A Beam Closed-Loop Cooperative Control Method Based on X-ray Source Control System
By constructing a feature tensor and an edge-end intelligent inference model, the adaptive beam stability output of the X-ray source control system under different operating conditions was realized, which solved the problem of insufficient beam stability in the existing technology and improved the stability and response speed of the control system.
Patent Information
- Application Number
- CN202611363035.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-03
- Publication Date
- 2026-11-17
AI Technical Summary
Existing X-ray source control systems lack beam stability under dynamic conditions and cannot adapt to rapid switching and environmental changes under different operating conditions, resulting in inconsistent control performance.
By constructing a feature tensor to represent the dynamic relationship of the cross-coupling between the high-voltage feedback signal and the filament current feedback signal, the dynamic confidence weight allocation of the control loop is calculated in real time using an edge-end intelligent inference model, an adaptive modulation solution timing vector is generated, and a closed-loop control link is formed through a predictive compensation matrix to achieve adaptive and stable output of the beam.
It improves the stability and robustness of X-ray source beam control under strongly nonlinear dynamic conditions, and enhances the response speed and suppression capability to sudden load changes and external disturbances.
Smart Images

Figure CN123113540A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of X-ray source control technology, and more specifically, to a beam closed-loop cooperative control method based on an X-ray source control system. Background Technology
[0002] X-ray sources are core components of medical imaging equipment and industrial non-destructive testing systems. The stability of their output beam directly affects image clarity and the reliability of test results. During the operation and control of an X-ray source, beam stability depends on the coordinated regulation of the high-voltage output and filament current by the control system. These two factors jointly determine the final dose rate and energy spectrum characteristics of the X-ray tube. However, in the actual physical processes of an X-ray source, the high-voltage electric field and thermionic emission are not completely independent but exhibit a mutually influential coupling characteristic. Changes on the high-voltage side alter the electric field distribution within the tube, thus affecting the motion characteristics of electrons emitted by the filament. Simultaneously, fluctuations in the filament current also influence the load state on the high-voltage side. This coupling relationship means that the X-ray source system is not inherently a strictly linear, time-invariant system but rather possesses time-varying and nonlinear characteristics.
[0003] To balance control accuracy and engineering feasibility, a series of X-ray source control schemes have been proposed in existing technologies. A typical implementation involves pre-setting a set of control parameters, embedding these parameters in the control program, and maintaining their values unchanged during equipment operation. The basic logic of this approach is to approximate the X-ray source system as a linear time-invariant model, and calculate a combination of control parameters that can accommodate typical operating conditions based on this model. When the X-ray source operates within the standard operating conditions covered in the design phase, the aforementioned method of adjusting the pre-set parameters can keep the system output within an acceptable deviation range, meeting the basic usage requirements of conventional application scenarios, and thus has been widely used in industrial practice. Although some existing schemes attempt to introduce adaptive control logic, their adaptive strategies still rely on a pre-set rule base or lookup table method for specific disturbance types, meaning they still fail to break through the constraints of the pre-set model framework. When operating condition combinations not covered by the rule base occur, their adjustment capability is limited.
[0004] However, with the increasing demands for imaging resolution and detection accuracy in medical imaging and industrial inspection, the actual operating modes of X-ray sources are becoming increasingly diverse. They frequently require rapid switching between different voltage and current levels, while the external environment and usage conditions of the equipment are constantly changing. Under these operating conditions, if pre-calibrated and fixed control parameters are continued to be used, the matching degree between the system's actual response characteristics and the preset model will decrease, resulting in inconsistent beam stability control performance under different operating conditions. Since the intervention priorities and adjustment strengths of the voltage regulation loop, current regulation loop, and integrated compensation loop need to be adjusted in real time according to changes in operating conditions, and pre-calibrated fixed parameters cannot achieve this real-time tracking, using fixed parameters when the X-ray source is operating under dynamic conditions will cause a deviation between the intervention capabilities of each control loop and the actual requirements, affecting the beam stability accuracy.
[0005] If the actual operating state can be quantified into a unified feature expression through multi-source sensor signals, and the intervention weights of each control loop can be dynamically allocated based on this feature expression using an inference model, while simultaneously predicting the fluctuation direction of the next cycle based on the deviation change trend and applying compensation in advance, it is hoped that the limitations of fixed parameter control can be overcome. How to enable the control strategy to adaptively adjust according to the current actual operating state of the X-ray source, so as to expand the effective application range of beam stabilization control, is a direction that those skilled in the art continue to focus on. Summary of the Invention
[0006] To achieve the above objectives, the present invention provides the following technical solution: A beam closed-loop cooperative control method based on an X-ray source control system includes the following steps: The first step is to acquire multi-source sensor signals of the X-ray source in real time and in parallel. The multi-source sensor signals include at least the high-voltage feedback signal, the filament current feedback signal, the actual beam current detection signal, and the die temperature detection signal. The multi-source sensor signals are then subjected to time-frequency domain feature extraction and normalization to construct a feature tensor that characterizes the current transient multi-physics coupling state. The feature tensor uniformly carries the cross-coupling dynamic relationship between the high-voltage feedback signal and the filament current feedback signal. The second step is to input the feature tensor into the intelligent inference model pre-deployed at the edge. The intelligent inference model calculates the dynamic confidence weight allocation array of each control loop in real time based on the working condition represented by the current feature tensor. The control loop includes at least a voltage regulation loop, a current regulation loop, and a comprehensive compensation loop. The dynamic confidence weight allocation array represents the intervention priority and regulation intensity allocation value of each control loop in the current control cycle. The third step is to generate an adaptive modulation and decomposition timing vector for the high-frequency inverter circuit within the current control cycle based on the dynamic confidence weight allocation array. The adaptive modulation and decomposition timing vector contains a set of timing parameters for at least one pulse width modulation drive signal and one pulse frequency modulation drive signal. The adaptive modulation and decomposition timing vector is then output to the high-frequency inverter circuit to drive the X-ray source to perform power output. The fourth step is to re-acquire the actual output X-ray beam parameters and high voltage output parameters after the power output is completed. Based on the deviation between the actual output X-ray beam parameters and the preset target beam parameters, the steady-state deviation cost functional of the current control cycle is calculated. The steady-state deviation cost functional is a quantitative evaluation value that comprehensively reflects the transient error, cumulative error and error change trend in the current control cycle. The fifth step is to input the steady-state deviation cost functional into the intelligent inference model. The intelligent inference model combines the steady-state deviation cost functional of the current control cycle with the historical deviation cost functional of at least one control cycle to predict the fluctuation trend of the next control cycle and generate a predictive compensation prior coefficient matrix accordingly. The sixth step involves feeding the predictive compensation prior coefficient matrix back to the feature extraction stage in the first step as a feedback compensation quantity. This corrects the reference bias and feature sensitivity weights when acquiring multi-source sensor signals in the next control cycle, thus forming a complete closed-loop control link consisting of feature tensor construction, dynamic confidence weight allocation array calculation, adaptive modulation solution time-series vector generation, steady-state deviation cost functional calculation, and predictive compensation prior coefficient matrix feedback. This enables adaptive and stable output of the X-ray source beam.
[0007] In a preferred embodiment, the multi-source sensing signals also include independent cathode filament power supply voltage signals and anode rotation drive current signals; the specific implementation of time-frequency domain feature extraction and normalization processing is as follows: For each signal in the multi-source sensing signal, a Fast Fourier Transform is performed to extract the frequency domain amplitude spectrum and frequency domain phase spectrum of the signal, while retaining the time domain waveform values. The DC component value, fundamental amplitude, and second harmonic amplitude of the signal in the frequency domain amplitude spectrum are used as the basic feature values of the signal. Each dimension of the basic feature values is then linearly scaled according to a preset normalization interval, so that each scaled dimension falls within a uniform numerical range, thereby obtaining the normalized feature values corresponding to each dimension in the feature tensor. In addition to carrying the cross-coupling dynamic relationship between the high-voltage feedback signal and the filament current feedback signal, the feature tensor also carries the correlation coupling relationship between the independent cathode filament power supply voltage signal and the anode rotation drive current signal.
[0008] In a preferred embodiment, the edge processor performs the following operations to construct the feature tensor: The curves of the high-voltage feedback signal and the filament current feedback signal over time are considered as response waveforms of two independent vibration modes. Cross-correlation convolution is performed on the two response waveforms to obtain their cross-correlation intensity sequence in the time domain. Peak extraction is then performed on the cross-correlation intensity sequence to obtain the peak time position and peak amplitude. The coupling response delay time between the high-voltage feedback and the filament current feedback is determined based on the peak time position, and the coupling response gain coefficient between the high-voltage feedback and the filament current feedback is determined based on the peak amplitude. The coupling response delay time and the coupling response gain coefficient are incorporated as independent dimensions into the feature tensor.
