Optimization method of proton accelerator vacuum system based on local large model
Patent Information
- Application Number
- CN202511580035.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-13
- Publication Date
- 2026-02-06
AI Technical Summary
[0004]然而，在真空系统中，大量使用价格高昂的真空规会增加物料成本，同时过多的真空规会导致安装端口、电缆布线、信号采集通道以及后续校准与维护工作的复杂性增加，降低了系统的整体可靠性
1. 本地部署大模型能保证数据隐私与安全，减少数据泄露风险，满足医疗行业合规性要求，还保障了系统的实时性与可靠性，消除网络传输延迟和潜在中断风险，维持加速器的稳定运行，同时降低本地部署的门槛与资源消耗；
Smart Images
Figure CN121483540A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field of accelerator vacuum systems, and in particular to an optimization method for a proton accelerator vacuum system based on a local large model. Background Technology
[0002] Proton therapy accelerators hold significant importance in the medical field, and with the continuous development of medical technology, their therapeutic effects on diseases such as cancer are becoming increasingly remarkable. Proton therapy can precisely concentrate energy at the tumor site, reducing damage to surrounding normal tissues and improving the effectiveness and safety of treatment. During the operation of a proton therapy accelerator, data processing and system stability are crucial, directly affecting the treatment outcome and patient safety. Simultaneously, the rapid development of information technology places higher demands on the data processing and system optimization of proton therapy accelerators.
[0003] In existing technologies, common methods for addressing data processing and system stability issues related to proton therapy accelerators include: For data processing, uploading data to public cloud servers for centralized processing and analysis, leveraging the powerful computing capabilities of cloud servers to complete complex large-scale model training and data analysis tasks; and for the construction and maintenance of the vacuum system, employing traditional vacuum equipment and monitoring methods, ensuring stable operation through regular manual inspection and calibration. Vacuum level monitoring primarily relies on a large number of vacuum gauges for real-time monitoring to guarantee the system's normal operation.
[0004] However, in vacuum systems, the extensive use of expensive vacuum gauges increases material costs. At the same time, too many vacuum gauges increase the complexity of installation ports, cable routing, signal acquisition channels, and subsequent calibration and maintenance work, thus reducing the overall reliability of the system. Summary of the Invention
[0005] In order to effectively reduce the hardware cost of the vacuum system of a proton therapy accelerator while ensuring data security and system reliability, this application provides an optimization method for the vacuum system of a proton accelerator based on a local large model.
[0006] The optimization method for a proton accelerator vacuum system based on a local large model provided in this application adopts the following technical solution: An optimization method for a proton accelerator vacuum system based on a local large-scale model includes the following steps: Deploy large models on local servers; A vacuum system is established, which includes multiple sets of ion pumps and multiple vacuum gauges, with the vacuum gauges located between each set of ion pumps. After the vacuum system stabilizes, the vacuum parameters of multiple ion pumps and vacuum gauges are collected and saved locally. The large model is trained based on the collected data to establish a mapping relationship between the ion pump operating parameters and the vacuum gauge readings, and a performance threshold is preset. When the error between the ion pump vacuum reading output by the large model and the vacuum gauge reading is less than the preset performance threshold, the number of vacuum gauges in the vacuum system is reduced; if the error is greater than the performance error threshold, the large model is trained again.
[0007] By adopting the above technical solutions, and by deploying a large model locally, sensitive medical data and equipment parameters are processed within the internal network, meeting the compliance requirements of the medical industry, while avoiding the risks of network latency and interruption. By establishing a mapping relationship between ion pumps and vacuum gauges, the number of expensive vacuum gauges is reduced after the accuracy of the large model reaches the standard, significantly reducing the system hardware cost. By setting preset performance thresholds and setting condition judgment processes, hardware adjustments are only made after the accuracy of the large model has been fully verified, ensuring system reliability.
[0008] Preferably, the locally deployed large model specifically includes: Deploy a large model using the Ollam framework on a local server, and enable data acquisition scripts and monitoring systems to communicate with the large model through the local API interface provided by the Ollam framework.
[0009] By adopting the above technical solutions, using the Ollam framework and standardized API interfaces, the technical threshold and resource consumption for local deployment of large models are significantly reduced, and a unified communication interface is provided, making it easy to integrate large models into existing data acquisition and monitoring systems.
[0010] Preferably, the process of collecting and saving the vacuum parameters of multiple ion pumps and vacuum gauges locally includes the following steps: The collected data is cleaned to remove outlier data points; The cleaned data is structured and stored locally. The data includes date, time, vacuum gauge number, vacuum gauge reading, ion pump number, and ion pump operating data.
