Partial discharge mode identification method based on ultrahigh frequency signal phase independent characteristics

Abstract

The invention discloses a partial discharge mode identification method based on ultrahigh frequency signal phase independent characteristics. The method comprises the following steps: generating a phase resolution partial discharge map; obtaining a pre-processing phase resolution partial discharge map with a uniform size; forming a phase-independent composite feature vector; obtaining a low-redundancy phase-independent discrimination feature vector; establishing a corresponding discriminant energy potential function set after the discriminant diffusion matrix is stabilized, and forming a discriminant energy potential function model; and determining a partial discharge mode to which the partial discharge sample to be detected belongs according to a minimum discriminant energy criterion, and outputting a partial discharge mode identification result. According to the method, the interference of noise features and weak discrimination features on the model can be remarkably reduced through adaptive dimensionality reduction, and the stability and robustness of the discrimination model under small sample and short-time monitoring conditions are improved.

Description

Technical Field

[0001] This invention relates to the field of partial discharge technology, and in particular to a partial discharge pattern recognition method based on the phase-independent characteristics of ultra-high frequency signals. Background Technology

[0002] Partial discharge, as a significant precursor to insulation defects within GIS (Gas Insulation System), plays a crucial role in preventing electrical accidents and ensuring the safe and stable operation of the power grid through early identification. Ultra-high frequency (UHF) detection technology, due to its strong anti-interference capability, high sensitivity, and suitability for online monitoring, has become one of the mainstream technologies for partial discharge monitoring in GIS. Traditional UHF discharge identification methods typically rely on the "discharge amplitude-power frequency phase" distribution characteristics reflected in PRPD (Power Line Discharge Phase Diagram) spectra, identifying discharge patterns by analyzing the statistical laws governing discharge pulses within different phase intervals.

[0003] However, in actual engineering sites, due to the complex operating environment of GIS equipment, the power frequency synchronization signal in the acquisition link is easily interfered with, voltage transformers have inherent phase errors, and the clock synchronization accuracy between detection devices is limited. As a result, the corresponding power frequency phase information in the PRPD spectrum often has unavoidable drift or even loss. On the other hand, on-site partial discharge often exhibits characteristics such as "infrequent, random and sparse", and the number of pulses is limited in a short period of time, making it difficult to form stable and complete statistical spectrum features. This seriously reduces the accuracy of identification based on phase correlation features, and the robustness and universality of the model are significantly restricted.

[0004] Existing technologies still have the following prominent problems: they are highly dependent on power frequency phase. When the phase synchronization error reaches the millisecond level, the characteristic distribution of the PRPD spectrum will be significantly distorted, and the characteristic boundaries between different types of partial discharges will become blurred, making it easy to misjudge. They are not adaptable to sporadic discharge pulses. Since traditional PRPD relies on long-term statistics, when the discharge exhibits low repetition rate characteristics, short-term PRPD is difficult to form a distinguishable discharge pattern. There are abundant interference sources under complex operating conditions. In substations, there are various electromagnetic interference sources such as communication equipment, power electronic devices, and radio transmitters. The pulse signals generated by these sources in the UHF band may overlap with the actual partial discharge signals in terms of amplitude-phase distribution, causing false triggering and misidentification.

[0005] As GIS equipment develops towards higher voltage levels and larger capacities, its internal structure becomes more complex. The propagation path and spectral structure of electromagnetic radiation generated by partial discharge exhibit stronger nonlinearity and diversity. Traditional identification methods that rely on single phase information can no longer meet the actual needs of high-reliability operation. Summary of the Invention

[0006] One objective of this invention is to propose a partial discharge pattern recognition method based on phase-independent features of ultra-high frequency signals. The adaptive dimensionality reduction of this invention can significantly reduce the interference of noise features and weak discriminative features on the model, and improve the stability and robustness of the discriminative model under small sample and short-time monitoring conditions.

[0007] A partial discharge pattern recognition method based on phase-independent characteristics of ultra-high frequency signals according to an embodiment of the present invention includes: The UHF receiver is used to collect ultra-high frequency time-domain pulse signals, and the ultra-high frequency time-domain pulse signals are mapped into two-dimensional coordinate data of amplitude and power frequency phase to generate phase-resolved partial discharge maps. Preprocessing is performed on the phase-resolved partial discharge spectrum, and resampling is performed according to a uniform resolution to obtain a preprocessed phase-resolved partial discharge spectrum of uniform size; Based on the preprocessed phase-resolved partial discharge spectrum, amplitude-phase statistical feature vectors, pulse number-phase statistical feature vectors, pulse number-amplitude statistical feature vectors, energy-phase statistical feature vectors, and frequency domain correlation feature vectors are extracted and spliced ​​in a preset order to form a phase-independent composite feature vector. Using a training dataset containing various labeled partial discharge samples, an adaptive feature dimensionality reduction process based on the contribution of class separability is performed on the phase-independent composite feature vector to obtain a low-redundancy phase-independent discriminative feature vector. Based on the low-redundancy phase-independent discrimination feature vector, the discrimination prototype center and discrimination diffusion matrix of different partial discharge modes are calculated respectively. After the discrimination diffusion matrix is ​​stabilized, the corresponding discrimination energy potential function set is established to form the discrimination energy potential function model. The low-redundancy phase-independent discrimination feature vector of the partial discharge sample to be tested is input into the discrimination energy potential function model. The partial discharge mode to which the partial discharge sample to be tested belongs is determined according to the minimum discrimination energy criterion, and the partial discharge mode identification result is output.

[0008] Optionally, mapping the ultra-high frequency time-domain pulse signal into amplitude-power frequency phase two-dimensional coordinate data includes: A UHF receiver using continuous sampling continuously acquires electromagnetic radiation signals generated by partial discharge inside a gas-insulated metal-enclosed switchgear within a preset acquisition time period to obtain ultra-high frequency time-domain pulse signals. For ultra-high frequency time-domain pulse signals, discrete sampling is performed at equal intervals at a set sampling frequency to obtain a set of ultra-high frequency discrete sampling sequences; In the ultra-high frequency discrete sampling sequence, all discharge pulse events are detected to form a set of discharge pulse events; Synchronously acquire the power frequency reference signal aligned with the UHF time-domain pulse signal on the time axis, and calculate the power frequency phase corresponding to each discharge pulse event based on the arrival time of each discharge pulse event and the phase information of the power frequency reference signal; All discharge pulse events are mapped according to their pulse amplitude and power frequency phase to obtain a set of two-dimensional coordinates of amplitude and power frequency phase. Each pair of two-dimensional coordinates of amplitude and power frequency phase corresponds to the amplitude of a discharge pulse event and its corresponding power frequency phase. Set the amplitude axis resolution of the phase-resolved partial discharge spectrum to 256 and the phase axis resolution to 256, and construct a two-dimensional counting matrix. Map the amplitude of each discharge pulse event to an amplitude index, and map the power frequency phase of each discharge pulse event to a phase index; For each discharge pulse event, the corresponding element position in the two-dimensional counting matrix is ​​incremented according to its amplitude index and phase index. The value of each element in the two-dimensional counting matrix is ​​used as the image pixel intensity to form a phase-resolved partial discharge map.

