Adaptive filtering and intelligent separation method for multi-mode partial discharge signals

Abstract

The invention discloses a self-adaptive filtering and intelligent separation method for a multi-mode partial discharge signal, and relates to the technical field of insulation state monitoring and signal processing of power equipment, and the method comprises the following steps: S1, collecting the partial discharge signal of the power equipment in real time, and synchronously obtaining an equipment operation parameter and an environment noise signal; according to the adaptive filtering and intelligent separation method for the multi-modal partial discharge signal, the problems of modal aliasing and feature distortion caused by signal and noise time-frequency coupling in a dynamic noise environment in a traditional method are effectively solved through a cooperative mechanism of dynamic coupling degree modeling and dual-channel adversarial decoupling. The dynamic coupling path sensing module quantifies interaction characteristics of noise and signals in real time, and realizes accurate suppression of high-frequency transient interference and low-frequency periodic noise in combination with a double-channel architecture of complex field phase sensitive filtering and vibration trajectory matched filtering.

Description

Technical Field

[0001] The present invention relates to the technical field of insulation status monitoring and signal processing of power equipment, and in particular to a method for adaptive filtering and intelligent separation of multi-modal partial discharge signals. Background Art

[0002] Partial discharge detection is a core method for assessing the insulation condition of power equipment. However, in actual operating conditions, the signals collected by sensors are often mixed with multiple discharge types and complex noise. For example, different modal signals such as corona discharge and internal discharge dynamically overlap with high-frequency electromagnetic interference and low-frequency mechanical vibration noise. Traditional adaptive filtering methods typically assume static noise and separate signal from noise through fixed frequency band segmentation or preset thresholds. However, in dynamic noise environments, the time-frequency coupling characteristics of signal and noise can shift nonlinearly with factors such as equipment load fluctuations and changes in ambient temperature and humidity. This dynamic coupling makes it difficult for traditional algorithms to accurately distinguish the cross-bands between signal and noise. For example, high-frequency transient interference may intrude into the main frequency range of the discharge pulse, while low-frequency mechanical vibration noise may overlap with the periodic discharge signal in the time domain. Even after conventional filtering, residual cross-interference components remain between the two. This residual noise further undermines the independence of the multimodal signals, causing the time-frequency characteristics of different discharge types to alias during the intelligent separation process. This can cause pattern recognition-based classification models to misjudge the discharge type or miss weak signals. Although existing improvement schemes attempt to improve separation accuracy by optimizing thresholds or introducing deep learning, they have not essentially broken through the static noise separation framework and are unable to track the dynamic coupling relationship between noise and signals in real time, resulting in cascade error accumulation in the filtering and separation links, seriously affecting diagnostic reliability. Summary of the Invention

[0003] (1) Technical problems solved

[0004] In view of the shortcomings of the existing technology, the present invention provides an adaptive filtering and intelligent separation method for multimodal partial discharge signals, which solves the problem of adaptive decoupling and high-fidelity separation of multimodal partial discharge signals in a dynamic noise environment.

[0005] (2) Technical solution

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for adaptive filtering and intelligent separation of multimodal partial discharge signals, comprising the following steps:

[0007] Step S1: real-time acquisition of partial discharge signals of power equipment, and simultaneous acquisition of equipment operating parameters and environmental noise signals;

[0008] Step S2: Calculate the dynamic coupling index based on the time-frequency matrix of the signal and noise, extract the time-frequency energy distribution through the sliding window, and generate a dynamic coupling strength quantization value based on the inverse correlation between the high-frequency interference energy gradient and the signal main frequency gradient;

[0009] Step S3: Based on the timing prediction results of the dynamic coupling index, the signal is divided into a high-frequency transient interference channel and a low-frequency periodic noise channel. Complex-domain phase-sensitive filtering and vibration trajectory matching filtering are used for noise suppression, respectively. The complex-domain phase-sensitive filtering maps the signal to the complex domain using a transient pulse compression encoder and dynamically adjusts the filter cutoff frequency. The vibration trajectory matching filtering constructs an adaptive notch filter bank based on prior knowledge of the equipment's mechanical vibration spectrum.

[0010] Step S4: Input the output signals of the high-frequency channel and the low-frequency channel into the adversarial modal separation network, and force the separated signals to meet the preset independent modal statistical distribution through the modal discriminator;

[0011] Step S5: The separation results are cross-validated based on physical feature compliance detection and multi-model classification confidence verification. If the pulse parameter exceeds the threshold or the classification confidence variance exceeds the standard, parameter recalibration is triggered and decoupling is re-iterated.

[0012] Preferably, the time series prediction of the dynamic coupling index is to model the time-frequency mutual information entropy gradient sequence through a bidirectional gated cyclic unit to predict the offset direction of the coupling path between the noise and the signal in the next three time windows. During the calculation of the time-frequency mutual information entropy gradient, the dynamic coupling strength is identified by comparing the rate of change of the energy distribution overlap area between the main frequency band of the signal and the noise frequency band; when the high-frequency interference energy gradient and the signal main frequency gradient show an inverse correlation, the coupling risk is determined to be increased, triggering the pre-adjustment of the cutoff frequency of the high-frequency channel filter; for low-frequency vibration noise, by analyzing the phase difference between the slope of the vibration signal envelope and the periodicity of the discharge signal, the possibility of the coupling path penetrating into the high frequency is predicted, and the notch filter parameters are optimized accordingly.

