A particle accelerator low-level control parameter prediction method and system

Abstract

The present application relates to the technical field of particle accelerator control, and discloses a particle accelerator low-level control parameter prediction method and system, comprising the following steps: step 1, obtaining low-level control related data and encapsulating the data as a parameter adjustment period level input sample set; step 2, determining a control action influence amount, a thermal drift lag amount and a channel mismatch conduction amount; step 3, generating a state differentiation vector, a state constraint label and an input sequence with a state label; step 4, performing three-branch bidirectional long short-term memory network coding fusion to obtain comprehensive memory features; step 5, obtaining a parameter adjustment context representation through attention weighting based on the comprehensive memory features and the state constraint label; step 6, obtaining a field building power and a plurality of channel incident and reflected power prediction values; and step 7, combining a target field building power to back-propagate a next parameter adjustment period control setting amount. The present application realizes prediction and determination of a next parameter adjustment period particle accelerator low-level control setting amount.

Description

Technical Field

[0001] This invention belongs to the field of particle accelerator control technology, specifically relating to a method and system for predicting low-level control parameters of a particle accelerator. Background Technology

[0002] Particle accelerators are core components of radiotherapy equipment, and their low-level control systems are responsible for tasks such as field establishment adjustment, power distribution, and state stability control. During equipment operation, the control system typically needs to continuously adjust the field establishment power, incident power, and reflected power around gain settings, phase settings, feed settings, and feedforward settings to ensure a stable and controllable output state during treatment. Especially during the open-loop parameter tuning stage, changes in different control settings directly affect the subsequent power chain results. Improper parameter tuning can easily cause field establishment deviations, uneven channel power distribution, or increased reflection levels, thereby affecting subsequent operational stability.

[0003] Existing parameter tuning methods often rely on empirical settings, partial corrections, or single-target calibrations, frequently treating changes in control settings, thermal state changes, and inter-channel power mismatches separately, lacking a unified characterization of the coupling relationships among multiple influencing factors. In reality, thermal state changes typically exhibit cross-cycle cumulative characteristics, channel mismatches propagate along the power chain, and distribution shifts occur between different channels; these factors also superimpose with the current parameter tuning action. If simple rules or uniform timing processing methods are still used, it is difficult to accurately predict the establishment power, incident power, and reflected power of the next parameter tuning cycle, and it is also not conducive to further deducing the control settings for the next parameter tuning cycle. Summary of the Invention

[0004] This invention provides a method and system for predicting low-level control parameters of particle accelerators, solving the technical problems in the background art.

[0005] This invention provides a method for predicting low-level control parameters of a particle accelerator, comprising the following steps:

[0006] Step 1: Obtain low-level control-related data for multiple consecutive tuning cycles during the open-loop tuning phase of the particle accelerator, encapsulate the data according to the tuning cycle, and obtain a tuning cycle-level input sample set. The tuning cycle-level input sample set includes control settings, thermal state parameters, and power chain result parameters.

[0007] Step 2: Based on the input samples of adjacent parameter tuning cycles, determine the control action influence, thermal drift hysteresis, and channel mismatch conduction for each parameter tuning cycle.

[0008] Step 3: Based on the influence of the control action, the thermal drift hysteresis, and the channel mismatch conduction, determine the state differentiation vector and state constraint label corresponding to each parameter tuning cycle, and construct the input sequence with state labels;

[0009] Step 4: Input the state-marked input sequence into the first bidirectional long short-term memory network coding branch, the second bidirectional long short-term memory network coding branch, and the third bidirectional long short-term memory network coding branch respectively to obtain control action memory features, hot drift memory features, and channel mismatch memory features, and fuse them to obtain comprehensive memory features;

[0010] Step 5: Based on the comprehensive memory features and state constraint labels, obtain the parameter tuning context representation through a state-constrained attention weighting structure;

[0011] Step 6: Input the parameter tuning context representation into a chain-consistent multi-branch output structure to obtain the field establishment power prediction value, the incident power prediction value of multiple channels, and the reflection power prediction value of multiple channels for the next parameter tuning cycle.

[0012] Step 7: Based on the parameter tuning context representation, the establishment power prediction value, the incident power prediction value of multiple channels, the reflection power prediction value of multiple channels, and the target establishment power, determine the multiple channel gain settings, multiple channel phase settings, feed settings, and multiple channel feedforward settings for the next parameter tuning cycle.

[0013] The present invention also provides a low-level control parameter prediction system for particle accelerators, comprising:

[0014] The sample construction module is used to acquire low-level control-related data of multiple consecutive tuning cycles during the open-loop tuning phase of the particle accelerator, encapsulate them according to the tuning cycle, and obtain a tuning cycle-level input sample set. The tuning cycle-level input sample set includes control setpoints, thermal state parameters, and power chain result parameters.

[0015] The state decomposition module is used to determine the control action influence, thermal drift hysteresis and channel mismatch conduction for each tuning cycle based on the input samples of adjacent tuning cycle levels.

[0016] The state labeling module is used to determine the state differentiation vector and state constraint label corresponding to each parameter tuning cycle based on the influence of the control action, the thermal drift hysteresis, and the channel mismatch conduction, and to construct the input sequence with state labels.

[0017] The branch coding module is used to input the state-marked input sequence into the first bidirectional long short-term memory network coding branch, the second bidirectional long short-term memory network coding branch, and the third bidirectional long short-term memory network coding branch respectively to obtain control action memory features, hot drift memory features, and channel mismatch memory features, and then fuse them to obtain comprehensive memory features;

[0018] The context extraction module is used to obtain a parametric context representation based on the comprehensive memory features and state constraint labels through a state-constrained attention weighting structure;

[0019] The power prediction module is used to input the parameter tuning context representation into a chain-consistent multi-branch output structure to obtain the field establishment power prediction value, the incident power prediction value of multiple channels, and the reflection power prediction value of multiple channels for the next parameter tuning cycle.

[0020] The parameter back-calculation module is used to determine the gain settings, phase settings, feed settings, and feedforward settings of multiple channels for the next parameter tuning cycle based on the parameter tuning context representation, the establishment power prediction value, the incident power prediction value of multiple channels, the reflection power prediction value of multiple channels, and the target establishment power.

[0021] The beneficial effects of this invention are as follows: This invention uses the parameter tuning cycle as the organizational unit, uniformly encapsulating control setpoints, thermal state quantities, and power chain result quantities, and further decomposing them into control action influence quantities, thermal drift hysteresis quantities, and channel mismatch conduction quantities, enabling the separate expression of influencing factors from different sources and with different change rhythms. Combined with state differentiation vectors, state constraint labels, bidirectional long short-term memory network encoding, and state-constrained attention weighted processing, historical information related to the current parameter tuning state can be extracted more accurately. Based on this, the field establishment power, incident power, and reflected power are jointly predicted, and the control setpoint for the next parameter tuning cycle is inferred. This improves the consistency between low-level control parameter prediction and the parameter tuning process, while also taking into account the overall field establishment results, multi-channel power distribution, and reflection suppression relationships, making subsequent parameter tuning more informed. Attached Figure Description

[0022] Figure 1 This is a flowchart of a method for predicting low-level control parameters of a particle accelerator according to the present invention. Detailed Implementation

[0023] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0024] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0025] like Figure 1 As shown, a method for predicting low-level control parameters of a particle accelerator includes the following steps:

[0026] Step 1: Obtain low-level control-related data for multiple consecutive tuning cycles during the open-loop tuning phase of the particle accelerator, encapsulate the data according to the tuning cycle, and obtain a tuning cycle-level input sample set. The tuning cycle-level input sample set includes control settings, thermal state parameters, and power chain result parameters.

