Partial discharge real-time monitoring method based on ultrahigh frequency signal and ultrasonic signal

Abstract

The invention relates to the technical field of discharge monitoring, and discloses a partial discharge real-time monitoring method based on an ultrahigh-frequency signal and an ultrasonic signal, and the method comprises the steps: obtaining an original monitoring data set which is synchronously collected during the operation of detected power equipment, and comprises an ultrahigh-frequency original signal sequence and an ultrasonic original signal sequence; processing the original monitoring data set to obtain a standard monitoring data set which comprises an ultrahigh frequency purification signal sequence and an ultrasonic purification signal sequence; extracting waveform features of the ultrahigh-frequency purification signal sequence, constructing an ultrahigh-frequency feature vector set, extracting sound field features of the ultrasonic purification signal sequence, and constructing an ultrasonic feature vector set; generating a joint feature matrix based on the ultrahigh frequency feature vector set and the ultrasonic feature vector set; and constructing a partial discharge identification model according to the joint feature matrix to obtain a partial discharge real-time diagnosis and monitoring result of the detected power equipment, thereby ensuring the partial discharge real-time diagnosis and monitoring precision, and improving the early identification accuracy and predictive maintenance level of the equipment insulation defect.

Description

Technical Field

[0001] This invention relates to the field of discharge monitoring technology, and more specifically, to a method for real-time monitoring of partial discharge based on ultra-high frequency signals and ultrasonic signals. Background Technology

[0002] Partial discharge online monitoring refers to a technology that uses real-time sensing of the insulation condition of power equipment to identify early insulation defects and assess their deterioration trends. It is widely used in condition-based maintenance of high-voltage power equipment. Partial discharge generates a series of physical phenomena, including light and sound, and chemical changes within and around the power equipment. These physical phenomena and chemical changes accompanying partial discharge can provide detection signals for monitoring the internal insulation condition of the power equipment.

[0003] Existing monitoring technologies often rely on offline detection and single-mode signal detection. These methods have significant shortcomings and cannot meet the needs of modern power grids: Offline detection is conducted in laboratories or under power outage conditions, which differ from the voltage and load environment during actual equipment operation, limiting the accuracy of the detection data; single-mode signals are susceptible to electromagnetic interference or environmental noise, resulting in insufficient recognition accuracy. Furthermore, the characteristics of a single signal are limited, making it difficult to comprehensively characterize the nature of the discharge. Moreover, existing technologies mostly focus on static threshold discrimination, lacking the ability to continuously track and dynamically warn of the insulation state evolution process, making it difficult to promptly capture key turning points that accelerate defect degradation. Summary of the Invention

[0004] This invention provides a method for real-time monitoring of partial discharge based on ultra-high frequency signals and ultrasonic signals. It aims to solve the technical problems of low accuracy of single-mode identification, insufficient multi-mode fusion, and lack of state evolution prediction capability in the prior art. It achieves the technical effects of complementary enhancement of dual-mode signals, fine classification of insulation status, and dynamic early warning of trend evolution. Thus, while ensuring the accuracy of real-time diagnosis and monitoring of partial discharge, it significantly improves the early identification accuracy of equipment insulation defects and the level of predictive maintenance.

[0005] To achieve the above objectives, the present invention provides a method for real-time monitoring of partial discharge based on ultra-high frequency signals and ultrasonic signals, comprising: Acquire the raw monitoring dataset synchronously collected during the operation of the power equipment under test, wherein the raw monitoring dataset includes UHF raw signal sequences and ultrasonic raw signal sequences; The original monitoring dataset is subjected to signal purification processing to obtain a standard monitoring dataset, wherein the standard monitoring dataset includes an ultra-high frequency purification signal sequence and an ultrasonic purification signal sequence; Waveform features are extracted from the ultra-high frequency purification signal sequence to construct an ultra-high frequency feature vector set; sound field features are extracted from the ultrasonic purification signal sequence to construct an ultrasonic feature vector set. Multimodal information fusion is performed based on the UHF feature vector set and the ultrasonic feature vector set to generate a joint feature matrix; A partial discharge identification model is constructed based on the joint feature matrix to obtain the real-time diagnostic monitoring results of partial discharge of the tested power equipment.

[0006] Furthermore, when acquiring the raw monitoring dataset synchronously collected during the operation of the tested power equipment, it includes: By using a synchronous triggering mechanism, the UHF sensor probe and the ultrasonic sensor probe are controlled to acquire signals under the same time reference, thus obtaining the initial UHF signal segment and the initial ultrasonic signal segment. The initial UHF signal segment and the initial ultrasonic signal segment are time-stamp aligned to generate a time synchronization signal pair. All time synchronization signal pairs within the acquisition period are integrated into the original monitoring dataset.

[0007] Furthermore, when performing signal purification processing on the original monitoring dataset to obtain a standard monitoring dataset, the process includes: Each UHF original signal in the UHF original signal sequence is subjected to noise basis processing to obtain the noise power level; An adaptive filter is constructed based on the noise power level to perform narrowband interference suppression on the original UHF signal, thereby obtaining a UHF filtered signal. The waveform amplitude of the ultra-high frequency filtered signal is normalized to map the peak-to-peak value of the signal to a uniform dimension range, thereby obtaining the ultra-high frequency purified signal sequence. Environmental noise is collected for each raw ultrasonic signal in the raw ultrasonic signal sequence to obtain a background sound field sample. A sound field cancellation model is constructed based on the background sound field samples, and environmental noise cancellation processing is performed on the original ultrasonic signal to obtain an ultrasonic noise-reduced signal. The ultrasonic noise reduction signal is subjected to sound intensity calibration processing to standardize the sound pressure amplitude to the same reference dimension, thereby obtaining the ultrasonic purification signal sequence.

