A method and system for partial discharge detection in cable branch boxes

Abstract

This application relates to the field of discharge detection technology, specifically to a method and system for partial discharge detection in cable branch boxes. The method includes: acquiring and preprocessing multi-channel real-time monitoring signals from the cable branch box to obtain an analysis signal; performing mode decomposition on the analysis signal to obtain multiple mode components, extracting statistical features characterizing periodicity and energy distribution patterns from each mode component, identifying and filtering out mode components containing narrowband interference; performing wavelet decomposition on the remaining mode components, adaptively adjusting the denoising threshold of the corresponding wavelet layer for filtering, and reconstructing the denoised effective signal; extracting statistical features from the effective signals under multiple channels to construct a fused feature vector, inputting the fused feature vector into a pre-trained partial discharge detection model, and outputting the partial discharge detection result of the cable branch box. This improves the accuracy of partial discharge detection.

Description

Technical Field

[0001] This application relates to the field of discharge detection technology, specifically to a partial discharge detection method and system for cable branch boxes. Background Technology

[0002] Cable branch boxes serve as crucial hubs in urban power distribution networks, connecting main lines to end users. Within a confined space, they house multiple incoming and outgoing lines and primary electrical equipment. Insulation aging or moisture absorption can easily trigger partial discharge. The high-energy electrons generated by partial discharge continuously damage the insulation, and long-term accumulation can lead to insulation breakdown or even the branch box exploding. Therefore, partial discharge detection in cable branch boxes is of paramount importance.

[0003] However, in actual partial discharge detection, the electromagnetic environment of the cable branch box is extremely complex, inevitably resulting in severe narrowband interference and high-frequency electromagnetic noise interference caused by external communication signals and the operation of surrounding electronic equipment. The distribution characteristics of these interference signals in the time and frequency domains often highly overlap and confuse with the weak partial discharge signals. Existing detection and filtering methods are insufficient to effectively suppress complex mixed interference while accurately preserving the detailed characteristics of partial discharge, leading to a high likelihood of misjudgment or missed detection during partial discharge detection. Summary of the Invention

[0004] To address the aforementioned technical problems, the purpose of this application is to provide a method and system for detecting partial discharge in cable branch boxes, the specific technical solution of which is as follows: In a first aspect, embodiments of this application provide a partial discharge detection method for a cable branch box, the method comprising the following steps: Acquire multi-channel real-time monitoring signals from the cable branch box and preprocess them to obtain analysis signals; The analysis signal is subjected to mode decomposition to obtain multiple mode components. Statistical features characterizing periodicity and energy distribution patterns are extracted from each mode component. Based on these statistical features, mode components containing narrowband interference are identified and filtered out. Wavelet decomposition is performed on the remaining modal components after filtering. Based on the amplitude distribution differences of pulse extrema points in each wavelet layer and the similarity of local pulse waveforms, the wavelet threshold of the corresponding wavelet layer is adaptively adjusted for filtering and the denoised effective signal is reconstructed. The statistical features of the effective signals under multiple channels are extracted and time-aligned to construct a fused feature vector. The fused feature vector is then input into a pre-trained partial discharge detection model to output the partial discharge detection results of the cable branch box.

[0005] In one embodiment, the step of extracting statistical features characterizing periodicity and energy distribution patterns from each modal component includes: Calculate the autocorrelation coefficients of each modal component within the preset hysteresis range, and mark the maximum autocorrelation coefficient corresponding to each modal component; Determine the kurtosis value of each modal component and the sum of all energy values ​​of each modal component; and perform linear fitting on each modal component to obtain the mean of all fitting residuals corresponding to each modal component. Based on the maximum autocorrelation coefficient, the kurtosis value, the summation and the mean, the statistical characteristics corresponding to each modal component are determined.

[0006] In one embodiment, the statistical characteristics corresponding to each modal component are: The maximum autocorrelation coefficient is negatively correlated, and the result of the negative correlation mapping, the kurtosis value, the sum and the mean are combined to form a multidimensional vector, which serves as the statistical feature corresponding to each modal component.

[0007] In one embodiment, identifying and filtering modal components containing narrowband interference includes: Obtain the modal components of the historical analysis signal of the cable branch box, and perform threshold segmentation on the multidimensional vectors based on the numerical distribution characteristics of the multidimensional vectors corresponding to all modal components of the historical analysis signal and the current analysis signal. The modal components corresponding to multidimensional vectors smaller than the segmentation threshold are considered as modal components containing narrowband interference, and the modal components containing narrowband interference are screened out from all modal components of the current analyzed signal.

