Multi-agent cooperative particle accelerator commissioning method, device, equipment and medium

Abstract

The application discloses a kind of multi-agent collaborative particle accelerator debugging method, device, equipment and medium, it is related to particle accelerator debugging technical field, by deploying multi-agent collaborative system based on large language model, state perception, task planning, simulation prediction and execution control are modularized and closed-loop cooperation, realize particle accelerator multi-parameter linkage optimization adjustment, while shortening the debugging cycle, improve the overall efficiency of particle accelerator operation maintenance.Therein method includes: using state perception intelligent agent to collect the multi-source data of accelerator control system in real time;Using task planning intelligent agent to analyze abnormal information, retrieve similar cases in knowledge base, generate preliminary debugging scheme containing multiple steps in combination with pre-set tool box;Using execution intelligent agent to convert the debugging scheme data predicted by simulation into control instruction and issue to accelerator control system, to debug through accelerator control system, obtain debugging result.

Description

Technical Field

[0001] This application relates to the field of particle accelerator commissioning technology, and in particular to a multi-agent collaborative particle accelerator commissioning method, apparatus, equipment and medium. Background Technology

[0002] High-current, high-power particle accelerators are complex, nonlinear mega-systems composed of tens of thousands of components, demanding extremely high levels of long-term stable operation. As the scale of large scientific facilities expands and operational requirements become increasingly stringent, traditional operation and maintenance models are becoming unsustainable. In recent years, large language models, with their powerful logical reasoning and knowledge integration capabilities, have provided a novel technological approach to solving the challenges of automated operation and maintenance of large facilities and achieving intelligent assisted decision-making.

[0003] Current particle accelerator commissioning methods primarily rely on manual operations, with expert teams using their experience to consult scattered documents, analyze monitoring data, and manually adjust parameters. Although rule-based or traditional machine learning-assisted tools have been introduced in some stages, these systems are mostly limited to local monitoring of single subsystems. When faced with complex and unknown faults, the lack of multi-role collaboration mechanisms between subsystems makes it difficult to achieve coordinated adjustments of multiple parameters under the global optimal goal, resulting in long commissioning cycles and severely restricting the operational efficiency of particle accelerators. Summary of the Invention

[0004] In view of this, this application provides a multi-agent collaborative particle accelerator debugging method, apparatus, equipment and medium. The main purpose is to solve the problem that the existing particle accelerator debugging method that relies on manual operation mode is difficult to achieve multi-parameter linkage adjustment under the global optimal goal, resulting in a long debugging cycle and seriously restricting the operating efficiency of particle accelerator.

[0005] Firstly, this application provides a multi-agent collaborative particle accelerator debugging method, wherein the method is collaboratively executed by a multi-agent system based on a large language model, the multi-agent system including at least a state-aware agent, a task-planning agent, a simulation-prediction agent, and an execution agent; the method includes: The state-aware intelligent agent collects multi-source data from the accelerator control system in real time. When an anomaly is detected in the particle accelerator state based on the multi-source data, the anomaly information is converted into a natural language description and sent to the task planning intelligent agent. The task planning agent is used to analyze the abnormal information, retrieve similar historical cases from the knowledge base, and generate a preliminary debugging plan with multiple steps by combining the pre-set toolbox. The simulation prediction agent is used to perform simulation prediction on the preliminary debugging scheme based on the physical model; if the simulation prediction result meets the set conditions, the debugging scheme data predicted by the simulation is generated; if the simulation prediction result does not meet the set conditions, the simulation prediction result is fed back to the task planning agent to re-plan the preliminary debugging scheme. The execution agent converts the simulation-predicted debugging scheme data into control commands and sends them to the accelerator control system for debugging, thereby obtaining the debugging results.

[0006] Secondly, a multi-agent collaborative particle accelerator debugging device is provided. The device deploys a multi-agent system based on a large language model, the multi-agent system including at least a state-aware agent, a task-planning agent, a simulation-prediction agent, and an execution agent; the device includes: The monitoring unit is used to collect multi-source data of the accelerator control system in real time using the state-aware intelligent agent. When an abnormality is detected in the state of the particle accelerator based on the multi-source data, the abnormal information is converted into a natural language description and sent to the task planning intelligent agent. The generation unit is used to analyze the abnormal information using the task planning agent, retrieve historical similar cases in the knowledge base, and generate a preliminary debugging plan containing multiple steps by combining a pre-set toolbox. The simulation prediction unit is used to perform simulation prediction on the preliminary debugging scheme based on the physical model using the simulation prediction agent; if the simulation prediction result meets the set conditions, the debugging scheme data predicted by the simulation is generated; if the simulation prediction result does not meet the set conditions, the simulation prediction result is fed back to the task planning agent to re-plan the preliminary debugging scheme. The debugging unit is used to convert the simulation-predicted debugging scheme data into control commands and send them to the accelerator control system through the execution agent to perform debugging and obtain debugging results.

[0007] Thirdly, a multi-agent collaborative particle accelerator debugging device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor. When the processor executes the program, it implements the aforementioned multi-agent collaborative particle accelerator debugging method.

[0008] Fourthly, a storage medium is provided on which a computer program is stored, which, when executed by a processor, implements the aforementioned multi-agent collaborative particle accelerator debugging method.

[0009] By employing the aforementioned technical solutions, this application provides a multi-agent collaborative particle accelerator debugging method, apparatus, equipment, and medium. Compared to existing methods relying on manual operation for particle accelerator debugging, this application deploys a multi-agent collaborative system based on a large language model. This system modularizes and coordinates state perception, task planning, simulation prediction, and execution control in a closed-loop manner. On one hand, it can collect and intelligently identify particle accelerator operational anomalies in real time, automatically converting anomaly information into standardized natural language descriptions, avoiding delays and misjudgments caused by manual intervention, and significantly improving the efficiency and accuracy of fault perception and problem transmission. On the other hand, by retrieving historical similar cases from a knowledge base and combining them with a pre-built algorithm toolbox to generate debugging schemes, it can fully reuse historical operation and maintenance experience, significantly shortening the generation cycle of debugging schemes. Simultaneously, the simulation prediction agent performs pre-verification of the debugging scheme based on the physical model, effectively preventing ineffective or degrading operations from directly affecting the physical equipment. Through the organic collaboration of multiple agents, multi-parameter linkage optimization and adjustment of the particle accelerator are achieved, shortening the debugging cycle while improving the overall efficiency of particle accelerator operation and maintenance.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a multi-agent collaborative particle accelerator debugging method in one embodiment of this application. Figure 2 yes Figure 1 A flowchart illustrating a specific implementation method of step 101; Figure 3 yes Figure 1 A flowchart illustrating a specific implementation method for step 102; Figure 4 yes Figure 1 A flowchart illustrating a specific implementation method for step 103; Figure 5 yes Figure 1 A flowchart illustrating a specific implementation method following step 103; Figure 6 yes Figure 1 A flowchart illustrating a specific implementation method following step 104; Figure 7This is a flowchart illustrating the multi-agent collaborative particle accelerator commissioning process in one embodiment of this application. Figure 8 This is a schematic diagram of the structure of a multi-agent collaborative particle accelerator debugging device in one embodiment of this application; Figure 9 This is a schematic diagram of the device structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0012] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0013] In related technologies, particle accelerator commissioning primarily relies on manual operation, with expert teams using their experience to consult scattered documents, analyze monitoring data, and manually adjust parameters. Although rule-based or traditional machine learning-assisted tools have been introduced in some stages, these systems are mostly limited to local monitoring of single subsystems. When faced with complex and unknown faults, the lack of multi-role collaboration mechanisms between subsystems makes it difficult to achieve multi-parameter linkage adjustments under the global optimal goal, resulting in long commissioning cycles and severely restricting the operational efficiency of particle accelerators.

[0014] To address this problem, this embodiment provides a multi-agent collaborative particle accelerator debugging method. This method is executed collaboratively by a multi-agent system based on a large language model. The multi-agent system includes at least a state-aware agent, a task-planning agent, a simulation-prediction agent, and an execution agent. Figure 1 As shown, the method includes the following steps: 101. The state-aware intelligent agent collects multi-source data from the accelerator control system in real time. When an anomaly is detected in the particle accelerator state based on the multi-source data, the anomaly information is converted into a natural language description and sent to the task planning intelligent agent.

