A control system for accelerator beam measurement and diagnostics

Abstract

The application discloses an accelerator beam measurement and diagnosis control system, comprising: a full-dimension perception module arranged at the inlet and outlet of an accelerator vacuum chamber and a focusing section, used for synchronously collecting beam parameters and multi-physical field environment data; a digital twin core module in communication connection with the full-dimension perception module, used for constructing and dynamically updating a full-element digital twin model covering the beam, equipment and environment based on the collected data, realizing real-time mapping of the beam state between the physical entity and the virtual model and parameter difference calculation, wherein the application breaks the isolated situation of past parameter measurement through the deployment of a collaborative perception network composed of multiple types of sensors, and the system synchronously obtains beam physical parameters, position, intensity, emission and key environment physical field data, electromagnetism, temperature, vacuum and radiation for the first time. This full-dimension perception capability breaks the isolated situation of past parameter measurement.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for particle accelerators, and particularly relates to a control system for accelerator beam measurement and diagnosis. Background Technology

[0002] The accelerator beam measurement and diagnostic control system is the core nerve center ensuring the stable, efficient, and safe operation of a particle accelerator. In existing technologies, such systems typically employ a hierarchical distributed architecture. The sensing layer consists of discretely deployed beam diagnostic equipment (such as beam position detectors, current transformers, and fluorescent screens), primarily used to monitor basic parameters such as beam position, intensity, and profile. The control core largely relies on feedback loops based on classical control theory (such as PID control), generating adjustment commands for actuators such as calibration magnets and RF power sources by comparing setpoints and measured values ​​of beam parameters. Furthermore, independent subsystems (such as vacuum monitoring, water cooling control, and radiation safety) operate through programmable logic controllers (PLCs), with limited hardwired interlocks between them and the beam control system. The data acquisition system records operating parameters, but analysis and optimization typically rely on offline processing by engineers. Overall, the existing technology achieves basic automation of accelerator operation, enabling the device to maintain routine operation under set conditions.

[0003] However, with the increasing demands on accelerator performance and the growing complexity of the operating environment, the existing technology system is gradually revealing its inherent shortcomings. First, its sensing dimension is limited, lacking synchronous monitoring and quantitative correlation of multi-physical field interference such as electromagnetic and temperature fields. This leads to decreased reliability of beam measurements under high-dynamic interference environments, and system resources cannot be intelligently allocated based on interference intensity. Second, the control strategy is rigid, relying on preset parameters and unable to adapt to changes in object characteristics caused by equipment aging and operating mode switching. Furthermore, it lacks the ability to autonomously learn and optimize strategies from historical data. Faced with complex multi-parameter coupled disturbances or gradual faults, the system can only respond passively, often resulting in oscillations or beam interruptions due to misadjustments. Fault recovery relies on manual troubleshooting, which is time-consuming. In addition, information is isolated between system modules, lacking an intelligent platform capable of integrating multi-source data and achieving global state simulation and collaborative optimization. This leads to severe "data silos," hindering continuous evolution of overall performance. Therefore, developing a new generation of beam control system capable of panoramic perception, intelligent decision-making, autonomous adaptation, and continuous optimization has become a critical technical problem urgently needing to be solved in this field. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a control system for accelerator beam measurement and diagnosis, which solves the problems of single sensing, fixed strategy, and inability to intelligently coordinate and autonomously evolve in the existing accelerator control system.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A control system for accelerator beam measurement and diagnosis, comprising:

[0007] The all-dimensional sensing module is deployed at the entrance and exit of the accelerator vacuum chamber and the focusing section to simultaneously acquire beam parameters and multi-physics environment data;

[0008] The core module of the digital twin is communicatively connected to the full-dimensional perception module. It is used to construct and dynamically update a full-element digital twin model covering the beam, equipment and environment based on the collected data, so as to realize the real-time mapping of the beam state and the calculation of parameter differences between the physical entity and the virtual model.

[0009] The edge-cloud collaborative computing module is deployed at the edge of the accelerator and is communicatively connected to the full-dimensional perception module and the digital twin core module, respectively. The edge is used to process the collected data in real time and dynamically calibrate the beam measurement value based on the output of the digital twin core module. The edge is also communicatively connected to the cloud deployed on a remote server. The cloud is used to iteratively optimize the model and calibration logic based on historical data.

[0010] The reinforcement learning optimization module, integrated into the digital twin core module, is used to autonomously learn and output the optimal beam control strategy based on the state information output by the digital twin core module.

[0011] The execution module is communicatively connected to the edge terminal and is used to execute beam correction and fault repair commands generated by the reinforcement learning optimization module and issued by the edge terminal.

[0012] Preferably, the all-dimensional sensing module includes a beam parameter sensor submodule and a multi-physics field sensor submodule;

[0013] The beam parameters collected by the beam parameter sensor submodule include beam position, beam intensity and beam emittance. It consists of a two-dimensional position sensitive detector deployed at the inlet and outlet of the accelerator vacuum chamber, a Faraday tube deployed in the focusing section and an emittance measurement unit.

[0014] The environmental data collected by the multi-physics sensor submodule includes electromagnetic field strength, cavity temperature, radiation dose, and vacuum level. It consists of a Hall sensor deployed near the magnet coil, a thermocouple embedded in the cavity surface, an ionization chamber deployed in the radiation-sensitive area, and a capacitive vacuum gauge connected to the vacuum pipeline.

[0015] Preferably, the core module of the digital twin adopts a hybrid modeling method based on multiphysics coupling equations and data-driven iteration; wherein, the multiphysics coupling equations are based on the accelerator three-dimensional geometric model embedding the coupling relationship between electromagnetic field, temperature field and vacuum field, and the data-driven iteration dynamically corrects the model parameters through the measured data received in real time from the full-dimensional sensing module;

[0016] The digital twin core module performs noise reduction on the measured data and then integrates it with the model's basic parameters to drive synchronous model updates. The noise reduction process uses a filtering algorithm to remove electromagnetic interference noise.