[0008] In one embodiment, the adaptive adjustment of the wavelet threshold corresponding to the wavelet layer includes: Identify the peak value of the detail coefficients within each wavelet layer, and determine the difference between the peak value and the average level of the detail coefficients within the wavelet layer, which is denoted as the first difference; Centered on each peak point, the local detail coefficient segments corresponding to each peak are extracted within the wavelet layer to determine the metric distance between local detail coefficient segments of different peaks within the wavelet layer; By comparing the degree of dispersion of the measured distance with the first difference, the comprehensive feature value of each wavelet layer is obtained, which is used to adjust the wavelet threshold of the corresponding wavelet layer.

[0009] In one embodiment, obtaining the comprehensive feature values ​​of each wavelet layer includes: Calculate the average value of all the first differences within the wavelet layer. The comprehensive feature value is positively correlated with the average value and negatively correlated with the degree of dispersion.

[0010] In one embodiment, the step of adjusting the wavelet threshold corresponding to the wavelet layer includes: The normalized comprehensive eigenvalues ​​of each wavelet layer are positively shifted and mapped to obtain the adaptive adjustment factor of each wavelet layer. The wavelet threshold of the corresponding wavelet layer is then calculated, with the specific expression as follows: In the formula, The wavelet threshold is the optimized wavelet coefficient of the j-th wavelet at the i-th level after wavelet decomposition of the k-th intrinsic mode component. The fixed wavelet threshold used after wavelet decomposition of the k-th intrinsic mode component. This is the adaptive adjustment factor for the i-th wavelet layer after wavelet decomposition of the k-th intrinsic mode component. This is the preset steepness adjustment factor.

[0011] In one embodiment, the adaptive adjustment factor is calculated as follows: Calculate the product of the normalized comprehensive eigenvalues ​​of each wavelet layer and the preset basic adjustment factor, wherein the adaptive adjustment factor is the sum of the product and the value 1.

[0012] In one embodiment, the statistical features in the fused feature vector include the mean, variance, kurtosis factor, peak factor, impulse factor, margin factor, waveform factor, and skewness of the effective signal under multiple channels.

[0013] Secondly, embodiments of this application also provide a partial discharge detection system for a cable branch box, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0014] This application has at least the following beneficial effects: This application addresses the pain point of narrowband interference being prevalent in cable branch box environments and easily masking weak partial discharge characteristics by introducing a priori screening mechanism based on modal statistical features. It extracts the intrinsic modal components of the signal through modal decomposition and jointly calculates multiple statistical features characterizing periodicity and energy distribution to quantitatively assess the regularity of narrowband interference. Compared to traditional filtering methods that rely on fixed-band hard cutoff, this application can adaptively identify and accurately remove narrowband interference components with severe spectral aliasing, thus purifying the input source for subsequent analysis at its root. Secondly, addressing the problem that external high-frequency environmental noise and partial discharge pulses exhibit highly similar behavior in the time-frequency domain, easily leading to "useful pulses being mistakenly deleted" or "high-frequency spikes being misjudged as partial discharge" during denoising, this application employs an adaptive wavelet threshold control strategy. By deeply analyzing the highly reproducible morphological pattern of partial discharge, characterized by "stationary-extreme abrupt change-stationary," and dynamically adjusting the soft threshold contraction of each wavelet layer based on the amplitude difference of extreme points and the similarity of local pulse waveforms, this approach breaks the constant smoothing deviation caused by the fixed threshold used in traditional denoising algorithms. While filtering out irregular high-frequency spikes, it achieves high-fidelity reconstruction of transient weak details of partial discharge. Finally, this application constructs a multi-dimensional fusion feature vector, which, combined with the powerful high-dimensional nonlinear classification capability of the pre-trained partial discharge detection model, significantly improves the online identification accuracy and anti-interference robustness of insulation defects (partial discharge) in cable branch boxes under complex distribution network conditions, effectively reducing the risk of equipment explosion due to insulation breakdown. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the steps of a partial discharge detection method for a cable branch box, as provided in one embodiment of this application. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by this application to achieve the intended purpose of the invention, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a partial discharge detection method and system for cable branch boxes proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0019] The following description, in conjunction with the accompanying drawings, details a specific scheme for a partial discharge detection method and system for cable branch boxes provided in this application.

[0020] Please see Figure 1 The diagram illustrates a flowchart of a partial discharge detection method for a cable branch box according to an embodiment of this application. The method includes the following steps: S1: Acquire multi-channel real-time monitoring signals from the cable branch box and perform preprocessing to obtain analysis signals.

[0021] To comprehensively and accurately monitor the insulation status inside the cable branch box, this embodiment employs a multi-type sensor array to simultaneously acquire multi-channel real-time monitoring signals from the cable branch box, and uses these signals as a basis for feature extraction and status determination. The specific data acquisition and deployment methods are as follows: First, a high-frequency current sensor is used to detect high-frequency pulse current signals. During installation, a high-frequency current sensor is mounted on the grounding wire of the cable branch box to acquire the high-frequency pulse current generated by the cable branch box in real time. Its sampling rate is configured to 100MHz, but the implementer can set it according to the actual situation.