[0015] In this embodiment, the state-aware agent, acting as the sensing front end of a multi-agent collaborative system, is responsible for continuously collecting multi-source heterogeneous data from the particle accelerator control system. Its data sources broadly cover the underlying control system architecture, such as control networks based on EPICS, TANGO, or DOOCS protocols, as well as various key beam diagnostic devices. Specific data acquisition targets include, but are not limited to: beam position monitors, AC current transformers, beam loss detectors, beam scrapers, vacuum gauges, temperature sensors, and level gauges. To address the noise and asynchronous nature of industrial field data, the agent incorporates signal processing and time alignment modules to process the raw acquired data in real time. The processing flow specifically includes sampling synchronization, digital filtering, outlier detection, and missing value completion, ultimately fusing the multi-source data to generate a standardized state vector.

[0016] The aforementioned state vector is equivalent to a real-time data structure containing multi-dimensional information, the main contents of which include: (1) key beam parameters, such as inlet and outlet current intensity, track eccentricity, beam loss rate, etc.; (2) equipment operating condition data, such as magnet power supply current, phase and amplitude readings of the radio frequency cavity; (3) environmental quantity data, such as tunnel temperature, cooling water level, etc.; (4) confidence score and precise timestamp corresponding to each data item. On this basis, the state-aware agent can use deep learning neural network models represented by long short-term memory networks as a technical means for particle accelerator state monitoring and trend prediction. By learning from historical normal operation data, a dynamic benchmark for state evolution is established, which can keenly capture the small drift of system operating condition or beam state. When the deviation between the real-time input state vector and the model prediction value exceeds the set threshold, that is, when it is determined that the system operating condition or beam state has an abnormal drift, the state-aware agent will immediately trigger an alarm mechanism and promptly send abnormal signals and related context information to the task planning agent.

[0017] During the specific process of transforming abnormal information, if the model detects an abnormality in a specific indicator, the state-aware agent will call a preset natural language generation prompt word template to transform the abstract state drift data into a natural language description that conforms to human logic.

[0018] Through the instantiation of the above template, structured data can be dynamically generated into specific natural language instructions. Taking a real-world scenario as an example, when the beam loss monitor detects that the beam loss rate exceeds 3%, the agent automatically extracts fields such as the current accelerator name, fault location, monitoring equipment type, specific indicator name, comparison relationship, and target threshold, and fills them into the following prompt template: "The value of [a certain segment or thermostat] of accelerator [accelerator name] exceeds the threshold, the indicator [indicator name] is [greater than / less than / equal to] the target threshold, and the online repair process for [problem name] needs to be initiated." Correspondingly, the generated description is: "The value of [superconducting segment] [beam loss detector] of accelerator [CAFE2] exceeds the threshold, the indicator [beam loss rate] is [greater than] [3%], and the online repair process for [beam loss] needs to be initiated." This natural language description not only clearly indicates the fault phenomenon and quantitative indicator but also implies the initial handling intention. It is then encapsulated into a message packet and sent to the task planning agent, thereby triggering the subsequent fault diagnosis and debugging scheme generation process.

[0019] 102. Utilize the task planning agent to parse the abnormal information, retrieve historical similar cases from the knowledge base, and generate a preliminary debugging plan containing multiple steps by combining the preset toolbox.

[0020] In this embodiment, the task planning agent, acting as the control terminal of the multi-agent collaborative system, first executes an anomaly information parsing task after receiving the anomaly description sent by the state-aware agent. Utilizing the semantic understanding capabilities of a large language model, the task planning agent extracts key fault features from the input text, including the faulty device identifier, current value, target threshold, and fault type. Based on the parsing results, the task planning agent constructs a structured query vector and searches for similar problems in a knowledge base. The knowledge base stores successful debugging cases, expert experience records, and fault handling reports accumulated during past accelerator operations. Through vector similarity matching, the agent quickly locates historical successful cases highly similar to the current key fault features, extracts their solution strategies, key parameter adjustment ranges, and final effects, and uses these as a reference benchmark for this task planning.

[0021] Subsequently, the task planning agent, based on a pre-built, callable toolkit, and combining current fault characteristics with retrieved historical cases, conducts in-depth problem analysis and process planning. The toolkit pre-integrates various core algorithm modules, including but not limited to beam trajectory correction algorithms, emittance measurement algorithms, transmission efficiency optimization algorithms, beam loss localization algorithms, and phase scanning algorithms. The task planning agent logically combines these tools according to the specific nature of the problem to be solved, breaking down the abstract debugging objective into a series of verifiable and executable sub-tasks. In this process, the task planning agent not only considers the function of individual tools but also comprehensively considers the causal dependencies between steps, ensuring that the generated process conforms to physical laws and operational specifications, forming a complete logical chain from diagnosis to repair.

[0022] Finally, the mission planning agent outputs a preliminary debugging plan comprising multiple steps. This plan is presented as structured data, clearly defining the execution purpose and tool invocation method for each step. Taking the detection of "LINAC-A L3 beam loss monitor value of 3.5%" as an example, the generated preliminary debugging plan is as follows: First, call the loss localization tool to analyze the beam loss data collected by all loss monitors in this segment to accurately locate the high-loss point; Second, call the orbit reconstruction tool to compare the real-time BPM data with the reference orbit to check for orbit drift; Third, if the orbit deviation is confirmed to be significant, call the local correction tool to calculate the adjustment amount of the correction coil; Fourth, call the simulation prediction tool to evaluate the expected effect of the candidate adjustment scheme on reducing beam loss in a virtual environment; Fifth, call the safety verification tool to perform compliance checks on the generated magnet current adjustment value and change rate to ensure that it does not exceed the equipment safety boundary; Sixth, call the execution compilation tool to convert the verified adjustment parameters into a low-level instruction sequence that the accelerator control system can recognize. Through this process, the task planning agent successfully transforms complex fault recovery tasks into standardized, automated multi-step execution schemes, which can then be used by subsequent agents for simulation verification and actual operation.

[0023] 103. The simulation prediction agent is used to perform simulation prediction on the preliminary debugging scheme based on the physical model.

[0024] In this embodiment, the simulation prediction agent, acting as a virtual verification terminal in a multi-agent collaborative system, immediately initiates a physical model-based simulation evaluation process upon receiving a preliminary debugging plan containing multiple steps submitted by the task planning agent. The simulation prediction agent first dynamically constructs or updates a high-fidelity beam dynamics model based on the accelerator's current real-time operating parameters, such as magnet current, RF phase, and beam energy. Subsequently, the agent invokes a pre-built set of professional simulation tools, including but not limited to modules for beam trajectory simulation, phase space evolution simulation, loss distribution prediction, and cavity voltage and phase response simulation, to perform independent numerical simulations of each candidate operation step in the preliminary debugging plan. By simulating the execution process of different operation plans in a virtual environment, the simulation prediction agent can quantitatively evaluate their impact on key performance indicators, specifically including changes in beam loss rate, convergence of trajectory offset, evolution trend of emittance, degree of improvement in matching quality, and improvement in current transmission efficiency.

[0025] After completing the simulation calculations, the simulation prediction agent generates a detailed quantitative evaluation report, which includes comparable prediction data such as the expected loss reduction ratio, the orbital error convergence curve, and the system stability trend. The simulation prediction agent compares these prediction results with the set conditions. Examples of set conditions include: the expected beam loss must be reduced to below 1%, the root mean square error of the orbit must be less than 0.5 mm, and all equipment parameters must be within the safety threshold range. If the simulation prediction results meet the set conditions, it indicates that the preliminary debugging plan is physically feasible and effective. The simulation prediction agent marks the preliminary debugging plan as the debugging plan data predicted by the simulation, along with the expected performance improvement indicators, and formally transmits it to the downstream execution agent, preparing for actual deployment. At the same time, the agent also records the characteristics of successful simulation cases to optimize subsequent prediction models.

[0026] If the simulation prediction results do not meet the set conditions—for example, the prediction shows that the beam loss can only be reduced to 2%, failing to reach the target; or a certain step causes the local magnetic field to exceed the safety limit; or all candidate solutions fail to achieve significant improvement—the simulation prediction agent will not blindly execute. Instead, it will trigger a closed-loop feedback mechanism, encapsulating detailed reasons for simulation failure, unmet indicator data, and specific risk points into structured feedback information and sending it back to the task planning agent. Accordingly, based on this feedback, the task planning agent, combined with other strategies or adjustment tool call logic in the knowledge base, will re-plan the initial debugging scheme, such as changing the correction algorithm, adjusting the step size, or introducing new constraints, and then resubmit it to the simulation prediction agent for a new round of iterative verification. Furthermore, the simulation prediction agent also undertakes the task of adaptive model updates: when there is a significant deviation between the subsequent actual execution results and the current simulation results, the simulation prediction agent will automatically call the model update tool to correct the beam model parameters using actual measurement data, thereby continuously improving the accuracy of subsequent simulation predictions.

