A current pulse generator noise reduction method and device

Abstract

The present invention discloses a current pulse generator noise reduction method and device, which relates to the technical field of current pulse generator noise reduction. Specifically, the method comprises the following steps: removing the main pulse waveform from the filter response sequence, extracting the filter residual response segments of multiple cycles, time-aligning and overlapping the residual response segments of multiple cycles, and constructing a residual energy superposition trend line; based on the residual energy superposition trend line, using a Savitzky-Golay sliding window polynomial fitting derivative algorithm to calculate the trend offset rate parameter, and using a wavelet packet decomposition and Shannon entropy calculation algorithm to generate a disturbance frequency domain diffusion parameter to reflect whether the filter residual energy superposition behavior evolves into self-excited background noise. The present invention solves the problem of difficulty in identifying the filter residual energy superposition under continuous narrow pulses, realizes the accurate identification and dynamic intervention control of self-excited background noise, and improves system stability.

Description

Technical Field

[0001] The present invention relates to the technical field of current pulse generator noise reduction, and in particular to a current pulse generator noise reduction method and device. Background Art

[0002] A current pulse generator is an electronic device used to generate current pulse signals with specific amplitude, frequency, duty cycle, and waveform characteristics. It is widely used in radar systems, communications systems, power electronics control, electromagnetic compatibility testing, and pulsed laser driving. In these applications, the accuracy and stability of the pulse signal are critical to system performance. However, the high-speed switching characteristics of current pulses can easily introduce noise issues such as electromagnetic interference (EMI), ground line noise, voltage spikes, and echo disturbances. This noise not only affects the signal integrity of the system itself but can also interfere with other sensitive circuits or communication modules, resulting in increased bit error rates, system instability, and even component damage. Therefore, to ensure reliable system operation and signal quality, effective noise reduction must be implemented in the current pulse generator's operating process. This noise generation can be suppressed at the source and its propagation and impact can be reduced through control strategies, filtering mechanisms, or structural optimization. This can improve the overall system's interference immunity and operational efficiency.

[0003] Existing noise reduction technologies for current pulse generators primarily address circuit structure optimization, signal processing, and electromagnetic compatibility (EMC) design to mitigate or suppress the noise interference caused by the high-speed switching of pulse signals. At the circuit level, components such as low-pass filters, common-mode inductors, and RC snubber circuits are often added to filter out high-frequency noise components and suppress voltage spikes. Furthermore, layout design involves optimizing routing, properly placing ground planes, and implementing power supply decoupling to reduce the propagation paths of electromagnetic interference. Regarding signal processing, some systems incorporate digital control algorithms or PWM modulation strategies to smooth current variations by adjusting pulse edge slopes and limiting mutation rates, thereby reducing interference caused by high-speed switching. Furthermore, in the EMC design phase, shielding structures, electromagnetic absorption materials, and appropriate housing grounding design are employed to further mitigate the release and coupling effects of electromagnetic radiation noise. The overall noise reduction process typically involves three core steps: noise source analysis, interference path identification and control, and protection of sensitive receivers. The goal is to establish a comprehensive noise suppression mechanism from the "source-path-end" level to enhance the stability and reliability of current pulse generators in complex systems.

[0004] The existing technology has the following deficiencies:

[0005] When a current pulse generator continuously outputs narrow pulses with extremely short pulse widths, especially when the pulse period is shorter than the filter's natural energy dissipation time, residual energy is generated within the filter. Because this residual energy is not fully released before the next pulse arrives, it is repeatedly superimposed, forming equivalent parasitic harmonics or low-amplitude self-excited interference. Because this background noise does not manifest as obvious spikes, distortion, or edge aberrations, but rather as a periodic rise in baseline noise, it is difficult to detect using traditional noise reduction methods such as fixed-band filtering, amplitude clamping, or edge amplitude monitoring. Existing current pulse generator noise reduction technologies cannot identify the formation of this self-excited background noise based on the residual energy superposition behavior of the filter under continuous narrow pulse output. This can lead to blind spots in the noise reduction method during dynamic operation, resulting in hidden interference caused by the superposition of pulse output. This can not only cause downstream circuits to misjudge the reference threshold, leading to false triggering or timing errors, but can also accumulate to form broad-spectrum electromagnetic interference, affecting the overall system's anti-interference performance and stability.

[0006] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0007] The object of the present invention is to provide a current pulse generator noise reduction method and device to solve the problems in the above background technology.

[0008] In order to achieve the above object, the present invention provides the following technical solution: a current pulse generator noise reduction method, specifically comprising the following steps:

[0009] S01. During the continuous narrow pulse output process of the current pulse generator, the pulse width, pulse interval and edge slope parameters of each pulse are collected to obtain the output voltage of the filter in the corresponding cycle, thereby forming a filter response sequence of multiple cycles;

[0010] S02. removing the main pulse waveform from the filter response sequence, extracting filter residual response segments of multiple cycles, performing time alignment and overlapping processing on the residual response segments of multiple cycles, and constructing a residual energy superposition trend line;

[0011] S03. Based on the residual energy superposition trend line, the Savitzky-Golay sliding window polynomial fitting derivative algorithm is used to calculate the trend offset rate parameter, and the wavelet packet decomposition and Shannon entropy calculation algorithm are used to generate the disturbance frequency domain diffusion parameter to reflect whether the residual energy superposition behavior of the filter evolves into self-excited background noise;

[0012] S04, mapping the trend offset rate parameter and the disturbance frequency domain diffusion parameter into a two-dimensional trajectory path, extracting trajectory deformation features using a similar cone path mapping method, and generating a state factor representing the stage of the self-excited background noise formation process;

[0013] S05. Performing stage determination based on the generated state factor and executing intervention measures according to the determination result, including adjusting the pulse output rhythm, inserting gap interference pulses, or applying rhythm disturbance control to block the continuous superposition of the residual energy of the filter.

[0014] Preferably, in S01, the following steps are specifically included:

[0015] During the continuous narrow pulse output process of the current pulse generator, the start time and end time of each pulse are collected based on the trigger clock, the pulse width is calculated by the start and end time, the pulse interval is calculated by the start time difference of adjacent pulses, and the edge slope is calculated by the start edge fitting curve;

[0016] Taking the start time of each pulse as the reference point, the voltage signal waveform at the output of the filter is obtained within a fixed sampling window, including the main pulse and the response segment after the main pulse;

[0017] Repeat the parameter acquisition and voltage acquisition operations for multiple pulse cycles, and record the data of each cycle according to the timestamp number;

[0018] The parameters and response voltage data of multiple cycles are arranged in a preset cycle order to form a filter response behavior sequence on a continuous time axis for subsequent residual energy analysis.

[0019] Preferably, in S02, the following steps are specifically included:

[0020] Determine the time interval corresponding to the main pulse in each cycle, calculate the start and end boundaries of the main pulse based on the pulse start time and pulse width, and remove the filter voltage data in the time interval corresponding to the main pulse in each cycle from the original response sequence;

[0021] In each cycle, a fixed-length residual observation window is set after the main pulse ends, and the voltage sampling points retained in the window are extracted as the residual response segment, and their relative time coordinates relative to the pulse starting point are uniformly recorded;

[0022] The residual response segments extracted from all cycles are aligned according to their start time and overlapped on a unified time axis to generate a cross-cycle residual response distribution set.

[0023] The voltage data corresponding to each time point in the residual response distribution set are averaged to obtain a residual energy evolution curve covering multiple cycles, and the residual energy superposition trend line is constructed based on this curve. The specific construction process is as follows: sliding window processing is performed on the residual energy evolution curve covering multiple cycles, and the data points in each window are fitted into local linear segments to obtain the local slope distribution in the continuous time interval; the slope sequence of multiple local linear segments is smoothed and filtered to eliminate the high-frequency noise offset caused by small disturbances and generate a continuous trend-guiding signal; the trend-guiding signal is sampled at a fixed time interval, and a third-order polynomial is used to fit the data interval before and after each sampling point to form a continuous interpolation expression of the trend curve; all interpolation points are connected to form a continuous change trajectory covering the entire residual response time axis, and this trajectory is used as the superposition trend line of the filter residual energy.

[0024] Preferably, in S03, the process of calculating the trend shift rate parameter based on the residual energy superimposed trend line includes the following steps:

[0025] The Savitzky-Golay sliding window process is performed on the residual energy superposition trend line, and a third-order polynomial fitting is performed within each fixed-length window to generate a local smooth curve;

[0026] At the center of each sliding window, the first-order derivative value of the curve is extracted to form the derivative sequence of the residual trend on the entire time axis;

[0027] The derivative sequence is processed by boundary extension to avoid the interference of edge window fitting distortion on the overall rate continuity;

[0028] The derivative series is smoothed by sliding average to suppress the interference of local mutations on the rate trend;

[0029] The current derivative is extracted at each fixed sampling time point as the trend shift rate parameter to characterize the change amplitude and direction of the residual energy superposition trend.

[0030] Preferably, in S03, the process of generating the disturbance frequency domain diffusion parameter based on the residual energy superposition trend line includes the following steps:

[0031] Multi-level wavelet packet decomposition is performed on the residual energy superposition trend line, and the db4 wavelet basis function is used to split the original signal into multiple sub-band signals;

[0032] Calculate the energy proportion of the corresponding sub-signal in each frequency band to form the energy distribution vector of the complete frequency band;

[0033] Applying Shannon entropy calculation method to the frequency band energy distribution vector, the energy distribution dispersion index is obtained;

[0034] Normalize the Shannon entropy value to a standard value range to facilitate horizontal comparison between different cycles;

[0035] The normalized entropy value is used as the disturbance frequency domain diffusion parameter to reflect the diffusion degree of residual energy in the frequency domain.