[0022] Secondly, ultrasonic sensors are used to detect the vibration ultrasonic signals generated during the operation of electrical equipment in real time. Specifically, the ultrasonic sensor probe is positioned at openings or discontinuities in the cable branch box enclosure, such as gaps, ventilation holes, and door edges, where sound waves can easily conduct, to maximize the capture of discharge sound pressure waves inside the box. The sampling rate is configured to 1MHz. In this embodiment, the ultrasonic sensor probe is positioned at the ventilation holes of the cable branch box enclosure.

[0023] Next, a transient ground voltage sensor is used to acquire the transient ground voltage signal in real time. Specifically, the transient ground voltage sensor is tightly attached to the inner wall or outer surface of the metal enclosure of the cable branch box to detect high-frequency electromagnetic waves excited by partial discharge and propagating along the metal armor surface. Its sampling rate is configured to 100MHz. In this embodiment, the transient ground voltage sensor is tightly attached to the outer surface of the metal enclosure near the cable joint.

[0024] Furthermore, the multi-channel real-time monitoring signals collected by each sensor will be timestamped and stored in the local cache under the same time base.

[0025] Considering that various types of sensors use high sampling rates when acquiring data, resulting in extremely large amounts of raw signal data, this places a huge computational burden on real-time signal processing. Therefore, this embodiment performs filtering and downsampling preprocessing on the acquired raw signals, significantly reducing the data volume while retaining effective partial discharge characteristic information.

[0026] Specifically, the acquired transient ground voltage signal is abbreviated as voltage signal. First, the real-time acquired high-frequency pulse current signal and transient ground voltage signal are used as inputs, and a bandpass filter is used for filtering. The lower limit of the passband is set to 500kHz and the upper limit of the passband is set to 30MHz to extract the main energy frequency band at which partial discharge occurs. Considering that the highest frequency component of the signal is close to 30MHz, if downsampling is performed, it will be difficult to reserve enough anti-aliasing protection frequency band, which will easily cause spectral aliasing and thus cause high-frequency characteristic distortion. Therefore, this embodiment does not perform downsampling to retain the complete signal details at the original sampling rate and outputs the preprocessed current signal and voltage signal.

[0027] Similarly, for the real-time acquired vibration ultrasonic signal, a bandpass filter is used for filtering, with the lower passband limit set to 20kHz and the upper passband limit set to 200kHz to retain the main energy frequency band of partial discharge in the ultrasonic signal; then, a decimation algorithm is used for downsampling, with the decimation factor set to 2, which significantly reduces the amount of data by half while ensuring that the ultrasonic features are not confused, and outputs the preprocessed ultrasonic signal.

[0028] The preprocessed current signal, voltage signal, and ultrasonic signal are all used as analysis signals.

[0029] S2, perform mode decomposition on the analysis signal to obtain multiple mode components, extract statistical features representing periodicity and energy distribution patterns from each mode component, and identify and filter out mode components containing narrowband interference based on the statistical features.

[0030] Because partial discharge pulses are transient, this embodiment divides the preprocessed analysis signals into time windows, setting the time window length to 20 milliseconds. Implementers can adjust this according to their specific circumstances. The actual environment where cable branch boxes are located typically contains various stable, continuous, periodic external electromagnetic signals, resulting in widespread narrowband interference in the detection signals, which easily overlaps significantly with the signal spectrum of partial discharge. Variational Mode Decomposition (VMD) can decompose mixed signals into several intrinsic mode components (IMFs) with different center frequencies and limited bandwidths, thereby effectively removing narrowband interference concentrated in the spectrum.

[0031] Taking any type of analysis signal within a single time window as an example, the analysis signal is used as input and decomposed using the VMD algorithm. To ensure that the decomposition effect is neither overfitting nor under-decomposing, the preset value of the number of modes K is set to 5 to 15, and in this embodiment, it is preferably set to 10. At the same time, the penalty factor of the VMD algorithm is set to 2000 to strictly limit the bandwidth of each mode component, ensure the spectral compactness of narrowband interference, and output 10 intrinsic mode components after decomposition. The variational mode decomposition (VMD) algorithm is a well-known existing technology, and the specific process will not be described in detail.

[0032] Because narrowband interference has an extremely concentrated spectrum, it can be completely and individually decomposed into a few specific intrinsic mode components. Therefore, it is necessary to first locate and filter out these interference components. The theoretical basis for this identification is that when the intrinsic mode components are dominated by narrowband interference, their waveforms usually exhibit smooth, regular periodic sine waves, with energy highly concentrated in a local area and relatively low total mode energy. Conversely, when the intrinsic mode components are dominated by partial discharge or external high-frequency electromagnetic noise, since both discharge and noise are randomly generated, their waveforms exhibit irregularly distributed transient spikes, with irregular energy distribution and global divergence, and relatively high total mode energy. Based on these differences in characteristics, all intrinsic mode components are quantitatively evaluated and screened.