[0017] Preferably, the core functions of the edge terminal include:

[0018] Receive the measured data from the full-dimensional perception module and drive the digital twin core module to perform real-time updates;

[0019] Based on the built-in multi-physics coupling interference logic, the beam measurement value is dynamically calibrated. The logic is implemented by establishing a quantitative mapping model between the beam measurement value and the environmental interference. The output of the mapping model is an interference comprehensive coefficient.

[0020] The sampling frequency of the all-dimensional sensing module is adaptively adjusted according to the interference comprehensive coefficient. The adjustment logic is as follows: a threshold for the interference comprehensive coefficient is preset; when the interference comprehensive coefficient calculated in real time exceeds the threshold, the sampling frequency is increased; when the interference comprehensive coefficient is lower than the threshold, the sampling frequency is decreased.

[0021] Preferably, the core functions of the cloud include:

[0022] Receive the desensitized historical data uploaded by the edge terminal, the historical data including beam parameters, environmental data and equipment control records;

[0023] Based on the historical data, the model parameters of the digital twin core module and the multi-physics coupling interference logic at the edge are iteratively optimized.

[0024] Preferably, the reinforcement learning optimization module adopts a deep deterministic policy gradient algorithm. Its state space includes beam parameters consisting of beam position, beam intensity, and beam emittance; environmental interference consisting of electromagnetic field intensity, cavity temperature, radiation dose, and vacuum level; and equipment operating status consisting of magnet current, radio frequency power, and vacuum valve opening. Its action space includes magnet current adjustment, radio frequency power optimization value, and vacuum valve opening adjustment.

[0025] Its reward function includes a positive reward term to encourage beam stability and a negative penalty term to suppress beam loss and system energy consumption;

[0026] The training process includes: the first stage of pre-training in the virtual environment constructed by the core module of the digital twin; and the second stage of online iterative optimization of the strategy by combining real-time data from the physical system at the edge.

[0027] Preferably, the execution module includes a beam correction magnet for beam trajectory correction, an radio frequency power source for beam energy regulation, and a vacuum valve for vacuum degree control.

[0028] Preferably, it also includes a fault pre-diagnosis submodule, which is integrated into the digital twin core module. By comparing the real-time parameter differences between the virtual beam and the physical beam, and combining the beam evolution model optimized in the cloud, it can determine two types of potential fault risks: orbital deviation and beam loss, and trigger an early warning.

[0029] After receiving the early warning information, the reinforcement learning optimization module outputs targeted self-repair and adjustment instructions, which are then executed by the execution module driven by the edge terminal.

[0030] Preferably, it also includes a human-machine collaboration interface, deployed in the cloud, for engineers to view the system's operating status and the evolution process of the digital twin model, and has the function of issuing policy correction instructions to the reinforcement learning optimization module.

[0031] Preferably, the data transmission between the full-dimensional sensing module and the edge terminal adopts a shielded transmission link, and the shielded transmission link is provided with an anti-electromagnetic interference shielding layer.

[0032] The technical effects and advantages of the control system for accelerator beam measurement and diagnosis of the present invention are as follows:

[0033] 1. This invention, through the deployment of a collaborative sensing network composed of multiple types of sensors, enables the system to simultaneously acquire beam physical parameters (position, intensity, emittance) and key environmental physical field data (electromagnetism, temperature, vacuum, radiation) for the first time. This "all-dimensional sensing" capability breaks the previous isolation in parameter measurement. Simultaneously, a dedicated anti-interference shielded transmission link fundamentally ensures data transmission fidelity in strong electromagnetic environments, providing an accurate and reliable raw data foundation for all subsequent advanced functions and solving the control inaccuracy problem caused by data distortion or missing data in traditional systems.

[0034] 2. This invention provides a virtual reference for identifying and quantifying the impact of environmental interference on measured values ​​by constructing and updating a digital twin model of multi-physics coupling in real time. Based on this, the interference logic model built into the edge can calculate dynamic calibration coefficients online and compensate for beam measurements in real time, significantly improving the absolute accuracy of measurements under complex operating conditions. Furthermore, the system can adaptively adjust the sampling frequency according to the real-time interference level (interference synthesis coefficient), optimizing the utilization efficiency of system computing and storage resources while ensuring the capture of critical data.

[0035] 3. This invention, through massive training in a virtual environment, has mastered the optimal control strategy under multi-variable and strongly coupled operating conditions. It not only achieves coordinated and optimized control of multiple actuators under normal operation, but also, based on the early warning from the fault pre-diagnosis submodule, autonomously generates and executes precise self-repair commands for potential faults such as track deviation and beam loss. This represents a fundamental shift from "passively responding to faults" to "actively predicting and repairing," greatly improving the continuity and stability of operation.

[0036] 4. This invention's edge-cloud collaborative architecture, along with cloud-based optimization and human-machine interface functions, enables the system to continuously learn. The edge ensures real-time control, while the cloud utilizes historical data for in-depth analysis and model iteration, feeding the optimization results back to the edge. This mechanism allows the system's digital twin model, calibration logic, and control strategy to continuously improve themselves, overcoming the shortcomings of traditional control systems where algorithms and parameters are fixed and unable to adapt to equipment aging and changes in operating modes.

