Method for accurately controlling influence of ignition of extraction system on proton treatment process

Abstract

The invention discloses a method for accurately controlling a proton treatment process. The method comprises the following steps: selecting a feed-in voltage fluctuation amplitude greater than or equal to U0 as a later fault prediction standard; data acquisition of leakage current, cavity vacuum degree and feed-in voltage value is carried out, and a time sequence sample database is established; constructing and training a time sequence prediction model based on LSTM; collecting the leakage current and vacuum degree data of the electrostatic deflection plate at the current t-th time point in real time; using the data collected at the current t-th time point as the input of a prediction model, and predicting the fluctuation amplitude of the feed-in voltage at the (t + 1)-th time point through the prediction model; after it is predicted that a fault is about to occur, a PLC sends an instruction to control a central area vertical deflection plate to turn off a beam; after it is predicted that the fault is eliminated, a reset signal is sent; according to the method, through the combination of two predictions, the factor that the beam current intensity is unstable in the treatment process is eliminated, treatment is immediately recovered after the factor that the beam current intensity is unstable is eliminated, and the precise control of the proton treatment process is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of accelerators, and in particular relates to a method for accurately controlling the influence of ignition of an extraction system on a proton therapy process. Background Art

[0002] With the advancement of technology, the development of economic foundation and the improvement of people's living standards, the demand for proton therapy is becoming more and more widespread. Internationally, the conventional proton therapy equipment requires superconducting cyclotron accelerator to produce protons. As the core component of the proton therapy device, the accelerator integrates multiple systems, including main magnet, high frequency, vacuum, ion source, beam extraction, beam line, treatment device, beam monitoring and control system.

[0003] The movement of proton beam in the cyclotron can be divided into three stages: injection, acceleration, and extraction. The electrostatic deflection plate is a key component of the cyclotron extraction system. It uses the high voltage generated by the high-voltage power supply to form a strong electric field, which causes the beam trajectory to shift, and then extracts the beam to the beam line outside the accelerator, and then transports the beam to the treatment head through the beam line.

[0004] The electrostatic deflection plate is composed of a cutting plate and a negative high-voltage electrode, and a very uniform electric field is generated between the two. When the electrostatic deflection plate is affected by certain environmental factors and its own physical properties, high-voltage breakdown spark discharge will occur. When high-voltage spark discharge occurs in the gap between the deflection plates, the leakage current increases, the stability of the cavity vacuum decreases, and the voltage drop in the circuit increases, thereby affecting the internal electric field of the electrostatic deflection plate, causing the flow intensity of the extraction beam from the electrostatic deflection plate to the treatment head to fluctuate, affecting the precision of the proton therapy process.

[0005] The difficulty in achieving precision in the proton therapy process lies in the contradiction between the instability of the beam intensity and the refinement of proton therapy. On the one hand, the instability of the beam intensity is unavoidable because environmental factors are irresistible. During the treatment process of more than ten hours a day, high-voltage breakdown and sparking are inevitable due to environmental factors or sudden changes in the physical properties of the device. On the other hand, during proton therapy, the beam intensity and beam-carrying time at each treatment position need to be precisely controlled. The instability of the beam intensity will lead to poor irradiation of some lesions, thus affecting the overall treatment effect. Summary of the invention

[0006] In view of the problems existing in the prior art, the present invention proposes a method for accurately controlling the influence of ignition of the extraction system on the proton therapy process, aiming to solve the problem of precision of the proton therapy process.

[0007] The present invention proposes the following technical solutions to solve the technical problems:

[0008] A method for accurately controlling the effect of sparking of an extraction system on a proton therapy process, characterized by comprising the following steps:

[0009] Step 1: Select the feed voltage fluctuation amplitude greater than or equal to U0 as the later fault prediction standard;

[0010] Step 2: The electrostatic deflection plate enters a voltage stabilizing working state;

[0011] Step 3: Collect data on leakage current, cavity vacuum and input voltage of the electrostatic deflection plate in the voltage-stabilized working state, and establish a time series sample database; the data collection period of the time series sample database is T ≥ the prediction model calculation time t 预测 + The time t during which the PLC issues a "pause" command when a fault occurs PLC + The maximum value of two times max; the two times are: the high voltage power supply voltage adjustment time t of the central area switching current 电源 and the treatment head response instruction time t 治疗头 , and the data sampling period T≤single ignition fault duration;