[0036] Preferably, in S03, the trend shift rate parameter and the disturbance frequency domain diffusion parameter are used to reflect whether the filter residual energy superposition behavior evolves into self-excited background noise, specifically:

[0037] On the same time axis, the trend offset rate parameter and the disturbance frequency domain diffusion parameter are combined into a two-dimensional time series vector sequence;

[0038] A sliding window cluster analysis is performed on the vector sequence to identify segments where both the trend growth rate and the frequency domain diffusion degree increase simultaneously within a continuous time interval. These segments are marked as suspected self-excitation segments, indicating that the residual energy is continuously increasing and showing a frequency domain diffusion tendency.

[0039] The duration and amplitude of each suspected self-excitation segment are counted to determine whether it exceeds the preset duration threshold and energy diffusion threshold;

[0040] If, in any suspected self-excited segment, the length of continuous sampling time is not less than the set minimum length threshold, and the average value of the trend offset rate parameter in the segment is greater than the trend rate threshold, and the average value of the disturbance frequency domain diffusion parameter is greater than the diffusion threshold, then it is determined that the filter is in the process of self-excited background noise evolution, and subsequent classification and intervention operations are triggered.

[0041] Preferably, in S04, the following steps are specifically included:

[0042] According to the time sequence, the trend deviation rate parameter corresponding to each moment is used as the horizontal coordinate, and the disturbance frequency domain diffusion parameter is used as the vertical coordinate. The two-dimensional trajectory path is constructed in sequence to form a set of trajectory segments reflecting the parameter change process.

[0043] In the two-dimensional trajectory path, the geometric features of each trajectory segment are extracted using a sliding window of fixed length, including local curvature, path bending rate, angular rotation trend and change rate;

[0044] The extracted geometric features are projected into the conical path mapping space, and the contour morphology of each trajectory segment in the space is analyzed and clustered to divide it into multiple continuous evolution state regions;

[0045] According to the evolution order, deformation amplitude and distribution characteristics of each state area in the trajectory, the state factor characterizing the stage of the self-excited background noise formation process is generated and serves as the basis for subsequent classification and intervention.

[0046] Preferably, in S05, the steps specifically include:

[0047] Based on the distribution position of the generated state factors in the multidimensional state space, the stage of the current filter residual energy superposition behavior is determined, and the state factors are divided into the initial fluctuation stage, the evolution critical stage and the continuous self-excitation stage;

[0048] When the state factor is in the initial fluctuation stage, a fixed-amplitude time interval disturbance is introduced into each continuous pulse cycle to change the rhythm arrangement of the continuous pulse output and prevent the filter from accumulating energy under constant excitation.

[0049] When the state factor is in the critical stage of evolution, a gap interference pulse with a set interval is inserted into the continuous pulse output sequence to change the pulse trigger timing and interrupt the residual energy superposition link in the filter;

[0050] When the state factor is in the continuous self-excitation stage, the rhythm disturbance control operation is performed, including setting the rhythm change rules within multiple continuous pulse output cycles, clearing the accumulated energy inside the filter by extending the pulse interval and inserting blank periods, and blocking the continuous formation of self-excited background noise.

[0051] Preferably, a current pulse generator noise reduction device includes a pulse characteristic analysis module, a residual energy trend modeling module, a dynamic evolution parameter extraction module, a state factor generation module, and a rhythm intervention control module;

[0052] The pulse characteristic analysis module collects the pulse width, pulse interval and edge slope parameters of each pulse during the continuous narrow pulse output of the current pulse generator, obtains the output voltage of the filter in the corresponding cycle, and forms a filter response sequence of multiple cycles;

[0053] The residual energy trend modeling module removes the main pulse waveform from the filter response sequence, extracts the filter residual response segments of multiple cycles, performs time alignment and overlap processing on the residual response segments of multiple cycles, and constructs the residual energy superposition trend line;

[0054] The dynamic evolution parameter extraction module calculates the trend offset rate parameter based on the residual energy superposition trend line using the Savitzky-Golay sliding window polynomial fitting derivative algorithm. It also generates the disturbance frequency domain diffusion parameter using wavelet packet decomposition and Shannon entropy calculation algorithms to reflect whether the residual energy superposition behavior of the filter evolves into self-excited background noise.

[0055] The state factor generation module maps the trend offset rate parameter and the disturbance frequency domain diffusion parameter into a two-dimensional trajectory path, extracts the trajectory deformation characteristics using the similar cone path mapping method, and generates a state factor that characterizes the stage of the self-excited background noise formation process;

[0056] The rhythm intervention control module makes stage judgment based on the generated state factor and executes intervention measures according to the judgment result, including adjusting the pulse output rhythm, inserting gap interference pulses or applying rhythm disturbance control to block the continuous superposition of residual energy of the filter.

[0057] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0058] 1. The current pulse generator noise reduction method provided by this invention first acquires pulse characteristic parameters and filter response data in real time during the continuous narrow pulse output process and removes the main pulse waveform to obtain a sequence of filter residual response behaviors spanning multiple cycles. This structured data processing method effectively overcomes the limitations of traditional methods that rely on amplitude spikes or fixed thresholds for identification, allowing the previously imperceptible background energy accumulation behavior to be accurately characterized and extracted. It establishes a dynamic observation foundation based on the evolution of residual response behavior, providing sufficient support for subsequent trend modeling and behavioral evolution identification.

[0059] 2. By incorporating an analytical mechanism combining Savitzky-Golay derivative fitting with wavelet packet decomposition, this approach can precisely characterize the rate of change and diffusion characteristics of residual energy in both the time and frequency domains. These two parameters are further mapped into two-dimensional trajectory paths. Using a conical path mapping method to extract trajectory deformation characteristics, this approach achieves process modeling and state partitioning of the residual energy's evolution from accumulation to self-excited interference. This processing path, independent of traditional frequency-domain filtering or fixed limiting rules, offers enhanced robustness and adaptability, making it particularly suitable for the dynamic identification and monitoring of covert background interference in nonlinear systems.

[0060] 3. After identifying that the filter is in a specific evolutionary stage, the present invention can trigger differentiated intervention strategies based on the generated state factors, such as adjusting the pulse rhythm, inserting gap pulses, or applying disturbance control, actively intervening in the continuous superposition chain of energy, thereby blocking the continuous formation of self-excited noise. This closed-loop control method, from identification to regulation, not only significantly improves the noise reduction capability of the current pulse system under dynamic operation, but also significantly enhances the system's anti-interference performance and output stability, ensuring the accuracy and security of signal transmission and event triggering in high-frequency and high-speed circuit environments. It has outstanding technical advantages and engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction to the drawings required for use in the embodiments will be given below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0062] Figure 1 The present invention is a flow chart of a current pulse generator noise reduction method and device.

[0063] Figure 2 This is a module schematic diagram of a current pulse generator noise reduction method and device of the present invention. DETAILED DESCRIPTION

[0064] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art.

[0065] The present invention provides Figure 1 A current pulse generator noise reduction method shown includes the following steps:

[0066] S01. During the continuous narrow pulse output process of the current pulse generator, the pulse width, pulse interval and edge slope parameters of each pulse are collected to obtain the output voltage of the filter in the corresponding cycle, thereby forming a filter response sequence of multiple cycles;

[0067] In this embodiment, in S01, the following steps are specifically included:

[0068] During the continuous narrow pulse output process of the current pulse generator, the start time and end time of each pulse are collected based on the trigger clock, the pulse width is calculated by the start and end time, the pulse interval is calculated by the start time difference of adjacent pulses, and the edge slope is calculated by the start edge fitting curve;

[0069] During the continuous narrow pulse output of a current pulse generator, the start and end times of each pulse are acquired based on a trigger clock. A high-precision time-to-digital converter (TDC) can be used to timestamp pulse edges. With picosecond resolution, the TDC is suitable for nanosecond pulse measurement. Pulse width is calculated by recording the rising edge time t_rise and falling edge time t_fall of the same pulse, and subtracting the difference (t_fall - t_rise) from the rising edge time. For example, if a pulse's rising edge occurs at 10.000ns and its falling edge occurs at 10.900ns, the pulse width is 0.900ns. The pulse interval is calculated by the difference between the rising edge times of two adjacent pulses. For example, if the first pulse starts at 10.000ns and the second pulse starts at 11.200ns, the interval is 1.200ns. The edge slope parameter can be digitally fitted using oscilloscope waveform data. For example, a least-squares linear fit method is used to establish a voltage-time fitting model for 3 to 5 sampling points before and after the rising edge. The slope is the slope of the fitted line. For example, if the voltage rises from 0 V to 3 V between 0 ns and 0.3 ns, the slope is 3 / 0.3 = 10 V / ns. This method uses an FPGA equipped with a time-determined value (TDC) and a high-bandwidth ADC to achieve edge capture and waveform readback, providing key parameters for subsequent identification of the filter's residual response.