[0037] 5. This invention, through a cloud-deployed human-machine collaborative interface, allows engineers to intuitively grasp the overall system status, understand the evolution of the digital twin and the decision-making logic of the intelligent agent, and make strategy corrections and inject experience. This not only ensures the ultimate supervisory and guiding rights of human experts over complex systems, but also transforms expert experience into AI-learnable knowledge through friendly interaction, achieving an efficient integration of artificial intelligence and human wisdom, and improving the manageability and trustworthiness of the entire system. Attached Figure Description

[0038] Figure 1 This is a flowchart of a control system for accelerator beam measurement and diagnosis proposed in this invention. Detailed Implementation

[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0040] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "includes..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0041] refer to Figure 1 This invention provides a control system for accelerator beam measurement and diagnosis. The system includes a full-dimensional sensing module, a digital twin core module, an edge-cloud collaborative computing module, a reinforcement learning optimization module, and an execution module. The full-dimensional sensing module synchronously collects beam parameters and multi-physics environment data; the digital twin core module constructs and dynamically updates a full-element digital twin model based on this data, realizing real-time mapping and difference calculation between the physical and virtual beams; the edge-cloud collaborative computing module performs real-time data processing, dynamic calibration of measurement values, and iterative optimization of the model; the reinforcement learning optimization module is embedded in the digital twin and is used to autonomously learn the optimal control strategy; the execution module ultimately executes correction and repair commands. All modules are interconnected, forming an intelligent closed loop of "sensing-modeling-decision-execution-optimization," achieving accurate measurement of beam state, intelligent control, fault prediction and self-repair, and continuous autonomous evolution of system performance.

[0042] Example 1

[0043] This embodiment provides a control system for accelerator beam measurement and diagnosis, used in an intelligent control system infrastructure based on edge-cloud collaboration. Specific implementation details include:

[0044] Purpose of implementation:

[0045] This embodiment aims to construct a basic architecture for a closed-loop beam control system capable of achieving full-cycle sensing, real-time simulation, intelligent decision-making, and precise execution. The core objective of this architecture is to address the fundamental shortcomings of traditional accelerator control systems, such as information isolation among subsystems (measurement, power supply, vacuum, and temperature control), slow control loop response (typically on the order of hundreds of milliseconds), and the inability to utilize historical data for self-iterative optimization. This will provide intelligent platform support for the high stability, high efficiency, and high availability of the beam.

[0046] Implementation System:

[0047] The system's basic physical and logical architecture constructed in this embodiment is as follows: Figure 1 As shown. This architecture follows a closed-loop paradigm of "perception-modeling-decision-execution-optimization," and its core components and collaborative working mechanisms are as follows:

[0048] Full-dimensional perception layer:

[0049] This layer consists of a sensor network widely deployed in the accelerator vacuum chamber, magnets, cavity, and radiation-sensitive areas. Its primary responsibility is to synchronously and in parallel acquire two types of core data: first, the physical parameters of the beam itself, including microsecond-level variations in beam position (lateral X, Y), intensity (current), and emittance (beam quality); second, multiphysics environment data of the accelerator operation, including electromagnetic field strengths that may interfere with the beam, cavity temperatures that cause equipment deformation, radiation doses crucial for safety, and vacuum levels affecting beam lifetime. All sensors are connected to a unified hardware interface and a synchronous clock signal to ensure data time alignment, laying the foundation for subsequent coupling analysis.

[0050] Edge intelligent control node:

[0051] This node is a high-performance industrial computing server, physically deployed in the accelerator tunnel or a nearby control room to ensure minimal network latency. It carries three core functions:

[0052] Real-time data aggregation and preprocessing: The raw data stream from the sensing layer is received via a high-speed data bus (such as PXIe or 10GbE) and then formatted, timestamped, and initially removed outliers.

[0053] Lightweight Digital Twin Engine: Runs a simplified and optimized accelerator digital twin model. This model can simulate and predict the beam's expected state (predicted position, predicted intensity) under current conditions within milliseconds, based on real-time input environmental data (such as magnetic field and temperature).

[0054] High real-time control command calculation: The beam state measured by the sensors is compared with the state predicted by the twin model. Combined with the built-in fast calibration algorithm and classical feedback control law (such as PID feedforward for track stabilization), fine-tuning commands for actuators such as the magnet and RF power source are generated. The decision-execution delay in this stage is strictly controlled within 1 millisecond to cope with rapid beam disturbances.

[0055] Cloud-based intelligent analysis and optimization platform:

[0056] This platform is deployed on a remote high-performance computing cluster and connected to edge nodes via an enterprise-grade fiber optic network. It is not designed for real-time control, but rather focuses on "offline" deep analysis and global optimization.

[0057] Data Lake and Knowledge Base: Receives and stores anonymized and compressed historical operational data (including raw data, control commands, equipment logs, and event records) uploaded from edge nodes of one or more accelerator devices, forming data assets that can be deeply mined.

[0058] Model Training and Optimization Factory: Utilizing the massive amounts of data in the data lake, the digital twin models deployed on edge nodes are periodically retrained to correct errors that may accumulate during long-term operation; at the same time, machine learning algorithms are used to analyze the optimal combination of control parameters under different operating modes to form a "strategy package".

[0059] Health management and trend prediction: Perform trend analysis on equipment operating parameters (such as magnet coil resistance and vacuum pump current) to achieve predictive maintenance.

[0060] Intelligent Decision-Making and Collaborative Execution Layer:

[0061] Intelligent Decision-Making Agent: An intelligent agent based on deep reinforcement learning algorithms is embedded as a software module within the digital twin engine of the edge node. It "thinks" and "trials and errors" in a virtual accelerator highly synchronized with the real environment, learning how to combine and adjust multiple actuators under different perturbations to achieve globally optimal beam quality.

[0062] The coordinating actuators include a multi-pole correction magnet power supply system for beam trajectory correction, an RF power amplifier system for beam energy and phase adjustment, and an array of gate valves and slide gate valves for maintaining vacuum. They receive digital commands from edge nodes and convert them into high-precision analog outputs.