[0009] Optionally, the preprocessed phase-resolved partial discharge map includes: Based on the two-dimensional counting matrix, an adaptive noise threshold removal process is performed to obtain the two-dimensional counting matrix after the adaptive noise threshold removal process. The phase resolution remapping is performed on the two-dimensional counting matrix after adaptive removal of the low noise threshold to form a phase-remapped two-dimensional counting matrix. The phase-remapped two-dimensional counting matrix is ​​subjected to pulse number normalization, amplitude normalization, and energy normalization to obtain the normalized two-dimensional counting matrix. The amplitude axis of the normalized two-dimensional counting matrix is ​​resampled at a uniform resolution to obtain a preprocessed phase-resolved partial discharge spectrum of uniform size.

[0010] Optionally, the step of splicing the vectors in a preset order to form a phase-independent composite feature vector includes: Based on the preprocessed phase-resolved partial discharge spectrum, the amplitude-phase statistical feature vector is extracted; Based on the preprocessed phase-resolved partial discharge spectrum, the pulse number-phase statistical feature vector is extracted; Based on the preprocessed phase-resolved partial discharge spectrum, the pulse number-amplitude statistical feature vector is extracted; Based on the preprocessed phase-resolved partial discharge spectrum, energy-phase statistical feature vectors are extracted; Based on the amplitude-phase statistical feature vector, frequency domain correlation feature vectors are extracted; Following a preset order, the amplitude-phase statistical feature vector, the pulse number-phase statistical feature vector, the pulse number-amplitude statistical feature vector, the energy-phase statistical feature vector, and the frequency domain correlation feature vector are connected end to end to obtain a phase-independent composite feature vector.

[0011] Optionally, the adaptive feature dimensionality reduction processing based on the contribution of class separability to the phase-independent composite feature vector includes: A training dataset is constructed based on phase-independent composite feature vectors; Each phase-independent composite feature vector in the training dataset is split into its feature components one by one. For each category and each feature dimension, the values ​​of all training samples in the corresponding category in the feature dimension are collected to form the set of values ​​for each category in each feature dimension. For each feature dimension, calculate the intra-class variance of all categories on the feature dimension, and sum the intra-class variances of all categories on the feature dimension to obtain the total intra-class dispersion of the feature dimension. For each feature dimension, calculate the global mean and the inter-class scatter of the i-th feature dimension; For each feature dimension, calculate the discriminant contribution factor; Sort the discriminant contribution factors of all feature dimensions in descending order. For the first few feature dimensions after sorting, add up their discriminant contribution factors in turn and calculate the cumulative contribution rate. When the cumulative contribution rate reaches the preset threshold, the number of selected feature dimensions is the final number of feature dimensions to be retained. Based on the final retained feature dimension, the feature components corresponding to the sorted ranking in the phase-independent composite feature vector are extracted sequentially to form a low-redundancy phase-independent discrimination feature vector.

[0012] Optionally, the discriminant energy potential function model includes: Based on low-redundancy phase-independent discriminant feature vectors, we construct the training samples required for the discriminant energy potential function model. For each training sample in each category, calculate the sample confidence weight; Based on the sample confidence weights, the low-redundancy phase-independent discriminant feature vectors of all training samples in the same category are weighted and summed to obtain the discriminant prototype center of the corresponding category of partial discharge mode; Based on the sample confidence weights and the discriminant prototype center, the discriminant diffusion matrix under each category is calculated, and the discriminant diffusion matrix is ​​subjected to structured shrinkage to obtain the structured shrinkage discriminant diffusion matrix. Based on the feature dimension, the contribution factor is determined, and a weighted matrix of the contribution factor is constructed. The structured shrinkage of the discriminative diffusion matrix, the weighted matrix of the discriminative contribution factor and the identity matrix are weighted, and then inverted to obtain the stable inverse diffusion matrix of the corresponding category of partial discharge mode. Based on the discriminant prototype center and the stabilized inverse diffusion matrix, a discriminant energy potential function corresponding to each category is established. The discriminant prototype center, the stabilized inverse diffusion matrix, and the discriminant energy potential function for each type of partial discharge mode are combined to form a set of discriminant energy potential functions, which serve as the discriminant energy potential function model.

[0013] Optionally, determining the partial discharge mode of the partial discharge sample to be tested according to the minimum discrimination energy criterion includes: Obtain low-redundancy phase-independent discriminant feature vectors for the partial discharge sample to be tested; Based on the calculation form of the discrimination energy potential function, the discrimination energy potential function value is calculated for each partial discharge mode category; The discriminant energy potential function values ​​calculated under all partial discharge mode categories are combined into a discriminant energy potential function value set. Based on the minimum discriminant energy criterion, the local partial discharge sample to be tested is determined according to the partial discharge mode category index. The corresponding partial discharge mode category definition rules are extracted from the pre-established correspondence table between partial discharge mode categories and physical defect types. The partial discharge sample to be tested is then classified according to the partial discharge generation mechanism, discharge pulse statistical characteristics, and phase-independent characteristic distribution characteristics.

[0014] Optionally, the pattern classification includes: Tip discharge mode: When the partial discharge sample to be tested belongs to the partial discharge mode category index, the discriminant prototype center is characterized by concentrated discharge pulse amplitude, single-peak distribution of amplitude-phase statistical features, pulse number-amplitude statistical features exceeding the first threshold in the high amplitude range, and low frequency components exceeding the second threshold in the frequency domain correlation features. Floating electrode discharge mode: When the partial discharge sample to be tested belongs to the partial discharge mode category index, the discriminant prototype center is characterized by a discharge pulse amplitude distribution range greater than the preset range, a pulse number-phase statistical feature that is dispersed in multiple phase intervals, an asymmetric energy-phase statistical feature, and the simultaneous presence of low and medium frequency components in the frequency domain correlation features. Insulation defect discharge mode: When the partial discharge sample to be tested belongs to the partial discharge mode category index, the discriminant prototype center is characterized by the discharge pulse energy accumulation exceeding the third threshold, the energy-phase statistical characteristics showing an increasing trend in a specific phase interval, and the pulse number-amplitude statistical characteristics forming a stable state in the medium amplitude interval.