[0027] Step 2: Based on the input samples of adjacent parameter tuning cycles, determine the control action influence, thermal drift hysteresis, and channel mismatch conduction for each parameter tuning cycle.

[0028] Step 3: Based on the influence of the control action, the thermal drift hysteresis, and the channel mismatch conduction, determine the state differentiation vector and state constraint label corresponding to each parameter tuning cycle, and construct the input sequence with state labels;

[0029] Step 4: Input the state-marked input sequence into the first bidirectional long short-term memory network coding branch, the second bidirectional long short-term memory network coding branch, and the third bidirectional long short-term memory network coding branch respectively to obtain control action memory features, hot drift memory features, and channel mismatch memory features, and fuse them to obtain comprehensive memory features;

[0030] Step 5: Based on the comprehensive memory features and state constraint labels, obtain the parameter tuning context representation through a state-constrained attention weighting structure;

[0031] Step 6: Input the parameter tuning context representation into a chain-consistent multi-branch output structure to obtain the field establishment power prediction value, the incident power prediction value of multiple channels, and the reflection power prediction value of multiple channels for the next parameter tuning cycle.

[0032] Step 7: Based on the parameter tuning context representation, the establishment power prediction value, the incident power prediction value of multiple channels, the reflection power prediction value of multiple channels, and the target establishment power, determine the multiple channel gain settings, multiple channel phase settings, feed settings, and multiple channel feedforward settings for the next parameter tuning cycle.

[0033] In one embodiment of the present invention, the system first acquires low-level control-related data for multiple consecutive tuning cycles during the open-loop parameter tuning phase of the particle accelerator. Then, it aggregates and encapsulates the low-level control-related data according to the tuning cycle to obtain a tuning cycle-level input sample set. Here, low-level control-related data refers to the data set that directly reflects the control input, thermal state changes, and power chain output results during the parameter tuning process; the tuning cycle refers to a continuous data interval from one control setting update to the next. For therapeutic particle accelerators, data from the open-loop parameter tuning phase more easily preserves the original correspondence between control settings and power response, facilitating both the subsequent establishment of predictive models and the reverse inference of control settings in the next tuning cycle.

[0034] Specifically, the system first acquires low-level control-related data. This data includes data corresponding to control settings, thermal state parameters, and power chain results. Subsequently, the system divides the low-level control-related data according to the start and end positions of each tuning cycle, obtaining the low-level control-related data corresponding to each tuning cycle, and establishes a binding relationship between each tuning cycle and its corresponding data. This binding relationship means that all types of data collected within the same tuning cycle belong to the same cycle identifier.

[0035] For example, if the gain setting is updated after the start of a parameter tuning cycle, the cooling state parameters, field establishment power, incident power, and reflected power collected during that cycle will all correspond to this setting update. This process ensures consistency in the sources of various data for subsequent calculations and avoids discrepancies in correspondence caused by cross-cycle data mixing.

[0036] The system extracts control settings, thermal state parameters, and power chain result parameters from the low-level control-related data corresponding to each parameter tuning cycle. The control settings include multiple channel gain settings, multiple channel phase settings, feed settings, and multiple channel feedforward settings; the thermal state parameters include the cooling state parameters on both sides; and the power chain result parameters include the establishment power, multiple channel incident power, and multiple channel reflected power. The multiple channels are a preset number; in this embodiment of the invention, the multiple channels are all four channels.

[0037] It should be noted that the control setpoint refers to the adjustment input actively provided by the low-level control system, the thermal state quantity refers to the data reflecting changes in the thermal environment of the device, and the power chain result quantity refers to the output result on the power transmission chain after the control input is applied. Extracting the cooling state quantities on the left and right sides separately facilitates the subsequent differentiation between overall thermal changes and left-right thermal imbalances; extracting the incident power and reflected power of the four channels separately also facilitates the subsequent identification of mismatch conduction relationships between the channels.

[0038] The system establishes a correspondence between the control setpoints, thermal state variables, and power chain results corresponding to the same parameter tuning cycle, and combines them according to a preset field order to obtain parameter tuning cycle-level input samples. Then, it arranges the parameter tuning cycle-level input samples according to the parameter tuning cycle order to obtain a parameter tuning cycle-level input sample set. The preset field order refers to a consistent field arrangement within each parameter tuning cycle sample. That is, the arrangement of similar data remains consistent between samples from previous and subsequent parameter tuning cycles. This facilitates subsequent difference calculations for adjacent parameter tuning cycles and allows subsequent state differentiation vector construction, bidirectional long short-term memory network encoding, and state-constrained attention calculations to directly call data objects in a unified format.

[0039] After the above steps, the multi-source data that was originally scattered throughout the parameter tuning process is organized into a sample set organized according to the parameter tuning cycle. Subsequently, when determining the influence of control actions, thermal drift hysteresis, and channel mismatch transmission, processing can be performed directly based on samples from the same parameter tuning cycle and samples from adjacent parameter tuning cycles. This ensures that the data input of the entire prediction chain remains consistent and that there is a clear correspondence between the subsequent output control setpoint and the front-end acquired data.

[0040] In one embodiment of the present invention, after forming a parameter tuning cycle-level input sample set, the system further determines the control action influence, thermal drift hysteresis, and channel mismatch conduction for each parameter tuning cycle based on adjacent parameter tuning cycle-level input samples. Here, the control action influence refers to the comprehensive change in the control setpoint relative to the previous parameter tuning cycle; the thermal drift hysteresis refers to the cumulative change in thermal state across adjacent parameter tuning cycles and across two parameter tuning cycles; and the channel mismatch conduction refers to the comprehensive mismatch caused by changes in the incident and reflection relationships between channels within the power chain. Through the above processing, the system separates the different influencing factors originally mixed in the control setpoint, thermal state, and power chain result into three categories of state variables, thereby providing a unified input for subsequent state differentiation vector construction.

[0041] Specifically, starting from the third tuning cycle, the system reads the current tuning cycle-level input samples, the previous tuning cycle-level input samples, and the input samples from the two previous tuning cycle-level input sample set, and establishes a reference relationship corresponding to the current tuning cycle. The reason for starting processing from the third tuning cycle is that the calculation of thermal drift hysteresis depends simultaneously on the thermal state quantities of the current tuning cycle, the previous tuning cycle, and the two previous tuning cycles. After this step, subsequent steps 22, 23, and 24 all call the corresponding samples under the same tuning cycle index, avoiding cross-cycle misalignment of data from different sources.

[0042] The system calculates the average absolute value of the corresponding differences in the gain settings of the four channels, the average absolute value of the corresponding differences in the phase settings of the four channels, the absolute value of the difference in the feed-in settings, and the average absolute value of the corresponding differences in the feed-forward settings of the four channels, based on the control settings in the current and previous parameter tuning cycle input samples. These four items are then weighted and summed according to preset weights to obtain the control action influence quantity corresponding to the k-th parameter tuning cycle. The control action influence quantity represents the comprehensive change in control settings at the current parameter tuning cycle relative to the previous cycle, reflecting the direct driving strength of the active parameter tuning action on subsequent power chain changes. The control action influence quantity can be expressed as: ,in, Indicates the first The influence of control actions corresponding to each parameter tuning cycle. Indicates the first The average of the absolute values ​​of the differences between the four channel gain settings in each tuning cycle and the previous tuning cycle. Indicates the first The average of the absolute values ​​of the differences between the four channel phase settings of each tuning cycle and the previous tuning cycle. Indicates the first The absolute value of the difference in the feed setting amount between each parameter tuning cycle and the previous parameter tuning cycle. Indicates the first The average of the absolute values ​​of the differences between the four channel feedforward settings in each tuning cycle and the previous tuning cycle. , , and This indicates the corresponding preset weight. In other words, the system first converts different types of control setting changes into four variable quantities, and then combines the four variable quantities into a single control action influence quantity.

