimum requirements of the task is not received within the preset time, or if the winning AI digital employee agent does not confirm acceptance of the task within the predetermined time, the system will automatically start a high-priority second bidding process. If bidding fails consecutively, the task will be forcibly attached to the manual control console and an alarm will be triggered.

10. The method according to claim 4, characterized in that, When the type of the task request is determined to be a pure software operation task, the TravelTime parameter in the bid value calculation formula is replaced with a constant minimum value or a network response latency indicator; the method also includes: when calling the plant management underlying software control interface, performing security verification on the operation command based on a dynamic permission token mechanism, wherein the dynamic permission token is generated by the global scheduling agent for a specific task and is only valid within a preset effective time window and a specified device address range.

Images