[0015] The beneficial effects of this invention are: (1) In the feature construction stage, the present invention does not directly use the power frequency phase as the discrimination criterion. Instead, it systematically extracts the amplitude-phase statistical features, pulse number-phase statistical features, pulse number-amplitude statistical features, and energy-phase statistical features based on the preprocessed phase-resolved partial discharge spectrum. Furthermore, the amplitude-phase statistical features are regarded as periodic sequences and frequency domain transformation is performed to construct frequency domain related features. By converting the phase distribution information into spectral amplitude features, the influence of the overall translation, drift and synchronization error of the power frequency phase on the stability of the features is effectively eliminated. Even when there is a shift, jitter or even local loss in the power frequency phase, the consistency and discriminability of the feature expression can still be maintained, which significantly improves the engineering applicability of partial discharge pattern recognition in the complex working conditions of actual substations.

[0016] (2) This invention introduces an adaptive feature dimensionality reduction mechanism based on the contribution of class separability, which effectively suppresses the interference of redundant features and improves the discrimination reliability under sparse discharge conditions. By calculating the intra-class dispersion and inter-class dispersion of each feature dimension during the training phase, a discrimination contribution factor is constructed, and the final retained feature dimension is adaptively determined based on the cumulative contribution rate, thereby achieving targeted compression of the feature space. This avoids the problem of traditional unsupervised dimensionality reduction methods such as principal component analysis ignoring class discrimination information, and makes the retained features have stronger statistical discrimination ability in the distinction of partial discharge modes. In scenarios where the number of partial discharge pulses is limited and the distribution is sparse, adaptive dimensionality reduction can significantly reduce the interference of noise features and weak discrimination features on the model, and improve the stability and robustness of the discrimination model under small sample and short-term monitoring conditions.

[0017] (3) In the discrimination stage, this invention makes multi-level improvements to the discrimination energy potential function model to address the characteristics of sporadic, random, and sparse partial discharge signals. By introducing a sample confidence weight mechanism, the training samples that are more representative in terms of pulse sparsity and frequency domain consistency have a greater impact on the discrimination prototype center and diffusion structure. By performing structured shrinkage processing on the discrimination diffusion matrix, the numerical stability of the model is enhanced when the discreteness within the category is large. Furthermore, a sparsity consistency penalty term is introduced into the discrimination energy potential function, making the model more sensitive to the subtle differences in the sparse structure of different partial discharge modes. In cases where the boundaries of complex discharge modes are blurred and the category distributions overlap, it can achieve more refined and reliable mode differentiation, significantly improving the accuracy and engineering robustness of partial discharge mode recognition. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is a flowchart of a partial discharge pattern recognition method based on the phase-independent characteristics of ultra-high frequency signals proposed in this invention; Figure 2 The PRPD map is a partial discharge pattern recognition method based on phase-independent features of ultra-high frequency signals proposed in this invention. Figure 3 The amplitude-phase characteristics of the tip discharge mode are presented in the partial discharge mode identification method based on the phase-independent characteristics of ultra-high frequency signals proposed in this invention. Figure 4 The amplitude-phase characteristics of the floating electrode discharge mode are presented in the partial discharge mode recognition method based on the phase-independent characteristics of ultra-high frequency signals proposed in this invention. Figure 5 The amplitude-phase characteristics of the insulation defect discharge mode are presented in the partial discharge mode identification method based on the phase-independent characteristics of ultra-high frequency signals proposed in this invention. Figure 6 This invention relates to the amplitude-phase characteristics of free particle discharge modes in a partial discharge mode recognition method based on the phase-independent characteristics of ultra-high frequency signals. Detailed Implementation

[0019] Example 1: Reference Figures 1-6 A partial discharge pattern recognition method based on phase-independent characteristics of ultra-high frequency signals, comprising: The UHF receiver is used to collect ultra-high frequency time-domain pulse signals, and the ultra-high frequency time-domain pulse signals are mapped into two-dimensional coordinate data of amplitude and power frequency phase to generate phase-resolved partial discharge maps. In this embodiment, mapping the ultra-high frequency time-domain pulse signal into two-dimensional coordinate data of amplitude and power frequency phase includes: A UHF receiver using continuous sampling continuously acquires electromagnetic radiation signals generated by partial discharge inside a gas-insulated metal-enclosed switchgear within a preset acquisition time period to obtain ultra-high frequency time-domain pulse signals. The preset acquisition time is no less than ten seconds, which is recorded as the continuous acquisition time of the UHF signal. When the continuous acquisition time of the UHF signal reaches the requirement, the UHF time-domain pulse signal is obtained. The UHF time-domain pulse signal refers to the amplitude of the UHF pulse signal acquired at each sampling time point.

[0020] For ultra-high frequency time-domain pulse signals, discrete sampling is performed at equal intervals at a set sampling frequency to obtain a set of ultra-high frequency discrete sampling sequences; The sampling frequency refers to the number of sample points collected per second. The discrete sampling point index is the number of each sample point. The UHF discrete sampling value refers to the amplitude of the UHF signal collected at that number.

[0021] In the ultra-high frequency discrete sampling sequence, all discharge pulse events are detected to form a set of discharge pulse events; For each discharge pulse event, the arrival time and amplitude of the pulse event in the UHF time-domain pulse signal are determined. The total number of discharge pulse events is equal to the number of pulse events detected within the acquisition time. The arrival time is the specific time point when each pulse event occurs, and the pulse amplitude is the magnitude of the event.

[0022] Synchronously acquire the power frequency reference signal aligned with the UHF time-domain pulse signal on the time axis, and calculate the power frequency phase corresponding to each discharge pulse event based on the arrival time of each discharge pulse event and the phase information of the power frequency reference signal; The power frequency reference signal is a reference signal that reflects the phase of the power frequency voltage. The power frequency phase indicates the relative position of the pulse event in the power frequency voltage cycle.

[0023] All discharge pulse events are mapped according to their pulse amplitude and power frequency phase to obtain a set of two-dimensional coordinates of amplitude and power frequency phase. Each pair of two-dimensional coordinates of amplitude and power frequency phase corresponds to the amplitude of a discharge pulse event and its corresponding power frequency phase. Set the amplitude axis resolution of the phase-resolved partial discharge spectrum to 256 and the phase axis resolution to 256, and construct a two-dimensional counting matrix. A two-dimensional counting matrix is ​​used to count the number of pulse events under different combinations of amplitude and phase intervals. Each element in the two-dimensional counting matrix represents the number of pulses at the corresponding amplitude and phase index combination position.

[0024] Map the amplitude of each discharge pulse event to an amplitude index, and map the power frequency phase of each discharge pulse event to a phase index; The amplitude index ranges from 1 to 256, and the phase index ranges from 1 to 256.

[0025] For each discharge pulse event, the corresponding element position in the two-dimensional counting matrix is ​​incremented according to its amplitude index and phase index. The value of each element in the two-dimensional counting matrix is ​​used as the image pixel intensity to form a phase-resolved partial discharge map.

[0026] The final value of each element represents the number of pulse events corresponding to the amplitude index and phase index.