[0008] Furthermore, when extracting waveform features from the ultra-high frequency purification signal sequence and constructing an ultra-high frequency feature vector set, the following steps are included: Perform time-frequency transformation processing on each UHF purification signal in the UHF purification signal sequence to generate an UHF time-frequency spectrum. Spectral statistical features are extracted from the ultra-high frequency time spectrum, including the concentration of spectral energy distribution, the dominant frequency offset index, and the high-frequency component attenuation coefficient. The ultra-high frequency purification signal is subjected to pulse envelope detection to extract pulse waveform parameters, including rise edge steepness, pulse width variation coefficient and half-wave duration. The spectral statistical features and the pulse waveform parameters are integrated into a partial discharge fingerprint feature tuple; All fingerprint feature tuples corresponding to the ultra-high frequency purification signals are aggregated into the ultra-high frequency feature vector set.

[0009] Furthermore, when extracting sound field features from the ultrasonic purification signal sequence and constructing an ultrasonic feature vector set, the following steps are included: Each ultrasonic purification signal in the ultrasonic purification signal sequence is decomposed by wavelet to obtain multi-scale sound field components. Energy distribution features are extracted from the multi-scale sound field components, including the energy proportion of each scale and the energy centroid offset. The arrival time difference of the ultrasonic purification signal is calculated to obtain the sound wave propagation delay parameter; Based on the sound wave propagation delay parameters, the sound source localization is calculated to obtain the spatial coordinate location value of the discharge source. The energy distribution characteristics are combined with the spatial coordinates of the discharge source to form a sound source localization feature vector; The sound source localization feature vectors corresponding to all ultrasonic purification signals are integrated into the ultrasonic feature vector set.

[0010] Furthermore, when performing multimodal information fusion based on the UHF feature vector set and the ultrasonic feature vector set to generate a joint feature matrix, the process includes: The ultra-high frequency feature vector set and the ultrasonic feature vector set are aligned in the time dimension to generate a synchronous feature pair sequence; Calculate the mutual information value between the UHF feature vector and the ultrasound feature vector in each synchronous feature pair, and retain synchronous feature pairs with mutual information values ​​greater than the benchmark threshold parameter to obtain a subset of strongly correlated feature pairs; Principal component dimensionality reduction is performed on the ultra-high frequency feature vectors in the subset of strongly correlated feature pairs to obtain ultra-high frequency compressed feature vectors; Linear discriminant analysis is performed on the ultrasonic feature vectors in the subset of strongly correlated feature pairs to obtain ultrasonic discriminant feature vectors; The ultra-high frequency compressed feature vector and the ultrasonic discrimination feature vector are concatenated to generate a primary fusion feature; The primary fusion features are subjected to a nonlinear mapping transformation to be mapped to a high-dimensional joint feature space to obtain the joint feature matrix.

[0011] Furthermore, when calculating the mutual information value between the UHF feature vector and the ultrasound feature vector in each synchronization feature pair, and retaining synchronization feature pairs with mutual information values ​​greater than a benchmark threshold parameter to obtain a subset of strongly correlated feature pairs, the following steps are included: The mutual information estimate between the UHF feature vector and the ultrasonic feature vector is calculated using the K-nearest neighbor-based entropy estimation method. All mutual information estimates are normalized and statistical distribution analysis is performed to construct the probability density distribution curve of mutual information values. Based on the robust statistical properties of the probability density distribution curve, the benchmark threshold parameter is determined; The number of feature pairs in the candidate strongly correlated feature pair set is counted. When the number of feature pairs is less than the preset minimum number of pairs, the baseline threshold parameter is gradually reduced by a preset step size until the number of feature pairs meets the preset minimum number of pairs, thus obtaining the final strongly correlated feature pair subset.

[0012] Furthermore, when performing a nonlinear mapping transformation on the primary fusion features to map them to a high-dimensional joint feature space to obtain the joint feature matrix, the process includes: The primary fusion feature set is input into the kernel principal component analysis model to calculate the covariance matrix of the primary fusion features in the kernel feature space; Find the eigenvectors corresponding to the first M largest eigenvalues ​​of the covariance matrix, and arrange the eigenvectors in descending order to construct the joint feature matrix.

[0013] Furthermore, when constructing a partial discharge identification model based on the joint feature matrix, the following steps are included: Multiple joint feature matrices were determined based on historical monitoring databases; The samples in the joint feature matrix are divided into a training data subset and a validation data subset; The deep neural network is iteratively trained using the subset of training data until the loss function converges to a preset error range to obtain the trained model. The deep neural network includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is equal to the dimension of the joint feature matrix, and the number of nodes in the output layer is equal to the preset number of discharge type categories. The generalization ability of the trained model is evaluated based on the subset of validation data, and the model whose evaluation accuracy exceeds a preset performance threshold is retained as the partial discharge identification model.

