Accelerator beam orbit control method and system based on neural network inverse model

Abstract

The application relates to the technical field of orbit control, and discloses an accelerator beam orbit control method and system based on a neural network inverse model, which comprises the following steps: acquiring beam orbit point position data output by a beam position probe and control signal data of a correction magnet power supply in real time during the operation of an SSMB storage ring; analyzing the beam orbit point position data and the control signal data according to an inverse mapping neural network model to generate a mapping drift judgment mark used for judging whether a mapping relationship between the beam orbit and the control signal has a drift trend; and the scheme enables the inverse mapping neural network model to track dynamic changes in the mapping relationship caused by the magnet excitation strength drift in the SSMB storage ring and the cumulative installation deviation of the equipment alignment in real time, dynamically decides the parameter updating time and the weight adjustment direction of the inverse mapping neural network model, and solves the problem that a traditional inverse model cannot track the mapping drift.

Description

Technical Field

[0001] This invention relates to the field of orbit control technology, and more specifically to an accelerator beam orbit control method and system based on a neural network inverse model. Background Technology

[0002] Currently, accelerator beam trajectory control based on neural network inverse models often uses a single hidden layer or a simple multi-level BP neural network to construct the inverse model. The beam trajectory position data collected by the beam position detector is used as input, and the control signal of the corrector power supply is used as output. The neural network inverse model is trained offline to fit the mapping relationship between the beam trajectory and the control signal, and the corresponding correction signal is output through the inverse model to realize the closed-loop control of the beam trajectory.

[0003] However, when the above control method is applied to the storage ring, the following defects still exist: The SSMB storage ring has extremely high requirements for the stability of the beam track and must meet the accuracy index of the residual closed track, which is usually below the micrometer level. Moreover, during the long-term operation of the storage ring, the drift of its magnet excitation intensity and the deviation of equipment collimation and installation will accumulate over time, causing the mapping relationship between the beam track and the control signal to change dynamically. The existing fixed-structure neural network inverse model lacks the ability to analyze the trend of mapping drift. It cannot dynamically analyze whether the model parameters need to be updated and the direction of weight adjustment based on the real-time deviation of the beam track and the mapping drift state. As a result, the inverse model cannot track the drift of the mapping relationship in real time, which reduces the stability control effect of the beam track. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an accelerator beam trajectory control method and system based on a neural network inverse model, which solves the aforementioned problems.

[0005] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0006] Accelerator beam trajectory control methods based on neural network inverse models include:

[0007] Step S1: Real-time acquisition of beam trajectory position data and control signal data of correction magnet power supply output by beam position detector during SSMB storage ring operation; analysis of beam trajectory position data and control signal data based on inverse mapping neural network model to generate mapping drift judgment label for judging whether there is a drift trend in the mapping relationship between beam trajectory and control signal.

[0008] Step S2: Based on the mapping drift judgment identifier, analyze the hierarchical response of the inverse mapping neural network model, and generate the model trigger threshold and the model hierarchy weights representing the changes in the mapping relationship between the beam trajectory and the control signal, respectively.

[0009] Step S3: Based on the beam trajectory point data and control signal data, analyze the deviation of the current beam trajectory and generate a deviation judgment value that represents the degree of deviation of the actual beam trajectory.

[0010] Step S4: Analyze the update status of the inverse mapping neural network model based on the model trigger threshold, mapping drift judgment flag, and deviation judgment value, and generate update determination coefficients to determine whether the parameters of the inverse mapping neural network model need to be updated;

[0011] Step S5: Based on the updated determination coefficients, model level weights, mapping drift judgment indicators, and deviation judgment values, perform collaborative calculations to analyze the adjustment direction of the inverse mapping neural network model and generate control commands.

[0012] Furthermore, based on the inverse mapping neural network model, the beam trajectory point data and control signal data are analyzed to generate a mapping drift judgment indicator to determine whether there is a drift trend in the mapping relationship between the beam trajectory and the control signal, including:

[0013] For the beam trajectory point data and control signal data, the time lag distribution is analyzed based on the inverse mapping neural network model to obtain the dynamic hysteresis disturbance representing the control response delay.

[0014] Furthermore, based on the inverse mapping neural network model, the beam trajectory point data and control signal data are analyzed to generate a mapping drift judgment indicator for determining whether there is a drift trend in the mapping relationship between the beam trajectory and the control signal. This also includes:

[0015] Based on the dynamic hysteresis disturbance, the responses of the beam trajectory along the X and Y axes are compared to generate a cross-variation coefficient representing the degree of lateral coupling distortion.

[0016] The cross-variance coefficient is calculated to generate a mapping drift judgment indicator used to determine whether there is a drift trend in the mapping relationship between the beam trajectory and the control signal.

[0017] Furthermore, based on the mapping drift judgment identifier, the hierarchical response of the inverse mapping neural network model is analyzed, and model trigger thresholds and model hierarchical weights representing changes in the mapping relationship between the beam trajectory and the control signal are generated, including:

[0018] Based on the mapping drift judgment identifier, the hierarchical perturbation factor representing the degree of disorder in the inter-layer response transmission is calculated for each hidden layer in the inverse mapping neural network model.

[0019] Based on the hierarchical perturbation factor, the update sensitivity of the inverse mapping neural network model and the differences in the mapping contributions of each layer are analyzed, and the model trigger threshold and the model hierarchical weights representing the changes in the mapping relationship between the beam trajectory and the control signal are obtained respectively.

