Security constraint optimization control method for beam scanning system

Abstract

The invention discloses a security constraint optimization control method for a beam scanning system, which relates to the technical field of particle acceleration and comprises the following steps of: loading a machine learning agent model to predict system performance and security constraint; collecting beam data in real time and calculating a uniformity index; by taking the index as an optimization target, dynamically constructing and updating a high-probability safety parameter set as an optimization feasible region based on the prediction and uncertainty of an agent model; by judging whether candidate points generated by the benchmark acquisition function are located in a security set or not, intelligent decision is made, and a known security point or a security extension boundary point is selected for evaluation; and updating the model by using new data and forming closed-loop feedback, thereby realizing autonomous optimization tuning of the beam uniformity on the premise of ensuring the safety of the whole process. According to the method, a'dynamic safety set 'is introduced as an absolute constraint boundary of a Bayesian optimization exploration process, so that a core safety bottleneck faced by a high-power beam scanning system in an automatic tuning process is solved.

Description

Technical Field

[0001] This invention relates to the field of particle acceleration technology, and more specifically to a safety constraint optimization control method for beam scanning systems. Background Technology

[0002] The use of high-intensity particle beams generated by cyclotrons to bombard specific targets to prepare medical radioisotopes has become a mainstream technological approach in modern nuclear medicine. In this production process, the beam scanning system, as a key subsystem ensuring uniform beam distribution on the target surface, directly determines the yield and quality of the product through the optimization and tuning of its scanning parameters. With the continuous increase in beam power, the need for automated and intelligent optimization of scanning parameters is becoming increasingly urgent. However, the safety issues inherent in the optimization process itself are also becoming prominent, becoming a core bottleneck restricting production efficiency and continuity. Currently, optimization methods attempted in this type of system mainly suffer from the following inherent limitations. Model-free global optimization algorithms, such as particle swarm optimization or genetic algorithms, while possessing powerful global search capabilities, are essentially unconstrained stochastic exploration processes. In high-value, high-risk safety-critical systems (such as accelerators), their blind exploration behavior can easily trigger safety constraints such as electromagnet power over-limit and beam loss overload, leading to emergency beam cutoff and unplanned production interruptions. This is unacceptable for commercial production lines that strive for continuous and stable operation. While gradient descent-based local optimization methods can achieve fast convergence, they are heavily reliant on the choice of initial point, prone to getting trapped in local optima, and cannot guarantee the safety of the entire optimization trajectory from the initial point to the optimum. Furthermore, while offline optimization methods based on high-fidelity physical simulations can accurately predict beam behavior, their enormous computational cost—up to hundreds of seconds per evaluation—fundamentally hinders their application in online control. Even traditional or existing constrained Bayesian optimization methods are typically designed to find a safe optimal solution, rather than guaranteeing the safety of each iteration in the optimization path. Therefore, they still allow high-risk exploration near constraint boundaries. This "trial-and-error" learning model remains too risky for the absolutely safe operating environment of industrial accelerators. In summary, existing technologies generally fail to achieve a fundamental balance between exploring the unknown parameter space to find a better solution and ensuring that the system never triggers safety interlocks. Their core flaw lies in the lack of an intrinsic mechanism that can guarantee the safety of the entire optimization path with a high probability. Summary of the Invention

[0003] To achieve a fundamental balance between exploring the unknown parameter space to find a better solution and ensuring that the system never triggers a safety interlock, this invention proposes a safety constraint optimization control method for beam scanning systems, comprising the following steps: S1: Load a machine learning proxy model for predicting the target performance and safety constraints of the beam scanning system; S2: Real-time acquisition of beam data during the operation of the beam scanning system, and calculation of target performance indicators characterizing the uniformity of beam distribution based on the beam data; S3: Taking the target performance index as the optimization objective, based on the prediction results and uncertainty estimation of the surrogate model, a safety set is dynamically constructed and updated as the feasible region for subsequent optimization exploration. This safety set consists of all scanning parameters that have a high probability of satisfying all safety constraints. S4: Perform Bayesian optimization decision under the constraints of the safe set. Determine the evaluation point for this iteration by judging whether the candidate parameter point generated by a benchmark acquisition function is within the safe set. If it is within the safe set, select the point for evaluation to utilize the known safe region. If it is outside the safe set, select an exploration point on the boundary of the current safe set for evaluation to safely expand the range of the safe set. S5: Update the surrogate model using the new measurement data obtained at the evaluation point, and return to step S2.