[0063] Implementation results:

[0064] By implementing the infrastructure described in this embodiment, the system achieves a fundamental capability enhancement. In terms of responsiveness, edge intelligent nodes improve the system's response time to sudden disturbances from the hundreds of milliseconds of traditional architectures to sub-milliseconds, effectively suppressing rapidly changing interference. Regarding collaboration, edge-cloud collaboration provides local control with a global perspective, while optimization models and strategies distributed from the cloud enable the edge controller's performance to continuously evolve over time, solving the problem of fixed parameters in traditional control systems. In terms of reliability, the "virtual-real comparison" capability provided by digital twins offers the system an unprecedented perspective on status monitoring and fault diagnosis. Overall, this architecture provides the essential infrastructure for accelerators to move from "automation" to "intelligence."

[0065] Example 2

[0066] This embodiment provides a control system for accelerator beam measurement and diagnosis, used for high-precision multi-physics sensing and strong anti-interference transmission schemes. Specific implementation details include:

[0067] Purpose of implementation:

[0068] This embodiment aims to build a data sensing and acquisition foundation for the upper-level intelligent control system, capable of resisting extreme electromagnetic interference within the accelerator hall and providing high-fidelity, highly synchronous raw data. It directly addresses the core pain point of sensor signal distortion and high transmission error rates under strong pulsed magnetic fields and radio frequency noise environments, which prevents measurement data from accurately reflecting the physical state.

[0069] Implementation System:

[0070] This embodiment constructs a distributed, heterogeneous sensor network and a dedicated transmission link.

[0071] Beam parameter sensing unit (core measurement):

[0072] Beam position monitoring: At critical locations on the beam transmission line, such as the inlet and outlet of each focusing cycle (cell), strip-electrode-based beam position detectors are installed. For example, a push-button type BPM with an impedance of 50Ω and a resolution better than 10μm is selected. Its analog output signal is immediately converted to digital signal locally after passing through a preamplifier.

[0073] Beam intensity monitoring: A Faraday tube made of non-magnetic material is installed at the end of the beamline or in a specific diagnostic box and connected to an integrating amplifier circuit (such as an electrometer-based operational amplifier) ​​with extremely high input impedance and low leakage current to achieve accurate measurement of the charge of a continuous or pulsed beam.

[0074] Beam emittance monitoring: A "moving target" trigradation measurement method is employed. This unit includes a precision platform driven by a stepper motor, on which three slit-fluorescent screen assemblies of different materials are mounted. By controlling the platform's movement, the beam sequentially passes through three slits at different locations. A downstream CCD camera captures beam spot images on the fluorescent screen, and a dedicated image processing algorithm is used to invert and calculate the beam's transverse normalized emittance.

[0075] Environmental field sensing unit (auxiliary but critical):

[0076] Electromagnetic field monitoring: Open-loop Hall current sensors are connected to the coil leads of all main magnets (dipole magnets and quadrupole magnets) to monitor the ripple and harmonics of the drive current. A triaxial fluxgate probe is installed near the air gap of the magnets to monitor changes in stray magnetic fields.

[0077] Temperature field monitoring: PT1000 platinum resistance temperature sensors are embedded in key thermal points such as the coupler window of the acceleration cavity, the inlet and outlet of the cooling water circuit of the magnet, and the flange of the vacuum chamber. The temperature measurement range covers -50°C to +200°C, with an accuracy of ±0.1°C.

[0078] Radiation field monitoring: Calibrated scintillator or ionization chamber type continuous radiation monitors are installed in areas where people are active, such as outside the beamline shielding door and at the entrance of the control room, with a range covering 1 nSv / h to 10 mSv / h.

[0079] Vacuum monitoring: In each independent section of the vacuum pipeline, a composite vacuum gauge (such as a combination of a Pirani gauge and a cold cathode ionization gauge) with both rough vacuum and high vacuum measurement capabilities is connected via CF35 flanges to achieve full range coverage from atmospheric pressure to 10^-7 Pa.

[0080] Strong anti-interference transmission and acquisition system:

[0081] One of the core innovations of this solution lies in its transmission link design:

[0082] Full-link shielding: All analog signal lines output by sensors use double-shielded coaxial cables (such as SMA interface phase-stable cables). The inner layer is a high-density copper mesh for shielding electric fields; the outer layer is aluminum foil for shielding low-frequency magnetic fields. The cable connectors at both ends achieve 360° full-circuit conductive connection with the equipment chassis.

[0083] Localized data acquisition: A robust, dedicated data acquisition enclosure is deployed near each sensor cluster. This enclosure is constructed entirely of sealed aluminum and features strict internal power partitioning and isolation between analog and digital grounds. Sensor signals undergo analog-to-digital conversion as close to the interface as possible after entering the enclosure, minimizing the path of susceptible analog signals.

[0084] Digital bus transmission: The digital signal, after being converted by the ADC, is uploaded to the edge intelligent node via a high-speed fieldbus (such as fiber-optic EtherCAT) with hardware CRC check and automatic retransmission mechanism. The fiber optic medium itself has natural electromagnetic immunity properties.

[0085] Implementation results:

[0086] After implementing this solution, the system data quality achieved a significant leap. In terms of interference immunity, during the strong electromagnetic pulses generated by the startup of nearby high-power RF systems (such as klystrons), the transient glitches in the key beam position signal were suppressed to below 0.5% of full scale, while in traditional solutions this value could exceed 10%. Regarding data synchronization, the timestamp synchronization accuracy of the entire system's sensor data is better than 100 nanoseconds, enabling precise causal correlation analysis between magnetic field fluctuations and beam position jitter collected from different physical locations. In terms of reliability, the bit error rate of the transmission link is below 10^-12, ensuring data integrity and credibility. This high-fidelity, high-reliability sensing layer becomes the "sensory system" that enables the stable operation of the entire intelligent control system.