[0012] Step 4: construct a time series prediction model based on LSTM, input the sample database into the model, and obtain the input-output model of the ignition fault prediction of the high-voltage extraction electrode of the cyclotron through model training, model evaluation, and model optimization. The input is the input data collected at the current t-th time point, and the output is the voltage fluctuation amplitude predicted by the model at the t+1-th time point. The actual time is the data sampling period T*t, and t at the t-th time point is the time series number.

[0013] Step 5: real-time acquisition of the electrostatic deflection plate leakage current and vacuum degree data at the current time point t;

[0014] Step 6: Use the data collected at the current time point t as the input of the prediction model, and predict the fluctuation amplitude of the feed-in voltage at the time point t+1 through the prediction model;

[0015] Step 7: Whether the fluctuation amplitude of the input voltage at the time point t+1 is greater than U0, if not, return to step 6, if yes, continue to step 8;

[0016] Step 8: After detecting that a fault is about to occur, a fault signal is issued, and a command is issued through the PLC to control the vertical deflection plate in the central area to turn off the beam, control the treatment head at the treatment end to pause, and record the beam pause time T*△t, where △t is the sum of the total time points from the pause to the resumption of treatment, and the initial value is 0;

[0017] Step 9: At the time point △t=△t+1, determine whether the fluctuation amplitude of the input voltage at the time point t+1 is greater than U0. If yes, loop through step 9; if no, proceed to step 10.

[0018] Step 10: After detecting that the fault is eliminated, a reset signal is issued, and a command is issued through the PLC to control the vertical deflection plate in the central area to start the beam and control the treatment head at the treatment end to resume work. At the same time, the pause time T*△t is fed back to the treatment head program, and the pause time T*△t is the time from the occurrence of the fault to the elimination of the fault; the beam current intensity in the working program of the treatment head changes from f(T*t) to f(T*tT*△t); wherein T*t is the current time;

[0019] Furthermore, the specific principles of the feed voltage fluctuation amplitude selected in step 1 are as follows:

[0020] 1) When the fluctuation amplitude of the input voltage is greater than U1, the beam load fluctuation at the treatment end reaches about 10%;

[0021] 2) When the leakage current is greater than a certain value and the corresponding input voltage fluctuation amplitude is greater than U2, the leakage current arcing will cause damage to the high-voltage feed ball head of the electrostatic deflection plate;

[0022] 3) Feed voltage fluctuation amplitude U0 = min(U1, U2).

[0023] Furthermore, the specific process of establishing a time series sample database in step three is as follows:

[0024] 1) Use a high-speed data acquisition card to collect the leakage current i, cavity vacuum degree p and input voltage value U under the voltage-stabilized working state as the original data sequence;

[0025] 2) Performing difference processing on the data cavity vacuum degree p and the input voltage value U to obtain a time series database, which includes the time sequence number t, the leakage current i, the cavity vacuum degree change △p and the input voltage fluctuation amplitude △U;

[0026] 3) The data is divided into a training set and a test set at a ratio of 1:1 between the first half of the time series and the second half of the time series, for training and testing the prediction model;

[0027] 4) Taking the leakage current i and the cavity vacuum degree change △p of the training set as input variables, and the input voltage fluctuation amplitude ΔU as the output variable, a time series sample database is formed;

[0028] 5) Normalize all features of the training set and test set;

[0029] 6) Convert the data of the training set into a supervised learning problem. By converting it into a supervised learning problem, the output variable at the t+1th time point can be predicted by the input variable at the tth time point.

[0030] Supplementary Note 1:

[0031] Each time point corresponds to an actual time. For example, if the time point t is 1, the time corresponding to the tth time point = T*t, T = 150ms. If the time point = t+1, the time corresponding to the t+1 time point = T*2 = 300ms.

[0032] Furthermore, the construction of the time series prediction model based on LSTM in step 4 is as follows:

[0033] 1) Build an LSTM model. The model is divided into an input layer, a hidden layer and an output layer. There are 2 neurons in the input layer, 50 neurons are initially selected in the hidden layer, and 1 neuron in the output layer.