[0009] In a preferred embodiment, the intelligent inference model is a lightweight deep neural network model pre-trained offline, and the model is deployed in an edge-embedded processor of the X-ray source control system; the calculation process of the dynamic confidence weight allocation array includes: The feature tensor is fed into the deep neural network model as input data to perform forward inference operations. The output layer of the deep neural network model is directly mapped to the confidence weight values of the voltage regulation circuit, current regulation circuit and comprehensive compensation circuit respectively.
[0010] In a preferred embodiment, the timing parameters of the pulse width modulation drive signal in the adaptive modulation demodulation timing vector include the duty cycle, switching frequency, and dead time; the timing parameters of the pulse frequency modulation drive signal include the period reference frequency and frequency offset; the adaptive modulation demodulation timing vector also includes the time phase offset of each timing parameter relative to the synchronous clock reference.
[0011] In a preferred embodiment, the steady-state deviation cost functional calculated for each historical control cycle is stored as a sequence of deviation values in chronological order; the specific method for obtaining the error change trend in the steady-state deviation cost functional is as follows: Extract the deviation value subsequence from the deviation value sequence that is at least one consecutive historical control cycle prior to the current control cycle; perform linear predictive coding analysis on the deviation value subsequence, and extract the power spectral density distribution characteristics of the deviation value subsequence based on the analysis results; then take the main peak frequency value in the power spectral density distribution characteristics as the dominant frequency of deviation fluctuation, and compare the dominant frequency of deviation fluctuation with the preset fluctuation frequency threshold. Generate a trend weighting coefficient based on the comparison results, and use the trend weighting coefficient to amplify or reduce the evaluation weight of the error change trend in the steady-state deviation cost functional; wherein the preset fluctuation frequency threshold is obtained by pre-calibration based on the rated operating frequency of the X-ray source control system.
[0012] In a preferred embodiment, before calculating the steady-state deviation cost functional, a target beam parameter tolerance band is preset. When the absolute value of the deviation between the actual output X-ray beam parameter and the preset target beam parameter is within the target beam parameter tolerance band, it is determined that the current control cycle has reached a stable output state. The step of generating the predictive compensation prior coefficient matrix is skipped, and the predictive compensation prior coefficient matrix is directly identified as a matrix with all elements being zero. The dimension of the matrix is the same as that of the predictive compensation prior coefficient matrix in the fifth step. The matrix with all elements being zero is used as the feedback compensation amount to be fed back to the first step, thereby keeping the current adaptive modulation decomposition timing vector unchanged.
[0013] In a preferred embodiment, when the intelligent inference model combines the steady-state deviation cost functional of the current control cycle with the historical deviation cost functional of at least one control cycle to predict the fluctuation trend of the next control cycle, it specifically performs the following operations: The steady-state deviation cost functional of the current control cycle and the historical steady-state deviation cost functional of each control cycle in at least one historical control cycle are arranged in chronological order to form a deviation cost time series graph. Local window normalization is performed on the deviation cost time series graph to normalize the amplitude of each element in the deviation cost time series graph relative to its local neighborhood mean. Then, a difference operation is performed on the normalized deviation cost time series graph to obtain the deviation change rate graph. The change rate value of the most recent time point in the deviation change rate graph is compared with the change rate value of the previous historical time point. If the comparison result shows that the most recent change rate value has a continuous increasing trend in the same direction relative to the previous change rate value, it is determined that the next control cycle will continue the same oscillation trend, and a compensation coefficient in the opposite direction to the same oscillation trend is assigned to the predictive compensation prior coefficient matrix.
[0014] In a preferred embodiment, when the predictive compensation prior coefficient matrix is fed back as a feedback compensation amount to the first step of the feature extraction stage, it is specifically used to perform the following correction operation: The original scaling transformation coefficients multiplied by each signal in the multi-source sensing signal before feature extraction are corrected to form a reference bias, and the weighting ratio of each feature dimension in the feature tensor during subsequent normalization processing is corrected to form a feature sensitivity weight.
[0015] The technical effects and advantages of this invention are as follows: This invention constructs a unified feature tensor to carry the dynamic relationship of cross-coupling between the high-voltage feedback signal and the filament current feedback signal. This quantifies the bidirectional physical coupling relationship between the high-voltage electric field and thermionic emission into a numerical input that can be directly processed by the edge-end intelligent inference model, thereby enabling the control system to have real-time perception capabilities of nonlinear coupling states. Based on this, the edge-end intelligent inference model calculates the dynamic confidence weight allocation array of each control loop in real time according to the actual operating conditions represented by the current feature tensor. This allows the intervention priority and adjustment intensity of the voltage regulation loop, current regulation loop, and integrated compensation loop to adaptively adjust with changes in operating conditions, rather than being fixed to preset values under a specific typical operating condition. When the X-ray source performs rapid switching between different voltage and current levels, the control weights of each control loop can be reallocated in real time according to changes in operating conditions. The control strategy no longer relies on a preset linear time-invariant model, but dynamically allocates the control weights of each loop according to the actual coupling state, fundamentally avoiding control mismatch problems caused by system characteristics deviating from the preset model, and achieving adaptive adjustment of the beam stabilization control strategy across the entire operating range.
[0016] This invention re-acquires X-ray beam parameters and high-voltage output parameters after the actual power output is completed, and calculates the steady-state deviation cost functional based on the deviation between the actual output value and the preset target value. This quantifies the control effect of each control cycle into a comprehensive evaluation index that includes transient error, cumulative error, and error change trend, providing a precise quantitative basis for subsequent adjustments to the control strategy. After this steady-state deviation cost functional is input into the intelligent inference model, the model combines the deviation cost functionals of the current cycle and historical cycles to predict the fluctuation trend of the next cycle and generates a predictive compensation prior coefficient matrix. This allows compensation in the opposite direction to the fluctuation trend to be applied before the control action of the next control cycle is executed. This shifts the control logic from adjusting based on feedback after the deviation occurs to predictive compensation based on the deviation change trend, improving the response speed and suppression capability of the control system to load changes and external disturbances.
[0017] This invention feeds back a predictive compensation prior coefficient matrix as a feedback compensation amount to the first step of the feature extraction stage. This corrects the reference bias and feature sensitivity weights when acquiring multi-source sensor signals in the next control cycle. The feedback compensation is applied at the very beginning of the closed-loop chain, before the sensor signal enters the feature extraction stage, rather than only at the actuator output. This mechanism allows the compensation information calculated based on the bias cost functional in the previous cycle to directly adjust the scaling reference of the original signal and the contribution ratio of each dimension of features in the feature tensor in subsequent cycles. This changes the system's sensitivity to sensor signals and the direction of feature extraction at the source level, ensuring that the feature tensor construction in the next cycle carries predictive correction information from the outset. This information is then sequentially passed to subsequent stages such as confidence weight allocation, modulation timing vector generation, and cost functional calculation, forming a deep closed-loop correction that spans the entire chain from perception to decision-making to execution to evaluation to feedback. This comprehensively improves the stability and robustness of X-ray source beam control under strongly nonlinear dynamic conditions. Attached Figure Description
[0018] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a schematic diagram of a beam closed-loop cooperative control method based on an X-ray source control system, as described in this invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0020] Reference Figure 1 The following examples were obtained: Example 1: A beam closed-loop cooperative control method based on an X-ray source control system, comprising the following steps: The first step is to acquire multi-source sensor signals of the X-ray source in real time and in parallel. The multi-source sensor signals include at least the high-voltage feedback signal, the filament current feedback signal, the actual beam current detection signal, and the die temperature detection signal. The multi-source sensor signals are then subjected to time-frequency domain feature extraction and normalization to construct a feature tensor that characterizes the current transient multi-physics coupling state. The feature tensor uniformly carries the cross-coupling dynamic relationship between the high-voltage feedback signal and the filament current feedback signal. The second step is to input the feature tensor into the intelligent inference model pre-deployed at the edge. The intelligent inference model calculates the dynamic confidence weight allocation array of each control loop in real time based on the working condition represented by the current feature tensor. The control loop includes at least a voltage regulation loop, a current regulation loop, and a comprehensive compensation loop. The dynamic confidence weight allocation array represents the intervention priority and regulation intensity allocation value of each control loop in the current control cycle. The third step is to generate an adaptive modulation and decomposition timing vector for the high-frequency inverter circuit within the current control cycle based on the dynamic confidence weight allocation array. The adaptive modulation and decomposition timing vector contains a set of timing parameters for at least one pulse width modulation drive signal and one pulse frequency modulation drive signal. The adaptive modulation and decomposition timing vector is then output to the high-frequency inverter circuit to drive the X-ray source to perform power output. The fourth step is to re-acquire the actual output X-ray beam parameters and high voltage output parameters after the power output is completed. Based on the deviation between the actual output X-ray beam parameters and the preset target beam parameters, the steady-state deviation cost functional of the current control cycle is calculated. The steady-state deviation cost functional is a quantitative evaluation value that comprehensively reflects the transient error, cumulative error and error change trend in the current control cycle. The fifth step is to input the steady-state deviation cost functional into the intelligent inference model. The intelligent inference model combines the steady-state deviation cost functional of the current control cycle with the historical deviation cost functional of at least one control cycle to predict the fluctuation trend of the next control cycle and generate a predictive compensation prior coefficient matrix accordingly. The sixth step involves feeding the predictive compensation prior coefficient matrix back to the feature extraction stage in the first step as a feedback compensation quantity. This corrects the reference bias and feature sensitivity weights when acquiring multi-source sensor signals in the next control cycle, thus forming a complete closed-loop control link consisting of feature tensor construction, dynamic confidence weight allocation array calculation, adaptive modulation solution time-series vector generation, steady-state deviation cost functional calculation, and predictive compensation prior coefficient matrix feedback. This enables adaptive and stable output of the X-ray source beam.
