Automatic identification method and system for steady operation stage of gyrotron and power prediction method

Abstract

The application provides a gyrotron steady-state operation stage automatic identification method, and belongs to the field of electron cyclotron resonance heating systems. Control signals and power waveforms of single discharge of a gyrotron are read, and a search range of a steady-state stage is positioned on the power waveforms based on control signal parameters. In the search range, sliding window scanning is performed, the relative standard deviation of power in each sliding window is calculated, and sliding windows with a relative standard deviation less than a threshold value are marked as stable windows. At least N one continuous stable window is identified as a continuous stable section, N a set number, and a continuous stable section with a total steady-state duration longer than a set duration is identified as a steady-state section. A steady-state section with the longest total steady-state duration is identified as a gyrotron steady-state operation stage. The application also provides a gyrotron steady-state operation stage automatic identification system and a gyrotron power prediction method. The problems of high cost, low efficiency and poor universality of a gyrotron steady-state identification method are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electron cyclotron resonance heating system, and particularly relates to a gyrotron steady-state operation stage automatic identification method and system and a power prediction method. BACKGROUND

[0002] As a high-power microwave source providing 1-2 MW power and 110-170 GHz frequency, the gyrotron is the core component of the electron cyclotron resonance heating system in the magnetic confinement nuclear fusion device. With the development of the International Thermonuclear Experimental Reactor (ITER) and the domestic fusion experimental device towards long pulse and high parameter operation, the single discharge time has exceeded hundreds of seconds or even thousands of seconds, and the operation reliability of the gyrotron has become a key bottleneck factor restricting the experimental process and plasma control ability. During the long pulse operation of the gyrotron, various failure modes may occur, including radio frequency oscillation instability, internal sparking, beam abnormality, power reflection and the like. These failures often lead to the forced termination of the experiment, seriously affecting the experimental efficiency and scientific output of the device. M. Thumm et al. found that the gyrotron failure is one of the main reasons for the unplanned shutdown of the long pulse electron cyclotron resonance heating (ECRH) system, and in some experimental activities, the experiment termination ratio caused by the gyrotron related problems is as high as more than 18%.

[0003] In order to improve the reliability and availability of the gyrotron system, researchers at home and abroad have carried out a large amount of failure prediction and diagnosis research work. In recent years, with the rapid development of machine learning and artificial intelligence technology, the application of neural network, support vector machine, ensemble learning and other algorithms to the gyrotron failure prediction has become a research hotspot. For example, J. She et al. used a deep learning algorithm to predict the failure of the gyrotron system on a test bench. N. I. Sapankevych et al. systematically summarized the application of support vector machine in time series prediction. H. B. Zhang et al. proposed a data-driven gyrotron abnormal radio frequency oscillation diagnosis method. These researches have promoted the development of intelligent monitoring technology of the gyrotron to a certain extent.

[0004] However, the existing technology still has the following significant problems and deficiencies:

[0005] (1) Existing machine learning prediction methods usually directly use the complete discharge record data when training the model, and fail to distinguish different physical stages of the cyclotron operation. In fact, a complete discharge process of the cyclotron contains multiple stages with completely different physical characteristics: the start-up transient stage usually lasts for 1-5 seconds, during which the parameters change rapidly, the power rises rapidly, and it behaves as a nonlinear mode competition process, which is extremely sensitive to initial conditions; the steady-state operation stage can last for tens to hundreds of seconds, during which the power fluctuates within ±2%, mainly controlled by the deterministic heat diffusion process. These two stages are dominated by different time scales and physical processes. Mixing them together to train the prediction model actually forces the algorithm to learn both transient dynamics and steady-state behavior, which are two completely different physical mechanisms, which will inevitably lead to a decline in model performance and prediction accuracy. However, in the published related research, there is no work on systematic evaluation of training strategies based on operation phase filtering, so it is impossible to determine the actual contribution of data quality improvement relative to algorithm architecture optimization.

[0006] (2) To realize the training based on the operation phase filtering, it is necessary to accurately identify the steady-state stage, but the existing steady-state identification methods have obvious limitations. There are mainly two types of steady-state identification methods:

[0007] One type is the threshold-based criterion method, such as using efficiency, reflection coefficient, and other derived parameters for judgment, but this method needs tedious empirical calibration for each power level and operation mode, and lacks universality and self-adaptive ability; the other type is the method relying on additional diagnostic equipment, such as using infrared thermal imaging, spectral diagnosis, and other special equipment, but this will significantly increase the system integration cost and technical complexity, and is difficult to implement on many existing devices. In addition, for the cyclotron test platform that generates continuous data streams with kilohertz sampling rate, the accumulated experimental data can reach thousands of discharges in months to years. Manual labeling of steady-state stages not only consumes a lot of time and costs, but also has subjective and inconsistent problems between different annotators. Therefore, there is an urgent need for a method that can automatically and accurately identify the steady-state stage without additional instruments.

[0008] (3) In recent years, the research trend in the field of machine learning has overemphasized the complexity and refinement of algorithm architecture, such as using deeper neural networks, introducing attention mechanisms, and using meta-heuristic optimization algorithms (such as whale optimization algorithm, bird swarm algorithm, bat algorithm, etc.) for hyperparameter tuning. However, few studies have systematically compared the impact of data quality improvement (such as training data filtering based on physical stages) and algorithm optimization (such as hyperparameter tuning) on prediction performance under actual operating conditions. When faced with complex engineering systems with heterogeneous operating mechanisms, the fundamental question of whether to prioritize resource investment in data management infrastructure or to prioritize the complexity improvement of algorithm architecture still lacks quantitative empirical evidence. If this is not clear, it may lead to the misallocation of research and development resources, affecting the direction and efficiency of technological development.

[0009] (4) For prediction models deployed in integrated control systems, strict real-time processing constraints must be met. Typical auxiliary heating control systems for fusion require a response delay of less than fifty milliseconds to achieve online fault prediction and timely preventive interventions. However, many existing computationally intensive methods, especially those using complex network architectures or meta-heuristic optimization, often struggle to meet this real-time requirement, limiting their deployment and application value in actual control systems.

[0010] In summary, the existing technology has many shortcomings in the identification of steady-state and power prediction of gyrotrons, and there is an urgent need to develop an innovative technical solution that can utilize existing control signals, without the need for additional instruments, automatically and accurately identify steady-state phases, meet real-time processing requirements, and systematically quantify the performance improvement of data quality improvement relative to algorithm optimization. Such technology not only has important significance for improving the reliability of gyrotron systems, but also provides valuable technical reference and methodological guidance for other complex engineering systems with heterogeneous operating mechanisms. SUMMARY

[0011] The technical problem to be solved by the present application is how to solve the problem of lack of self-adaptive ability or dependence on additional diagnostic equipment, which leads to time-consuming, high cost, and low efficiency in the existing gyrotron steady-state identification method.

[0012] The present application solves the above technical problems by the following technical solution: a gyrotron steady-state operation phase automatic identification method, the method comprising:

[0013] reading the control signal and power waveform of a single discharge of the gyrotron, positioning the search range of the steady-state phase on the power waveform based on the control signal parameters; performing sliding window scanning within the search range, calculating the relative standard deviation of the power in each sliding window, and marking the sliding window with a relative standard deviation less than the threshold as a stable window; marking at least NA series of consecutive stable windows are identified as consecutive stable segments. N To set the quantity, continuous stable segments with a total steady-state duration exceeding the set duration are identified as steady-state segments, and the steady-state segment with the longest total steady-state duration is identified as the steady-state operation stage of the gyrotube.