[0033] Based on the above analysis, taking the k-th intrinsic mode component as an example, when this intrinsic mode component is dominated by narrowband interference, its waveform usually exhibits high periodicity; conversely, if this component mainly contains partial discharge pulses or external high-frequency noise, its waveform often exhibits random and irregular fluctuations. Based on this difference, the amplitude sequence composed of all numerical elements in the k-th intrinsic mode component is extracted as input, and its periodicity is quantified by calculating its autocorrelation coefficient sequence.

[0034] Specifically, the maximum delay step size (i.e., the maximum period threshold) for autocorrelation calculation is set to one-quarter of the total sampling length of the intrinsic mode component to ensure a sufficient number of sample pairs participating in the correlation calculation, thus making the statistical results stable and reliable. After obtaining the autocorrelation coefficient sequence, the trivial autocorrelation peaks at a delay step size of 0 (this value is always 1) are removed, and the maximum value of the autocorrelation coefficient is extracted within the remaining delay range.

[0035] The larger the obtained maximum autocorrelation coefficient value, the stronger the periodicity of the k-th intrinsic mode component, indicating a greater likelihood that it contains narrowband interference. To maintain a positive correlation with subsequent irregular features, a negative correlation mapping is performed on the maximum autocorrelation coefficient, and the result of this negative correlation mapping is defined as the out-of-order characteristic value of the k-th intrinsic mode component. In this embodiment, the negative correlation mapping is performed by calculating the reciprocal of the maximum autocorrelation coefficient and using this reciprocal as the out-of-order characteristic value. To avoid computational overflow caused by the maximum autocorrelation coefficient being 0, a very small positive number of 0.01 is added to the denominator to ensure that the denominator value is not zero. Implementers can choose other feasible negative correlation mapping methods, such as taking the negative of the maximum autocorrelation coefficient and then performing a non-negative transformation to obtain disordered characteristic values.

[0036] Furthermore, considering the significant differences in the probability distribution characteristics of waveform amplitudes: in mode components dominated by narrowband interference, the waveforms are mostly regular quasi-sine waves, with relatively uniform amplitude distribution and clustered within a certain range, lacking prominent transient spikes, thus exhibiting low kurtosis values; while in mode components dominated by partial discharge or external high-frequency noise interference, the presence of numerous irregular transient high-energy discharge pulses or spikes leads to extremely uneven amplitude distribution, and the probability density at the tail (extreme value region) increases significantly, thus exhibiting high kurtosis values. Based on these differences, the kurtosis value of the amplitude sequence of each intrinsic mode component is calculated sequentially, and the kurtosis value calculated for the k-th intrinsic mode component is denoted as... This is used to characterize the significance of spike abrupt changes in the signal component.

[0037] Considering that the energy of narrowband interference is often concentrated in a specific continuous frequency band, its energy fluctuations are smooth and its distribution is relatively concentrated in the time domain. Except for the local dominant region, the energy is extremely low in other time periods. However, under the conditions of partial discharge and external high-frequency noise, the discharge pulses flash randomly, and the energy distribution shows extremely irregular and discrete global distribution characteristics. Moreover, the total accumulated energy value is often high.

[0038] Therefore, the instantaneous energy value (i.e., the square of the amplitude) corresponding to each sampling point in the k-th intrinsic mode component is first calculated. Next, the sequence of instantaneous energies from all sampling points in the k-th intrinsic mode component is used as input, and a first-order linear fit is performed using the least squares method to obtain the baseline energy trend line for that mode component. Subsequently, the absolute fitting residual between the instantaneous energy value of each sampling point and the corresponding point on the baseline energy trend line is calculated. For narrowband interference with gently concentrated energy, the residual deviation from the trend line is usually small; however, for partial discharge or noise components containing a large number of random abrupt pulses, the residual fluctuations are extremely drastic.

[0039] Based on this, the mean of the absolute fitting residuals of all sampling points in the k-th intrinsic mode component is calculated. This quantifies the degree of irregular fluctuation in energy distribution; simultaneously, it calculates the cumulative sum of energy values ​​at all sampling points of the k-th intrinsic mode component. It is used to evaluate the overall energy intensity of the component.

[0040] Based on the above multi-dimensional analysis features, in order to accurately locate and filter out intrinsic mode components with narrowband interference, a multi-dimensional feature vector representing the anti-interference characteristics of the k-th intrinsic mode component is constructed. .

[0041] When narrowband interference exists in the intrinsic mode components, the waveform exhibits a significant periodicity, uniform amplitude distribution without obvious sudden spikes, and highly concentrated energy at a specific frequency. Based on this characteristic, the disordered eigenvalues ​​corresponding to the narrowband interference components... kurtosis value The mean of the absolute fitting residuals Total sum All will be at a low level.

