Multi-agent resource scheduling method based on large language model in industrial scene

Abstract

The invention discloses a multi-agent resource scheduling method based on a large language model in an industrial scene, and the method comprises the following steps: 1, setting initial scheduling parameters and user input of a multi-agent scheduling system; 2, the planning agent performs intention recognition, feature value extraction and coding; and 3, the scheduling agent receives the scheduling coding information, selects a scheduling algorithm tool and generates a scheduling scheme. And 4, the fault detection agent receives the initial scheduling scheme and receives key parameters of the working agent. And 5, the fault detection agent judges that a fault exists and optimization exists, if yes, the scheduling agent is fed back, and the step 3 is executed again, and if not, an optimal scheduling result is output. According to the method, a multi-agent architecture driven by a large language model is adopted, and a dynamic algorithm tool selection mechanism and a prompt project are introduced. Minimization of task total time consumption and maximization of fault detection accuracy are taken as optimization objectives, and adaptivity of industrial scene resource scheduling and system reliability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent scheduling technology, and in particular to a multi-agent resource scheduling method based on a large language model in industrial scenarios. Background Technology

[0002] Resource scheduling in industrial scenarios is a complex dynamic optimization problem that requires coordination of multiple types of equipment. Traditional methods have three major drawbacks: poor environmental adaptability: fixed resource scheduling algorithms are difficult to respond dynamically; delayed fault response: equipment fault detection relies on human experience, multimodal data fusion is difficult, and delayed fault feedback leads to scheduling interruptions; and high system coupling: scheduling and detection modules operate in isolation and lack collaborative optimization mechanisms.

[0003] Existing research attempts to improve flexibility through multi-agent systems, but traditional multi-agent systems have fixed decision-making logic, cannot understand natural language instructions, and struggle to handle complex spatiotemporal constraints. The emergence of large language models (GLAMs) offers a new solution. GLAMs are deep learning models trained on large-scale text data, capable of understanding, generating, and manipulating natural language. Their core technology is the Transformer architecture built on a self-attention mechanism. Currently, with the continuous expansion of model parameters and training data, GLAMs have demonstrated powerful language generation and understanding capabilities. Among them, OpenAI's Generative Pre-trained Transformer (GPT) series is particularly noteworthy, and the iterative development of GPT series models (such as GPT-2 and GPT-3) has further propelled the development of the field of natural language processing.

[0004] With the continuous development of large language models, the concept of intelligent agents driven by large language models has emerged. These agents can understand and process complex natural language tasks, achieving instructions or goals through mechanisms such as autonomous learning, reasoning, planning, and knowledge reasoning. However, ensuring the interpretability of their decision-making processes remains a significant challenge. Currently, large language model-based agents are widely used in dialogue systems, task execution, and knowledge management, demonstrating strong autonomy and flexibility. In this research direction, agents perceive the environment and store information as memory, thus giving rise to large language model-driven multi-agent systems, further enhancing the capabilities of individual large language model agents. In large language model-driven multi-agent systems, multiple agents can communicate and collaborate, ultimately obtaining more comprehensive and accurate research results.

[0005] In the field of industrial resource scheduling, the core advantage of large language model-driven intelligent agents lies in their natural language understanding capabilities, enabling them to handle complex scheduling tasks. In production systems, these agents can act as scheduling control centers, interacting with human operators in natural language and dynamically adjusting production plans and resource allocation by describing equipment and task status. Research shows that large language model-driven intelligent agents exhibit high adaptability when dealing with changing production environments.

[0006] However, existing agents driven by large language models are still limited by computing power when handling large-scale tasks, especially under complex constraints, where the generated scheduling schemes may not be optimal. Furthermore, although agents driven by large language models possess some reasoning ability when dealing with multi-task scheduling, their optimization capabilities are still insufficient when handling highly complex mathematical programming problems. During resource scheduling, the reasoning efficiency of these agents decreases significantly with increasing problem complexity, posing a pressing problem to be solved in large-scale industrial production systems.

[0007] In the field of fault detection, large language model-driven agents, with their powerful natural language parsing capabilities, can extract key information from massive amounts of fault logs and reports and generate anomaly prediction models. By integrating with industrial equipment, these agents can learn fault patterns from sensor data and generate fault reports or maintenance suggestions through language.

[0008] However, the fault detection performance of large language model-driven agents remains limited when processing multimodal data. For example, when faced with multimodal inputs of sensor data and text descriptions, the model exhibits weak fusion capabilities, easily leading to information loss or false alarms. Furthermore, large language model-driven agents rely on large amounts of historical data for fault prediction, but when the data is incomplete or of low quality, the model's predictive ability significantly decreases. The fault detection accuracy of large language model-driven agents drops significantly when faced with data noise and changes in the equipment environment, posing a major challenge in practical industrial applications. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies by providing a multi-agent resource scheduling method based on a large language model in industrial scenarios. This method effectively solves problems such as insufficient adaptability of resource scheduling and low real-time performance of fault detection in industrial scenarios by constructing a multi-agent architecture driven by a large language model and a dynamic interaction paradigm, thereby significantly improving system reliability and task execution efficiency.

[0010] The objective of this invention is achieved through the following technical solution: a multi-agent resource scheduling method based on a large language model in industrial scenarios, comprising the following steps: (1) Receive the initial scheduling data of the target task and generate formatted codes through the routing coding agent; (2) Based on the formatted encoding, the scheduling agent calls the heuristic scheduling algorithm tool to generate a preliminary resource scheduling scheme; (3) The fault detection agent collects real-time indicator data of the working agent and receives the resource scheduling scheme generated by the scheduling agent, and performs fault diagnosis in conjunction with the dynamic time warping algorithm; if no fault is diagnosed, the scheduling scheme is output; if a fault is diagnosed, fault information is generated. (4) The working intelligent agent verifies the fault information and combines it with its own fault document to feed back parameter adjustment instructions; (5) The scheduling agent adjusts the scheduling parameters based on the feedback from the working agent and re-outputs the scheduling scheme; (6) Repeat steps (3) to (5) until the scheduling scheme diagnosis is fault-free, and output the scheduling scheme.