[0027] Preprocessing is performed on the phase-resolved partial discharge spectrum, and resampling is performed according to a uniform resolution to obtain a preprocessed phase-resolved partial discharge spectrum of uniform size; In this embodiment, the preprocessing of the phase-resolved partial discharge spectrum includes: Based on the two-dimensional counting matrix, an adaptive noise threshold removal process is performed to obtain the two-dimensional counting matrix after the adaptive noise threshold removal process. In Example 1, the phase-resolved partial discharge spectrum is expressed in the form of a two-dimensional counting matrix. The amplitude axis resolution of the two-dimensional counting matrix is ​​256, and the phase axis resolution is 256. The value of each matrix element represents the number of pulse events at the corresponding amplitude index and phase index combination position. Both the amplitude index and the phase index range from one to 256, and each combination of amplitude index and phase index corresponds to a pulse event count value.

[0028] The adaptive removal of the noise floor threshold is specifically performed by traversing all phase indices. For each phase index, the maximum amplitude index with a pulse event count greater than zero under that phase index is found. The minimum value among all maximum amplitude indices under all phase indices is determined as the amplitude index corresponding to the noise floor threshold. For matrix elements with amplitude indices less than the noise floor threshold, their pulse event count is set to zero. For matrix elements with amplitude indices not less than the noise floor threshold, their original pulse event count is retained.

[0029] The phase resolution remapping is performed on the two-dimensional counting matrix after adaptive removal of the low noise threshold to form a phase-remapped two-dimensional counting matrix. In Example 1, the phase resolution remapping requantizes the phase axis resolution from 256 to 64, and counts the pulse events in the original phase index range 1 to 256, according to the grouping and mapping of the phase intervals, and adds them to the new phase index range 1 to 64. The number of pulse events under each new quantized phase index is equal to the sum of the original number of all pulse events mapped to the new index.

[0030] The phase-remapped two-dimensional counting matrix is ​​subjected to pulse number normalization, amplitude normalization, and energy normalization to obtain the normalized two-dimensional counting matrix. The amplitude axis of the normalized two-dimensional counting matrix is ​​resampled at a uniform resolution to obtain a preprocessed phase-resolved partial discharge spectrum of uniform size.

[0031] In Example 1, the normalized amplitude axis is re-divided into 256 equally spaced amplitude intervals according to the preset target amplitude axis resolution. The pulse event counts that fall within the same target amplitude interval in the original amplitude axis are merged and mapped so that each target amplitude interval corresponds to a unique amplitude index. At the same time, the pulse event counts corresponding to each quantized phase index in the phase axis are rearranged according to the amplitude index after amplitude resampling to form a two-dimensional array with 256 amplitude indices and 256 phase indices. The value corresponding to each combination position of amplitude index and phase index in the two-dimensional array is used as the pixel intensity to construct a preprocessed phase-resolved partial discharge spectrum of uniform size.

[0032] Based on the preprocessed phase-resolved partial discharge spectrum, amplitude-phase statistical feature vectors, pulse number-phase statistical feature vectors, pulse number-amplitude statistical feature vectors, energy-phase statistical feature vectors, and frequency domain correlation feature vectors are extracted and spliced ​​in a preset order to form a phase-independent composite feature vector. In this embodiment, phase-independent composite feature vectors are formed by splicing them in a preset order, including: Based on the preprocessed phase-resolved partial discharge spectrum, the amplitude-phase statistical feature vector is extracted; For each phase index, find all amplitude indices with pixel intensities greater than zero, select the largest amplitude index, and map the largest amplitude index to the interval between zero and one through linear normalization. Arrange the amplitude normalization results under all phase indices in sequence to obtain an amplitude-phase statistical feature vector of length 256. If all pixel intensities under a certain phase index are zero, the corresponding normalized amplitude is set to zero.

[0033] Based on the preprocessed phase-resolved partial discharge spectrum, the pulse number-phase statistical feature vector is extracted; The pixel intensity values ​​under all amplitude indices corresponding to the phase index are summed up, and the sum of the pulse counts under all phase indices is arranged in order to obtain a pulse count-phase statistical feature vector with a length of 256.

[0034] Based on the preprocessed phase-resolved partial discharge spectrum, the pulse number-amplitude statistical feature vector is extracted; The pixel intensity values ​​under all phase indices corresponding to the amplitude index are summed up, and the sum of the pulse counts under all amplitude indices is arranged in order to obtain a pulse count-amplitude statistical feature vector with a length of 256.

[0035] Based on the preprocessed phase-resolved partial discharge spectrum, energy-phase statistical feature vectors are extracted; Add the normalized pulse power accumulation values ​​under all amplitude indices corresponding to the phase index to obtain the energy accumulation result under the phase index. Arrange the energy accumulation results under all phase indices in order to obtain an energy-phase statistical feature vector with a length of 256. Based on the amplitude-phase statistical feature vector, frequency domain correlation feature vectors are extracted; The amplitude-phase statistical eigenvector is regarded as a periodic sequence of length 256. According to the principle that the fundamental frequency of the power frequency is 50 Hz, the periodic sequence is subjected to discrete Fourier transform to obtain the corresponding discrete spectrum sequence. The amplitudes of six frequency components corresponding to 50 Hz, 100 Hz, 150 Hz, 200 Hz, 250 Hz and 300 Hz are selected and arranged in ascending order according to the fundamental frequency of the power frequency to form a frequency domain correlation eigenvector of length six.

[0036] Following a preset order, the amplitude-phase statistical feature vector, the pulse number-phase statistical feature vector, the pulse number-amplitude statistical feature vector, the energy-phase statistical feature vector, and the frequency domain correlation feature vector are connected end to end to obtain a phase-independent composite feature vector.

[0037] In Example 1, all feature components of the amplitude-phase statistical feature vector are used as the starting part of the composite feature vector according to their phase index order. Then, all feature components of the pulse number-phase statistical feature vector are appended to the starting part in the same phase index order. Then, all feature components of the pulse number-amplitude statistical feature vector are appended to the aforementioned feature components in the amplitude index order. Then, all feature components of the energy-phase statistical feature vector are appended to the aforementioned feature components in the phase index order. Finally, the feature values ​​corresponding to each frequency component in the frequency domain correlation feature vector are appended to the aforementioned feature components in a preset frequency order from low frequency to high frequency, resulting in a phase-independent composite feature vector with a length of 1030.

[0038] Using a training dataset containing various labeled partial discharge samples, an adaptive feature dimensionality reduction process based on the contribution of class separability is performed on the phase-independent composite feature vector to obtain a low-redundancy phase-independent discriminative feature vector. In this embodiment, adaptive feature dimensionality reduction based on the contribution of class separability is performed on the phase-independent composite feature vector, including: A training dataset is constructed based on phase-independent composite feature vectors; Each phase-independent composite feature vector contains 1,030 feature components. The training dataset contains multiple partial discharge mode categories, and each category corresponds to several training samples. For each training sample, there is a phase-independent composite feature vector corresponding to its respective partial discharge mode category.