[0027] 104. The execution agent is used to convert the simulation-predicted debugging scheme data into control commands and send them to the accelerator control system for debugging, so as to obtain the debugging results.

[0028] In this embodiment, the executing agent, acting as the active terminal of a multi-agent collaborative system, is responsible for converting the simulation-predicted debugging scheme data, after task planning, simulation prediction, and safety verification, into actual control commands that can be issued. First, the executing agent performs structured parsing of the abstract operation commands generated by the higher-level agent. Examples of abstract operation commands include: adjusting the Q5 magnet to +0.8 A and correcting the RF cavity phase to -5 degrees. The executing agent has a built-in communication adaptation module for a specific accelerator control system, which can automatically convert these high-level semantic commands into specific command formats recognizable by the lower-level devices. These commands cover power current settings, adjustment rate limits, synchronization trigger flags, and timeout protection parameters. Through this conversion process, the abstract debugging scheme is mapped into a series of precise, timing-controllable low-level operation sequences, ready to be issued to the accelerator control system via a standard communication interface.

[0029] During command issuance and debugging execution, the execution agent initiates a closed-loop tracking mechanism to comprehensively monitor the real-time response of critical equipment. The execution agent not only confirms whether the setpoint has been successfully sent to the controller, but also continuously polls or subscribes to feedback signals from the control system, focusing on: whether the control system returns an error code, whether the actual equipment value converges to the target value range within a specified time, and whether the fluctuation of the feedback signal is within the expected tolerance range. If, at any moment during execution, a critical parameter is detected to deviate from the normal range—for example, the current rise rate is too fast, the equipment response times out, or the feedback value oscillates and diverges—the execution agent will immediately trigger a safety circuit breaker mechanism: suspending all subsequent operation steps, issuing an emergency anomaly warning to the safety monitoring agent, and automatically performing a rollback operation or freezing the current action state to restore the equipment to a safe baseline, thereby effectively avoiding potential equipment damage or beam loss risks.

[0030] Once all steps have been successfully executed and the system feedback is stable, the executing agent is responsible for collecting and organizing the final debugging results. The executing agent archives all data from the execution process, including detailed information across multiple dimensions: first, the execution result determination, clearly indicating whether the task was successful, partially successful, or required manual intervention; second, high-resolution real-time feedback curves from the device, fully recording the dynamic response process from the moment the command was issued to the system reaching steady state; third, a detailed deviation analysis report, quantifying the static and dynamic tracking errors between the actual action parameters and the target setpoints; and fourth, microsecond-level global timestamps to ensure time alignment of multi-source data, as well as a snapshot of the environmental state at the moment of execution. This high-value data is written into the system's historical knowledge base, not only for tracing the current debugging task but also providing real-world empirical data support for subsequent model adaptive updates and planning strategy optimization. Through this mechanism, the executing agent successfully achieved the crucial implementation from virtual solutions to physical actions, completing a closed loop of planning, prediction, safety, and execution, ensuring the controllability, traceability, and high reliability of the accelerator's automated debugging process.

[0031] The method provided in this application, by deploying a multi-agent collaborative system based on a large language model, modularizes and coordinates state perception, task planning, simulation prediction, and execution control. On the one hand, it can collect and intelligently identify particle accelerator operational anomalies in real time, automatically converting anomaly information into standardized natural language descriptions, avoiding delays and misjudgments caused by manual intervention, and significantly improving the efficiency and accuracy of fault perception and problem transmission. On the other hand, by retrieving historical similar cases from a knowledge base and combining them with a pre-built algorithm toolbox to generate debugging plans, it can fully reuse historical operation and maintenance experience, significantly shortening the generation cycle of debugging plans. At the same time, the simulation prediction agent performs pre-verification of the debugging plan based on the physical model, effectively avoiding ineffective or degrading operations directly affecting the physical equipment. Through the organic collaboration of multiple agents, multi-parameter linkage optimization and adjustment of the particle accelerator can be achieved, shortening the debugging cycle and improving the overall efficiency of particle accelerator operation and maintenance.

[0032] To further clarify how a state-aware agent achieves anomaly detection and information transmission, the specific implementation process of step 101 above will be described in detail below, using a specific implementation scenario. Specifically, as follows... Figure 2 As shown, step 101 includes the following steps: 201. Real-time acquisition of multi-source data from the accelerator control system, and trend prediction of the multi-source data using a long short-term memory network model.

[0033] 202. When the multi-source data is detected to deviate from the predicted trend by more than a preset residual threshold or exceeds the physical safety threshold, it is determined that there is an abnormality in the state of the particle accelerator, and the abnormal information is extracted.

[0034] 203. Instantiate the abnormal information and fill it into the corresponding placeholder in the preset natural language prompt word template to generate a standardized natural language description, and send the standardized natural language description to the task planning agent.

[0035] In one specific application scenario of this embodiment, a state-aware agent is deployed at the control network node of a particle accelerator. It collects real-time heterogeneous data from multiple sources, including a beam position monitor, a magnet power supply current sensor, a vacuum meter, and a radio frequency cavity temperature probe, via a high-frequency interface. To accurately identify potential risks from the massive and noisy real-time data stream, the system incorporates a Long Short-Term Memory (LSTM) network model pre-trained with historical fault data. This model receives multi-source data over a past time series using a sliding window approach, leveraging its unique gating mechanism to capture long-term dependencies within the time series. This allows for high-precision prediction of the changing trends of key accelerator parameters over several future time steps, forming a dynamic baseline.

[0036] Once the real-time collected data flows into the system, the computing module immediately calculates the residual between the actual observed value and the predicted value output by the Long Short-Term Memory network model. The system employs a dual anomaly detection mechanism: firstly, if the absolute value of the residual of one or more key parameters continuously exceeds a preset residual threshold based on statistical distribution, it indicates an unexpected deviation from the equipment's operating trajectory; secondly, if any sensor reading directly exceeds a hard safety threshold defined by physical limits, an alarm is immediately triggered regardless of the predicted trend. Once either of these conditions is met, the state-aware agent immediately determines that the particle accelerator's current state is abnormal, and locks down the time window of the anomaly, the specific equipment number involved, the degree of deviation, and the current environmental context information, completing the structured extraction of the anomaly information.

[0037] During the alarm command generation phase, a natural language generation strategy is employed to improve the efficiency of human-computer interaction and downstream intelligent agents' understanding. Extracted structured anomaly information, such as specific details like a 15% drop in the current of quadrupole magnet No. 3 within 0.5 seconds or a sudden increase in vacuum chamber pressure exceeding safety limits, is automatically instantiated and filled into corresponding placeholders in a pre-designed natural language prompt template. This natural language prompt template is optimized to generate standardized natural language descriptions with correct grammar, clear logic, and causal inference suggestions. For example, it might state that an abnormal drop in the current of quadrupole magnet No. 3 could lead to beam trajectory distortion, and it is recommended to immediately check power supply stability. Finally, this standardized natural language description is encapsulated into a message packet and sent in real-time to the task planning agent. This serves as a key input for the agent to reassess the current task status, generate emergency intervention strategies, or adjust subsequent debugging steps, thus achieving seamless integration from low-level perception to high-level decision-making.

[0038] Furthermore, after the state-aware agent completes anomaly detection and sends the anomaly information, described in standardized natural language, to the task planning agent, the task planning agent will then proceed with solution planning based on this anomaly information. To more clearly illustrate how this agent progresses from receiving anomaly information to generating a preliminary debugging plan, specifically... Figure 3 As shown, step 102 includes the following steps: 301. Receive and parse the abnormal information in the form of natural language, and extract the faulty equipment identifier and the abnormal indicator features of the deviation prediction trend degree or the situation of exceeding the physical safety threshold.

[0039] 302. Obtain system topology association information corresponding to abnormal device identifiers based on the static configuration model of the accelerator control system, and combine the system topology association information with the abnormal feature indicators to construct a comprehensive retrieval feature.

[0040] 303. Based on the comprehensive retrieval features, retrieve historical similar cases with similar abnormal indicator features and matching system topology relationships in the knowledge base, and obtain the handling strategy information corresponding to the historical similar cases.

[0041] 304. Based on the handling strategy information and the abnormal indicator characteristics, select suitable debugging operation units from the preset toolbox, and sort and combine the selected debugging operation units according to the execution dependency to generate a preliminary debugging plan containing multiple steps.

[0042] In this embodiment, the task planning agent, acting as the decision-making center of the accelerator's intelligent operation and maintenance system, immediately initiates a deep analysis process upon receiving anomaly information in natural language form from the state-aware agent. Utilizing named entity recognition and dependency parsing algorithms from natural language processing, the agent accurately removes redundant descriptions from the text and extracts key structured data, including the specific device identifier that experienced the malfunction and quantitative anomaly indicator features.