[0087] Example 3

[0088] This embodiment provides a control system for accelerator beam measurement and diagnosis, and a data-driven dynamic calibration and adaptive resource scheduling method. Specific implementation details include:

[0089] Purpose of implementation:

[0090] This embodiment aims to address two key issues: "how to utilize environmental data to perform online, dynamic compensation of beam measurements to improve their realism and accuracy," and "how to intelligently allocate data acquisition and processing bandwidth when system resources are limited." Its goal is to enable the system to maintain optimal beam state observation capabilities in complex and changing environments.

[0091] Implementation System:

[0092] In this embodiment, a software algorithm module is deployed in the edge intelligent control node described in Embodiment 1.

[0093] Dynamic calibration coefficient generation module:

[0094] This module is the core of achieving accurate measurements. It is based on a data-driven "environmental interference - measurement error" mapping model.

[0095] Model Construction: During the system debugging phase, controllable environmental disturbances (such as adjusting cooling water temperature and fine-tuning the bias magnetic field) were intentionally introduced, while high-precision reference beam measurements (provided by independent, more precise but slower diagnostic equipment) and readings from various environmental sensors were recorded. After collecting sufficient data, a hybrid algorithm combining ridge regression and gradient boosting decision trees was used for training. The final model is a function: [K_x, K_y, K_i] = F(B_x, B_y, B_z,T1, T2, ..., D), where K is the calibration coefficient for position (X, Y) and intensity (I), the inputs are the three-dimensional magnetic field B, the temperatures T at multiple key points, and the radiation dose D, and F is the core component for achieving real-time high-precision beam calibration.

[0096] Online Application: During system operation, this module reads environmental sensor data in real time, substitutes it into function F, and calculates the current dynamic calibration coefficient K. Subsequently, the original beam measurement value M_raw is corrected to M_calibrated = K. M_raw.

[0097] Interference assessment and adaptive sampling control module:

[0098] Interference Composite Index Calculation: The real-time interference composite index is defined as I = sqrt( (K_x-1)² + (K_y-1)² + (K_i-1)² ). This index quantitatively reflects the overall interference level of the current environment on the beam measurement system.

[0099] Adaptive sampling strategy: Three preset thresholds are used: a high interference threshold I_high (e.g., 0.1), a low interference threshold I_low (e.g., 0.02), and an emergency threshold I_emergency (e.g., 0.5). The system has a built-in state machine.

[0100] Normal mode (I_low < I < I_high): All sensors sample at a reference frequency (e.g., 1 kHz).

[0101] High Interference Mode (I >= I_high): Trigger upgrade. The sampling frequency of the beam position detector (BPM) is increased to 10kHz, and the temperature and magnetic field sensors are increased to 500Hz to capture rapid changes; at the same time, the current data is marked as "high interference condition data" to the cloud platform.

[0102] Low-interference mode (I <= I_low): Triggers degradation. The BPM sampling frequency is reduced to 100 Hz, and non-critical environmental sensors are reduced to 1 Hz to save storage space and CPU cycles.

[0103] Emergency Mode (I >= I_emergency): Triggers a "snapshot" record. The system automatically saves 10 seconds of raw, uncompressed waveform data before and after the trigger for in-depth post-incident fault analysis.

[0104] Implementation results:

[0105] By implementing the method of this embodiment, the system achieves dual optimization in measurement accuracy and resource efficiency. Regarding measurement accuracy, in a strong magnetic field fluctuation event lasting several seconds due to a power line switch, the dynamic calibration module successfully compensated for the drift error of the beam position reading from approximately 2.5 mm before calibration to within 0.3 mm, achieving a deviation elimination rate of 88%. In terms of resource efficiency, the adaptive sampling mechanism allows the system to operate in a low-power "low-interference mode" for up to 95% of its stable operation time, reducing overall data storage requirements by approximately 60% and lowering the average CPU load of edge nodes by 35%, freeing up valuable resources for running more complex digital twin and artificial intelligence algorithms.

[0106] Example 4

[0107] This embodiment provides a control system for accelerator beam measurement and diagnosis, used for predictive maintenance and self-repair based on reinforcement learning and digital twins. Specific implementation details include:

[0108] Purpose of implementation:

[0109] This embodiment aims to empower the control system with the ability to predict potential faults and proactively repair them before they occur or worsen, thereby minimizing unplanned downtime and transforming the traditional passive "fault response" operation and maintenance model into a proactive "predictive intervention" protection model.

[0110] Implementation System:

[0111] This embodiment is a high-level application of the intelligent decision-making body in Embodiment 1, and includes two stages: offline training and online operation.

[0112] Offline training: Learning "medical skills" in a digital twin:

[0113] Fault Scenario Library Construction: First, a rich virtual fault scenario library is built on a cloud platform. Scenario sources include: 1) historical real-world fault event records; 2) fault modes derived from physical failure models (such as deformation of the magnet coil caused by Lorentz force, material fatigue); and 3) artificially designed extreme boundary conditions (such as slow deterioration of vacuum level, gradual decrease in cooling efficiency). The library ultimately contains over 100 different fault or performance degradation modes.

[0114] Reinforcement learning agent training: The agent is trained using a proximal policy optimization algorithm. The state space S includes: beam parameters, environmental parameters, and critical equipment states (such as power ripple and cooling water flow) from the twin model. The action space A consists of adjustment commands (continuous values) for each actuator. The reward function R is carefully designed as: R = R_stability - α R_loss-β R_energy - γ R_act, where R_stability rewards beam stability, R_loss penalizes beam loss, R_energy penalizes high energy consumption, and R_act penalizes excessively large and frequent actions to encourage smooth control. The agent undergoes millions of simulations in a fault scenario library, learning to identify early fault characteristics and take optimal intervention measures. Here, α is a non-negative weighted hyperparameter of the beam loss penalty term, used to quantitatively adjust the severity of the agent's penalty for beam loss; a larger value indicates a higher priority for beam loss control. β is a non-negative weighted hyperparameter of the energy consumption penalty term, used to quantitatively adjust the severity of the agent's penalty for high system energy consumption; a larger value indicates a more stringent requirement for energy-saving operation. γ is a non-negative weighted hyperparameter of the action penalty term (unrelated to the general discount factor in reinforcement learning), used to quantitatively adjust the severity of the agent's penalty for excessively large and frequent actuator adjustments; a larger value encourages the agent to adopt a smooth control strategy.