[0014] Furthermore, when obtaining the real-time diagnostic monitoring results of partial discharge of the tested power equipment, the following are included: The newly determined joint feature matrix in the next stage is input into the partial discharge identification model, and the probability distribution sequence of each type of discharge is obtained through forward propagation calculation. The category with the highest probability value in the probability distribution sequence is selected as the result of the real-time diagnostic monitoring of partial discharge.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention discloses a method for real-time monitoring of partial discharge based on UHF and ultrasonic signals. The method acquires raw monitoring datasets synchronously collected during the operation of the tested power equipment, including UHF and ultrasonic raw signal sequences. The raw monitoring datasets are processed to obtain a standard monitoring dataset, including UHF purification signal sequences and ultrasonic purification signal sequences. Waveform features of the UHF purification signal sequences are extracted to construct a UHF feature vector set, and sound field features of the ultrasonic purification signal sequences are extracted to construct an ultrasonic feature vector set. A joint feature matrix is ​​generated based on the UHF and ultrasonic feature vector sets. A partial discharge identification model is constructed based on the joint feature matrix to obtain real-time diagnostic monitoring results of partial discharge in the tested power equipment. This method ensures the accuracy of real-time diagnostic monitoring of partial discharge and significantly improves the early identification accuracy and predictive maintenance level of equipment insulation defects. Attached Figure Description

[0016] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a real-time partial discharge monitoring method based on ultra-high frequency signals and ultrasonic signals in an embodiment of the present invention is shown. Detailed Implementation

[0017] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0018] In the description of this application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0019] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0020] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0021] The following is a description of preferred embodiments of the present invention in conjunction with the accompanying drawings.

[0022] like Figure 1 As shown, embodiments of the present invention disclose a method for real-time monitoring of partial discharge based on ultra-high frequency signals and ultrasonic signals, including: S110: Obtain the raw monitoring dataset synchronously collected during the operation of the power equipment under test, wherein the raw monitoring dataset includes UHF raw signal sequence and ultrasonic raw signal sequence; S120: Perform signal purification processing on the original monitoring dataset to obtain a standard monitoring dataset, wherein the standard monitoring dataset includes an ultra-high frequency purification signal sequence and an ultrasonic purification signal sequence; S130: Extract waveform features from the ultra-high frequency purification signal sequence to construct an ultra-high frequency feature vector set; extract sound field features from the ultrasonic purification signal sequence to construct an ultrasonic feature vector set. S140: Multimodal information fusion is performed based on the UHF feature vector set and the ultrasonic feature vector set to generate a joint feature matrix; S150: Construct a partial discharge identification model based on the joint feature matrix to obtain the real-time diagnostic monitoring results of partial discharge of the tested power equipment.

[0023] In some embodiments of this application, the process of acquiring the raw monitoring dataset synchronously collected during the operation of the power equipment under test includes: By using a synchronous triggering mechanism, the UHF sensor probe and the ultrasonic sensor probe are controlled to acquire signals under the same time reference, thus obtaining the initial UHF signal segment and the initial ultrasonic signal segment. The initial UHF signal segment and the initial ultrasonic signal segment are time-stamp aligned to generate a time synchronization signal pair. All time synchronization signal pairs within the acquisition period are integrated into the original monitoring dataset.

[0024] In this embodiment, a UHF sensor probe and an ultrasonic sensor probe are deployed at the coaxial detection location of the power equipment under test. The UHF sensor probe operates in the frequency band of 300MHz to 3000MHz, and the ultrasonic sensor probe has a frequency response range of 20kHz to 200kHz. The coaxial detection location refers to a detection point located no more than 10cm apart on both sides of the same basin insulator of GIS equipment or on the same tank wall of a transformer.

[0025] In this embodiment, the synchronization triggering mechanism employs GPS clock synchronization or fiber optic triggering synchronization, with a time synchronization accuracy better than 10 ns. The initial UHF signal segment length is 10,000 sampling points, and the initial ultrasonic signal segment length is 5,000 sampling points. Timestamp alignment processing uses a cross-correlation algorithm to find the optimal time delay, with an alignment error not exceeding one sampling period. The time synchronization signal pair consists of a pair of UHF and ultrasonic signals with the same time base label. The acquisition period is set to 1 hour, with 10 sets of time synchronization signal pairs acquired in each period. The integrated original monitoring dataset contains 10 pairs of signals.

[0026] The beneficial effects of the above technical solution are: through coaxial deployment, synchronous acquisition and precise alignment mechanism, the strict correspondence between the two-modal signals in the spatiotemporal dimension is ensured, providing a high-quality data foundation for subsequent feature fusion.

[0027] In some embodiments of this application, when performing signal purification processing on the original monitoring dataset to obtain a standard monitoring dataset, the following steps are included: Each UHF original signal in the UHF original signal sequence is subjected to noise basis processing to obtain the noise power level; An adaptive filter is constructed based on the noise power level to perform narrowband interference suppression on the original UHF signal, thereby obtaining a UHF filtered signal. The waveform amplitude of the ultra-high frequency filtered signal is normalized to map the peak-to-peak value of the signal to a uniform dimension range, thereby obtaining the ultra-high frequency purified signal sequence. Environmental noise is collected for each raw ultrasonic signal in the raw ultrasonic signal sequence to obtain a background sound field sample. A sound field cancellation model is constructed based on the background sound field samples, and environmental noise cancellation processing is performed on the original ultrasonic signal to obtain an ultrasonic noise-reduced signal. The ultrasonic noise reduction signal is subjected to sound intensity calibration processing to standardize the sound pressure amplitude to the same reference dimension, thereby obtaining the ultrasonic purification signal sequence.

[0028] In this embodiment, the noise floor processing employs power spectral density analysis to statistically analyze the average power within the frequency band during non-discharge periods. The noise power level is expressed in dBm. The adaptive filter uses a notch filter based on the LMS algorithm. The notch bandwidth is dynamically adjusted according to the interference signal bandwidth. After narrowband interference suppression, the signal-to-interference ratio is improved by no less than 15dB, resulting in an ultra-high frequency filtered signal.