[0070] Taking the start time of each pulse as the reference point, the voltage signal waveform at the output of the filter is obtained within a fixed sampling window, including the main pulse and the response segment after the main pulse;

[0071] This can be achieved using a high-speed analog-to-digital converter (ADC) in conjunction with external trigger control logic. Specifically, an external pulse trigger source (such as a trigger marker signal output by an FPGA) is set in the system. Whenever a pulse starts, the ADC starts acquiring data. A fixed-length sampling window is defined, for example, to acquire waveform data of 1.5 ns each time. The sampling window's starting point is aligned with the pulse start time, and the sampling interval is determined by the ADC sampling rate. For example, using a 10 GSa / s (10 billion samples per second) ADC, a data point is acquired every 0.1 ns, for a total of 15 points. The sampling window should be large enough to cover the entire main pulse and the filter's natural response following the main pulse. The main pulse typically lasts 0.9 ns, while the filter's fallback may last 0.4 to 0.6 ns. After data acquisition, this waveform segment is temporarily stored in a buffer and packaged with the pulse timestamp to form the complete voltage response of the pulse event. This approach accurately tracks and captures the "main pulse + tail response" at the filter output, facilitating subsequent isolation and modeling of residual disturbances. To improve real-time performance, a ring buffer structure can be used in conjunction with DMA to quickly transmit the collected waveforms to the back-end processing unit for further analysis.

[0072] Repeat the parameter acquisition and voltage acquisition operations for multiple pulse cycles, and record the data of each cycle according to the timestamp number;

[0073] This can be achieved through FPGA control logic coupled with a cyclic sampling scheduling mechanism. Specifically, while the pulse generator continuously outputs narrow pulses, the FPGA receives the start trigger signal for each pulse in real time. Each trigger initiates a data acquisition process, which includes reading the start and end times of the current pulse from the time-determined clock (TDC), calculating parameters such as pulse width, interval, and edge slope, and controlling a high-speed ADC to read the filter output voltage waveform according to a preset sampling window. After each cycle is acquired, the pulse parameters and corresponding voltage waveform data are combined into a data packet, timestamped with the pulse trigger time. For example, the data is recorded as "Cycle #N, Start Time Tn, Parameter Set Pn, Voltage Sequence Wn." To ensure data sequentiality and real-time performance, a FIFO buffer can be set up within the FPGA to write data packets in trigger order, or a cyclic write structure can be used to automatically increment the data packets. The batches can then be written to a high-speed cache (such as DDR) via a DMA interface for subsequent processing. This approach enables stable acquisition of pulse data for hundreds or even thousands of consecutive cycles, providing high-resolution dynamic data support for constructing residual response trend lines and identifying background noise evolution.

[0074] The parameters and response voltage data of multiple cycles are arranged in a preset cycle order to form a filter response behavior sequence on a continuous time axis for subsequent residual energy analysis.

[0075] This can be achieved through structured data concatenation and index management. Specifically, after completing parameter acquisition and voltage waveform acquisition for each pulse cycle, the system constructs the data for that cycle into data units consisting of a timestamp, a pulse parameter vector, and a voltage sample sequence. These units are then written sequentially into an ordered buffer structure, such as a two-dimensional array, a circular buffer queue, or a chain structure, in chronological order of trigger time. During each write, the data units are automatically sorted by timestamp (or stored sequentially if the acquisition trigger sequence is consistent), ensuring strict logical temporal continuity of the data. Furthermore, the system can set a unified reference clock and remap the voltage samples in each data unit onto a global time axis. For example, the starting position of the voltage sample for the Nth cycle can be calibrated to T0 + N × ΔT, where ΔT is the pulse period. This constructs a continuous filter response behavior matrix across cycles and samples. This sequence not only reflects the filter's transient response to the pulse excitation within each cycle but also fully preserves the continuity and superposition of responses between cycles, providing a structured and time-consistent data foundation for subsequent extraction of residual energy trends. This processing can be automatically completed through the data scheduling logic preset in the embedded processor or FPGA, and is suitable for high-density, high-timeliness pulse data sequence management.

[0076] During the continuous narrow pulse output of a current pulse generator, the pulse width, pulse interval, and edge slope parameters of each pulse are collected, and the filter output voltage within the corresponding cycle is obtained. This multi-cycle filter response sequence is formed to analyze the dynamic evolution of the filter response under high-speed excitation conditions. Due to the high-density and high-steepness switching characteristics of continuous narrow pulses, the filter may not fully recover to a steady state after each pulse excitation, resulting in energy accumulation between adjacent cycles and potential residual perturbations. By systematically collecting these parameters, it is possible to accurately restore the mapping between pulse excitation and filter response, specifically identifying whether residual energy is not fully released due to excessively fast pulses or steep edges. The pulse width and interval reflect the pulse rhythm structure, the edge slope is related to the excitation intensity, and the filter output voltage reflects the transient response and residual effects within each cycle. Organizing this information into a multi-cycle response sequence provides accurate and continuous dynamic input for subsequently extracting the filter residual energy accumulation trend and constructing a background noise evolution model, which is a prerequisite for achieving highly reliable noise reduction identification.

[0077] S02. removing the main pulse waveform from the filter response sequence, extracting filter residual response segments of multiple cycles, performing time alignment and overlapping processing on the residual response segments of multiple cycles, and constructing a residual energy superposition trend line;

[0078] In this embodiment, in S02, the following steps are specifically included:

[0079] Determine the time interval corresponding to the main pulse in each cycle, calculate the start and end boundaries of the main pulse based on the pulse start time and pulse width, and remove the filter voltage data in the time interval corresponding to the main pulse in each cycle from the original response sequence;

[0080] "Determining the time interval corresponding to the main pulse in each cycle, calculating the main pulse start and end boundaries based on the pulse start time and pulse width, and removing the filter voltage data within this time period from the original response sequence" can be achieved through pulse trigger timing control and a data window marking mechanism. Specifically, during the operation of the current pulse generator, the system records the start time T_start and pulse width W of each pulse event in real time, thereby determining the end time of the main pulse as T_end = T_start + W. After obtaining this time interval, all voltage data with sampling times within the interval [T_start, T_end] in the acquired filter output voltage sequence are marked as the main pulse segment using the timestamp index. Subsequently, data cleaning processes remove all sampling points within this time period or replace them with zero values ​​to maintain timeline integrity. A common implementation structure includes a timestamp-based voltage sequence array. Before extracting the residual response segments, the start and end time indexes are used in each cycle array to perform segment stripping. For example, if the main pulse start time of a cycle is 10ns and the pulse width is 0.9ns, the main pulse interval is [10ns, 10.9ns], and the system deletes all voltage data within this interval from the response sequence of the current cycle. The core purpose of this operation is to accurately distinguish the main excitation portion from the subsequent residual disturbance portion. This ensures that subsequent trend analysis is based only on the natural release behavior of the filter rather than the results of active excitation, improving the accuracy of identifying residual energy evolution.

[0081] In each cycle, a fixed-length residual observation window is set after the main pulse ends, and the voltage sampling points retained in the window are extracted as the residual response segment, and their relative time coordinates relative to the pulse starting point are uniformly recorded;

[0082] "In each cycle, a fixed-length residual observation window is set after the main pulse ends. The voltage sampling points retained within this window are extracted as residual response segments, and their relative time coordinates relative to the pulse start point are uniformly recorded." This can be achieved through a method based on time window truncation and unified time normalization. Specifically, after the system completes the identification and elimination of the main pulse time interval [T_start, T_end], it immediately extends the time window ΔT of a set length from T_end in each cycle. For example, if it is set to 20ns, the observation window interval is [T_end, T_end+20ns]. The system extracts all voltage data sampled within this time period from the filter voltage response sequence as residual response segments. Furthermore, to facilitate unified data overlay processing across different cycles, the sampling points in each residual response segment have their original absolute timestamps (e.g., 13.5ns) converted to relative times (e.g., 3.5ns) relative to the current pulse start time T_start, thereby achieving time normalization. This normalization operation ensures that the residual response segments can be calculated using a unified relative time base in subsequent alignment and overlap analysis, even if the actual start time of each pulse is different. For example, if the pulse in the first cycle starts at 10 ns and the main pulse ends at 10.9 ns, the extraction window is [10.9 ns, 30.9 ns], where the relative time of any sampling point t is t − 10 ns. This method accurately captures the residual response behavior of the filter during its natural release after the main excitation ends, providing critical data for identifying energy accumulation and trend evolution.

[0083] The residual response segments extracted from all cycles are aligned according to their start time and overlapped on a unified time axis to generate a cross-cycle residual response distribution set.

[0084] This can be achieved through relative time normalization, array stacking, and uniform interpolation alignment. Specifically, in the previous stage, the time axes of all residual response segments are converted to relative time coordinates relative to their respective pulse onsets. Therefore, the residual response segments of each cycle can be considered as voltage sequences based on the same time base (e.g., with the pulse onset at 0 ns). The system constructs all residual segments of each cycle into a two-dimensional matrix, where each row represents the residual response of a cycle, and the column index represents a uniform relative time point (e.g., 200 points sampled at 0.1 ns intervals between 0 ns and 20 ns). If the actual sampling points of different cycles are not perfectly aligned, linear interpolation or high-order spline interpolation can be used to resample the data within each cycle to fill in the corresponding time points, thereby achieving consistent alignment. Subsequently, by stacking the data column by column, aggregating the voltage values ​​corresponding to multiple cycles at a uniform time point, a "cross-cycle residual response distribution set" matrix is ​​formed. This matrix can not only be used for subsequent trend extraction but also reflects the volatility and stability of the response at each time point. For example, the 50th column of the matrix aggregates the residual response values ​​at the 5th ns position across all cycles. This can be used to calculate the average response trend at that point in time and also to statistically analyze its variance, supporting the identification of the dynamic evolution of energy superposition. This processing method can integrate non-stationary interference information from multiple cycles into a unified time frame and is the core foundation for building a reliable trend identification model.