[0043] To improve the accuracy of fault location and the relevance of solution generation, the task planning agent further utilizes the static configuration model of the accelerator control system. This static configuration model records in detail the physical connections between devices, signal transmission paths, and energy dependency logic. Based on the parsed faulty device identifier, the task planning agent traverses the static configuration model to obtain the local system topology association information of the abnormal device, such as the upstream power supply status, downstream beam diagnostic element readings, and the availability of parallel redundant devices. Subsequently, this topology association information is weighted and fused with the previously extracted anomaly indicator features to construct a multi-dimensional comprehensive retrieval feature vector. This combination not only considers the appearance of the fault itself but also incorporates the propagation and impact range of the fault within the complex accelerator network.

[0044] Based on the constructed comprehensive retrieval features, the task planning agent performs high-dimensional vector similarity retrieval in the built-in knowledge base. The knowledge base stores tens of thousands of past debugging and maintenance records, each annotated with fault characteristics, a system topology snapshot, and a final verified effective handling strategy. By calculating cosine similarity or Euclidean distance, historical similar cases with highly similar anomaly indicator features and matching system topology relationships are selected. For these highly similar cases, the task planning agent extracts their corresponding handling strategy information, including the correction algorithms used, the range of adjusted parameters, and the risk points avoided. By summarizing and weighted voting on the handling strategies of multiple similar cases, the direction of the handling strategy for the current fault is determined.

[0045] After determining the direction of the handling strategy, the mission planning agent enters the solution generation phase. The system maintenance has a pre-built toolbox that encapsulates various standardized debugging operation units, such as atomic-level functional modules for adjusting the quadrupole magnet gradient, performing beam closure correction, switching backup power channels, and initiating vacuum replenishment procedures. Based on the summarized handling strategy information and the current specific anomaly characteristics, the mission planning agent evaluates the contribution and potential risks of each operation unit to fault repair, selecting a suitable set of debugging operation units. Considering the strict timing constraints of accelerator operation, the mission planning agent performs topological sorting and logical combination of the selected operation units based on the data dependencies and physical causal logic between them, generating a preliminary debugging plan containing multiple ordered steps.

[0046] In practical applications, to ensure the technical feasibility and operational safety of the preliminary debugging plan before actual execution, and to avoid damage to the particle accelerator or beam degradation caused by improper operation, after the task planning agent generates the preliminary debugging plan, the simulation prediction agent will conduct virtual simulation and effect prediction of the aforementioned preliminary debugging plan based on the accelerator physical model. Specifically, such as... Figure 4 As shown, step 103 includes the following steps: 401. Obtain the current real-time operating status data of the particle accelerator as the initial conditions for simulation, and map each debugging operation unit included in the preliminary debugging scheme to the key physical variables and their time-series change curves in the accelerator physical model.

[0047] 402. Drive the accelerator physical model to perform multi-step simulation calculations in a virtual environment according to the time-series change curve, so as to simulate the dynamic response behavior of the particle accelerator during the execution of the preliminary debugging scheme, and generate a pre-simulation state sequence containing the trajectory of the key physical variables over time.

[0048] 403. Based on the pre-run state sequence, analyze the final steady-state characteristics of the particle accelerator after executing the preliminary debugging scheme and the instantaneous extreme value characteristics during the execution process.

[0049] 404. Determine whether the final steady-state characteristics meet the preset target operating indicators, and detect whether the instantaneous extreme value characteristics violate physical safety constraints, and generate simulation prediction results containing feasibility judgment conclusions and expected effect quantitative data.

[0050] In this embodiment, upon receiving the preliminary debugging scheme generated by the task planning agent, the simulation prediction agent first synchronously obtains a snapshot of the current real-time operating state from the accelerator's distributed control system as the initial boundary conditions for the simulation. This data includes, but is not limited to, the current beam intensity, energy dissipation, emittance, and the setpoints of each key magnet. Subsequently, the simulation prediction agent invokes the built-in scheme parsing engine to precisely map the abstract debugging operation units in the preliminary debugging scheme to the key physical variables in the underlying accelerator physical model and their expected change curves in the time dimension, thereby constructing a virtual control excitation signal.

[0051] After initialization and mapping, the simulation predictive agent drives a high-precision accelerator physics model to run in an isolated virtual computing environment. The physics model receives the aforementioned time-series variation curves as input and performs multi-step extrapolation calculations at minute time steps, dynamically simulating the transport and acceleration processes of charged particles under changing electromagnetic fields, as well as potential loss mechanisms. During this extrapolation, the system not only calculates the evolution of the beam's core parameters but also simultaneously simulates the thermodynamic and electromagnetic responses of the device itself. After the extrapolation, a detailed pre-simulation state sequence is generated, which fully records the time-varying trajectories of key physical variables from the start of the scheme execution to its expected end.

[0052] Finally, the simulation prediction agent integrates the above analysis and comparison results to generate a structured simulation prediction result. This simulation prediction result not only includes clear feasibility judgment conclusions, but also includes detailed quantitative data on expected effects, such as a predicted reduction of 40% in beam loss rate or a peak transient heat load margin of only 5%.

[0053] If the initial debugging plan is deemed infeasible or has safety risks, the system will automatically mark the specific time step in which the problem occurred and the corresponding non-compliant physical variables. This provides accurate feedback for the task planning agent to backtrack and correct the initial debugging plan, ensuring that only initial debugging plans that have been verified as safe and effective in the virtual environment can proceed to the next stage of physical execution.

[0054] In practical applications, to further ensure the absolute safety of the particle accelerator's physical equipment and beam operation, and to avoid potential safety hazards in actual operation due to simulation deviations or oversights in boundary conditions, the multi-agent system constructed in this application also introduces a safety monitoring agent to perform a rigorous safety verification process on the aforementioned debugging scheme data. Furthermore, such as... Figure 5 As shown, after step 103, the method further includes the following steps: 501. Use the security monitoring agent to perform a security boundary review on the simulation-predicted debugging scheme data.

[0055] 502. If the key physical variables in the debugging scheme data fall within the allowable range defined by the multidimensional constraint set, the safety boundary review is deemed to have passed, and the authorized execution method of the debugging scheme data is determined according to the operational risk level.

[0056] 503. If the key physical variable exceeds the static physical limit threshold, deviates from the dynamic operation constraint range, or matches the taboo state pattern, the security boundary review is deemed to have failed. A violation report containing the violation dimension and the corrected safe operation range is generated and fed back to the task planning agent to instruct the task planning agent to replan the preliminary debugging scheme with the corrected safe operation range as a hard constraint.

[0057] After the simulation prediction agent completes the virtual simulation of the initial debugging plan and outputs the debugging plan data predicted by the simulation, the safety boundary review procedure is immediately initiated. This safety monitoring agent maintains internal safety boundaries, defined as a multi-dimensional set of constraints, including at least: static physical limit thresholds, dynamic operational constraint ranges, and taboo state modes. In other words, the safety boundary is not a single numerical limitation, but a comprehensive protection network covering equipment physical limits, operational stability requirements, and historical failure modes. Static physical limit thresholds involve insurmountable hardware red lines such as the saturation current of the magnet coil, the withstand voltage limit of the vacuum chamber, and the rated power of the power supply; dynamic operational constraint ranges define the maximum rate of beam trajectory change, the slope limit of energy ramp, and other dynamic ranges to ensure stable system transition; while taboo state modes are specific parameter combinations summarized from historical accident cases, such as resonance risk at a specific frequency or beam loss modes under a specific magnetic field configuration.

[0058] Specifically, the safety monitoring agent first performs a time-series scan and feature extraction on the key physical variables contained in the simulation-predicted debugging scheme data, comparing them one by one with the multi-dimensional constraint set. This process uses pattern recognition algorithms to detect whether there are any signs in the scheme data falling into any of the aforementioned prohibited areas.

[0059] If the review results show that the key physical variables in the debugging plan data strictly fall within the allowable range defined by the multidimensional constraint set throughout the entire pre-run timeline, the safety monitoring agent determines that the safety boundary review has passed. At this point, the safety monitoring agent further assesses the operational risk level based on the margin between the variables and the safety boundary during the plan execution process. For low-risk plans with high margins, full execution permission is automatically granted, allowing the execution agent to directly issue instructions; for high-risk plans with low margins, a restricted authorization execution method is determined, such as requiring the activation of a high-frequency real-time monitoring mode during execution, or setting more stringent intermediate breakpoints for manual confirmation, thereby achieving graded and classified refined safety control.