[0020] Furthermore, based on the beam trajectory location data and control signal data, the deviation of the current beam trajectory is analyzed, and a deviation judgment value representing the degree of actual beam trajectory deviation is generated, including:

[0021] Based on the X-axis offset and Y-axis offset in the beam track position data and the output current change rate in the control signal data, the direction of change of current control and track offset is analyzed to obtain the transient mismatch coefficient representing the severity of the error.

[0022] The transient mismatch coefficient is analyzed to determine the deviation of the current beam trajectory and generate a deviation judgment value that represents the degree of deviation of the actual beam trajectory.

[0023] Furthermore, based on the model trigger threshold, mapping drift judgment flag, and deviation judgment value, the update status of the inverse mapping neural network model is analyzed, and update determination coefficients are generated to determine whether the parameters of the inverse mapping neural network model need to be updated, including:

[0024] Based on the status of the mapping drift judgment flag and the deviation judgment value, the duration and slope of their simultaneous activation are analyzed to generate a co-deterioration coefficient representing the deterioration trend of mapping drift and orbital deviation.

[0025] Furthermore, based on the model trigger threshold, mapping drift judgment flag, and deviation judgment value, the update status of the inverse mapping neural network model is analyzed, and update determination coefficients are generated to determine whether the parameters of the inverse mapping neural network model need to be updated. This also includes:

[0026] Obtain the cumulative number of updates of the inverse mapping neural network model, fuse the co-deterioration coefficient with the cumulative number of updates, and generate the inverse mapping confidence decay rate, which represents the degree of decay of the current inverse mapping confidence of the model.

[0027] The co-deterioration coefficient, inverse mapping confidence decay rate, and model trigger threshold are analyzed to generate update determination coefficients for determining whether the parameters of the inverse mapping neural network model need to be updated.

[0028] Furthermore, based on the updated determination coefficients, model level weights, mapping drift judgment indicators, and deviation judgment values, a collaborative calculation is performed to analyze the adjustment direction of the inverse mapping neural network model and generate control commands, including:

[0029] Based on the updated determination coefficients, the mapping drift judgment flag, and the deviation judgment value, the comprehensive direction driving factor representing the current control direction correction requirement is calculated.

[0030] Furthermore, based on the updated determination coefficients, model level weights, mapping drift judgment indicators, and deviation judgment values, a collaborative calculation is performed to analyze the adjustment direction of the inverse mapping neural network model and generate control commands. This also includes:

[0031] By comparing the comprehensive directional driving factors with the model hierarchical weights, the adjustment direction of each layer is analyzed, and control instructions are generated.

[0032] Furthermore, an accelerator beam trajectory control system based on a neural network inverse model, applied to the above control method, includes:

[0033] The drift judgment unit is used to acquire in real time the beam trajectory position data and the control signal data of the correction magnet power supply output by the beam position detector during the operation of the SSMB storage ring. It analyzes the beam trajectory position data and control signal data according to the inverse mapping neural network model and generates a mapping drift judgment label to determine whether there is a drift trend in the mapping relationship between the beam trajectory and the control signal.

[0034] The model analysis unit is used to analyze the hierarchical response of the inverse mapping neural network model based on the mapping drift judgment identifier, and generate the model trigger threshold and the model hierarchical weights representing the changes in the mapping relationship between the beam trajectory and the control signal.

[0035] The deviation analysis unit is used to analyze the deviation of the current beam trajectory based on the beam trajectory position data and control signal data, and generate a deviation judgment value that represents the degree of deviation of the actual beam trajectory.

[0036] The model update unit is used to analyze the update status of the inverse mapping neural network model based on the model trigger threshold, mapping drift judgment flag and deviation judgment value, and generate update determination coefficients to determine whether the parameters of the inverse mapping neural network model need to be updated.

[0037] The control analysis unit performs collaborative calculations based on the updated determination coefficients, model level weights, mapping drift judgment indicators, and deviation judgment values ​​to analyze the adjustment direction of the inverse mapping neural network model and generate control commands.

[0038] In summary, the present invention has the following main beneficial effects:

[0039] By performing time-series lag distribution analysis on beam trajectory point data and control signal data using an inverse mapping neural network model, dynamic hysteresis disturbances are obtained. Based on this, the response differences between the beam trajectory X-axis and Y-axis are compared to generate cross-variation coefficients, thereby obtaining a mapping drift judgment indicator. This enables accurate judgment of whether there is a drift trend in the mapping relationship between the beam trajectory and the control signal. At the same time, based on the mapping drift judgment indicator, the hierarchical perturbation factors of each hidden layer of the inverse mapping neural network model are analyzed, and the model trigger threshold and the model hierarchical weight of each neuron are calculated, thereby reflecting the degree of disorder in the mapping relationship and the difference in contribution of each layer.

[0040] By analyzing the X-axis and Y-axis offsets in the beam trajectory point data and the output current change rate in the control signal data, a deviation judgment value is generated to accurately reflect the degree of deviation of the actual beam trajectory. The number of consecutive simultaneous activation steps and the change slope of the mapping drift judgment mark and the deviation judgment value are integrated into a cooperative deterioration coefficient. Combined with the cumulative update times, the inverse mapping confidence decay rate is obtained, and finally, the update determination coefficient is generated to determine whether the parameters of the inverse mapping neural network model need to be updated. When an update is required, the comprehensive direction driving factor and the model level weights are analyzed collaboratively, and control commands are output for each neuron. This scheme enables the inverse mapping neural network model to track the dynamic changes in the mapping relationship caused by the drift of the magnet excitation intensity in the SSMB storage ring and the accumulation of equipment collimation installation deviation in real time. It dynamically decides the timing of parameter updates and the direction of weight adjustment of the inverse mapping neural network model, solving the problem that traditional inverse models cannot track mapping drift. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the steps of the accelerator beam trajectory control method based on a neural network inverse model of the present invention.