[0085] The voltage data corresponding to each time point in the residual response distribution set are averaged to obtain a residual energy evolution curve covering multiple cycles, and the residual energy superposition trend line is constructed based on this curve. The specific construction process is as follows: sliding window processing is performed on the residual energy evolution curve covering multiple cycles, and the data points in each window are fitted into local linear segments to obtain the local slope distribution in the continuous time interval; the slope sequence of multiple local linear segments is smoothed and filtered to eliminate the high-frequency noise offset caused by small disturbances and generate a continuous trend-guiding signal; the trend-guiding signal is sampled at a fixed time interval, and a third-order polynomial is used to fit the data interval before and after each sampling point to form a continuous interpolation expression of the trend curve; all interpolation points are connected to form a continuous change trajectory covering the entire residual response time axis, and this trajectory is used as the superposition trend line of the filter residual energy.

[0086] The method "averages the voltage data corresponding to each time point in the residual response distribution set to obtain a residual energy evolution curve covering multiple cycles, and uses this curve to construct a residual energy overlay trend line" can be achieved through a combination of time-point averaging, sliding window local modeling, slope sequence filtering, and polynomial fitting interpolation. This step is highly feasible for data processing and capable of modeling evolutionary trends. Specifically, the voltage values ​​at uniform relative time points (e.g., 0.1ns, 0.2ns, etc.) across all cycles in the residual response distribution set are averaged to form a single one-dimensional time series representing the average evolution of residual energy over multiple cycles. Next, a sliding window is applied to this one-dimensional evolution curve, with each window length set to a certain number of sampling points (e.g., 7 points). Within each window, a local linear curve is fitted using the least squares method, and its slope is extracted as the trend rate within that interval. This results in a set of time-varying "local slope series" that reflect the accelerating or stable sections of residual energy evolution. Because these slope values ​​may introduce high-frequency fluctuations due to small perturbations or discrete sampling, the entire slope sequence is further smoothed (e.g., using moving average filtering, wavelet denoising, or exponential weighting) to remove unstructured high-frequency noise and extract a stable trend-oriented signal. Key sampling points are then selected at fixed intervals (e.g., every 1 ns) on this smoothed trend signal. A certain number of data points are taken on either side of each sampling point to construct local data segments. A third-order polynomial fitting algorithm is then used to curve fit these data to obtain the trend expression function for each key point. By continuously concatenating all fitted segments, a continuous trajectory covering the entire residual response timeline is formed. This trajectory is the superimposed trend line of the filter residual energy. This trend line not only preserves the dominant characteristics of energy evolution but also effectively suppresses background disturbances. It can truly reflect the dynamic accumulation behavior of the filter under continuous narrow pulse excitation, providing a structural criterion for subsequent background noise identification and intervention strategy selection.

[0087] The main pulse waveform is removed from the filter response sequence, and the filter residual response segments over multiple cycles are extracted. These segments are then time-aligned and overlapped to construct a residual energy superposition trend line. This approach accurately identifies potential energy accumulation within the filter under continuous narrow pulse excitation. Because the main pulse of a current pulse generator in high-frequency output mode has high amplitude and sharp edges, it induces a significant transient response in the filter. This energy release process requires a certain time window. However, if the next pulse arrives before the residual energy of the previous pulse has fully dissipated, the residual energy will periodically superimpose, evolving into a weak but persistent background disturbance. This type of disturbance does not manifest as noticeable pulse distortion and is often overlooked by traditional filtering or edge detection methods. Therefore, the main pulse excitation itself must be removed, and analysis must be focused solely on the natural release segment of the filter following the main pulse. Time alignment and overlap after multi-cycle sampling accumulates sufficient statistical samples to eliminate the influence of individual pulse jitter, thereby extracting the cross-cycle energy evolution trend. This trend line reveals whether the filter exhibits a cumulative increase in residual energy under continuous excitation, providing a key basis for determining whether it has entered a potential self-excitation state. Therefore, the purpose of setting up the entire step is to remove the main excitation interference, extract the real release behavior, and provide the input basis for subsequent intervention judgment through trend modeling.

[0088] S03. Based on the residual energy superposition trend line, the Savitzky-Golay sliding window polynomial fitting derivative algorithm is used to calculate the trend offset rate parameter, and the wavelet packet decomposition and Shannon entropy calculation algorithm are used to generate the disturbance frequency domain diffusion parameter to reflect whether the residual energy superposition behavior of the filter evolves into self-excited background noise;

[0089] In this embodiment, in S03, the process of calculating the trend shift rate parameter based on the residual energy superimposed trend line includes the following steps:

[0090] The Savitzky-Golay sliding window process is performed on the residual energy superposition trend line, and a third-order polynomial fitting is performed within each fixed-length window to generate a local smooth curve;

[0091] Savitzky-Golay sliding window processing is performed on the residual energy superimposed trend line. Specifically, adjacent data segments are sequentially intercepted along the time axis of the entire residual energy trend curve using a preset fixed-length sliding window. Within each data segment, a third-order polynomial is fitted to the data points using the Savitzky-Golay filtering algorithm, with the window center as the reference. The Savitzky-Golay sliding window is a local smoothing algorithm that preserves the data trend characteristics while eliminating high-frequency fluctuation noise by fitting a low-order polynomial within the sliding window. Unlike simple moving average filtering, it better preserves the local slope, peaks, and curvature of the original signal. In this application, the polynomial curve generated by fitting each window is used to extract the smooth variation of the trend within that interval, thereby constructing multiple local smooth segments. These segments are then spliced ​​together to form a continuous and smooth representation of the entire residual energy trend, providing the basis for subsequent calculation of derivative sequences and analysis of energy evolution rates. Furthermore, this method is highly sensitive to weak energy variations under continuous narrow pulse excitation and can effectively capture potential cumulative growth trends in the filter residual response.

[0092] At the center of each sliding window, the first-order derivative value of the curve is extracted to form the derivative sequence of the residual trend on the entire time axis;

[0093] Extracting the first-order derivative of the curve at the center of each sliding window involves calculating the rate of change at the center of the smooth curve generated by fitting within the sliding window. This operation essentially determines whether the residual energy is increasing, decreasing, or remaining stable at the current moment. In practice, numerical differentiation or difference calculations are performed to extract the data points adjacent to the center point within the window, and their voltage changes and time intervals are used to estimate the rate of change at that location. As the window slides along the entire trend line, the rates of change at all center points are extracted and recorded one by one, ultimately forming a complete "derivative sequence"—the rate trajectory of trend change. This trajectory reflects the accumulation or dissipation of the filter's residual energy under continuous pulse excitation and is a key indicator for identifying whether energy is continuously accumulating and evolving into background noise. This method can transform the "shape" of energy changes into a recognizable dynamic "trend," laying the foundation for subsequent background noise analysis.

[0094] The derivative sequence is processed by boundary extension to avoid the interference of edge window fitting distortion on the overall rate continuity;

[0095] Boundary extension of a derivative sequence involves expanding and smoothing the edges of a trend curve (the start and end points) using data compensation strategies to avoid inaccurate fitting or sudden derivative changes due to insufficient data at the edges of the sliding window. Specific implementations include mirror extension, constant value extension, or polynomial extrapolation. For example, at the curve's starting point, the first few data points can be "mirrored" forward according to their changing trend to simulate the trend at an earlier point in the timeline, thus providing sufficient data support for the sliding window. Similarly, at the end point, the simulated values ​​can be extended backward to ensure that the terminal window still has complete data input. Boundary extension improves the accuracy of the sliding fit at the start and end points of the timeline, ensuring that the derivative sequence does not experience abrupt jumps at the beginning and end due to missing fitting data. This allows the construction of a more continuous, smooth, and stable trend rate trajectory, ensuring that the entire derivative sequence has a unified physical meaning across the entire time domain, providing a reliable foundation for subsequent trend analysis and clustering.

[0096] The derivative series is smoothed by sliding average to suppress the interference of local mutations on the rate trend;

[0097] Sliding average smoothing of a derivative sequence involves calculating the local mean of the derivative values ​​at each moment in the entire trend rate trajectory using a sliding window, thereby reducing the impact of sudden jitter or individual outliers on overall trend judgment. Specifically, in the derivative sequence, with a point as the center, several adjacent derivative values ​​are taken before and after it, and the arithmetic average of these values ​​is calculated. This result is used as the new derivative value for the center point. The window is then slid sequentially, and this process is repeated, ultimately resulting in a smoother derivative curve. Sliding average smoothing is a commonly used data denoising method that effectively suppresses sharp changes caused by electrical transients, measurement errors, or fitting edge effects, making trend changes more coherent and analyzable. In the noise reduction and identification process of current pulse generators, using this smoothing method can help filter out rate fluctuations that are not physically meaningful, thereby ensuring a more stable and reliable process for determining the residual energy superposition trend.

[0098] The current derivative is extracted at each fixed sampling time point as the trend shift rate parameter to characterize the change amplitude and direction of the residual energy superposition trend.