[0034] 2) Input the training set data of the time series sample database into the LSTM model for training, and obtain a training set data model that predicts the output data at the next moment based on the input data at the previous moment;

[0035] 3) Use the test set data of the time series sample database to test the training set data model training results, and evaluate the accuracy of the training set data model through testing.

[0036] Furthermore, the accuracy of the training set data model in step 4 (process 3) is evaluated, and the specific steps are as follows:

[0037] 1) Use the prediction results and test set data to calculate the error loss;

[0038] 2) When the error loss exceeds the threshold, the training set data model is adjusted and optimized.

[0039] The adjustment and optimization of the training set data model is specifically as follows: the model is optimized by adjusting the number of neurons included in the hidden layer to reduce the prediction error loss, and a trained input-output model for predicting the ignition fault of the high-voltage lead-out electrode of the cyclotron is obtained. The trained input-output model for predicting the ignition fault of the high-voltage lead-out electrode of the cyclotron can predict the output variable feed voltage fluctuation amplitude △U at the t+1th time point by inputting the input variables leakage current i and cavity vacuum degree change △p collected at the current tth time point.

[0040] Furthermore, the specific process of step eight is as follows:

[0041] 1) After detecting that a fault is about to occur, the voltage difference of the vertical deflection plate in the central area is increased through PLC control, forming an electric field to shut down the beam in the central area, and at the same time, the PLC controls the proton accelerator treatment head end to suspend operation;

[0042] 2) Record the pause time T*△t (the initial value of △t is 0) from the occurrence of the fault to the elimination of the fault.

[0043] Furthermore, the specific process of step ten is as follows:

[0044] 1) After the fault is detected and eliminated, the voltage difference of the vertical deflection plate in the central area is controlled to drop to 0 after a set time delay through PLC, so that the beam in the central area stops being turned off and turned on again;

[0045] 2) At the same time, the PLC controls the start of the proton accelerator treatment head and feeds back the pause time T*△t to the treatment head program. The beam current intensity in the treatment head working program changes from f(T*t) to f(T*tT*△t), that is, the treatment work resumes normal when the fault is eliminated under the voltage stabilization working state of the electrostatic deflection plate; the T*t is the current time; the pause time T*△t is the time from the occurrence of the fault to the elimination of the fault.

[0046] Furthermore, the delay of a set time is the data collection period T of the time series sample database minus the prediction model calculation time t 预测 -When a fault occurs, the PLC issues a "pause" command for a certain time t PLC -The maximum value of the two times max.

[0047] Advantages and effects of the present invention

[0048] 1. The present invention combines "two predictions" and uses a time series model to predict that a fault will occur at time point t+1, and suspends treatment at time point t; uses a time series model to predict that the fault has been eliminated at time point t+1, and resumes treatment at time point t+1, thereby eliminating factors of unstable beam current intensity during the treatment process, and immediately resuming treatment after eliminating the factors of unstable beam current intensity, thereby achieving precise control of the proton therapy process.

[0049] 2. The present invention accurately designs the data collection cycle T of the sample database so that before a fault occurs, the PLC has enough time to issue a "pause" signal, the high-voltage power supply in the central area has enough time to shut off the current, and the treatment head has enough time to respond to the PLC's command to stop treatment; and after the fault is eliminated, the PLC has enough time to issue a "resume treatment" signal, the high-voltage power supply in the central area has enough time to turn on the current, and the treatment head has enough time to respond to the PLC's command to continue treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of a method for accurately controlling a proton therapy process according to the present invention.