[0014] This invention first uses control signal parameters to coarsely locate the steady-state position over a time range, narrowing the computational scope and reducing the risk of misjudgment. Then, a sliding window method is employed to scan the search range, calculating the relative standard deviation of power within each window. Sliding windows with a relative standard deviation less than a threshold are marked as stable windows. A dimensionless relative standard deviation index is used as a power-level-independent stability criterion to perform local stability dynamic scanning of the power waveform within the coarse location range. The dimensionless normalized form forms the mathematical basis for achieving cross-power adaptation. Finally, a continuity constraint is applied, requiring at least... N A continuous stable window is identified as a continuous stable segment, and a continuous stable segment with a total steady-state duration exceeding a set duration is identified as a true steady-state segment, excluding random fluctuations. This invention achieves adaptive identification across power and time through dimensionless normalization and physical constraints, without relying on additional diagnostic equipment. The identification speed is 84 times faster than manual annotation, solving the problems of high cost, low efficiency, and poor versatility of gyrotron steady-state identification methods.

[0015] Preferably, the process of locating the search range of the steady-state stage on the power waveform based on the control signal parameters includes: determining the total pulse duration by the control signal parameters, determining the proportion of the search range to the total pulse duration based on the actual duration of the steady-state stage of the gyrotube, and locating the search range of the steady-state stage on the power waveform based on the proportion.

[0016] Preferably, the actual duration of the steady-state phase of the gyrotube accounts for 30% to 70% of the total pulse duration, and the search range accounts for 15% to 85% of the total pulse duration.

[0017] Based on physical experience and statistical analysis of extensive experimental data, this invention indicates that the actual duration of the steady-state phase of a gyrotron typically accounts for 30% to 70% of the total pulse duration. Therefore, this invention extends the search range to the total pulse duration based on this proportion. The percentage is set to 15% to 85%. Based on this percentage, the search range for the steady-state phase is located on the power waveform. By narrowing the search range for the steady-state phase to 70% of the middle region of the power pulse, a coarse location of the steady-state phase time range is achieved. This effectively avoids unnecessary calculations during the transient phases of startup and shutdown, improving algorithm efficiency and reducing the risk of misjudgment.

[0018] Preferably, the total number of sliding windows within the search range is:

[0019]

[0020] wherein, is the length of the search range, is the size of the sliding window, is the step size of the sliding window scan, denotes the floor function.

[0021] Preferably, the relative standard deviation of the power within each sliding window is:

[0022]

[0023] wherein, is the first sliding window, is the relative standard deviation of the power within the sliding window, is the total number of sliding windows within the search range, is the first sliding window, is the mean of all powers within the first sliding window,

[0024] The present application uses the relative standard deviation of the power within the sliding window as the power stability criterion. First, it eliminates the influence of the absolute value of the power, so that the same threshold can be applied to operation at different power levels. Second, as a relative index, the relative standard deviation automatically normalizes the noise effect related to signal amplitude, avoiding the problem that when using the absolute standard deviation as the criterion, high-power operation will be misjudged as unstable due to the increase in the absolute value of the noise. When the relative standard deviation is less than the threshold, it indicates that the sum of the contributions of the intrinsic fluctuations of the physical process and the measurement noise of the measurement system is within an acceptable small range, and the macroscopic state of the system is in quasi-equilibrium, satisfying the physical definition of steady-state operation. On the contrary, in the transient stage, the rapid change in power far exceeds the contributions of the intrinsic fluctuations and the measurement noise, resulting in a significant increase in the relative standard deviation RSD, usually more than ten percent. Therefore, by setting an appropriate relative standard deviation threshold, the transient and steady-state two physical stages can be effectively distinguished.

[0025] Preferably, the size of the sliding window is 1000 points corresponding to a time length of 0.1 seconds, and the step size of the sliding window scan is 100 points corresponding to 0.01 seconds, and the threshold value ranges from 2% to 5%, and the setting time ranges from 0.2 times to 0.5 times of the total duration of the pulse.

[0026] Preferably, the size of the sliding window For 1000 points corresponding to 0.1 seconds of time length, first, the window is large enough to contain multiple measurement noise periods. For a 10kHz sampling frequency, 1000 points contain a complete statistical sample, ensuring the reliability of the RSD estimate and preventing false positives due to accidental noise spikes; second, the window is small enough to capture second-level state transitions, meeting the need for accurate positioning of transient and steady-state boundaries, as the transition from startup transient to steady state is usually completed within a few seconds. The step size of the sliding window scan For 100 points corresponding to 0.01 seconds, the step size achieves a good balance between time resolution and computational efficiency, providing sufficient time resolution to accurately capture the boundary position while avoiding redundant calculations caused by a small step size, allowing the algorithm to run in real time on a regular workstation.

[0027] Preferably, the steady-state start point of the gyrotron steady-state running phase is the start time of the first sliding window of the steady-state segment with the longest total steady-state length, and the steady-state end point of the gyrotron steady-state running phase is the end time of the last sliding window of the steady-state segment with the longest total steady-state length.

[0028] Preferably, the zero time of the power waveform to the steady-state start point is the startup transient phase of the gyrotron discharge, the steady-state start point to the steady-state end point is the steady-state running phase of the gyrotron discharge, and the steady-state end point to the end time of the pulse total duration is the shutdown transient phase of the gyrotron discharge. The startup transient phase, steady-state running phase and shutdown transient phase together form a U-shaped pulse waveform.

[0029] The present application determines the steady-state start point and the steady-state end point according to the gyrotron steady-state running phase, accurately positions the steady-state boundary, and then divides the complete discharge process into the startup transient phase, steady-state running phase and shutdown transient phase based on the steady-state boundary. By analyzing the U-shaped pulse waveform, the accurate time boundary of the three phases and the statistical characteristic parameters of each phase can be obtained. This clear division based on physical phases provides high-quality data screening basis for subsequent data analysis, model training and fault diagnosis.

[0030] The present application also provides a gyrotron steady-state running phase automatic identification system, which comprises:

[0031] The coarse positioning module is used to read the control signal and power waveform of a single discharge of the gyrotron, and to position the search range of the steady-state phase on the power waveform based on the control signal parameters;

[0032] The fine positioning module is used to perform sliding window scanning within the search range, calculate the relative standard deviation of the power in each sliding window, and mark the sliding window with a relative standard deviation less than a threshold value as a stable window;

[0033] The steady-state identification module is used to mark at least one stable window as a steady-state segment, and to determine the steady-state start point and the steady-state end point of the gyrotron steady-state running phase based on the steady-state segment with the longest total steady-state length.N a continuous stable window is identified as a continuous stable section, N a set number is set, and a continuous stable section with a total stable time length exceeding the set time length is identified as a steady state section, and a steady state section with the longest total stable time length is identified as a gyrotron steady state operation phase.

[0034] The present application also provides a gyrotron power prediction method based on the gyrotron steady state operation phase automatic identification method, which comprises: preprocessing the control parameters of the gyrotron steady state operation phase to obtain input features; training a power prediction model with the input features as input and the gyrotron output power as output label to obtain the trained power prediction model; and inputting the control parameters of the running gyrotron into the trained power prediction model to predict the gyrotron output power.

[0035] The gyrotron power prediction method uses steady state phase data to train the power prediction model, which requires the model to learn the slowly changing quasi-equilibrium behavior, capture the small amplitude stable fluctuations of the power, and only need to learn a single physical mechanism, i.e. steady state dynamics. Considering that the power prediction task in practical application is usually carried out under steady state conditions, i.e. predicting the power at a certain time in the steady state phase, the steady state phase data is used as training data, which is consistent with the physical mechanism and statistical distribution of the test data. The model can more effectively learn the internal rules of steady state dynamics, and can avoid problems such as gradient conflict, mode confusion and abnormal error distribution caused by mixing of transient and steady state heterogeneous data.