[0042] To achieve dimensional uniformity, for each intrinsic mode component obtained after modal decomposition of the same type of analysis signal within a single time window, the vector elements at the same position in their multidimensional feature vectors are normalized using the maximum-minimum normalization method. Since narrowband interference in the cable branch box environment exhibits spatiotemporal stability, a 10-minute reference signal from the historical operation process is pre-acquired. Multidimensional feature vectors of all intrinsic mode components within this time period are extracted using the method described above, and the sum of the four normalized features in each multidimensional feature vector is calculated. The sum of all acquired features is used as the input sample set, and the Otsu thresholding method is used to calculate the maximum inter-class variance, automatically obtaining the optimal threshold for distinguishing between "narrowband interference" and "non-narrowband interference." The reference signal and the analysis signal are of the same type.

[0043] During real-time detection, the sum of all normalized features in the multidimensional feature vector of each intrinsic mode component of the current analysis signal is calculated. If the sum is less than the above-mentioned optimal threshold, the intrinsic mode component is determined to be dominated by narrowband interference and is removed. Only the remaining intrinsic mode components are retained for subsequent partial discharge feature depth processing.

[0044] S3. Wavelet decomposition is performed on the remaining modal components after filtering. Based on the amplitude distribution differences of pulse extrema points in each wavelet layer and the similarity of local pulse waveforms, the wavelet threshold of the corresponding wavelet layer is adaptively adjusted for filtering and the denoised effective signal is reconstructed.

[0045] The remaining intrinsic mode components after filtering out narrowband interference are used as inputs, and an 8-level decomposition is performed using the db10 wavelet basis. The high-frequency detail coefficients corresponding to each intrinsic mode component in each level are extracted as the target objects for subsequent partial discharge feature analysis. The wavelet basis function and the number of decomposition levels can be set by the implementer according to the actual situation, and this embodiment does not impose any restrictions on them.

[0046] In terms of physical behavior, if partial discharge occurs, the signal will generate transient spike pulses in the corresponding frequency band. The detail coefficients of the wavelet layer will suddenly increase in amplitude and reach extreme values ​​at the moment of discharge, and tend to zero at non-discharge moments, exhibiting a sparse pulse distribution overall. Conversely, external high-frequency electromagnetic interference caused by white noise or high-power instruments will cause the signal to fluctuate frequently and irregularly. The detail coefficients of the wavelet layer will exhibit high-frequency glitches with dense fluctuation peaks and lack significant abrupt extreme values.

[0047] Based on the aforementioned characteristic differences, taking the i-th layer detail coefficient of the k-th intrinsic mode component as an example, a peak-finding algorithm is used to extract peak points. To eliminate dense high-frequency spikes and spurious peaks, an amplitude filtering threshold is set, and only points with an absolute amplitude value greater than this threshold are selected as valid peak points. In this embodiment, the peak-finding algorithm uses the AMPD algorithm; implementers can use other peak-finding algorithms, such as the derivative method. In this embodiment, the amplitude filtering threshold is set to three times the overall standard deviation of the corresponding layer detail coefficients.

[0048] Subsequently, the absolute value of the difference between each effective peak point and the mean amplitude of all detail coefficients in that layer is calculated and denoted as the first difference. The average of all calculated absolute values ​​of difference is then denoted as the mean difference feature. , The larger the value, the farther the extracted peak value deviates from the overall baseline, indicating that the wavelet coefficients of this layer are more likely to have typical transient extreme value characteristics of partial discharge.

[0049] When partial discharge occurs, its typical form on the corresponding wavelet layer detail coefficients is "stationary-extreme abrupt change-stationary," and the waveform fluctuation characteristics of different discharge pulses generated by the same discharge source before and after the abrupt change are highly similar. Conversely, glitches in the pure noise layer are often random and irregular. Based on this, taking each effective peak point as the center, the range of N detail coefficients before and after it is taken as the characteristic interval of the corresponding effective peak point. If it is close to the beginning or end of the signal, resulting in less than N detail coefficients on one side, it is truncated according to the actual number of remaining points. The detail coefficients in each characteristic interval are arranged in a time series to form the local morphology sequence corresponding to each effective peak point. In this embodiment, N is set to 20, and the implementer can set it according to the actual situation.

[0050] To quantify waveform similarity and avoid random errors caused by a single reference point, valid peak points are sorted from largest to smallest based on their absolute peak value, and the top-ranked peak points are selected. There are several significant peak points in this embodiment. The value is set to 5; if the total number of valid peak points is less than 5, the actual number is used. When the selected number is greater than 1, the Frechet distance between each pair of local morphological sequences corresponding to these significant peak points is calculated; further, the standard deviation of all the above Frechet distances is calculated. Specifically, if there is only one valid peak point in the layer, then the standard deviation is directly calculated. Assign a minimum value of 0.01. As one method of distance measurement, the Fraser distance can be used by the implementer to choose other available feasible distance measurement methods. Standard deviation Used to measure the dispersion of the Fraser distance.