[0051] Figure 2 This is a schematic diagram of the application of the time series number variation of the present invention. DETAILED DESCRIPTION

[0052] Design principle of the present invention

[0053] 1. Innovation of the present invention: The innovation lies in: combining the time series prediction model with refined proton therapy. Conventional methods only use models to predict the occurrence of faults, but merely predicting the occurrence of faults cannot be used for refined proton therapy. Because when the model predicts the occurrence of a fault at the t+1th time point at the tth time point, it is also the time to stop the treatment. If the treatment is stopped midway, the entire treatment process cannot be completed, which not only fails to achieve the purpose of refined treatment but also terminates the treatment. The present invention not only uses a time series model to predict the occurrence of a fault at the t+1th time point, but also predicts the elimination of the fault at the t+1th time point. When it is known at the tth time point that a fault will occur at the t+1th time point, the PLC controller shuts off the treatment of the treatment head at the t+1th time point. When it is predicted that the fault will be eliminated at the t+1th time point, the PLC controller completes the restoration of the treatment head at the t+1th time point. Predicting the occurrence of a fault at the t+1 time point and predicting the elimination of the fault at the t+1 time point are born together. The prediction of the occurrence of a fault can be preceded by the prediction of the elimination of the fault. This time sequence cannot be reversed, otherwise, the prediction of the elimination of the fault at the t+1 time point is inaccurate. In the standard definition of the fault, the feed voltage fluctuation amplitude greater than or equal to U0 is selected as the fault prediction standard. The selection of the standard depends on the need for the treatment effect of the accelerator treatment end and the need to protect the high-voltage feed ball head of the electrostatic deflection plate.

[0054] 2. Several key points of the present invention

[0055] One of the key points: the model should not only predict when a failure occurs, but also when the failure is eliminated;

[0056] Key point 2: Troubleshooting should be done “in advance” and recovery treatment should be done “immediately”. Troubleshooting in advance means troubleshooting at time point t rather than at time point t+1. Because the fault has already occurred at time point t+1, it is no longer helpful to eliminate it. Recovery treatment should be done “immediately”, which means resuming treatment at time point t+1 rather than at time point t, because the fault has not been completely eliminated at time point t, and treatment cannot be resumed before the fault is completely eliminated.

[0057] Key point three: A. Selection of model training input data: The present invention uses leakage current and vacuum degree as input data for model training. The principle is: although leakage current will be generated when sparking occurs, the measurement of leakage current is to measure a certain point, which may have the one-sidedness of local measurement, so the measurement of vacuum degree is added. When sparking occurs, the vacuum degree inside the entire vacuum chamber will change, but there are multiple reasons for the change in vacuum degree, not just the sparking that causes the vacuum degree to change, so the leakage current and vacuum degree should be measured together. B. Selection of model training output data: The present invention uses the change in feed voltage as the output data for model training. The changes in leakage current and vacuum degree are the causes of voltage changes, and the voltage changes are the results of the changes in leakage current and vacuum degree, so the change in feed voltage is used as the output data for model training.

[0058] The fourth key point is the establishment of the time series database. The time series database data is not the original collected data, but a part of the data is sorted out as the time series database data, that is, the difference data. For example, the leakage current i of the time series database is the directly collected data, while the cavity vacuum degree p and the input voltage value U are the difference data respectively;

[0059] The fifth key point is: the determination of the data collection period T. The data collection period T cannot be too large or too small. It must be combined with the calculation time of the model and the response time of multiple hardware in the proton therapy system. Specifically: the prediction model calculation time, the time when the PLC sends a "pause" command when a fault occurs, the high-voltage power supply voltage adjustment time of the central area switching current, the treatment head response command time, and the duration of a single ignition fault. Among these multiple times, it is difficult to determine which time to use as the data collection time, and it must be considered comprehensively. The data collection period T of the time series sample database of the present invention ≥ the prediction model calculation time t 预测 + The time t during which the PLC issues a "pause" command when a fault occurs PLC + The maximum value of two times max; the two times are: the high voltage power supply voltage adjustment time t of the central area switching current 电源 and the treatment head response instruction time t 治疗头 , and the data sampling period T≤single ignition fault duration;

[0060] Based on the above principles, the present invention relates to a method for accurately controlling the effect of the ignition of the extraction system on the proton therapy process, such as Figure 1 , Figure 2 As shown, its characteristics are: comprising the following steps:

[0061] Step 1: Select the feed voltage fluctuation amplitude greater than or equal to U0 as the later fault prediction standard;

[0062] Step 2: The electrostatic deflection plate enters a voltage stabilizing working state;