[0115] Online operation: Real-time diagnostics and self-repair:

[0116] Predictive diagnostic unit: When running online, this unit continuously calculates a "health index" H, which is a weighted sum of multiple sub-indicators (such as orbital closure deviation, emission increase trend, and vacuum deterioration rate). When H falls below a preset threshold, an early warning is triggered.

[0117] Self-Repair Decision-Making and Execution: Upon triggering an early warning, the reinforcement learning agent is activated. It receives the current state and, based on a trained policy network, outputs a "repairing" action sequence within milliseconds. For example, when an early trend of "a slight decrease in the magnetic field gradient of quadrupole magnet Q5 due to temperature rise" is diagnosed, the agent might output the following actions: slightly increase the current of the adjacent correction magnet of Q5 to compensate for the change in focusing force, and simultaneously fine-tune the opening of the magnet's cooling water valve to enhance heat dissipation, forming a closed-loop "diagnosis-compensation-root cause mitigation" strategy. This instruction sequence, after security verification, is executed by the edge node.

[0118] Implementation results:

[0119] The implementation of this embodiment elevates system reliability to a new level. During a three-month continuous operation experiment, the system successfully predicted two slow leaks in the vacuum seal and one degradation event in the magnet power supply filter. Warnings were issued and compensation measures were initiated before any noticeable decline in beam quality (with lead times of 36 hours and 120 hours, respectively), preventing unplanned experimental interruptions. Regarding self-healing capabilities, for random, low-intensity transient disturbances in the beam trajectory (such as those caused by external micro-vibrations), the system completes diagnosis and outputs correction commands within 5 milliseconds, attenuating the disturbance amplitude by more than 90%, making it virtually imperceptible to the experimental users. Statistics show that the application of this system has reduced the accelerator's annual unplanned downtime by approximately 65%, significantly improving equipment utilization.

[0120] Example 5

[0121] This embodiment provides a control system for accelerator beam measurement and diagnosis, used in a cloud-based federated learning optimization and human-machine fusion decision-making platform. Specific implementation details include:

[0122] Purpose of implementation:

[0123] This embodiment aims to address the issues of limited data from a single device and low ceiling for model optimization. By constructing a transparent and trustworthy human-computer interaction interface, it organically integrates the experience and knowledge of human experts into the artificial intelligence decision-making cycle, thereby achieving the fusion of "collective intelligence" and "human wisdom" and ensuring that complex systems evolve in the optimal direction under controlled and trustworthy conditions.

[0124] Implementation System:

[0125] 2.1 Cloud-based Federated Learning Optimization Engine

[0126] Architecture: A federated learning architecture of "central server - multiple edge clients" is adopted. Each edge node of the accelerator device acts as a client and has local data; the cloud platform acts as a coordination server.

[0127] Workflow: Every training cycle (e.g., weekly), the cloud server distributes the current global model (e.g., dynamically calibrated model F) to each client (edge ​​node). Each client trains the model locally using the newly generated data, calculates the updated gradients of the model parameters (not the original data), and uploads the encrypted gradients to the cloud. The cloud aggregates the gradients from all clients, updates the global model, and then distributes the new model. This process repeats continuously.

[0128] Effect: This allows the operational experience from multiple accelerators with different energy ranges and structures to be pooled together to train a more robust and generalizable general model, while strictly protecting the data privacy of each device.

[0129] Human-Machine Integration Decision-Making Platform (Digital Cockpit):

[0130] This platform is an advanced Web-GIS visualization system that provides engineers with a one-stop monitoring and intervention portal.

[0131] Panoramic visualization: Using a 3D digital twin model as the base map, real-time data streams are overlaid: the beam travels through the pipe like a light strip, with its color representing intensity and its width representing emissivity; the temperature distribution of devices such as magnets and cavities is displayed in the form of a heat map; key parameters are displayed as floating trend curves.

[0132] AI Decision-Making Perspective: Features an "AI Strategy Interpretation" panel. When a reinforcement learning agent makes a critical decision (such as significantly adjusting radio frequency power), the panel displays the weighted ranking of the state factors on which the decision was based, the expected short-term and long-term gains (displayed through "counterfactual reasoning" simulations), and the successful history of the strategy in similar past scenarios, greatly enhancing the interpretability of AI and the trust of engineers.

[0133] Expert knowledge injection channel: Engineers can use the "rule editor" to write domain knowledge rules in an "if-then" format. For example: "If the vacuum level is below 1e-5 Pa and continues to deteriorate, then prioritize checking the areas corresponding to vacuum gauges 5 and 7." These rules will be converted into an additional reward and integrated into the training environment of the reinforcement learning agent, thereby gently guiding the AI's learning direction and making its decisions consistent with physical common sense and operational experience.

[0134] Implementation results:

[0135] Through cloud-based federated learning, a model used to predict emission changes over time, after three months of joint training with three devices, showed an average 28% reduction in the root mean square error (RMSE) of its predictions for unknown operating conditions compared to models trained independently by each device. The human-machine fusion platform has fundamentally changed the way engineers work. When dealing with a complex multi-parameter coupled oscillation event, engineers used the decision-making perspective function to quickly understand the AI's intention to suppress the oscillation by adjusting an unconventional combination of magnets. Based on experience, they confirmed the safety of the strategy, approved its execution, and successfully quelled the oscillation. This process saved over 80% of troubleshooting time compared to traditional manual trial-and-error methods. The platform has established a new and efficient collaborative relationship of "human supervision, AI execution, and mutual learning."