[0036] The power prediction model trained using steady state phase data has significantly better performance than the model trained using mixed data and the model optimized by algorithm, and does not increase the computational cost of training. It only needs to run a steady state identification algorithm once in the data preparation stage, thereby determining that for complex engineering systems with heterogeneous operation mechanisms, under the condition of limited resources, high-quality data stage identification, screening and cleaning tools should be developed by investing in data management infrastructure first, rather than blindly pursuing the complexity of algorithm architecture and the fine optimization of hyperparameters based on mixed heterogeneous data. This physical-oriented data management methodology can be extended to other engineering systems with heterogeneous operation mechanisms. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The flowchart of the gyrotron steady state operation phase automatic identification method provided for the embodiment 1 of the present application;

[0038] Figure 2 The several-shaped pulse waveform diagram in the gyrotron steady state operation phase automatic identification method provided for the embodiment 1 of the present application;

[0039] Figure 3This is a schematic diagram illustrating the evolution of the relative standard deviation of power within each sliding window in the automatic identification method for steady-state operation of a gyrotube provided in Embodiment 1 of the present invention.

[0040] Figure 4 This is a schematic diagram of the automatic identification system for the steady-state operation stage of a gyrotube provided in Embodiment 2 of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0042] Example 1

[0043] like Figure 1 As shown, this embodiment provides an automatic identification method for the steady-state operation phase of a gyrotube, including the following steps:

[0044] Step 1: Read the control signal and power waveform of a single discharge of the gyrotron from the ECRH control system, and locate the search range of the steady-state stage on the power waveform based on the control signal parameters; the control signal is the Integral Pulse Length Control Signal (LPLEN), and the IPLEN parameter records the preset total pulse duration of the ECRH control system. The power waveform is a time series of output power with a sampling frequency of 10kHz. Output power time series The length is N Equal to the total pulse duration Multiply by 10,000 points. The control signal and power waveform data of a single discharge of the gyrotron are existing data of the ECRH control system, requiring no additional hardware or diagnostic system. This is the basis for the zero-cost identification achieved in this invention.

[0045] The process of locating the search range of the steady-state phase on the power waveform based on control signal parameters includes: determining the total pulse duration from the control signal parameters. The search range is determined as a percentage of the total pulse duration based on the actual duration of the steady-state phase of the gyrotron. According to physical experience and extensive experimental data, the actual duration of the steady-state phase typically accounts for 30% to 70% of the total pulse duration. This invention, based on this percentage, extends the search range to the entire pulse duration. The proportion on the power waveform is set to 15% to 85%. The search range of the steady state stage is positioned according to the proportion. By narrowing the search range of the steady state stage to the region of the middle seventy percent of the power pulse, the coarse positioning of the time range of the steady state stage is realized, which can effectively avoid unnecessary calculation in the transient stage of starting and shutting down, improve the efficiency of the algorithm, and reduce the risk of misjudgment.

[0046] The physical principle of IPLEN as the basis for coarse positioning lies in that the operation process of the gyrotron contains three stages with different physical mechanisms and time scales. The starting transient stage lasts for one to five seconds, is dominated by nonlinear processes such as electromagnetic field establishment and particle beam injection, and the power rapidly rises and fluctuates greatly. The steady state operation stage can last for tens of seconds to hundreds of seconds, is controlled by quasi-equilibrium processes such as heat diffusion, and the power is stable within two percent. The shutdown transient stage lasts for one to three seconds, involves energy dissipation processes, and the power rapidly decreases. The IPLEN parameter reflects the total preset pulse duration of the control system, while the steady state stage as the main part of the pulse, its actual duration usually accounts for thirty to seventy percent of IPLEN. This proportional relationship has relative stability, and remains basically unchanged in a wide range of operation from 0.1 MW to 1.2 MW of different power levels and 0.1 seconds to 500 seconds of different pulse lengths, and does not change significantly with specific operation modes. Therefore, by analyzing IPLEN, the approximate position range of the steady state stage can be automatically inferred, and the search area is limited to the period of about seventy percent of the middle of the pulse, without the need to recalibrate parameters for different power levels or pulse lengths, which is the physical basis for realizing cross-power adaptive recognition. In addition, it realizes zero-cost automatic recognition without any additional instruments and equipment.

[0047] Step 2, sliding window scanning is performed in the search range, the relative standard deviation of the power in each sliding window is calculated, and the sliding window with a relative standard deviation less than a threshold value is marked as a stable window; the threshold value ranges from 2% to 5%, and the threshold value in this embodiment is 5%.

[0048] The total number of sliding windows in the search range which is equal to the length of the search range minus the window size divided by the step size, rounded down, plus 1, that is:

[0049]

[0050] wherein, is the length of the search range, is the size of the sliding window, is the step size of the sliding window scanning, represents rounding down. The size of the sliding window and the step size of the sliding window scanning The size of the sliding window is set according to the need The step length of the sliding window scanning is 0.1 s corresponding to 1000 points The step length of the sliding window scanning is 0.01 s corresponding to 100 points

[0051] The first The first The relative standard deviation of the power in each sliding window is:

[0052]

[0053] Wherein, The relative standard deviation of the power in the first sliding window, , The total number of sliding windows in the search range, The mean of all powers in the first sliding window, The standard deviation of each power in the first sliding window.

[0054] The mean of all powers in the first sliding window is:

[0055]

[0056] Wherein, represents the gyrotron power at the th time point in the first sliding window, The size of the sliding window, The size of the sliding window is 1000 points corresponding to a time length of 0.1 s, .

[0057] The standard deviation of each power in the first sliding window is:

[0058]

[0059] Based on the physical characteristics that the power fluctuation in the steady state stage usually remains within the range of plus or minus two percent, and the total relative standard deviation after the power fluctuation in the steady state stage plus the measurement noise contribution should be less than five percent, the threshold value is selected as 5% in this embodiment. When the relative standard deviation in the sliding window is less than 5%, the sliding window is marked as stable, otherwise it is marked as unstable.

[0060] The mathematical principle of RSD as a stability criterion is reflected in its two key advantages as a dimensionless index:​

[0061] (1) Power level independence. The relative standard deviation is defined as the ratio of the standard deviation to the mean, i.e. RSD equals the standard deviation divided by the mean σ μ This ratio form eliminates the effect of the absolute value of the power, so that the same threshold value can be applied to different power levels of operation.

[0062]

[0063] where, is the standard deviation of all samples, is the mean of all samples, is the mean of all samples, is the value of the i th sample, n is the total number of samples.

[0064] For example, if the power fluctuation is 2 kW at 100 kW operation, the RSD is 2 percent; while if the power fluctuation is 20 kW at 1 MW operation, the RSD is also 2 percent. Although the absolute fluctuation amplitude of the two is ten times different, the RSD values are the same, which indicates that both have the same relative stability, and the physical meaning of the fluctuation level is plus or minus 2 percent. This feature makes the steady state judgment threshold of 5 percent automatically adapt to a wide power range from 0.1 MW to 1.2 MW, spanning an order of magnitude without recalibration, and also applies to different pulse lengths from 0.1 second to 500 seconds, spanning three orders of magnitude of operating conditions.