[0004] This invention fundamentally solves the core safety bottleneck faced by high-power beam scanning systems in automated tuning by introducing a "dynamic safety set" as the absolute constraint boundary of the Bayesian optimization exploration process.

[0005] Furthermore, the surrogate model is a Gaussian process regression model.

[0006] Furthermore, the proxy model is constructed by using scanning parameters as input to the proxy model and the predicted values ​​of the target performance index and one or more security constraint indexes as output to train the proxy model.

[0007] Further, in step S4, the security set is defined as follows at the t-th iteration: In the formula, Let be the safe set for the t-th iteration. For candidate parameter points, For the entire scan parameter space, Let i be the i-th security constraint function. The lower confidence bound function is obtained through the surrogate model, based on all observation data up to the t-th iteration. The preset confidence level parameter, The total number of safety constraints.

[0008] Furthermore, in step S4, the benchmark acquisition function is a strategy that maximizes the confidence bound on the objective function.

[0009] Furthermore, in step S4, selecting an exploration point on the boundary of the safe set specifically involves maximizing the lower confidence bound of any safety constraint function on the boundary of the safe set.

[0010] Furthermore, in step S4, a step size constraint mechanism is introduced to restrict the next candidate parameter point to be evaluated to a preset distance range from the current optimal parameter point.

[0011] Furthermore, the preset distance is an Euclidean distance threshold defined in the normalized scan parameter space.

[0012] Furthermore, in step S2, the center position deviation of the beam trajectory is calculated based on the beam data, and when the center position deviation exceeds a preset threshold, closed-loop stabilization adjustment of the beam trajectory is performed.

[0013] Furthermore, it also includes a safety monitoring process that runs asynchronously and in parallel with the optimization control method. This process independently and continuously monitors key safety parameters and triggers the highest priority hardware protection instruction when any monitored parameter exceeds its safety threshold.

[0014] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The present invention proposes a safety constraint optimization control method for beam scanning system, which uses a machine learning proxy model to accelerate high-fidelity physical simulation in milliseconds, thereby realizing rapid evaluation and uncertainty quantification of target performance and multiple safety constraints; (2) By dynamically constructing and maintaining a high-probability safe parameter feasible domain, and designing a dual-mode decision-making mechanism of "exploitation-safe expansion", we can ensure that each parameter exploration behavior of the optimization algorithm is strictly limited to the safe set, thereby actively avoiding the risk of triggering safety interlock due to blind exploration. (3) By integrating step size constraint mechanism and parallel independent safety monitoring, the stability and robustness of the control system in complex industrial scenarios are further enhanced, ultimately providing support for the continuous, safe and efficient production of high-power radioactive isotopes. Attached Figure Description

[0015] Figure 1 This is a step diagram of a safety constraint optimization control method for a beam scanning system. Detailed Implementation

[0016] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0017] In high-power particle accelerator applications, especially in the commercial production of medical radioisotopes, the control precision and operational reliability of the beam scanning system directly determine production efficiency and equipment safety. With the continuous increase in beam power, the uniformity of beam distribution on the target surface faces increasingly stringent requirements, while traditional parameter optimization methods for achieving this goal face fundamental technical bottlenecks. Neither global model-free optimization algorithms nor methods based on local gradients can guarantee with high probability that each iteration satisfies complex safety constraints while exploring optimal parameter configurations. This cycle of "exploration-trigger interlocking-interruption" has become a core obstacle restricting the continuous and stable operation of high-value accelerator devices 24 / 7. Existing technological approaches, either constrained by the binary opposition of safety and efficiency or limited by the computational time of high-fidelity physical simulations, consistently lack a systematic solution that can theoretically ensure the safety of the entire optimization trajectory while simultaneously possessing computational efficiency and global optimization capabilities.

[0018] In this technological context, such as Figure 1 As shown, this invention proposes a safety constraint optimization control method for beam scanning systems, the core steps of which include: S1: Load a machine learning proxy model for predicting the target performance and safety constraints of the beam scanning system; S2: Real-time acquisition of beam data during the operation of the beam scanning system, and calculation of target performance indicators characterizing the uniformity of beam distribution based on the beam data; S3: Taking the target performance index as the optimization objective, based on the prediction results and uncertainty estimation of the surrogate model, a safety set is dynamically constructed and updated as the feasible region for subsequent optimization exploration. This safety set consists of all scanning parameters that have a high probability of satisfying all safety constraints. S4: Perform Bayesian optimization decision under the constraints of the safe set. Determine the evaluation point for this iteration by judging whether the candidate parameter point generated by a benchmark acquisition function is within the safe set. If it is within the safe set, select the point for evaluation to utilize the known safe region. If it is outside the safe set, select an exploration point on the boundary of the current safe set for evaluation to safely expand the range of the safe set. S5: Update the surrogate model using the new measurement data obtained at the evaluation point, and return to step S2.