[0042] Figure 2 This is a schematic diagram of the accelerator beam trajectory control system based on the neural network inverse model of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] refer to Figure 1 and Figure 2 Accelerator beam trajectory control methods based on neural network inverse models include:

[0045] Step S1: Real-time acquisition of beam trajectory position data and control signal data of correction magnet power supply output by beam position detector during SSMB storage ring operation; analysis of beam trajectory position data and control signal data based on inverse mapping neural network model to generate mapping drift judgment label for judging whether there is a drift trend in the mapping relationship between beam trajectory and control signal.

[0046] The beam trajectory location data includes: the X-axis offset and Y-axis offset of the beam trajectory, etc.

[0047] Control signal data includes: power supply output voltage, output current, and output current change rate, etc.

[0048] Step S2: Based on the mapping drift judgment identifier, analyze the hierarchical response of the inverse mapping neural network model, and generate the model trigger threshold and the model hierarchy weights representing the changes in the mapping relationship between the beam trajectory and the control signal, respectively.

[0049] Step S3: Based on the beam trajectory point data and control signal data, analyze the deviation of the current beam trajectory and generate a deviation judgment value that represents the degree of deviation of the actual beam trajectory.

[0050] Step S4: Analyze the update status of the inverse mapping neural network model based on the model trigger threshold, mapping drift judgment flag, and deviation judgment value, and generate update determination coefficients to determine whether the parameters of the inverse mapping neural network model need to be updated;

[0051] Step S5: Based on the updated determination coefficients, model level weights, mapping drift judgment indicators, and deviation judgment values, perform collaborative calculations to analyze the adjustment direction of the inverse mapping neural network model and generate control commands.

[0052] In one embodiment, the beam trajectory point data and control signal data are analyzed according to the inverse mapping neural network model to generate a mapping drift judgment flag for determining whether there is a drift trend in the mapping relationship between the beam trajectory and the control signal, including:

[0053] For the beam trajectory position data and control signal data, the time lag distribution is analyzed based on the inverse mapping neural network model to obtain the dynamic hysteresis disturbance representing the control response delay. Specifically, the inverse mapping neural network model is a three-layer feedforward network used to inversely deduce the timing characteristics of the control signal for the trajectory from the current beam trajectory position data. The input layer contains two neurons, which receive the current beam trajectory X-axis offset and Y-axis offset, respectively. Both are normalized to eliminate dimensional differences. The hidden layer contains ten neurons, each using the hyperbolic tangent activation function to extract the nonlinear inverse mapping characteristics between the trajectory and the control signal. The output layer contains three neurons, corresponding to the three key parameters of the control signal: power supply output voltage, output current, and output current change rate. The model is trained using historical data, and the connection weights and biases are adjusted through error backpropagation. After training, the inverse mapping neural network model can predict the control signal that generates the trajectory based on the input trajectory offset.

[0054] At the current sampling moment, acquire the beam trajectory point data and acquire the actual control signal data corresponding to each of the past twenty sampling moments. Pair the current trajectory data with the actual control signals of the above twenty historical moments to form twenty test samples.

[0055] Each sampling time is 1ms. In the SSMB storage ring, after the correction magnet power supply receives the control signal, there is an inherent delay in the establishment of the magnetic field and the actual deviation of the beam trajectory, which is usually a few milliseconds to tens of milliseconds. Therefore, 20 milliseconds is used here to ensure that the maximum possible lag falls within the window and to avoid omissions.

[0056] For each test sample, the orbital data is input into the inverse mapping neural network model to obtain the three control signal parameters predicted by the model output layer. Then, the mean square error between the predicted value and the actual control signal in the test sample is calculated. The smaller the mean square error, the closer the actual control signal used at this historical moment is to the control signal that the model believes should be under the current orbital state. This indicates that the response effect of the control signal matches the current orbit exactly after a certain number of steps.

[0057] The reciprocals of the twenty mean squared error values ​​are taken respectively, and these twenty reciprocals constitute twenty likelihood values. Each likelihood value corresponds to a lag number, ranging from 1 to 20. Then, the sum of these twenty likelihood values ​​is calculated, and each likelihood value is divided by the sum to obtain the normalized weights. At this point, each lag number is assigned a weight between 0 and 1, and the sum of all weights is 1, forming a discrete probability distribution about the lag number.

[0058] Multiply each lag step by its corresponding weight to obtain twenty products, then sum these products to obtain the average lag step; calculate the difference between each lag step and the average lag step, multiply the square of the difference by the weight corresponding to that lag step to obtain twenty weighted square differences, sum these twenty weighted square differences to obtain the variance, and take the square root of the variance to obtain the standard deviation, which reflects the fluctuation of the lag step around the average.

[0059] Adding the average lag step count to the standard deviation yields the dynamic hysteresis disturbance that represents the delay in control response.