[0065] (2) Noise normalization feature. In an actual measurement system, the measurement noise is usually proportional to the signal amplitude, i.e. the signal-to-noise ratio remains relatively constant within a certain range. At low power operation, the absolute value of the noise is small, but the proportion relative to the signal remains the same. At high power operation, the absolute value of the noise is large, but the signal-to-noise ratio remains the same. As a relative indicator, RSD automatically normalizes the noise effect related to signal amplitude, avoiding the problem of misjudgment of high power operation as unstable due to the increase in absolute noise value when using absolute standard deviation as a criterion. The power fluctuation in the steady state stage mainly includes two parts of contribution: the intrinsic fluctuation of the physical process, such as the random disturbance caused by thermal fluctuation, and the measurement noise of the measurement system. When RSD is less than 5 percent, it indicates that the sum of the two contributions is within an acceptable small range, and the macroscopic state of the system is in quasi-equilibrium, satisfying the physical definition of steady state operation. On the contrary, in the transient stage, the rapid change of power far exceeds the contribution of intrinsic fluctuation and measurement noise, resulting in a significant increase in RSD, usually more than 10 percent, so that the two physical stages of transient and steady state can be effectively distinguished by setting an appropriate RSD threshold.

[0066] ​The dynamic response principle of the sliding window method lies in calculating the local statistics point by point by moving a fixed-width window on the time axis, realizing the dynamic tracking of the time-varying characteristics of the signal. The sliding window method moves a fixed-width window on the time axis with a fixed step, independently calculates the local statistics such as mean and standard deviation for each window position, obtains the trajectory of the statistics evolution over time, and captures the time-varying characteristics of the signal. This dynamic analysis method has two key advantages over global statistics: first, it can capture the time-varying characteristics of the signal, and for the non-stationary process containing different physical stages such as the discharge of the gyrotron, the global statistics will average the differences of each stage, while the sliding window can identify the local stability change; second, it can accurately locate the time of state transition, by observing the change trajectory of the local RSD from high to low and then to high, the transition boundaries of transient state to steady state and steady state to transient state can be accurately marked.

[0067] The selection of window size needs to balance between time resolution and statistical reliability. The smaller the window, the higher the time resolution can capture faster changes, but the estimation accuracy of the statistics will decrease due to the small number of samples and is easily affected by random noise. The larger the window, the more reliable the statistics estimation, but it will smooth out the rapid changes and reduce the time resolution. The present application selects a window size of 1000 points corresponding to 0.1 seconds, which is based on the following two considerations:

[0068] (1) The window is large enough to contain multiple measurement noise periods. For a sampling frequency of 10 kHz, 1000 points contain a complete statistical sample, ensuring the reliability of RSD estimation and preventing misjudgment due to accidental noise spikes.

[0069] (2) The window is small enough to capture state transitions on the order of seconds, meeting the requirement for accurate positioning of the transient and steady state boundaries, because the transition process from the start-up transient state to the steady state usually takes a few seconds. The step size is set to 100 points corresponding to 10 milliseconds, which means that the window moves forward every 10 milliseconds and recalculates the RSD. This step size setting strikes a good balance between time resolution and computational efficiency, providing a fine enough time resolution to accurately capture the boundary position, while avoiding redundant calculations caused by too small step size, so that the algorithm can run in real time on an ordinary workstation.

[0070] Step 3, applying continuity constraints, identifying at least N consecutive stable windows as a continuous stable segment, N for a set number, N the sequence length of the consecutive stable windows exceeds a set duration, and identifying the continuous stable segment with the longest total steady state duration as the steady state segment, and identifying the steady state segment with the longest total steady state duration as the steady state running phase of the gyrotron.

[0071] Since a single stable window is not enough to determine that the system enters a real steady state stage, it may just be a casual fluctuation or a short-term stability in the transient process, therefore the present application requires N the length of the sequence of continuous stable windows to be greater than or equal to the set duration In the present application, the set number N is 30, that is, at least 30 continuous sliding windows satisfying the stability criterion are required to be identified as a continuous stable section. The continuity constraint can effectively exclude casual power fluctuations and ensure that the identified steady state section has a real physical meaning. 30 continuous stable windows correspond to a duration of at least 3 seconds, and the set duration is 3 seconds. The window sequence length of the continuous stable section is the total steady state duration, and the continuous stable section with a total steady state duration exceeding the set duration is identified as a steady state section. The steady state proportion under different pulse lengths is different: about 35%-65% for short pulses (<10 seconds), about 40%-70% for medium pulses (10-100 seconds), and about 45%-75% for long pulses (>100 seconds). In the present application, the range of the set duration is determined to be 0.2 to 0.5 times the total pulse duration, which can adapt to a wider range of operating scenarios. In the present embodiment, the set duration is 0.3 times the total pulse duration. After scanning is completed in the entire coarse positioning range, the algorithm identifies all steady state sections that satisfy the above continuity condition (the window sequence length of the continuous stable window exceeds the set duration and the total steady state duration of the continuous stable section exceeds the set duration). If multiple steady state sections satisfying the condition are found, the algorithm selects the steady state section with the longest total steady state duration (steady state duration) as the real steady state stage, i.e., the cyclotron steady state operation stage. This is because the longest continuous stable section is most likely to correspond to the real steady state operation state dominated by quasi-equilibrium processes such as thermal diffusion. This continuity constraint ensures that the identified steady state stage indeed enters the steady state dynamics mechanism with a characteristic time scale of seconds to hundreds of seconds, rather than still being in the transient dynamics with a characteristic time scale of milliseconds to seconds.

[0072] The necessity of continuity constraint comes from the need to distinguish accidental fluctuations from true steady state. A single window of RSD less than 5% is not enough to determine that the system has entered steady state, because there can be a power temporary stabilization phenomenon with sub-second duration at the end of the start-up transient when the power is close to the target value or at the beginning of the shut-down transient when the power starts to decrease, but this does not mean that the system has really entered the steady state running stage, but only an accidental phenomenon in the transient process. The continuity constraint requires that at least 30 consecutive windows, i.e. 3 seconds, meet the condition of RSD less than 5%, which filters out accidental fluctuations from the time scale and ensures that only truly continuous stabilization can be identified as steady state. From the perspective of physical time scale, the characteristic time scale of quasi-equilibrium processes such as gyrotron thermal diffusion is on the order of seconds, and the requirement of at least 3 seconds of continuous stabilization can ensure that the system has really entered the steady state dynamics dominated by thermal diffusion, rather than the transient dynamics dominated by mode competition and beam regulation. In addition, the algorithm also requires that the total steady state time must be greater than or equal to 30 times IPLEN, which is based on the statistical law that the typical steady state occupies 30% to 70% in the physical experience, and also ensures that the amount of steady state data is sufficient for subsequent detailed analysis and model training.

[0073] The present application can also accurately locate the steady state boundary according to the identified steady state running stage of the gyrotron: the steady state start point of the steady state running stage of the gyrotron is the start time of the first sliding window of the steady state segment with the longest total steady state time, and the steady state end point of the steady state running stage of the gyrotron is the end time of the last sliding window of the steady state segment with the longest total steady state time. After determining the longest stable segment that meets the continuity constraint, the algorithm defines the start time of the first window of the stable segment as the steady state start point, which marks the end of the start-up transient and the beginning of the power fluctuation within ±2%. The steady state end point is defined as the end time of the last window of the stable segment, which marks the beginning of the shut-down transient and the time when the power starts to decrease rapidly.

[0074] Based on these two accurately positioned boundary points, the algorithm automatically divides the complete discharge process into three physical stages: the zero time of the power waveform to the steady state start point is the start-up transient stage of the gyrotron discharge, the steady state start point to the steady state end point is the steady state running stage of the gyrotron discharge, and the steady state end point to the end time of the pulse total duration is the shut-down transient stage of the gyrotron discharge. This clear division based on physical stages provides a high-quality data screening basis for subsequent data analysis, model training and fault diagnosis.