[0063] Step 3: Collect data on leakage current, cavity vacuum and input voltage of the electrostatic deflection plate in the voltage-stabilized working state, and establish a time series sample database; the data collection period of the time series sample database is T ≥ the prediction model calculation time t 预测 + The time t during which the PLC issues a "pause" command when a fault occurs PLC + The maximum value of two times max; the two times are: the high voltage power supply voltage adjustment time t of the central area switching current 电源 and the treatment head response instruction time t 治疗头 , and the data sampling period T≤single ignition fault duration;

[0064] Step 4: construct a time series prediction model based on LSTM, input the sample database into the model, and obtain the input-output model of the ignition fault prediction of the high-voltage extraction electrode of the cyclotron through model training, model evaluation, and model optimization. The input is the input data collected at the current t-th time point, and the output is the voltage fluctuation amplitude predicted by the model at the t+1-th time point. The actual time is the data sampling period T*t, and t at the t-th time point is the time series number.

[0065] Step 5: real-time acquisition of the electrostatic deflection plate leakage current and vacuum degree data at the current time point t;

[0066] Step 6: Use the data collected at the current time point t as the input of the prediction model, and predict the fluctuation amplitude of the feed-in voltage at the time point t+1 through the prediction model;

[0067] Step 7: Whether the fluctuation amplitude of the input voltage at the time point t+1 is greater than U0, if not, return to step 6, if yes, continue to step 8;

[0068] Step 8: After detecting that a fault is about to occur, a fault signal is issued, and a command is issued through the PLC to control the vertical deflection plate in the central area to turn off the beam, control the treatment head at the treatment end to pause, and record the beam pause time T*△t, where △t is the sum of the total time points from the pause to the resumption of treatment, and the initial value is 0;

[0069] Step 9: At the time point △t=△t+1, determine whether the fluctuation amplitude of the input voltage at the time point t+1 is greater than U0. If yes, loop through step 9; if no, proceed to step 10.

[0070] Step 10: After detecting that the fault is eliminated, a reset signal is issued, and a command is issued through the PLC to control the vertical deflection plate in the central area to start the beam and control the treatment head at the treatment end to resume work. At the same time, the pause time T*△t is fed back to the treatment head program, and the pause time T*△t is the time from the occurrence of the fault to the elimination of the fault; the beam current intensity in the working program of the treatment head changes from f(T*t) to f(T*tT*△t); wherein T*t is the current time;

[0071] Furthermore, the specific principles of the feed voltage fluctuation amplitude selected in step 1 are as follows:

[0072] 1) When the fluctuation amplitude of the input voltage is greater than U1, the beam load fluctuation at the treatment end reaches about 10%;

[0073] 2) When the leakage current is greater than a certain value and the corresponding input voltage fluctuation amplitude is greater than U2, the leakage current arcing will cause damage to the high-voltage feed ball head of the electrostatic deflection plate;

[0074] 3) Feed voltage fluctuation amplitude U0 = min(U1, U2).

[0075] Furthermore, the specific process of establishing a time series sample database in step three is as follows:

[0076] 1) Use a high-speed data acquisition card to collect the leakage current i, cavity vacuum degree p and input voltage value U under the voltage-stabilized working state as the original data sequence;

[0077] 2) Performing difference processing on the data cavity vacuum degree p and the input voltage value U to obtain a time series database, which includes the time sequence number t, the leakage current i, the cavity vacuum degree change △p and the input voltage fluctuation amplitude △U;

[0078] 3) The data is divided into a training set and a test set at a ratio of 1:1 between the first half of the time series and the second half of the time series, for training and testing the prediction model;

[0079] 4) Taking the leakage current i and the cavity vacuum degree change △p of the training set as input variables, and the input voltage fluctuation amplitude ΔU as the output variable, a time series sample database is formed;

[0080] 5) Normalize all features of the training set and test set;

[0081] 6) Convert the data of the training set into a supervised learning problem. By converting it into a supervised learning problem, the output variable at the t+1th time point can be predicted by the input variable at the tth time point.

[0082] Supplementary Note 1:

[0083] Each time point corresponds to an actual time. For example, if the time point t is 1, the time corresponding to the tth time point = T*t, T = 150ms. If the time point = t+1, the time corresponding to the t+1 time point = T*2 = 300ms.