[0099] Extracting the current derivative at each fixed sampling point as the trend shift rate parameter involves sampling the smoothed derivative sequence at pre-set intervals. The derivative value at each sampling point is then extracted to quantify the rate and direction of change in the filter residual energy superposition trend within that time period. In practice, the time axis can be divided into a number of equal sampling points, for example, sampling every 10 microseconds. The derivative value corresponding to each sampling point is recorded at that location, forming a discretized rate parameter sequence. This approach is necessary because the accumulation or release of residual energy is not continuous and uniform, but rather fluctuates dynamically with factors such as the pulse rhythm and the filter recovery state. Extracting rate information at these representative time points not only captures the microscopic evolution of the trend but also eliminates interference from invalid high-frequency noise, providing a stable and accurate basis for subsequent identification of phenomena such as persistent energy accumulation and self-excited evolution. This parameter ultimately serves as a key input for determining the evolution trend of background noise and is a crucial component of the overall noise reduction evaluation mechanism.

[0100] In this embodiment, in S03, the process of generating the disturbance frequency domain diffusion parameter based on the residual energy superposition trend line includes the following steps:

[0101] Multi-level wavelet packet decomposition is performed on the residual energy superposition trend line, and the db4 wavelet basis function is used to split the original signal into multiple sub-band signals;

[0102] Performing multi-level wavelet packet decomposition on the residual energy superimposed trend line involves recursively decomposing the trend signal at multiple scales, resulting in a refined representation of its frequency domain features at varying resolutions. The decomposition is based on the db4 wavelet basis function, a member of the Daubechies wavelet family with excellent smoothness and tight support, making it suitable for processing signals of medium complexity. The original trend curve is then subjected to a first-level wavelet packet decomposition, dividing it into low-frequency and high-frequency sub-signals of equal width. Each sub-signal is then subjected to a second and even third level of decomposition, ultimately splitting the signal into multiple sub-bands with uniform frequency coverage and progressively higher resolution. Unlike traditional wavelet transforms, wavelet packet decomposition not only refines the high-frequency signal but also performs equivalent processing on the low-frequency portion, making it more suitable for analyzing trend data with uncertain energy distribution density. This method allows the energy contribution of each sub-signal in each frequency band to be analyzed independently, comprehensively revealing the frequency domain dispersion of the filter's residual energy, laying the foundation for subsequent entropy calculation and perturbation characteristic identification. The key to this process is to achieve the conversion of time series trends into frequency structures, thereby capturing hidden interference evolution signs that are not easy to visually identify in the time domain.

[0103] Calculate the energy proportion of the corresponding sub-signal in each frequency band to form the energy distribution vector of the complete frequency band;

[0104] Calculating the energy contribution of the corresponding sub-signal in each frequency band involves calculating the energy contribution of each sub-signal within the overall trend curve, obtained after multi-level wavelet packet decomposition. This constructs a set of vectors reflecting the energy distribution characteristics in the frequency domain. Specifically, each sub-signal in each frequency band is first squared to obtain the sum of the squared amplitudes of all sampling points, which serves as the energy value for that frequency band. The energy values ​​for all frequency bands are then summed to obtain the total energy. Finally, each energy value is divided by the total energy to obtain the energy contribution. For example, if three-layer wavelet packet decomposition yields eight frequency bands, A1-A8, with corresponding energies E1-E8, and a total energy of Etotal, the energy contribution of each frequency band is E1 / Etotal, E2 / Etotal, and so on, resulting in a frequency band energy distribution vector of length 8. This vector reflects the distribution structure of the filter's residual energy across different frequency bands. If the energy contribution of certain high-frequency bands continues to increase, this may indicate energy diffusion in the frequency domain, suggesting a noise evolution trend. In this way, not only can the diffusion degree of frequency domain disturbance be quantified, but also a reliable data basis is provided for subsequent entropy calculation.

[0105] Applying Shannon entropy calculation method to the frequency band energy distribution vector, the energy distribution dispersion index is obtained;

[0106] The Shannon entropy calculation method is applied to the band energy distribution vector to quantify the "dispersion" of the filter's residual energy distribution across different frequency bands—that is, whether the energy is concentrated in a few frequency bands or widely dispersed. Specifically, each term in the band energy distribution vector is first used as a component in a probability distribution, with the energy contribution of each band being considered its "probability of occurrence." Each contribution is then substituted into the Shannon entropy formula. The core of this method is to perform the "−p × log(p)" calculation on all probability components and sum the results. Shannon entropy essentially measures the uncertainty or evenness of information distribution. When energy is concentrated in a few frequency bands, the entropy is low; conversely, when energy is evenly distributed across multiple frequency bands, the entropy is high. For example, if the energy in a decomposition is mostly concentrated in two frequency bands, the entropy is low, indicating a clear perturbation structure. If the energy is distributed nearly evenly across all frequency bands, the entropy is high, indicating that the frequency domain energy is diffuse and exhibits high uncertainty. This entropy value, as a parameter of the diffusion of disturbance in the frequency domain, directly reflects the possibility that the residual energy of the filter is transforming from regular disturbance to broad-spectrum noise. It is one of the key references for identifying self-excited background noise.

[0107] Normalize the Shannon entropy value to a standard value range to facilitate horizontal comparison between different cycles;

[0108] Normalizing the Shannon entropy value to a standard value range is to eliminate the absolute entropy value differences between different periods or different data segments due to differences in signal length, number of frequency bands, or energy distribution structure, so that the entropy values ​​between periods are comparable, which facilitates horizontal comparison and analysis of the perturbation diffusion trend. In specific implementation, the theoretical minimum and maximum values ​​of the Shannon entropy are first determined. The minimum value is usually 0 (indicating that the energy is completely concentrated in one frequency band), and the maximum value is log(N), where N is the number of frequency bands (for example, when there are 8 frequency bands, the maximum entropy value is log(8)≈2.08). Then, the entropy value H of the current period is substituted into the normalization formula: H_norm = H / log(N), so that the result is limited to the standard value range between 0 and 1. The closer the normalized entropy value is to 1, the more uniform the energy distribution is and the more broad-spectrum the perturbation is; the closer it is to 0, the more concentrated the energy is and the perturbation is structured. This normalization process not only ensures a unified evaluation scale when analyzing multi-period residual behavior, but also serves as a reliable indicator for subsequent state classification and intervention judgment, thereby improving the dynamic adaptability and algorithm stability of the overall noise reduction method.

[0109] The normalized entropy value is used as the disturbance frequency domain diffusion parameter to reflect the diffusion degree of residual energy in the frequency domain.

[0110] The normalized entropy value is used as the frequency-domain diffusion parameter for disturbances to quantitatively reflect the frequency-domain diffusion of the filter's residual energy and serve as one of the core criteria for determining the presence of self-excited background noise. Specifically, the Shannon entropy value calculated and normalized for each cycle is recorded as the frequency-domain diffusion parameter corresponding to that cycle. This constitutes a time-evolving diffusion index sequence that is used to dynamically monitor the changing trend of the frequency-domain energy distribution. This is necessary because under special pulse operating conditions, where the filter has not fully recovered and energy accumulates, self-excited interference does not appear as a distinct spike, but rather as a "background noise" with a spread spectrum. Traditional methods cannot identify this change, but this parameter reveals the evolution of energy from concentration to diffusion, thereby determining whether the system is entering a stage of potential instability or crosstalk risk. By tracking the trend of this parameter, early identification of hidden interference and intervention can be achieved, improving the intelligence and predictiveness of the noise reduction process for the current pulse generator.

[0111] In this embodiment, in S03, the trend shift rate parameter and the disturbance frequency domain diffusion parameter are used to reflect whether the filter residual energy superposition behavior evolves into self-excited background noise, specifically:

[0112] On the same time axis, the trend offset rate parameter and the disturbance frequency domain diffusion parameter are combined into a two-dimensional time series vector sequence;

[0113] Combining the trend shift rate parameter and the perturbation frequency domain diffusion parameter into a two-dimensional time series vector sequence on the same time axis is intended to integrate the time-domain evolution trend and frequency-domain diffusion characteristics of the filter residual energy into a dynamic analysis framework, thereby more comprehensively identifying the formation trajectory of self-excited background noise. Specifically, a time index is first established for each fixed time sampling point (e.g., every 10 microseconds). Then, at that time point, the corresponding trend shift rate parameter value (indicating the speed and direction of residual energy change) and the perturbation frequency domain diffusion parameter value (indicating the degree of frequency-domain energy diffusion) are extracted, and these two values ​​are combined to form a two-dimensional vector (v1, v2). Over time, the two-dimensional vectors corresponding to all sampling points are sequentially arranged to form a two-dimensional time series vector sequence, i.e., each moment contains a coordinate pair representing "energy change - frequency domain diffusion". This sequence can not only be visualized as a two-dimensional trajectory plot, intuitively reflecting the parameter evolution path, but also serves as the input for subsequent trajectory morphology analysis and state classification. In this way, the system can dynamically capture the linkage changes of residual energy superposition behavior in two dimensions, make up for the limitations of single indicator judgment, and improve the recognition accuracy of complex interference evolution processes.

[0114] A sliding window cluster analysis is performed on the vector sequence to identify segments where both the trend growth rate and the frequency domain diffusion degree increase simultaneously within a continuous time interval. Such segments are marked as suspected self-excitation segments, indicating that the residual energy is continuously increasing and showing a tendency of frequency domain diffusion.