[0011] Further, step (2) includes the following sub-steps: (2.1) Integrate heuristic scheduling algorithms, establish a classification label system according to the characteristics of the applicable scenarios of the algorithms, and define the algorithm fitness function. F ( A , S ),when F ( A , S ) ≥ α The time markers are used as candidate algorithms, where A For algorithm type, S For scene feature vectors, α To adapt the threshold; (2.2) Extracting scene feature parameters from formatted encoding S = [ n , m , t max , N min ], combined with specific scenarios and formulas A matched = argmax{ F ( A , S Select Algorithm A matched ;in, n For the total number of tasks, m For the total number of devices, t max For the maximum task time, N min Minimum output (2.3) For the selected algorithm A matchedThe scheduling agent automatically generates an initial parameter set based on the constraints in the formatted encoding; (2.4) The scheduling agent invokes the algorithm tool to execute the selected algorithm. A matched Generate a resource scheduling scheme.

[0012] Further, step (2.3) includes: (2.3.1) For each task k Generation processing time t k ,make sure t min(k) ≤ t k ≤ t max(k) ;in t min(k) = m k - s k , t max(k) = m k + s k , t min(k) For the task k The shortest processing time t min(k) For the task k The longest processing time, m k For the task k The historical average processing time, s k Standard deviation; (2.3.2) Initialize the device m task set S m = ∅ Total processing time T m = 0, set the initial load threshold based on the average load rate of devices in historical tasks. T load(m) = l × C m λ is the load factor. C m For equipment m Maximum load capacity, satisfying Σ t k ( k ∈ S m ) ≤ Tload(m) ; (2.3.3) Regarding the time interval Δt and the minimum output quantity N min Set dynamic verification threshold N alert = β × N min , β This is the early warning coefficient.

[0013] Further, step (2.4) includes: (2.4.1) Traverse all tasks and generate the corresponding symbols in step (2.3.1). t min(k) ≤ t k ≤ t max(k) The task processing time is recorded and stored in the time matrix. T = [ t 1, t 2, ..., t n ]; (2.4.2) For each device m Initialize task set S m Total processing time T m Initialize to the total time of tasks within the set. T m = Σ t k for k ∈ S m Synchronously record to the device load tracking table; (2.4.3) The task allocation is iteratively optimized using the particle swarm optimization algorithm. In each iteration, the task-device mapping relationship is adjusted according to the speed update formula, and the total processing time of the device is updated. T m' = T m + t k At the same time, the total global processing time T total Updated to the maximum completion time for all devices. T total = max( T 1, T 2, ..., T m ); (2.4.4) In the time window loop, verify each time interval. Δt Number of tasks completed within N completed Has it been achieved? N min ,like N completed < N alert This triggers an alert and marks it as a solution optimization point. Steps 4-3 are repeated to regenerate task assignments; this continues until all tasks are completed. N completed > N alert Finally, the scheduling scheme is output.

[0014] Furthermore, the workflow of the fault detection agent includes the following steps: S1. Fault detection agent collects real-time operational metrics uploaded by the agent. X ( t )=[ x 1(t) , x 2(t) , ..., x d(t) The collected data is preprocessed to generate a standardized index sequence. X norm(t) = ( X ( t ) - m X ) / s X ;in d As an indicator dimension, t For timestamps, m X The mean of the indicators. s X Standard deviation; S2. Obtain resource allocation parameters from the resource scheduling scheme generated by the scheduling agent. Y = [ y 1, y 2, ..., y d ]; S3. The fault detection agent invokes the dynamic time warping algorithm tool to compare the standardized indicator sequence with the equipment normal operation benchmark data sequence in the industrial knowledge base. X ref(t) Perform time series similarity comparison and calculate the deviation value: DTW( X norm , X ref ) = min{Σ| X norm(i) -X ref(j) | × w ( i , j )}, when DTW( X norm , X ref )> i The system determines when a fault exists and calculates its severity. F level = DTW( X norm , X ref )) / i , F level A level assessment is triggered when the value is >1; among which... w ( i , j ) represents the path weights in dynamic programming. i The preset deviation threshold; S4. If a fault exists, generate fault information. F info = { y d , S The confirmed fault information is fed back to the scheduling agent via the RFIP protocol, triggering the scheduling scheme optimization process, where the fault level is determined by... L = ceil( F level The system is divided into sections, and the scheduling agent performs tasks based on different fault levels. L The scheduling scheme is optimized through a self-loop; among which... L =1, 2, or 3 L When the value is 3, a new suitable scheduling algorithm is selected, and the process repeats three more times to select the optimal scheduling scheme after the loop. L When <3, the scheduling agent performs... L In the next iteration, select the optimal scheduling scheme after the previous iteration; S5. After the scheduling agent generates an optimized solution, repeat the diagnostic process of steps S1-S3 to verify the effectiveness of the optimized solution in resolving the fault. If there is no fault, output the scheduling optimization solution; if there is a fault, execute step S4 to optimize the scheduling solution.

[0015] Furthermore, the interaction logic of RFIP includes: Step 1: The scheduling agent generates a preliminary resource scheduling scheme. S scheme Then, it automatically synchronizes to the fault detection agent, and simultaneously triggers the working agent to perform indicator analysis. W metricsData collection ensures that the fault detection agent can acquire data in real time. S scheme Related indicator data; Step 2: After receiving the preliminary plan, the fault detection agent calls the dynamic time warping algorithm to... S scheme Includes time series data and W metrics Data similarity analysis is performed; when the deviation value exceeds the preset working agent fault document... S doc If the threshold set in the code is reached, a fault is considered to exist, and relevant indicator data is returned. W metrics-s and documents S doc-s If the threshold is not exceeded, it is determined that there is no fault, and a scheduling scheme is output. S scheme .

[0016] Step 3: The fault detection agent generates fault information. F info Then, it is prioritized for distribution to the corresponding working agent, and the working agent queries. W metrics Perform a matching verification; if the verification matches, provide feedback. S scheme Adjustment suggestions; if verification is inconsistent, repeat step 2. Step 4: After receiving the adjustment suggestions, the scheduling agent combines them with the fault information generated by the fault detection agent. F info The scheduling agent determines the fault information. F info Generate an optimized scheduling scheme S scheme-pre Then it synchronizes again to the fault detection agent until the fault detection agent detects no fault warnings, and then outputs the final scheduling plan. S scheme-final .