[0029] In this embodiment, the waveform amplitude normalization process adopts the maximum-minimum normalization method, which maps the peak-to-peak value of the signal to the normalized dimension range of [-1, 1] to eliminate the sensitivity differences of different sensors.

[0030] In this embodiment, environmental noise is collected during the off-duty period, with a background sound field sample duration of 60 seconds and 120,000 sampling points. The sound field cancellation model employs an adaptive finite impulse response filter structure, with a filter order of 128 and a total of 129 weighting coefficients (including one bias term). The samples are divided into 120 segments, with 1,000 sampling points in each segment serving as the training sample set.

[0031] The specific implementation steps for constructing the sound field cancellation model are as follows: Initialization phase: Initialize all weight coefficients of the filter to zero, set the initial value of the learning step size parameter to 0.01, and set the convergence tolerance value to 1 multiplied by 10. -6 The maximum number of iterations is set to 5000.

[0032] Iterative training phase: Perform the following operations on each background sound field sample: The sample segment is used as the desired output signal; the synchronously acquired raw ultrasonic signal is used as the filter input signal sequence; the filter output value is calculated, which is equal to the inner product of the current weight coefficient vector and the input signal vector, where the input signal vector is composed of the 129 sampling points before the current time; the error signal is calculated, which is equal to the desired output signal minus the filter output value; the weight coefficients are updated, and the new weight coefficients are equal to the original weight coefficients plus the product of the learning step size parameter, the error signal, and the input signal vector; the root mean square error of the error signal is calculated, and convergence is determined when the decrease in root mean square error is less than the convergence tolerance value in 10 consecutive iterations.

[0033] Parameter solidification stage: When the filter has been trained on all 120 samples and the root mean square error value is stable below 0.05, the final weight coefficients are solidified as the core parameters of the sound field cancellation model, and the sound field cancellation model is obtained.

[0034] In this embodiment, the sound intensity calibration process adopts the standard sound source calibration method, which standardizes the sound pressure amplitude to the decibel value under the reference dimension of 20μPa, and controls the calibration error within ±1.5dB, finally obtaining a standardized ultrasonic purification signal sequence.

[0035] The beneficial effects of the above technical solution are: through purification processes such as adaptive filtering, amplitude normalization and sound field cancellation, the influence of environmental interference and sensor differences is effectively eliminated, ensuring the dimensional uniformity and comparability of the dual-mode signals.

[0036] In some embodiments of this application, the process of extracting waveform features from the UHF purification signal sequence and constructing a UHF feature vector set includes: Perform time-frequency transformation processing on each UHF purification signal in the UHF purification signal sequence to generate an UHF time-frequency spectrum. Spectral statistical features are extracted from the ultra-high frequency time spectrum, including the concentration of spectral energy distribution, the dominant frequency offset index, and the high-frequency component attenuation coefficient. The ultra-high frequency purification signal is subjected to pulse envelope detection to extract pulse waveform parameters, including rise edge steepness, pulse width variation coefficient and half-wave duration. The spectral statistical features and the pulse waveform parameters are integrated into a partial discharge fingerprint feature tuple; All fingerprint feature tuples corresponding to the ultra-high frequency purification signals are aggregated into the ultra-high frequency feature vector set.

[0037] In this embodiment, the time-frequency transformation process employs a short-time Fourier transform, with a Hanning window selected as the window function. The window length is 512 sampling points, and the overlap rate is 50%. The concentration of spectral energy in the center frequency band is calculated as the spectral energy distribution concentration, with a value ranging from 0 to 1. The dominant frequency offset index reflects the offset of the discharge dominant frequency relative to the standard frequency, expressed as a percentage. The high-frequency component attenuation coefficient describes the attenuation slope of frequency components above 1000MHz, with units of dB / GHz.

[0038] In this embodiment, the pulse envelope detection uses the Hilbert transform method to extract the signal envelope. The rising edge steepness is defined as the reciprocal of the time it takes for the envelope to rise from 10% to 90% of its peak value. The pulse width variation coefficient is the ratio of the standard deviation to the mean of the pulse width. The half-wave duration is the time length corresponding to half the pulse width.

[0039] In this embodiment, the concentration of spectral energy distribution, the main frequency offset index, and the high-frequency component attenuation coefficient are arranged in a fixed order to form the front-end feature sequence; the rising edge steepness, the pulse width variation coefficient, and the half-wave duration are added to the back-end position in sequence to form a partial discharge fingerprint feature tuple consisting of six feature parameters arranged in sequence.

[0040] In this embodiment, the above 6 features are extracted from each UHF purification signal to form a fingerprint feature tuple, and the 10 groups of signals are aggregated to form a 10×6 UHF feature vector set.

[0041] The beneficial effects of the above technical solution are: by combining time-frequency domain feature extraction and pulse waveform parameter analysis, a fingerprint feature system that can comprehensively characterize the electromagnetic wave characteristics of partial discharge is constructed, providing a multi-dimensional basis for accurate identification of discharge types.