[0043] The system averages the cooling state quantities on both sides of each tuning cycle based on the thermal state quantities in the current tuning cycle input sample, the previous tuning cycle input sample, and the input samples from the two previous tuning cycles, to obtain the corresponding average thermal state quantity. Then, it calculates the absolute values ​​of the differences between the current tuning cycle's average thermal state quantity and the previous tuning cycle's average thermal state quantity, the absolute values ​​of the differences between the current tuning cycle's average thermal state quantity and the average thermal state quantities from the two previous tuning cycles, and the absolute values ​​of the differences between the left and right sides of the current tuning cycle's cooling state quantities. Finally, it weights and sums these three items according to preset weights to obtain the first... The thermal drift hysteresis corresponds to each parameter tuning cycle. The thermal drift hysteresis represents the cumulative change in thermal state within adjacent tuning cycles and across two tuning cycles, reflecting the delayed impact of thermal state changes on subsequent parameter tuning results. The thermal drift hysteresis can be expressed as: ,in, Indicates the first Thermal drift hysteresis corresponding to each parameter tuning cycle Indicates the first Average thermal state parameters over one parameter adjustment cycle This represents the average thermal state parameters of the previous parameter tuning cycle. This represents the average thermal state parameters over the previous two parameter adjustment cycles. Indicates the first The left-side cooling state quantity of each parameter adjustment cycle. Indicates the first The right-side cooling status quantity of each parameter tuning cycle , and This represents the corresponding preset weight. Here, we consider not only the thermal state changes between adjacent tuning cycles, but also the thermal state changes across two tuning cycles, as well as the imbalance of the left and right thermal states within the same tuning cycle. Therefore, the obtained thermal drift hysteresis can simultaneously reflect the continuity of thermal changes and the differences in thermal distribution.

[0044] Based on the power chain results in the current parameter tuning cycle input samples, the system first averages the incident power of each channel to obtain the average incident power of the four channels; then it calculates the reflection coefficient, incident power deviation, and incident-reflection difference of each channel, and sums them according to preset weights to obtain the single-channel mismatch conduction; finally, it averages the single-channel mismatch conduction to obtain the first channel's average incident power. Channel mismatch conductance corresponding to each tuning cycle. Channel mismatch conductance represents the comprehensive deviation of each channel in terms of incident power, reflected power, and inter-channel distribution within the current tuning cycle, and is used to reflect the conduction level of mismatch within the power chain across multiple channels. The channel mismatch conductance can be expressed as: ,in, Indicates the first Channel mismatch conductance corresponding to each parameter tuning cycle Indicates the first The first parameter tuning cycle The reflection coefficient of each channel Indicates the first The first parameter tuning cycle Incident power deviation of each channel Indicates the first The first parameter tuning cycle The difference between incident and reflected light in each channel. , and This represents the corresponding preset weight. The reflection coefficient is determined by dividing the reflected power of the corresponding channel by the sum of the incident power of the corresponding channel and the smallest positive number. The incident power deviation is determined by the absolute value of the difference between the incident power of the corresponding channel and the average incident power of the four channels. The incident-reflection difference is determined by the absolute value of the difference between the incident power of the corresponding channel and the reflected power of the corresponding channel. Through the above processing, the system unifies the incident and reflection relationships within a single channel and the distribution differences between multiple channels into a single channel mismatch conductance.

[0045] The system establishes a correspondence between the control action influence, thermal drift hysteresis, and channel mismatch conduction for the current parameter tuning cycle, obtaining the state output for the current parameter tuning cycle. This state output serves as the direct input for subsequent state differentiation vectors and state constraint labels.

[0046] Through the above processing, we can distinguish the effects of rapid control changes, slow thermal drift, and channel mismatch during the parameter tuning process. We can also provide a state basis with the same period, name, and reference relationship for the subsequent construction of input sequences with state labels. This makes the prediction process of low-level control parameters of the particle accelerator closer to the change law in the actual parameter tuning scenario of the treatment device.

[0047] In one embodiment of the present invention, after obtaining the control action influence, thermal drift hysteresis, and channel mismatch conduction for each tuning cycle, the system further determines the state differentiation vector and state constraint label for each tuning cycle, and constructs an input sequence with state labels based on this. Here, the state differentiation vector refers to a state description object formed by combining the control action influence, thermal drift hysteresis, and channel mismatch conduction in a fixed order within the same tuning cycle; the state constraint label refers to the dominant proportion result calculated further based on the state differentiation vector, used to indicate which type of state has a more prominent influence in the current tuning cycle. Through the above processing, the system organizes the originally dispersed state quantities into state objects that can be compared and directly participate in subsequent feature fusion.

[0048] Specifically, the system sequentially combines the control action influence, thermal drift hysteresis, and channel mismatch conduction quantities corresponding to each parameter tuning cycle to obtain the state differentiation vector for each cycle. This sequential combination means that the three types of state variables maintain a consistent arrangement throughout all parameter tuning cycles. In other words, the positional relationship of the control action influence, thermal drift hysteresis, and channel mismatch conduction quantities within the state differentiation vector remains unchanged in any given parameter tuning cycle. After this step, subsequent processing no longer calls the three original state variables separately, but instead uniformly calls the single state object, the state differentiation vector.

[0049] The system maps the state differentiation vector corresponding to each parameter tuning cycle to preset mapping weights and uses this mapping with preset biases to determine the state mapping result. The state mapping result is then normalized to obtain the state constraint markers corresponding to each parameter tuning cycle. These state constraint markers include the proportion of control action dominance, the proportion of thermal drift hysteresis dominance, and the proportion of channel mismatch conduction dominance. The state constraint markers can be represented as follows: ,in, Indicates the first The proportion of control actions dominant in each parameter tuning cycle. Indicates the first The proportion of thermal drift hysteresis dominance corresponding to each parameter tuning cycle. Indicates the first The proportion of channel mismatch propagation dominance corresponding to each parameter tuning cycle. This indicates the preset mapping weights. This indicates the preset bias. Indicates the first The system first maps the state differentiation vectors according to preset mapping weights and preset biases, and then converts the mapping results to a uniform scale to obtain the three dominant proportions. This not only reflects which dominant state the current parameter tuning period is closer to, but also allows the three dominant proportions to be directly used as weights in feature fusion in subsequent processing.

[0050] It should be noted that the mapping weights represent the degree of participation of the three types of state variables in the state mapping process, the biases represent the overall adjustment amount of the mapping result, and the normalization process means organizing the state mapping result into a unified range that can be directly compared. After normalization, the proportions of control action dominance, thermal drift hysteresis dominance, and channel mismatch propagation dominance are expressed on the same scale. Subsequently, whether it is the fusion of bidirectional long short-term memory network coding branches or the attention weighted structure constrained by state, this set of state constraint labels can be directly called.

[0051] Using the current tuning cycle as the endpoint, the system sequentially reads the tuning cycle-level input samples, state differentiation vectors, and state constraint labels corresponding to the current tuning cycle and a predetermined number of consecutive tuning cycles prior to it, and establishes the correspondence between these input samples, state differentiation vectors, and state constraint labels within the same tuning cycle. The predetermined number of consecutive cycles refers to the length of historical tuning cycles pre-set during the sequence construction phase. In other words, the system does not only utilize the data from the current tuning cycle but also simultaneously introduces a continuous historical tuning cycle data segment preceding the current tuning cycle, enabling the subsequent network to identify the state evolution process within the time frame.