[0021] When acquiring multi-source sensor signals in real-time parallel during X-ray source operation, there are high-voltage feedback signals, filament current feedback signals, actual beam current detection signals, and die temperature detection signals. The high-voltage feedback signal typically ranges from 40 kV to 150 kV, the filament current feedback signal from 3 amps to 6 amps, the actual beam current detection signal from 0.5 mA to 10 mA, and the die temperature detection signal from 20°C to 80°C. The multi-source sensor signals also include independent cathode filament power supply voltage signals and anode rotation drive current signals. These are acquired from different channels, reflecting the voltage and current states in the filament power supply circuit, respectively. The anode rotation drive current signal reflects the load current change during anode target rotation, typically ranging from 0.2 amps to 1.5 amps.
[0022] For each sensor signal, time-frequency domain feature extraction and normalization are performed. During time-frequency domain feature extraction, a Fast Fourier Transform (FFT) is performed on each signal. The sampling frequency of the FFT is set to 2.56 times the highest frequency component of the signal, and the number of sampling points is set to 1024 or 2048. This setting is based on the sampling criteria commonly used in industrial frequency converters, which allows for a 28% engineering safety margin beyond the Nyquist sampling theorem's requirement of more than twice the frequency. This margin accommodates the transition band width of the anti-aliasing filter in the analog-to-digital converter and the short-time offset that may occur in the highest frequency component of the actual signal during operating condition fluctuations. The 2.56 multiple is a commonly used empirical value in power electronics control for distinguishing between fundamental and carrier signals. After performing the FFT, the frequency domain amplitude spectrum and frequency domain phase spectrum of each signal are obtained, while the time domain waveform values of each signal are retained as supplementary time-domain features. The frequency domain amplitude spectrum reflects the energy distribution of the signal at different frequency points, the frequency domain phase spectrum reflects the phase difference between each frequency component, and the time domain waveform value reflects the amplitude trajectory of the signal as time changes.
[0023] The DC component, fundamental frequency amplitude, and second harmonic amplitude are extracted from the frequency domain amplitude spectrum as the basic characteristic values of this signal. The DC component value is the value of the frequency domain amplitude spectrum at 0 Hz, representing the DC bias level of the signal; the fundamental frequency amplitude is the amplitude of the frequency domain amplitude spectrum at the fundamental frequency of the signal, which is determined by the periodic frequency of the signal; the second harmonic amplitude is the amplitude of the frequency domain amplitude spectrum at twice the fundamental frequency. For example, when the fundamental frequency of the high-voltage feedback signal is 20 kHz, the fundamental frequency amplitude is taken as the amplitude of the frequency domain amplitude spectrum at 20 kHz, and the second harmonic amplitude is taken as the amplitude of the frequency domain amplitude spectrum at 40 kHz.
[0024] Each dimension of the basic feature values is linearly scaled according to a preset normalization interval, which is either -1.0 to +1.0 or 0.0 to 1.0. During linear scaling, the maximum and minimum values of each basic feature value in historical data are first calculated. Then, the current value is subtracted from the minimum value and divided by the difference between the maximum and minimum values to obtain an initial ratio between zero and one. This initial ratio is then multiplied by the width of the normalization interval and the lower limit of the interval is added to obtain the normalized feature value. After normalization, the values of all dimensions fall within a unified numerical interval, eliminating differences in dimensions and amplitude between different signals.
[0025] The feature tensor is composed of normalized feature values across its various dimensions. This feature tensor contains multiple dimensions, one part of which represents the normalized feature values of each sensing signal, and another part represents the quantized values of the cross-coupling dynamic relationship. These two parts together form the complete feature tensor. The feature tensor carries the cross-coupling dynamic relationship between the high-voltage feedback signal and the filament current feedback signal. This cross-coupling dynamic relationship refers to the fact that fluctuations in the high-voltage feedback signal modulate the filament current feedback signal, while changes in the filament current feedback signal also inversely affect the ripple component in the high-voltage feedback signal. The feature tensor also carries the correlation and coupling relationship between the independent cathode filament supply voltage signal and the anode rotation drive current signal.
[0026] The term "cross-coupling" is used because both the high-voltage feedback signal and the filament current feedback signal are control inputs of the X-ray tube, and there is a bidirectional feedback loop between them. A change in one will be transmitted to the other through a physical mechanism, forming a closed-loop effect. Therefore, "cross-coupling" is used to describe the mutual influence between these bidirectional control quantities. The term "correlated coupling" is used because the cathode filament supply voltage is the input voltage of the filament circuit, and the anode rotation drive current is the load current of the anode rotation motor. There is no direct bidirectional control loop between them; instead, an indirect causal relationship is generated through changes in the equivalent impedance of the X-ray tube. That is, an increase in the cathode filament supply voltage leads to an increase in emitted electrons and a decrease in tube impedance, which in turn causes a corresponding change in the anode rotation drive current. Therefore, "correlated coupling" is used to describe this indirect causal relationship to distinguish it from the bidirectional closed-loop effect of cross-coupling.
[0027] The edge processor is an embedded microprocessor in the X-ray source control system. Its main frequency is no less than 200 MHz and it has a built-in hardware floating-point unit. The feature tensor is constructed by the edge processor. The cross-coupling dynamic relationship means that the fluctuation of the high voltage feedback signal will modulate the filament current feedback signal, and the change of the filament current feedback signal will also affect the ripple component in the high voltage feedback signal.
[0028] The curves of the high-voltage feedback signal versus time and the curves of the filament current feedback signal versus time are considered as response waveforms of two independent vibration modes. The curve of the high-voltage feedback signal is taken from a continuously sampled voltage value sequence within the current control cycle, and the curve of the filament current feedback signal is taken from a continuously sampled current value sequence within the same time window. The length of the sampling time window is set to ten control cycles, with each control cycle corresponding to a time length of 200 microseconds, thus the total window length is two milliseconds. The basis for considering these two signals as vibration mode response waveforms is that the high-voltage circuit and the filament circuit exhibit damped oscillation characteristics similar to damped vibration when subjected to disturbances.
[0029] Cross-correlation convolution is performed on the two response waveforms to obtain their cross-correlation intensity sequence in the time domain. The cross-correlation convolution operation slides the response waveform of the filament current feedback signal along the time axis. At each sliding position, the high-voltage feedback signal waveform and the filament current feedback signal waveform after the slide are multiplied point by point and summed to obtain the cross-correlation intensity value corresponding to that position. The sliding step size is set to the time interval between two sampling points, i.e., two microseconds. As the sliding position moves from the beginning to the end of the window, the cross-correlation intensity value corresponding to each position is calculated sequentially. All cross-correlation intensity values are arranged in order of sliding position to form a cross-correlation intensity sequence. The length of this cross-correlation intensity sequence is the number of sampling points minus the sliding offset plus one. The number of sampling points is two milliseconds divided by two microseconds, resulting in one thousand points. Therefore, the cross-correlation intensity sequence contains nine hundred and ninety-nine values.
[0030] A peak extraction operation is performed on this cross-correlation intensity sequence to obtain the peak time position and peak amplitude. The peak extraction operation compares each value in the cross-correlation intensity sequence with its adjacent values. When a value is greater than both its preceding and following values, this value is marked as a local peak. The largest value among all local peaks is selected as the global peak. The sliding position corresponding to this global peak is the peak time position, and the value of this global peak is the peak amplitude. The peak time position is represented by a sample index, ranging from 0 to 998. When the sample index corresponding to the peak time position is 35, it means that the filament current feedback signal is delayed by 35 sampling steps relative to the high-voltage feedback signal.