[0042] In some embodiments of this application, the process of extracting sound field features from the ultrasonic purification signal sequence and constructing an ultrasonic feature vector set includes: Each ultrasonic purification signal in the ultrasonic purification signal sequence is decomposed by wavelet to obtain multi-scale sound field components. Energy distribution features are extracted from the multi-scale sound field components, including the energy proportion of each scale and the energy centroid offset. The arrival time difference of the ultrasonic purification signal is calculated to obtain the sound wave propagation delay parameter; Based on the sound wave propagation delay parameters, the sound source localization is calculated to obtain the spatial coordinate location value of the discharge source. The energy distribution characteristics are combined with the spatial coordinates of the discharge source to form a sound source localization feature vector; The sound source localization feature vectors corresponding to all ultrasonic purification signals are integrated into the ultrasonic feature vector set.

[0043] In this embodiment, the wavelet decomposition uses the db4 wavelet basis function and the number of decomposition layers is set to 5, resulting in a total of 6 multi-scale sound field components, namely the approximate component A5 and the detail components D1-D5.

[0044] In this embodiment, the energy proportion at each scale is calculated as the percentage of the sum of squares of wavelet coefficients in the total energy, and the energy centroid offset reflects the distribution center position of energy at different scales, with a value range of 1-5.

[0045] In this embodiment, the time difference of arrival is calculated using the peak detection method of the cross-correlation function, achieving a detection accuracy of 8 ns. The acoustic wave propagation delay parameters include the delay value and the delay variance, with units of μs.

[0046] In this embodiment, the sound source localization calculation adopts the time-delay-based spherical intersection algorithm. Assuming the sound speed is 340m / s, the estimated three-dimensional spatial coordinates (x, y, z) of the discharge source inside the device are calculated with a coordinate accuracy of ±5cm.

[0047] In this embodiment, seven energy distribution characteristic parameters can be obtained, namely, the energy proportion at six scales and the energy centroid offset, which are arranged in a fixed order to form the first-stage characteristic sequence. The estimated three-dimensional spatial coordinate values, including the horizontal, vertical, and triangular coordinate values, are sequentially added to the second-stage position to form a feature vector composed of ten characteristic parameters arranged in sequence, thus obtaining the sound source localization feature vector. The 10 sets of ultrasonic purification signals are integrated to form a 10×10 ultrasonic feature vector set.

[0048] The beneficial effects of the above technical solution are: through wavelet multi-scale analysis and sound source localization calculation, the spatial location of the discharge source is accurately estimated, providing key acoustic characteristic parameters for fault location and visualization.

[0049] In some embodiments of this application, when performing multimodal information fusion based on the UHF feature vector set and the ultrasonic feature vector set to generate a joint feature matrix, the following steps are included: The ultra-high frequency feature vector set and the ultrasonic feature vector set are aligned in the time dimension to generate a synchronous feature pair sequence; Calculate the mutual information value between the UHF feature vector and the ultrasound feature vector in each synchronous feature pair, and retain synchronous feature pairs with mutual information values ​​greater than the benchmark threshold parameter to obtain a subset of strongly correlated feature pairs; Principal component dimensionality reduction is performed on the ultra-high frequency feature vectors in the subset of strongly correlated feature pairs to obtain ultra-high frequency compressed feature vectors; Linear discriminant analysis is performed on the ultrasonic feature vectors in the subset of strongly correlated feature pairs to obtain ultrasonic discriminant feature vectors; The ultra-high frequency compressed feature vector and the ultrasonic discrimination feature vector are concatenated to generate a primary fusion feature; The primary fusion features are subjected to a nonlinear mapping transformation to be mapped to a high-dimensional joint feature space to obtain the joint feature matrix.

[0050] In this embodiment, the time dimension alignment adopts the nearest neighbor matching method. Based on the acquisition time of the UHF signal, the feature vectors of the ultrasonic signal with a time difference of less than 100ms are selected to form a synchronization feature pair, and a synchronization feature pair sequence with a length of 8 is generated.

[0051] In this embodiment, principal component dimensionality reduction retains the principal components with a cumulative contribution rate of 90%, compressing the 6-dimensional UHF eigenvector into a 4-dimensional UHF compressed eigenvector. Linear discriminant analysis extracts two optimal discriminant components, reducing the 10-dimensional ultrasonic eigenvector to a 2-dimensional ultrasonic discriminant eigenvector.

[0052] In this embodiment, the feature cascading adopts a serial connection method to form a 6-dimensional primary fusion feature vector. Specifically, the ultra-high frequency compressed feature vector obtained in the previous steps is connected end-to-end with the ultrasonic discrimination feature vector in a fixed order to form a new feature vector.

[0053] The beneficial effects of the above technical solution are: through the cascade fusion strategy of mutual information filtering, differentiated dimensionality reduction and nonlinear mapping, the correlation and complementarity of heterogeneous features are effectively improved, and the optimal expression of bimodal features in a unified space is achieved.

[0054] In some embodiments of this application, when calculating the mutual information value between the UHF feature vector and the ultrasonic feature vector in each synchronization feature pair, and retaining synchronization feature pairs with mutual information values ​​greater than a benchmark threshold parameter to obtain a subset of strongly correlated feature pairs, the following steps are included: The mutual information estimate between the UHF feature vector and the ultrasonic feature vector is calculated using the K-nearest neighbor-based entropy estimation method. All mutual information estimates are normalized and statistical distribution analysis is performed to construct the probability density distribution curve of mutual information values. Based on the robust statistical properties of the probability density distribution curve, the benchmark threshold parameter is determined; The number of feature pairs in the candidate strongly correlated feature pair set is counted. When the number of feature pairs is less than the preset minimum number of pairs, the baseline threshold parameter is gradually reduced by a preset step size until the number of feature pairs meets the preset minimum number of pairs, thus obtaining the final strongly correlated feature pair subset.