[0019] Specifically, the implementation of the method of this invention is typically based on a typical beam scanning system hardware platform, which mainly includes a particle accelerator, a scanning magnet, a beam position monitor, a beam profile monitor, and a central control unit that executes the control algorithm. However, it should be particularly noted that the core of this invention lies in the innovation of its control method, which itself does not depend on any specific hardware configuration. Any beam scanning system capable of providing the necessary beam parameter sensing and executing scanning magnet parameter adjustments can apply the method described in this invention through software or firmware upgrades.

[0020] The core process of this invention begins with system initialization and preparation. Upon power-up of the control unit, a comprehensive operational clearance check process is automatically executed, serving as the first line of defense for system safety. This process systematically checks the status of all critical subsystems, including but not limited to: confirming that the temperature of the scanning magnet coil and its cooling water flow rate are within preset safety thresholds; verifying that sensing units such as the beam profile monitor are correctly positioned and that their background noise level is below technical requirements; checking the operational status of the signal generator and magnetic field power supply; and receiving and confirming the heartbeat signal emitted by the system self-test. If any necessary condition is missing, the system will prohibit beam activation, thereby eliminating potential safety hazards at the source.

[0021] After passing the runtime license check, the system loads one or more pre-trained machine learning surrogate models from non-volatile storage. These models serve as the computational engine for subsequent efficient and secure optimization decisions. Building this surrogate model is a crucial preliminary step, and its standard construction method includes two main phases: sparse sampling and model training.

[0022] First, using a validated high-fidelity beam transport physics simulation program, an initial training dataset covering the main operating modes of the system is generated by sparse sampling within the required scanning parameter space (e.g., scanning magnet current amplitude, scanning frequency ratio, waveform phase difference, etc., covering the X and Y directions). Then, a machine learning algorithm is used to train this dataset, using the combination of scanning parameters as model input and the target performance indicators and various safety constraint indicators obtained from simulation calculations as model output, thereby fitting the mapping relationship between scanning parameters and the complex nonlinear response of the beam. In a preferred embodiment of the invention, the surrogate model employs a Gaussian process regression algorithm, as it can naturally provide uncertainty estimates for predicted values, laying a solid foundation for the subsequent construction of the safety set. Finally, the trained surrogate model becomes a high-precision evaluator capable of providing millisecond-level rapid predictions of beam uniformity and safety constraints.

[0023] After successfully loading the proxy model, the operator can set the initial scanning waveform parameters and control target values ​​through the human-machine interface. For example, the initial scanning frequency can preferably be set in the range of 1kHz to 10kHz, with 5kHz as the reference frequency. At the same time, the operator needs to set specific control targets, such as the desired beam distribution uniformity target value. Subsequently, the system enters an autonomous, continuously iterative optimization control loop.

[0024] The core of this optimized control loop lies in the synchronous acquisition of multi-source beam data and the real-time calculation of key parameters. The control unit synchronously acquires multi-channel data streams from different sensing units at a predetermined frequency. These data specifically include: the total beam intensity signal from the non-interceptor beam transformer, used to monitor the total beam dose; the beam center position signal from the beam position monitor, used to determine the beam trajectory alignment; and raw two-dimensional beam intensity distribution image data from the downstream beam profile monitor, which is the direct basis for evaluating the control effect. Based on these real-time acquired raw data, the system calculates two key physical quantities using a built-in algorithm as the basis for subsequent intelligent decisions. The first key quantity is the beam center deviation, obtained by calculating the difference between the real-time beam center position signal and the preset center setpoint, providing real-time beam trajectory offset information. The second, and more crucial, quantity is the target performance index characterizing the uniformity of the beam distribution. The target performance index is calculated by calling an image processing algorithm to analyze the effective irradiation area of ​​the two-dimensional beam intensity distribution image and calculate the standard deviation of the intensity values ​​of all pixels within that area. Compared with the average The ratio, i.e., uniformity The smaller the value of this indicator, the more uniform the beam distribution and the better the control effect.