[0017] This invention also provides a multi-agent resource scheduling and fault monitoring system driven by a large language model, comprising: a scheduling agent, a fault detection agent, a working agent, and a routing coding agent; The routing coding agent is used to receive initial scheduling data for the target task and generate formatted codes. The scheduling agent is used to generate a preliminary resource scheduling scheme by calling a heuristic scheduling algorithm tool according to the formatted encoding; and if the generated resource scheduling scheme is diagnosed as faulty, the scheduling parameters are adjusted according to the feedback of the working agent, and the scheduling scheme is re-outputted until the fault detection agent diagnoses no fault and outputs the final scheduling scheme. The fault detection agent receives real-time indicator data from the working agent and resource scheduling schemes generated by the scheduling agent, and performs fault diagnosis in conjunction with the dynamic time warping algorithm; if a fault is diagnosed, fault information is generated. The working intelligent agent is used to collect and store real-time indicator data of the working intelligent agent; and to verify fault information and provide parameter adjustment instructions based on its own fault documentation.

[0018] Furthermore, it also includes: An integrated scheduling agent algorithm toolkit containing multiple heuristic scheduling algorithms. The scheduling agent autonomously selects algorithm tools based on scenario characteristics, including parameter optimization and scheme generation modules; specifically: The scheduling agent autonomously selects algorithm tools based on scene characteristics and identifies scene feature parameters through formatted encoding. S feat Output matching heuristics from the algorithm tool library A matched The parameter optimization module is invoked to automatically generate an initial parameter set based on the constraints in the formatted encoding; then the scheme generation module invokes the selected heuristic algorithm. A matched Generate scheduling scheme S scheme ; The system constructs a fault detection intelligent agent, including a data acquisition interface, a feature engineering module, a dynamic time warping algorithm execution unit, and a fault assessment module; specifically: Receive indicator data stored by the working intelligent agent through the data acquisition interface. W metrics The scheduling scheme generated by the scheduling agent S scheme The feature engineering module is invoked to preprocess the collected data and generate a standardized index sequence. X norm(t) The dynamic time warping algorithm execution unit and fault assessment module are used to call the dynamic time warping algorithm to detect whether there are problems with the indicator data and to perform fault assessment, generating fault information. F info ; Configure the working intelligent agent fault document storage module: collect its own operating parameters and respond to fault verification commands; specifically: Based on scene feature parameters S feat Configuration scenario-related fault documentation S doc and indicator data W metrics Simultaneously, it receives fault verification instructions from the fault scheduling agent; if a fault exists, it returns relevant indicator data. Wmetrics-s and corresponding fault documentation S doc-s If no fault is found, return a fault-free flag. S safe .

[0019] The present invention also provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, is used to implement a multi-agent resource scheduling method based on a large language model in an industrial scenario according to any one of claims 1-6.

[0020] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the multi-agent resource scheduling method based on a large language model in an industrial scenario as described in any one of claims 1-6.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) In the prior art, multi-agent systems often face problems such as black box decision logic and reliance on manual intervention for iteration. However, the present invention relies on the semantic understanding and reasoning capabilities of a large language model to enable the agent to autonomously complete the entire process of "problem analysis - solution generation - self-evaluation - dynamic iteration".

[0022] (2) In the problem understanding stage, the big language model gives the agent the ability to interpret the deep semantics of complex scenarios and can accurately capture the implicit constraints in the task requirements. In the solution generation stage, with the help of the knowledge transfer and logical deduction advantages of the big language model, the agent can generate a global optimal solution that is more in line with the actual scenario, rather than a local compromise solution. The self-evaluation and iteration mechanism, through the natural language model of the decision-making process, makes the entire reasoning path traceable and explainable, solving the pain point of the obscure and difficult-to-understand decision logic of traditional agents. (3) Develop a resource scheduling and fault detection interaction paradigm (RFIP) to build a self-regulation and feedback loop for multi-agent systems. However, the interaction of traditional multi-agent systems is mostly limited to the one-way transmission of task instructions, and resource allocation and fault handling are often in a state of separation, resulting in problems such as system response delay and resource waste. Attached Figure Description

[0023] Figure 1 This is a flowchart of a multi-agent scheduling method for container terminals based on a large language model; Figure 2 This is a schematic diagram of container terminal transportation; Figure 3 This is a schematic diagram of RFIP. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings. Unless otherwise specified, the features of the following embodiments and implementations can be combined with each other.

[0025] This invention provides a multi-agent resource scheduling method based on a large language model in an industrial scenario, comprising the following steps: (1) Receive the initial scheduling data of the target task and generate formatted codes through the routing coding agent; (2) Based on the formatted encoding, the scheduling agent calls the heuristic scheduling algorithm tool to generate a preliminary resource scheduling scheme; (3) The fault detection agent collects real-time indicator data of the working agent and receives the resource scheduling scheme generated by the scheduling agent, and performs fault diagnosis in conjunction with the dynamic time warping algorithm; if no fault is diagnosed, the scheduling scheme is output; if a fault is diagnosed, fault information is generated. (4) The working intelligent agent verifies the fault information and combines it with its own fault document to feed back parameter adjustment instructions; (5) The scheduling agent adjusts the scheduling parameters based on the feedback from the working agent and re-outputs the scheduling scheme; (6) Repeat steps (3) to (5) until the scheduling scheme diagnosis is fault-free, and output the scheduling scheme.

[0026] Specifically, in step (1), a routing coding agent is developed to realize the structured transformation and encoding generation of initial task data; specifically, it includes: structured task scheduling sequence, individual task scheduling time, task priority and other initial scheduling data, forming scheduling codes. S code . Specifically, step (2) includes the following sub-steps: (2.1) Integrate heuristic scheduling algorithms, establish a classification label system according to the characteristics of the applicable scenarios of the algorithms, and define the algorithm fitness function. F ( A , S ),when F ( A , S ) ≥ α ( α Algorithms ∈ [0.6, 0.8] are marked as candidate algorithms, where A For algorithm type, S For scene feature vectors, α To adapt the threshold; (2.2) Extracting scene feature parameters from formatted encoding S = [ n , m , t max ,N min ], combined with specific scenarios and formulas A matched = argmax{ F ( A , S Select Algorithm A matched ;in, n For the total number of tasks, m For the total number of devices, t max For the maximum task time, N min Minimum output (2.3) For the selected algorithm A matched The scheduling agent automatically generates an initial parameter set based on the constraints in the formatted encoding; (2.4) The scheduling agent invokes the algorithm tool to execute the selected algorithm. A matched Generate a resource scheduling scheme.