[0013] Preferably, the specific implementation of the complex domain phase sensitive filtering includes:

[0014] Nonlinear phase compensation is performed on both the real and imaginary parts of the signal, and the complex domain filter's cutoff frequency is offset and compensated in combination with the dynamic coupling degree prediction value, preserving the steepness of the pulse leading edge and suppressing the cross-coupling components in the high-frequency interference band. During the complex domain mapping stage, nonlinear phase compensation is applied to both the real and imaginary parts of the signal, with the compensation coefficient dynamically adjusted based on the phase derivative of the instantaneous frequency to ensure the time delay consistency of the high-frequency pulse waveform. When high-frequency noise is detected intruding into the main frequency band, the filter's cutoff frequency coverage is expanded based on the coupling path prediction results, while the adjustment step size is limited to prevent pulse waveform blunting. To address instantaneous frequency jumps caused by sudden load changes, the phase derivative calculation window length is shortened to improve tracking sensitivity. Phase alignment of the real and imaginary parts is checked during the signal reconstruction stage, and compensation is iteratively performed if the deviation exceeds the limit.

[0015] Preferably, the specific implementation of the vibration trajectory matched filtering includes:

[0016] The notch center frequency is dynamically generated based on the equipment vibration fundamental frequency and its harmonic components. The notch bandwidth is adaptively adjusted using the slope of the vibration signal envelope, and the time-domain amplitude of the residual component is corrected using a Kalman filter. The notch center frequency of the adaptive notch filter group is dynamically generated based on the equipment vibration fundamental frequency and its harmonic components, and the notch bandwidth is adaptively adjusted with the slope of the vibration signal envelope. When the slope exceeds a threshold, the bandwidth is expanded and the notch depth is increased; conversely, the bandwidth is compressed to focus on the primary vibration frequency. The time-domain amplitude of the vibration noise is predicted using a Kalman filter, and the state equation is updated based on the residual component of the notch at the previous moment, dynamically correcting the amplitude compensation coefficient. When a periodic discharge signal is detected to overlap with the vibration noise in the time domain, the notch filter group delay parameter is adjusted based on the phase difference to stagger the overlapping interval.

[0017] Preferably, the adversarial modality separation network is implemented in the following manner:

[0018] A modal discriminator is set at the output end of the high-frequency channel and the low-frequency channel, respectively. The discriminator presupposes the physical constraints that the pulse interval of the corona discharge obeys the Poisson distribution and the periodicity of the internal discharge obeys the Gaussian distribution.

[0019] The difference between the separation signal and the target modal distribution is calculated based on KL divergence, and the standard deviation of the pulse interval is introduced as an additional constraint. The filter parameters are then inversely optimized using an adversarial loss function. The modal discriminator presupposes the physical statistical constraints that corona discharges follow a Poisson distribution and internal discharges follow a Gaussian distribution. The difference between the separation signal and the target distribution is quantified using KL divergence, and the standard deviation of the pulse interval is introduced as an additional constraint. When the statistical characteristics deviate from the target distribution, the filter parameters are inversely optimized to restore the discharge type characteristics. To address distribution shifts caused by sudden changes in environmental noise, the coefficient of variation of the statistical parameters is monitored through a sliding window, triggering a robust optimization mode and relaxing the distribution tolerance range. For unknown modal interference, a manual labeling and model parameter update mechanism is initiated.

[0020] Preferably, the physical feature compliance detection includes:

[0021] The threshold range is set based on the quantitative relationship between the rise time of the partial discharge pulse and the size of the insulation defect. When the rise time of the separated signal pulse is less than 1ns and there is no subsequent oscillation waveform, it is judged as a distorted signal and recalibration is triggered. The physical feature compliance detection module sets a dynamic threshold based on the coordinated consistency of the pulse rise time, peak amplitude, and half-wave width. When the parameters exceed the limit, it prioritizes checking whether the filtering strength of the high-frequency and low-frequency channels is unbalanced, and triggers parameter recalibration. The multi-model classification confidence verification module analyzes the modal aliasing risk through the confidence variance of the time-domain and frequency-domain classification models. When the variance exceeds the standard, it returns to the dynamic coupling modeling stage for re-decoupling. For abnormal signals with low matching degree of historical data, manual auxiliary annotation and model iterative optimization process are initiated.

[0022] When a sudden mechanical shock or abrupt change in temperature and humidity is detected, the system switches to broadband notch mode or anti-temperature drift mode by monitoring the instantaneous peak value of the vibration signal and the environmental sensor data, limiting the parameter adjustment range to prevent misoperation. When the hardware acquisition channel is abnormal, the system automatically switches to the backup channel and resets the compensation parameters. If convergence is still not possible after multiple iterations, the model is frozen and static filtering of the historical optimal parameters is enabled, while an alarm signal is triggered to wait for human intervention.

[0023] Preferably, the multi-model classification confidence verification includes:

[0024] The separated signals are input into the pre-trained time-domain convolutional network and frequency-domain graph neural network in parallel, and the confidence variance of the classification results of each model is counted. When the variance exceeds the preset threshold, it is determined that modal aliasing exists.

[0025] Preferably, the nonlinear phase compensation is achieved by the following steps:

[0026] The phase derivative of the instantaneous frequency of the signal is extracted, and a compensation coefficient inversely proportional to the real part phase change rate is applied to the imaginary part of the complex domain signal to ensure the delay consistency of the high-frequency pulse waveform.

[0027] Preferably, the adjustment rule of the notch bandwidth is: when the slope of the vibration signal envelope exceeds the set threshold, the notch bandwidth is expanded to cover the diffusion range of the vibration noise harmonics; when the slope of the envelope is lower than the threshold, the notch bandwidth is compressed to the matching range of the focused main vibration frequency.

[0028] Preferably, the optimization process of the adversarial loss function includes:

[0029] The distribution parameters of the modal discriminator are updated synchronously in each iteration, so that the standard deviation of the pulse interval of corona discharge converges to the theoretical variance of Poisson distribution, and the periodic error of internal discharge converges to the standard deviation range of Gaussian distribution.