[0052] The system uses the input samples, state differentiation vectors, and state constraint labels corresponding to the same tuning cycle within the current tuning cycle and a predetermined number of previous tuning cycles as sequence elements, arranging them in tuning cycle order to obtain a state-labeled input sequence. Here, a sequence element refers to the smallest input unit formed by uniformly mapping the tuning cycle-level input samples, state differentiation vectors, and state constraint labels within a single tuning cycle. Through this process, the system generates not just a simple sample sequence, nor simply a simple state sequence, but a state-labeled input sequence that simultaneously carries both sample and state information.

[0053] Following the steps outlined above, the system further organizes the control action influence, thermal drift hysteresis, and channel mismatch conduction into state differentiation vectors and state constraint labels, which, together with the periodic input samples from the parameter tuning process, form a state-labeled input sequence. This embeds the state differences during parameter tuning into the temporal input structure and ensures that subsequent bidirectional long short-term memory network encoding and state-constrained attention calculations are based on unified, continuous, and distinguishable dominant state data. Consequently, the prediction of low-level control parameters for the particle accelerator more closely reflects the state evolution during actual parameter tuning.

[0054] In one embodiment of the present invention, after constructing an input sequence with state labels, the system further inputs the input sequence with state labels into a first bidirectional long short-term memory network coding branch, a second bidirectional long short-term memory network coding branch, and a third bidirectional long short-term memory network coding branch, respectively, to obtain control action memory features, thermal drift memory features, and channel mismatch memory features. These are then fused to obtain a comprehensive memory feature, which is used to characterize the comprehensive parameter tuning state of the current tuning cycle. Here, the control action memory feature refers to the feature result extracted by the first bidirectional long short-term memory network coding branch based on the control action change law; the thermal drift memory feature refers to the feature result extracted by the second bidirectional long short-term memory network coding branch based on the thermal state cross-cycle change law; the channel mismatch memory feature refers to the feature result extracted by the third bidirectional long short-term memory network coding branch based on the channel mismatch conduction law within the power chain; and the comprehensive memory feature is a single feature object formed by uniformly fusing the three types of memory features under the three dominant proportion constraints corresponding to the current parameter tuning cycle.

[0055] Specifically, the system reads the state-labeled input sequence corresponding to the current parameter tuning cycle, as well as the control action dominance ratio, thermal drift hysteresis dominance ratio, and channel mismatch propagation dominance ratio corresponding to the current parameter tuning cycle. Here, the state-labeled input sequence refers to a sequence object arranged in the order of the parameter tuning cycle, where each sequence element simultaneously contains a parameter tuning cycle-level input sample, a state differentiation vector, and a state constraint label; the control action dominance ratio, thermal drift hysteresis dominance ratio, and channel mismatch propagation dominance ratio are the state constraint labels obtained after normalization in step 32.

[0056] It should be noted that in this step, the system reads not only the input sequence itself, but also the three dominant proportions corresponding to the current parameter tuning period. This is because the subsequent fusion process of the three types of memory features is not a fixed proportion fusion, but a differentiated processing based on the state constraint label of the current parameter tuning period.

[0057] The system inputs the state-marked input sequence into the first, second, and third bidirectional long short-term memory (LSTM) network coding branches, respectively, to obtain control action memory features, hot drift memory features, and channel mismatch memory features, respectively. These three branches are three independently configured temporal coding paths. In other words, the system does not split the input sequence into three different sequences for separate processing; instead, it sends the same state-marked input sequence into three different coding branches, allowing each branch to learn different types of change patterns from the same sequence.

[0058] The system employs a three-tiered encoding approach: the first bidirectional long short-term memory (LSTM) branch focuses on characterizing the temporal impact of control setting changes in adjacent tuning cycles; the second branch focuses on characterizing the cumulative impact of thermal state changes across multiple tuning cycles; and the third branch focuses on characterizing the propagation impact of channel mismatch in the power chain. This approach results in three distinct memory features for the same state-labeled input sequence across the three encoding paths. For example, when control setting changes rapidly while thermal state changes slowly during a tuning process, the first LSM branch is more likely to retain the temporal differences related to control setting changes, while the second branch is more likely to retain information about the continued changes in thermal state. Through this processing, the system can extract memory results for control actions, thermal drift, and channel mismatch separately from the same input sequence, rather than mixing all state changes into a single memory result.

[0059] The system multiplies the control action memory feature by the control action dominance ratio, the thermal drift memory feature by the thermal drift hysteresis dominance ratio, and the channel mismatch memory feature by the channel mismatch conduction dominance ratio. These three products are then added together to obtain the comprehensive memory feature. The comprehensive memory feature can be expressed as: ,in, Indicates the first The comprehensive memory features corresponding to each parameter tuning cycle Indicates the first The proportion of control actions dominant in each parameter tuning cycle. Indicates the first The proportion of thermal drift hysteresis dominance corresponding to each parameter tuning cycle. Indicates the first The proportion of channel mismatch propagation dominance corresponding to each parameter tuning cycle. Indicates the first The memory characteristics of control actions corresponding to each parameter tuning cycle Indicates the first Thermal drift memory characteristics corresponding to each parameter tuning cycle Indicates the first The channel mismatch memory features correspond to each tuning cycle. The memory features that the current tuning cycle is closer to the dominant state will have a higher proportion in the overall memory features. This preserves the feature information extracted by each of the three coding branches and ensures that the final output overall memory features reflect the dominant state relationship of the current tuning cycle.

[0060] Through the above steps, the system converts the state-labeled input sequence into control action memory features, hot drift memory features, and channel mismatch memory features via three bidirectional long short-term memory network encoding branches. These features are then fused under the influence of the state constraint label corresponding to the current parameter tuning cycle, resulting in a comprehensive memory feature. This processing preserves the temporal representation capabilities of different state sources while making the comprehensive memory feature more closely reflect the actual dominant state of the current parameter tuning cycle. This provides a unified and stable feature input for subsequent extraction of parameter tuning context representations based on a state-constrained attention-weighted structure.

[0061] In one embodiment of the present invention, after obtaining the comprehensive memory features, the system further combines state constraint labels and obtains a parameter tuning context representation through a state-constrained attention weighting structure. Here, the parameter tuning context representation refers to the context feature results obtained by filtering and aggregating from historical parameter tuning cycles under the current parameter tuning cycle. It is not a simple superposition of all historical information, but a feature representation formed under the joint constraints of the similarity of comprehensive memory features and the state dominance relationship. Through the above processing, the system retains historical memories related to the current parameter tuning cycle while suppressing historical segments that deviate significantly from the current dominant state, thereby providing a unified input for subsequent field power prediction values, incident power prediction values ​​for the four channels, and reflected power prediction values ​​for the four channels.

[0062] Specifically, the system reads the comprehensive memory features corresponding to the current tuning cycle and the comprehensive memory features corresponding to the historical tuning cycles, as well as the proportions of control actions, thermal drift hysteresis, and channel mismatch propagation corresponding to the current and historical tuning cycles. Here, the historical tuning cycle refers to a preset number of consecutive tuning cycles selected backward from the current tuning cycle. In other words, under the current tuning cycle, the system not only reads the current comprehensive state result but also simultaneously reads the memory results and state constraint results within the same historical window. This ensures that the objects being compared subsequently are within the same time frame and that subsequent state bias calculations and memory comparisons are based on the same batch of tuning cycle objects.