[0027] Specifically, step (2.3) includes: (2.3.1) For each task k Generation processing time t k ,make sure t min(k) ≤ t k ≤ t max(k) ;in t min(k) = m k - s k , t max(k) = m k + s k , t min(k) For the task k The shortest processing time t min(k) For the task k The longest processing time, m k For the task k The historical average processing time, s k Standard deviation; (2.3.2) Initialize the device m task setS m = ∅ Total processing time T m = 0, set the initial load threshold based on the average load rate of devices in historical tasks. T load(m) = l × C m λ is the load factor, λ∈[0.7,0.9], C m For equipment m Maximum load capacity, satisfying Σ t k ( k ∈ S m ) ≤ T load(m) ; (2.3.3) Regarding the time interval Δt and the minimum output quantity N min Set dynamic verification threshold N alert = β × N min , β As the early warning coefficient, β ∈[1.1,1.3].

[0028] Specifically, step (2.4) includes: (2.4.1) Traverse all tasks and generate the corresponding symbols in step (2.3.1). t min(k) ≤ t k ≤ t max(k) The task processing time is recorded and stored in the time matrix. T = [ t 1, t 2, ..., t n ]; (2.4.2) For each device m Initialize task set S m Total processing time T m Initialize to the total time of tasks within the set. T m = Σ t k for k ∈ S mSynchronously record to the device load tracking table; (2.4.3) The task allocation is iteratively optimized using the particle swarm optimization algorithm. In each iteration, the task-device mapping relationship is adjusted according to the speed update formula, and the total processing time of the device is updated. T m' = T m + t k At the same time, the total global processing time T total Updated to the maximum completion time for all devices. T total = max( T 1, T 2, ..., T m ); (2.4.4) In the time window loop, verify each time interval. Δt Number of tasks completed within N completed Has it been achieved? N min ,like N completed < N alert This triggers an alert and marks it as a solution optimization point. Steps 4-3 are repeated to regenerate task assignments; this continues until all tasks are completed. N completed > N alert Finally, the scheduling scheme is output.

[0029] Taking the particle swarm optimization algorithm as an example, the population size is initialized... With maximum number of iterations Generate task-equipment allocation scheme ( For equipment (task set), and calculate the total processing time of the device. ( For the task Processing time, to meet Total global execution time The following are examples of algorithm tools: Specifically, the workflow of the fault detection agent includes the following steps: S1. Fault detection agent collects real-time operational metrics uploaded by the agent. X ( t )=[ x 1(t) , x2(t) , ..., x d(t) The collected data is preprocessed to generate a standardized index sequence. X norm(t) = ( X ( t ) - m X ) / s X ;in d As an indicator dimension, t For timestamps, m X The mean of the indicators. s X Standard deviation; S2. Obtain resource allocation parameters from the resource scheduling scheme generated by the scheduling agent. Y = [ y 1, y 2, ..., y d ]; S3. The fault detection agent invokes the dynamic time warping algorithm tool to compare the standardized indicator sequence with the equipment normal operation benchmark data sequence in the industrial knowledge base. X ref(t) Perform time series similarity comparison and calculate the deviation value: DTW( X norm , X ref ) = min{Σ| X norm(i) - X ref(j) | × w ( i , j )}, when DTW( X norm , X ref )> i The system determines when a fault exists and calculates its severity. F level = DTW( X norm , X ref )) / i , F level A level assessment is triggered when the value is >1; among which... w ( i , j ) represents the path weights in dynamic programming. i To set a preset deviation threshold, i∈[0.1,0.3]; S4. If a fault exists, generate fault information. F info = { y d , S The confirmed fault information is fed back to the scheduling agent via the RFIP protocol, triggering the scheduling scheme optimization process, where the fault level is determined by... L = ceil( F level ) division ( L =1 indicates a minor fault. L =2 indicates a moderate fault. L =3 indicates a severe fault (ceil() is a function that rounds up), and the scheduling agent performs different actions based on the fault level. L The scheduling scheme is optimized through a self-looping mechanism. S5. After the scheduling agent generates an optimized solution, repeat the diagnostic process of steps S1-S3 to verify the effectiveness of the optimized solution in resolving the fault. If there is no fault, output the scheduling optimization solution; if there is a fault, execute step S4 to optimize the scheduling solution.

[0030] Specifically, the interaction logic of RFIP (resource scheduling and fault detection interaction paradigm) includes: Step 1: The scheduling agent generates a preliminary resource scheduling scheme. S scheme Then, it automatically synchronizes to the fault detection agent, and simultaneously triggers the working agent to perform indicator analysis. W metrics Data collection ensures that the fault detection agent can acquire data in real time. S scheme Related indicator data; Step 2: After receiving the preliminary plan, the fault detection agent calls the dynamic time warping algorithm to... S scheme Includes time series data and W metrics Data similarity analysis is performed; when the deviation value exceeds the preset working agent fault document... S doc If the threshold set in the code is reached, a fault is considered to exist, and relevant indicator data is returned. W metrics-s and documents S doc-s If the threshold is not exceeded, it is determined that there is no fault, and a scheduling scheme is output. S scheme .

[0031] Step 3: The fault detection agent generates fault information. F info Then, it is prioritized for distribution to the corresponding working agent, and the working agent queries. W metrics Perform a matching verification; if the verification matches, provide feedback. S scheme Adjustment suggestions; if verification is inconsistent, repeat step 2. Step 4: After receiving the adjustment suggestions, the scheduling agent combines them with the fault information generated by the fault detection agent. F info The large language model in the scheduling agent reselects a suitable heuristic scheduling algorithm from the heuristic scheduling algorithm toolkit to generate an optimized scheduling scheme. S scheme-pre Then it synchronizes again to the fault detection agent until the fault detection agent detects no fault warnings, and then outputs the final scheduling plan. S scheme-final .

[0032] In one embodiment, if the working intelligent agent receives fault information Then, call the rules in the local fault document. (e.g., "vibration frequency > 2Hz → mechanical failure") Perform matching verification. If real-time indicators... Satisfy the rules Then the fault is confirmed and parameter adjustment suggestions are generated. (e.g., load reduction ratio) Time window correction value If they do not match, provide feedback on how to troubleshoot the problem and supplement the internal status data of the equipment (such as lubricating oil cleanliness and motor temperature).