[0039] Each phase-independent composite feature vector in the training dataset is split into its feature components one by one. For each category and each feature dimension, the values ​​of all training samples in the corresponding category in the feature dimension are collected to form the set of values ​​for each category in each feature dimension. For each feature dimension, calculate the intra-class variance of all categories on the feature dimension, and sum the intra-class variances of all categories on the feature dimension to obtain the total intra-class dispersion of the feature dimension. In Example 1, the intra-class mean of all categories on the feature dimension is calculated. The intra-class mean is the sum of the values ​​of all training samples in the feature dimension under the category divided by the number of training samples in the category. The intra-class variance is obtained by dividing the sum of the squares of the differences between the values ​​of all training samples in the feature dimension under the category and the intra-class mean of the category by the number of training samples in the category.

[0040] For each feature dimension, calculate the global mean and the inter-class scatter of the i-th feature dimension; The global mean is the sum of the values ​​of all training samples in all categories along the feature dimension divided by the total number of training samples in all categories. For each category, the number of training samples in that category is multiplied by the square of the difference between the intra-class mean and the global mean. The values ​​of all categories are summed to obtain the inter-class dispersion of the feature dimension.

[0041] For each feature dimension, calculate the discriminant contribution factor; The discriminant contribution factor is equal to the inter-class dispersion of the feature dimension divided by the sum of the total intra-class dispersion of the feature dimension and the stability factor. The discriminant contribution factor is used to measure the contribution of the feature dimension to the class separability. The stability factor is set to a constant of 1 multiplied by 10 to the power of negative 6 to avoid instability caused by a denominator of zero.

[0042] Sort the discriminant contribution factors of all feature dimensions in descending order. For the first few feature dimensions after sorting, add up their discriminant contribution factors in turn and calculate the cumulative contribution rate. When the cumulative contribution rate reaches the preset threshold, the number of selected feature dimensions is the final number of feature dimensions to be retained. The cumulative contribution rate is equal to the sum of the contribution factors of the selected feature dimensions divided by the sum of the contribution factors of all feature dimensions. The threshold for the cumulative contribution rate is set between 85% and 95%.

[0043] Based on the final retained feature dimension, the feature components corresponding to the sorted ranking in the phase-independent composite feature vector are extracted sequentially to form a low-redundancy phase-independent discrimination feature vector.

[0044] Based on the low-redundancy phase-independent discrimination feature vector, the discrimination prototype center and discrimination diffusion matrix of different partial discharge modes are calculated respectively. After the discrimination diffusion matrix is ​​stabilized, the corresponding discrimination energy potential function set is established to form the discrimination energy potential function model. In this embodiment, the determination of the energy potential function model includes: Based on low-redundancy phase-independent discriminant feature vectors, we construct the training samples required for the discriminant energy potential function model. Each training sample belongs to a certain partial discharge mode category. All training samples are divided into several categories according to the partial discharge mode category. Each category contains several low-redundancy phase-independent discriminant feature vectors.

[0045] For each training sample in each category, calculate the sample confidence weight; The calculation of sample confidence weights comprehensively considers two factors: first, the feature components in the low-redundancy phase-independent discriminant feature vector used to describe the stability of partial discharge pulse count distribution; and second, the feature components used to describe the stability of partial discharge frequency domain correlation. For the first part, a set of pulse sparsity-related feature components is extracted, and the absolute values ​​of the components are averaged to obtain a pulse sparsity scalar. For the second part, a set of frequency domain consistency-related feature components is extracted, and the absolute values ​​of the components are averaged using normalized weighting to obtain a frequency domain consistency scalar. The sample confidence weights are obtained by weighted exponential mapping between the pulse sparsity scalar and the frequency domain consistency scalar, and the sum of the normalized confidence weights of all samples of the same type is 1.

[0046] ; ; ; in, For impulse sparsity scalars, For frequency domain consistency scalar, This is the set of feature components in the low-redundancy, phase-independent discrimination feature vector used to describe the stability of the partial discharge pulse count distribution. This is the set of characteristic components used to describe the stability of the frequency domain correlation of partial discharge. for The Each feature component To serve as a stable constant to avoid a denominator of zero, and The weighting coefficients are used to adjust the contribution ratio of the impulse sparsity scalar and the frequency domain consistency scalar in the sample confidence weights.

[0047] Based on the sample confidence weights, the low-redundancy phase-independent discriminant feature vectors of all training samples in the same category are weighted and summed to obtain the discriminant prototype center of the corresponding category of partial discharge mode; The discriminant prototype center reflects the statistical mean of the corresponding category in the low-redundancy phase-independent discriminant feature space, and is closely related to the physical characteristics of the partial discharge signal.

[0048] Based on the sample confidence weights and the discriminant prototype center, the discriminant diffusion matrix under each category is calculated, and the discriminant diffusion matrix is ​​subjected to structured shrinkage to obtain the structured shrinkage discriminant diffusion matrix. The discriminant diffusion matrix is ​​used to measure the distribution structure of the corresponding category in the low-redundancy phase-independent discriminant feature space. To enhance the generalization and stability of the model under sporadic, random, and sparse partial discharge pulse conditions, the discriminant diffusion matrix is ​​subjected to structured shrinkage. The structured shrinkage process is to linearly combine the original discriminant diffusion matrix with its diagonal elements according to the shrinkage coefficient to obtain the structured shrinkage discriminant diffusion matrix, so that the discriminant diffusion matrix maintains reversibility and comparability when the discrete characteristics within the category are prominent.

[0049] ; in, To determine the prototype center, To determine the diffusion matrix, Let be the number of training samples for the k-th type of partial discharge mode, and construct the structured, shrunken discriminant diffusion matrix: ; in, For the first The contraction coefficient of the partial discharge-like mode, The discriminant diffusion matrix represents the result of the discriminant diffusion matrix. A diagonal matrix composed of the diagonal elements.

[0050] Based on the feature dimension, the contribution factor is determined, and a weighted matrix of the contribution factor is constructed. The structured shrinkage of the discriminative diffusion matrix, the weighted matrix of the discriminative contribution factor and the identity matrix are weighted, and then inverted to obtain the stable inverse diffusion matrix of the corresponding category of partial discharge mode. The discrimination contribution weighting matrix is ​​a diagonal matrix, and each diagonal element is obtained by normalizing the discrimination contribution factor of the corresponding feature dimension. The stabilized inverse diffusion matrix is ​​used to enhance the weight of the feature dimension with higher class separability in the discrimination energy potential function, thereby improving the model's ability to discriminate complex features of partial discharge.

[0051] Based on the discriminant prototype center and the stabilized inverse diffusion matrix, a discriminant energy potential function is established for each category.