[0030] Through dynamic coupling quantitative modeling, the interactive characteristics of noise and signal are perceived in real time, and dual-channel noise suppression is achieved by using complex domain filtering and vibration trajectory matching. Then, the independence of discharge type characteristics is enhanced through adversarial modal separation. Finally, the reliability of the separation results is ensured through dual verification of physical laws and data models. A closed-loop feedback of "perception-decoupling-separation-verification" is formed between the modules, and the contradiction between signal fidelity and modal independence in dynamic noise environments is systematically resolved.

[0031] (3) Beneficial effects

[0032] The present invention provides a method for adaptive filtering and intelligent separation of multimodal partial discharge signals. It has the following beneficial effects:

[0033] (1) This adaptive filtering and intelligent separation method for multimodal partial discharge signals effectively solves the modal aliasing and feature distortion problems caused by the time-frequency coupling of signals and noise in dynamic noise environments, through the synergistic mechanism of dynamic coupling modeling and dual-channel adversarial decoupling. The dynamic coupling path perception module quantifies the interactive characteristics of noise and signal in real time, and combines the dual-channel architecture of complex domain phase-sensitive filtering and vibration trajectory matching filtering to achieve precise suppression of high-frequency transient interference and low-frequency periodic noise. The adversarial modal separation network, through physical statistical constraints and adversarial training mechanisms, forces the separated signals to meet the independent distribution characteristics of the discharge type, significantly improving the separation accuracy and anti-interference ability of multimodal signals.

[0034] (2) This adaptive filtering and intelligent separation method for multimodal partial discharge signals addresses the problem of insufficient diagnostic reliability caused by cascade error accumulation in traditional algorithms by introducing a dual guarantee mechanism of physical feature compliance detection and multi-model confidence verification, ensuring that the separation results comply with the dual constraints of physical laws and data models. Under complex operating conditions of power equipment, this method can adaptively track the dynamic coupling changes of noise and signals, reducing the rate of misjudgment and missed detection, providing high-confidence data support for early warning and precise positioning of insulation defects, while reducing the need for manual intervention and improving the practicality and engineering deployment efficiency of intelligent diagnostic systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0036] Figure 2 This is a control logic timing diagram of the present invention. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0038] See also Figure 1 and Figure 2 The present invention provides a technical solution: a method for adaptive filtering and intelligent separation of multimodal partial discharge signals, comprising the following steps:

[0039] Step S1: real-time acquisition of partial discharge signals of power equipment, and simultaneous acquisition of equipment operating parameters and environmental noise signals;

[0040] Step S2: Calculate the dynamic coupling index based on the time-frequency matrix of the signal and noise, extract the time-frequency energy distribution through the sliding window, and generate a dynamic coupling strength quantization value based on the inverse correlation between the high-frequency interference energy gradient and the signal main frequency gradient;

[0041] Step S3: Based on the timing prediction results of the dynamic coupling index, the signal is divided into a high-frequency transient interference channel and a low-frequency periodic noise channel. Complex-domain phase-sensitive filtering and vibration trajectory matching filtering are used for noise suppression, respectively. The complex-domain phase-sensitive filtering maps the signal to the complex domain using a transient pulse compression encoder and dynamically adjusts the filter cutoff frequency. The vibration trajectory matching filtering constructs an adaptive notch filter bank based on prior knowledge of the equipment's mechanical vibration spectrum.

[0042] Step S4: Input the output signals of the high-frequency channel and the low-frequency channel into the adversarial modal separation network, and force the separated signals to meet the preset independent modal statistical distribution through the modal discriminator;

[0043] Step S5: The separation results are cross-validated based on physical feature compliance detection and multi-model classification confidence verification. If the pulse parameter exceeds the threshold or the classification confidence variance exceeds the standard, parameter recalibration is triggered and decoupling is re-iterated.

[0044] The time series prediction of the dynamic coupling index uses a bidirectional gated cyclic unit to model the time-frequency mutual information entropy gradient sequence, and predicts the coupling path offset direction of the noise and signal in the next three time windows. It should be further explained that in the specific implementation process, in the time series prediction process of the dynamic coupling index, a bidirectional gated cyclic unit is used to model the time-frequency mutual information entropy gradient sequence, and the time-frequency energy distribution characteristics of the current and historical moments are extracted through a sliding window to capture the coupling trend of the signal and noise in the time domain and frequency domain; based on the forward and backward time series dependency analysis of the bidirectional gated cyclic unit, the coupling path offset direction of the noise and signal in the next three time windows is predicted, including two typical modes of high-frequency electromagnetic interference spreading to low frequency or low-frequency mechanical vibration noise penetrating to high frequency; according to the prediction results, when the coupling path spreads to low frequency, the complex domain filter cutoff frequency of the high-frequency transient interference channel is adjusted in advance, The high-frequency suppression range is expanded to avoid the intrusion of the main frequency band, and the notch bandwidth of the low-frequency channel is reduced to match the slowly varying characteristics of the vibration noise. When the coupling path penetrates into the high frequency, the center frequency offset of the low-frequency notch filter is dynamically increased, and the cutoff frequency adjustment step of the high-frequency filter is shortened to ensure the fidelity of the high-frequency pulse waveform. For working conditions with sudden load changes or rapid fluctuations in ambient temperature and humidity, the correlation between the equipment operating parameters and the noise signal is monitored in real time, and a short-time coupling strength correction coefficient is added to the prediction model. If the rate of change of the dynamic coupling degree index in two consecutive time windows exceeds the preset threshold, the online update of the parameters of the coupling path prediction model is triggered to avoid the failure of historical data dependence due to sudden changes in working conditions.