[0055] In this embodiment, the nearest neighbor parameter K is dynamically adjusted according to the number of samples, ranging from 3 to 5. For example, when the number of samples is 8, K is set to 3. The mutual information estimate is the Euclidean distance between feature vectors.

[0056] In this embodiment, the normalization process uses max-min normalization, which maps to the dimensionless interval [0, 1].

[0057] In this embodiment, the probability density distribution curve is fitted using a Gaussian kernel function.

[0058] In this embodiment, the baseline threshold parameter is 1.5 times the difference between the quartile and the median of the probability density distribution curve. For example, if the upper quartile is 0.62 and the median is 0.45, the difference is 0.17, and 1.5 times this difference yields a baseline threshold parameter of 0.255.

[0059] In this embodiment, the preset step size is 0.1, and the preset minimum number of pairs is set to 6. Due to insufficient quantity, the baseline threshold parameter is gradually reduced to 0.155 with a step size of 0.1.

[0060] The beneficial effects of the above technical solution are: through adaptive K-nearest neighbor mutual information calculation and dynamic threshold adjustment mechanism, highly correlated synchronous feature pairs can be effectively screened out, redundant information can be eliminated, and the data quality and robustness of multimodal fusion can be improved.

[0061] In some embodiments of this application, when performing a nonlinear mapping transformation on the primary fusion features to map them to a high-dimensional joint feature space to obtain the joint feature matrix, the following steps are included: The primary fusion feature set is input into the kernel principal component analysis model to calculate the covariance matrix of the primary fusion features in the kernel feature space; Find the eigenvectors corresponding to the first M largest eigenvalues ​​of the covariance matrix, and arrange the eigenvectors in descending order to construct the joint feature matrix.

[0062] In this embodiment, the primary fusion feature set is fed into the kernel principal component analysis model in batches as input data. The kernel principal component analysis model uses the radial basis function kernel function as the nonlinear mapping function, and the kernel bandwidth parameter is automatically determined to be 0.38 based on the historical sample distribution. The kernel function values ​​between each pair of input samples are calculated to construct an 8x8 kernel matrix, and the kernel matrix is ​​centered to obtain the centered kernel matrix as the covariance matrix in the kernel feature space.

[0063] In this embodiment, for example, the covariance matrix yields 8 eigenvalues, arranged in descending order as [2.85, 1.92, 1.45, 0.78, 0.65, 0.41, 0.23, 0.11]. M is set to 3, and the first 3 largest eigenvalues ​​are selected, corresponding to the first three eigenvectors. Based on the three retained eigenvectors, the final joint feature matrix is ​​constructed. The joint feature matrix is ​​a two-dimensional feature matrix with dimensions of 8 rows and 2 columns.

[0064] The beneficial effects of the above technical solution are: by analyzing the covariance matrix of the kernel feature space, it is ensured that the joint feature matrix retains the maximum discriminative information and that each dimension is strictly consistent, which effectively improves the input quality and classification performance of the subsequent recognition model, lays the foundation for the partial discharge diagnosis of the power equipment under test, and provides reliable data support.

[0065] In some embodiments of this application, constructing a partial discharge identification model based on the joint feature matrix includes: Multiple joint feature matrices were determined based on historical monitoring databases; The samples in the joint feature matrix are divided into a training data subset and a validation data subset; The deep neural network is iteratively trained using the subset of training data until the loss function converges to a preset error range to obtain the trained model. The deep neural network includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is equal to the dimension of the joint feature matrix, and the number of nodes in the output layer is equal to the preset number of discharge type categories. The generalization ability of the trained model is evaluated based on the subset of validation data, and the model whose evaluation accuracy exceeds a preset performance threshold is retained as the partial discharge identification model.

[0066] In some embodiments of this application, obtaining the real-time diagnostic monitoring results of partial discharge of the tested power equipment includes: The newly determined joint feature matrix in the next stage is input into the partial discharge identification model, and the probability distribution sequence of each type of discharge is obtained through forward propagation calculation. The category with the highest probability value in the probability distribution sequence is selected as the result of the real-time diagnostic monitoring of partial discharge.

[0067] In this embodiment, 1200 sets of sample data of known discharge types are extracted from the historical monitoring database. Each set of samples contains a complete pair of UHF and ultrasonic signals, generating a corresponding joint feature matrix. The samples of this joint feature matrix are randomly divided into a training data subset and a validation data subset at a ratio of 8:2. The training data subset contains 960 sets of samples, and the validation data subset contains 240 sets of samples. Each set of samples corresponds to a joint feature matrix.

[0068] In this embodiment, the deep neural network adopts a three-layer architecture: a 2-node input layer, a 10-node hidden layer, and a 6-node output layer. The number of input layer nodes (2) corresponds to the dimension of the joint feature matrix; the number of hidden layer nodes (10) is determined by a grid search method, traversing the training with a stride of 2 within a range of 6 to 14 nodes, and selecting the 10 nodes with the highest verification accuracy as the final structure; the number of output layer nodes (6) corresponds to the number of six preset discharge types: tip discharge, suspended discharge, surface discharge, air gap discharge, particle discharge, and insulation defect. The hidden layer activation function is a linear rectified function, and the output layer activation function is a softmax function.

[0069] In this embodiment, the iterative training process employs batch gradient descent with a batch size of 32 and an initial learning rate of 0.001. After every 50 batch training iterations, if the validation loss does not decrease for three consecutive iterations, the learning rate is halved, with a minimum learning rate of 0.00001. The loss function used is the cross-entropy loss function, with a preset error range set to a loss value less than 0.001 or a training epoch of 300. An early stopping mechanism is employed during training; if the validation set accuracy improves by less than 0.1% over 30 consecutive training epochs, training is terminated early, resulting in the trained model.