[0051] Standard deviation The smaller the standard deviation, the higher the consistency of the extracted pulse waveform morphology, and the more it conforms to the characteristics of partial discharge; conversely, if the standard deviation is larger... A larger value indicates that the fluctuations around the extracted maximum value are irregular and are more likely to be random noise.

[0052] Furthermore, based on the aforementioned mean difference characteristics With the standard deviation Calculate the comprehensive feature value of the i-th wavelet layer, which is used to adjust the denoising threshold of the corresponding wavelet layer. Specifically, calculate the mean difference feature. with standard deviation The ratio is used as the comprehensive feature value of the i-th wavelet layer. During the fractional calculation, when the denominator is 0, a parameter tuning factor can be added to the denominator to prevent it from being zero, ensuring the calculation result is meaningful. This parameter tuning factor is a very small positive number. For example, the value of this parameter tuning factor can be 0.01. Its specific value can be set by the implementer according to the actual situation; this application embodiment does not impose specific limitations.

[0053] The maximum-minimum normalization method is used to normalize the comprehensive eigenvalues ​​corresponding to all wavelet layers of the k-th intrinsic mode component, so as to obtain the normalized result of the comprehensive eigenvalues ​​of each wavelet layer.

[0054] Using a fixed threshold in conventional wavelet thresholding can easily lead to the erroneous deletion of useful signals or the retention of noise. In this embodiment, in order to preserve the transient weak features of partial discharge to the maximum extent while denoising, adaptive intervention is required based on the feature attributes of each wavelet layer: if the partial discharge features contained in the detail coefficients of a certain layer are more significant, that is, the normalized comprehensive feature value is closer to 1, the wavelet denoising threshold of that layer should be adaptively reduced to weaken the contraction force in order to avoid the smooth erasure of weak discharge pulses and prevent missed detection; conversely, if a certain layer is dominated by external irregular noise, the denoising threshold of that layer should be adaptively maintained to ensure that the noise is effectively filtered out.

[0055] Based on the above control logic, taking the detail coefficients of the i-th layer of the k-th intrinsic mode component as an example, a soft threshold denoising algorithm based on adaptive adjustment of normalized synthetic eigenvalues ​​is used for filtering. The wavelet threshold of the i-th layer wavelet coefficients is calculated, and the specific expression is as follows: In the formula, The wavelet threshold is the optimized wavelet coefficient of the j-th wavelet at the i-th level after wavelet decomposition of the k-th intrinsic mode component. The fixed wavelet threshold used after wavelet decomposition of the k-th intrinsic mode component. This is the adaptive adjustment factor for the i-th wavelet layer after wavelet decomposition of the k-th intrinsic mode component. The steepness adjustment factor is a preset value. The adaptive adjustment factor is obtained by positively shifting the normalized comprehensive feature value of the i-th wavelet layer. In this embodiment, the product of the normalized comprehensive feature value of the i-th wavelet layer and the preset basic adjustment factor is calculated. The adaptive adjustment factor is the sum of this product and the value 1. The basic adjustment factor has a value range of (1,2), and in this embodiment, the basic adjustment factor is set to 1.2. The steepness adjustment factor has a value range of (0,1), and in this embodiment, the steepness adjustment factor is set to 0.25.

[0056] For the fixed wavelet threshold, this embodiment uses the fixed threshold rule sqtwolog to obtain it.

[0057] After obtaining the optimized wavelet threshold, targeted filtering, truncation, and shrinkage processing are performed on the wavelet coefficients at different levels in the intrinsic mode components. Specifically: Compare the current wavelet coefficients The absolute value and the corresponding calculated optimized wavelet threshold ,like The wavelet coefficients were determined to be primarily noise, and were adjusted to 0. If the wavelet coefficients contain valid high-frequency characteristics of partial discharge, then the formula is used to determine that the wavelet coefficients contain effective high-frequency characteristics of partial discharge. The process involves shrinking the material, where sign() is the sign function. It represents the j-th wavelet coefficient of the i-th wavelet layer after wavelet decomposition of the k-th intrinsic mode component.

[0058] The above-mentioned flexible contraction method based on adaptive optimized threshold not only overcomes the problem of reconstructed signal oscillation caused by the discontinuity at the cutoff point of the traditional hard threshold function, but also corrects the excessive smoothing deviation of useful signal caused by the use of constant attenuation limit in the traditional soft threshold. Thus, while thoroughly filtering out high-frequency random noise, it restores the true transient details of partial discharge in cable branch boxes with high fidelity.