[0045] The specific implementation of complex domain phase-sensitive filtering includes:

[0046] The real and imaginary parts of the signal are subjected to nonlinear phase compensation respectively, and the cutoff frequency of the complex domain filter is offset compensated in combination with the dynamic coupling degree prediction value, so as to retain the steep characteristic of the pulse leading edge and suppress the cross-coupling components in the high-frequency interference band.

[0047] It should be further explained that, in the specific implementation process, in the complex domain phase-sensitive filtering process, the time domain signal of the high-frequency transient interference channel is first mapped to the complex domain, and nonlinear phase compensation is performed on the real and imaginary parts of the signal respectively. By extracting the phase derivative of the instantaneous frequency of the signal, the imaginary part compensation coefficient is dynamically adjusted so that the imaginary part phase change rate is inversely proportional to the real part phase change rate, thereby ensuring the time delay consistency of the high-frequency pulse waveform; based on the dynamic coupling degree prediction result, when the high-frequency interference energy gradient and the signal main frequency gradient show an enhanced inverse correlation, it is determined that the risk of high-frequency noise invading the main frequency band increases. At this time, the cutoff frequency of the complex domain filter is extended to the high frequency direction, and the high-frequency suppression bandwidth is increased to cover the cross-coupling frequency band. At the same time, the cutoff frequency adjustment step is dynamically reduced according to the real-time monitoring value of the pulse front steepness to avoid excessive filtering and resulting in pulse waveform blunting; when the signal main frequency gradient shows an enhanced inverse correlation, it is determined that the risk of high-frequency noise invading the main frequency band increases. At this time, the cutoff frequency of the complex domain filter is extended to the high frequency direction, and the high-frequency suppression bandwidth is increased to cover the cross-coupling frequency band. At the same time, the cutoff frequency adjustment step is dynamically reduced according to the real-time monitoring value of the pulse front steepness to avoid excessive filtering and resulting in pulse waveform blunting. When there is intermittent electromagnetic interference near the frequency, a dynamic weighting factor is introduced into the complex domain filter based on the predicted value of the coupling path offset direction to selectively attenuate the signal components in the edge area of ​​the main frequency band, thereby retaining the pulse energy integrity in the core area of ​​the main frequency. For instantaneous high-frequency noise mutations caused by load fluctuations, the correlation between the current change rate of the equipment and the noise spectrum offset is monitored in real time. If the current change rate exceeds the threshold and the main peak of the noise spectrum migrates to the low frequency, the phase compensation coefficient of the complex domain filter is triggered to be adaptively reset, giving priority to ensuring the fidelity of the pulse leading edge characteristics. In addition, in the signal reconstruction stage, the phase alignment of the real and imaginary parts of the filtered complex domain signal is checked. If the phase deviation of the rising edge of the pulse is detected to exceed the allowable range, the nonlinear compensation parameters are recalculated and the complex domain filtering is iteratively performed until the pulse waveform meets the preset time domain distortion tolerance threshold.

[0048] The specific implementation of vibration trajectory matched filtering includes:

[0049] The notch center frequency is dynamically generated according to the fundamental vibration frequency of the equipment and its harmonic components. The notch bandwidth is adaptively adjusted through the slope of the vibration signal envelope, and the time domain amplitude of the residual component is corrected in combination with the Kalman filter.

[0050] It should be further explained that, in the specific implementation process, in the vibration trajectory matching filtering process of the low-frequency periodic noise channel, the distribution characteristics of the fundamental frequency and its harmonic components are first extracted based on the historical data of the mechanical vibration spectrum of the equipment, and the notch center frequency matching the current working conditions is dynamically generated; according to the change in the slope of the envelope of the vibration signal collected in real time, the intensity fluctuation trend of the mechanical vibration is judged. If the slope of the envelope exceeds the preset threshold, it is determined that the vibration noise energy is positively concentrated in the low-frequency area. At this time, the notch bandwidth is extended to 1.5 times the fundamental frequency to cover the harmonic diffusion range. Conversely, if the slope is lower than the threshold, the notch bandwidth is compressed to 0.8 times the fundamental frequency to focus on the main vibration frequency; for the vibration frequency offset caused by temperature and humidity changes, a vibration noise time-domain amplitude prediction model is constructed through Kalman filtering, and the correlation between the notch residual component of the previous moment and the current vibration acceleration signal is used to update the state equation to dynamically correct the residual noise. amplitude compensation coefficient; when it is detected that the periodic discharge signal and vibration noise overlap in the time domain, the group delay parameter of the notch filter is adjusted in real time according to the phase difference between the two. If the phase of the discharge signal leads the vibration noise, the phase lag of the notch filter is increased to stagger the overlapping interval, otherwise the lag is reduced to preserve the integrity of the discharge pulse; for vibration spectrum distortion caused by sudden mechanical shock, by monitoring the instantaneous kurtosis value of the vibration signal, if the kurtosis exceeds 3 times the standard deviation of the normal operating condition, it is temporarily switched to broadband notch mode and the Kalman filter correction is suspended, and adaptive tracking is resumed after the shock ends; in addition, in the signal reconstruction stage after notch filtering, the time domain waveform of the residual component is periodically checked. If a residual oscillation component consistent with the equipment rotation period is detected, the fine-tuning mechanism of the notch center frequency is triggered, and the fundamental frequency value is shifted by 0.5% step by step until the amplitude of the residual component drops below the noise baseline.

[0051] The adversarial modality separation network is implemented in the following way:

[0052] A modal discriminator is set at the output end of the high-frequency channel and the low-frequency channel, respectively. The discriminator presupposes the physical constraints that the pulse interval of the corona discharge obeys the Poisson distribution and the periodicity of the internal discharge obeys the Gaussian distribution.

[0053] The difference between the separated signal and the target modal distribution is calculated based on the KL divergence, and the standard deviation of the pulse interval is introduced as an additional constraint term. The filtering parameters are reversely optimized through the adversarial loss function.