[0052] The input to the discriminant energy potential function is the low-redundancy phase-independent discriminant feature vector of the sample to be discriminated, and the output is the weighted distance between the sample and the current class prototype center. The discriminant energy potential function consists of two parts: the first part is the quadratic distance between the sample and the class prototype center, which is weighted by the stabilized inverse diffusion matrix and is used to measure the fit with the distribution of the class in the feature space; the second part is the L1 norm distance between the sample and the class prototype center, multiplied by the sparsity consistency penalty coefficient, which is used to enhance the sensitivity to sporadic sparsity differences in partial discharge and enable the model to better distinguish easily confused partial discharge modes.

[0053] ; in, Representing vectors Norm, The sparsity consistency penalty coefficient is used. This represents the stabilized inverse diffusion matrix.

[0054] The discriminant prototype center, the stabilized inverse diffusion matrix, and the discriminant energy potential function for each type of partial discharge mode are combined to form a set of discriminant energy potential functions, which serve as the discriminant energy potential function model.

[0055] The low-redundancy phase-independent discrimination feature vector of the partial discharge sample to be tested is input into the discrimination energy potential function model. The partial discharge mode to which the partial discharge sample to be tested belongs is determined according to the minimum discrimination energy criterion, and the partial discharge mode identification result is output.

[0056] In this embodiment, the partial discharge mode of the partial discharge sample to be tested is determined according to the minimum discrimination energy criterion, including: Obtain low-redundancy phase-independent discriminant feature vectors for the partial discharge sample to be tested; Based on the calculation form of the discrimination energy potential function, the discrimination energy potential function value is calculated for each partial discharge mode category; The discriminant energy potential function values ​​calculated under all partial discharge mode categories are combined into a discriminant energy potential function value set, and the partial discharge mode category index to which the partial discharge sample to be tested belongs is determined based on the minimum discriminant energy criterion. The minimum discrimination energy criterion is as follows: among all partial discharge mode categories, select the category with the smallest discrimination energy potential function value, and its category index is the category to which the partial discharge sample to be tested belongs.

[0057] Based on the partial discharge mode category index, the corresponding partial discharge mode category definition rules are extracted from the pre-established correspondence table between partial discharge mode categories and physical defect types. The partial discharge samples to be tested are then classified according to the partial discharge generation mechanism, discharge pulse statistical characteristics, and phase-independent characteristic distribution features.

[0058] In this embodiment, pattern classification includes: Tip discharge mode: When the partial discharge sample to be tested belongs to the partial discharge mode category index, the discriminant prototype center is characterized by concentrated discharge pulse amplitude, single-peak distribution of amplitude-phase statistical features, pulse number-amplitude statistical features exceeding the first threshold in the high amplitude range, and low frequency components exceeding the second threshold in the frequency domain correlation features. Floating electrode discharge mode: When the partial discharge sample to be tested belongs to the partial discharge mode category index, the discriminant prototype center is characterized by a discharge pulse amplitude distribution range greater than the preset range, a pulse number-phase statistical feature that is dispersed in multiple phase intervals, an asymmetric energy-phase statistical feature, and the simultaneous presence of low and medium frequency components in the frequency domain correlation features. Insulation defect discharge mode: When the partial discharge sample to be tested belongs to the partial discharge mode category index, the discriminant prototype center is characterized by the discharge pulse energy accumulation exceeding the third threshold, the energy-phase statistical characteristics showing an increasing trend in a specific phase interval, the pulse number-amplitude statistical characteristics forming a stable distribution in the medium amplitude interval, and the frequency domain correlation characteristics having periodic characteristics. Free particle discharge mode: When the partial discharge sample to be tested belongs to the partial discharge mode category index, the discriminant prototype center is characterized by discrete distribution of discharge pulse amplitude and pulse number, random distribution of pulse number-phase statistical characteristics, higher fluctuation of energy-phase statistical characteristics than preset fluctuation, and the proportion of high frequency components in frequency domain correlation characteristics exceeds the fourth threshold.

[0059] Example 2: In a routine online monitoring task for partial discharge in gas-insulated metal-enclosed switchgear, the implementers conducted ultra-high frequency (UHF) acquisition of partial discharge signals for the bus section of equipment A. The acquisition device was set to a sampling frequency of 2 GHz, with a single continuous acquisition duration of 12 seconds. During one acquisition cycle, the device detected a total of 5320 discharge pulse events, and all signals were saved in the form of raw time-domain pulses.

[0060] The implementer first preprocessed the original signal, employing an adaptive noise threshold removal process. Within all 256 phase intervals, a pulse with an amplitude index of 13 was detected to be greater than zero at the 22nd phase index. Further analysis was conducted across all phase indices to determine the maximum amplitude index with a pulse count greater than zero; the minimum value was 13. Therefore, the number of all pulse events with amplitude indices less than 13 was reset to zero, while other indices remained unchanged. After this step, the original number of pulse events was reduced from 5320 to 4892, effectively eliminating noise.

[0061] The implementer requantized the denoised 2D counting matrix from 256 to 64 along the phase axis, expanding the phase interval width from 1.41° to 5.625°. The amplitude axis dimension remained unchanged at 256. Based on the denoising + remapping matrix, the maximum pulse count was normalized. In this sampling, the maximum pulse count was 38, and the normalized value for each pixel was between 0 and 1. The amplitude normalization range was statistically [0.12, 0.96], corresponding to the amplitude distribution range. During energy normalization, the maximum pulse energy was 3.28 μJ, and the energy of each element in the matrix was normalized according to the maximum energy, with the normalized matrix value range being [0, 1].

[0062] After preprocessing, the system automatically extracts features: For each phase index, it finds all amplitude indices with pixel intensities greater than zero, with the largest amplitude indices being 13, 14, 15, ..., 71, etc. After linear normalization and mapping to [0,1], it obtains an amplitude-phase statistical feature vector of length 256. Then, it sums the amplitude pixel intensities under each phase index to obtain a 256-dimensional pulse number-phase statistical feature vector. It sums the phase pixel intensities under each amplitude index to obtain a 256-dimensional pulse number-amplitude statistical feature vector. Finally, it sums the energy under each phase index to obtain a 256-dimensional energy-phase statistical feature vector. Finally, it performs a discrete Fourier transform on the amplitude-phase statistical feature vector to extract the amplitudes of the six frequency components corresponding to 50Hz, 100Hz, 150Hz, 200Hz, 250Hz, and 300Hz, forming a 6-dimensional frequency domain correlation feature vector. This results in a composite feature vector of length 1030.

[0063] For this sampling scenario, the system uses a historical sample set from the past 6 months for training. The sample set includes 50 sets of tip discharges (average number of events 5500, normalized amplitude peak values ​​all above 0.85), 50 sets of floating electrode discharges (average number of events 5700, wide pulse distribution, low frequency dominance in the frequency domain), 50 sets of insulation defect discharges (average number of events 4100, energy characteristics dominance), and 50 sets of free particle discharges (average number of events 1600, sparse pulses, significant high frequency components in the frequency domain).