[0136] Comparative Example 1

[0137] This comparative model provides a traditional discrete PID control and planned maintenance solution, specifically including:

[0138] Implementation plan:

[0139] This comparative example represents a traditional industrial control scheme that is still widely used today. Its core consists of multiple independent, discrete control loops:

[0140] Beam trajectory control: Feedback is provided by several beam position detectors, and each horizontal / vertical plane is controlled by an independent PID controller that calculates the correction magnet current. The controller parameters (P, I, D) are manually tuned by engineers during the machine commissioning phase and remain essentially fixed thereafter.

[0141] Each subsystem operates independently: the vacuum system, water cooling system, and magnet power supply system each have their own PLC controller and HMI interface. They only transmit "ready" and "fault" switching signals through simple hard wiring, without in-depth data sharing and collaborative optimization.

[0142] Data recording and analysis are disconnected: the data acquisition system is independent of the control system, mainly recording slow trends. Data analysis is conducted offline and retrospectively, and cannot be used for real-time decision-making.

[0143] Maintenance mode: Strict periodic preventive maintenance is adopted, with shutdowns at fixed intervals (such as quarterly) to routinely inspect and replace all equipment, regardless of the actual operating status of the equipment.

[0144] Analysis of implementation effects and limitations:

[0145] Traditional solutions are adequate under stable and ideal operating conditions. However, their inherent limitations become apparent during complex, long-term operation.

[0146] Weak anti-interference capability: The PID controller cannot distinguish between the actual beam offset and the reading drift caused by environmental interference (such as magnetic fields and temperature). When interference occurs, the controller may "malfunction," potentially causing oscillations that require manual intervention to resolve.

[0147] Fault handling is passive and slow: the system lacks predictive capabilities. An alarm and shutdown are only triggered when a fault reaches a shutdown threshold (such as complete beam loss or complete vacuum failure). Subsequent fault localization relies entirely on the engineer's experience, involving step-by-step troubleshooting, resulting in a mean time to recovery (MTTR) of several hours or even days.

[0148] Performance cannot be continuously optimized: Fixed PID parameters cannot adapt to changes in object characteristics caused by equipment aging, operation mode switching, etc., resulting in slow degradation of beam quality over time until the next manual readjustment.

[0149] High maintenance costs and low efficiency: Planned maintenance may lead to "over-maintenance" (replacing still healthy parts) or "under-maintenance" (failing to replace deteriorated parts in time), resulting in high overall maintenance costs and still being unable to avoid unplanned downtime.

[0150] Compared to Examples 1-5 and Comparative Example 1, the intelligent control system constructed in Examples (1-5) of this invention presents a stark contrast to the traditional solution (Comparative Example 1) in terms of core concepts, technical approaches, and final performance, representing a generational leap. In summary:

[0151] Traditional solutions employ discrete subsystems (beam control, vacuum, temperature control, etc.) mechanically assembled, resulting in isolated data and control systems and creating "information silos." This invention, however, constructs a unified and integrated intelligent platform based on edge-cloud collaboration. It achieves global fusion of sensory data, real-time digital twin simulation, and unified decision-making by intelligent agents, treating the accelerator as a fully perceptible, simulated, and optimized organic whole.

[0152] Traditional methods only perform limited measurements on the beam itself, with environmental data used solely for safety alarms. This invention establishes a multi-physics synchronous sensing network that not only measures the beam with high precision but also simultaneously captures all key environmental interference quantities (electromagnetic, temperature, vacuum, and radiation). Furthermore, for the first time, through shielded transmission and dynamic calibration techniques, it establishes a quantitative compensation relationship between environmental interference and measurement errors, transforming the measurement system from "fragile" to "robust."

[0153] Traditional data systems are only used for post-event backtracking. This invention achieves data-driven online adaptive optimization: at the edge, data is cleaned in real time through dynamic model calibration, and sampling resources are intelligently scheduled according to the degree of interference; in the cloud, federated learning is used to aggregate experience from multiple devices and continuously iterate and optimize the global model. Data transforms from a "record" into the "fuel" driving system evolution.

[0154] Traditional solutions rely on PID controllers with fixed parameters, which cannot handle complex coupled disturbances and whose strategies never evolve. This invention introduces an intelligent decision-making entity based on digital twins and reinforcement learning. It simulates millions of fault scenarios in a virtual environment, learns predictive intervention and self-healing strategies involving multi-actuator collaboration, and can absorb human experience through a human-machine interface, enabling the control system to have the ability to learn autonomously and continuously optimize.

[0155] Traditional operations and maintenance rely on rigid, periodic inspections and reactive responses after failures, resulting in long downtimes and high costs. This invention achieves predictive health management and autonomous recovery. Through virtual-to-real comparison and trend analysis, the system can issue early warnings tens of hours before a failure occurs, and the intelligent agent automatically executes compensation or repair instructions, significantly reducing unplanned downtime and transforming the operations and maintenance model from a "firefighting team" to a "health care provider."

[0156] Traditional solutions can operate under stable and ideal conditions, but they exhibit poor anti-interference capabilities, slow recovery, and continuous performance degradation when faced with complex interference, long-term operation, and equipment aging. This invention, through systematic innovation, achieves high-precision stable operation under high dynamic interference (accuracy improvement of over 85%), fault prediction and self-repair (reducing unplanned downtime by approximately 65%), and continuous optimization of energy efficiency throughout the entire lifecycle (reducing energy consumption by approximately 7%). This marks a significant step for accelerator control, moving from an "automation" stage reliant on human experience to a new "autonomy" stage driven by data and intelligence.