[0025] After obtaining the aforementioned key parameters, the system executes two levels of control tasks in parallel. Among these, the uniformity autonomous tuning based on secure Bayesian optimization is the core embodiment of the invention's intelligence and safety capabilities. First, as a high-priority and parallel process, the system performs closed-loop stabilization adjustment of the beam trajectory. When the calculated absolute value of the beam center deviation exceeds the operator-preset threshold... At this time, the system immediately generates a corresponding correction signal and outputs it to the guide magnet power supply upstream of the beamline, realizing rapid correction of the beam trajectory. This process ensures that the beam can accurately hit the center region of the target surface, providing a stable foundation for subsequent uniformity optimization. Meanwhile, the core uniformity optimization process continues. This process operates entirely within the safety set framework constructed and dynamically maintained in the aforementioned steps. The safety set is precisely defined as a subset of parameters in each iteration, within which the surrogate model predictions of all key safety constraint indicators satisfy the safe operating conditions with a very high probability. Specifically, in the t-th iteration, the safety set... Mathematically defined as: , in, For a specific candidate parameter point, For the entire scan parameter space, Let i be the i-th security constraint function. Based on all observation data up to the t-th iteration, the lower confidence bound function of the i-th constraint function is calculated using the surrogate model. The preset confidence level parameter, The total number of safety constraints. This definition ensures that any constraint belonging to... The parameter points have their risk of violating any safety constraint controlled to a very low level.

[0026] Based on this dynamic safety set, the system executes a decision logic called the safety acquisition function, which is the essence of the secure Bayesian optimization algorithm. Its decision process is as follows: First, the algorithm calculates a baseline candidate point that does not consider the current safety set. This point is typically obtained by maximizing the upper confidence bound of the objective function, aiming to find the point with the greatest potential for improvement. Next, the algorithm determines... Is it located in the current security set? within. like If this indicates that the most promising point is itself safe, the algorithm will directly select that point for evaluation; this is the "exploitation" mode, designed for rapid convergence. If If the algorithm fails to find a safe point, it will not risk trying that unsafe point. Instead, it will intelligently switch to "safe exploration" mode. In this mode, the algorithm selects an exploration point on the boundary of the current safe set that can most safely expand the range of the safe set. . Specifically, It is determined by maximizing the lower confidence bound of a certain security constraint function on the current security set boundary, i.e. This mechanism ensures that the optimization process can make full use of known safe areas to achieve rapid convergence, while also cautiously exploring outwards to continuously expand the boundaries of safe operations without crossing the safety red line.

[0027] To further enhance the engineering practicality and robustness of the optimization process, this embodiment of the invention also introduces a step size constraint mechanism in the aforementioned safety optimization decision-making. This mechanism requires that, in each iteration, the next candidate parameter point to be evaluated (whether a utilization point or a safety exploration point) must be limited to a preset distance range from the current optimal parameter point. In a specific implementation, this preset distance is defined as a Euclidean distance threshold in the normalized scan parameter space. For example, this threshold can be set to 10% of the total range of the parameter space. This mechanism effectively prevents beam trajectory oscillations and system instability caused by parameter mutations, ensuring the smoothness and stability of the optimization process.

[0028] Selected assessment points After performing parameter adjustments, the system uses the latest real measurement data obtained at the new parameter point to synchronously update the parameters of the machine learning surrogate model. This step enables the model to learn online and evolve self-evolving, allowing it to continuously approximate the dynamic response characteristics of the real physical system, thus ensuring the long-term effectiveness and adaptability of the optimization strategy. After completing the model update, the system automatically returns to the data acquisition and calculation steps, initiating a new round of optimization iterations, thereby forming a secure closed-loop feedback control system with self-learning and continuous optimization capabilities.

[0029] Furthermore, as a crucial safety redundancy design, this embodiment of the invention also includes a layered safety monitoring process that runs asynchronously and in parallel with the aforementioned optimized control loop. This process is independent of the main control loop and provides ultimate, physical-level safety assurance. It comprises two layers: first, continuous permission monitoring, which continuously verifies whether all critical operating conditions are consistently met during system operation; and second, rapid protection execution, serving as the final barrier for system safety. This subsystem independently and continuously monitors critical safety parameters such as beam overflow rate and scanning magnetic field parameter deviation. Once any abnormality is detected, such as beam overflow exceeding 1% of the total beam current, or scanning magnetic field parameters deviating from preset values ​​by more than 10%, this subsystem will bypass all conventional control loops and directly trigger the highest-priority emergency beam cutoff command, immediately stopping the beam supply, thereby forming the fastest and most reliable physical protection barrier for the accelerator device and target station equipment.