[0070] In this embodiment, the specific process of evaluating the generalization ability of the trained model based on the validation data subset is as follows: 240 sets of joint feature matrix samples from the validation data subset are sequentially input into the trained model. The predicted discharge type for each set of samples is calculated through forward propagation and compared with the actual discharge type label. The number of correctly predicted samples is counted. For example, if the validation accuracy is 92.5%, and the preset performance threshold is set to 90%, since the validation accuracy exceeds the preset performance threshold, all weight parameters and network structure of the currently trained model are retained, and a partial discharge recognition model is solidified. If the validation accuracy does not reach the threshold, the learning rate parameter is adjusted and retraining is performed until the performance requirements are met.

[0071] In this embodiment, the new joint feature matrix constructed in the next stage refers to the new joint feature matrix constructed from the monitoring data collected in real time from the power equipment under test during the online monitoring stage. The new joint feature matrix is ​​input into the partial discharge identification model, and through forward propagation calculation of the input layer, hidden layer, and output layer, the six nodes of the output layer generate probability values ​​corresponding to the six discharge types, for example, the probability distribution sequence is [0.03, 0.78, 0.02, 0.04, 0.01, 0.03].

[0072] In this embodiment, the above probability distribution sequence is numerically compared, and the maximum probability value is identified as 0.78, which corresponds to the second category mentioned above, namely, suspension discharge. Therefore, the discharge type of the current monitoring result is determined to be suspension discharge.

[0073] The beneficial effects of the above technical solution are as follows: by dividing the training and validation subsets by historical data, the sufficiency of model training and the objectivity of evaluation are ensured; by using deep neural networks to automatically learn the complex mapping relationship between features and discharge types, the subjectivity of manual rule setting is avoided; and by combining UHF signals and ultrasonic signals to obtain a joint feature matrix, the accuracy of real-time diagnosis and monitoring of partial discharge is further guaranteed.

[0074] In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0075] Although the invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the embodiments disclosed in this invention can be combined with each other in any way. The fact that not all of these combinations are described in this specification is merely for the sake of brevity and resource conservation.

[0076] It will be understood by those skilled in the art that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for real-time monitoring of partial discharge based on ultra-high frequency signals and ultrasonic signals, characterized in that, include: Acquire the raw monitoring dataset synchronously collected during the operation of the power equipment under test, wherein the raw monitoring dataset includes UHF raw signal sequences and ultrasonic raw signal sequences; The original monitoring dataset is subjected to signal purification processing to obtain a standard monitoring dataset, wherein the standard monitoring dataset includes an ultra-high frequency purification signal sequence and an ultrasonic purification signal sequence; Waveform features are extracted from the ultra-high frequency purification signal sequence to construct an ultra-high frequency feature vector set; sound field features are extracted from the ultrasonic purification signal sequence to construct an ultrasonic feature vector set. Multimodal information fusion is performed based on the UHF feature vector set and the ultrasonic feature vector set to generate a joint feature matrix; A partial discharge identification model is constructed based on the joint feature matrix to obtain the real-time diagnostic monitoring results of partial discharge of the tested power equipment.

2. The method for real-time monitoring of partial discharge based on ultra-high frequency signals and ultrasonic signals according to claim 1, characterized in that, When acquiring the raw monitoring dataset synchronously collected during the operation of the power equipment under test, the following should be included: By using a synchronous triggering mechanism, the UHF sensor probe and the ultrasonic sensor probe are controlled to acquire signals under the same time reference, thus obtaining the initial UHF signal segment and the initial ultrasonic signal segment. The initial UHF signal segment and the initial ultrasonic signal segment are time-stamp aligned to generate a time synchronization signal pair. All time synchronization signal pairs within the acquisition period are integrated into the original monitoring dataset.

3. The method for real-time monitoring of partial discharge based on ultra-high frequency signals and ultrasonic signals according to claim 1, characterized in that, When performing signal purification processing on the original monitoring dataset to obtain a standard monitoring dataset, the process includes: Each UHF original signal in the UHF original signal sequence is subjected to noise basis processing to obtain the noise power level; An adaptive filter is constructed based on the noise power level to perform narrowband interference suppression on the original UHF signal, thereby obtaining a UHF filtered signal. The waveform amplitude of the ultra-high frequency filtered signal is normalized to map the peak-to-peak value of the signal to a uniform dimension range, thereby obtaining the ultra-high frequency purified signal sequence. Environmental noise is collected for each raw ultrasonic signal in the raw ultrasonic signal sequence to obtain a background sound field sample. A sound field cancellation model is constructed based on the background sound field samples, and environmental noise cancellation processing is performed on the original ultrasonic signal to obtain an ultrasonic noise-reduced signal. The ultrasonic noise reduction signal is subjected to sound intensity calibration processing to standardize the sound pressure amplitude to the same reference dimension, thereby obtaining the ultrasonic purification signal sequence.