[0157] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0158] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A control system for accelerator beam measurement and diagnosis, characterized in that, include: The all-dimensional sensing module is deployed at the entrance and exit of the accelerator vacuum chamber and the focusing section to simultaneously acquire beam parameters and multi-physics environment data; The digital twin core module, which is communicatively connected to the all-dimensional sensing module, is used to construct and dynamically update a full-element digital twin model covering the beam, equipment, and environment based on the collected data. This enables real-time mapping of the beam state and calculation of parameter differences between the physical entity and the virtual model. The digital twin core module adopts a hybrid modeling method based on multi-physics coupling equations and data-driven iteration. The multi-physics coupling equations embed the coupling relationships of electromagnetic field, temperature field, and vacuum field based on the accelerator's three-dimensional geometric model. The data-driven iteration dynamically corrects the model parameters through real-time received measured data from the all-dimensional sensing module. The digital twin core module performs noise reduction on the measured data and then integrates it with the model's basic parameters to drive synchronous model updates. The noise reduction process uses a filtering algorithm to remove electromagnetic interference noise. The edge-cloud collaborative computing module, deployed locally on the accelerator's edge, communicates with both the full-dimensional sensing module and the digital twin core module. The edge module processes the collected data in real time and dynamically calibrates beam measurements based on the output of the digital twin core module. It also communicates with a cloud server deployed on a remote server, which iteratively optimizes the model and calibration logic based on historical data. The core functions of the edge module include: Receive the measured data from the full-dimensional perception module and drive the digital twin core module to perform real-time updates; Based on the built-in multi-physics coupling interference logic, the beam measurement value is dynamically calibrated. The logic is implemented by establishing a quantitative mapping model between the beam measurement value and the environmental interference. The output of the mapping model is an interference comprehensive coefficient. The sampling frequency of the all-dimensional sensing module is adaptively adjusted according to the interference comprehensive coefficient. The adjustment logic is as follows: a threshold for the interference comprehensive coefficient is preset; when the interference comprehensive coefficient calculated in real time exceeds the threshold, the sampling frequency is increased; when the interference comprehensive coefficient is lower than the threshold, the sampling frequency is decreased. The reinforcement learning optimization module, integrated into the digital twin core module, is used to autonomously learn and output the optimal beam control strategy based on the state information output by the digital twin core module. The execution module is communicatively connected to the edge terminal and is used to execute beam correction and fault repair commands generated by the reinforcement learning optimization module and issued by the edge terminal.

2. The control system for accelerator beam measurement and diagnosis as described in claim 1, characterized in that, The all-dimensional sensing module includes a beam parameter sensor submodule and a multi-physics field sensor submodule; The beam parameters collected by the beam parameter sensor submodule include beam position, beam intensity and beam emittance. It consists of a two-dimensional position sensitive detector deployed at the inlet and outlet of the accelerator vacuum chamber, a Faraday tube deployed in the focusing section and an emittance measurement unit. The environmental data collected by the multi-physics sensor submodule includes electromagnetic field strength, cavity temperature, radiation dose, and vacuum level. It consists of a Hall sensor deployed near the magnet coil, a thermocouple embedded in the cavity surface, an ionization chamber deployed in the radiation-sensitive area, and a capacitive vacuum gauge connected to the vacuum pipeline.

3. The control system for accelerator beam measurement and diagnosis as described in claim 1, characterized in that, The core functions of the cloud include: Receive the desensitized historical data uploaded by the edge terminal, the historical data including beam parameters, environmental data and equipment control records; Based on the historical data, the model parameters of the digital twin core module and the multi-physics coupling interference logic at the edge are iteratively optimized.

4. The control system for accelerator beam measurement and diagnosis as described in claim 1, characterized in that, The reinforcement learning optimization module adopts a deep deterministic policy gradient algorithm. Its state space includes beam parameters consisting of beam position, beam intensity and beam emittance, environmental interference consisting of electromagnetic field intensity, cavity temperature, radiation dose and vacuum degree, and equipment operating state consisting of magnet current, radio frequency power and vacuum valve opening. Its action space includes magnet current adjustment, radio frequency power optimization value and vacuum valve opening adjustment. Its reward function includes a positive reward term to encourage beam stability and a negative penalty term to suppress beam loss and system energy consumption; Its training process includes: the first stage is pre-training in the virtual environment constructed by the core module of the digital twin; The second stage involves online iterative optimization of the strategy by combining real-time data from the physical system at the edge.

5. The control system for accelerator beam measurement and diagnosis as described in claim 1, characterized in that, The execution module includes a beam correction magnet for beam trajectory correction, an RF power source for beam energy regulation, and a vacuum valve for vacuum degree control.

6. A control system for accelerator beam measurement and diagnosis as described in claim 1 or 5, characterized in that, It also includes a fault pre-diagnosis submodule, which is integrated into the digital twin core module. By comparing the real-time parameter differences between the virtual beam and the physical beam, and combining the beam evolution model optimized in the cloud, it can determine two types of potential fault risks: orbital deviation and beam loss, and trigger an early warning. After receiving the early warning information, the reinforcement learning optimization module outputs targeted self-repair and adjustment instructions, which are then executed by the execution module driven by the edge terminal.

7. The control system for accelerator beam measurement and diagnosis as described in claim 1, characterized in that, It also includes a human-machine collaboration interface, deployed in the cloud, which allows engineers to view the system's operating status and the evolution process of the digital twin model, and has the function of issuing policy correction instructions to the reinforcement learning optimization module.

8. The control system for accelerator beam measurement and diagnosis as described in claim 1, characterized in that, Data transmission between the full-dimensional sensing module and the edge terminal adopts a shielded transmission link, and the shielded transmission link is provided with an anti-electromagnetic interference shielding layer.

Images