[0075] The distinguishing mechanism between start-up transient and steady state is the key to the accuracy of the identification algorithm. At the end of the start-up transient, the power has approached the target steady value, and it seems to have reached the steady state from the absolute value of the power. Why is it not mistaken for steady state operation by the algorithm? This is mainly based on the essential differences in the fluctuation characteristics and physical mechanisms. From the perspective of fluctuation characteristics, although the average power at the end of the start-up transient is close to the target value, the power is still in the process of fine tuning and fine climbing, and the ratio of the fluctuation amplitude to the average value, i.e. RSD, is usually still greater than 5%. Or even if the instantaneous RSD drops below 5%, it still shows an unstable fluctuation pattern, such as quasi-periodic oscillation or random jump, which cannot meet the requirement of stability for 30 consecutive windows. In contrast, the power fluctuation in the real steady state operation stage has been reduced to a small range of plus or minus 2%, and the RSD can be continuously and stably maintained below 5% or even below 2%, showing a small random fluctuation around the average value.

[0076] From the perspective of physical mechanisms, the start-up transient stage is dominated by processes such as nonlinear electromagnetic field mode competition and electron beam current injection adjustment. Even if the macroscopic power is close to the steady value, the microscopic electromagnetic field mode distribution and electron beam velocity distribution in each cavity are still rapidly evolving and have not yet reached the final quasi-equilibrium configuration. This rapid evolution of internal state in the macroscopic power is reflected in the form of large relative fluctuations and unstable fluctuation patterns. The steady state operation stage is dominated by slow-varying quasi-equilibrium processes such as thermal diffusion and radiation transmission. The evolution time scale of each physical quantity is much larger than the measurement time scale, and the internal state of the system has reached a quasi-equilibrium state, with the temperature field and electromagnetic field distribution in each cavity being basically stable. At this time, the power fluctuation is mainly caused by two unavoidable random factors, i.e. thermodynamic fluctuations and measurement noise, and the fluctuation amplitude is reduced to a minimum level. It is precisely due to the essential differences in these two aspects, i.e. large and unstable fluctuations, and the lack of continuity, that the power at the end of the start-up transient, although close to the steady state level, can still be accurately identified as a transient state by the algorithm, and will not be confused with the real steady state operation.

[0077] See Figure 2 and Figure 3The start-up transient phase, steady-state operation phase, and turn-off transient phase together constitute a complete "U"-shaped pulse waveform. Analyzing this pulse waveform allows for the precise time boundaries of the three phases and the statistical characteristic parameters of each phase. For the start-up transient phase, the average power ramp-up rate and the maximum value of the power relative standard deviation are calculated. Typically, this is characterized by a rapid power ramp-up, with the power relative standard deviation usually exceeding 10%, reflecting nonlinear transient processes such as electromagnetic field establishment and mode competition. For the steady-state operation phase, parameters such as the average power, power standard deviation, and average power relative standard deviation are statistically analyzed. A typical characteristic is that the power fluctuates stably within ±2%, and the average power relative standard deviation is less than 5%, reflecting a stable operating state dominated by quasi-equilibrium processes such as thermal diffusion. For the turn-off transient phase, parameters such as the power decrease rate and the change in the power relative standard deviation are calculated. This shows a rapid power decrease, with the power relative standard deviation increasing again to over 10%. This complete identification and quantitative characterization based on physical phases allows the three phases of the "U"-shaped pulse waveform to be precisely divided not only temporally but also accurately characterized physically.

[0078] For applications requiring batch processing of thousands of discharge data points, this invention offers excellent parallelization and computational efficiency. Since the steady-state identification task for each discharge is completely independent, the algorithm can be executed in parallel on multi-core workstations. For the identification task of a single discharge, the processing time on a typical 8-core Intel workstation is approximately 0.3 to 0.5 seconds. This makes the total processing time for 2500 discharges approximately 2.1 hours, significantly faster than the 168 hours (one week) required for manual annotation, representing an 84-fold speedup. The algorithm's output includes the time boundaries of the three stages for each discharge, statistical characteristic parameters for each stage, and identification quality indicators such as the average RSD and continuity score of the steady-state segment. These results can be directly used for subsequent data analysis, model training, and quality assessment, providing a practical tool for the automated processing of large-scale experimental data.

[0079] This invention first utilizes IPLEN parameters to coarsely locate the steady-state position over a time range. Taking a search range that accounts for 15% to 85% of the total pulse duration as an example, this invention sets the search range to 0.15 times... Up to 0.85 times The interval is reduced, the calculation range is reduced, and the misjudgment risk is reduced. Then, a sliding window method is used to scan the sliding window in the search range, the relative standard deviation RSD of power in each sliding window is calculated, the RSD is calculated point by point, and the stability of each window is judged, the sliding window with a relative standard deviation less than a threshold value is marked as a stable window, a dimensionless relative standard deviation RSD index (RSD is equal to the standard deviation divided by the mean value) is used as a stability criterion independent of the power level, and the power waveform in the coarse positioning range is dynamically scanned for local stability. The dimensionless normalization form is the mathematical basis for realizing cross-power adaptation. Then, continuity constraints are applied, and the length of the window sequence of continuous stable windows is required to exceed a set duration, and the total length of the stable state exceeds a set duration. The continuous stable section identifies the real stable section and excludes accidental fluctuations. The length of the window sequence of continuous stable windows exceeds the set duration, that is, the number of continuous stable windows is greater than or equal to the set number (in the present application, the set number is 30). The present application realizes cross-power and cross-time adaptive identification through dimensionless normalization and physical constraints, has the characteristics of zero hardware cost and high automation expert-level accuracy, and solves the problems of high cost, low efficiency and poor universality of the prior art.

[0080] In addition, the accurate positioning of the stable state starting point is the starting time of the first stable window, and the stable state ending point is the ending time of the last stable window, so as to realize automatic division of the three stages of start transient state, stable state and shutdown transient state.

[0081] The present application realizes zero-cost high-precision stable state identification, and the identification performance reaches the expert level. The present application uses the existing IPLEN control signal and power waveform data of the control system to realize stable state identification, and completely does not need additional hardware investment. Compared with the prior art which depends on special instruments such as infrared thermal imaging systems and spectral diagnostic devices, the present application saves hundreds of thousands of hardware costs and system integration costs, and avoids the maintenance cost and reliability risk of additional diagnostic systems.

[0082] In terms of automation efficiency, the identification speed of the present application is 84 times faster than manual labeling, and the processing time of 2500 discharges is reduced from 168 hours of manual labeling, i.e. one week of working time, to 2.1 hours of automatic processing of the algorithm, greatly improving the data processing efficiency and solving the bottleneck problem of time-consuming and high cost of manual labeling.

[0083] In terms of recognition accuracy, the steady-state recognition accuracy of the application reaches 89.1%, the Fleiss' kappa coefficient is 0.87 close to the expert annotation level of "almost complete agreement", the false positive rate is only 3.2%, that is, the proportion of mistaking transient state as steady state is very low, and the false negative rate is 7.7%, that is, the proportion of missing real steady state is also very low. Such high accuracy not only reaches the quality level of expert annotation, but also completely eliminates the subjective differences between different annotators and the inconsistency of annotation results due to the use of objective mathematical criteria, ensuring the quality stability of large-scale data processing.

[0084] More importantly, the application uses the dimensionless RSD index as a stability criterion, so that a single threshold of 5% can automatically adapt to a wide power range of 0.1 megawatt to 1.2 megawatt spanning an order of magnitude and different pulse lengths of 0.1 seconds to 500 seconds spanning three orders of magnitude, without the need to recalibrate parameters for different operating conditions. This cross-power and cross-time adaptive capability is a significant advantage that existing absolute threshold-based methods cannot achieve.