[0033] Structured work feedback: The working agent encapsulates the verification results into structured feedback. ,in This is a fault confirmation indicator. This indicates the current operating status of the equipment. Feedback is synchronized to the scheduling agent via the RFIP communication protocol, ensuring low latency and integrity of information transmission.

[0034] The scheduling parameters are dynamically adjusted, and the scheduling agent is based on... Re-optimize the scheduling scheme: If it is a minor fault ( Adjust the equipment load threshold. ( For load factor, , (This refers to the maximum load capacity of the equipment). If it is a moderate fault ( The load threshold is adjusted to Load_max(m) = λ × C_m for λ ∈ [0.5, 0.7]. Tasks whose time exceeds 60% of the adjusted threshold are migrated to low-load devices. Simultaneously, according to... T window' = T window + Δ time ×1.5 Correct the time window; then recalculate the total global time including the shortened monitoring interval and verify synchronously. If the deviation exceeds the limit, repeat the migration until it passes.

[0035] If it is a serious fault ( This triggers an algorithm tool switch (e.g., from particle swarm optimization to genetic algorithm), and reallocates the task set. ,make sure This applies to all devices.

[0036] After optimization, the total global execution time was recalculated. And synchronize it to the fault detection intelligent agent for secondary verification.

[0037] Specifically, the design of prompt word templates for intelligent agents involves the following steps: Step 1: Role Definition and Core Objectives. "As a {smart agent type}, it needs to leverage the industrial knowledge base {Knowledge_Base} and real-time data interaction to complete {core functions}: {Scheduling smart agent: resource optimization and solution generation; Fault detection smart agent: equipment status monitoring and anomaly diagnosis; Working smart agent: task execution and parameter feedback; Routing coding smart agent: task data parsing and standardized coding}, ultimately supporting the system to achieve the goals of minimizing total time {T_total} and maximizing fault detection accuracy {F}." Step 2: Basic Task Instructions: "Receive input data: {Scheduling agent: formatted encoding {Encoded_Data}; Fault detection agent: real-time indicators {Real_Time_Indicators} and scheduling scheme parameters; Working agent: scheduling instructions {Schedule_Plan}; Routing coding agent: initial task data {Raw_Data}}. Execute the basic functions: {Scheduling agent: select a heuristic algorithm to generate a preliminary scheme; Fault detection agent: compare with baseline data to identify anomalies; Working agent: execute according to the scheme and collect parameters; Routing coding agent: extract features to generate standardized codes}, and output basic results {Scheme / Fault Warning / Status Data / Encoding Results}." Step 3: Constraints and Parameter Specifications. "Operational Constraints: {Scheduling Agent: Task Priority {Task_Priority}, Device Load Threshold {Load_Max}, Time Window [{t_min}, {t_max}]; Fault Detection Agent: Normal Range [{Norm_min}, {Norm_max}], Deviation Threshold {θ} (Applicable to DTW Algorithm), Fault Level Classification Rule {L=ceil(DTW / θ)}; Working Agent: Parameter Acquisition Frequency {Δt}, Local Fault Rules {Local_Rules}; Routing Encoding Agent: Required Fields {Required_Fields}, Encoding Standard {Format_Standard}}. Algorithm Tools: {Scheduling Agent: Algorithm Tool Library {Algorithms}; Fault Detection Agent: Diagnostic Algorithms such as Dynamic Time Warping; Working Agent: Local Rule Matching Engine; Routing Encoding Agent: Feature Extraction and Verification Tool}. These must be called according to the specifications and the parameter sources must be labeled." Step 4: Dynamic Interaction and Feedback Adaptation. "Real-time feedback processing: Receive interaction information {feedback content} from the {collaborative agent}: {Scheduling agent: Fault detection result {Fault_Info}, worker agent adjustment instructions; Fault detection agent: Worker agent verification data, scheduling scheme update; Worker agent: Fault warning {Fault_Warning}, scheduling parameter change; Routing coding agent: Scheduling agent coding verification opinion}. Based on the feedback, dynamically adjust: {Scheduling agent: Reselection algorithm optimization scheme; Fault detection agent: Secondary diagnosis to confirm the fault; Worker agent: Execute parameter adjustments and provide feedback on the effect; Routing coding agent: Correct coding and supplement missing fields}, outputting the adjusted results and basis, until the condition of {No fault monitoring / Optimal solution output} is met." In one embodiment, such as Figure 1 and Figure 2 As shown, taking container terminal transportation as an example, the process includes the following steps: Step 1. Construct a scheduling mathematical model for a container terminal scenario, comprising three core components: yard cranes, automated guided vehicles (AGVs), and quay cranes; assuming... N ={1, 2, …} represents the set of natural numbers. i and j The representative indexes, symbols, and definitions are shown in Table 1. Table 1 Symbols and Definitions Table 1 provides the symbols and definitions required to construct the dual-loop model of the automated guided vehicle (AGV), including the definition of task points, the Gaussian distribution of the quay crane operation fitting, the transportation mode of the yard crane, and key parameters such as the waiting time of the AGV under different conditions.

[0038] The variables and symbols are defined as follows: The relevant constraints are as follows: Formulas (3.6) to (3.7) describe the transport of containers by quay cranes. i The shortest time and the longest time.

[0039] Equation (3.8) describes the uncertainty of quay crane operations, indicating that whether it is a container loading or unloading operation, the quay crane will [ , Complete container delivery within the specified timeframe. i Transportation.

[0040] Formulas (3.9) to (3.10) describe the process of transporting containers in the yard. Formula (3.9) indicates that when the yard cranes of the target yard cannot be prepared in advance, the yard cranes must move from the position of the previous container to the I / O area in order to grab the container. Therefore, the time for loading the container is the time for the yard cranes to move to the I / O area plus the time for the cranes to grab the container. When the yard cranes of the target yard can be prepared in advance, the time for loading the container is the time for the cranes to grab the container.