[0031] The coupling response delay time between the high-voltage feedback and the filament current feedback is determined based on the peak position. The coupling response delay time is equal to the sample index of the peak position multiplied by the time length corresponding to the sampling step size. When the sample index of the peak position is 35 and the sampling step size is 2 microseconds, the coupling response delay time is 35 multiplied by 2 microseconds, which equals 70 microseconds. The delay time reflects the time required for the high-voltage side fluctuation to be transmitted to the filament current loop.
[0032] The coupling response gain coefficient between the high-voltage feedback and the filament current feedback is determined based on the peak amplitude. The coupling response gain coefficient is equal to the peak amplitude divided by the product of the root mean square (RMS) values of the high-voltage feedback signal and the filament current feedback signal. The RMS value of the high-voltage feedback signal is calculated by taking the square root of the sum of the squares of all voltage samples within the current window, divided by the number of sampling points. The RMS value of the filament current feedback signal is calculated similarly. When the peak amplitude is 0.85, the RMS value of the high-voltage feedback signal is 120 kV, and the RMS value of the filament current feedback signal is 5.5 amperes, the coupling response gain coefficient is 85 divided by the product of 120 and 5.5, which is 85 divided by 660, resulting in 0.129. The gain coefficient reflects the proportionate change in the filament current feedback signal caused by a one-unit change in the high-voltage feedback signal.
[0033] The coupling response delay time and coupling response gain coefficient are incorporated as independent dimensions into the feature tensor. Based on the existing normalized feature values in each dimension, the feature tensor adds two independent dimensions to carry the coupling response delay time and coupling response gain coefficient values, respectively. The coupling response delay time is measured in microseconds, and the coupling response gain coefficient is a dimensionless value. Therefore, the feature tensor actually contains two types of coupling relationship information: one is the dynamic cross-coupling relationship extracted by cross-correlation convolution operations, and the other is the static association relationship reflected by the normalized feature values.
[0034] The intelligent inference model is a lightweight deep neural network model pre-trained offline and deployed in an edge-embedded processor of the X-ray source control system. This edge-embedded processor is a microcontroller unit on the X-ray source control motherboard, with a built-in hardware accelerator for performing matrix multiplication and addition operations. The processor's clock speed is no less than 200 MHz, and its built-in random access memory (RAM) capacity is no less than 512 kilobytes, while its flash memory capacity is no less than 2 megabytes, used to store the model's network weights and biases. The lightweight deep neural network model refers to a compact network structure with no more than 3 hidden layers and no more than 64 neurons per layer. The total number of parameters in this structure is controlled within 50,000, meeting the storage and computing resource limitations of the embedded processor. The model adopts a fully connected feedforward network architecture. The number of nodes in the input layer equals the total dimension of the feature tensor. The hidden layers use rectified linear units (RCUs) as activation functions, setting input values less than zero to zero and leaving input values greater than or equal to zero unchanged. The output layer uses a linear activation function to directly output three consecutive values.
[0035] The calculation process for the dynamic confidence weight allocation array is as follows: The feature tensor is fed into the deep neural network model as input data for forward inference. During forward inference, the normalized feature values of each dimension in the feature tensor are first multiplied by the weight matrix from the input layer to the first hidden layer, and then a bias vector is added to obtain the weighted sum vector of the first hidden layer. Each element of the weighted sum vector is then processed through a modified linear unit activation function to obtain the output vector of the first hidden layer. The output vector of the first hidden layer is then multiplied by the weight matrix between the second hidden layer and a bias vector is added. After processing by an activation function, the output vector of the second hidden layer is obtained, and so on until the output layer. The output layer directly outputs three values, which are mapped to the confidence weight values corresponding to the voltage regulation circuit, the current regulation circuit, and the comprehensive compensation circuit, respectively.
[0036] The confidence weight values range from 0.0 to 1.0. The sum of the three confidence weight values equals 1.0, which is normalized to one. The confidence weight value of the voltage regulation circuit reflects the reliability of stabilizing the beam current by adjusting the high voltage under the current operating conditions. The confidence weight value of the current regulation circuit reflects the reliability of stabilizing the beam current by adjusting the filament current. The confidence weight value of the comprehensive compensation circuit reflects the reliability of simultaneously adjusting the high voltage and filament current with additional feedforward compensation.
[0037] The offline training process of the deep neural network model is completed in the host computer of the X-ray source control system. After training, the network weight parameters and bias parameters are burned into the flash memory of the edge embedded processor in binary format. The training dataset used for offline training comes from the operation records of the X-ray source under various operating conditions. These operation records contain feature tensor samples and their corresponding optimal confidence weight assignment labels. The acquisition of feature tensor samples covers the entire operating voltage range and the entire operating current range of the X-ray source. For the voltage range of 40 kV to 150 kV, one set of samples is acquired in 5 kV increments. For the current range of 0.5 mA to 10 mA, one set of samples is acquired in 0.5 mA increments. Ten sets of samples are acquired under each combination of voltage and current, for a total of 4600 sets of samples (23 voltage levels multiplied by 20 current levels multiplied by 10 times per set). In addition to the feature tensor, each sample also records the confidence weight allocation combination that optimizes the beam stability under the operating condition. The optimal confidence weight allocation combination is selected as the label after trying different weight combinations one by one under the operating condition and measuring the actual beam stability index.
[0038] The training dataset is divided into a training set and a validation set in an 8:2 ratio. The training set contains 3680 samples used to update network weights, while the validation set contains 920 samples used to monitor overfitting during training. Ten samples are collected for each operating condition point, and the average of the three samples with the best stability index is selected as the label for that condition point to reduce the impact of random errors from single measurements on label accuracy. Mean squared error (MSE) is used as the loss function during training. MSE is the average of the squared differences between the three predicted confidence weights of the network output layer and the three true confidence weights of the training labels. The feature tensors of the training input have already undergone dimensionless scaling in the first step of normalization. The three confidence weights of the training labels themselves range from 0.0 to 1.0, and their sum equals 1.0, which are normalized values; therefore, no additional normalization is required for the training data. An adaptive moment estimation (IME) optimization algorithm was used to update the network weights and biases. The initial learning rate was set to 0.001, and it was decayed by multiplying by 0.9 after every 100 training epochs, with a total of 1000 training epochs. After each training epoch, the loss function on the validation set was calculated. Training was terminated early when the loss function on the validation set stopped decreasing for 50 consecutive epochs to prevent overfitting. After training, the network weights and biases were quantized to 8 bits, converting 32-bit floating-point numbers to 8-bit integers to reduce model storage space and improve the forward inference speed of the edge processor. The mean squared error of the confidence weights in the inference output on the validation set after quantization was less than 0.01 compared to before quantization, indicating that the impact of quantization on inference accuracy was within an acceptable range.
[0039] The adaptive modulation decomposition timing vector is generated within the current control cycle based on the dynamic confidence weight allocation array. This vector, geared towards the high-frequency inverter circuit, contains a set of timing parameters for both the pulse width modulation (PWM) drive signal and the pulse frequency modulation (PWM) drive signal. The PWM drive signal adjusts the amplitude of the high-frequency inverter circuit's output power; its switching frequency is determined based on the intersection frequency of the switching losses of the insulated-gate bipolar transistor (IGBT) and the transformer core losses. The PWM drive signal adjusts the output frequency of the high-frequency inverter circuit to match the operating point of the resonant circuit; its periodic reference frequency is determined based on the resonant frequencies of the resonant inductor and capacitor. Both work together to achieve power amplitude adjustment and resonant efficiency optimization, respectively. The high-frequency inverter circuit is a full-bridge DC-AC converter with a DC bus voltage of 300V to 600V. Its output is connected to the primary winding of the high-voltage transformer of the X-ray tube. During adaptive modulation and timing vector generation, the edge-embedded processor reads three weight values from the dynamic confidence weight allocation array and determines the corresponding adjustment amplitude for the pulse width modulation (PWM) channel and pulse frequency modulation (PWM) channel based on the confidence weights of the voltage regulation loop, current regulation loop, and comprehensive compensation loop, respectively. When the confidence weight of the voltage regulation loop is 0.7, the confidence weight of the current regulation loop is 0.2, and the confidence weight of the comprehensive compensation loop is 0.1, the PWM drive signal undertakes the main adjustment task, and the PWM drive signal undertakes the auxiliary adjustment task. The PWM and PWM channels share the same synchronous clock reference, which is generated by a timer unit inside the edge-embedded processor at a frequency of 20 MHz.