[0084] Furthermore, the construction of the time series prediction model based on LSTM in step 4 is as follows:

[0085] 1) Build an LSTM model. The model is divided into an input layer, a hidden layer and an output layer. There are 2 neurons in the input layer, 50 neurons are initially selected in the hidden layer, and 1 neuron in the output layer.

[0086] 2) Input the training set data of the time series sample database into the LSTM model for training, and obtain a training set data model that predicts the output data at the next moment based on the input data at the previous moment;

[0087] 3) Use the test set data of the time series sample database to test the training set data model training results, and evaluate the accuracy of the training set data model through testing.

[0088] Furthermore, the accuracy of the training set data model in step 4 (process 3) is evaluated, and the specific steps are as follows:

[0089] 1) Use the prediction results and test set data to calculate the error loss;

[0090] 2) When the error loss exceeds the threshold, the training set data model is adjusted and optimized.

[0091] The adjustment and optimization of the training set data model is specifically as follows: the model is optimized by adjusting the number of neurons included in the hidden layer to reduce the prediction error loss, and a trained input-output model for predicting the ignition fault of the high-voltage lead-out electrode of the cyclotron is obtained. The trained input-output model for predicting the ignition fault of the high-voltage lead-out electrode of the cyclotron can predict the output variable feed voltage fluctuation amplitude △U at the t+1th time point by inputting the input variables leakage current i and cavity vacuum degree change △p collected at the current tth time point.

[0092] Furthermore, the specific process of step eight is as follows:

[0093] 1) After detecting that a fault is about to occur, the voltage difference of the vertical deflection plate in the central area is increased through PLC control, forming an electric field to shut down the beam in the central area, and at the same time, the PLC controls the proton accelerator treatment head end to suspend operation;

[0094] 2) Record the pause time T*△t (△t is the initial value of 0) from the occurrence of the fault to the elimination of the fault.

[0095] Furthermore, the specific process of step ten is as follows:

[0096] 1) After the fault is detected and eliminated, the voltage difference of the vertical deflection plate in the central area is controlled to drop to 0 after a set time delay through PLC, so that the beam in the central area stops shutting down and restarting;

[0097] 2) At the same time, the PLC controls the start of the proton accelerator treatment head and feeds back the pause time T*△t to the treatment head program. The beam current intensity in the treatment head working program changes from f(T*t) to f(T*tT*△t), that is, the treatment work resumes normal when the fault is eliminated under the voltage stabilization working state of the electrostatic deflection plate; the T*t is the current time; the pause time T*△t is the time from the occurrence of the fault to the elimination of the fault.

[0098] Furthermore, the delay of a set time is the data collection period T of the time series sample database minus the prediction model calculation time t 预测 -When a fault occurs, the PLC issues a "pause" command for a certain time t PLC -The maximum value of the two times max.

[0099] Example 1: Calculation of sampling period and delay time

[0100] 1) Calculation of sampling period: Step 1: Sum prediction model calculation time + PLC control reaction time + high-voltage power supply voltage adjustment time of the central area opening and closing current: If the prediction model calculation time is 100ms, the PLC control reaction time is 10ms, and the high-voltage power supply voltage adjustment time of the central area opening and closing current tpower supply is 20ms, then the prediction model calculation time + PLC control reaction time + high-voltage power supply voltage adjustment time of the central area opening and closing current = 130mm; Step 2: Determine the sampling period: Since the single ignition failure time is usually greater than 200ms, the sampling period is taken between 130mm and 200ms. In this embodiment, the sampling period is taken as 150mm≤single ignition failure duration 200ms; Since the high-voltage power supply voltage adjustment time t of the central area opening and closing current 电源 and the treatment head response instruction time t 治疗头 Compared with the time, the high voltage power supply voltage regulation time t of the central area switching current 电源 Therefore, after comparing the two, the high voltage power supply voltage adjustment time of the central area switching current is 20ms.

[0101] The reason why the high voltage power supply voltage regulation time of the central area switching current is selected as 20ms is that it is consistent with the treatment head response instruction time t 治疗头 Compared with the time, the high voltage power supply voltage adjustment time of the central area switching current is 20ms, and the treatment head response command time t治疗头 The time is less than the high voltage power supply voltage regulation time of the center area switching current.