[0041] Formula (3.10) indicates that when the yard cranes in the target yard cannot be prepared in advance, the yard cranes must move from the position of the previous container to the position of the target container, grab the container, move it to the I / O area, and then use the gantry crane to unload the container. The time required to unload the container is the time it takes for the yard crane to reach the position of the target container (the maximum of the time it takes for the yard crane to move to the row of the target container and the time it takes for the gantry crane to move to the column of the container in the constraints) plus the time it takes for the gantry crane to grab the target container plus the time it takes for the crane to move to the I / O area (the maximum of the time it takes for the yard crane to move to the I / O area and the time it takes for the gantry crane to move to the I / O area in the constraints) plus the time it takes for the gantry crane to load and unload the container. When the yard cranes in the target yard can be prepared in advance, the time it takes to unload the container is the time it takes for the gantry crane to unload the container.

[0042] Equation (3.11) describes LTQ. For containers... i LTQ is a quay crane that completes container handling. i-1 The time of transport.

[0043] Formulas (3.12) to (3.14) describe the specific situations of automated guided vehicles (AGVs) entering and exiting quay cranes. Formula (3.12) indicates that the time interval between two consecutive AGVs leaving the quay crane is the time it takes for the second AGV to complete loading. Formula (3.13) describes the waiting time for the AGV to arrive at the first quay crane for unloading, which is the sum of the waiting time at the previous quay crane and the unloading time at the previous quay crane. Formula (3.14) describes the waiting time for the AGV to arrive at the second quay crane for loading, which is the sum of the waiting time at the previous quay crane and the loading time at the previous quay crane.

[0044] Formulas (3.15) and (3.16) describe the specific situation of Automated Guided Vehicles (AGVs) entering and leaving the yard. Formula (3.15) describes the waiting time before the AGVs arrive at the yard to unload containers. Specifically, when the target yard can be prepared in advance, the waiting time is 0. When the target yard cannot be prepared in advance, there are two situations. If the yard is in the process of loading containers, the waiting time is the loading waiting time of the previous task point plus the loading time minus the time interval between the arrival of the previous task point and the yard. This result may be less than 0, so the result needs to be compared with 0 and the maximum value is taken. If the yard is in the process of unloading containers, the waiting time is the unloading waiting time of the previous task point plus the unloading completion time minus the time interval between the arrival of the previous task point and the yard. This result may also be less than 0, so the result needs to be compared with 0 and the maximum value is taken.

[0045] Formula (3.16) describes the time required for the automated guided vehicle (AGV) to complete unloading at the transport task point. Specifically, if the target yard can be prepared in advance, the unloading completion time is the time it takes for the gantry crane to pick up the container in the I / O area. If the target yard cannot be prepared in advance, there are two situations: when the previous operation in the yard was a loading operation, since the yard crane will stop in the I / O area after the operation, the unloading time is the time it takes for the gantry crane to pick up the container; when the previous operation in the yard was a unloading operation, the unloading time is the maximum of the time it takes for the crane and gantry crane to move to the row and column of the special position, respectively, plus the time it takes for the gantry crane to pick up the container.

[0046] Step 2. Employ a large language model (LLM) to drive a multi-agent system, comprising four types of agents: scheduling agents, fault detection agents, working agents, and routing coding agents. In this system, LLM-driven agents collaborate to achieve resource scheduling and fault detection functions. Agent types include resource scheduling agents, fault detection agents, and working agents (WAs)—the latter referring to agents that reproduce the task scenario within the system. Taking a container terminal scenario as an example, working agents can be divided into three categories: yard crane agents, automated guided vehicle agents, and quay crane agents, as shown in the following steps: Step 2-1, Set up the scheduling agent container scheduling algorithm tool: Step 2-2: Set the fault detection agent's detection algorithm to the time dynamic warping algorithm: Identify the common detection parameters included in the scene: Table 1. Crane Failures in the Yard Table 2 AGV Faults Table 3. Faults of quay cranes Step 3: The encoding agent receives the initial parameters and encodes them. As the interface between physical world data and the multi-agent system, the encoding agent is responsible for converting the initial parameters of the container terminal scenario into a standardized, structured encoding format, providing unified data input for the scheduling agent and the fault detection agent. The specific steps are as follows: Step 3-1: Initial parameter reception and classification. The coded agent receives multi-source initial parameters from the physical system and classifies them into four categories according to data type: Basic task parameters: Container information: total number of containers (num_containers), IDC_i for each container (i=1,2,...,num_containers), priority (P_i) (P_i∈{1,2,3}, 1 is the highest), and target stacking area (Area_i). Time constraints: minimum transport time (min_time), maximum transport time (max_time), time interval (time_interval), minimum output (min_output) Equipment resource parameters: Automated Guided Vehicles (AGVs): Quantity num_agvs, IDA_j of each AGV (j=1,2,...,num_agvs), Maximum load Load_{A_j} Yard cranes: quantity num_yard_cranes, IDY_k for each yard crane (k=1,2,...,num_yard_cranes) Quay cranes: quantity num_quay_cranes, IDQ_l for each quay crane (l=1,2,...,num_quay_cranes) Fault detection baseline parameters: Based on the normal parameter range of the equipment in Table 1-3, a baseline parameter matrix Base is constructed: Base = { "yard_crane": {"vibration": [0.5, 2], "oil_particle": [18,16,13], ...}, "agv": {"signal_strength": [-60, -30], "battery_voltage": [48,54], ...}, "quay_crane": {"boom_angle": [0, 30], "cable_tension": [10, 20],...}} Scene association parameters: Dock layout: Dist matrix between areas (Dist[Area_i][Area_j] represents the distance from area i to area j) Equipment association rules: such as "Shore crane Q_l can only cooperate with AGVA_1-A_5". Step 3-2: Parameter standardization and format conversion. The coded agent converts the original parameters into structured codes that the agent can parse using the following rules: Numerical standardization: The time parameter is normalized as follows: t_norm = (t - min_time) / (max_time - min_time) (t_norm∈[0,1]) Physical quantities such as equipment load and distance should be converted to industry standard units (e.g., distance should be standardized to meters, and time to seconds). Encoding format definition: JSON structure is adopted as the general encoding format, containing four core fields: task_info, resource_info, fault_base, and scene_rules. An example framework is shown below: { "task_info": { "containers": [{"id": "C1", "priority": 1, "target_area": ​​"A3", "t_min": 60, "t_max": 180}, ...], "constraints": {"time_interval": 300, "min_output": 5}}, "resource_info": { "agvs": [{"id": "A1","max_load": 20}, ...], "yard_cranes": [{"id": "Y1"}, ...], "quay_cranes": [{"id": "Q1"}, ...]}, "fault_base": {...}, / / Corresponding to the Base matrix in step 3-1 "scene_rules": {"cooperation": [{"quay_crane": "Q1", "agvs": ["A1", "A2"]}, ...]}} Steps 3-4: Output the encoding result The final output structured encoding Encoded_Data must satisfy: Compatibility: The container scheduling algorithm (step 2-1) can be directly parsed by the schedulable agent to extract input parameters such as num_containers and num_agvs. Scalability: The reserved field ext is used to add new scenario parameters (such as weather impact coefficient). Security: Hash verification (Hash(Encoded_Data)) ensures that data transmission has not been tampered with. After completing the above steps, the coded agent enters a listening state, waiting to receive parameter change instructions or the agent's encoding query request.