[0040] The timing parameters of the pulse width modulation (PWM) drive signal in the adaptive modulation decomposition timing vector include the duty cycle, switching frequency, and dead time. The duty cycle is the percentage of the high-level duration of the PWM drive signal within the entire switching cycle, ranging from 10% to 90%, with a typical value of 50%. The specific value of the duty cycle is dynamically adjusted based on the confidence weight of the voltage regulation loop. When the confidence weight of the voltage regulation loop is higher than 0.6, the duty cycle is increased or decreased by no more than 20% from the typical value. The adjustment is calculated as follows: the target value of the duty cycle equals the typical value multiplied by one, plus the product of the difference between the voltage regulation loop confidence weight and 0.6, and a proportionality coefficient of 0.1. This proportionality coefficient is obtained by fitting experimental data. When the confidence weight is 0.7, the adjustment is positive 10% of the typical value; when the confidence weight is 0.9, the adjustment is positive 20% of the typical value; intermediate values are calculated using a linear ratio, with the slope of this linear relationship being the typical value multiplied by 0.1. The switching frequency is set based on the intersection frequency of two opposing trend curves in the high-frequency inverter circuit: the switching loss of the insulated gate bipolar transistor increases with increasing frequency, and the transformer core loss decreases with increasing frequency. The total loss reaches its minimum at this intersection frequency. The equipment manual provides the data points of these two loss curves. Plotting the two curves on the same coordinate system and reading the x-coordinate of their intersection point is the switching frequency. The intersection frequency is set to 20 kHz, and the switching frequency range is from 15 kHz to 30 kHz.
[0041] The dead time is set based on the turn-off delay of the insulated gate bipolar transistor plus a safety margin. The equipment manual provides a curve showing the change of transistor turn-off delay with collector current and junction temperature. Under the conditions of maximum collector current and maximum allowable junction temperature, the maximum turn-off delay is read as 1.5 microseconds. The safety margin is taken as 1.0 microsecond, which is 70% of this maximum value. The two are added together to obtain a dead time of 2.5 microseconds. The value range of the dead time is 2 microseconds to 4 microseconds.
[0042] The timing parameters of the pulse frequency modulation drive signal in the adaptive modulation demodulation timing vector include the periodic reference frequency and the frequency offset. The periodic reference frequency is set based on the frequency point where the inductive reactance of the resonant inductor and the capacitive reactance of the resonant capacitor in the resonant circuit of the high-frequency inverter circuit are equal, i.e., the frequency value when the inductive reactance value equals the capacitive reactance value. The equipment manual gives the nominal inductive reactance value of the resonant inductor and the nominal capacitive reactance value of the resonant capacitor. Based on the equality relationship between the two, this frequency point is calculated to be 25 kHz, and the periodic reference frequency is set to 25 kHz. The frequency offset is the increase or decrease relative to the periodic reference frequency, and its value is between -5 kHz and +5 kHz, with a typical offset of +1 kHz or -1 kHz.
[0043] The specific value of the frequency offset is determined according to the following piecewise linear relationship based on the confidence weight of the current regulation loop: When the confidence weight is between 0.3 and 0.6, the frequency offset is 0 kHz. When the confidence weight is greater than 0.6, the frequency offset is a positive offset between +1 kHz and +5 kHz that is proportional to the confidence weight. When the confidence weight is less than 0.3, the frequency offset is a negative offset between -1 kHz and -5 kHz that is proportional to the absolute value of the confidence weight. This piecewise linear relationship was obtained by fitting experimental data. The experimental data shows that when the confidence weight is higher than 0.6, using a positive offset can reduce the beam ripple by about 15%. The period of the pulse frequency modulation drive signal is the reference period corresponding to the period reference frequency plus or minus the period change value corresponding to the frequency offset. The reference period is 25 kHz, which corresponds to 40 microseconds. When the frequency offset is +1 kHz, the period change value is 40 microseconds multiplied by 25 kHz, divided by 26 kHz, and the difference is 1.54 microseconds. The actual period is 40 microseconds plus 1.54 microseconds, which is 41.54 microseconds.
[0044] The adaptive modulation decomposition timing vector also includes the time phase offset of each timing parameter relative to the synchronous clock reference. The time phase offset refers to the time delay between the rising edge of the pulse width modulation drive signal and the pulse frequency modulation drive signal and the rising edge of the synchronous clock reference. This delay is determined by the confidence weight of the integrated compensation loop. The time phase offset is set based on the linear mapping relationship between the confidence weight of the integrated compensation loop and the offset. This mapping relationship is as follows: The offset is equal to the confidence weight multiplied by half the switching cycle. When the confidence weight is 0, the offset is 0 microseconds; when the confidence weight is 1.0, the offset is half the switching cycle, i.e., 25 microseconds. Intermediate values are calculated linearly. The slope of this linear mapping is determined by the maximum allowable phase shift of the high-frequency inverter circuit in the equipment manual. The maximum phase shift cannot exceed 50% of the switching cycle; exceeding this value will result in commutation failure of the inverter circuit. When the confidence weight of the integrated compensation circuit is 0.5, the time phase offset is set to one-quarter of the synchronous clock reference period. That is, for a switching frequency of 20 kHz, the period is 50 microseconds, and the time phase offset is 12.5 microseconds. The value of the time phase offset ranges from 0 to the entire switching cycle. When the confidence weight of the integrated compensation circuit increases from 0.5 to 0.8, the time phase offset linearly increases to half the switching cycle, i.e., 25 microseconds.
[0045] The time phase offset creates a fixed phase difference between multiple drive signals, preventing current surges in the high-frequency inverter circuit when pulse width modulation (PWM) and pulse frequency modulation (PWM) switch simultaneously. All timing parameters in the adaptive modulation decoding timing vector are arranged in an ordered sequence according to the clock reference. PWM and PWM drive signals are independently output to the corresponding driver chip inputs in the high-frequency inverter circuit. The pulse width modulation output pin of the edge-embedded processor outputs a logic-level drive signal to the driver chip input. The driver chip receives this logic-level signal and executes... Electrical isolation and power amplification are performed to generate a gate drive pulse with sufficient drive current and matching voltage level. The high-level amplitude of the gate drive pulse is set to 15 volts to match the gate turn-on voltage requirement of the insulated gate bipolar transistor, and the low-level amplitude is set to -5 volts to ensure reliable turn-off. This causes the power switching devices in the high-frequency inverter circuit to turn on and off according to the set timing sequence, thereby converting the DC bus voltage into a high-frequency AC voltage and outputting it to the primary side of the high-voltage transformer. After rectification, the secondary side of the high-voltage transformer outputs a controllable DC high voltage, which is applied to both ends of the X-ray tube to drive the X-ray source to perform power output.
[0046] After the power output is completed, the actual output X-ray beam parameters and high-voltage output parameters are reacquired. The actual output X-ray beam parameters are read by the beam detector in the X-ray source control system, and the high-voltage output parameters are read by the high-voltage divider. Both are digital values obtained by analog-to-digital conversion. The preset target beam parameter is the desired beam value set by the operator according to the scanning protocol. For example, the preset target beam parameter is 5 mA for chest CT scans and 8 mA for industrial casting inspection. The deviation between the actual output X-ray beam parameter and the preset target beam parameter is equal to the actual value minus the preset value. A positive deviation indicates that the actual beam is higher than the target value, and a negative deviation indicates that the actual beam is lower than the target value. The steady-state deviation cost functional is a quantitative evaluation value that comprehensively reflects the transient error, cumulative error, and error change trend within the current control cycle. The transient error is the absolute value of the real-time deviation in the current control cycle, the cumulative error is the sum of the absolute values of all deviations from the start time to the current control cycle, and the error change trend is the rate at which the deviation increases or decreases over time.
[0047] Before calculating the steady-state deviation cost functional, a target beam parameter tolerance band is pre-defined. This tolerance band is based on the beam stability requirements of X-ray imaging quality. According to the beam stability index specified in the X-ray source equipment manual, the beam stability requirement is ±5% of the preset target beam parameter. Therefore, the tolerance band is the interval formed by extending 5% above and below the preset target beam parameter. When the preset target beam parameter is 5 mA, the tolerance band is between 4.75 mA and 5.25 mA. When the absolute value of the deviation between the actual output X-ray beam parameters and the preset target beam parameters is within the tolerance band of the target beam parameters, for example, if the actual output value is 5.1 mA and the absolute value of the deviation is 0.1 mA, which is between 4.75 mA and 5.25 mA, then it is determined that the current control cycle has reached a stable output state. At this time, the step of generating the predictive compensation prior coefficient matrix is skipped, and the predictive compensation prior coefficient matrix is directly identified as a matrix in which all elements are zero. The dimension of this matrix is the same as the dimension of the predictive compensation prior coefficient matrix in the fifth step. This matrix in which all elements are zero is used as the feedback compensation amount to be fed back to the first step, so as to keep the current adaptive modulation solution timing vector unchanged.