[0102] 2) Calculation of delay time: If the sampling period is 150ms, the prediction model calculation time is 100ms, the PLC control response time is 10ms, and the high-voltage power supply voltage adjustment time of the central area switching current is t 电源 20ms, when the single ignition failure time is usually greater than 200ms, the delay is a set time = sampling period 150ms-prediction model calculation time 100ms-PLC control response time 10ms-high-voltage power supply voltage adjustment time of the central area switching current 20ms=20ms.

[0103] It should be emphasized that the above specific embodiments are merely explanations of the present invention, and they are not limitations of the present invention. After reading this specification, those skilled in the art can make modifications to the above embodiments without any creative contribution as needed, but as long as they are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A method for accurately controlling the effect of ignition of an extraction system on a proton therapy process, characterized in that: The following steps are involved: Step 1: Select the feed voltage fluctuation amplitude greater than or equal to U0 as the later fault prediction standard; Step 2: The electrostatic deflection plate enters a voltage stabilizing working state; Step 3: Collect data on leakage current, cavity vacuum and input voltage of the electrostatic deflection plate in the voltage-stabilized working state, and establish a time series sample database; the data collection period of the time series sample database is T ≥ the prediction model calculation time t 预测 + The time t during which the PLC issues a "pause" command when a fault occurs PLC + The maximum value of two times max; the two times are: the high voltage power supply voltage adjustment time t of the central area switching current 电源 and the treatment head response instruction time t 治疗头 , and the data sampling period T≤single ignition fault duration; Step 4: construct a time series prediction model based on LSTM, input the sample database into the model, and obtain the input-output model of the ignition fault prediction of the high-voltage extraction electrode of the cyclotron through model training, model evaluation, and model optimization. The input is the input data collected at the current t-th time point, and the output is the voltage fluctuation amplitude predicted by the model at the t+1-th time point. The actual time is the data sampling period T*t, and t at the t-th time point is the time series number. Step 5: real-time acquisition of the electrostatic deflection plate leakage current and vacuum degree data at the current time point t; Step 6: Use the data collected at the current time point t as the input of the prediction model, and predict the fluctuation amplitude of the feed-in voltage at the time point t+1 through the prediction model; Step 7: Whether the fluctuation amplitude of the input voltage at the time point t+1 is greater than U0, if not, return to step 6, if yes, continue to step 8; Step 8: After detecting that a fault is about to occur, a fault signal is issued, and a command is issued through the PLC to control the vertical deflection plate in the central area to turn off the beam, control the treatment head at the treatment end to pause, and record the beam pause time T*△t, where △t is the sum of the total time points from the pause to the resumption of treatment, and the initial value is 0; Step 9: At the time point △t=△t+1, determine whether the fluctuation amplitude of the input voltage at the time point t+1 is greater than U0. If yes, loop through step 9; if no, proceed to step 10. Step 10. After detecting that the fault has been eliminated, a reset signal is issued, and a command is issued through the PLC to control the vertical deflection plate in the central area to start the beam and control the treatment head at the treatment end to resume work. At the same time, the pause time T*△t is fed back to the treatment head program, and the pause time T*△t is the time from the occurrence of the fault to the elimination of the fault; the beam current intensity in the working program of the treatment head changes from f(T*t) to f(T*tT*△t); wherein T*t is the current time.

2. According to claim 1, a method for accurately controlling the effect of ignition of the extraction system on the proton therapy process, characterized in that: The specific principles of the feed voltage fluctuation amplitude selected in step 1 are as follows: 1) When the fluctuation amplitude of the input voltage is greater than U1, the beam load fluctuation at the treatment end reaches about 10%; 2) When the leakage current is greater than a certain value and the corresponding input voltage fluctuation amplitude is greater than U2, the leakage current arcing will cause damage to the high-voltage feed ball head of the electrostatic deflection plate; 3) Feed voltage fluctuation amplitude U0 = min(U1, U2).