[0063] The system determines the query vector based on the comprehensive memory features corresponding to the current parameter tuning period, and determines the key vector and value vector based on the comprehensive memory features corresponding to each historical parameter tuning period. Here, the query vector refers to the feature object representing the retrieval requirement of the current parameter tuning period; the key vector refers to the feature object that can be matched in each historical parameter tuning period; and the value vector refers to the feature object that actually participates in the aggregation after being selected in a historical parameter tuning period. In other words, the query vector indicates what kind of memory the current parameter tuning period hopes to find from history, the key vector indicates what memory features each historical parameter tuning period has, and the value vector indicates the content that each historical parameter tuning period can ultimately provide to the current parameter tuning period. Specifically, the query vector is determined by combining the query vector mapping weight with the comprehensive memory features corresponding to the current parameter tuning period; the key vector is determined by combining the key vector mapping weight with the comprehensive memory features corresponding to the historical parameter tuning period; and the value vector is determined by combining the value vector mapping weight with the comprehensive memory features corresponding to the historical parameter tuning period. Through the above processing, the system further transforms the original comprehensive memory features into three types of vector objects suitable for attention weighting.

[0064] The system multiplies the proportions of control actions as the dominant force in the current tuning cycle and the proportions of thermal drift lag as the dominant force, and the proportions of channel mismatch propagation as the dominant force. These three products are then weighted and summed according to their respective preset adjustment coefficients to obtain the state constraint bias term. This state constraint bias term refers to the overall consistency between the current tuning cycle and a certain historical tuning cycle in terms of the three dominant state relationships. If the distributions of control action dominance, thermal drift lag dominance, and channel mismatch propagation dominance between a certain historical tuning cycle and the current tuning cycle are closer, the corresponding state constraint bias term is higher; conversely, it is lower. This approach considers both the similarity of historical memories and the similarity between historical states and the current state, ensuring that the subsequent weighting result depends not only on the features themselves but also on the constraints of state relationships.

[0065] The system calculates the similarity by performing an inner product between the query vector and the key vectors corresponding to each historical tuning period. This similarity represents the closeness in feature space between the query vector for the current tuning period and the key vector for a given historical tuning period. A higher similarity indicates that the historical tuning period is closer to the current tuning period in terms of overall memory. Subsequently, the system divides the similarity by the square root of the attention dimension and adds it to the state constraint bias term to obtain the initial attention value. This initial attention value represents the initial weight base after considering both memory closeness and state consistency. Then, each initial attention value is exponentially normalized to obtain the attention weight. This attention weight represents the proportion of contribution made by each historical tuning period in the context construction process of the current tuning period.

[0066] Finally, the system multiplies the value vectors corresponding to each historical parameter tuning period with the corresponding attention weights and sums them to obtain the parameter tuning context representation. This parameter tuning context representation can be expressed as: ,in, This represents the tuning context corresponding to the current tuning cycle. This represents the query vector corresponding to the current parameter tuning cycle. This represents the key vector corresponding to the historical parameter tuning period. This represents the key vector corresponding to the nth historical parameter tuning period during the normalization summation process. This represents the state constraint bias term between the current parameter tuning period k and the historical parameter tuning period j. This represents the state constraint bias term between the current parameter tuning period k and the historical parameter tuning period n. Represents the attention dimension. Indicates a consecutive preset quantity. This represents the value vector corresponding to the historical parameter tuning period, k represents the index of the current parameter tuning period, j represents the index of the historical parameter tuning period currently involved in the attention weight calculation, and n represents the index of the historical parameter tuning period used in the normalization summation process. Both of these correspond to the historical parameter tuning periods in the previous preset number of parameter tuning periods. exp represents the exponential function, and T represents the transpose operation. The parameter tuning context representation refers to the unified result formed after filtering, weighting, and aggregating the feature information from the historical parameter tuning periods that is more consistent with the current state and closer to the current memory in the current parameter tuning period.

[0067] After the above steps, the system further transforms the integrated memory features and state constraint labels into a parameter tuning context representation. This not only preserves the memory information that is truly relevant to the current parameter tuning period from the historical parameter tuning periods, but also avoids introducing historical results with large state differences into the subsequent prediction process, thereby making the subsequent power chain prediction closer to the current parameter tuning state.

[0068] In one embodiment of the present invention, after obtaining the parameter tuning context representation, the system further inputs the parameter tuning context representation into a chain-consistent multi-branch output structure to obtain the predicted values ​​of the establishment power, the predicted values ​​of the incident power of the four channels, and the predicted values ​​of the reflected power of the four channels for the next parameter tuning cycle. The chain-consistent multi-branch output structure refers to a structure that, based on the same parameter tuning context representation, outputs along the establishment power prediction path, the incident power prediction path, and the reflected power prediction path respectively, and then continues to perform consistency checks on the correlation between the three types of results after output. This not only obtains the power chain prediction results for the next parameter tuning cycle but also preserves the sequential and distributional relationships within the power chain, providing a more complete basis for subsequent control setting back-calculation.

[0069] Specifically, the system inputs the parameter tuning context representation into the establishment power prediction mapping network, the incident power prediction mapping network, and the reflection power prediction mapping network, respectively, to obtain the establishment power prediction value, the incident power prediction value of the four channels, and the reflection power prediction value of the four channels for the next parameter tuning cycle. Here, the establishment power prediction value refers to the overall establishment result for the next parameter tuning cycle calculated by the system based on the current parameter tuning state; the incident power prediction value of the four channels refers to the system's estimate of the result of the four channels on the power injection side for the next parameter tuning cycle; and the reflection power prediction value of the four channels represents the system's estimate of the result of the four channels on the reflection side for the next parameter tuning cycle. Through the above processing, a single parameter tuning context representation is transformed into three types of power prediction results for the next parameter tuning cycle. That is, subsequent steps no longer process abstract feature objects, but directly process the power objects corresponding to the parameter tuning results.

[0070] The system sums the predicted incident power values ​​of the four channels and the predicted reflected power values ​​of the four channels, then subtracts the latter from the former to obtain the net transmitted power prediction. This net transmitted power prediction represents the actual power level transmitted along the power chain in the next parameter tuning cycle, after deducting reflected backflow. Subsequently, the system subtracts the predicted field power value from the product of the net transmitted power prediction and the preset mapping coefficient, taking the absolute value to obtain the field-established power consistency residual. This residual represents the degree of deviation between the predicted field power value and the net power chain transmission result. A smaller residual indicates a more consistent relationship between the predicted field power value and the overall power chain transmission relationship; a larger residual indicates a significant deviation between the two. Through this process, the system compares the field-established result and the overall power chain transmission result under the same verification relationship.

[0071] The system subtracts the predicted incident power of the corresponding channel from the predicted reflected power of each channel, retains the positive-zero differences, and sums them to obtain the channel order consistency residual. This channel order consistency residual represents the degree of anomaly in the power order relationship between channels. More specifically, when the predicted reflected power of a channel is higher than the predicted incident power of the corresponding channel, that channel contributes to the channel order consistency residual; if the predicted reflected power is not higher than the predicted incident power of the corresponding channel, then that channel does not contribute additional accumulation to this residual. Through this processing, the system can extract the power order relationship within each channel separately, rather than relying solely on judgments at the total power level.