[0048] The steady-state deviation cost functional calculated for each historical control cycle is stored as a sequence of deviation values in chronological order. This sequence is stored in a pre-defined circular buffer within the random access memory of the edge-embedded processor. The buffer's capacity is set to store the steady-state deviation cost functional values for the most recent 100 control cycles. When the number of stored values exceeds 100, the oldest data is overwritten by the newest data. The specific method for obtaining the error change trend in the steady-state deviation cost functional is to extract a subsequence of deviation values from the deviation value sequence for at least one consecutive historical control cycle preceding the current control cycle. The length of the extracted deviation value subsequence is set to the length of the 20 consecutive historical control cycles preceding the current control cycle. That is, the deviation values from the (N-20)th to the (N-1)th control cycle are taken to form the subsequence, where N is the sequence number of the current control cycle. Linear predictive coding analysis is performed on this deviation value subsequence. For example, when the duration of a single control cycle is 200 microseconds, the sampling time window corresponding to the 20 consecutive historical control cycles is 4 milliseconds, the sampling frequency is 5000 Hz, and the frequency resolution is 250 Hz, which is sufficient to distinguish the differences in frequency components above the threshold of 300 Hz. The subsequence length of 20 points meets the requirements of subsequent frequency comparison operations.
[0049] Linear predictive coding analysis uses a linear combination of the preceding values in a subsequence of deviation values to predict the following value. The prediction coefficients are obtained by minimizing the sum of squared prediction errors. This is achieved by constructing an autocorrelation matrix (the Topplitz matrix) that is the product of the prediction order and the prediction order. The element in the i-th row and j-th column is the autocorrelation coefficient between the i-th and j-th values in the deviation subsequence. The prediction coefficients are then obtained by multiplying the inverse of this autocorrelation matrix by the autocorrelation vector. The prediction order is set to one-tenth of the length of the deviation subsequence. When the length of the deviation subsequence is 20, the prediction order is set to 2. In this case, the autocorrelation matrix is a 2x2 square matrix. The condition for the existence of its inverse matrix is that its determinant is not zero, meaning the two autocorrelation coefficients are not equal. This condition is always satisfied in real signals, therefore there is no risk of division by zero. After linear predictive coding analysis is completed, a set of prediction coefficients and corresponding prediction error sequences are obtained. Based on these analysis results, the power spectral density distribution characteristics of the deviation value subsequence are extracted. The power spectral density distribution characteristics are used to calculate the frequency response of the prediction error filter. The denominator of the filter's transfer function is 1 minus the sum of the products of each prediction coefficient and the corresponding order delay operator. After substituting the frequency variable into the transfer function, the amplitude is taken and then squared to obtain the power spectral density value. The power spectral density sequence is obtained by calculating it sequentially at all frequency points. Then, the sequence is traversed from smallest to largest frequency to find the frequency point with the largest amplitude as the main peak frequency value. This main peak frequency value is the dominant frequency of deviation fluctuation.
[0050] For example, when the power spectral density distribution of the deviation value subsequence reaches its peak at 3 Hz, the dominant frequency of the deviation fluctuation is 3 Hz. This dominant frequency is compared with a preset fluctuation frequency threshold. The preset fluctuation frequency threshold is set based on a 50% safety margin, which is the maximum of the cooling fan rotation frequency and the anode rotation frequency in the X-ray source. The typical value of the cooling fan rotation frequency is 50 Hz to 150 Hz, and the typical value of the anode rotation frequency is 100 Hz to 200 Hz. The maximum value of the two is 200 Hz, which, after a 50% increase, becomes 300 Hz. Therefore, the preset fluctuation frequency threshold is 300 Hz. This threshold makes the comparison results practically distinguishable: when the dominant frequency is less than or equal to 300 Hz, the fluctuation falls within the range of low-frequency disturbances caused by known mechanical rotation; when the dominant frequency is greater than 300 Hz, it indicates abnormal fluctuations exceeding the known disturbance frequency range, requiring close monitoring by the controller.
[0051] The comparison results are divided into two categories: when the dominant frequency of deviation fluctuation is less than or equal to the preset fluctuation frequency threshold, the fluctuation belongs to the known low-frequency disturbance range, and the trend weighting coefficient is taken as a small value of 0.3; when the dominant frequency of deviation fluctuation is greater than the preset fluctuation frequency threshold, the fluctuation belongs to the abnormal medium-high frequency disturbance range, and the trend weighting coefficient is taken as a large value of 0.9. The trend weighting coefficient is set to 0.3 because, in experiments where the trend term weight was gradually increased from 0 to 1.0 under typical operating conditions, the beam ripple suppression effect was optimal when the weight was 0.3. The value of 0.9 is set because, in experiments simulating abnormal fluctuation conditions, the system response time was shortest when the weight was 0.9. Increasing the weight to 1.0 did not significantly improve the response time but significantly increased the overshoot. Therefore, 0.9 is the optimal value. When the trend weighting coefficient is 0.9, the evaluation weight of the error change trend in the steady-state deviation cost functional is amplified, making the controller more concerned with the rate of change of the current deviation. Conversely, when the trend weighting coefficient is 0.3, the contribution of the error change trend to the steady-state deviation cost functional is reduced, and the controller mainly focuses on transient and cumulative errors.
[0052] The deviation cost time series graph is an ordered sequence formed by arranging the steady-state deviation cost functional of the current control cycle and the historical steady-state deviation cost functionals corresponding to each control cycle in at least one historical control cycle in chronological order. Each element in the sequence corresponds to the steady-state deviation cost functional value of one control cycle. The historical steady-state deviation cost functionals are stored values taken from the deviation value sequence in chronological order. The deviation value sequence is stored in the circular buffer of the edge embedded processor. When the current control cycle number is N, the deviation cost time series map includes the steady-state deviation cost functional of the Nth cycle and the historical steady-state deviation cost functionals of the (N-1), (N-2)th to (NM)th cycles, where M is the preset number of historical cycles. The value of M is based on the minimum number of control cycles required to cover a complete cycle of mechanical disturbance in the X-ray source. The typical value of the cooling fan rotation cycle is 20 milliseconds, and the duration of a single control cycle is 200 microseconds. Therefore, M is set to 100. This time window also includes multiple complete cycles of the anode rotation frequency (100 Hz to 200 Hz corresponds to 5 milliseconds to 10 milliseconds).
[0053] Local window normalization is performed on the deviation cost time series graph to normalize the magnitude of each element relative to its local neighborhood mean. Local window normalization involves taking L elements before and after each element in the deviation cost time series graph to form a local window, calculating the arithmetic mean of all elements within this local window, subtracting the local window mean from the original value of the element to obtain the difference, and then dividing this difference by the standard deviation of the elements within the local window to obtain the normalized value. The value of the local window length L is determined by testing the trend prediction accuracy within a window length range of 3 to 21. When the window length is 9, the trend prediction accuracy reaches its maximum value of 92%, therefore L is set to 9. At this point, the window span is 9 control cycles, or 1.8 milliseconds, which is much smaller than the 20 millisecond time window covered by M. The normalization result reflects local deviation changes without being affected by the global trend. When a local window is located at the edge of the graph, there are insufficient neighboring elements at the edge. In this case, the mean and standard deviation are calculated using the actual elements within the window. The calculation can be performed if there are at least three actual elements within the window; otherwise, the normalization process is skipped, and the original values remain unchanged. This boundary processing only affects the first element at the beginning of the graph and the last element at the end. These two elements do not participate in subsequent trend determination operations and therefore do not affect the final determination result. For example, in a time series graph of deviation cost, the values at three consecutive time points are 0.85, 0.92, and 1.01. For the middle element 0.92, one element before and one element after it are taken to form a window containing the values 0.85, 0.92, and 1.01. The window mean is 0.93, and the window standard deviation is 0.08. After normalization, the value is 0.92 minus 0.93, which equals -0.01. Dividing this by 0.08 gives -0.125.
[0054] A difference operation is performed on the normalized deviation cost time series graph to obtain a deviation change rate graph. The difference operation involves subtracting the values of two adjacent time points in the normalized deviation cost time series graph; the difference between the values of the later and earlier time points is taken as the rate of change between those two time points. All differences are arranged in chronological order to form the deviation change rate graph. The deviation change rate graph is one element shorter than the normalized deviation cost time series graph. When three consecutive values in the normalized deviation cost time series graph are -0.125, 0.250, and 0.620, the first rate of change is 0.250 minus -0.125, which equals 0.375, and the second rate of change is 0.620 minus 0.250, which also equals 0.370. Both of these rates of change are positive and close in value, indicating that the deviation is continuously increasing.