[0011] By adopting the above technical solutions, abnormal data points are effectively identified and removed through the data cleaning process, thereby improving the quality and reliability of training data; and by using structured storage, the traceability and analyzability of the data are ensured.
[0012] Preferably, the large model is trained based on the collected vacuum parameters of the ion pump and the vacuum parameters of the vacuum gauge, including the following steps: The structured collected data was randomly divided into training set, validation set and test set according to a certain proportion; A mapping function F is set within the large model, and the predicted vacuum reading output by the large model is equal to F (the vacuum reading of the ion pump), so that the predicted vacuum reading is approximately equal to the vacuum reading of the vacuum gauge. The internal parameters of the large model are adjusted using data from the training set, and the difference between the self-reading vacuum degree of the ion pump and the actual reading of the vacuum gauge is quantified by the mean square error (MSE) loss function. During training, the mean absolute error and root mean square error are monitored simultaneously on the validation set.
[0013] By adopting the above technical solutions and rationally dividing the training set, validation set, and test set, the comprehensiveness and objectivity of the large model evaluation are ensured; the mapping function F is clearly defined to establish an accurate large model of the relationship between ion pump parameters and vacuum readings; and through the synchronous monitoring of multiple error indicators, fine control of the training process and prevention of overfitting are achieved.
[0014] As a preferred option, this also includes encapsulating the trained large model into a standardized service, and calling the large model directly to output the vacuum degree parameters via a script, specifically including the following steps: Set up a monitoring script to periodically read the ion pump's operating data; The trained large model is deployed as a network service via an API interface. The network service listens for requests from monitoring scripts, receives the operating data of the ion pump as input, and returns the predicted vacuum reading as output.
[0015] By adopting the above technical solutions, the trained large model is encapsulated as a standard network service, enabling rapid deployment and invocation of the large model's capabilities; and by periodically reading data from the monitoring script, real-time monitoring of the system status is achieved.
[0016] Preferably, the monitoring script will simultaneously display the predicted vacuum reading, the operating data of the ion pump, and the vacuum reading of the vacuum pump on the monitoring interface, and / or store them in the database.
[0017] By adopting the above technical solution, key parameters can be displayed synchronously through the monitoring interface, making it easy for operators to keep track of the system status in real time.
[0018] As a preferred option, this also includes setting a data update cycle and using the incremental data collected within the cycle to fine-tune the deployed large model.
[0019] By adopting the above technical solutions, a regular large model update mechanism is established to ensure that the large model can adapt to changes in the system state.
[0020] Preferably, the step of setting a data update period and fine-tuning the deployed large model using incremental data collected within the period specifically includes: Set up an empty incremental dataset; After each data update cycle ends, the synchronous operation data of all ion pumps and vacuum gauges within the data update cycle are automatically collected and stored in the incremental dataset; The currently deployed large model is fine-tuned using the incremental dataset to generate a new candidate large model for evaluation; The performance of the new large model is evaluated on validation data collected at the same time that were not used for training. When the error between the predicted vacuum reading output by the new large model and the vacuum reading of the vacuum gauge is less than the performance threshold, the new large model is automatically deployed online.
[0021] By adopting the above technical solutions, the large model can be continuously self-optimized and upgraded, maintaining the accuracy and reliability of the system in long-term operation.
[0022] In summary, this application includes at least one of the following beneficial technical effects: 1. Local deployment of large models can ensure data privacy and security, reduce the risk of data leakage, meet the compliance requirements of the medical industry, ensure the real-time performance and reliability of the system, eliminate network transmission delays and potential interruption risks, maintain the stable operation of the accelerator, and reduce the threshold and resource consumption of local deployment. 2. Establish a vacuum system that includes ion pump groups and vacuum gauges. The ion pumps can generate and maintain a high vacuum environment, and the vacuum gauges can accurately detect the vacuum pressure in real time, providing real value references for system status evaluation and large model training. Furthermore, the grouped arrangement of ion pumps and the interspersed arrangement of vacuum gauges facilitates batch data acquisition and monitoring of local vacuum environments. 3. Adjusting the large model prediction parameters based on the difference between the ion pump set operating data and the vacuum gauge reading can minimize the difference between the two, ensure that the large model has good generalization ability, guarantee the accuracy of data matching, thereby reducing the number of vacuum gauges required and lowering costs. Attached Figure Description
[0023] Figure 1 This is a flowchart illustrating the implementation process of this application. Detailed Implementation
[0024] The following combination Figure 1 This application will be described in further detail.