[0064] The system calculates the intra-class variance and inter-class dispersion for each class of historical samples. The top 92 feature components with the highest contribution rate to the discrimination factor have a cumulative contribution rate of 89.2% and are retained as the final discrimination features. For each category, all training samples are weighted according to the sparsity and frequency domain consistency of the features to generate the discrimination prototype center and the structured shrinkage-diffusion matrix, with a shrinkage coefficient of 0.27 and a sparsity consistency penalty coefficient of 0.19.

[0065] The implementer assembled the 1030-dimensional feature vectors obtained from this batch of sampling into a 92-dimensional low-redundancy, phase-independent discriminative feature vector according to the selected feature components. After being input into the discriminative energy potential function model, the model automatically outputs the discriminative energy potential function values ​​for four categories, as follows: Tip discharge mode: 0.073; floating electrode discharge mode: 0.178; insulation defect discharge mode: 0.145; free particle discharge mode: 0.202.

[0066] The system identifies the category with the smallest energy potential function value as the tip discharge mode and automatically outputs the identification result of the partial discharge sample as tip discharge.

[0067] Throughout the identification process, the system simultaneously compared the traditional PRPD discrimination method. Due to intermittent loss of synchronization in the equipment's synchronization signal during this batch of sampling, the traditional method lacked effective phase for some pulse events during power frequency phase mapping, resulting in a significant drift in the final PRPD spectrum. Manual experience led to the classification as tip or floating electrode discharge, with a confidence level of 0.61. However, actual disassembly and inspection revealed that the discharge point was located at a tip, consistent with the identification method of this invention. This demonstrates that the traditional method suffers from ambiguity in classification and insufficient confidence.

[0068] Furthermore, the system processed the other three batches of sampling data in the same way. The statistics are as follows: Batch 1, number of pulse events 1720, model classifies it as free particle discharge with confidence level 0.93, traditional method fails to output category with confidence level 0.28.

[0069] Batch 2, number of pulse events 3986, model identified it as insulation defect discharge with confidence level 0.91, traditional method output category as tip discharge with confidence level 0.55.

[0070] Batch 3, pulse event count 5875, model identified as floating electrode discharge, confidence level 0.95, traditional method is consistent with the present invention, confidence level 0.82.

[0071] In all simulated test batches, the method of this invention can output clear identification results even in the absence of power frequency phase synchronization or sparse discharge pulses, and the confidence level is consistently higher than 0.90; the traditional method failed to provide discrimination results in 2 batches under synchronization loss and sparse scenarios, and the output results of the other 2 batches deviated from the actual results or had a confidence level lower than 0.60.

[0072] The entire sample identification process takes no more than 2.5 seconds. The automated identification output includes detailed information such as: the original sample number, the principal component distribution of each feature, the discriminant energy potential function score, the final pattern category, and the discrimination confidence level. Detailed features of a set of samples are as follows: Original sample number: S-2023-1119, number of pulse events: 3986, normalized peak value of amplitude-phase feature: 0.83, maximum value of energy-phase feature: 0.95, amplitude of 50Hz component in frequency domain: 0.74, amplitude of 300Hz component in frequency domain: 0.17, main category to be identified: insulation defect discharge, confidence level: 0.91.

[0073] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for partial discharge pattern recognition based on phase-independent characteristics of ultra-high frequency signals, characterized in that, include: The UHF receiver is used to collect ultra-high frequency time-domain pulse signals, and the ultra-high frequency time-domain pulse signals are mapped into two-dimensional coordinate data of amplitude and power frequency phase to generate phase-resolved partial discharge maps. Preprocessing is performed on the phase-resolved partial discharge spectrum, and resampling is performed according to a uniform resolution to obtain a preprocessed phase-resolved partial discharge spectrum of uniform size; Based on the preprocessed phase-resolved partial discharge spectrum, amplitude-phase statistical feature vectors, pulse number-phase statistical feature vectors, pulse number-amplitude statistical feature vectors, energy-phase statistical feature vectors, and frequency domain correlation feature vectors are extracted and spliced ​​in a preset order to form a phase-independent composite feature vector. Using a training dataset containing various labeled partial discharge samples, an adaptive feature dimensionality reduction process based on the contribution of class separability is performed on the phase-independent composite feature vector to obtain a low-redundancy phase-independent discriminative feature vector. Based on the low-redundancy phase-independent discrimination feature vector, the discrimination prototype center and discrimination diffusion matrix of different partial discharge modes are calculated respectively. After the discrimination diffusion matrix is ​​stabilized, the corresponding discrimination energy potential function set is established to form the discrimination energy potential function model. The low-redundancy phase-independent discrimination feature vector of the partial discharge sample to be tested is input into the discrimination energy potential function model. The partial discharge mode to which the partial discharge sample to be tested belongs is determined according to the minimum discrimination energy criterion, and the partial discharge mode identification result is output.

2. The partial discharge pattern recognition method based on the phase-independent characteristics of ultra-high frequency signals according to claim 1, characterized in that, The process of mapping ultra-high frequency time-domain pulse signals into two-dimensional coordinate data of amplitude and power frequency phase includes: A UHF receiver using continuous sampling continuously acquires electromagnetic radiation signals generated by partial discharge inside a gas-insulated metal-enclosed switchgear within a preset acquisition time period to obtain ultra-high frequency time-domain pulse signals. For ultra-high frequency time-domain pulse signals, discrete sampling is performed at equal intervals at a set sampling frequency to obtain a set of ultra-high frequency discrete sampling sequences; In the ultra-high frequency discrete sampling sequence, all discharge pulse events are detected to form a set of discharge pulse events; Synchronously acquire the power frequency reference signal aligned with the UHF time-domain pulse signal on the time axis, and calculate the power frequency phase corresponding to each discharge pulse event based on the arrival time of each discharge pulse event and the phase information of the power frequency reference signal; All discharge pulse events are mapped according to their pulse amplitude and power frequency phase to obtain a set of two-dimensional coordinates of amplitude and power frequency phase. Each pair of two-dimensional coordinates of amplitude and power frequency phase corresponds to the amplitude of a discharge pulse event and its corresponding power frequency phase. Set the amplitude axis resolution of the phase-resolved partial discharge spectrum to 256 and the phase axis resolution to 256, and construct a two-dimensional counting matrix. Map the amplitude of each discharge pulse event to an amplitude index, and map the power frequency phase of each discharge pulse event to a phase index; For each discharge pulse event, the corresponding element position in the two-dimensional counting matrix is ​​incremented according to its amplitude index and phase index. The value of each element in the two-dimensional counting matrix is ​​used as the image pixel intensity to form a phase-resolved partial discharge map.