[0055] The rate of change value at the most recent time point in the deviation rate of change spectrum is compared with the rate of change value at the previous historical time point. If the comparison result shows that the most recent rate of change value has a continuous increasing trend in the same direction relative to the previous rate of change value, it is determined that the next control cycle will continue the trend of fluctuation in the same direction, and a compensation coefficient in the opposite direction of the trend of fluctuation is assigned to the predictive compensation prior coefficient matrix. The most recent time point refers to the last element in the deviation rate of change spectrum, and the previous historical time point refers to the second to last element in the deviation rate of change spectrum. The condition for determining a continuous increasing trend in the same direction is that the most recent rate of change value and the previous rate of change value are both positive or both negative, and the absolute value of the most recent rate of change value is greater than the absolute value of the previous rate of change value. When the most recent rate of change value is 0.370 and the previous rate of change value is 0.375, both are positive values and 0.370 is less than 0.375, and the absolute value has not increased. Therefore, it is determined that the continuous increasing trend in the same direction is not satisfied, and the compensation coefficient in the predictive compensation prior coefficient matrix is set to zero.
[0056] When the most recent rate of change is 0.420 and the previous rate of change is 0.375, both are positive and 0.420 is greater than 0.375, indicating a continuous increasing trend in the same direction. This means the deviation will continue to increase. In this case, a compensation coefficient in the opposite direction to the same-direction fluctuation trend, i.e., a negative compensation coefficient, is assigned to the predictive compensation prior coefficient matrix.
[0057] The specific value of the compensation coefficient is determined proportionally to the difference between the most recent rate of change and the previous rate of change. The ratio is such that for every 0.01 increase in the rate of change difference, the absolute value of the compensation coefficient increases by 0.01. When the rate of change difference is 0.045, the compensation coefficient is -0.045, and the maximum absolute value does not exceed 0.20. If it exceeds 0.20, it is clamped to 0.20. The slope and limiting value of this ratio were obtained through experimental testing under the conditions of a rated operating voltage of 100 kV and a rated operating current of 5 mA for the X-ray source. The test results show that an absolute value of the compensation coefficient exceeding 0.20 will lead to overcompensation and cause reverse fluctuations; therefore, the limiting value is 0.20. When the compensation coefficient is negative, it is written into the correction term of the corresponding modulation parameter in the predictive compensation prior coefficient matrix. In the next step, it is used to reduce the duty cycle or increase the frequency offset to suppress further beam current increase. When the compensation coefficient is positive, it indicates a continuous decreasing trend in the same direction. A positive compensation coefficient is used to increase the duty cycle or decrease the frequency offset to suppress further beam current decrease.
[0058] The predictive compensation prior coefficient matrix serves as feedback compensation to the first step of feature extraction. Each element in this matrix corresponds to a specific value for a correction operation. The dimension of the predictive compensation prior coefficient matrix is equal to the total dimension of the feature tensor plus the total number of channels in the multi-source sensor signals. The first part of the matrix corresponds to the scaling transformation coefficient correction values for each channel of the sensor signals, and the latter part corresponds to the weighted ratio correction values for each dimension of the feature tensor. When the total dimension of the feature tensor is 32 and the total number of multi-source sensor signal channels is 5, the predictive compensation prior coefficient matrix contains 37 elements, of which the first 5 elements are the scaling transformation coefficient correction values, and the last 32 elements are the weighted ratio correction values.
[0059] The original scaling conversion coefficients multiplied by each signal in the multi-source sensing signal before feature extraction are corrected to form a reference bias. The original scaling conversion coefficients are fixed ratios multiplied when converting the original physical quantity values of each sensing signal into uniform dimensionless values. Each sensing signal corresponds to an independent original scaling conversion coefficient. The original physical quantity unit of the high-voltage feedback signal is kilovolt, and its original scaling conversion factor is typically 0.001, converting 100 kilovolts to 0.1; the original physical quantity unit of the filament current feedback signal is ampere, and its original scaling conversion factor is typically 0.1, converting 5 amperes to 0.5; the original physical quantity unit of the actual beam current detection signal is milliampere, and its original scaling conversion factor is typically 1.0, converting 5 milliamperes to 5.0; the original physical quantity unit of the die temperature detection signal is degree Celsius, and its original scaling conversion factor is typically 0.01, converting 50 degrees Celsius to 0.5; the original physical quantity unit of the independent cathode filament power supply voltage signal is volt, and its original scaling conversion factor is typically 0.01, converting 20 volts to 0.2; the original physical quantity unit of the anode rotation drive current signal is ampere, and its original scaling conversion factor is typically 0.2, converting 0.5 amperes to 0.1.
[0060] The position of the scaling transformation coefficient correction value in the predictive compensation prior coefficient matrix indicates that the current scaling transformation coefficient needs to be increased, while a negative value indicates that the current scaling transformation coefficient needs to be decreased. The correction value is directly added to the current scaling transformation coefficient, and the result is used as the scaling transformation coefficient for the next cycle. When the original scaling transformation coefficient of the high-voltage feedback signal is 0.001 and the corresponding correction value is -0.0001, the corrected scaling transformation coefficient is 0.0009. The 100 kV high-voltage signal acquired in the next cycle will be converted to 0.09 instead of 0.1, which is equivalent to adding a negative reference bias to the high-voltage feedback signal before feature extraction, causing the overall high-voltage feature value to shift downward. Conversely, when the correction value is +0.0002, the corrected scaling transformation coefficient is 0.0012, and the 100 kV high-voltage signal is converted to 0.12, which is equivalent to adding a positive reference bias, causing the overall high-voltage feature value to shift upward. Each correction value is limited to no more than 20% of the current value of the original scaling factor. This limit is based on experimental data conducted under rated operating voltage of 100 kV and rated operating current of 5 mA. Experiments show that a correction exceeding 20% will cause the sensor signal to exceed the optimal range of the analog-to-digital converter. The optimal range is 20% to 80% of the full scale of the analog-to-digital converter. When this range is exceeded, the sampling resolution decreases by about 30%. Therefore, the limit is set at 20%.
[0061] The weighting ratio of each feature dimension in the modified feature tensor during subsequent normalization is used to form the feature sensitivity weight. The weighting ratio of each feature dimension during subsequent normalization is the contribution proportion of the normalized feature value of that dimension in the feature tensor, and the sum of the weighting ratios of all dimensions equals 1.0. When the feature tensor contains 32 features, the initial value of the weighting ratio of each dimension is 1 divided by 32, which equals 0.03125, meaning that each dimension has an equal contribution proportion in the initial state. The position of the corresponding weighting ratio correction value in the predictive compensation prior coefficient matrix is indicated. A positive value indicates that the weighting ratio of the current dimension needs to be increased, and a negative value indicates that the weighting ratio of the current dimension needs to be decreased. The correction value is directly added to the current weighting ratio, and the result is used as the weighting ratio for the next cycle. When the current weighting ratio of the first dimension feature corresponding to the high-voltage feedback signal is 0.03125 and the corresponding correction value is +0.005, the corrected weighting ratio is 0.03625. This dimension feature has a larger contribution ratio in the normalization process, which enhances the influence of the high-voltage feedback signal on all subsequent calculations.
[0062] When the current weighting ratio of the second-dimensional feature corresponding to the filament current feedback signal is 0.03125 and the corresponding correction value is -0.003, the corrected weighting ratio is 0.02825. This dimension of the feature has a smaller contribution weight during normalization, thus reducing the influence of the filament current feedback signal on all subsequent calculations. After the weighting ratio correction is completed, the corrected weighting ratios of all dimensions are summed to obtain a total. Then, the weighting ratio of each dimension is divided by this total, so that the sum of all corrected weighting ratios is equal to 1.0 again. This normalization operation is a mandatory constraint, ensuring that the contribution weight of each dimension of the feature tensor always meets the normalization condition in subsequent calculations. The correction range of the weighting ratio shall not exceed ±50% of the current weighting ratio value. This limit is based on experimental data conducted under the conditions of rated operating voltage of 100 kV and rated operating current of 5 mA. The experiment shows that when the correction range exceeds 50%, the contributions of each dimension of the feature tensor become unbalanced, causing a sharp jump in the output confidence weight allocation array of the intelligent inference model, with the jump range exceeding 30%. Therefore, the limit is set at 50%.
[0063] When the current weighting ratio of a certain feature dimension is 0.03125 and the correction value is +0.02, since 0.02 exceeds 0.03125 multiplied by 50% (0.015625), the correction value is clamped to 0.015625. The corrected weighting ratio is 0.03125 plus 0.015625, which equals 0.046875. Through this feedback correction operation, each element in the predictive compensation prior coefficient matrix acts on the original scaling transformation coefficients and the weighting ratios of each dimension of the feature tensor of each sensor signal. This ensures that the multi-source sensor signals acquired in the next step already carry correction information based on the previous cycle bias cost functional when entering the feature extraction stage, thereby adjusting the benchmark and sensitivity of feature extraction at the source. After the five steps of feature tensor construction, dynamic confidence weight allocation array calculation, adaptive modulation solution time-series vector generation, steady-state deviation cost functional calculation, and predictive compensation prior coefficient matrix feedback are executed in sequence, the feedback quantity is sent back to the input of the feature tensor construction step, forming a complete closed-loop control link.