[0054] It should be further explained that, in the specific implementation process, during the physical feature compliance detection process, the dynamic threshold range is set according to the quantitative relationship between the local discharge pulse rise time and the insulation defect size. When the separated signal pulse rise time is less than 1ns and no subsequent oscillation waveform is detected, it is determined that the waveform distortion is caused by high-frequency filtering overshoot or low-frequency notch residue; at this time, the parameter recalibration mechanism is triggered to first check whether the complex domain filter cutoff frequency adjustment step of the high-frequency channel exceeds the allowable range. If the step exceeds 80% of the preset maximum value, the cutoff frequency is retracted to the average of the first three time windows and the adjustment step is reduced to 50%; if the pulse rise time is within the threshold range but there is a periodic oscillation aftermath, it is determined that the vibration trajectory matching filter of the low-frequency channel has failed, and the instantaneous estimate of the vibration fundamental frequency is recalculated and Reset the offset of the notch center frequency to eliminate the residual oscillation component; for changes in the dielectric constant of the insulating material caused by sudden changes in temperature and humidity, dynamically adjust the baseline value of the rise time threshold. When the ambient humidity rises by 10% and the temperature fluctuates by more than 5°C, relax the upper threshold limit to 1.2ns to match the changes in material properties; during the recalibration process, if the pulse parameters are still detected to be out of standard after two consecutive calibrations, it is determined that the multimodal coupling is not completely decoupled, and it is forced to switch to the full-band original signal to re-execute dynamic coupling modeling and dual-channel filtering; in addition, the pulse waveform is jointly verified for multiple parameters, including the coordinated consistency of rise time, peak amplitude, and half-wave width. If the deviation of any two of the three exceeds 20%, it is determined that the separated signal has irreversible distortion, the current window data is discarded, and the historical optimal parameters are enabled to reconstruct the output signal.

[0055] Physical characteristics compliance testing includes:

[0056] The threshold range is set according to the quantitative relationship between the rise time of the partial discharge pulse and the size of the insulation defect. When the rise time of the separated signal pulse is less than 1ns and there is no subsequent oscillation waveform, it is judged as a distorted signal and recalibration is triggered.

[0057] It should be further explained that, in the specific implementation process, during the physical feature compliance detection process, the dynamic threshold range is set according to the quantitative relationship between the local discharge pulse rise time and the insulation defect size. When the separated signal pulse rise time is less than 1ns and no subsequent oscillation waveform is detected, it is determined that the waveform distortion is caused by high-frequency filtering overshoot or low-frequency notch residue; at this time, the parameter recalibration mechanism is triggered to first check whether the complex domain filter cutoff frequency adjustment step of the high-frequency channel exceeds the allowable range. If the step exceeds 80% of the preset maximum value, the cutoff frequency is retracted to the average of the first three time windows and the adjustment step is reduced to 50%; if the pulse rise time is within the threshold range but there is a periodic oscillation aftermath, it is determined that the vibration trajectory matching filter of the low-frequency channel has failed, and the instantaneous estimate of the vibration fundamental frequency is recalculated and Reset the offset of the notch center frequency to eliminate the residual oscillation component; for changes in the dielectric constant of the insulating material caused by sudden changes in temperature and humidity, dynamically adjust the baseline value of the rise time threshold. When the ambient humidity rises by 10% and the temperature fluctuates by more than 5°C, relax the upper threshold limit to 1.2ns to match the changes in material properties; during the recalibration process, if the pulse parameters are still detected to be out of standard after two consecutive calibrations, it is determined that the multimodal coupling is not completely decoupled, and it is forced to switch to the full-band original signal to re-execute dynamic coupling modeling and dual-channel filtering; in addition, the pulse waveform is jointly verified for multiple parameters, including the coordinated consistency of rise time, peak amplitude, and half-wave width. If the deviation of any two of the three exceeds 20%, it is determined that the separated signal has irreversible distortion, the current window data is discarded, and the historical optimal parameters are enabled to reconstruct the output signal.

[0058] Multi-model classification confidence verification includes:

[0059] The separated signals are input into the pre-trained time-domain convolutional network and frequency-domain graph neural network in parallel, and the confidence variance of the classification results of each model is counted. When the variance exceeds the preset threshold, it is determined that modal aliasing exists.

[0060] It should be further explained that, in the specific implementation process, in the process of multi-model classification confidence verification, the separated signal is input into the pre-trained time-domain convolutional network and frequency-domain graph neural network in parallel to extract the time-domain pulse sequence characteristics and frequency-domain energy distribution characteristics of the signal respectively; when the corona discharge confidence output by the time-domain convolutional network is higher than 90%, and the internal discharge confidence of the frequency-domain graph neural network exceeds 85%, it is determined that there is a conflict in the cognition of the signal modality by the two types of models, and the variance values ​​of the two types of confidence are further calculated; if the variance exceeds the preset threshold, it is determined that the current signal has modal aliasing, triggering the re-decoupling instruction and returning to the dynamic coupling degree modeling stage; for the time-domain pulse feature distortion caused by high-frequency transient interference residues, by monitoring the confidence fluctuation amplitude of the time-domain convolutional network output, if the fluctuation amplitude exceeds 30% in three consecutive time windows, it is determined that the time-domain feature extraction fails, and the time-domain model is temporarily frozen. The model is validated based on the classification results of the frequency domain model. When the ambient noise spectrum mutates, the frequency band weight distribution of the frequency domain graph neural network is dynamically adjusted. If the frequency of the noise main peak shifts to within ±10% of the signal main frequency, the weight coefficient of the corresponding frequency band is reduced by 50% to suppress the impact of noise interference on the classification results. For the frequency domain energy tailing phenomenon caused by incomplete filtering of low-frequency vibration noise, the frequency domain graph neural network is used to detect the energy proportion of the subharmonic component. If the subharmonic energy exceeds 15% of the fundamental frequency energy, it is determined that the low-frequency channel filtering is insufficient, and the notch depth of the vibration trajectory matched filter is adjusted first. If the credibility of the verification result is insufficient, a similarity matching mechanism of historical separation success cases is introduced. If the matching degree between the time-frequency joint features of the current signal and the historical success case library is less than 60%, it is determined to be a new aliasing mode, and the manual auxiliary annotation process is initiated and the model parameter library is updated.