[0025] This application discloses an optimization method for a proton accelerator vacuum system based on a local large model.
[0026] Reference Figure 1An optimization method for a proton accelerator vacuum system based on a local large model includes the following steps: S1: Locally deployed large model.
[0027] Building large models on local servers ensures data privacy and security. Proton therapy accelerators handle sensitive medical data and critical equipment parameters; localized processing ensures all training data and large model calculations are completed within the internal network, physically isolated from the public internet to minimize the risk of data leakage and meet the stringent compliance requirements of the healthcare industry. Furthermore, local deployment guarantees the system's real-time performance and reliability, minimizing network transmission latency and potential interruption risks, and maintaining the accelerator's stable operation.
[0028] Specifically, this application uses a DeepSeekR1 distillation version (such as 32BQ4) as the large model and deploys it using frameworks such as Ollam. DeepSeek-R1's excellent inference capabilities ensure that it can accurately learn patterns from complex device data, while quantization technology and Ollam significantly reduce the threshold and resource consumption for local deployment, enabling advanced large models to serve industrial-grade application scenarios efficiently and stably. Ollam also provides a standardized local API interface, allowing subsequently developed data acquisition scripts and monitoring systems to easily communicate with the local large model and integrate into existing automated processes.
[0029] S2: Establish a vacuum system.
[0030] The vacuum system comprises multiple vacuum gauges and multiple ion pumps. The ion pumps continuously adsorb gas molecules through electrophysical means, generating and maintaining the required high vacuum or even ultra-high vacuum environment. The vacuum gauges are used to detect the vacuum pressure of the vacuum system in real time and accurately, providing a realistic reference for system status assessment and large model training.
[0031] Multiple ion pumps are grouped together, with multiple vacuum gauges positioned between each group. Arranging a large number of ion pumps in groups allows for batch data acquisition while simultaneously creating a vacuum environment. Interspersing higher-precision vacuum gauges ensures that each gauge can monitor the localized vacuum environment created by the combined action of its surrounding group of ion pumps.
[0032] S3: After the vacuum system stabilizes, collect the vacuum parameters of multiple ion pumps and the vacuum gauge, and save them locally.
[0033] When the vacuum gauge readings, ion pump operating current, and voltage no longer show significant and continuous changes, the vacuum system enters a stable state. After the vacuum system enters a stable state, the vacuum gauge readings and ion pump operating data are collected. The ion pump operating data includes the ion pump operating current, ion pump operating voltage, and ion pump vacuum reading. Synchronous acquisition ensures that each set of data completely records the system's state at a specific moment.
[0034] Next, the collected data is structured and stored. All collected data is indexed by timestamps and stored locally using a standardized structure (such as CSV or JSON). Each data point includes the date, time, vacuum gauge number, vacuum reading for each gauge, ion pump number, and transport data for each pump. This structured storage not only ensures data traceability and facilitates subsequent querying and analysis by time or device, but also provides convenience for efficient reading and processing of large-scale models.
[0035] Preferably, data cleaning is also required during the data collection process. Due to factors such as transient equipment failures, sensor malfunctions, or human error, outliers may be mixed in with the raw data. Therefore, these unreasonable data points must be identified and removed manually or by setting threshold rules.
[0036] S4; Adjust the large model based on the collected vacuum parameters of the ion pump and the vacuum gauge.
[0037] The structured collected data was randomly divided into training, validation, and test sets in an 8:1:1 ratio. The training set was used to learn the parameters of the large model; the validation set was used to evaluate the performance of the large model in real time during training to adjust hyperparameters and prevent overfitting; and the test set was used for a final unbiased evaluation of the generalization ability of the large model after training. A mapping function F was set within the large model such that the predicted vacuum reading output by the large model = F(vacuum reading of the ion pump) ≈ vacuum reading of the vacuum gauge, and the learning rate, batch size, and number of training epochs were set.
[0038] After training begins, the large model continuously adjusts its internal parameters using the training set data to minimize the difference between the ion pump's self-reading vacuum level and the actual reading of the vacuum gauge. This difference can be quantified using a loss function (such as mean squared error, MSE) to minimize the average of the squared differences between the predicted and actual values across all training samples. Specifically, for a training batch containing N samples, the formula for calculating the mean squared error is: MSE = (1 / N) * Σ(predicted vacuum reading ᵢ - vacuum level ᵢ)², where i represents the i-th training sample currently being calculated. A larger mean squared error indicates a less accurate prediction by the large model; a smaller mean squared error indicates a more accurate prediction.