[0059] For the extremely complex electromagnetic environment inside cable branch boxes, traditional wavelet thresholding denoising often leads to serious misjudgments because it cannot distinguish between narrowband interference, random noise, and real partial discharge. The core logic of this application lies in fully exploring the differences in statistical distribution characteristics of each detail coefficient layer after wavelet decomposition under partial discharge (manifested as significant extreme pulses with strong morphological reproducibility) and high-frequency electromagnetic noise (manifested as dense distribution and random fluctuations), thereby performing precise adaptive weighted adjustment of the threshold shrinkage function for each layer.

[0060] Specifically, for each intrinsic mode component after variational mode decomposition and narrowband interference removal, it is sequentially input into the aforementioned adaptive wavelet filtering process. First, it is filtered through an optimized threshold... Directional filtering is performed on the wavelet coefficients of each layer to output the wavelet coefficient sequence after noise removal; then, the signal is reconstructed by inverse wavelet transform to obtain the preprocessed intrinsic mode components.

[0061] Finally, all intrinsic mode components of the same type of analytical signal within a single time window, after denoising, are superimposed and accumulated to restore the filtered reconstructed signal, which is denoted as the effective signal. Through this preprocessing system, high-fidelity reconstructed signals of current, transient ground voltage, and vibration ultrasonic waves are obtained, providing a clean and reliable input feature set for subsequent accurate detection of partial discharge based on multi-source data fusion.

[0062] S4. Statistical features of effective signals under multiple channels are extracted and time-aligned to construct a fused feature vector. The fused feature vector is then input into a pre-trained partial discharge detection model, and the partial discharge detection results of the cable branch box are output.

[0063] Based on existing open-source partial discharge noise-free and noise-contaminated datasets, current signals, voltage signals, and vibration / ultrasonic signals, along with their corresponding labels, are extracted. The labels include two categories: "normal" and "partial discharge." To ensure the consistency of the feature distribution of the model input data, the extracted current, voltage, and vibration / ultrasonic signals are first preprocessed sequentially through the aforementioned variational mode decomposition and adaptive wavelet filtering process to obtain clean reconstructed signals, which are then assigned the classification labels "normal" and "partial discharge," respectively.

[0064] During the feature extraction stage, spatiotemporal alignment of multiple sensors must be ensured. Specifically, under the same analysis time window, time-domain statistical features are extracted from the preprocessed reconstructed current signal, reconstructed voltage signal, and reconstructed ultrasonic signal, including: mean, variance, kurtosis factor, peak factor, impulse factor, margin factor, waveform factor, and skewness. If a channel fails to capture effective signal fluctuations within its corresponding analysis window, the eight feature values ​​corresponding to that channel are padded with zeros. Through the above alignment operation, the features of the three signals are concatenated in channel order to robustly construct a 24-dimensional multi-source fusion feature vector.

[0065] A training set is constructed using a large number of labeled 24-dimensional multi-source fusion feature vectors, which are then used as input to a Support Vector Machine (SVM) for training the classification model. Since partial discharge features are highly nonlinear in multidimensional space, a radial basis function (RBF) kernel is selected, with a fixed penalty coefficient C of 10 and a kernel parameter Gamma of 0.1. Through supervised training, a high-precision partial discharge detection model is finally solidified and output. Training the Support Vector Machine (SVM) is a well-known existing technique; implementers can choose other feasible machine learning models, and this embodiment does not impose any restrictions on this.

[0066] During the real-time monitoring phase, the current signal, vibration ultrasonic signal, and transient ground voltage signal of the cable branch box are acquired in real time through a multi-type sensor array, and the aforementioned denoising preprocessing operation based on VMD decomposition and adaptive wavelet filtering is performed to obtain a high-fidelity reconstructed signal set.

[0067] Subsequently, strictly following the time window alignment and zero-padding rules of the training phase, statistical features of the three reconstructed signals are extracted in real time and spliced ​​to generate a real-time 24-dimensional multi-source fusion feature vector. This feature vector is then input into the trained and solidified partial discharge detection model (SVM) for forward inference calculation, and the current state classification result of the cable branch box (labeled as normal or partial discharge) is output in real time.

[0068] When the model determines that the current state is "partial discharge", it immediately triggers the early warning feedback mechanism, thereby realizing the early detection and closed-loop monitoring of insulation aging and discharge defects in cable branch boxes, and completing the real-time detection operation of partial discharge.

[0069] Based on the same inventive concept as the above methods, this application also provides a partial discharge detection system for cable branch boxes, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above methods for partial discharge detection of cable branch boxes.

[0070] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments of this specification have been described above. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0071] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0072] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the principles of this application should be included within the protection scope of this application.