[0061] Nonlinear phase compensation is achieved by the following steps:

[0062] The phase derivative of the instantaneous frequency of the signal is extracted, and a compensation coefficient inversely proportional to the real part phase change rate is applied to the imaginary part of the complex domain signal to ensure the delay consistency of the high-frequency pulse waveform.

[0063] It should be further explained that, in the specific implementation process, in the process of nonlinear phase compensation of complex domain phase sensitive filtering, the phase derivative of the instantaneous frequency of the signal is extracted to calculate the difference in the phase change rate of the real part and the imaginary part in real time, and the compensation coefficient is dynamically allocated according to the direction of the difference; when the real part phase change rate is higher than the imaginary part, a compensation coefficient inversely proportional to the change rate difference is applied to the imaginary part, so that the imaginary part phase is accelerated to match the real part phase trend, and vice versa, the imaginary part compensation coefficient is reduced to delay the phase offset; for the phase nonlinear distortion caused by the sudden change of the high-frequency interference energy gradient, by monitoring the attenuation rate of the pulse front steepness, if the rate exceeds the preset critical value, it is determined that the phase compensation is insufficient, and the adaptive incremental adjustment of the compensation coefficient is triggered until the delay fluctuation of the pulse front converges to the allowable range; for the instantaneous frequency jump caused by load fluctuation, the calculation window length of the phase derivative is synchronously corrected according to the dynamic coupling degree prediction result, if If the prediction shows an increased risk of high-frequency noise intrusion, the window length is shortened to 1 / 4 of the current period to improve phase tracking sensitivity. During the signal reconstruction phase, the real and imaginary parts of the compensated complex domain signal are checked for phase alignment. If the peak moments of the real and imaginary parts of the rising edge of the pulse are detected to deviate by more than 0.1 microseconds, the compensation coefficient is recalculated and phase compensation is performed iteratively. The direction of the deviation is recorded as a reference for initializing the subsequent compensation coefficient. In response to the difference in medium propagation delay caused by changes in temperature and humidity, the baseline value of the phase derivative is dynamically corrected using environmental sensor data. If the ambient humidity rises by 15% and the temperature drops by more than 10°C, the baseline phase derivative is multiplied by 1.2 times the correction factor to offset the impact of changes in medium characteristics. If the phase alignment requirements cannot be met after multiple iterations, it is determined that the hardware acquisition channel is abnormal, and the backup signal channel is switched to and all compensation parameters are reset to ensure the continuous stability of the filtering process.

[0064] The adjustment rules for the notch bandwidth are:

[0065] When the slope of the vibration signal envelope exceeds the set threshold, the notch bandwidth is expanded to 1.5 times the fundamental frequency; when the slope of the envelope is lower than the threshold, the notch bandwidth is compressed to 0.8 times the fundamental frequency.

[0066] It should be further explained that, in the specific implementation process, during the adjustment of the notch bandwidth of the vibration trajectory matching filter, the instantaneous rate of change of the slope of the vibration signal envelope is monitored in real time. When the slope exceeds the set threshold, it is determined that the vibration noise energy is positively concentrated in the low-frequency harmonic area. At this time, the notch bandwidth is expanded to 1.5 times the fundamental frequency to cover the harmonic diffusion range, and the notch depth is increased to 120% of the original value to enhance low-frequency suppression; if the slope is lower than the threshold and the energy fluctuation of the main peak of the vibration spectrum is less than 5%, it is determined that the vibration noise is in a stable state, and the notch bandwidth is compressed to 0.8 times the fundamental frequency and the notch depth is reduced to 80% to avoid excessive filtering and attenuation of the discharge signal fundamental frequency; for the instantaneous offset of the vibration fundamental frequency caused by the sudden change of the equipment load, the Kalman filter is used to predict the fundamental frequency change trend of the next time window. If the predicted offset exceeds 2% of the current fundamental frequency, the tracking step of the notch center frequency is adjusted in advance, and During the bandwidth expansion phase, an offset compensation is added to prevent notch mismatch. When a sudden rise in ambient temperature is detected, causing the slope of the vibration signal envelope to increase abnormally sharply, the anti-temperature drift mode is temporarily enabled, and the slope judgment threshold is dynamically increased to 1.3 times the standard value. The maximum expansion amplitude of the notch bandwidth is limited to 1.2 times the fundamental frequency to prevent temperature drift interference from falsely triggering the filter. For intermittent broadband vibration noise caused by mechanical looseness, the spectrum flatness of the residual component after notching is analyzed. If the flatness index exceeds 0.7 and lasts for three time windows, it is determined to be broadband noise leakage, and the mode is switched to multi-stage series notch mode and the bandwidth overlap area is increased step by step. After the notch filter parameters are updated, the integrity of the periodic discharge pulse of the separated signal is reversely verified. If the standard deviation of the pulse interval increases by more than 15% compared with that before filtering, it is determined that the waveform is distorted due to excessive notching, and the system reverts to the previous parameter set and reduces the bandwidth adjustment amplitude to 50% of the original value.