[0060] In one embodiment, the analysis of beam trajectory point data and control signal data based on an inverse mapping neural network model to generate a mapping drift judgment flag for determining whether there is a drift trend in the mapping relationship between the beam trajectory and the control signal further includes:

[0061] Based on the dynamic hysteresis disturbance, the responses of the beam orbit X-axis and Y-axis are compared to generate a cross-variation coefficient representing the degree of lateral coupling distortion. Specifically, the X-axis offset and Y-axis offset of the beam orbit are regarded as two independent channels, and the response of each channel is calculated, that is, the dynamic hysteresis disturbance is calculated.

[0062] The absolute value of the difference between the dynamic hysteresis disturbance on the X-axis and the dynamic hysteresis disturbance on the Y-axis is divided by the sum of the two dynamic hysteresis disturbances, and the calculation result is normalized to the 0-1 interval to generate a cross-variation coefficient representing the degree of lateral coupling distortion. A cross-variation coefficient close to 0 indicates that the delay of the control response in the two lateral directions is almost equal, and the lateral coupling distortion is slight. If the cross-variation coefficient is close to 1, it indicates that the delay in one direction is significantly greater than that in the other direction, indicating severe lateral coupling distortion.

[0063] The cross-variance coefficient is calculated to generate a mapping drift judgment identifier to determine whether there is a drift trend in the mapping relationship between the beam trajectory and the control signal. Specifically, this includes: establishing two historical queues, each with the same length as the above sampling time, which is 20 sampling times. The first queue stores the cross-variance coefficients of the most recent 20 sampling times, and the second queue stores the first-order absolute difference of the cross-variance coefficients of the most recent 20 sampling times. The first-order absolute difference is the absolute value of the cross-variance coefficient of the current sampling time minus the cross-variance coefficient of the previous sampling time. The first-order absolute difference is mainly used to reflect the degree of change of the coefficients between adjacent times.

[0064] Calculate the median of the first queue as the current short-term baseline value; use the median of the first queue as the baseline scale and the median of the second queue as the noise amplitude. The detection threshold is then the sum of the baseline scale and the noise amplitude. This calculation allows the detection threshold to increase adaptively when the noise level rises, avoiding false judgments triggered by normal fluctuations. When the noise level is very low, the detection threshold is mainly determined by the baseline scale, maintaining sensitivity to drift.

[0065] Calculate the deviation of the current cross-variance coefficient from the short-term baseline value, i.e., the absolute value of the difference between the two. If the deviation is greater than the detection threshold, it is determined that there is a significant lateral coupling distortion anomaly, and the mapping drift judgment flag is directly set to 1. Otherwise, calculate the sum of five consecutive deviations and the sum of five consecutive detection thresholds. Calculating the sum of five deviations and the sum of five detection thresholds can quickly indicate the trend of continuous offset. If the sum of deviations is greater than the sum of detection thresholds, it indicates that there is a continuous offset, and the mapping drift judgment flag is set to 1. Otherwise, the mapping drift judgment flag is set to 0, indicating that there is no drift trend. The mapping drift judgment flag is mainly used to determine whether there is a drift trend in the mapping relationship between the beam trajectory and the control signal.

[0066] By accurately analyzing the temporal lag distribution of the control response using an inverse mapping neural network model, dynamic hysteresis disturbances are obtained, and cross-variation coefficients and mapping drift judgment indicators are derived. This scheme can adaptively determine the degree of lateral coupling distortion and dynamically adjust the detection threshold to avoid misjudgments caused by normal fluctuations. At the same time, by comparing the sum of continuous deviation amplitudes with the sum of detection thresholds, the continuous offset trend is identified, thereby determining whether the inverse model needs to be updated and the direction of weight adjustment. This improves the tracking capability of mapping relationship drift, ensures that the residual orbit closure accuracy of the beam trajectory is maintained below the micrometer level, and enhances the control stability under long-term operation.

[0067] In one embodiment, based on the mapping drift judgment identifier, the hierarchical response of the inverse mapping neural network model is analyzed to generate model trigger thresholds and model hierarchical weights representing changes in the mapping relationship between the beam trajectory and the control signal, including:

[0068] Based on the mapping drift judgment flag, the hierarchical perturbation factor representing the degree of disorder in the inter-layer response transmission is calculated for each hidden layer in the inverse mapping neural network model. Specifically, this includes: establishing a drift-free benchmark library; collecting the output activation values ​​of each neuron in the hidden layer at sampling times when the mapping drift judgment flag is continuously 0; calculating the mean and standard deviation of the output activation values ​​of each neuron; and simultaneously calculating the overall standard deviation of the output activation values ​​of the ten neurons in the hidden layer.

[0069] Regardless of whether the current mapping drift judgment flag is 0 or 1, for each neuron, divide the absolute value of the difference between its current activation value and the mean by the standard deviation of that neuron to obtain the deviation factor. Then sum the ten deviation factors to obtain the total deviation of the hidden layer.

[0070] Calculate the standard deviation among the ten activation values ​​of the hidden layer at the current moment, divide the standard deviation by the overall standard deviation recorded during the drift-free period to obtain the intra-layer dispersion factor; multiply the total deviation by the intra-layer dispersion factor to obtain the hierarchical perturbation factor of the hidden layer. When the hierarchical perturbation factor > 1, it indicates that there is significant disorder in the inter-layer response transmission; when the hierarchical perturbation factor ≤ 1, it is considered as normal fluctuation.