[0030] In summary, the safety constraint optimization control method for beam scanning systems proposed in this invention achieves a fundamental breakthrough in technical effectiveness. By introducing a safety Bayesian optimization framework based on machine learning surrogate models and dynamic safety sets, it successfully overcomes the long-standing core challenge of balancing safety and efficiency in the automated tuning process of high-power beam scanning systems. This method not only proactively avoids the risk of safety interlocking triggered by blind algorithmic exploration by strictly confining optimization exploration behavior within the safety parameter space with high probability, providing a reliable guarantee for fully automated continuous operation, but also accelerates time-consuming high-fidelity physical simulation to millisecond-level evaluation using surrogate models, breaking down the barrier that high-precision models cannot be used for online optimization. Ultimately, this method promotes a fundamental shift in the control paradigm from static open-loop, relying on human experience to dynamic closed-loop, autonomous intelligent decision-making. It can continuously maintain a high degree of uniformity in beam distribution on the target surface while satisfying dozens to hundreds of complex safety constraints, thereby significantly improving the automation level of the production process, batch-to-batch consistency, and overall system robustness.

[0031] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0032] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0033] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0034] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

Claims

1. A safety-constrained optimization control method for a beam scanning system, characterized in that, The method comprises the steps of: S1: loading a machine learning surrogate model for predicting the performance of a target and safety constraints of a beam scanning system; S2: collecting beam data in real time during the operation of the beam scanning system, and calculating a target performance index representing the uniformity of the beam distribution based on the beam data; S3: dynamically constructing and updating a safety set as the feasible region for subsequent optimization exploration based on the prediction results and uncertainty estimation of the surrogate model, the safety set being composed of all scanning parameters that are highly likely to satisfy all safety constraints; S4: performing a Bayesian optimization decision under the constraints of the safety set, and deciding the evaluation point of this iteration by judging whether a candidate parameter point generated by a benchmark acquisition function is located within the safety set: if yes, selecting the point for evaluation to utilize the known safe region, and if not, selecting an exploration point on the boundary of the current safety set for evaluation to safely expand the range of the safety set; S5: updating the surrogate model using new measurement data obtained at the evaluation point, and returning to step S2.

2. The safety-constrained optimization control method for a beam scanning system of claim 1, wherein, The surrogate model is a Gaussian process regression model.

3. The safety-constrained optimization control method for a beam scanning system of claim 1, wherein, The surrogate model is constructed in the following manner: taking the scanning parameters as the input of the surrogate model, and taking the predicted values of the target performance index and one or more safety constraint indexes as the output of the surrogate model, the surrogate model is trained.

4. The safety-constrained optimization control method for a beam scanning system of claim 1, wherein, In the S4 step, the safety set at the tthiteration is defined as: wherein is the safety set for the tth iteration, is the candidate parameter point, is the entire scan parameter space, is the ith safety constraint function, is the lower confidence bound function obtained by the surrogate model based on all observation data up to the tth iteration, is a preset confidence level parameter, is the total number of safety constraints.

5. The safety-constrained optimization control method for a beam scanning system of claim 1, wherein, In the S4 step, the benchmark acquisition function is a strategy of maximizing the upper confidence bound of the objective function.

6. The safety-constrained optimization control method for a beam scanning system of claim 1, wherein, In the S4 step, the selection of the exploration point on the boundary of the safety set is specifically: maximizing the lower confidence bound of any safety constraint function on the boundary of the safety set.

7. The method for safety-constrained optimization control of a beam scanning system of claim 1, wherein, In the S4 step, a step constraint mechanism is further introduced to limit the next candidate parameter point to be evaluated within a preset distance range from the current optimal parameter point.

8. The safety-constrained optimization control method for a beam scanning system of claim 7, wherein, The preset distance is an Euclidean distance threshold defined in the normalized scanning parameter space.

9. The method for safety-constrained optimization control of a beam scanning system of claim 1, wherein, In the S2 step, the center position deviation of the beam orbit is also calculated based on the beam data, and when the center position deviation exceeds a preset threshold, a closed-loop stable adjustment of the beam orbit is performed.

10. The method of claim 1, wherein, A safety monitoring process running asynchronously and in parallel with the optimization control method is further included, which independently and continuously monitors key safety parameters, and triggers a hardware protection instruction of the highest priority when any monitored parameter exceeds its safety threshold.

Images