[0064] The above algorithms or formulas are all dimensionless and numerical calculations, and the results are obtained by software simulation based on a large amount of collected data to obtain the most recent real-world results. The preset parameters are set by those skilled in the art according to the actual situation.
[0065] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0066] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0067] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0068] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A beam closed-loop cooperative control method based on an X-ray source control system, characterized in that, Includes the following steps: The first step is to acquire multi-source sensor signals of the X-ray source in real time and in parallel. The multi-source sensor signals include at least the high-voltage feedback signal, the filament current feedback signal, the actual beam current detection signal, and the die temperature detection signal. The multi-source sensor signals are then subjected to time-frequency domain feature extraction and normalization to construct a feature tensor that characterizes the current transient multi-physics coupling state. The feature tensor uniformly carries the cross-coupling dynamic relationship between the high-voltage feedback signal and the filament current feedback signal. The second step is to input the feature tensor into the intelligent inference model pre-deployed at the edge. The intelligent inference model calculates the dynamic confidence weight allocation array of each control loop in real time based on the working condition represented by the current feature tensor. The control loop includes at least a voltage regulation loop, a current regulation loop, and a comprehensive compensation loop. The dynamic confidence weight allocation array represents the intervention priority and regulation intensity allocation value of each control loop in the current control cycle. The third step is to generate an adaptive modulation and decomposition timing vector for the high-frequency inverter circuit within the current control cycle based on the dynamic confidence weight allocation array. The adaptive modulation and decomposition timing vector contains a set of timing parameters for at least one pulse width modulation drive signal and one pulse frequency modulation drive signal. The adaptive modulation and decomposition timing vector is then output to the high-frequency inverter circuit to drive the X-ray source to perform power output. The fourth step is to re-acquire the actual output X-ray beam parameters and high voltage output parameters after the power output is completed. Based on the deviation between the actual output X-ray beam parameters and the preset target beam parameters, the steady-state deviation cost functional of the current control cycle is calculated. The steady-state deviation cost functional is a quantitative evaluation value that comprehensively reflects the transient error, cumulative error and error change trend in the current control cycle. The fifth step is to input the steady-state deviation cost functional into the intelligent inference model. The intelligent inference model combines the steady-state deviation cost functional of the current control cycle with the historical deviation cost functional of at least one control cycle to predict the fluctuation trend of the next control cycle and generate a predictive compensation prior coefficient matrix accordingly. The sixth step is to use the predictive compensation prior coefficient matrix as a feedback compensation amount and feed it back to the feature extraction stage in the first step to correct the reference bias and feature sensitivity weight when acquiring multi-source sensor signals in the next control cycle.
2. The beam closed-loop coordinated control method based on an X-ray source control system according to claim 1, characterized in that, The multi-source sensing signals also include independent cathode filament power supply voltage signals and anode rotation drive current signals. The specific implementation methods for time-frequency domain feature extraction and normalization processing are as follows: For each signal in the multi-source sensing signal, perform a Fast Fourier Transform to extract the frequency domain amplitude spectrum and frequency domain phase spectrum of the signal, while retaining the time domain waveform values of the signal; use the DC component value, fundamental amplitude value, and second harmonic amplitude value in the frequency domain amplitude spectrum of the signal as the basic characteristic values of the signal. Then, each dimension of the basic feature value is linearly scaled according to a preset normalization interval, so that each scaled dimension value falls within a uniform value range, thereby obtaining the normalized feature values corresponding to each dimension in the feature tensor. In addition to carrying the cross-coupling dynamic relationship between the high voltage feedback signal and the filament current feedback signal, the feature tensor also carries the correlation coupling relationship between the independent cathode filament power supply voltage signal and the anode rotation drive current signal.
3. The beam closed-loop coordinated control method based on an X-ray source control system according to claim 2, characterized in that, The edge processor performs the following operations to construct the feature tensor: The curves of the high-voltage feedback signal versus time and the curves of the filament current feedback signal versus time are respectively regarded as the response waveforms of two independent vibration modes. Perform cross-correlation convolution on the two response waveforms to obtain their cross-correlation intensity sequence in the time domain; Then, perform peak extraction on the cross-correlation intensity sequence to obtain the peak time position and peak amplitude; The coupling response delay time between the high voltage feedback and the filament current feedback is determined based on the peak time position, and the coupling response gain coefficient between the high voltage feedback and the filament current feedback is determined based on the peak amplitude. The coupling response delay time and coupling response gain coefficient are incorporated as independent dimensions into the feature tensor.
4. The beam closed-loop coordinated control method based on an X-ray source control system according to claim 3, characterized in that, The intelligent inference model is a lightweight deep neural network model pre-trained offline. The model is deployed in an edge-embedded processor of the X-ray source control system. The calculation process of the dynamic confidence weight allocation array includes: The feature tensor is fed into the deep neural network model as input data to perform forward inference operations. The output layer of the deep neural network model is directly mapped to the confidence weight values of the voltage regulation circuit, current regulation circuit and comprehensive compensation circuit respectively.
5. The beam closed-loop coordinated control method based on an X-ray source control system according to claim 4, characterized in that, The timing parameters of the pulse width modulation drive signal in the adaptive modulation solution timing vector include the duty cycle, switching frequency, and dead time. The timing parameters of the pulse frequency modulation drive signal include the periodic reference frequency and the frequency offset; The adaptive modulation solution timing vector also includes the time phase offset of each timing parameter relative to the synchronous clock reference.
6. The beam closed-loop coordinated control method based on an X-ray source control system according to claim 5, characterized in that, The steady-state deviation cost functional calculated for each historical control cycle is stored as a sequence of deviation values in chronological order. The specific method for obtaining the error change trend in the steady-state deviation cost functional is as follows: Extract the deviation value subsequence from the deviation value sequence that is at least one consecutive historical control cycle before the current control cycle; perform linear predictive coding analysis on the deviation value subsequence, and extract the power spectral density distribution characteristics of the deviation value subsequence based on the analysis results; The dominant frequency value in the power spectral density distribution characteristics is taken as the dominant frequency of deviation fluctuation. The dominant frequency of deviation fluctuation is compared with the preset fluctuation frequency threshold. Based on the comparison result, a trend weighting coefficient is generated. The trend weighting coefficient is used to amplify or reduce the evaluation weight of the error change trend in the steady-state deviation cost functional.
7. The beam closed-loop coordinated control method based on an X-ray source control system according to claim 6, characterized in that, Before calculating the steady-state deviation cost functional, the target beam parameter tolerance band is pre-defined; When the absolute value of the deviation between the actual output X-ray beam parameters and the preset target beam parameters is within the tolerance band of the target beam parameters, it is determined that the current control cycle has reached a stable output state. The step of generating the predictive compensation prior coefficient matrix is skipped, and the predictive compensation prior coefficient matrix is directly identified as a matrix in which all elements are zero. The dimension of the matrix is the same as that of the predictive compensation prior coefficient matrix in the fifth step. The matrix in which all elements are zero is used as the feedback compensation amount to be fed back to the first step.
8. The beam closed-loop coordinated control method based on an X-ray source control system according to claim 7, characterized in that, When the intelligent inference model combines the steady-state deviation cost functional of the current control cycle with the historical deviation cost functional of at least one control cycle to predict the fluctuation trend of the next control cycle, it specifically performs the following operations: The steady-state deviation cost functional of the current control cycle and the historical steady-state deviation cost functional of each control cycle in at least one historical control cycle are arranged in chronological order to form a deviation cost time series graph. Local window normalization is performed on the deviation cost time series graph to normalize the magnitude of each element in the deviation cost time series graph relative to the mean of its local neighborhood. Perform a difference operation on the normalized time series plot of deviation cost to obtain the plot of deviation change rate; The rate of change at the most recent time point in the deviation rate of change spectrum is compared with the rate of change at the previous historical time point. If the comparison result shows that the most recent rate of change has a continuous increasing trend in the same direction relative to the previous rate of change, it is determined that the next control cycle will continue the trend of fluctuation in the same direction, and a compensation coefficient in the opposite direction to the trend of fluctuation in the same direction is assigned to the predictive compensation prior coefficient matrix.
9. A beam closed-loop coordinated control method based on an X-ray source control system according to claim 8, characterized in that, When the predictive compensation prior coefficient matrix is fed back as the feedback compensation amount to the first step of feature extraction, it is specifically used to perform the following correction operations: The original scaling transformation coefficients multiplied by each signal in the multi-source sensing signal before feature extraction are corrected to form a reference bias, and the weighting ratio of each feature dimension in the feature tensor during subsequent normalization processing is corrected to form a feature sensitivity weight.