[0071] Based on the hierarchical perturbation factor, the update sensitivity of the inverse mapping neural network model and the differences in the mapping contributions of each layer are analyzed. The model trigger threshold and the model hierarchical weights representing the changes in the mapping relationship between the beam trajectory and the control signal are obtained. The specific calculation formulas are as follows: ;

[0072] In the formula, Indicates the model trigger threshold. This means that, prior to the current moment, the hierarchical perturbation factors calculated from the ten most recent sampling moments where the mapping drift judgment flag is consecutively zero constitute a sequence, which is the set of hierarchical perturbation factors. , Indicates taking The median of Indicates calculation The absolute deviation of the median is calculated by first taking the absolute value of the difference between each level of disturbance factor and the median, obtaining ten absolute values, and then calculating the median of these absolute values. The sample skewness coefficient is obtained by dividing the third central moment of the ten level perturbation factors by the 1.5th power of the second central moment of the ten level perturbation factors.

[0073] The specific formula for calculating the model hierarchy weights is as follows: ;

[0074] In the formula, This represents the model layer weights of the i-th neuron. This represents the degree of deviation of the i-th neuron, that is, the degree to which the current activation value deviates from the historical baseline. , and Let represent the mean and standard deviation of the activation value of the i-th neuron in the drift-free benchmark library, respectively. This represents the activation value of the neuron at the current moment. The standard deviation of the ten activations in the hidden layer at the current time is represented by tanh, where tanh represents the hyperbolic tangent function, and e represents the natural constant. This represents the percentile rank of the activation value of the i-th neuron within the same neuron activation value sequence in the drift-free benchmark library, with a value ranging from 0 to 1. Indicates the summation index. Indicates the first The percentile ranking of a neuron's own activation value within the same neuron activation value sequence in a drift-free benchmark library. That is the first The degree of deviation of each neuron.

[0075] By using the mapping drift judgment flag to trigger the layer-by-layer analysis of the hidden layer response of the inverse mapping neural network model, the hierarchical perturbation factor is calculated, thereby reflecting the degree of disorder in the inter-layer response transmission. At the same time, the model trigger threshold and model hierarchical weights are calculated, which can accurately determine whether the inverse mapping neural network model needs to be updated and the direction of adjustment of each level weight. This enables real-time tracking of the drift of the mapping relationship between the beam trajectory and the control signal, avoiding control failure caused by magnet excitation drift and equipment collimation deviation accumulation in the long-term operation of the fixed structure model.

[0076] In one embodiment, based on beam trajectory position data and control signal data, the deviation of the current beam trajectory is analyzed to generate a deviation judgment value representing the degree of actual beam trajectory deviation, including:

[0077] Based on the X-axis offset and Y-axis offset in the beam trajectory point data and the output current change rate in the control signal data, the direction of change of current control and trajectory offset is analyzed to obtain the transient mismatch coefficient representing the severity of the error. Specifically, this includes: calculating the difference between the X-axis offset and the Y-axis offset between the current moment and the previous moment, and then taking the square root of the sum of the squares of these two differences to obtain the composite amplitude of the trajectory offset change.

[0078] To determine whether the direction of change of control current is consistent with the direction of change of track offset, specifically: if the product of the difference in offset in the X direction and the rate of change of current, and the product of the difference in offset in the Y direction and the rate of change of current are both non-negative, then they are considered to be consistent in direction; otherwise, they are not consistent in direction.

[0079] Divide the synthesized amplitude by the absolute value of the output current change rate and normalize the calculation result to the 0-1 range to obtain the ratio. If the direction of the control current change is consistent with the direction of the track offset change, the transient mismatch coefficient is this ratio; if the directions are inconsistent, the transient mismatch coefficient is the sum of 1 and the ratio. The larger the transient mismatch coefficient, the more severe the instantaneous following error.

[0080] The transient mismatch coefficients are analyzed to determine the deviation of the current beam trajectory and generate a deviation judgment value representing the degree of deviation of the actual beam trajectory. Specifically, this includes: taking all transient mismatch coefficients calculated at all sampling times corresponding to the drift-free reference library as reference samples, calculating the geometric mean and interquartile range of the reference samples; for the transient mismatch coefficient at the current time, calculating the ratio of the transient mismatch coefficient to the geometric mean, and then taking the logarithm to the base 10 to obtain the logarithmic deviation exponent.

[0081] If at least one of the current transient mismatch coefficient and the transient mismatch coefficient at the previous moment is zero, the relative rate of change is defined as 0. Otherwise, the difference between the current transient mismatch coefficient and the transient mismatch coefficient at the previous moment is calculated and then divided by the harmonic mean of the two to obtain the relative rate of change. The logarithmic deviation exponent is multiplied by the relative rate of change to obtain the composite distortion factor.

[0082] The current composite distortion factor is compared with the upper quartile of the benchmark sample. If the current composite distortion factor is greater than the upper quartile, the deviation judgment value is directly 1, indicating that the actual beam trajectory has deviated significantly. Otherwise, the deviation judgment value is 0, indicating that the actual beam trajectory deviation is within a reasonable range.

[0083] By accurately calculating the consistency between the composite amplitude of the track offset change and the direction of the control current change in the X-axis and Y-axis offset data from the beam track position data and the output current change rate in the control signal data, a transient mismatch coefficient is generated, which can reflect the severity of the instantaneous following error. Combined with the transient mismatch coefficient and the composite distortion factor, the deviation judgment value is finally calculated. This scheme can sensitively identify the instantaneous deviation state between the beam track and the control signal. Even if the magnet excitation intensity drifts or the equipment collimation installation deviation accumulates over a long period of time, it can judge the actual deviation degree of the beam track in real time, avoiding control failure due to excessive track deviation.