[0115] Sliding window cluster analysis is performed on this vector sequence to identify continuous segments on the dynamic time axis where both the trend excursion rate and frequency domain diffusion increase simultaneously. These segments are known as "suspicious self-excitation segments" where self-excited background noise may be evolving. Specifically, a fixed-length sliding window (e.g., containing 5 to 10 consecutive sampling points) is set. During each sliding window, a set of two-dimensional vectors within the window is extracted. A clustering algorithm (such as DBSCAN or K-means) is then applied to the vector data within the window, dividing them into several clusters based on their distribution characteristics in two-dimensional space. Clusters of vectors that are concentrated in intervals with high trend excursion rate and high diffusion values ​​are then identified, and their presence in multiple sliding windows is determined. If a vector cluster maintains this "rapid growth combined with frequency domain diffusion" pattern over a continuous period of time, the segment is labeled a "suspicious self-excitation segment." "Sliding window cluster analysis" involves repeatedly clustering the data within a sliding window on the time axis, dynamically tracking the evolution of the clustering structure. This method is more adaptable than static threshold judgment, can more accurately identify subtle signs of background noise evolution, and achieve early identification of nonlinear and self-driven interference trends.

[0116] The duration and amplitude of each suspected self-excitation segment are counted to determine whether it exceeds the preset duration threshold and energy diffusion threshold;

[0117] Duration and amplitude statistics for each suspected self-excitation segment are used to further quantitatively assess the initially flagged suspicious areas, ensuring that the identified background noise evolution behavior is sufficiently persistent and intense, thereby avoiding misidentification due to occasional disturbances. Specifically, the number of consecutive sampling points along the time axis for each suspected segment is counted, and the actual duration of the segment is calculated based on the sampling period. The maximum and average values ​​of the trend deviation rate parameter and the disturbance frequency domain diffusion parameter within the segment are then extracted to measure the energy variation amplitude and frequency domain diffusion intensity within the segment. These statistical results are then compared with pre-set thresholds. If the duration exceeds a minimum duration threshold (e.g., 50 μs) and the average diffusion parameter or trend rate parameter exceeds a set energy diffusion threshold (e.g., a normalized value of 0.7), the segment is classified as a "significant self-excitation segment." This approach effectively eliminates short-term, low-amplitude disturbances, improves the accuracy of identifying real, evolving background noise, and provides more reliable data for subsequent classification, response, and intervention strategies.

[0118] If, in any suspected self-excited segment, the length of continuous sampling time is not less than the set minimum length threshold, and the average value of the trend offset rate parameter in the segment is greater than the trend rate threshold, and the average value of the disturbance frequency domain diffusion parameter is greater than the diffusion threshold, then it is determined that the filter is in the process of self-excited background noise evolution, and subsequent classification and intervention operations are triggered.

[0119] To accurately determine whether a filter is experiencing self-excited background noise, quantitative analysis is performed on each "suspected self-excited segment" and a decision logic is established using three key thresholds. First, the length of continuous sampling time within the segment—the time span from the start to the end of the segment—is calculated and compared with a set minimum length threshold. This threshold is typically determined based on the average recovery time of the filter's residual energy under natural dissipation conditions, aiming to eliminate false positives caused by short, sporadic disturbances. Next, the average value of the trend excursion rate parameter within the segment is calculated and compared with the trend rate threshold. This threshold reflects the significance of residual energy growth and is typically determined by measuring the normal fluctuation range of the trend rate through extensive experimental measurements, with a safety margin added. Finally, the average value of the disturbance frequency domain spread parameter within the same segment is calculated and compared with the spread threshold. This threshold indicates whether the spectral energy has begun to expand into higher-order frequency bands, indicating the propagation of the interference. Only when all three criteria are met simultaneously—namely, a sufficiently long period, a continuously increasing energy trend, and an intensified frequency disturbance—can the filter be considered to have failed to complete natural decay, entering an excitation path of energy superposition and potentially evolving into a stable background interference state. This judgment method not only effectively eliminates atypical noise but also proactively identifies interference with potential evolutionary trends, ensuring accurate triggering for subsequent classification and intervention operations, thereby enhancing the robustness and practicality of the entire noise reduction solution.

[0120] When a current pulse generator continuously outputs narrow pulses, if the pulse period is shorter than the natural energy dissipation time of the filter, the filter will be frequently excited by incompletely decayed energy. This causes residual energy to accumulate over multiple cycles, resulting in imperceptible self-excited background noise. Because this type of background noise does not manifest as obvious spikes or distortion, its evolution is subtle and gradual, making it difficult to effectively identify with traditional noise reduction methods. Therefore, in-depth analysis based on the residual energy superposition trend line is necessary. Firstly, a trend shift rate parameter is extracted using a Savitzky-Golay sliding window polynomial fitting derivative algorithm to capture the rate and direction of change in the energy accumulation trend. Secondly, a disturbance frequency domain diffusion parameter is generated using wavelet packet decomposition combined with Shannon entropy calculation to characterize whether the spectral energy has spread to higher frequency bands and thus characterize the propagation trend of the interference. These two parameters, based on the time variation and frequency domain perturbation dimensions, quantify whether the filter's residual energy superposition behavior has evolved into persistent and diffuse background interference. They are key to achieving high-precision identification and triggering intervention, and are the core innovation of this noise reduction method that distinguishes it from traditional techniques.

[0121] S04, mapping the trend offset rate parameter and the disturbance frequency domain diffusion parameter into a two-dimensional trajectory path, extracting trajectory deformation features using a similar cone path mapping method, and generating a state factor representing the stage of the self-excited background noise formation process;

[0122] In this embodiment, in S04, the following steps are specifically included:

[0123] According to the time sequence, the trend deviation rate parameter corresponding to each moment is used as the horizontal coordinate, and the disturbance frequency domain diffusion parameter is used as the vertical coordinate. The two-dimensional trajectory path is constructed in sequence to form a set of trajectory segments reflecting the parameter change process.

[0124] This step constructs a two-dimensional trajectory path reflecting the temporal evolution of the filter's residual energy stacking behavior by pairing the trend offset rate parameter with the perturbation frequency domain diffusion parameter by timestamp. Specifically, the trend offset rate parameter and perturbation frequency domain diffusion parameter calculated within each sampling period are time-indexed, ensuring a one-to-one temporal order. Then, using the trend offset rate parameter as the horizontal axis and the perturbation frequency domain diffusion parameter as the vertical axis, these coordinate points are connected in chronological order on a two-dimensional plane to form a temporally ordered trajectory. This trajectory essentially reflects the coupling relationship between the filter's residual energy stacking trend (speed) and the frequency domain perturbation (complexity) during the dynamic process. For example, when residual energy accumulates rapidly and is accompanied by increased frequency domain diffusion, the trajectory will shift to the upper right. Conversely, if energy changes slowly or diffusion weakens, the trajectory may converge to the lower left. By dividing these continuous trajectories into segments, a trajectory segment set is formed, providing the basis for subsequent geometric feature extraction and stage state analysis. This path construction process not only retains the joint evolution information of key parameters, but also has strong visualization and pattern recognition potential.

[0125] In the two-dimensional trajectory path, the geometric features of each trajectory segment are extracted using a sliding window of fixed length, including local curvature, path bending rate, angular rotation trend and change rate;

[0126] In this step, a fixed-length sliding window is set on the two-dimensional trajectory path to extract local geometric features segment by segment, which are used to analyze detailed trends in parameter evolution. Specifically, the sliding window length is first set (for example, to contain 5 to 7 consecutive trajectory points). The window is then slid across the trajectory path in chronological order, extracting the set of trajectory points within the current window each time. For each window, the following features are extracted using geometric calculation methods: local curvature reflects the degree of trajectory curvature within the interval and can be estimated by fitting circular arcs or second-order derivatives; path curvature measures the frequency of trajectory reversals or turns, typically calculated as the standard deviation of path direction changes or the maximum angular change; angular rotation trend assesses the consistency or torsional tendency of the trajectory, for example, by using the angle between the line connecting the first and last points and the midpoint vector; and rate of change quantifies the ratio of trajectory distance traveled within the window to time, reflecting the degree of acceleration or deceleration of the overall parameter change. Once these features are extracted, they can be used as feature descriptors for the trajectory segment for subsequent similarity analysis and state recognition. This method can accurately characterize the geometric characteristics of the evolution trajectory of self-excited background noise without destroying the time sequence, providing a structured basis for subsequent state judgment.

[0127] The extracted geometric features are projected into the conical path mapping space, and the contour morphology of each trajectory segment in the space is analyzed and clustered to divide it into multiple continuous evolution state regions;

[0128] In this step, the geometric features of each trajectory segment extracted in the previous stage (such as local curvature, path curvature, angular rotation trend, and rate of change) are uniformly mapped into a high-dimensional geometric space called the conical path mapping space, enhancing the distinguishability of the morphologies of different trajectory segments. This is achieved as follows: First, the feature vectors of each trajectory segment are normalized to ensure that features of different dimensions have a relatively balanced influence in the projection. Next, these normalized vectors are embedded in a conical mapping space constructed based on the principle of "continuous morphological evolution." In this space, trajectory segments close to the cone axis represent "stable evolution," while those away from the axis represent "dramatic change." In this space, the projection of each trajectory segment exhibits a specific contour morphology, such as spirals, return paths, bends, or diffusion paths. Subsequently, a clustering algorithm (such as density-based DBSCAN or continuous distribution K-means) is applied to the projection results of all trajectory segments based on similarity, grouping segments with similar geometric morphological trends into the same cluster. This ultimately forms several continuous evolution state regions, each representing a potential stage of energy superposition evolution. In this way, not only can a clear state flow structure be extracted from complex parameter changes, but potential mutation points or pattern turning points can also be identified, providing a structured basis for the subsequent generation of state factors.