[0085] In addition, the application can meet the real-time deployment requirements and has practical engineering application value. The end-to-end processing delay of the steady-state recognition algorithm of the application is only 15 milliseconds, which is much lower than the 50-millisecond response time constraint required by the auxiliary heating control system of fusion, which enables the algorithm to be deployed in an online control system to realize real-time steady-state determination and data filtering, providing technical support for online fault prediction and timely preventive measures. The processing throughput of the algorithm reaches 320 times per second, which is sufficient to meet the real-time data stream processing requirements of the control system and will not become a bottleneck in the data processing link. In terms of resource occupation, the memory occupation of the algorithm running is less than 500 MB, which can run smoothly on a standard 8-core workstation without the need for a dedicated high-performance computing cluster or GPU accelerator, which greatly reduces the deployment threshold and operation and maintenance cost. From the perspective of engineering reliability, the application uses mature signal processing methods such as sliding window and statistical analysis, the algorithm logic is clear without complex external dependencies, and it is easy to integrate into existing industrial control system frameworks such as EPICS, and the code implementation is stable and reliable. These characteristics make the application not only innovative in scientific research, but also highly deployable and practical in engineering practice.

[0086] Embodiment 2

[0087] Referring to Figure 4 The embodiment provides a gyrotron steady-state operation phase automatic recognition system, which comprises:

[0088] The coarse positioning module is used to read the control signal and power waveform of a single discharge of the gyrotron, and to locate the search range of the steady-state stage on the power waveform based on the control signal parameters. The process of locating the search range of the steady-state stage on the power waveform based on the control signal parameters includes: determining the total pulse duration by the control signal parameters, determining the proportion of the search range to the total pulse duration based on the actual duration of the steady-state stage of the gyrotron, and locating the search range of the steady-state stage on the power waveform based on the proportion.

[0089] The actual duration of the steady-state phase of the gyrotube accounts for 30% to 70% of the total pulse duration, and the search range accounts for 15% to 85% of the total pulse duration.

[0090] The fine-positioning module performs a sliding window scan within the search range, calculates the relative standard deviation of power within each sliding window, and marks sliding windows with a relative standard deviation less than a threshold as stable windows; the size of the sliding window... The step size of the sliding window scan corresponds to a time length of 0.1 seconds for 1000 points. 100 points correspond to 0.01 seconds, and the threshold ranges from 2% to 5%. The total number of sliding windows within the search range. for:

[0091]

[0092] in, The length of the search range, The size of the sliding window. The step size for the sliding window scan. This indicates rounding down to the nearest integer.

[0093] The relative standard deviation of power within each sliding window is:

[0094]

[0095] in, For the first The relative standard deviation of power within the sliding window. , The total number of sliding windows within the search range. For the first The average of all power values ​​within the sliding window. For the first The standard deviation of each power within the sliding window.

[0096] Steady-state identification module, used to identify at least N A series of consecutive stable windows are identified as consecutive stable segments. NThe number of set is set, and the continuous stable segment with a total stable duration longer than the set duration is identified as a stable segment, and the stable segment with the longest total stable duration is identified as a gyrotron stable operation stage. The set duration ranges from 0.2 times to 0.5 times of the total pulse duration.

[0097] The stable start point of the gyrotron stable operation stage is the start time of the first sliding window of the stable segment with the longest duration, and the stable end point of the gyrotron stable operation stage is the end time of the last sliding window of the stable segment with the longest duration. The zero time of the power waveform to the stable start point is the start transient stage of the gyrotron discharge, the stable start point to the stable end point is the stable operation stage of the gyrotron discharge, and the stable end point to the end time of the total pulse duration is the off transient stage of the gyrotron discharge. The start transient stage, the stable operation stage and the off transient stage jointly constitute a pulse waveform in the shape of several characters.

[0098] Embodiment 3

[0099] The embodiment provides a gyrotron power prediction method, based on the gyrotron stable operation stage automatic identification method in embodiment 1, the gyrotron power prediction method comprising the following steps:

[0100] Step 4, preprocessing the control parameters of the gyrotron stable operation stage, the preprocessing is a normalization operation, and an input feature is obtained; the input feature includes multiple physical quantities such as cathode voltage, body current, anode current, magnetic field strength, cooling water flow, etc., which form an input sequence in time sequence.

[0101] Step 5, training a power prediction model with the input feature as input and the gyrotron output power as output label to obtain the trained power prediction model; the power prediction model of the embodiment adopts a standard long short-term memory network (LSTM) architecture, the network structure is two-layer LSTM containing 64 hidden units in each layer, and the output layer is a full connection layer outputting the predicted gyrotron output power. The training process uses Adam optimizer, the learning rate is set to 0.001, the batch size is 32, and the mean square error is used as the loss function for back propagation training.

[0102] Step 6, inputting the control parameters of the running gyrotron into the trained power prediction model to predict the gyrotron output power.

[0103] The stable stage data identified by the embodiment 1 of the application is used to train the machine learning power prediction model, avoiding problems such as gradient conflict, mode confusion, error distribution anomaly caused by mixing of transient and stable heterogeneous data.

[0104] The mechanism of improving prediction performance by steady-state filtering training originates from the principle of consistency between training data and test data distribution. The generalization performance of a machine learning model essentially depends on the matching degree between the training data distribution and the test data distribution. When the two distributions are consistent, the model can effectively generalize the patterns learned from the training data to the test data. When the training data mixes samples with completely different physical mechanisms and dynamics characteristics, the model is forced to fit two completely different input-output mapping relationships in the same parameter space, which will lead to performance degradation in multiple aspects.

[0105] Firstly, there is a gradient conflict problem. Transient samples require the model to learn the fast-changing nonlinear dynamics that capture the rapid climb of power and the mode competition process, while steady-state samples require the model to learn the slowly changing quasi-equilibrium behavior that captures the small stable fluctuations of power. These two types of samples produce mutually contradictory gradient update directions during backpropagation, reducing the efficiency and convergence quality of the optimization process.

[0106] Secondly, there is a pattern confusion problem. The model may mistake the mode competition behavior in the transient stage for normal fluctuation patterns in the steady-state stage, or mistake the small fluctuations in the steady-state stage for rapid changes in the transient stage, leading to internal representation confusion and inaccurate learning of the true physical laws in each stage.

[0107] Thirdly, there is an error distribution anomaly problem. Due to the high nonlinearity and sensitivity to initial conditions of the transient process, the prediction error of transient samples is usually significantly larger than that of steady-state samples. When mixed training is used, the overall error distribution of the model presents a long-tailed feature with a much higher kurtosis than a Gaussian distribution, which means that there is a high probability of significant deviation in the prediction result. By using steady-state filtering training, the model only needs to learn a single physical mechanism, i.e., steady-state dynamics, avoiding all the above problems. Since power prediction tasks are usually performed in steady-state conditions in practical applications, i.e., predicting the power at a certain time in the steady-state stage, the physical mechanism and statistical distribution of the training data and test data are consistent, and the model can more effectively learn the internal laws of steady-state dynamics, showing higher prediction accuracy, smaller error variance, and more reliable error distribution in the test stage. The essence of this method is to improve the quality and internal consistency of training data through physically oriented data management, improving model performance from the data source rather than the algorithm architecture level, embodying the technical concept of "data first, algorithm for use".

[0108] Comparative Experiment 1

[0109] To verify the effectiveness of the automatic identification method of the gyrotron steady-state operation phase in Embodiment 1, this embodiment quantitatively evaluates the accuracy and practical value of the automatic identification method of the gyrotron steady-state operation phase in Embodiment 1 through a comparative experiment. The basic idea of verification is: if the steady-state identification is accurate and can effectively separate the heterogeneous physical phases, then the prediction model trained only using the steady-state data should be significantly better than the model trained using mixed data; on the contrary, if the identification is inaccurate and causes the mixing of transient data or the omission of part of the steady-state data, then the performance improvement will not be obvious or even possibly negative.