[0084] In one embodiment, the update status of the inverse mapping neural network model is analyzed based on the model trigger threshold, mapping drift judgment flag, and deviation judgment value. Update determination coefficients are then generated to determine whether the parameters of the inverse mapping neural network model need to be updated, including:

[0085] Based on the status of the mapping drift judgment flag and the deviation judgment value, analyze the duration and change slope of their simultaneous activation, and generate a co-deterioration coefficient representing the deterioration trend of mapping drift and orbit deviation. Specifically, the following steps are taken: if the mapping drift judgment flag is 1 and the deviation judgment value is also 1, it is considered as simultaneous activation. Starting from the current moment, trace back along the time axis to find the longest sequence that continuously satisfies simultaneous activation. Record the number of sampling points contained in the continuous sequence, which is the number of continuous activation steps. The number of continuous activation steps represents the duration of simultaneous activation.

[0086] Calculate the slope of change within the continuous sequence, that is, calculate the difference between the deviation judgment values ​​at adjacent sampling times within the continuous sequence, sum the absolute values ​​of these differences, and then divide by the difference between the sequence length and 1 to obtain the frequency of change of the deviation judgment value, which is the slope of change. The larger the slope of change, the more frequently the deviation judgment value fluctuates during the continuous activation period, reflecting the unstable deterioration trend.

[0087] Multiply the number of consecutive activation steps by the slope of change, and normalize the result to the 0-1 interval to obtain the co-deterioration coefficient, which represents the deterioration trend of mapping drift and orbital deviation. The larger the co-deterioration coefficient, the more significant the co-deterioration trend of mapping drift and orbital deviation.

[0088] In one embodiment, the process of analyzing the update status of the inverse mapping neural network model based on the model trigger threshold, mapping drift judgment flag, and deviation judgment value, and generating update determination coefficients for determining whether the parameters of the inverse mapping neural network model need to be updated, further includes:

[0089] To obtain the cumulative number of updates of the inverse mapping neural network model, the co-deterioration coefficient is fused with the cumulative number of updates to generate the inverse mapping confidence decay rate, which represents the degree of decay of the current inverse mapping confidence of the model. Specifically, starting from the beginning of the use of the inverse mapping neural network model, the number of parameter updates of each inverse mapping neural network model is accumulated to obtain the cumulative number of updates.

[0090] The co-deterioration coefficient at the current moment is fused with the cumulative number of updates. If the co-deterioration coefficient is zero, the inverse mapping confidence decay rate is also zero, indicating that the confidence does not decay when no deterioration occurs.

[0091] Otherwise, first calculate the natural logarithm of the co-deterioration coefficient, which can compress the numerical range of the deterioration trend and avoid extreme values ​​dominating the results. Then, multiply the value after taking the natural logarithm by the cumulative number of updates to obtain the intermediate product value. The intermediate product value reflects the amplification effect of the number of updates on the confidence decay. Then, divide the intermediate product value by the sum of the cumulative number of updates and 1, and normalize the calculation result to the 0-1 interval to obtain the inverse mapping confidence decay rate, which represents the degree of confidence decay of the current inverse mapping of the model. The larger the value of the inverse mapping confidence decay rate, the more severe the confidence decay of the current model inverse mapping.

[0092] The co-deterioration coefficient, inverse mapping confidence decay rate, and model trigger threshold are analyzed to generate update determination coefficients for determining whether the parameters of the inverse mapping neural network model need to be updated. Specifically, the co-deterioration coefficient is multiplied by the inverse mapping confidence decay rate to obtain the comprehensive deterioration intensity. The comprehensive deterioration intensity mainly reflects the severity of the co-deterioration of the current mapping drift and trajectory deviation, as well as the degree of confidence decay caused by the accumulation of historical model updates.

[0093] The overall deterioration intensity is compared with the model trigger threshold. If the overall deterioration intensity is greater than the model trigger threshold, it is determined that the current model parameters can no longer accurately describe the mapping relationship between the beam trajectory and the control signal. The coefficient of determination is updated to 1, and the model parameters are updated. Otherwise, the coefficient of determination is updated to 0, indicating that the current mapping relationship is still within an acceptable range, and the current model parameters are kept unchanged.

[0094] In one embodiment, based on the updated determination coefficients, model level weights, mapping drift judgment flags, and deviation judgment values, a collaborative calculation is performed to analyze the adjustment direction of the inverse mapping neural network model and generate control commands, including:

[0095] Based on the updated determination coefficient, the mapping drift judgment flag, and the deviation judgment value, calculate the comprehensive direction driving factor that represents the current control direction correction requirement. Specifically, when the updated determination coefficient is 0, it means that there is no need to adjust the model parameters. Directly output the control command that keeps the parameters of the current mapping neural network model unchanged, and end this step.

[0096] If the updated certainty coefficient is 1, the calculation continues. The two identifiers, the mapping drift judgment identifier and the deviation judgment value, are added together to obtain a sum. This sum reflects the overall evidence strength of the two abnormal states, namely, mapping drift and orbit deviation, at the current moment.

[0097] The cumulative sampling count of the inverse mapping neural network model is obtained, and the sum of the values ​​calculated at each sampling time point of the inverse mapping neural network model is added together and then divided by the cumulative sampling count to obtain the dynamic baseline strength.