[0047] Step 4: Perform cooperative scheduling based on the resource scheduling and fault detection interaction paradigm (RFIP). Figure 3 This paper presents the RFIP framework for an LLM-driven multi-agent system. In this framework, a large language model serves as the foundation for the agents, enabling them to autonomously parse tasks and collaborate with other agents. Each agent dynamically selects algorithms, evaluates constraints, and adapts to environmental feedback using a generative large language model. The steps are as follows: Step 4-1: Initial data and files are first input into the coding agent. This agent processes the data according to the workflow of the actual task scenario, ultimately generating agent flow code, which is then transmitted to RFIP. Step 4-2: The scheduling agent and the fault detection agent use this code as input data to perform resource scheduling and fault detection.

[0048] Step 4-3: Recreate the key facilities in the task scenario as working agents, each equipped with its own data and fault documentation library. Step 4-4: The fault information processed by the working intelligent agent will be fed back to the scheduling intelligent agent for further optimization of the resource scheduling scheme, ultimately realizing the system's self-adjustment and iteration, and outputting the optimal scheduling strategy.

[0049] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, is used to implement the above-described multi-agent resource scheduling method based on a large language model in an industrial scenario.

[0050] This invention provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned multi-agent resource scheduling method based on a large language model in an industrial scenario.

[0051] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0052] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure one One or more processes and / or boxes Figure one A device that provides the functions specified in one or more boxes.

[0053] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure one One or more processes and / or boxes Figure one The function specified in one or more boxes.

[0054] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure one One or more processes and / or boxes Figure one The steps of the function specified in one or more boxes.

Claims

1. A method for multi-agent resource scheduling based on a large language model in an industrial scene, characterized in that, Comprising the following steps: (1) accept the initial scheduling data of the target task, and generate a formatted code through a routing coding agent; (2) according to the formatted code, the scheduling agent calls a heuristic scheduling algorithm tool to generate a preliminary resource scheduling scheme; (3) the fault detection agent collects real-time index data of the working agent and the resource scheduling scheme generated by the scheduling agent, and cooperates with the dynamic time warping algorithm to perform fault diagnosis; If no fault is diagnosed, the scheduling scheme is output; If a fault is diagnosed, fault information is generated; (4) the working agent verifies the fault information and adjusts the parameter adjustment instruction in combination with the self-fault document feedback parameter; (5) the scheduling agent adjusts the scheduling parameters according to the feedback of the working agent, and re-outputs the scheduling scheme; (6) repeat steps (3) to (5) until the scheduling scheme is diagnosed without fault, and output the scheduling scheme.

2. The method of claim 1, wherein, The step (2) comprises the following sub-steps: (2.1) Integrated heuristic scheduling algorithm, establish classification tag system according to algorithm applicable scene characteristics, define algorithm adaptation degree function F ( A , S ), when F ( A , S ) ≥ α Mark as candidate algorithm, wherein A Algorithm type S Scene feature vector α Adaptation threshold (2.2) Extracting scene feature parameters in format coding S = [ n , m , t max , N min ], combined with specific scenes through formula A matched = argmax{ F ( A , S )} selected algorithm A matched ; wherein, n is the total number of tasks, m is the total number of devices, t max is the maximum task duration, N min is the minimum output (2.3) for the selected algorithm A matched the scheduling agent automatically generates an initial parameter set based on constraints in the formatted encoding; (2.4) the scheduling agent calls the algorithm tool to execute the selected algorithm A matched generates the resource scheduling scheme.

3. The method of claim 2, wherein, The step (2.3) comprises: (2.3.1) for each task k the generation processing time t k , ensuring t min(k) ≤ t k ≤ t max(k) ; wherein t min(k) = μ k - σ k , t max(k) = μ k + σ k , t min(k) is the shortest time for the task k to be processed, t min(k) is the longest time for the task k to be processed, μ k is the historical average processing time for the task k , σ k is the standard deviation; (2.3.2) Initialization of the device m set of tasks S m = ∅ total processing time T m = 0, the initial load threshold is set in combination with the average load rate of the device in the historical tasks T load(m) = λ × C m , λ is the load coefficient, C m is the maximum load capacity of the device m , and Σ t k ( k ∈ S m ) ≤ T load(m) ; (2.3.3) for the time interval Δt and the minimum output quantity N min , a dynamic check threshold is set N alert = β × N min , β is a pre-warning coefficient.

4. The method of claim 3, wherein, The step (2.4) comprises: (2.4.1) Traverse all tasks, store the task processing time of the tasks that meet the condition of step (2.3.1) generated in the time matrix t min(k) ≤ t k ≤ t max(k) ; and store to the time matrix T = [ t 1, t 2,..., t n ]; (2.4.2) for each device m Initialize the set of tasks S m , the total processing time T m Initialize to the sum of the times of the tasks within the set T m = Σ t k for k ∈ S m , synchronize the record to the device load tracking table; (2.4.3) Iteratively optimize task allocation by particle swarm algorithm, adjust task-device mapping relationship according to velocity update formula in each iteration, update total processing time of device T m' = T m + t k , and update the global total processing time T total to the maximum value of the completion time of all devices T total = max( T 1, T 2,..., T m ); (2.4.4) In the time window cycle, check each time interval Δt Number of tasks completed N completed Whether to reach N min , if N completed < N alert Trigger an early warning and mark it as a scheme optimization point, repeat step 4-3, regenerate task allocation; until all N completed > N alert , finally output the scheduling scheme.