[0060] Specifically, in determining the authorized execution method of the debugging scheme data based on the operational risk level, the safety margin between the key physical variables and the static physical limit threshold in the safety boundary is calculated. The safety margin represents the relative distance between the current variable value and the physical limit. According to the preset risk classification mapping table, the safety margin is divided into a low-risk range or a high-risk range. If the safety margin corresponding to the key physical variable falls into the low-risk range, the safety monitoring agent generates an execution authorization instruction and sends it to the execution agent. If the safety margin corresponding to the key physical variable falls into the high-risk range, the execution process is paused, a confirmation request containing the current safety margin value and potential risk warnings is generated and sent to the user terminal, and the execution authorization instruction is only generated and sent to the execution agent after receiving a manual confirmation instruction from the user terminal.

[0061] The aforementioned safety monitoring agent can determine the calculated safety margin range based on a preset risk classification mapping table. This risk classification mapping table is formulated based on long-term operational experience data of the accelerator and the equipment's tolerance characteristics, dividing the continuous safety margin value space into low-risk and high-risk ranges. When the safety margin value is large, it indicates that the key physical variables have sufficient buffer space during operation, and even minor environmental disturbances or equipment fluctuations will not reach the physical limits; this is classified as a low-risk range. Conversely, if the safety margin value is small, it means that the scheme is operating close to the limit, requiring extremely high control precision and environmental stability; any slight deviation may lead to violations, and this is classified as a high-risk range. For debugging schemes determined to be in the low-risk range, the safety monitoring agent implements an automated rapid passage strategy; for debugging schemes determined to be in the high-risk range, the safety monitoring agent initiates a human-machine collaborative prudent execution strategy.

[0062] If the review finds that key physical variables exceed static physical limit thresholds, deviate from dynamic operational constraint ranges, or match forbidden state patterns, the security monitoring agent immediately determines that the security boundary review has failed. A detailed violation report is then generated, clearly indicating the specific dimensions of the violation and calculating a corrected safe operating range based on the current state. This violation report is immediately fed back to the task planning agent, instructing it to treat the corrected safe operating range as an unshakeable hard constraint and to iterate and plan the initial debugging scheme again using the knowledge base and toolbox. This closed-loop review and feedback mechanism ensures that any instructions issued to physical devices undergo rigorous security verification.

[0063] In practical application scenarios, to quantify the debugging effect, accumulate effective operation and maintenance experience, and provide optimization support for subsequent debugging tasks, the multi-agent system of this application is also configured with a performance evaluation agent. This performance evaluation agent is mainly used to quantify the actual effect of this debugging. Furthermore, such as... Figure 6 As shown, after step 104, the method further includes the following steps: 601. Using the performance evaluation agent to collect real-time equipment operation data during the debugging process, calculate the improvement range of key performance indicators, use statistical analysis methods to perform significance testing on the improvement range, and generate a quantitative evaluation report including significance level and confidence interval.

[0064] 602. Based on the quantitative evaluation report, the setting tuple data, the state embedding vector of the initial state before execution, and the operation embedding vector of the debugging scheme data are jointly stored in the vectorized knowledge base.

[0065] 603. In response to subsequent debugging tasks, the historical debugging scheme data with similarity higher than a preset threshold in the vectorized knowledge base are weighted and fused according to the current state embedding vector to generate a weighted reference strategy.

[0066] 604. The quantitative evaluation report, as the latest evaluation result, is sent to the task planning agent along with the weighted reference strategy to instruct the task planning agent to optimize or modify the parameters of the candidate operation scheme by combining historical experience with the latest evaluation result.

[0067] After the execution agent converts the simulation-predicted debugging scheme data into control commands and sends them to the accelerator control system, completing one round of physical debugging and obtaining actual debugging results, the performance evaluation agent captures real-time equipment operation data of each key node of the accelerator during the debugging process. Based on this real-time data, the performance evaluation agent calculates the improvement of key performance indicators such as beam intensity and stability relative to before debugging, and further uses statistical analysis methods to test the significance of this improvement, generating a detailed quantitative evaluation report. This report not only includes specific numerical improvements but also clearly indicates the statistical significance level and confidence interval, thereby objectively determining whether the debugging was truly effective.

[0068] After obtaining the quantitative evaluation report, the system enters the knowledge accumulation phase. Based on the quantitative evaluation report, a set of tuple data is constructed. This tuple consists of the quantitative evaluation report itself, the initial state before execution, and the debugging scheme data implemented in this instance. The initial state before execution specifically refers to the equipment snapshot captured and locked from the accelerator control system before the control command is issued. Using a pre-trained feature extraction model, the initial state before execution is encoded to generate state embedding vectors, and the debugging scheme data is encoded to generate operation embedding vectors. The set of tuple data, state embedding vectors, and operation embedding vectors are then jointly stored in a vectorized knowledge base, and an association index between states and operations is established. It is worth noting that the set of tuple data is only marked as a successful historical experience when the quantitative evaluation report shows a significant improvement, thus ensuring that the knowledge base stores only valid and reliable debugging cases.

[0069] When the system responds to subsequent debugging tasks, it enters the knowledge reuse and strategy generation stage. First, the device operating parameters for the current debugging task are collected as the current initial state, and the same pre-trained feature extraction model is used to generate the current state embedding vector. Then, in the vectorized knowledge base, the similarity between the current state embedding vector and the historical state embedding vectors corresponding to all past successful historical experiences is calculated. Further filtering is performed to identify historical similar operating conditions with similarity exceeding a preset threshold, and the historical operation embedding vectors associated with these operating conditions and the corresponding historical debugging scheme data are extracted. Weights are assigned based on the similarity between historical and current operating conditions, and these historical debugging schemes are weighted and fused to generate a weighted reference strategy. This weighted reference strategy accurately represents the historically validated effective parameter distribution or operation sequence under similar operating conditions, providing a data-driven reference benchmark for the current task.

[0070] Finally, the quantitative evaluation report generated during this debugging process is sent to the task planning agent along with the weighted reference strategy generated above. Upon receiving this information, the task planning agent combines historical experience with the latest evaluation results to optimize or modify the parameters of candidate operation schemes. Optimization includes prioritizing schemes consistent with high-frequency operation characteristics in the weighted reference strategy; that is, the system will increase the selection priority of schemes consistent with high-frequency operation characteristics in the weighted reference strategy, thus favoring historically successful operation modes. Parameter modification includes using the effective operation parameter range in the weighted reference strategy as a strong constraint; that is, the system uses the effective operation parameter range extracted from the weighted reference strategy as a strong constraint condition to limit the parameter values ​​of candidate schemes. Through this closed-loop feedback mechanism, a complete process from data acquisition, performance evaluation, knowledge storage to strategy reuse is realized, significantly improving the intelligence level of accelerator debugging.

[0071] In the closed-loop debugging process of a multi-agent system, after the performance evaluation agent generates a quantitative evaluation report, it immediately enters the knowledge solidification stage. Specifically, based on the quantitative evaluation report, the initial state before execution, and the debugging scheme data, a setpoint tuple data is constructed; using a pre-trained feature extraction model, the initial state before execution is encoded to generate a state embedding vector, and the debugging scheme data is encoded to generate an operation embedding vector; the setpoint tuple data, state embedding vectors, and operation embedding vectors are jointly stored in a vectorized knowledge base, and an association index between state embedding vectors and operation embedding vectors is established in the vectorized knowledge base. Among them, the initial state before execution is a snapshot of the device locked before the control command is issued, the debugging scheme data contains specific parameter adjustments and operation timing, and the quantitative evaluation report serves as the evaluation label for the tuple. The three together constitute a complete debugging case unit.

[0072] Specifically, a pre-trained feature extraction model can be invoked to vectorize the structured data. This model is specifically trained for accelerator equipment status and debugging operations, enabling it to map high-dimensional equipment parameters and operational commands to a low-dimensional vector space. Specifically, the model encodes the initial state before execution, generating a state embedding vector that characterizes the equipment's operating condition; simultaneously, it encodes the debugging scheme data, generating an operation embedding vector that characterizes the operational features. These two vectors perform feature compression on the initial debugging case from both the state and operation dimensions.

[0073] Understandably, during the storage of state embedding vectors and operation embedding vectors, the knowledge base establishes an association index between the state embedding vectors and operation embedding vectors, enabling the distance in the vector space to reflect the similarity of physical conditions. This storage method not only preserves the original debugging data and evaluation results but also achieves efficient state and operation matching through vector indexing.