[0072] The system averages the predicted incident power values ​​from the four channels, calculates the absolute value of the difference between the predicted incident power value of each channel and the average value, and sums them to obtain the incident distribution equilibrium residual. The incident distribution equilibrium residual can be expressed as: ,in, This represents the incident distribution equilibrium residual corresponding to the next parameter tuning period. Indicates the next parameter tuning cycle number Predicted incident power values ​​for each channel This represents the predicted incident power values ​​for the four channels. The incident distribution balance residual here represents the degree of dispersion of the incident power distribution among the four channels. If the predicted incident power values ​​of the four channels are closer to the average level, the residual is small; if one channel is significantly higher or lower than the others, the residual is larger. Subsequently, the system establishes a correspondence between the field establishment power prediction values, the predicted incident power values ​​of the four channels, the predicted reflected power values ​​of the four channels, the field establishment power consistency residual, the channel sequence consistency residual, and the incident distribution balance residual.

[0073] After this step, the system not only retains the three types of power prediction results for the next parameter tuning cycle, but also simultaneously retains the corresponding three types of consistency verification results. When performing control setting back-calculation subsequently, the total power relationship, single-channel sequence relationship, and multi-channel distribution relationship can be referenced simultaneously. The resulting prediction results have a more complete structure and are closer to the way the overall state of the power chain is judged during actual parameter tuning.

[0074] In one embodiment of the present invention, after obtaining the parameter tuning context representation, the establishment power prediction value, the incident power prediction value of multiple channels, and the reflection power prediction value of multiple channels, the system further combines the target establishment power to determine the gain settings, phase settings, feed settings, and feedforward settings of multiple channels for the next parameter tuning cycle. Here, the target establishment power refers to the establishment target expected to be achieved in the next parameter tuning cycle; the establishment power prediction value, the incident power prediction value of multiple channels, and the reflection power prediction value of multiple channels represent the system's estimate of the power chain result for the next parameter tuning cycle under the current parameter tuning state. Through subsequent processing, the system further converts the power chain prediction result into directly executable control settings, thereby seamlessly connecting the prediction process and the parameter tuning process.

[0075] Specifically, the system reads the tuning context representation, multiple channel gain settings, multiple channel phase settings, feed settings, and multiple channel feedforward settings corresponding to the current tuning cycle, as well as the establishment power prediction, incident power prediction, reflection power prediction, and target establishment power corresponding to the next tuning cycle. The tuning context representation here represents the comprehensive context result formed under the filtering of historical tuning cycles and state constraints for the current tuning cycle. The multiple channel gain settings, multiple channel phase settings, feed settings, and multiple channel feedforward settings corresponding to the current tuning cycle serve as the update benchmark for generating the control settings for the next tuning cycle. After this step, the state object, prediction object, target object, and benchmark object required for subsequent control setting back calculation are uniformly aggregated into the same processing chain.

[0076] The system subtracts the predicted establishment power from the target establishment power to obtain the establishment power deviation. This deviation represents the magnitude and direction of the difference between the target and predicted establishment power levels. A large deviation indicates a significant deviation from the target. Next, the system averages the predicted incident power values ​​from multiple channels and subtracts the predicted incident power value from each channel's average to obtain the incident balance deviation for each channel. This deviation represents the degree of deviation in incident power distribution for each channel; in other words, it reflects whether a channel is higher or lower than the overall average level. Then, the system uses the predicted reflection power for each channel as the reflection suppression deviation. This deviation represents the degree to which the corresponding channel needs to be suppressed on the reflection side. Through these processes, the system further transforms the power chain prediction results for the next parameter tuning cycle into three types of deviations for control backpropagation: deviations at the establishment level, deviations at the inter-channel balance level, and deviations at the single-channel reflection level.

[0077] The system inputs the parameter tuning context representation, establishment power deviation, incident balance deviation of each channel, and reflection suppression deviation of each channel into the control parameter back-mapping network to obtain multiple channel gain corrections, multiple channel phase corrections, feed-in corrections, and multiple channel feedforward corrections. The control parameter back-mapping network represents a network structure that maps the current tuning state and the deviation object of the next tuning cycle to control corrections. In one implementation, a multi-layer fully connected regression network can be used. A multi-layer fully connected regression network is a regression network structure that includes at least an input layer, a hidden layer, and an output layer. The input layer receives the back-mapping input vector, the hidden layer jointly represents the state information and deviation information from different sources, and the output layer provides correction results corresponding one-to-one with each control setting. Multiple channel gain corrections represent the magnitude of adjustment required for each channel gain setting, multiple channel phase corrections represent the magnitude of adjustment required for each channel phase setting, the feed-in correction represents the magnitude of adjustment required for the feed-in setting, and multiple channel feedforward corrections represent the magnitude of adjustment required for each channel feedforward setting. In other words, after this step, the system no longer remains at the deviation judgment level but further transforms the deviation object into a specific control correction object.

[0078] The system adds the corresponding correction values ​​to the multiple channel gain settings, multiple channel phase settings, feed settings, and multiple channel feedforward settings for the current parameter tuning cycle, respectively, to obtain the multiple channel gain settings, multiple channel phase settings, feed settings, and multiple channel feedforward settings for the next parameter tuning cycle. This addition update means that the control settings from the previous parameter tuning cycle are used as a basis, and the correction results obtained from the current back-calculation are added to form the final control settings for the next parameter tuning cycle. Through this step, the system completes the transformation from the parameter tuning context representation and power chain prediction results to the control settings for the next parameter tuning cycle.

[0079] Through the above steps, the system integrates the status information of the current tuning cycle, the power prediction result of the next tuning cycle, and the target establishment power into the control backpropagation process. First, it generates the establishment power deviation, the incident balance deviation of each channel, and the reflection suppression deviation of each channel. Then, the control parameter backpropagation mapping network generates correction values ​​and updates the control settings for the next tuning cycle. This establishes a clear correspondence between the tuning context representation formed in the previous stage and the control settings formed in the next stage. When executing control in the next tuning cycle, the updated gain settings, phase settings, feed settings, and feedforward settings for multiple channels can be directly invoked.

[0080] The present invention also provides a low-level control parameter prediction system for particle accelerators, comprising:

[0081] The sample construction module is used to acquire low-level control-related data of multiple consecutive tuning cycles during the open-loop tuning phase of the particle accelerator, encapsulate them according to the tuning cycle, and obtain a tuning cycle-level input sample set. The tuning cycle-level input sample set includes control setpoints, thermal state parameters, and power chain result parameters.

[0082] The state decomposition module is used to determine the control action influence, thermal drift hysteresis and channel mismatch conduction for each tuning cycle based on the input samples of adjacent tuning cycle levels.

[0083] The state labeling module is used to determine the state differentiation vector and state constraint label corresponding to each parameter tuning cycle based on the influence of the control action, the thermal drift hysteresis, and the channel mismatch conduction, and to construct the input sequence with state labels.

[0084] The branch coding module is used to input the state-marked input sequence into the first bidirectional long short-term memory network coding branch, the second bidirectional long short-term memory network coding branch, and the third bidirectional long short-term memory network coding branch respectively to obtain control action memory features, hot drift memory features, and channel mismatch memory features, and then fuse them to obtain comprehensive memory features;

[0085] The context extraction module is used to obtain a parametric context representation based on the comprehensive memory features and state constraint labels through a state-constrained attention weighting structure;

[0086] The power prediction module is used to input the parameter tuning context representation into a chain-consistent multi-branch output structure to obtain the field establishment power prediction value, the incident power prediction value of multiple channels, and the reflection power prediction value of multiple channels for the next parameter tuning cycle.