5. The method of claim 1, wherein, The working process of the fault detection agent comprises the following steps: S1. The fault detection intelligent agent collects the real-time operation indicators uploaded by the working intelligent agent X ( t )=[ x 1(t) , x 2(t) ,..., x d(t) ] and pre-processes the collected data to generate a standardized index sequence X norm(t) = ( X ( t ) - μ X ) / σ X ; wherein d is the index dimension, t is the timestamp, μ X is the index mean, σ X is the standard deviation; S2. Obtain the resource allocation parameter in the resource scheduling scheme generated by the scheduling agent Y = [ y 1, y 2,..., y d ]; S3. The fault detection agent invokes the dynamic time warping algorithm tool to compare the standardized indicator sequence with the equipment normal operation benchmark data sequence in the industrial knowledge base. X ref(t) Perform time series similarity comparison and calculate the deviation value: DTW( X norm , X ref )= min{Σ| X norm(i) - X ref(j) | × w ( i , j )}, when DTW( X norm , X ref ) > θ The system determines when a fault exists and calculates its severity. F level = DTW( X norm , X ref )) / θ , F level A level assessment is triggered when the value is >1; among which... w ( i , j ) represents the path weights in dynamic programming. θ The preset deviation threshold; S4. If a fault exists, generate fault information. F info = { y d , S The confirmed fault information is fed back to the scheduling agent via the RFIP protocol, triggering the scheduling scheme optimization process, where the fault level is determined by... L = ceil( F level The system is divided into sections, and the scheduling agent performs tasks based on different fault levels. L The scheduling scheme is optimized through a self-loop; among which... L =1, 2, or 3 L When the value is 3, a new suitable scheduling algorithm is selected, and the process repeats three more times to select the optimal scheduling scheme after the loop. L When < 3, the scheduling agent performs L In the next iteration, select the optimal scheduling scheme after the previous iteration; S5. After the scheduling agent generates an optimized scheme, the diagnosis process of steps S1-S3 is repeated to verify the effect of the optimized scheme on solving the fault, and if there is no fault, the scheduling optimization scheme is output, and if there is a fault, step S4 is executed to optimize the scheduling scheme.

6. The method of claim 5, wherein, The interaction logic of RFIP comprises: Step 1: The scheduling agent generates a preliminary resource scheduling scheme S scheme Afterwards, automatically synchronize to the fault detection agent, and trigger the working agent to perform the indicators W metrics Collection, to ensure that the fault detection agent can obtain S scheme Associated indicator data; Step 2: After receiving the preliminary scheme, the fault detection intelligent agent calls the dynamic time warping algorithm to analyze the similarity of the time series data and the data of the working intelligent agent S scheme and the working intelligent agent W metrics When the deviation value exceeds the threshold value set in the working intelligent agent fault document, it is considered that there is a fault, and the relevant index data and the document are returned S doc and the document are returned W metrics-s If the threshold value is not exceeded, it is determined that there is no fault, and the scheduling scheme is output S doc-s ; If the threshold value is not exceeded, it is determined that there is no fault, and the scheduling scheme is output S scheme ; Step 3: Fault detection intelligent agent generates fault information F info After that, it is distributed to the corresponding working intelligent agent first, which queries W metrics Performs matching verification, and if the verification is consistent, it feeds back S scheme Adjusts the suggestion, and if the verification is inconsistent, repeats step 2; Step 4: After receiving the adjustment suggestion, the dispatching agent combines the fault information generated by the fault detection agent F info According to the fault information F info The dispatching agent generates an optimized dispatching scheme S scheme-pre And synchronizes it again to the fault detection agent until the fault detection agent monitors no fault warning and outputs the final dispatching scheme S scheme-final .

7. A large language model driven multi-agent resource scheduling and fault monitoring system, characterized in that, Comprising: a scheduling agent, a fault detection agent, a working agent and a routing coding agent; a routing coding agent for accepting initial scheduling data of a target task and generating a formatted code a scheduling agent for calling a heuristic scheduling algorithm tool to generate a preliminary resource scheduling scheme according to the formatted code; and if the generated resource scheduling scheme is diagnosed with a fault, adjusting the scheduling parameters according to the feedback of the working agent, re-outputting the scheduling scheme, until the fault detection agent diagnoses no fault, and outputting the final scheduling scheme; a fault detection agent for receiving real-time index data of the working agent and the resource scheduling scheme generated by the scheduling agent, and cooperating with the dynamic time warping algorithm to perform fault diagnosis; If a fault is diagnosed, fault information is generated; a working agent for collecting and storing real-time index data of the working agent; and verifying the fault information and adjusting the parameter adjustment instruction in combination with the self-fault document feedback parameter.

8. The system of claim 7, wherein, Further comprising: an integrated scheduling agent algorithm tool library containing multiple heuristic scheduling algorithms, a parameter optimization module and a scheme generation module, wherein the scheduling agent selects the algorithm tool according to the scene characteristics; specifically: The scheduling agent autonomously selects an algorithm tool according to the scene characteristics to identify the scene characteristic parameters through formatted coding S feat Output the matched heuristic algorithm from the algorithm tool library A matched ; calling the parameter optimization module to automatically generate an initial parameter set based on the constraint conditions in the formatted code; The scheme generation module then calls the selected heuristic algorithm A matched generating a scheduling scheme S scheme ; constructing a fault detection agent data collection interface, a feature engineering module, a dynamic time warping algorithm execution unit and a fault evaluation module; specifically: Receiving the index data stored by the working agent through the data acquisition interface W metrics And the scheduling scheme generated by the scheduling agent S scheme ; calling the feature engineering module to preprocess the collected data and generate standardized index sequences X norm(t) ; dynamic time warping algorithm execution unit and fault evaluation module, for calling dynamic time warping algorithm to detect whether there is a problem with the index data and perform fault evaluation to generate fault information F info ; configuring a working agent fault document storage module: collecting its own running parameters and responding to fault verification instructions; specifically: According to the scene characteristic parameters S feat , configure the scene-related fault document S doc and index data W metrics , receive the fault verification instruction sent by the fault scheduling intelligent agent at the same time, and return the related index data if there is a fault W metrics-s and the corresponding fault document S doc-s , and return the no-fault flag if there is no fault S safe .

9. A computer-readable storage medium having stored thereon a program, characterized in that, When the program is executed by a processor, it is used to implement the multi-agent resource scheduling method based on a large language model in an industrial scene according to any one of claims 1-6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the multi-agent resource scheduling method based on a large language model in an industrial scene according to any one of claims 1-6.

Images