[0110] The comparative experiment design includes two experimental groups. Experimental group one uses mixed data training, that is, the training data contains all the data of the complete discharge process including the startup transient, steady-state operation and shutdown transient three phases, which simulates the conventional practice in the prior art of not distinguishing the physical phases and directly using all the data. Experimental group two uses steady-state filtering training, that is, the training data only contains the data of the steady-state operation phase extracted by the automatic identification method of the gyrotron steady-state operation phase in Embodiment 1, and excludes the startup and shutdown transient phases. The two experimental groups use the same LSTM network architecture, the same hyperparameter configuration, the same training and test data division method, and the only difference is whether the training data is filtered by the steady-state. The LSTM network configuration is a standard implementation of the power prediction task and is widely used in existing literature, so it is suitable as a benchmark model for comparative experiments. The performance evaluation indicators include the coefficient of determination, the MAPE average absolute percentage error, and the statistical characteristics of the error distribution (such as kurtosis and skewness), the coefficient of determination is used to measure the goodness of fit of the model, the MAPE average absolute percentage error is used to measure the relative error of the prediction, and the statistical characteristics of the error distribution are used to evaluate the prediction reliability.

[0111] The experimental results show that if the steady-state identification is accurate, the performance of the model trained by the steady-state filtering will be significantly improved. Specifically, from 0.201 to 0.284, with an improvement of 41.3%, the MAPE is reduced from 14.8% to 11.1%, with a reduction of 25.0%, and the error distribution is improved from the long-tailed distribution with a kurtosis of 5.82 to the nearly Gaussian distribution with a kurtosis of 3.21. These performance improvements directly prove the effectiveness of the steady-state identification algorithm. If the identification is inaccurate and a large amount of transient data is mixed or important steady-state data is omitted, the quality of the training data will not be better than the mixed data or even possibly worse, and the model performance will not improve. Therefore, the performance comparison of the power prediction task is actually a practical verification means of the steady-state identification algorithm, which confirms the accuracy and practical value of the identification algorithm from the application perspective. In addition, the time of steady-state filtering training remains 45 minutes, which is the same as the mixed data training, indicating that the improvement of data quality does not increase the computing cost, which is of great significance for actual deployment.

[0112] The performance difference between the model trained by the steady-state data and the model trained by the mixed data is compared as The improvement of the coefficient of determination and the MAPE average absolute percentage error verifies the effectiveness and accuracy of the steady-state identification algorithm. The application verification from the practical point of view confirms the practical value of the identification algorithm, and provides a high-quality data basis for data-driven modeling of the gyrotron.

[0113] The effectiveness of the steady-state identification algorithm is verified by the power prediction application, and a high-quality data basis is laid for prediction modeling. Compared with the model trained by the mixed stage data, the performance of the LSTM power prediction model trained by the steady-state data identified by the application is significantly improved. Specifically, The coefficient of determination is improved from 0.201 to 0.284, with an improvement of 41.3%, which indicates that the goodness of fit and the explanation ability of the model are greatly enhanced. The MAPE average absolute percentage error is reduced from 14.8% to 11.1%, with a reduction of 25.0%, which means that the relative error of the prediction is reduced by one fourth, and the prediction accuracy is significantly improved. The error distribution characteristics are also significantly improved, from the long-tailed distribution kurtosis 5.82 to the nearly Gaussian distribution kurtosis 3.21, which indicates that the reliability of the prediction result is enhanced, and the probability of significant deviation from the prediction is greatly reduced. These performance improvements directly prove the effectiveness and accuracy of the steady-state identification algorithm of the application.

[0114] From a logical point of view, if the steady-state identification is inaccurate, resulting in a large amount of transient data mixed in the training data or important steady-state data missing, the quality of the training data will not be better than or even worse than the mixed data, so the model performance cannot be significantly improved. Therefore, the significant improvement in the power prediction task confirms from the practical application point of view that the identification algorithm indeed accurately separates the two heterogeneous physical phases of steady-state and transient, and extracts high-quality steady-state data, providing a training data basis with a single physical mechanism and strong internal consistency for prediction modeling. In addition, it is worth noting that the steady-state filtering training time remains at 45 minutes, which is the same as the mixed data training, which means that the data quality improvement does not increase the training computational cost, and only needs to run the steady-state identification algorithm once in the data preparation stage, which is very beneficial to actual deployment.

[0115] Comparative experiment 2

[0116] In order to evaluate the performance contribution of the data quality improvement method proposed in the application, i.e. data filtering based on steady-state identification, relative to the traditional algorithm optimization method such as hyperparameter tuning, this embodiment designs a comparative experiment of the system, and compares and evaluates the data quality improvement and the algorithm optimization. The experiment uses the control variable method to decompose the performance improvement into two independent contribution factors of data quality improvement and algorithm optimization.

[0117] Experiment A evaluates the contribution of data quality improvement. In this experiment, the LSTM network architecture and all hyperparameter configurations are kept completely unchanged, only the source of training data is switched from mixed-phase data to steady-state filtered data. By comparing the performance difference of the model under two kinds of training data, the performance improvement brought by data quality improvement can be directly quantified as .

[0118] Experiment B evaluates the contribution of algorithm optimization. In this experiment, the training data is kept unchanged and still uses mixed-phase data, but the hyperparameters of LSTM are optimized using meta-heuristic algorithms such as whale optimization algorithm, including the number of hidden units, learning rate, batch size, etc. By comparing the performance difference of the model under default hyperparameters and optimized hyperparameters, the performance improvement brought by algorithm optimization can be quantified as .

[0119] The comparison results show that the contribution of data quality improvement is much greater than that of algorithm optimization. Specifically is about times, data quality improvement improves R² by 41.3%, while algorithm optimization only improves 1.4%, even in some cases it is negative. More importantly, the time cost of data quality improvement is almost zero, only running the steady-state identification algorithm once, while the algorithm optimization needs a lot of hyperparameter search iterations, and the training time increases by 4 to 5 times. This comparison experiment quantitatively answers a basic question: under the constraint of limited resources, should we invest more in developing high-quality data screening and cleaning tools for data management infrastructure or in the complexity and fine optimization of algorithm architecture? The experimental results clearly show that for complex engineering systems with heterogeneous operating mechanisms, when the data itself contains serious heterogeneity, i.e. mixed with different physical phases and dynamic mechanisms, the performance contribution of data quality improvement is much greater than that of algorithm optimization. Resources should be invested in data management rather than blindly pursuing the complexity of algorithms.

[0120] The methodological significance of this finding is that it can be generalized to other engineering systems with heterogeneous operating mechanisms. For example, particle accelerators include three physically distinct phases of injection, acceleration, and storage; laser devices include preheating, main pulse, and cooling phases; power systems include startup, normal operation, and overload phases; and chemical reaction processes include reaction startup, stable reaction, and reaction termination phases. When developing prediction models or diagnostic algorithms for these systems, the comparison experiment results of the present invention provide a clear basis for decision-making: priority should be given to developing data phase identification and filtering tools similar to the steady-state identification algorithm of the present invention to improve the quality and physical consistency of training data, rather than blindly optimizing algorithm architecture based on mixed heterogeneous data. This physically oriented data management methodology can provide valuable reference and guidance for the development of related technologies.