[0098] The difference between the current sum and the dynamic baseline strength is calculated. This difference is the comprehensive direction driving factor that represents the current control direction correction requirement. If the comprehensive direction driving factor is positive, it means that the current anomalous evidence strength exceeds the long-term average level, and the control direction needs to be positively adjusted in weight, that is, to enhance the response of the inverse mapping neural network model to the current mapping relationship. Otherwise, it means that the anomalous strength is lower than the long-term average level, and the weight needs to be negatively adjusted in weight, that is, to suppress the response of the inverse mapping neural network model to the current mapping relationship.

[0099] In one embodiment, the method further includes performing collaborative calculations based on updated determination coefficients, model level weights, mapping drift judgment flags, and deviation judgment values ​​to analyze the adjustment direction of the inverse mapping neural network model and generate control commands.

[0100] The comprehensive direction-driving factor is compared with the model hierarchical weights to analyze the adjustment direction of each layer and generate control instructions. Specifically, at the current sampling time, the model hierarchical weights of all ten neurons in the hidden layer are collected, the median of these model hierarchical weights is calculated, and it is used as a dynamic boundary threshold. The median is chosen here instead of the average value because the median is not sensitive to abnormal weight values ​​and can better represent the typical level of weights.

[0101] For each neuron in the hidden layer, its model layer weight is compared with the dynamic boundary threshold. If the model layer weight is greater than the dynamic boundary threshold, it indicates that the contribution of the neuron is higher than the typical level of the current layer and plays a dominant role in the mapping relationship. Its weight adjustment direction should be consistent with the control direction of the comprehensive direction driving factor, and then the output direction is consistent with the control command.

[0102] If the model layer weights are less than the dynamic boundary threshold, it indicates that the contribution of the neuron is lower than the typical level. Its weight adjustment direction should be opposite to the control direction of the comprehensive direction driving factor, and then the output direction of the control command should be opposite.

[0103] If the model layer weights are equal to the dynamic boundary threshold, the weights corresponding to that neuron will not be adjusted, and a control command will be output to keep the parameters of the current mapped neural network model unchanged.

[0104] The co-deterioration coefficient is calculated by continuously activating the mapping drift judgment indicator and the deviation judgment value simultaneously, and the change slope is used to calculate the inverse mapping confidence decay rate. This rate is then compared with the model trigger threshold to generate the update determination coefficient, thereby accurately determining whether the parameters of the inverse mapping neural network model need to be updated. When an update is required, the mapping drift judgment indicator and deviation judgment value are further analyzed to obtain the comprehensive direction driving factor. Based on the comparison between the model layer weight of each neuron and the median within the layer, control commands are output for each neuron. This scheme enables the inverse mapping neural network model to dynamically decide the timing of parameter updates and the adjustment direction of each weight based on the real-time drift state of the beam trajectory and control signal mapping relationship, effectively solving the problem that fixed structure inverse models cannot track mapping drift.

[0105] In one embodiment, an accelerator beam trajectory control system based on a neural network inverse model is applied to the above-described control method, including:

[0106] The drift judgment unit is used to acquire in real time the beam trajectory position data and the control signal data of the correction magnet power supply output by the beam position detector during the operation of the SSMB storage ring. It analyzes the beam trajectory position data and control signal data according to the inverse mapping neural network model and generates a mapping drift judgment label to determine whether there is a drift trend in the mapping relationship between the beam trajectory and the control signal.

[0107] The model analysis unit is used to analyze the hierarchical response of the inverse mapping neural network model based on the mapping drift judgment identifier, and generate the model trigger threshold and the model hierarchical weights representing the changes in the mapping relationship between the beam trajectory and the control signal.

[0108] The deviation analysis unit is used to analyze the deviation of the current beam trajectory based on the beam trajectory position data and control signal data, and generate a deviation judgment value that represents the degree of deviation of the actual beam trajectory.

[0109] The model update unit is used to analyze the update status of the inverse mapping neural network model based on the model trigger threshold, mapping drift judgment flag and deviation judgment value, and generate update determination coefficients to determine whether the parameters of the inverse mapping neural network model need to be updated.

[0110] The control analysis unit performs collaborative calculations based on the updated determination coefficients, model level weights, mapping drift judgment indicators, and deviation judgment values ​​to analyze the adjustment direction of the inverse mapping neural network model and generate control commands.