3. The partial discharge pattern recognition method based on the phase-independent characteristics of ultra-high frequency signals according to claim 1, characterized in that, The preprocessed phase-resolved partial discharge map includes: Based on the two-dimensional counting matrix, an adaptive noise threshold removal process is performed to obtain the two-dimensional counting matrix after the adaptive noise threshold removal process. The phase resolution remapping is performed on the two-dimensional counting matrix after adaptive removal of the low noise threshold to form a phase-remapped two-dimensional counting matrix. The phase-remapped two-dimensional counting matrix is ​​subjected to pulse number normalization, amplitude normalization, and energy normalization to obtain the normalized two-dimensional counting matrix. The amplitude axis of the normalized two-dimensional counting matrix is ​​resampled at a uniform resolution to obtain a preprocessed phase-resolved partial discharge spectrum of uniform size.

4. The partial discharge pattern recognition method based on the phase-independent characteristics of ultra-high frequency signals according to claim 1, characterized in that, The step of splicing the vectors in a preset order to form a phase-independent composite feature vector includes: Based on the preprocessed phase-resolved partial discharge spectrum, the amplitude-phase statistical feature vector is extracted; Based on the preprocessed phase-resolved partial discharge spectrum, the pulse number-phase statistical feature vector is extracted; Based on the preprocessed phase-resolved partial discharge spectrum, the pulse number-amplitude statistical feature vector is extracted; Based on the preprocessed phase-resolved partial discharge spectrum, energy-phase statistical feature vectors are extracted; Based on the amplitude-phase statistical feature vector, frequency domain correlation feature vectors are extracted; Following a preset order, the amplitude-phase statistical feature vector, the pulse number-phase statistical feature vector, the pulse number-amplitude statistical feature vector, the energy-phase statistical feature vector, and the frequency domain correlation feature vector are connected end to end to obtain a phase-independent composite feature vector.

5. The partial discharge pattern recognition method based on the phase-independent characteristics of ultra-high frequency signals according to claim 1, characterized in that, The adaptive feature dimensionality reduction processing based on the contribution of class separability to the phase-independent composite feature vector includes: A training dataset is constructed based on phase-independent composite feature vectors; Each phase-independent composite feature vector in the training dataset is split into its feature components one by one. For each category and each feature dimension, the values ​​of all training samples in the corresponding category in the feature dimension are collected to form the set of values ​​for each category in each feature dimension. For each feature dimension, calculate the intra-class variance of all categories on the feature dimension, and sum the intra-class variances of all categories on the feature dimension to obtain the total intra-class dispersion of the feature dimension. For each feature dimension, calculate the global mean and the inter-class scatter of the i-th feature dimension; For each feature dimension, calculate the discriminant contribution factor; Sort the discriminant contribution factors of all feature dimensions in descending order. For the first few feature dimensions after sorting, add up their discriminant contribution factors in turn and calculate the cumulative contribution rate. When the cumulative contribution rate reaches the preset threshold, the number of selected feature dimensions is the final number of feature dimensions to be retained. Based on the final retained feature dimension, the feature components corresponding to the sorted ranking in the phase-independent composite feature vector are extracted sequentially to form a low-redundancy phase-independent discrimination feature vector.

6. The partial discharge pattern recognition method based on phase-independent characteristics of UHF signals according to claim 1, characterized in that, The discriminant energy potential function model includes: Based on low-redundancy phase-independent discriminant feature vectors, we construct the training samples required for the discriminant energy potential function model. For each training sample in each category, calculate the sample confidence weight; Based on the sample confidence weights, the low-redundancy phase-independent discriminant feature vectors of all training samples in the same category are weighted and summed to obtain the discriminant prototype center of the corresponding category of partial discharge mode; Based on the sample confidence weights and the discriminant prototype center, the discriminant diffusion matrix under each category is calculated, and the discriminant diffusion matrix is ​​subjected to structured shrinkage to obtain the structured shrinkage discriminant diffusion matrix. Based on the feature dimension, the contribution factor is determined, and a weighted matrix of the contribution factor is constructed. The structured shrinkage of the discriminative diffusion matrix, the weighted matrix of the discriminative contribution factor and the identity matrix are weighted, and then inverted to obtain the stable inverse diffusion matrix of the corresponding category of partial discharge mode. Based on the discriminant prototype center and the stabilized inverse diffusion matrix, a discriminant energy potential function corresponding to each category is established. The discriminant prototype center, the stabilized inverse diffusion matrix, and the discriminant energy potential function for each type of partial discharge mode are combined to form a set of discriminant energy potential functions, which serve as the discriminant energy potential function model.

7. The partial discharge pattern recognition method based on phase-independent characteristics of UHF signals according to claim 1, characterized in that, The process of determining the partial discharge mode of the partial discharge sample to be tested according to the minimum discrimination energy criterion includes: Obtain low-redundancy phase-independent discriminant feature vectors for the partial discharge sample to be tested; Based on the calculation form of the discrimination energy potential function, the discrimination energy potential function value is calculated for each partial discharge mode category; The discriminant energy potential function values ​​calculated under all partial discharge mode categories are combined into a discriminant energy potential function value set. Based on the minimum discriminant energy criterion, the local partial discharge sample to be tested is determined according to the partial discharge mode category index. The corresponding partial discharge mode category definition rules are extracted from the pre-established correspondence table between partial discharge mode categories and physical defect types. The partial discharge sample to be tested is then classified according to the partial discharge generation mechanism, discharge pulse statistical characteristics, and phase-independent characteristic distribution characteristics.

8. The partial discharge pattern recognition method based on the phase-independent characteristics of ultra-high frequency signals according to claim 7, characterized in that, The pattern classification includes: Tip discharge mode: When the partial discharge sample to be tested belongs to the partial discharge mode category index, the discriminant prototype center is characterized by concentrated discharge pulse amplitude, single-peak distribution of amplitude-phase statistical features, pulse number-amplitude statistical features exceeding the first threshold in the high amplitude range, and low frequency components exceeding the second threshold in the frequency domain correlation features. Floating electrode discharge mode: When the partial discharge sample to be tested belongs to the partial discharge mode category index, the discriminant prototype center is characterized by a discharge pulse amplitude distribution range greater than the preset range, a pulse number-phase statistical feature that is dispersed in multiple phase intervals, an asymmetric energy-phase statistical feature, and the simultaneous presence of low and medium frequency components in the frequency domain correlation features. Insulation defect discharge mode: When the partial discharge sample to be tested belongs to the partial discharge mode category index, the discriminant prototype center is characterized by the discharge pulse energy accumulation exceeding the third threshold, the energy-phase statistical characteristics showing an increasing trend in a specific phase interval, the pulse number-amplitude statistical characteristics forming a stable distribution in the medium amplitude interval, and the frequency domain correlation characteristics having periodic characteristics. Free particle discharge mode: When the partial discharge sample to be tested belongs to the partial discharge mode category index, the discriminant prototype center is characterized by discrete distribution of discharge pulse amplitude and pulse number, random distribution of pulse number-phase statistical characteristics, higher fluctuation of energy-phase statistical characteristics than preset fluctuation, and the proportion of high frequency components in frequency domain correlation characteristics exceeds the fourth threshold.

Images