3. According to claim 1, a method for accurately controlling the effect of ignition of the extraction system on the proton therapy process is characterized by: The specific process of establishing a time series sample database in step 3 is as follows: 1) Use a high-speed data acquisition card to collect the leakage current i, cavity vacuum degree p and input voltage value U under the voltage-stabilized working state as the original data sequence; 2) Performing difference processing on the data cavity vacuum degree p and the input voltage value U to obtain a time series database, which includes the time sequence number t, the leakage current i, the cavity vacuum degree change △p and the input voltage fluctuation amplitude △U; 3) The data is divided into a training set and a test set at a ratio of 1:1 between the first half of the time series and the second half of the time series, for training and testing the prediction model; 4) Taking the leakage current i and the cavity vacuum degree change △p of the training set as input variables, and the input voltage fluctuation amplitude ΔU as the output variable, a time series sample database is formed; 5) Normalize all features of the training set and test set; 6) Convert the data of the training set into a supervised learning problem. By converting it into a supervised learning problem, the output variable at the t+1th time point can be predicted by the input variable at the tth time point.

4. According to claim 1, a method for accurately controlling the effect of ignition of the extraction system on the proton therapy process, characterized in that: The construction of the time series prediction model based on LSTM in step 4 is as follows: 1) Build an LSTM model. The model is divided into an input layer, a hidden layer and an output layer. There are 2 neurons in the input layer, 50 neurons are initially selected in the hidden layer, and 1 neuron in the output layer. 2) Input the training set data of the time series sample database into the LSTM model for training, and obtain a training set data model that predicts the output data at the next moment based on the input data at the previous moment; 3) Use the test set data of the time series sample database to test the training set data model training results, and evaluate the accuracy of the training set data model through testing.

5. According to claim 4, a method for accurately controlling the effect of ignition of the extraction system on the proton therapy process is characterized by: The accuracy of the training set data model in step 4 (3) is evaluated, and the specific steps are as follows: 1) Use the prediction results and test set data to calculate the error loss; 2) When the error loss exceeds the threshold, the training set data model is adjusted and optimized.

6. According to claim 5, a method for accurately controlling the effect of ignition of the extraction system on the proton therapy process is characterized by: The adjustment and optimization of the training set data model is specifically as follows: the model is optimized by adjusting the number of neurons included in the hidden layer to reduce the prediction error loss, and a trained input-output model for predicting the ignition fault of the high-voltage lead-out electrode of the cyclotron is obtained. The trained input-output model for predicting the ignition fault of the high-voltage lead-out electrode of the cyclotron can predict the output variable feed voltage fluctuation amplitude △U at the t+1th time point by inputting the input variables leakage current i and cavity vacuum degree change △p collected at the current tth time point.

7. According to claim 1, a method for accurately controlling the effect of ignition of the extraction system on the proton therapy process, characterized in that: The specific process of step eight is as follows: 1) After detecting that a fault is about to occur, the voltage difference of the vertical deflection plate in the central area is increased through PLC control, forming an electric field to shut down the beam in the central area, and at the same time, the PLC controls the proton accelerator treatment head end to suspend operation; 2) Record the pause time T*△t from the occurrence of the fault to the elimination of the fault, and the initial value of △t is 0.

8. According to claim 1, a method for accurately controlling the effect of ignition of the extraction system on the proton therapy process, characterized in that: The specific process of step ten is as follows: 1) After the fault is detected and eliminated, the voltage difference of the vertical deflection plate in the central area is controlled to drop to 0 after a set time delay through PLC, so that the beam in the central area stops being turned off and turned on again; 2) At the same time, the PLC controls the start of the proton accelerator treatment head and feeds back the pause time T*△t to the treatment head program. The beam current intensity in the treatment head working program changes from f(T*t) to f(T*tT*△t), that is, the treatment work resumes normal when the fault is eliminated under the voltage stabilization working state of the electrostatic deflection plate; the T*t is the current time; the pause time T*△t is the time from the occurrence of the fault to the elimination of the fault.

9. A method for accurately controlling the effect of ignition of the extraction system on the proton therapy process according to claim 8, characterized in that: The delay of a set time is the data collection period T of the time series sample database minus the prediction model calculation time t 预测 -When a fault occurs, the PLC issues a "pause" command for a certain time t PLC -The maximum value of the two times max.

Images