[0111] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An accelerator beam trajectory control method based on a neural network inverse model, characterized in that, include: Step S1: Real-time acquisition of beam trajectory position data and control signal data of the correction magnet power supply output by the beam position detector during the operation of the SSMB storage ring; analysis of beam trajectory position data and control signal data based on the inverse mapping neural network model; generation of mapping drift judgment flag to determine whether there is a drift trend in the mapping relationship between the beam trajectory and the control signal; including: for beam trajectory position data and control signal data, analysis of the time lag distribution based on the inverse mapping neural network model to obtain the dynamic hysteresis disturbance representing the control response delay. Based on the dynamic hysteresis disturbance, the responses of the beam trajectory along the X and Y axes are compared to generate a cross-variation coefficient representing the degree of lateral coupling distortion. The cross-change coefficient is calculated to generate a mapping drift judgment indicator used to determine whether there is a drift trend in the mapping relationship between the beam trajectory and the control signal; Step S2: Based on the mapping drift judgment identifier, analyze the hierarchical response of the inverse mapping neural network model, and generate the model trigger threshold and the model hierarchical weights representing the changes in the mapping relationship between the beam trajectory and the control signal, including: based on the mapping drift judgment identifier, calculate each hidden layer in the inverse mapping neural network model to obtain the hierarchical perturbation factor representing the degree of disorder in the inter-layer response transmission. Based on the hierarchical perturbation factor, the update sensitivity of the inverse mapping neural network model and the differences in the mapping contribution of each layer are analyzed, and the model trigger threshold and the model hierarchical weights representing the changes in the mapping relationship between the beam trajectory and the control signal are obtained respectively. Step S3: Based on the beam trajectory point data and control signal data, analyze the deviation of the current beam trajectory and generate a deviation judgment value that represents the degree of deviation of the actual beam trajectory. Step S4: Analyze the update status of the inverse mapping neural network model based on the model trigger threshold, mapping drift judgment flag, and deviation judgment value, and generate update determination coefficients to determine whether the parameters of the inverse mapping neural network model need to be updated; Step S5: Based on the updated determination coefficients, model level weights, mapping drift judgment indicators, and deviation judgment values, perform collaborative calculations to analyze the adjustment direction of the inverse mapping neural network model and generate control commands.

2. The accelerator beam trajectory control method based on a neural network inverse model according to claim 1, characterized in that, Based on beam trajectory location data and control signal data, the deviation of the current beam trajectory is analyzed, and a deviation judgment value representing the degree of actual beam trajectory deviation is generated, including: Based on the X-axis offset and Y-axis offset in the beam track position data and the output current change rate in the control signal data, the direction of change of current control and track offset is analyzed to obtain the transient mismatch coefficient representing the severity of the error. The transient mismatch coefficient is analyzed to determine the deviation of the current beam trajectory, and a deviation judgment value representing the degree of deviation of the actual beam trajectory is generated.

3. The accelerator beam trajectory control method based on a neural network inverse model according to claim 2, characterized in that, Based on the model trigger threshold, mapping drift judgment flag, and deviation judgment value, the update status of the inverse mapping neural network model is analyzed, and update determination coefficients are generated to determine whether the parameters of the inverse mapping neural network model need to be updated, including: Based on the status of the mapping drift judgment flag and the deviation judgment value, the duration and slope of their simultaneous activation are analyzed to generate a co-deterioration coefficient representing the deterioration trend of mapping drift and orbital deviation.

4. The accelerator beam trajectory control method based on a neural network inverse model according to claim 3, characterized in that, The update status of the inverse mapping neural network model is analyzed based on the model trigger threshold, mapping drift judgment flag, and deviation judgment value. Update determination coefficients are generated to determine whether the parameters of the inverse mapping neural network model need to be updated. This also includes: Obtain the cumulative number of updates of the inverse mapping neural network model, fuse the co-deterioration coefficient with the cumulative number of updates, and generate the inverse mapping confidence decay rate, which represents the degree of decay of the current inverse mapping confidence of the model. The co-deterioration coefficient, inverse mapping confidence decay rate, and model trigger threshold are analyzed to generate update determination coefficients for determining whether the parameters of the inverse mapping neural network model need to be updated.

5. The accelerator beam trajectory control method based on a neural network inverse model according to claim 4, characterized in that, Based on the updated certainty coefficients, model level weights, mapping drift judgment indicators, and deviation judgment values, a collaborative calculation is performed to analyze the adjustment direction of the inverse mapping neural network model and generate control commands, including: Based on the updated determination coefficients, mapping drift judgment flags, and deviation judgment values, a comprehensive direction driving factor representing the current control direction correction requirement is calculated.

6. The accelerator beam trajectory control method based on a neural network inverse model according to claim 5, characterized in that, Based on the updated certainty coefficients, model hierarchy weights, mapping drift judgment indicators, and deviation judgment values, a collaborative calculation is performed to analyze the adjustment direction of the inverse mapping neural network model and generate control commands. This also includes: By comparing the comprehensive directional driving factors with the model hierarchical weights, the adjustment direction of each layer is analyzed, and control instructions are generated.

7. An accelerator beam trajectory control system based on a neural network inverse model, applied in the control method described in any one of claims 1-6, characterized in that, include: The drift judgment unit is used to acquire in real time the beam trajectory position data and the control signal data of the correction magnet power supply output by the beam position detector during the operation of the SSMB storage ring. It analyzes the beam trajectory position data and control signal data according to the inverse mapping neural network model and generates a mapping drift judgment label to determine whether there is a drift trend in the mapping relationship between the beam trajectory and the control signal. The model analysis unit is used to analyze the hierarchical response of the inverse mapping neural network model based on the mapping drift judgment identifier, and generate the model trigger threshold and the model hierarchical weights representing the changes in the mapping relationship between the beam trajectory and the control signal. The deviation analysis unit is used to analyze the deviation of the current beam trajectory based on the beam trajectory position data and control signal data, and generate a deviation judgment value that represents the degree of deviation of the actual beam trajectory. The model update unit is used to analyze the update status of the inverse mapping neural network model based on the model trigger threshold, mapping drift judgment flag and deviation judgment value, and generate update determination coefficients to determine whether the parameters of the inverse mapping neural network model need to be updated. The control analysis unit performs collaborative calculations based on the updated determination coefficients, model level weights, mapping drift judgment indicators, and deviation judgment values ​​to analyze the adjustment direction of the inverse mapping neural network model and generate control commands.

Images