[0087] The parameter back-calculation module is used to determine the gain settings, phase settings, feed settings, and feedforward settings of multiple channels for the next parameter tuning cycle based on the parameter tuning context representation, the establishment power prediction value, the incident power prediction value of multiple channels, the reflection power prediction value of multiple channels, and the target establishment power.

[0088] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0089] The embodiments of the present invention have been described above, but the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of the present embodiments, all of which are within the protection scope of the present embodiments.

Claims

1. A method for predicting low-level control parameters of a particle accelerator, characterized in that, Includes the following steps: Step 1: Obtain low-level control-related data for multiple consecutive tuning cycles during the open-loop tuning phase of the particle accelerator, encapsulate the data according to the tuning cycle, and obtain a tuning cycle-level input sample set. The tuning cycle-level input sample set includes control settings, thermal state parameters, and power chain result parameters. Step 2: Based on the input samples of adjacent parameter tuning cycles, determine the control action influence, thermal drift hysteresis, and channel mismatch conduction for each parameter tuning cycle. Step 3: Based on the influence of the control action, the thermal drift hysteresis, and the channel mismatch conduction, determine the state differentiation vector and state constraint label corresponding to each parameter tuning cycle, and construct the input sequence with state labels; wherein, the state constraint label includes the proportion of control action dominance, the proportion of thermal drift hysteresis dominance, and the proportion of channel mismatch conduction dominance. Step 4: Input the state-marked input sequence into the first bidirectional long short-term memory network coding branch, the second bidirectional long short-term memory network coding branch, and the third bidirectional long short-term memory network coding branch respectively to obtain control action memory features, hot drift memory features, and channel mismatch memory features. Then, under the action of the state constraint mark corresponding to the current parameter tuning cycle, the fusion is completed to obtain the comprehensive memory features. Step 5: Based on the comprehensive memory features and state constraint labels, obtain the parameter tuning context representation through a state-constrained attention weighting structure; Step 6: Input the parameter tuning context representation into a chain-consistent multi-branch output structure to obtain the field establishment power prediction value, the incident power prediction value of multiple channels, and the reflection power prediction value of multiple channels for the next parameter tuning cycle. Step 7: Based on the parameter tuning context representation, the establishment power prediction value, the incident power prediction value of multiple channels, the reflection power prediction value of multiple channels, and the target establishment power, determine the multiple channel gain settings, multiple channel phase settings, feed settings, and multiple channel feedforward settings for the next parameter tuning cycle.

2. The method for predicting low-level control parameters of a particle accelerator according to claim 1, characterized in that, Acquire low-level control-related data for multiple consecutive tuning cycles during the open-loop parameter tuning phase of the particle accelerator, encapsulate the data according to the tuning cycle, and obtain a tuning cycle-level input sample set, including: Step 11: Obtain low-level control related data, wherein the low-level control related data includes control setting data, thermal state data, and power chain result data; according to the start and end positions of each parameter adjustment cycle in the low-level control related data, divide the low-level control related data to obtain the low-level control related data corresponding to each parameter adjustment cycle, and establish the binding relationship between each parameter adjustment cycle and the corresponding data; Step 12: Extract control setting, thermal state and power chain result from the low-level control related data corresponding to each parameter adjustment cycle. The control setting includes multiple channel gain setting, multiple channel phase setting, feed setting and multiple channel feedforward setting. The thermal state includes the cooling state of the left and right sides. The power chain result includes the field power, multiple channel incident power and multiple channel reflected power. Step 13: Establish a correspondence between the control setpoint, thermal state quantity and power chain result quantity corresponding to the same parameter tuning cycle, and combine them according to the preset field order to obtain the parameter tuning cycle level input sample; arrange the parameter tuning cycle level input samples according to the parameter tuning cycle order to obtain the parameter tuning cycle level input sample set.

3. The method for predicting low-level control parameters of a particle accelerator according to claim 1, characterized in that, Based on the input samples from adjacent tuning cycles, determine the control action influence, thermal drift hysteresis, and channel mismatch conduction for each tuning cycle, including: Step 21: Starting from the third parameter tuning cycle, read the current parameter tuning cycle level input sample, the previous parameter tuning cycle level input sample, and the input samples from the previous two parameter tuning cycle levels from the parameter tuning cycle level input sample set, and establish a reference relationship corresponding to the current parameter tuning cycle. Step 22: Based on the control settings in the current tuning cycle input sample and the previous tuning cycle input sample, calculate the absolute value of the corresponding difference between the gain setting value, phase setting value and feedforward setting value of each channel and average them. Calculate the absolute value of the difference between the feed-in setting values. Multiply the above four items by the corresponding preset weights and add them together to obtain the control action influence amount. Step 23: Based on the thermal state quantities in the current tuning cycle level input sample, the previous tuning cycle level input sample, and the two previous tuning cycle level input samples, average the cooling state quantities on the left and right sides of each tuning cycle to obtain the corresponding average thermal state quantity; calculate the absolute value of the difference between the current tuning cycle average thermal state quantity and the previous tuning cycle average thermal state quantity, the absolute value of the difference between the current tuning cycle average thermal state quantity and the two previous tuning cycle average thermal state quantities, and the absolute value of the difference between the current tuning cycle left and right sides cooling state quantities, multiply the above three items by the corresponding preset weights, and then add them together to obtain the thermal drift hysteresis. Step 24: Based on the power chain results in the current parameter tuning cycle input samples, first average the incident power of each channel to obtain the average incident power of multiple channels; then divide the reflected power of each channel by the sum of the incident power of the corresponding channel and the smallest positive number, take the absolute value of the difference between the incident power of each channel and the average incident power of multiple channels, and take the absolute value of the difference between the incident power of each channel and the reflected power of the corresponding channel to obtain the reflection coefficient, the incident power deviation, and the incident-reflection difference of each channel; multiply the above three items by the corresponding preset weights and add them together to obtain the single-channel mismatch conduction of each channel; average the single-channel mismatch conduction of each channel to obtain the channel mismatch conduction. Step 25: Establish the correspondence between the control action influence, thermal drift hysteresis and channel mismatch conduction corresponding to the current parameter tuning cycle, and obtain the state output result corresponding to the current parameter tuning cycle.

4. The method for predicting low-level control parameters of a particle accelerator according to claim 1, characterized in that, Based on the control action influence, thermal drift hysteresis, and channel mismatch conduction, determine the state differentiation vector and state constraint label corresponding to each parameter tuning cycle, and construct the input sequence with state labels, including: Step 31: Combine the control action influence, thermal drift hysteresis and channel mismatch conduction corresponding to each parameter tuning cycle in sequence to obtain the state differentiation vector corresponding to each parameter tuning cycle. Step 32: Map the state differentiation vector corresponding to each parameter tuning cycle to the preset mapping weights and determine the state mapping result together with the preset bias; normalize the state mapping result to obtain the state constraint label corresponding to each parameter tuning cycle. Step 33: Taking the current tuning cycle as the endpoint, read the tuning cycle-level input samples, state differentiation vectors and state constraint labels corresponding to the current tuning cycle and the previous preset number of tuning cycles in the tuning cycle order, and establish the correspondence between the tuning cycle-level input samples, state differentiation vectors and state constraint labels under the same tuning cycle. Step 34: Take the input samples, state differentiation vectors and state constraint labels corresponding to the same tuning cycle in the current tuning cycle and the previous preset number of tuning cycles as sequence elements, arrange them in the order of the tuning cycles, and obtain the input sequence with state labels.