[0074] Furthermore, when a new debugging task is triggered, in response to subsequent debugging tasks, the current device operating parameters of the subsequent debugging task are collected as the current initial state, and a pre-trained feature extraction model is used to generate a current state embedding vector. In the vectorized knowledge base, the similarity between the current state embedding vector and the historical state embedding vectors corresponding to each successful historical experience is calculated, and historical similar working conditions with similarity higher than a preset threshold are selected. Historical operation embedding vectors associated with historical similar working conditions and corresponding historical debugging scheme data are extracted, weights are assigned according to the similarity, and a weighted reference strategy is generated by fusion.

[0075] Specifically, during the similarity retrieval operation in the vectorized knowledge base, the cosine similarity or Euclidean distance between the current state embedding vector and each historical state embedding vector in the knowledge base that is marked as a successful historical experience is calculated. By setting a preset similarity threshold, historical similarities that highly match the current situation are filtered out.

[0076] To generate the optimal reference strategy, weights are assigned based on the similarity between historical and current operating conditions, with higher similarity historical cases receiving greater weight. Finally, these weighted historical debugging schemes are fused and calculated to generate the final weighted reference strategy. This strategy integrates historical successes, characterizing validated parameter distributions or operational sequences under similar operating conditions, providing accurate decision-making support for the task planning agent.

[0077] In practical applications, this embodiment describes a multi-agent collaborative particle accelerator debugging workflow. This process is executed collaboratively by a multi-agent system based on a large language model. For details, please refer to [link to documentation]. Figure 7 As shown, the operation of the multi-agent system begins with the real-time monitoring by the state-aware agent. The state-aware agent continuously collects low-level operational data from the accelerator control system and generates state vectors reflecting the system's health status in real time through its integrated signal processing module. When the state-aware agent detects an abnormal trend or key parameters exceeding preset thresholds, it converts the numerical anomaly data into a fault report described in natural language. This report includes a specific timestamp and relevant device IDs and is sent to the task planning agent as the initial input to trigger downstream actions.

[0078] Upon receiving a fault report, the task planning agent immediately initiates the knowledge retrieval and task planning phase. First, it uses vector retrieval technology to search the knowledge base for historical cases with similar characteristics to the current fault. Based on the retrieved successful experiences and the current fault characteristics, the task planning agent utilizes the reasoning capabilities of a large language model to generate a preliminary debugging plan containing several steps using a pre-built algorithm toolbox. This plan is currently only a logical suggestion and needs to be validated before execution.

[0079] The initial debugging plan is then transmitted to the simulation prediction agent, entering the physical simulation and effect prediction phase. The simulation prediction agent constructs a digital twin environment based on a real-time updated beam dynamics model, performing a virtual simulation of the plan. The simulation prediction agent quantitatively calculates the expected effects after the plan's execution, for example, predicting a reduction in beam loss to 0.5% or a convergence of the orbital RMS error to 0.1 mm. If the simulation prediction results show that the plan is ineffective or may lead to a deterioration of performance indicators, feedback will be sent to the mission planning agent requesting a replanning. If the simulation prediction results meet expectations, the plan and related prediction data will be packaged and submitted to the safety monitoring agent.

[0080] The safety monitoring agent reviews each received plan, comparing it against the equipment's physical limits and a safety operation rule base. If any step exceeds the safety boundaries, the process will immediately terminate, and the specific reason for the violation will be reported to the planning agent, achieving automatic blocking. If the plan passes the safety review, it will be executed in a tiered manner based on the risk level of the operation.

[0081] For low-risk operations, such as minor track corrections, the system directly authorizes automatic execution. For high-risk operations, such as significant adjustments to the working point, the system will pause and send a confirmation request to a human operator; execution will only proceed after human authorization is obtained.

[0082] The authorized solution enters the execution agent. The execution agent translates the abstract task steps into low-level instructions that the control system can recognize, and distributes them to the hardware devices for debugging according to the specified time sequence. During the debugging process, the execution agent maintains bidirectional communication with the state-aware agent. Once it detects that the device response is inconsistent with expectations, the execution agent will immediately initiate a one-click rollback function to restore the machine to the safe state before execution, ensuring the robustness of the operation and maintenance process.

[0083] After the debugging task is completed, the system state before and after execution is compared with the knowledge base agent in the performance evaluation. Statistical methods are used to quantify the net benefit of this operation. The complete triple data of fault phenomenon, solution strategy and final effect is standardized and stored in the vector database. As the running time increases and the knowledge base continues to expand, the decision-making ability and accuracy of the multi-agent system in dealing with complex faults will continue to improve, ultimately realizing the evolution from assisted operation and maintenance to autonomous operation and maintenance.

[0084] Furthermore, as a specific implementation of the above method, embodiments of this application provide a multi-agent collaborative particle accelerator debugging device, such as... Figure 8 As shown, the device deploys a multi-agent system based on a large language model, which includes at least a state-aware agent, a task-planning agent, a simulation-prediction agent, and an execution agent; the device includes: Monitoring unit 71 is used to collect multi-source data of the accelerator control system in real time using the state-aware intelligent agent. When an abnormality is detected in the state of the particle accelerator based on the multi-source data, the abnormal information is converted into a natural language description and sent to the task planning intelligent agent. The generation unit 72 is used to analyze the abnormal information using a task planning agent, retrieve historical similar cases in the knowledge base, and generate a preliminary debugging plan containing multiple steps by combining a pre-set toolbox. The simulation prediction unit 73 is used to perform simulation prediction on the preliminary debugging scheme based on the physical model using the simulation prediction agent; if the simulation prediction result meets the set conditions, the debugging scheme data predicted by the simulation is generated; if the simulation prediction result does not meet the set conditions, the simulation prediction result is fed back to the task planning agent to re-plan the preliminary debugging scheme. The debugging unit 74 is used to convert the simulation-predicted debugging scheme data into control commands and send them to the accelerator control system through the execution intelligent agent, so as to perform debugging through the accelerator control system and obtain the debugging results.

[0085] The multi-agent collaborative particle accelerator debugging device provided in this invention deploys a multi-agent collaborative system based on a large language model. This system modularizes and coordinates state perception, task planning, simulation prediction, and execution control in a closed-loop manner. On one hand, it can collect and intelligently identify particle accelerator operational anomalies in real time, automatically converting anomaly information into standardized natural language descriptions. This avoids delays and misjudgments caused by manual intervention, significantly improving the efficiency and accuracy of fault perception and problem transmission. On the other hand, it leverages a knowledge base to retrieve similar historical cases and combines them with a pre-built algorithm toolbox to generate debugging solutions. This fully utilizes historical operation and maintenance experience, significantly shortening the debugging solution generation cycle. Simultaneously, the simulation prediction agent performs pre-verification of the debugging solution based on the physical model, effectively preventing ineffective or degrading operations from directly affecting the physical equipment. Through the organic collaboration of multiple agents, multi-parameter linkage optimization and adjustment of the particle accelerator are achieved, shortening the debugging cycle while improving the overall efficiency of particle accelerator operation and maintenance.

[0086] In specific application scenarios, the monitoring unit is specifically used for: Real-time acquisition of multi-source data from the accelerator control system, and trend prediction of the multi-source data using a long short-term memory network model; When the multi-source data is detected to deviate from the predicted trend by more than a preset residual threshold or exceeds the physical safety threshold, it is determined that there is an abnormality in the particle accelerator state, and the abnormal information is extracted. The abnormal information is instantiated and filled into the corresponding placeholder in the preset natural language prompt word template to generate a standardized natural language description, and the standardized natural language description is sent to the task planning agent.

[0087] In specific application scenarios, the generation unit is specifically used for: Receive and parse the abnormal information in the form of natural language, and extract the faulty equipment identifier and the abnormal indicator features of the offset prediction trend or the situation of exceeding the physical safety threshold; Based on the static configuration model of the accelerator control system, the system topology association information corresponding to the abnormal device identifier is obtained, and the system topology association information is combined with the abnormal feature index to construct a comprehensive retrieval feature. Based on the comprehensive retrieval features, historical similar cases with similar abnormal indicator features and matching system topology relationships are retrieved from the knowledge base to obtain the handling strategy information corresponding to the historical similar cases. Based on the handling strategy information and the abnormal indicator characteristics, suitable debugging operation units are selected from the preset toolbox, and the selected debugging operation units are sorted and combined according to the execution dependency to generate a preliminary debugging plan containing multiple steps.