[0039] During training, key statistical metrics, such as mean absolute error (MAE) and root mean square error (MSE), need to be monitored simultaneously on the validation set. These statistical metrics provide a quantitative assessment of the prediction accuracy of the large model. The purpose of monitoring is to ensure that the large model not only performs well on the training set, but more importantly, it maintains high accuracy on unseen validation data, i.e., it has good generalization ability, thereby ensuring the accuracy of data matching.
[0040] After training, set a performance threshold. For example, setting a performance threshold for a large model requires that the mean absolute error (MAE) is less than a*10. -n Pa, and the mean square error MSE is less than k*0 -n Pa², where the mean absolute error is used to control the average error level, and the mean square error is used to strictly limit the occurrence of individual excessive errors. When the predicted vacuum reading calculated by the large model based on the vacuum reading of the ion pump and the mean absolute error M and mean square error of the vacuum gauge reading are stable within the performance threshold, the system can output a vacuum reading that can replace the physical vacuum gauge based on the operating parameters of the ion pump.
[0041] S5: Reduce the number of vacuum gauges required.
[0042] Based on the aforementioned large-scale model, the vacuum system can be hardware reconfigured. In the vacuum piping, the deployment density of ion pumps within each group can be increased, while the number of vacuum gauges installed within that group can be reduced. For example, the original design used one vacuum gauge for every two ion pumps. After optimization, this can be adjusted to one vacuum gauge for every five ion pumps. The vacuum level data at the monitoring points previously handled by the removed vacuum gauges can now be calculated in real-time by the trained large-scale model based on the operating parameters of the newly added ion pumps in that area.
[0043] Since vacuum gauges are high-cost precision sensors in vacuum systems, reducing their number directly reduces the material cost of the vacuum system. Furthermore, fewer vacuum gauges mean fewer installation ports, fewer cabling routes, fewer signal acquisition channels, and less subsequent calibration and maintenance work, improving the overall reliability of the system and reducing the complexity and cost of long-term operation and maintenance.
[0044] S6: After the parameters of the vacuum system are trained, the large model is called to directly output the vacuum parameters and monitor the system's operation.
[0045] S61: Extract the trained large model from the development environment and encapsulate it into a standardized service that can be called by external programs.
[0046] Specifically, a monitoring script needs to be set up first. This script should periodically read the latest ion pump operating data from the vacuum control system's database or data acquisition system. Then, using the local API interfaces provided by frameworks such as Ollama, the large model is deployed as a network service. This service listens for requests from the monitoring script, receives real-time ion pump parameters as input, and returns predicted vacuum readings as output. Furthermore, the monitoring script can display the predicted vacuum readings along with other key system parameters on the monitoring interface and / or store them in a database for historical tracking and analysis.
[0047] S62: Continuously monitor the working status and performance of large models.
[0048] After the large-scale model is put into online operation, observe whether the vacuum degree prediction value output by the large-scale model is smooth and reasonable, and whether there are any abnormal jumps or situations that contradict common sense physics. Secondly, periodically compare the predicted vacuum degree readings output by the large-scale model with the vacuum degree readings of the vacuum gauge, calculate the real-time prediction error, and use this to continuously verify its long-term accuracy.
[0049] S7: Set the data update cycle for automatic data updates.
[0050] First, an empty incremental dataset is set up. In this embodiment, the data update cycle is set to 24 hours. After each data update cycle, the synchronous operation data of all ion pumps and vacuum gauges within the data update cycle are automatically collected and stored in the incremental dataset.
[0051] The system uses incremental datasets to fine-tune the currently deployed large model. Unlike resource-intensive training from scratch, incremental training builds upon the existing knowledge of the large model by making small adjustments to its parameters with new data, allowing it to learn and adapt to the latest state of the system. This approach is computationally efficient and has minimal impact on daily operations.
[0052] After training is complete, the system will automatically evaluate the performance of the new large model on validation data collected concurrently that was not used in the training. Only when the prediction error (such as mean absolute error) of the new large model is consistently lower than or equal to that of the current online large model will the system automatically deploy the new large model online to replace the old version and achieve a transition.
[0053] The system uses incremental datasets to fine-tune the currently deployed large model. Unlike resource-intensive training from scratch, incremental training builds upon the existing knowledge of the large model by making small adjustments to the parameters of the mapping function with new data, allowing it to learn and adapt to the latest state of the system. This approach is computationally efficient and has minimal impact on daily operations.