[0067] The optimization process of the adversarial loss function includes:

[0068] The distribution parameters of the modal discriminator are updated synchronously in each iteration, so that the standard deviation of the pulse interval of corona discharge converges to the theoretical variance of Poisson distribution, and the periodic error of internal discharge converges to the standard deviation range of Gaussian distribution.

[0069] It should be further explained that, in the specific implementation process, in the process of optimizing the adversarial loss function, the deviation degree of the pulse interval standard deviation of the corona discharge from the theoretical variance of the Poisson distribution is monitored in real time, and the weight coefficient of the KL divergence in the adversarial loss function is dynamically adjusted; when the pulse interval standard deviation exceeds 15% of the theoretical variance, it is determined that the random characteristics of the high-frequency channel are insufficiently retained, and the KL divergence weight is increased to 1.2 times the initial value to strengthen the distribution matching constraint, while reducing the optimization weight of the periodic error term; for the periodic peak-to-peak error of the internal discharge, if it exceeds the range of three times the standard deviation of the Gaussian distribution, it is determined that the group delay parameter of the low-frequency channel notch filter is mismatched, and the gradient of the periodic error term is propagated back to the low-frequency channel through adversarial training, and the tracking step size of the notch center frequency is dynamically adjusted; when the statistical characteristics of the corona discharge and the internal discharge deviate from the target distribution synchronously, a joint evaluation mechanism of the distribution coupling degree is introduced. If the product of the two types of deviation exceeds the preset threshold, then The physical constraints of the modal discriminator are determined to be invalid, the adversarial training is suspended, and the historical optimal distribution parameters are reloaded; in response to the abnormal increase in the randomness of the pulse interval caused by load fluctuations, the moving average of the coefficient of variation is calculated through a sliding window. If the average value increases by more than 20% in three consecutive time windows, the robust optimization mode is triggered, the theoretical variance tolerance range of the Poisson distribution is relaxed to 1.3 times, and the correction frequency of the periodic error is increased; the distribution parameter threshold of the discriminator is synchronously updated in each iteration, so that the standard deviation of the pulse interval of the corona discharge gradually converges to the theoretical variance range of the Poisson distribution, and at the same time, the peak value of the periodic error of the internal discharge is forced to be compressed to within twice the standard deviation of the Gaussian distribution; for the case where the parameters cannot converge under extreme working conditions, if the standard deviation still exceeds the target range after five iterations, it is determined to be environmental interference or hardware abnormality, the current model parameters are frozen and switched to the static filtering mode based on the historical optimal solution, and an alarm signal is triggered to wait for manual intervention and calibration.

[0070] When implementing adaptive filtering and intelligent separation of multimodal partial discharge signals in a dynamic noise environment, the system first collects mixed signals in real time through sensors, while simultaneously acquiring equipment operating parameters and ambient noise signals as auxiliary inputs. Based on the time-frequency energy distribution characteristics of the signal and noise, a sliding window is used to extract the time-frequency mutual information entropy gradient sequence at the current and historical moments. A bidirectional gated recurrent unit is used to model temporal dependencies and predict the offset direction of the noise-signal coupling path within multiple future time windows.

[0071] Based on the prediction results, the signal is divided into two independent channels: high-frequency transient interference and low-frequency periodic noise. The high-frequency channel uses complex-domain phase-sensitive filtering technology to map the time-domain signal to the complex domain and implement nonlinear phase compensation. By dynamically adjusting the filter cutoff frequency, high-frequency interference is suppressed from invading the main frequency band while retaining the steepness of the pulse leading edge.

[0072] The low-frequency channel constructs an adaptive notch filter bank based on prior knowledge of the equipment's mechanical vibration spectrum. By matching the vibration noise trajectory in real time and correcting the residual component amplitude, it eliminates the time-domain overlap between the low-frequency noise and the periodic discharge signal. The output signals of the two channels are input into the adversarial modal separation network. By presetting the modal discriminator with the physical statistical characteristics of different discharge types, the separated signals are forced to satisfy the randomness of the corona discharge pulse interval and the periodic regularity of the internal discharge. The adversarial loss function is used to reversely optimize the filter parameters. The separated signals are subject to a double verification mechanism. The physical feature compliance detection module sets a dynamic threshold based on the quantitative relationship between the partial discharge pulse parameters and insulation defects, and triggers parameter recalibration when waveform distortion is detected. The multi-model classification confidence verification module inputs the signals into the time-domain and frequency-domain classification models in parallel, determines the modal aliasing risk through confidence variance analysis, and returns to iterative decoupling.

[0073] In response to working conditions such as load fluctuations, sudden changes in temperature and humidity, the correlation between equipment parameters and noise signals is monitored in real time, and the coupling degree prediction model and filter parameter adjustment step size are dynamically corrected; when sudden mechanical shock or unknown modal interference is detected, it switches to anti-interference mode and starts the manual intervention protocol to ensure system robustness; all modules form a closed-loop feedback, and through the synergy of dynamic coupling perception, dual-channel decoupling filtering, adversarial modal separation and double verification, high-fidelity separation of multimodal signals in dynamic noise environments is achieved.

[0074] By combining dynamic coupling modeling with dual-channel adversarial decoupling, the proposed method effectively addresses the modal aliasing and feature distortion issues inherent in traditional methods in dynamic noise environments, caused by the time-frequency coupling of signals and noise. The dynamic coupling path perception module quantifies the interaction between noise and signals in real time. Combining a dual-channel architecture of complex-domain phase-sensitive filtering and vibration trajectory matching filtering, the proposed method achieves precise suppression of high-frequency transient interference and low-frequency periodic noise. The adversarial modal separation network, through physical statistical constraints and adversarial training, enforces the independent distribution of discharge types in the separated signals, significantly improving the separation accuracy and anti-interference capabilities of multimodal signals.