[0121] The performance comparison mechanism of data quality improvement and algorithm optimization is based on the scientific experiment design principle of control variable method. In the data quality improvement experiment, all algorithm related factors are kept unchanged, including network architecture, LSTM layer number, hidden unit number, hyperparameter learning rate, batch size, optimizer selection, training process, training round number, early stopping strategy, etc. The only variable changed is the source of training data, which is switched from mixed phase data to steady state filtered data. By comparing the performance indicators of the model in the two cases, the performance improvement can be completely attributed to the improvement of data quality, and the performance benefit brought by data filtering can be quantitatively measured In the algorithm optimization experiment, the training data is kept unchanged and still uses mixed phase data, but the meta-heuristic optimization algorithm such as whale optimization algorithm, particle swarm optimization algorithm or Bayesian optimization is used to search and optimize the hyperparameters of the model, to find the optimal combination of hyperparameters on the training data set. By comparing the performance difference of the model under the default hyperparameter configuration and the optimized hyperparameter configuration, the performance improvement can be attributed to algorithm optimization, and the performance benefit brought by hyperparameter tuning can be quantitatively measured .

[0122] After the completion of the two experiments, the and are directly compared. Combined with the calculation time cost and resource investment of the two methods, the basic technical route selection problem of whether to prioritize investment in data management infrastructure or algorithm architecture optimization under the constraint of limited resources can be quantitatively answered. From the physical intuition point of view, when the data itself contains serious heterogeneity, that is, mixed with different running stages dominated by different physical laws, there are fundamental inconsistencies and contradictions in the data. This data level problem cannot be fundamentally solved by algorithm level optimization, no matter how complex and sophisticated the algorithm architecture and how fine the hyperparameter tuning is. On the contrary, if the data quality is high, that is, the physical mechanism is single and the internal consistency is strong, and the noise level is low, even a relatively simple algorithm architecture and a rough hyperparameter setting can achieve good prediction effect. The experimental results of the present invention verify this physical intuition. In the gyrotron power prediction task, the performance contribution of data quality improvement is about 30 times that of algorithm optimization. This finding has important methodological significance for guiding the technical route of prediction modeling of similar engineering systems, and clearly indicates that resources should be prioritized for data management rather than blindly pursuing algorithm complexity.

[0123] The present application provides a quantitative basis for technical path decision-making, which has important methodological significance and promotional value. The present application quantitatively evaluates the performance contribution of data quality improvement (based on accurate steady-state identification) and algorithm optimization (based on hyperparameter tuning) through systematic comparative experiments. The experimental results show that the performance contribution of data quality improvement is about 30 times that of algorithm hyperparameter optimization. Specifically, data quality improvement improves R² by 41.3%, while algorithm optimization improves R² by only 1.4% or even negative in some experiments. More importantly, the hyperparameter optimization of meta-heuristic algorithms requires a large number of iterative searches, and the training time increases by 4 to 5 times, while the computational cost of data quality improvement is almost zero, only running a steady-state identification algorithm once. This comparative result clearly shows that for complex engineering systems with heterogeneous operating mechanisms, under limited resource conditions, it should prioritize the development of high-quality data management infrastructure, including phase identification, screening, and cleaning tools, rather than blindly pursuing the complexity of algorithm architecture and the fine optimization of hyperparameters based on mixed heterogeneous data. This physical-oriented data management methodology can be extended to other engineering systems with heterogeneous operating mechanisms, such as particle accelerators containing injection, acceleration, and storage stages, laser devices containing preheating, main pulse cooling, and three stages, power systems containing startup, normal, and overload stages, and chemical reaction processes containing reaction startup, stabilization, reaction termination, and other multiple engineering fields. For these systems in developing prediction models or diagnostic algorithms, the experimental results and methodology of the present application provide a scientific basis for technical route selection and resource investment decisions, which has a wide guiding value.

[0124] The above examples are only used to illustrate the technical solutions of the present application, but not to limit it; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An automatic identification method for the steady-state operation phase of a gyrotube, characterized in that: the method The method comprises the following steps: The control signal and power waveform of a single discharge of the readout cyclotron are read, and a search range of a steady state phase is located on the power waveform based on a control signal parameter; sliding window scanning is performed in the search range, the relative standard deviation of power in each sliding window is calculated, and the sliding window with a relative standard deviation less than a threshold value is marked as a stable window; at least N one continuous stable window is identified as a continuous stable section, N a set number, and the continuous stable section with a steady state total duration exceeding a set duration is identified as a steady state section, and the steady state section with the longest steady state total duration is identified as a steady state operation phase of the cyclotron.

2. The gyrotron steady-state operation phase automatic identification method of claim 1, wherein: The process of locating the search range of the steady state phase on the power waveform based on the control signal parameters comprises: determining the total pulse duration by the control signal parameters, determining the proportion of the search range in the total pulse duration according to the actual duration of the steady state phase of the gyrotron, and locating the search range of the steady state phase on the power waveform according to the proportion.

3. The gyrotron steady-state operation phase automatic identification method of claim 2, wherein: The proportion of the actual duration of the steady state phase of the gyrotron in the total pulse duration is 30% to 70%, and the proportion of the search range in the total pulse duration is 15% to 85%.

4. The gyrotron steady-state operation phase automatic identification method of claim 1, wherein: Total number of sliding windows within search range For: wherein, is the length of the search range, is the size of the sliding window, is the step size of the sliding window scan, denotes the floor function.

5. The gyrotron steady-state operation phase automatic identification method of claim 1, wherein: The relative standard deviation of the power in each sliding window is: wherein, is the first relative standard deviation of all powers within the sliding window, , is the total number of sliding windows within the search range, is the first is the mean of all powers within the sliding window, is the first is the standard deviation of each power within the sliding window.

6. The gyrotron steady-state operating phase automatic identification method of claim 1, wherein: Size of the sliding window 1000 points correspond to a time length of 0.1 seconds, the step size of the sliding window scan 100 points correspond to 0.01 seconds, the threshold value ranges from 2% to 5%, the setting time length ranges from 0.2 times to 0.5 times of the total duration of the pulse.

7. The gyrotron steady-state operation phase automatic identification method of claim 1, wherein: The starting time of the steady state phase of the steady state running phase is the starting time of the first sliding window of the steady state section with the longest total duration, and the ending time of the steady state phase of the steady state running phase is the ending time of the last sliding window of the steady state section with the longest total duration.

8. The gyrotron steady-state operation phase automatic identification method of claim 7, wherein: The zero time of the power waveform to the steady state starting point is the start transient phase of the gyrotron discharge, the steady state starting point to the steady state ending point is the steady state running phase of the gyrotron discharge, and the steady state ending point to the total pulse duration ending time is the off transient phase of the gyrotron discharge, and the start transient phase, the steady state running phase and the off transient phase jointly constitute a pulse waveform in the shape of several characters.

9. A system for automatic identification of steady-state operation phase of a gyrotron, characterized by: The system comprises: A coarse positioning module is configured to read the control signal and the power waveform of a single discharge of the gyrotron, and locate the search range of the steady state phase on the power waveform based on the control signal parameters; An accurate positioning module is configured to perform sliding window scanning in the search range, calculate the relative standard deviation of the power in each sliding window, and mark the sliding window with a relative standard deviation less than a threshold as a stable window. A steady state recognition module is configured to recognize at least N a plurality of continuous stable windows as continuous stable segments, N a set number, and recognize a continuous stable segment with a steady state total duration exceeding a set duration as a steady state segment, and recognize a steady state segment with the longest steady state total duration as a steady state operation phase of the gyrotron.

10. A gyrotron power prediction method based on the automatic identification method of the steady-state operating phase of a gyrotron according to any one of claims 1-8, characterized in that: The method comprises: Preprocessing the control parameters of the steady state running phase of the gyrotron to obtain input features; Training a power prediction model by taking the input features as input and the output power of the gyrotron as output label to obtain the trained power prediction model, and inputting the control parameters of the running gyrotron into the trained power prediction model to predict the output power of the gyrotron.

Images