[0088] In specific application scenarios, the simulation prediction unit is specifically used for: The current real-time operating status data of the particle accelerator is obtained as the initial conditions for simulation, and each debugging operation unit included in the preliminary debugging scheme is mapped to the key physical variables and their time-series change curves in the accelerator physical model. The accelerator physical model is driven to perform multi-step simulation calculations in a virtual environment according to the time-series change curve, so as to simulate the dynamic response behavior of the particle accelerator during the execution of the preliminary debugging scheme and generate a pre-simulation state sequence containing the trajectory of the key physical variables over time. Based on the aforementioned pre-run state sequence, the final steady-state characteristics of the particle accelerator after executing the initial debugging scheme and the instantaneous extreme value characteristics during the execution process are analyzed. Determine whether the final steady-state characteristics meet the preset target operating indicators, and detect whether the instantaneous extreme value characteristics violate physical safety constraints, and generate simulation prediction results that include feasibility judgment conclusions and quantitative data of expected effects.

[0089] In specific application scenarios, the multi-agent system also includes a security monitoring agent, and the device further includes: The review unit is used to review the safety boundary of the simulation-predicted debugging scheme data using the safety monitoring agent after the generation of the simulation-predicted debugging scheme data; the safety boundary is defined as a multi-dimensional constraint set, including at least: static physical limit threshold, dynamic operation constraint interval, and taboo state mode; The first determination unit is used to determine that the safety boundary review is passed if the key physical variables in the debugging scheme data fall within the allowable range defined by the multidimensional constraint set, and to determine the authorized execution method of the debugging scheme data according to the operational risk level. The second determination unit is further configured to determine that the security boundary review has failed if the key physical variable exceeds the static physical limit threshold, deviates from the dynamic operation constraint range, or matches the taboo state pattern, generate a violation report containing the violation dimension and the corrected safe operation range, and feed it back to the task planning agent to instruct the task planning agent to replan the preliminary debugging scheme with the corrected safe operation range as a hard constraint condition.

[0090] In specific application scenarios, the first determination unit is specifically used for: Calculate the safety margin between the key physical variable and the static physical limit threshold in the safety boundary, wherein the safety margin characterizes the relative distance between the current variable value and the physical limit; Based on a preset risk classification mapping table, the safety margin is divided into a low-risk range or a high-risk range; If the safety margin corresponding to the key physical variable falls into the low-risk range, the safety monitoring agent generates an execution authorization instruction and sends it to the execution agent. If the safety margin corresponding to the key physical variable falls into the high-risk range, the execution process is suspended, a confirmation request containing the current safety margin value and potential risk warnings is generated and sent to the user terminal, and an execution authorization instruction is generated and sent to the execution agent only after receiving a manual confirmation instruction from the user terminal.

[0091] In specific application scenarios, the multi-agent system further includes a performance evaluation agent, and the device further includes: The computing unit is used to convert the simulation-predicted debugging scheme data into control commands and send them to the accelerator control system for debugging through the accelerator control system. After obtaining the debugging results, the performance evaluation agent collects real-time equipment operation data during the debugging process, calculates the improvement range of key performance indicators, uses statistical analysis methods to perform significance testing on the improvement range, and generates a quantitative evaluation report including significance level and confidence interval. The storage unit is used to store the setting tuple data, the state embedding vector of the initial state before execution, and the operation embedding vector of the debugging scheme data into the vectorized knowledge base based on the quantitative evaluation report. The setting tuple data is marked as a successful historical experience only when the quantitative evaluation report shows a significant improvement. The initial state before execution is the equipment state collected and locked from the accelerator control system before the control command is issued. The setting tuple data consists of the quantitative evaluation report, the initial state before execution, and the debugging scheme data. The weighted fusion unit is used to respond to subsequent debugging tasks by weighting and fusing historical debugging scheme data with similarity higher than a preset threshold in the vectorized knowledge base according to the current state embedding vector, and generating a weighted reference strategy. The weighted reference strategy represents the distribution of effective parameters or operation sequence verified in history under similar working conditions. The sending unit is used to send the quantitative evaluation report as the latest evaluation result, together with the weighted reference strategy, to the task planning agent, so as to instruct the task planning agent to combine historical experience and the latest evaluation result to optimize or modify the parameters of the candidate operation schemes; the optimization includes increasing the selection priority of the schemes that are consistent with the high-frequency operation characteristics in the weighted reference strategy, and the parameter modification includes using the effective operation parameter range in the weighted reference strategy as a strong constraint. The storage unit is specifically used for: Based on the quantitative evaluation report, the initial state before execution, and the debugging scheme data, construct the set tuple data; Using a pre-trained feature extraction model, the initial state before execution is encoded to generate a state embedding vector, and the debugging scheme data is encoded to generate an operation embedding vector. The set tuple data, the state embedding vector, and the operation embedding vector are jointly stored in a vectorized knowledge base, and an association index between the state embedding vector and the operation embedding vector is established in the vectorized knowledge base; The weighted fusion unit is specifically used for: In response to subsequent debugging tasks, the current device operating parameters of the subsequent debugging tasks are collected as the current initial state, and the current state embedding vector is generated using the pre-trained feature extraction model. In the vectorized knowledge base, the similarity between the current state embedding vector and the historical state embedding vectors corresponding to each successful historical experience is calculated, and historical similar working conditions with similarity higher than a preset threshold are filtered out. Extract the historical operation embedding vectors associated with the historical similar working conditions and the corresponding historical debugging scheme data, assign weights according to the similarity, and fuse them to generate a weighted reference strategy.

[0092] Based on the above-mentioned multi-agent collaborative particle accelerator debugging method, this application embodiment also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the above-mentioned multi-agent collaborative particle accelerator debugging method.

[0093] Based on this understanding, the technical solution of this application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, or portable hard drive), and includes several instructions to cause a computer device (such as a personal computer, server, or network device) to execute the methods described in the various implementation scenarios of this application.

[0094] Based on the above-described multi-agent collaborative particle accelerator debugging method, in a virtual device embodiment, to achieve the above objectives, this application embodiment also provides a physical device for multi-agent collaborative particle accelerator debugging, specifically a computer, smartphone, tablet computer, smartwatch, server, or network device, etc., the physical device including a storage medium and a processor; the storage medium is used to store computer programs; the processor is used to execute the computer programs to implement the above-described multi-agent collaborative particle accelerator debugging method.

[0095] Optionally, the physical device may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0096] In an exemplary embodiment, see Figure 9 The aforementioned physical device includes a communication bus, a processor, a memory, and a communication interface. It may also include an input / output interface and a display device. The various functional units can communicate with each other via the bus. The memory stores a computer program, and the processor executes the program stored in the memory to perform the multi-agent cooperative particle accelerator debugging method described in the above embodiments.

[0097] Those skilled in the art will understand that the physical device structure for multi-agent collaborative particle accelerator debugging provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0098] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the physical device for the multi-agent collaborative particle accelerator debugging, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0099] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented using software plus necessary general-purpose hardware platforms, or it can be implemented in hardware. By applying the technical solution of this application, compared with the existing methods, this application deploys a multi-agent collaborative system based on a large language model, modularly dividing and coordinating state perception, task planning, simulation prediction, and execution control, realizing multi-parameter linkage optimization and adjustment of the particle accelerator, shortening the debugging cycle, and improving the overall efficiency of particle accelerator operation and maintenance.

[0100] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or can be modified to be located in one or more apparatuses different from this embodiment. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0101] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A multi-agent collaborative particle accelerator debugging method, characterized in that, The method is executed collaboratively by a multi-agent system based on a large language model, wherein the multi-agent system includes at least a state-aware agent, a task-planning agent, a simulation-prediction agent, and an execution agent; the method includes: The state-aware intelligent agent collects multi-source data from the accelerator control system in real time. When an anomaly is detected in the particle accelerator state based on the multi-source data, the anomaly information is converted into a natural language description and sent to the task planning intelligent agent. The task planning agent is used to analyze the abnormal information, retrieve similar historical cases from the knowledge base, and generate a preliminary debugging plan with multiple steps by combining the pre-set toolbox. The simulation prediction agent is used to perform simulation prediction on the preliminary debugging scheme based on the physical model; if the simulation prediction result meets the set conditions, the debugging scheme data predicted by the simulation is generated; if the simulation prediction result does not meet the set conditions, the simulation prediction result is fed back to the task planning agent to re-plan the preliminary debugging scheme. The execution agent converts the simulation-predicted debugging scheme data into control commands and sends them to the accelerator control system for debugging, thereby obtaining the debugging results.

2. The multi-agent collaborative particle accelerator debugging method according to claim 1, characterized in that, The process of utilizing the state-aware intelligent agent to collect multi-source data from the accelerator control system in real time, and when an anomaly is detected in the particle accelerator state based on the multi-source data, converting the anomaly information into a natural language description and sending it to the task planning intelligent agent, includes: Real-time acquisition of multi-source data from the accelerator control system, and trend prediction of the multi-source data using a long short-term memory network model; When the multi-source data is detected to deviate from the predicted trend by more than a preset residual threshold or exceeds the physical safety threshold, it is determined that there is an abnormality in the particle accelerator state, and the abnormal information is extracted. The abnormal information is instantiated and filled into the corresponding placeholder in the preset natural language prompt word template to generate a standardized natural language description, and the standardized natural language description is sent to the task planning agent.