[0075] By introducing a dual guarantee mechanism of physical feature compliance detection and multi-model confidence verification, this approach addresses the diagnostic reliability issues inherent in traditional algorithms due to cascaded error accumulation, ensuring that separation results adhere to the dual constraints of both physical laws and data models. Under complex operating conditions of power equipment, this method can adaptively track the dynamic coupling of noise and signal, reducing false positives and missed detections. This provides highly reliable data support for early warning and precise location of insulation defects, while reducing the need for manual intervention and improving the practicality and engineering deployment efficiency of intelligent diagnostic systems.

[0076] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0077] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for adaptive filtering and intelligent separation of multimodal partial discharge signals, characterized in that: The following steps are involved: Step S1: real-time acquisition of partial discharge signals of power equipment, and simultaneous acquisition of equipment operating parameters and environmental noise signals; Step S2: Calculate the dynamic coupling index based on the time-frequency matrix of the signal and noise, extract the time-frequency energy distribution through the sliding window, and generate a dynamic coupling strength quantization value based on the inverse correlation between the high-frequency interference energy gradient and the signal main frequency gradient; Step S3: Based on the timing prediction results of the dynamic coupling index, the signal is divided into a high-frequency transient interference channel and a low-frequency periodic noise channel. Complex-domain phase-sensitive filtering and vibration trajectory matching filtering are used for noise suppression, respectively. The complex-domain phase-sensitive filtering maps the signal to the complex domain using a transient pulse compression encoder and dynamically adjusts the filter cutoff frequency. The vibration trajectory matching filtering constructs an adaptive notch filter bank based on prior knowledge of the equipment's mechanical vibration spectrum. Step S4: Input the output signals of the high-frequency channel and the low-frequency channel into the adversarial modal separation network, and force the separated signals to meet the preset independent modal statistical distribution through the modal discriminator; Step S5: The separation results are cross-validated based on physical feature compliance detection and multi-model classification confidence verification. If the pulse parameter exceeds the threshold or the classification confidence variance exceeds the standard, parameter recalibration is triggered and decoupling is re-iterated.

2. The method for adaptive filtering and intelligent separation of multimodal partial discharge signals according to claim 1, characterized in that: The time series prediction of the dynamic coupling index models the time-frequency mutual information entropy gradient sequence through a bidirectional gated recurrent unit to predict the coupling path offset direction of the noise and signal in the next three time windows.

3. The method for adaptive filtering and intelligent separation of multimodal partial discharge signals according to claim 1, characterized in that: The implementation of the complex domain phase sensitive filtering includes: The real and imaginary parts of the signal are subjected to nonlinear phase compensation respectively, and the cutoff frequency of the complex domain filter is offset compensated in combination with the dynamic coupling degree prediction value, so as to retain the steep characteristic of the pulse leading edge and suppress the cross-coupling components in the high-frequency interference band.

4. The method for adaptive filtering and intelligent separation of multimodal partial discharge signals according to claim 1, characterized in that: The implementation of the vibration trajectory matched filtering includes: The notch center frequency is dynamically generated according to the fundamental vibration frequency of the equipment and its harmonic components. The notch bandwidth is adaptively adjusted through the slope of the vibration signal envelope, and the time domain amplitude of the residual component is corrected in combination with the Kalman filter.

5. The method for adaptive filtering and intelligent separation of multimodal partial discharge signals according to claim 1, characterized in that: The adversarial modality separation network is implemented as follows: A modal discriminator is set at the output end of the high-frequency channel and the low-frequency channel, respectively. The discriminator presupposes the physical constraints that the pulse interval of the corona discharge obeys the Poisson distribution and the periodicity of the internal discharge obeys the Gaussian distribution. The difference between the separated signal and the target modal distribution is calculated based on the KL divergence, and the standard deviation of the pulse interval is introduced as an additional constraint term. The filtering parameters are reversely optimized through the adversarial loss function.

6. The method for adaptive filtering and intelligent separation of multimodal partial discharge signals according to claim 1, characterized in that: The physical feature compliance detection includes: setting a threshold range based on the quantitative relationship between the partial discharge pulse rise time and the insulation defect size, and when the rise time of the separated signal pulse is less than 1ns and there is no subsequent oscillation waveform, it is determined to be a distorted signal and triggers recalibration.

7. The method for adaptive filtering and intelligent separation of multimodal partial discharge signals according to claim 1, characterized in that: The multi-model classification confidence verification includes: inputting the separated signals into the pre-trained time domain convolutional network and frequency domain graph neural network in parallel, counting the confidence variance of the classification results of each model, and determining the presence of modal aliasing when the variance exceeds a preset threshold.

8. The method for adaptive filtering and intelligent separation of multimodal partial discharge signals according to claim 3, characterized in that: The nonlinear phase compensation is achieved through the following process, including: extracting the phase derivative of the instantaneous frequency of the signal, applying a compensation coefficient inversely proportional to the real part phase change rate to the imaginary part of the complex domain signal, and ensuring the delay consistency of the high-frequency pulse waveform.

9. The method for adaptive filtering and intelligent separation of multimodal partial discharge signals according to claim 4, characterized in that: The adjustment rule of the notch bandwidth is: when the slope of the vibration signal envelope exceeds the set threshold, the notch bandwidth is expanded to cover the diffusion range of the vibration noise harmonics; when the slope of the envelope is lower than the threshold, the notch bandwidth is compressed to the matching range of the focused main vibration frequency.

10. The method for adaptive filtering and intelligent separation of multimodal partial discharge signals according to claim 5, characterized in that: The optimization process of the adversarial loss function includes: The distribution parameters of the modal discriminator are updated synchronously in each iteration, so that the standard deviation of the pulse interval of corona discharge converges to the theoretical variance of Poisson distribution, and the periodic error of internal discharge converges to the standard deviation range of Gaussian distribution.

Images