[0054] After training is complete, the system will automatically evaluate the performance of the new large model on validation data collected concurrently that was not used in the training. Only when the prediction error (such as mean absolute error) of the new large model is consistently lower than or equal to that of the current online large model will the system automatically deploy the new large model online, replacing the old version and achieving a seamless transition.
[0055] The implementation principle of the optimization method for a proton accelerator vacuum system based on a local large-scale model in this application is as follows: A locally deployed large-scale model is used to learn the relationship between the operating parameters of the ion pump and the readings of a high-precision vacuum gauge. When the prediction accuracy of the large-scale model is sufficiently high, the virtual readings of the large-scale model are used to replace part of the physical vacuum gauge, thereby reducing the system hardware cost while ensuring monitoring accuracy.
[0056] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.
Claims
1. An optimization method for a proton accelerator vacuum system based on a local large-scale model, characterized in that: Includes the following steps: Deploy large models on local servers; A vacuum system is established, which includes multiple sets of ion pumps and multiple vacuum gauges, with the vacuum gauges located between each set of ion pumps. After the vacuum system stabilizes, the vacuum parameters of multiple ion pumps and vacuum gauges are collected and saved locally. The large model is trained based on the collected data to establish a mapping relationship between the ion pump operating parameters and the vacuum gauge readings, and a performance threshold is preset. When the error between the ion pump vacuum reading output by the large model and the vacuum gauge reading is less than the preset performance threshold, the number of vacuum gauges in the vacuum system is reduced; if the error is greater than the performance error threshold, the large model is trained again.
2. The optimization method for a proton accelerator vacuum system based on a local large model according to claim 1, characterized in that: The local deployment model specifically includes: Deploy a large model using the Ollam framework on a local server, and enable data acquisition scripts and monitoring systems to communicate with the large model through the local API interface provided by the Ollam framework.
3. The optimization method for a proton accelerator vacuum system based on a local large model according to claim 1, characterized in that: The process of collecting and saving the vacuum parameters of multiple ion pumps and vacuum gauges locally includes the following steps: The collected data is cleaned to remove outlier data points; The cleaned data is structured and stored locally. The data includes date, time, vacuum gauge number, vacuum gauge reading, ion pump number, and ion pump operating data.
4. The optimization method for a proton accelerator vacuum system based on a local large model according to claim 1, characterized in that: The large model is trained based on the collected vacuum parameters of the ion pump and the vacuum gauge, including the following steps: The structured collected data was randomly divided into training set, validation set and test set according to a certain proportion; A mapping function F is set within the large model, and the predicted vacuum reading output by the large model is equal to F (the vacuum reading of the ion pump), so that the predicted vacuum reading is approximately equal to the vacuum reading of the vacuum gauge. The internal parameters of the large model are adjusted using data from the training set, and the difference between the self-reading vacuum degree of the ion pump and the actual reading of the vacuum gauge is quantified by the mean square error (MSE) loss function. During training, the mean absolute error and root mean square error are monitored simultaneously on the validation set.
5. The optimization method for a proton accelerator vacuum system based on a local large model according to claim 4, characterized in that: This also includes packaging the trained large model into a standardized service, which allows scripts to call the large model to directly output vacuum parameters. Specifically, this involves the following steps: Set up a monitoring script to periodically read the ion pump's operating data; The trained large model is deployed as a network service via an API interface. The network service listens for requests from monitoring scripts, receives the operating data of the ion pump as input, and returns the predicted vacuum reading as output.
6. The optimization method for a proton accelerator vacuum system based on a local large model according to claim 5, characterized in that: The monitoring script will simultaneously display the predicted vacuum reading, the operating data of the ion pump, and the vacuum reading of the vacuum pump on the monitoring interface, and / or store them in the database.
7. The optimization method for a proton accelerator vacuum system based on a local large model according to claim 1, characterized in that: It also includes setting a data update cycle and using the incremental data collected within the cycle to fine-tune the deployed large model.
8. The optimization method for a proton accelerator vacuum system based on a local large model according to claim 7, characterized in that: The step of setting a data update period and fine-tuning the deployed large model using incremental data collected within the period specifically includes: Set up an empty incremental dataset; After each data update cycle ends, the synchronous operation data of all ion pumps and vacuum gauges within the data update cycle are automatically collected and stored in the incremental dataset; The currently deployed large model is fine-tuned based on the incremental dataset to generate a new candidate large model to be evaluated. The performance of the new large model is evaluated on validation data collected at the same time that were not used for training. When the error between the predicted vacuum reading output by the new large model and the vacuum reading of the vacuum gauge is less than the performance threshold, the new large model is automatically deployed online.