[0129] According to the evolution order, deformation amplitude and distribution characteristics of each state area in the trajectory, the state factor characterizing the stage of the self-excited background noise formation process is generated and serves as the basis for subsequent classification and intervention.

[0130] In this step, a state factor characterizing the stage of the self-excited background noise formation process is generated by analyzing the evolution sequence, deformation amplitude, and distribution characteristics of each state region along the two-dimensional trajectory. The specific implementation method is as follows: First, the clustered state regions are arranged in chronological order along the trajectory to construct a state sequence reflecting the energy evolution process. Next, the duration (duration) and deformation amplitude indicators (such as average curvature and maximum curvature) of each state region in the trajectory are calculated to quantify the stability and intensity of the characteristics of each stage. These quantitative features are then normalized and combined to form a multidimensional state description vector. This description vector can be compressed into a low-dimensional state factor using dimensionality reduction techniques such as principal component analysis (PCA) or t-SNE, which serves as a compact representation of the overall behavior stage. Finally, this state factor is mapped to multiple defined noise formation stages (such as "untriggered," "disturbance accumulation," "critical turning point," and "self-excited noise generation") to determine the current filter residual energy evolution state. This state factor can not only be used to determine subsequent classification labels, but also serve as a basis for triggering intervention measures to achieve active regulation of the current pulse background noise formation process.

[0131] The purpose of performing this step is to abstract the dynamic evolution of the filter's residual energy under continuous narrow pulse excitation from the original numerical parameter level into interpretable and classifiable stage states. Since the trend offset rate parameter reflects the change in the speed of energy accumulation and the disturbance frequency domain diffusion parameter reflects the breadth of energy distribution in the frequency domain, mapping these two parameters into a two-dimensional trajectory path can intuitively capture their co-evolutionary relationship in the time dimension. The conical path mapping method is used to further extract the geometric deformation characteristics of the trajectory and divide it into multiple continuous state regions. This can effectively reveal the key turning points and process stages of the filter's residual energy evolution from the normal recovery stage to the self-excited noise stage. The final generated state factor not only has the ability to distinguish the stages of complex dynamic behavior, but also provides a clear target classification basis for subsequent intervention measures. It is the core bridge to break the bottleneck of traditional noise reduction "fuzzy recognition" and achieve precise noise reduction control.

[0132] S05. Performing stage determination based on the generated state factor and executing intervention measures according to the determination result, including adjusting the pulse output rhythm, inserting gap interference pulses, or applying rhythm disturbance control to block the continuous superposition of the residual energy of the filter.

[0133] In this embodiment, S05 specifically includes:

[0134] Based on the distribution position of the generated state factors in the multidimensional state space, the stage of the current filter residual energy superposition behavior is determined, and the state factors are divided into the initial fluctuation stage, the evolution critical stage and the continuous self-excitation stage;

[0135] The following method can be used to determine the stage of the filter's residual energy stacking behavior based on the distribution of state factors in a multidimensional state space: First, the state factors are constructed from the deformation characteristics of the trajectory path, typically representing the dynamic changes of each trajectory segment using multiple numerical dimensions, such as trajectory curvature, rotation amplitude, and rate of change. To facilitate subsequent stage determination, all state factors are projected into a predefined multidimensional state space, with each dimension corresponding to a standardized representation of a characteristic quantity. Within this space, three stage determination regions are predefined: the initial fluctuation stage region covers state points with low characteristic amplitudes and slow trajectory changes; the critical evolution stage region covers state points with medium characteristic amplitudes and significant trajectory bending but no continuous rate of change; and the continuous self-excitation stage region is concentrated in state clusters with high characteristic amplitudes, severe trajectory twisting, and continuous deflection. By matching the current state factor's position in this multidimensional space and determining its nearest determination region, the current stage of the filter's residual energy stacking evolution can be identified. This method can be implemented using a support vector machine (SVM) classifier, K-nearest neighbor (KNN) discrimination, or Euclidean distance-based nearest cluster identification, effectively improving the state recognition resolution and anti-interference capabilities. The stage determination achieved in this way provides a foundation for the subsequent development of differentiated intervention strategies.

[0136] When the state factor is in the initial fluctuation stage, a fixed-amplitude time interval disturbance is introduced into each continuous pulse cycle to change the rhythm arrangement of the continuous pulse output and prevent the filter from accumulating energy under constant excitation.

[0137] When the state factor is in the critical stage of evolution, a gap interference pulse with a set interval is inserted into the continuous pulse output sequence to change the pulse trigger timing and interrupt the residual energy superposition link in the filter;

[0138] When the state factor is in the continuous self-excitation stage, the rhythm disturbance control operation is performed, including setting the rhythm change rules within multiple continuous pulse output cycles, clearing the accumulated energy inside the filter by extending the pulse interval and inserting blank periods, and blocking the continuous formation of self-excited background noise.

[0139] The following is a detailed explanation of how and why interventions are implemented for these three status factors:

[0140] When the state factor is in the initial fluctuation stage: rhythm adjustment can be achieved by dynamically perturbing the pulse period. Specifically, while maintaining the stability of the overall pulse sequence, a small time perturbation of the ±ΔT level is introduced to each continuous pulse period. The perturbation amplitude can be generated using a pseudo-random sequence, such as based on the linear congruential method or M-sequence generator, so that the pulse interval is slightly offset within a reasonable range. This perturbation method can effectively break the resonant accumulation path caused by continuous equal-periodic pulses on the filter and prevent the energy from forming synchronous excitations within the filter. This measure is taken because a stable residual superposition chain is usually not formed in the initial fluctuation stage, and a periodic response only appears under perturbation conditions. Therefore, the development of the problem can be effectively suppressed by slightly disturbing the rhythm.

[0141] When the state factor reaches a critical stage of evolution, a gap interference pulse insertion strategy can be employed. This is achieved by intentionally inserting a "gap" pulse every N normal output cycles in the pulse sequence. This means that no actual current is output during this cycle, with only the trigger signal or a blank pulse signal being transmitted. This approach can be implemented by setting a gap beat table within the drive controller and pre-programming the rhythm control bits. This aims to proactively disrupt the superposition path formed within the filter, similar to the interference method of "breaking the current resonant chain," preventing the further superposition and expansion of residual energy along the time axis. This stage is a critical state before self-excitation occurs, so effective intervention should be implemented through moderate rhythm reconstruction.

[0142] When the state factor is in the sustained self-excitation phase, a more robust rhythmic perturbation control should be implemented. This is achieved by setting multiple consecutive output cycles as a perturbation rhythm unit. Within this unit, the pulse interval is gradually extended (e.g., linearly or exponentially) and one or two complete blank periods are inserted in between or at the end. This ensures that the filter output is completely free of input pulses, creating a complete energy release window. Furthermore, adjustable PWM duty cycle control or a digital delay line module is used to precisely manage pulse timing, making the perturbation operation controllable and predictable. This approach simulates the behavior of an "actively draining filter," clearing accumulated internal energy and restoring its steady state. This intervention method is suitable for scenarios where self-excited background noise has clearly established and exhibits stable periodicity, and is a key measure to curb its continued evolution.

[0143] like Figure 2 A current pulse generator noise reduction device shown includes a pulse characteristic analysis module, a residual energy trend modeling module, a dynamic evolution parameter extraction module, a state factor generation module, and a rhythm intervention control module;

[0144] The pulse characteristic analysis module collects the pulse width, pulse interval and edge slope parameters of each pulse during the continuous narrow pulse output of the current pulse generator, obtains the output voltage of the filter in the corresponding cycle, and forms a filter response sequence of multiple cycles;

[0145] The residual energy trend modeling module removes the main pulse waveform from the filter response sequence, extracts the filter residual response segments of multiple cycles, performs time alignment and overlap processing on the residual response segments of multiple cycles, and constructs the residual energy superposition trend line;

[0146] The dynamic evolution parameter extraction module calculates the trend offset rate parameter based on the residual energy superposition trend line using the Savitzky-Golay sliding window polynomial fitting derivative algorithm. It also generates the disturbance frequency domain diffusion parameter using wavelet packet decomposition and Shannon entropy calculation algorithms to reflect whether the residual energy superposition behavior of the filter evolves into self-excited background noise.

[0147] The state factor generation module maps the trend offset rate parameter and the disturbance frequency domain diffusion parameter into a two-dimensional trajectory path, extracts the trajectory deformation characteristics using the similar cone path mapping method, and generates a state factor that characterizes the stage of the self-excited background noise formation process;

[0148] The rhythm intervention control module makes stage judgment based on the generated state factor and executes intervention measures according to the judgment result, including adjusting the pulse output rhythm, inserting gap interference pulses or applying rhythm disturbance control to block the continuous superposition of residual energy of the filter.

[0149] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0150] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in the embodiments of this application are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0151] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean 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 the present application.

[0152] Those skilled in the art will appreciate that the units and algorithm steps of each example 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 performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel 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.