Claims

1. A method for detecting partial discharge in a cable branch box, characterized in that, The method includes the following steps: Acquire multi-channel real-time monitoring signals from the cable branch box and preprocess them to obtain analysis signals; The analysis signal is subjected to mode decomposition to obtain multiple mode components. Statistical features characterizing periodicity and energy distribution patterns are extracted from each mode component. Based on these statistical features, mode components containing narrowband interference are identified and filtered out. Wavelet decomposition is performed on the remaining modal components after filtering. Based on the amplitude distribution differences of pulse extrema points in each wavelet layer and the similarity of local pulse waveforms, the wavelet threshold of the corresponding wavelet layer is adaptively adjusted for filtering and the denoised effective signal is reconstructed. The statistical features of the effective signals under multiple channels are extracted and time-aligned to construct a fused feature vector. The fused feature vector is then input into a pre-trained partial discharge detection model to output the partial discharge detection results of the cable branch box.

2. The partial discharge detection method for a cable branch box as described in claim 1, characterized in that, The extraction of statistical features characterizing periodicity and energy distribution patterns from each modal component includes: Calculate the autocorrelation coefficients of each modal component within the preset hysteresis range, and mark the maximum autocorrelation coefficient corresponding to each modal component; Determine the kurtosis value of each modal component and the sum of all energy values ​​of each modal component; and perform linear fitting on each modal component to obtain the mean of all fitting residuals corresponding to each modal component. Based on the maximum autocorrelation coefficient, the kurtosis value, the summation and the mean, the statistical characteristics corresponding to each modal component are determined.

3. The partial discharge detection method for a cable branch box as described in claim 2, characterized in that, The statistical characteristics corresponding to each modal component are as follows: The maximum autocorrelation coefficient is negatively correlated, and the result of the negative correlation mapping, the kurtosis value, the sum and the mean are combined to form a multidimensional vector, which serves as the statistical feature corresponding to each modal component.

4. The partial discharge detection method for a cable branch box as described in claim 3, characterized in that, The identification and filtering of modal components containing narrowband interference includes: Obtain the modal components of the historical analysis signal of the cable branch box, and perform threshold segmentation on the multidimensional vectors based on the numerical distribution characteristics of the multidimensional vectors corresponding to all modal components of the historical analysis signal and the current analysis signal. The modal components corresponding to multidimensional vectors smaller than the segmentation threshold are considered as modal components containing narrowband interference, and the modal components containing narrowband interference are screened out from all modal components of the current analyzed signal.

5. The partial discharge detection method for a cable branch box as described in claim 1, characterized in that, The adaptive adjustment of the wavelet threshold corresponding to the wavelet layer includes: Identify the peak value of the detail coefficients within each wavelet layer, and determine the difference between the peak value and the average level of the detail coefficients within the wavelet layer, which is denoted as the first difference; Centered on each peak point, the local detail coefficient segments corresponding to each peak are extracted within the wavelet layer to determine the metric distance between local detail coefficient segments of different peaks within the wavelet layer; By comparing the degree of dispersion of the measured distance with the first difference, the comprehensive feature value of each wavelet layer is obtained, which is used to adjust the wavelet threshold of the corresponding wavelet layer.

6. The partial discharge detection method for a cable branch box as described in claim 5, characterized in that, The process of obtaining the comprehensive feature values ​​of each wavelet layer includes: Calculate the average value of all the first differences within the wavelet layer. The comprehensive feature value is positively correlated with the average value and negatively correlated with the degree of dispersion.

7. The partial discharge detection method for a cable branch box as described in claim 5, characterized in that, The wavelet threshold for adjusting the corresponding wavelet layer includes: The normalized comprehensive eigenvalues ​​of each wavelet layer are positively shifted and mapped to obtain the adaptive adjustment factor of each wavelet layer. The wavelet threshold of the corresponding wavelet layer is then calculated, with the specific expression as follows: In the formula, The wavelet threshold is the optimized wavelet coefficient of the j-th wavelet at the i-th level after wavelet decomposition of the k-th intrinsic mode component. The fixed wavelet threshold used after wavelet decomposition of the k-th intrinsic mode component. This is the adaptive adjustment factor for the i-th wavelet layer after wavelet decomposition of the k-th intrinsic mode component. This is the preset steepness adjustment factor.

8. The partial discharge detection method for a cable branch box as described in claim 7, characterized in that, The calculation process of the adaptive adjustment factor is as follows: Calculate the product of the normalized comprehensive eigenvalues ​​of each wavelet layer and the preset basic adjustment factor, wherein the adaptive adjustment factor is the sum of the product and the value 1.

9. A partial discharge detection method for a cable branch box as described in claim 1, characterized in that, The statistical features in the fused feature vector include the mean, variance, kurtosis factor, peak factor, impulse factor, margin factor, waveform factor, and skewness of the effective signal under multiple channels.

10. A partial discharge detection system for a cable branch box, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-9.

Images