4. The method for real-time monitoring of partial discharge based on ultra-high frequency signals and ultrasonic signals according to claim 1, characterized in that, When extracting waveform features from the ultra-high frequency purification signal sequence and constructing an ultra-high frequency feature vector set, the following steps are included: Perform time-frequency transformation processing on each UHF purification signal in the UHF purification signal sequence to generate an UHF time-frequency spectrum. Spectral statistical features are extracted from the ultra-high frequency time spectrum, including the concentration of spectral energy distribution, the dominant frequency offset index, and the high-frequency component attenuation coefficient. The ultra-high frequency purification signal is subjected to pulse envelope detection to extract pulse waveform parameters, including rise edge steepness, pulse width variation coefficient and half-wave duration. The spectral statistical features and the pulse waveform parameters are integrated into a partial discharge fingerprint feature tuple; All fingerprint feature tuples corresponding to the ultra-high frequency purification signals are aggregated into the ultra-high frequency feature vector set.

5. The method for real-time monitoring of partial discharge based on ultra-high frequency signals and ultrasonic signals according to claim 1, characterized in that, When extracting sound field features from the ultrasonic purification signal sequence and constructing an ultrasonic feature vector set, the following steps are included: Each ultrasonic purification signal in the ultrasonic purification signal sequence is decomposed by wavelet to obtain multi-scale sound field components. Energy distribution features are extracted from the multi-scale sound field components, including the energy proportion of each scale and the energy centroid offset. The arrival time difference of the ultrasonic purification signal is calculated to obtain the sound wave propagation delay parameter; Based on the sound wave propagation delay parameters, the sound source localization is calculated to obtain the spatial coordinate location value of the discharge source. The energy distribution characteristics are combined with the spatial coordinates of the discharge source to form a sound source localization feature vector; The sound source localization feature vectors corresponding to all ultrasonic purification signals are integrated into the ultrasonic feature vector set.

6. The method for real-time monitoring of partial discharge based on ultra-high frequency signals and ultrasonic signals according to claim 1, characterized in that, When performing multimodal information fusion based on the UHF feature vector set and the ultrasonic feature vector set to generate a joint feature matrix, the process includes: The ultra-high frequency feature vector set and the ultrasonic feature vector set are aligned in the time dimension to generate a synchronous feature pair sequence; Calculate the mutual information value between the UHF feature vector and the ultrasound feature vector in each synchronous feature pair, and retain synchronous feature pairs with mutual information values ​​greater than the benchmark threshold parameter to obtain a subset of strongly correlated feature pairs; Principal component dimensionality reduction is performed on the ultra-high frequency feature vectors in the subset of strongly correlated feature pairs to obtain ultra-high frequency compressed feature vectors; Linear discriminant analysis is performed on the ultrasonic feature vectors in the subset of strongly correlated feature pairs to obtain ultrasonic discriminant feature vectors; The ultra-high frequency compressed feature vector and the ultrasonic discrimination feature vector are concatenated to generate a primary fusion feature; The primary fusion features are subjected to a nonlinear mapping transformation to be mapped to a high-dimensional joint feature space to obtain the joint feature matrix.

7. The method for real-time monitoring of partial discharge based on ultra-high frequency signals and ultrasonic signals according to claim 6, characterized in that, When calculating the mutual information value between the UHF feature vector and the ultrasound feature vector in each synchronization feature pair, and retaining synchronization feature pairs with mutual information values ​​greater than a benchmark threshold parameter to obtain a subset of strongly correlated feature pairs, the following are included: The mutual information estimate between the UHF feature vector and the ultrasonic feature vector is calculated using the K-nearest neighbor-based entropy estimation method. All mutual information estimates are normalized and statistical distribution analysis is performed to construct the probability density distribution curve of mutual information values. Based on the robust statistical properties of the probability density distribution curve, the benchmark threshold parameter is determined; The number of feature pairs in the candidate strongly correlated feature pair set is counted. When the number of feature pairs is less than the preset minimum number of pairs, the baseline threshold parameter is gradually reduced by a preset step size until the number of feature pairs meets the preset minimum number of pairs, thus obtaining the final strongly correlated feature pair subset.

8. The method for real-time monitoring of partial discharge based on ultra-high frequency signals and ultrasonic signals according to claim 6, characterized in that, When performing a nonlinear mapping transformation on the primary fusion features to map them to a high-dimensional joint feature space to obtain the joint feature matrix, the process includes: The primary fusion feature set is input into the kernel principal component analysis model to calculate the covariance matrix of the primary fusion features in the kernel feature space; Find the eigenvectors corresponding to the first M largest eigenvalues ​​of the covariance matrix, and arrange the eigenvectors in descending order to construct the joint feature matrix.

9. The method for real-time monitoring of partial discharge based on ultra-high frequency signals and ultrasonic signals according to claim 1, characterized in that, When constructing a partial discharge identification model based on the joint feature matrix, the following steps are included: Multiple joint feature matrices were determined based on historical monitoring databases; The samples in the joint feature matrix are divided into a training data subset and a validation data subset; The deep neural network is iteratively trained using the subset of training data until the loss function converges to a preset error range to obtain the trained model. The deep neural network includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is equal to the dimension of the joint feature matrix, and the number of nodes in the output layer is equal to the preset number of discharge type categories. The generalization ability of the trained model is evaluated based on the subset of validation data, and the model whose evaluation accuracy exceeds a preset performance threshold is retained as the partial discharge identification model.

10. The method for real-time monitoring of partial discharge based on ultra-high frequency signals and ultrasonic signals according to claim 1, characterized in that, When obtaining the real-time diagnostic monitoring results of partial discharge of the tested power equipment, the following are included: The newly determined joint feature matrix in the next stage is input into the partial discharge identification model, and the probability distribution sequence of each type of discharge is obtained through forward propagation calculation. The category with the highest probability value in the probability distribution sequence is selected as the result of the real-time diagnostic monitoring of partial discharge.

Images