[0153] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0154] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0155] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0156] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A current pulse generator noise reduction method, characterized in that: The specific steps include: S01. During the continuous narrow pulse output process of the current pulse generator, the pulse width, pulse interval and edge slope parameters of each pulse are collected to obtain the output voltage of the filter in the corresponding cycle, thereby forming a filter response sequence of multiple cycles; S02. removing the main pulse waveform from the filter response sequence, extracting filter residual response segments of multiple cycles, performing time alignment and overlapping processing on the residual response segments of multiple cycles, and constructing a residual energy superposition trend line; S03. Based on the residual energy superposition trend line, the Savitzky-Golay sliding window polynomial fitting derivative algorithm is used to calculate the trend offset rate parameter, and the wavelet packet decomposition and Shannon entropy calculation algorithm are used to generate the disturbance frequency domain diffusion parameter to reflect whether the residual energy superposition behavior of the filter evolves into self-excited background noise; S04, mapping the trend offset rate parameter and the disturbance frequency domain diffusion parameter into a two-dimensional trajectory path, extracting trajectory deformation features using a similar cone path mapping method, and generating a state factor representing the stage of the self-excited background noise formation process; S05. Performing stage determination based on the generated state factor and executing intervention measures according to the determination result, including adjusting the pulse output rhythm, inserting gap interference pulses, or applying rhythm disturbance control to block the continuous superposition of the residual energy of the filter.

2. A current pulse generator noise reduction method according to claim 1, characterized in that: In S01, the following steps are specifically included: During the continuous narrow pulse output process of the current pulse generator, the start time and end time of each pulse are collected based on the trigger clock, the pulse width is calculated by the start and end time, the pulse interval is calculated by the start time difference of adjacent pulses, and the edge slope is calculated by the start edge fitting curve; Taking the start time of each pulse as the reference point, the voltage signal waveform at the output of the filter is obtained within a fixed sampling window, including the main pulse and the response segment after the main pulse; Repeat the parameter acquisition and voltage acquisition operations for multiple pulse cycles, and record the data of each cycle according to the timestamp number; The parameters and response voltage data of multiple cycles are arranged in a preset cycle order to form a filter response behavior sequence on a continuous time axis for subsequent residual energy analysis.

3. The current pulse generator noise reduction method according to claim 1, characterized in that: In S02, the following steps are specifically included: Determine the time interval corresponding to the main pulse in each cycle, calculate the start and end boundaries of the main pulse based on the pulse start time and pulse width, and remove the filter voltage data in the time interval corresponding to the main pulse in each cycle from the original response sequence; In each cycle, a fixed-length residual observation window is set after the main pulse ends, and the voltage sampling points retained in the window are extracted as the residual response segment, and their relative time coordinates relative to the pulse starting point are uniformly recorded; The residual response segments extracted from all cycles are aligned according to their start time and overlapped on a unified time axis to generate a cross-cycle residual response distribution set. The voltage data corresponding to each time point in the residual response distribution set are averaged to obtain a residual energy evolution curve covering multiple cycles, and the residual energy superposition trend line is constructed based on this curve. The specific construction process is as follows: sliding window processing is performed on the residual energy evolution curve covering multiple cycles, and the data points in each window are fitted into local linear segments to obtain the local slope distribution in the continuous time interval; the slope sequence of multiple local linear segments is smoothed and filtered to eliminate the high-frequency noise offset caused by small disturbances and generate a continuous trend-guiding signal; the trend-guiding signal is sampled at a fixed time interval, and a third-order polynomial is used to fit the data interval before and after each sampling point to form a continuous interpolation expression of the trend curve; all interpolation points are connected to form a continuous change trajectory covering the entire residual response time axis, and this trajectory is used as the superposition trend line of the filter residual energy.

4. The current pulse generator noise reduction method according to claim 1, characterized in that: In S03, the process of calculating the trend shift rate parameter based on the residual energy superimposed trend line includes the following steps: The Savitzky-Golay sliding window process is performed on the residual energy superposition trend line, and a third-order polynomial fitting is performed within each fixed-length window to generate a local smooth curve; At the center of each sliding window, the first-order derivative value of the curve is extracted to form the derivative sequence of the residual trend on the entire time axis; The derivative sequence is processed by boundary extension to avoid the interference of edge window fitting distortion on the overall rate continuity; The derivative series is smoothed by sliding average to suppress the interference of local mutations on the rate trend; The current derivative is extracted at each fixed sampling time point as the trend shift rate parameter to characterize the change amplitude and direction of the residual energy superposition trend.

5. The current pulse generator noise reduction method according to claim 1, characterized in that: In S03, the process of generating the disturbance frequency domain diffusion parameter based on the residual energy superposition trend line includes the following steps: Multi-level wavelet packet decomposition is performed on the residual energy superposition trend line, and the db4 wavelet basis function is used to split the original signal into multiple sub-band signals; Calculate the energy proportion of the corresponding sub-signal in each frequency band to form the energy distribution vector of the complete frequency band; Applying Shannon entropy calculation method to the frequency band energy distribution vector, the energy distribution dispersion index is obtained; Normalize the Shannon entropy value to a standard value range to facilitate horizontal comparison between different cycles; The normalized entropy value is used as the disturbance frequency domain diffusion parameter to reflect the diffusion degree of residual energy in the frequency domain.

6. The current pulse generator noise reduction method according to claim 1, characterized in that: In S03, the trend offset rate parameter and the disturbance frequency domain diffusion parameter are used to reflect whether the residual energy superposition behavior of the filter evolves into self-excited background noise, specifically: On the same time axis, the trend offset rate parameter and the disturbance frequency domain diffusion parameter are combined into a two-dimensional time series vector sequence; A sliding window cluster analysis is performed on the vector sequence to identify segments where both the trend growth rate and the frequency domain diffusion degree increase simultaneously within a continuous time interval. These segments are marked as suspected self-excitation segments, indicating that the residual energy is continuously increasing and showing a frequency domain diffusion tendency. The duration and amplitude of each suspected self-excitation segment are counted to determine whether it exceeds the preset duration threshold and energy diffusion threshold; If, in any suspected self-excited segment, the length of continuous sampling time is not less than the set minimum length threshold, and the average value of the trend offset rate parameter in the segment is greater than the trend rate threshold, and the average value of the disturbance frequency domain diffusion parameter is greater than the diffusion threshold, then it is determined that the filter is in the process of self-excited background noise evolution, and subsequent classification and intervention operations are triggered.

7. The current pulse generator noise reduction method according to claim 1, characterized in that: In S04, the following steps are specifically included: According to the time sequence, the trend deviation rate parameter corresponding to each moment is used as the horizontal coordinate, and the disturbance frequency domain diffusion parameter is used as the vertical coordinate. The two-dimensional trajectory path is constructed in sequence to form a set of trajectory segments reflecting the parameter change process. In the two-dimensional trajectory path, the geometric features of each trajectory segment are extracted using a sliding window of fixed length, including local curvature, path bending rate, angular rotation trend and change rate; The extracted geometric features are projected into the conical path mapping space, and the contour morphology of each trajectory segment in the space is analyzed and clustered to divide it into multiple continuous evolution state regions; According to the evolution order, deformation amplitude and distribution characteristics of each state area in the trajectory, the state factor characterizing the stage of the self-excited background noise formation process is generated and serves as the basis for subsequent classification and intervention.

8. The current pulse generator noise reduction method according to claim 1, characterized in that: In S05, specifically including: Based on the distribution position of the generated state factors in the multidimensional state space, the stage of the current filter residual energy superposition behavior is determined, and the state factors are divided into the initial fluctuation stage, the evolution critical stage and the continuous self-excitation stage; When the state factor is in the initial fluctuation stage, a fixed-amplitude time interval disturbance is introduced into each continuous pulse cycle to change the rhythm arrangement of the continuous pulse output and prevent the filter from accumulating energy under constant excitation. When the state factor is in the critical stage of evolution, a gap interference pulse with a set interval is inserted into the continuous pulse output sequence to change the pulse trigger timing and interrupt the residual energy superposition link in the filter; When the state factor is in the continuous self-excitation stage, the rhythm disturbance control operation is performed, including setting the rhythm change rules within multiple continuous pulse output cycles, clearing the accumulated energy inside the filter by extending the pulse interval and inserting blank periods, and blocking the continuous formation of self-excited background noise.

9. A current pulse generator noise reduction device, used to implement a current pulse generator noise reduction method according to any one of claims 1 to 8, characterized in that: It includes pulse characteristic analysis module, residual energy trend modeling module, dynamic evolution parameter extraction module, state factor generation module and rhythm intervention control module; The pulse characteristic analysis module collects the pulse width, pulse interval and edge slope parameters of each pulse during the continuous narrow pulse output of the current pulse generator, obtains the output voltage of the filter in the corresponding cycle, and forms a filter response sequence of multiple cycles; The residual energy trend modeling module removes the main pulse waveform from the filter response sequence, extracts the filter residual response segments of multiple cycles, performs time alignment and overlap processing on the residual response segments of multiple cycles, and constructs the residual energy superposition trend line; The dynamic evolution parameter extraction module calculates the trend offset rate parameter based on the residual energy superposition trend line using the Savitzky-Golay sliding window polynomial fitting derivative algorithm. It also generates the disturbance frequency domain diffusion parameter using wavelet packet decomposition and Shannon entropy calculation algorithms to reflect whether the residual energy superposition behavior of the filter evolves into self-excited background noise. The state factor generation module maps the trend offset rate parameter and the disturbance frequency domain diffusion parameter into a two-dimensional trajectory path, extracts the trajectory deformation characteristics using the similar cone path mapping method, and generates a state factor that characterizes the stage of the self-excited background noise formation process; The rhythm intervention control module makes stage judgment based on the generated state factor and executes intervention measures according to the judgment result, including adjusting the pulse output rhythm, inserting gap interference pulses or applying rhythm disturbance control to block the continuous superposition of residual energy of the filter.

Images