3. The multi-agent collaborative particle accelerator debugging method according to claim 1, characterized in that, The process involves using the task planning agent to analyze the anomaly information, retrieving similar historical cases from the knowledge base, and generating a preliminary debugging plan with multiple steps using a pre-built toolbox, including: Receive and parse the abnormal information in the form of natural language, and extract the faulty equipment identifier and the abnormal indicator features of the offset prediction trend or the situation of exceeding the physical safety threshold; Based on the static configuration model of the accelerator control system, the system topology association information corresponding to the abnormal device identifier is obtained, and the system topology association information is combined with the abnormal feature index to construct a comprehensive retrieval feature. Based on the comprehensive retrieval features, historical similar cases with similar abnormal indicator features and matching system topology relationships are retrieved from the knowledge base to obtain the handling strategy information corresponding to the historical similar cases. Based on the handling strategy information and the abnormal indicator characteristics, suitable debugging operation units are selected from the preset toolbox, and the selected debugging operation units are sorted and combined according to the execution dependency to generate a preliminary debugging plan containing multiple steps.

4. The multi-agent collaborative particle accelerator debugging method according to claim 1, characterized in that, The step of using the simulation prediction agent to perform simulation prediction of the preliminary debugging scheme based on the physical model includes: The current real-time operating status data of the particle accelerator is obtained as the initial conditions for simulation, and each debugging operation unit included in the preliminary debugging scheme is mapped to the key physical variables and their time-series change curves in the accelerator physical model. The accelerator physical model is driven to perform multi-step simulation calculations in a virtual environment according to the time-series change curve, so as to simulate the dynamic response behavior of the particle accelerator during the execution of the preliminary debugging scheme and generate a pre-simulation state sequence containing the trajectory of the key physical variables over time. Based on the aforementioned pre-run state sequence, the final steady-state characteristics of the particle accelerator after executing the initial debugging scheme and the instantaneous extreme value characteristics during the execution process are analyzed. Determine whether the final steady-state characteristics meet the preset target operating indicators, and detect whether the instantaneous extreme value characteristics violate physical safety constraints, and generate simulation prediction results that include feasibility judgment conclusions and quantitative data of expected effects.

5. The multi-agent cooperative particle accelerator debugging method according to any one of claims 1-4, characterized in that, The multi-agent system further includes a safety monitoring agent, and after generating the simulation-predicted debugging scheme data, the method further includes: The safety monitoring agent is used to conduct a safety boundary review on the simulation-predicted debugging scheme data; the safety boundary is defined as a multi-dimensional constraint set, including at least: static physical limit threshold, dynamic operation constraint interval, and taboo state mode; If the key physical variables in the debugging scheme data fall within the allowable range defined by the multidimensional constraint set, the safety boundary review is deemed to have passed, and the authorized execution method of the debugging scheme data is determined according to the operational risk level. If the key physical variable exceeds the static physical limit threshold, deviates from the dynamic operation constraint range, or matches the taboo state pattern, the security boundary review is deemed to have failed. A violation report containing the violation dimension and the corrected safe operation range is generated and fed back to the task planning agent to instruct the task planning agent to replan the preliminary debugging scheme with the corrected safe operation range as a hard constraint.

6. The multi-agent collaborative particle accelerator debugging method according to claim 5, characterized in that, The step of determining the authorized execution method for the debugging plan data based on the operational risk level includes: Calculate the safety margin between the key physical variable and the static physical limit threshold in the safety boundary, wherein the safety margin characterizes the relative distance between the current variable value and the physical limit; Based on a preset risk classification mapping table, the safety margin is divided into a low-risk range or a high-risk range; If the safety margin corresponding to the key physical variable falls into the low-risk range, the safety monitoring agent generates an execution authorization instruction and sends it to the execution agent. If the safety margin corresponding to the key physical variable falls into the high-risk range, the execution process is suspended, a confirmation request containing the current safety margin value and potential risk warnings is generated and sent to the user terminal, and an execution authorization instruction is generated and sent to the execution agent only after receiving a manual confirmation instruction from the user terminal.

7. The multi-agent cooperative particle accelerator debugging method according to any one of claims 1-4, characterized in that, The multi-agent system further includes a performance evaluation agent. After the execution agent converts the simulation-predicted debugging scheme data into control commands and sends them to the accelerator control system for debugging, and obtains the debugging results, the method further includes: The performance evaluation agent collects real-time equipment operation data during the debugging process, calculates the improvement range of key performance indicators, uses statistical analysis methods to perform significance testing on the improvement range, and generates a quantitative evaluation report including significance level and confidence interval. Based on the quantitative evaluation report, the setting tuple data, the state embedding vector of the initial state before execution, and the operation embedding vector of the debugging scheme data are jointly stored in the vectorized knowledge base. Only when the quantitative evaluation report shows a significant improvement, the setting tuple data is marked as a successful historical experience. The initial state before execution is the equipment state collected and locked from the accelerator control system before the control command is issued. The setting tuple data consists of the quantitative evaluation report, the initial state before execution, and the debugging scheme data. In response to subsequent debugging tasks, the historical debugging scheme data with similarity higher than a preset threshold in the vectorized knowledge base are weighted and fused according to the current state embedding vector to generate a weighted reference strategy. The weighted reference strategy represents the distribution of effective parameters or operation sequence verified in history under similar working conditions. The quantitative evaluation report, as the latest evaluation result, is sent to the task planning agent along with the weighted reference strategy to instruct the task planning agent to optimize or modify the parameters of candidate operation schemes by combining historical experience and the latest evaluation result. The optimization includes increasing the selection priority of schemes that are consistent with the high-frequency operation characteristics in the weighted reference strategy, and the parameter modification includes using the effective operation parameter range in the weighted reference strategy as a strong constraint. The process of storing the tuple data, the state embedding vector of the initial state before execution, and the operation embedding vector of the debugging scheme data together in the vectorized knowledge base based on the quantitative evaluation report includes: Based on the quantitative evaluation report, the initial state before execution, and the debugging scheme data, construct the set tuple data; Using a pre-trained feature extraction model, the initial state before execution is encoded to generate a state embedding vector, and the debugging scheme data is encoded to generate an operation embedding vector. The set tuple data, the state embedding vector, and the operation embedding vector are jointly stored in a vectorized knowledge base, and an association index between the state embedding vector and the operation embedding vector is established in the vectorized knowledge base; In response to subsequent debugging tasks, the historical debugging scheme data with similarity higher than a preset threshold in the vectorized knowledge base are weighted and fused according to the current state embedding vector to generate a weighted reference strategy, including: In response to subsequent debugging tasks, the current device operating parameters of the subsequent debugging tasks are collected as the current initial state, and the current state embedding vector is generated using the pre-trained feature extraction model. In the vectorized knowledge base, the similarity between the current state embedding vector and the historical state embedding vectors corresponding to each successful historical experience is calculated, and historical similar working conditions with similarity higher than a preset threshold are filtered out. Extract the historical operation embedding vectors associated with the historical similar working conditions and the corresponding historical debugging scheme data, assign weights according to the similarity, and fuse them to generate a weighted reference strategy.

8. A multi-agent collaborative particle accelerator debugging device, characterized in that, The device is equipped with a multi-agent system based on a large language model, the multi-agent system including at least a state-aware agent, a task-planning agent, a simulation-prediction agent, and an execution agent; the device includes: The monitoring unit is used to collect multi-source data of the accelerator control system in real time using the state-aware intelligent agent. When an abnormality is detected in the state of the particle accelerator based on the multi-source data, the abnormal information is converted into a natural language description and sent to the task planning intelligent agent. The generation unit is used to analyze the abnormal information using the task planning agent, retrieve historical similar cases in the knowledge base, and generate a preliminary debugging plan containing multiple steps by combining a pre-set toolbox. The simulation prediction unit is used to perform simulation prediction on the preliminary debugging scheme based on the physical model using the simulation prediction agent; if the simulation prediction result meets the set conditions, the debugging scheme data predicted by the simulation is generated; if the simulation prediction result does not meet the set conditions, the simulation prediction result is fed back to the task planning agent to re-plan the preliminary debugging scheme. The debugging unit is used to convert the simulation-predicted debugging scheme data into control commands and send them to the accelerator control system through the execution agent to perform debugging and obtain debugging results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images