5. The method for predicting low-level control parameters of a particle accelerator according to claim 1, characterized in that, The state-marked input sequence is input into the first, second, and third bidirectional long short-term memory (LSTM) network coding branches, respectively, to obtain control action memory features, hot drift memory features, and channel mismatch memory features. These features are then fused under the state constraint markers corresponding to the current parameter tuning cycle to obtain comprehensive memory features, including: Step 41: Read the input sequence with status markers corresponding to the current parameter tuning cycle, as well as the proportion of control action dominance, thermal drift hysteresis dominance, and channel mismatch conduction dominance corresponding to the current parameter tuning cycle. Step 42: Input the state-marked input sequence into the first bidirectional long short-term memory network coding branch, the second bidirectional long short-term memory network coding branch, and the third bidirectional long short-term memory network coding branch respectively to obtain control action memory features, hot drift memory features, and channel mismatch memory features respectively. Step 43: Multiply the control action memory feature by the control action dominance ratio, multiply the thermal drift memory feature by the thermal drift hysteresis dominance ratio, multiply the channel mismatch memory feature by the channel mismatch conduction dominance ratio, and add the three products to obtain the comprehensive memory feature.

6. The method for predicting low-level control parameters of a particle accelerator according to claim 1, characterized in that, Based on the comprehensive memory features and state constraint labels, a parametric context representation is obtained through a state-constrained attention weighting structure, including: Step 51: Read the comprehensive memory features corresponding to the current parameter tuning cycle and the comprehensive memory features corresponding to the historical parameter tuning cycle, as well as the control action dominance ratio, thermal drift hysteresis dominance ratio, and channel mismatch conduction dominance ratio corresponding to the current parameter tuning cycle and the historical parameter tuning cycle. Step 52: Determine the query vector based on the comprehensive memory features corresponding to the current parameter tuning cycle, and determine the key vector and value vector based on the comprehensive memory features corresponding to each historical parameter tuning cycle. Step 53: Multiply the dominant proportion of control actions corresponding to the current parameter tuning cycle and the historical parameter tuning cycle, multiply the dominant proportion of thermal drift hysteresis, multiply the dominant proportion of channel mismatch propagation, and then sum the three products by weighting them according to the corresponding preset adjustment coefficients to obtain the state constraint bias term. Step 54: Calculate the inner product of the query vector and the key vector corresponding to each historical parameter tuning period to obtain the similarity; divide the similarity by the square root of the attention dimension and add it to the state constraint bias term to obtain the original attention value; perform exponential normalization on each original attention value to obtain the attention weight; multiply the value vector corresponding to each historical parameter tuning period with the corresponding attention weight and sum them to obtain the parameter tuning context representation.

7. The method for predicting low-level control parameters of a particle accelerator according to claim 1, characterized in that, By representing the parameter tuning context as an input chain-consistent multi-branch output structure, the predicted field power, the predicted incident power of multiple channels, and the predicted reflection power of multiple channels for the next parameter tuning cycle are obtained, including: Step 61: Input the parameter tuning context representation into the field power prediction mapping network, the incident power prediction mapping network and the reflection power prediction mapping network respectively to obtain the field power prediction value, the incident power prediction value of multiple channels and the reflection power prediction value of multiple channels for the next parameter tuning cycle. Step 62: Sum the predicted incident power values ​​of multiple channels, sum the predicted reflected power values ​​of multiple channels, subtract the latter from the former to obtain the net transmitted power prediction; subtract the net transmitted power prediction from the product of the preset mapping coefficient and the field-establishing power prediction value, and take the absolute value to obtain the field-establishing power consistency residual. Step 63: Subtract the predicted incident power value of the corresponding channel from the predicted reflected power value of each channel, keep the difference greater than zero and sum them up to obtain the channel sequence consistency residual; Step 64: Average the predicted incident power values ​​of multiple channels, calculate the absolute value of the difference between the predicted incident power value of each channel and the average value, and sum them to obtain the incident distribution balance residual; and establish a correspondence between the predicted field power value, the predicted incident power value of multiple channels, the predicted reflection power value of multiple channels, the field power consistency residual, the channel sequence consistency residual, and the incident distribution balance residual.

8. The method for predicting low-level control parameters of a particle accelerator according to claim 1, characterized in that, Based on the parameter tuning context, the establishment power prediction, the incident power prediction for multiple channels, the reflected power prediction for multiple channels, and the target establishment power, determine the gain settings, phase settings, feed settings, and feedforward settings for multiple channels in the next parameter tuning cycle, including: Step 71: Read the parameter tuning context representation, multiple channel gain settings, multiple channel phase settings, feed settings and multiple channel feedforward settings corresponding to the current parameter tuning cycle, as well as the field establishment power prediction value, multiple channel incident power prediction value, multiple channel reflection power prediction value and target field establishment power corresponding to the next parameter tuning cycle. Step 72: Subtract the predicted establishment power from the target establishment power to obtain the establishment power deviation; average the predicted incident power values ​​of multiple channels, and then subtract the predicted incident power value of each channel from the average value to obtain the incident balance deviation of each channel; use the predicted reflection power value of each channel as the reflection suppression deviation of each channel. Step 73: Input the parameter tuning context representation, field establishment power deviation, incident balance deviation of each channel, and reflection suppression deviation of each channel into the control parameter back-mapping network to obtain multiple channel gain corrections, multiple channel phase corrections, feed corrections, and multiple channel feedforward corrections; Step 74: Add the multiple channel gain settings, multiple channel phase settings, feed settings and multiple channel feedforward settings corresponding to the current parameter tuning cycle to the corresponding correction values ​​to obtain the multiple channel gain settings, multiple channel phase settings, feed settings and multiple channel feedforward settings for the next parameter tuning cycle.

9. A low-level control parameter prediction system for a particle accelerator, characterized in that, The method for predicting low-level control parameters of a particle accelerator as described in any one of claims 1-8 includes: The sample construction module is used to acquire low-level control-related data of multiple consecutive tuning cycles during the open-loop tuning phase of the particle accelerator, encapsulate them according to the tuning cycle, and obtain a tuning cycle-level input sample set. The tuning cycle-level input sample set includes control setpoints, thermal state parameters, and power chain result parameters. The state decomposition module is used to determine the control action influence, thermal drift hysteresis and channel mismatch conduction for each tuning cycle based on the input samples of adjacent tuning cycle levels. The state labeling module is used to determine the state differentiation vector and state constraint label corresponding to each parameter tuning cycle based on the influence of the control action, the thermal drift hysteresis, and the channel mismatch conduction, and to construct the input sequence with state labels; wherein, the state constraint label includes the control action dominance ratio, the thermal drift hysteresis dominance ratio, and the channel mismatch conduction dominance ratio. The branch coding module is used to input the state-marked input sequence into the first bidirectional long short-term memory network coding branch, the second bidirectional long short-term memory network coding branch, and the third bidirectional long short-term memory network coding branch respectively to obtain control action memory features, hot drift memory features, and channel mismatch memory features, and to complete the fusion under the action of the state constraint mark corresponding to the current parameter tuning cycle to obtain comprehensive memory features; The context extraction module is used to obtain a parametric context representation based on the comprehensive memory features and state constraint labels through a state-constrained attention weighting structure; The power prediction module is used to input the parameter tuning context representation into a chain-consistent multi-branch output structure to obtain the field establishment power prediction value, the incident power prediction value of multiple channels, and the reflection power prediction value of multiple channels for the next parameter tuning cycle. The parameter back-calculation module is used to determine the gain settings, phase settings, feed settings, and feedforward settings of multiple channels for the next parameter tuning cycle based on the parameter tuning context representation, the establishment power prediction value, the incident power prediction value of multiple channels, the reflection power prediction value of multiple channels, and the target establishment power.

Images