A method and system for detecting insulation defects of a ring main unit based on a support vector machine

Abstract

The application belongs to the technical field of defect detection based on machine learning, and particularly relates to a ring main unit insulation defect detection method and system based on a support vector machine. The scheme realizes progressive diagnosis of ring main unit insulation defects from "whether a defect exists, what type of defect it is and where the defect is located" through organic combination of multi-source heterogeneous signal collection, multi-domain feature extraction, three-layer progressive fusion and cascade support vector machine classification. The scheme effectively improves the integrity and complementarity of detection information through systematic fusion of multi-source heterogeneous signals. Secondly, the three-layer progressive fusion compresses original high-dimensional features into compact fusion features, retaining multi-source complementary information and environmental self-adaptive capability. Meanwhile, the cascade support vector machine transmits decision function values between levels, enabling the rear classifier to utilize the classification prior of the front layer, further improving the accuracy of each level of classification.

Description

Technical Field

[0001] This invention belongs to the field of defect detection technology based on machine learning, specifically relating to a method and system for detecting insulation defects in ring main units based on support vector machines. Background Technology

[0002] Ring main units (RNBs) are critical equipment in power distribution networks, and their insulation condition directly affects the reliability of power supply. Partial discharge detection is the primary method for assessing insulation defects in RNBs. Common detection signals include transient earth voltage (TEV) signals, ultrasonic signals, and ultra-high frequency (UHF) signals. Existing technologies have seen some research on insulation defect detection methods based on support vector machines (SVMs), such as using single-class SVMs for fault identification in RNBs or classifying partial discharge modes based on support vector data description algorithms.

[0003] However, existing technologies still have the following shortcomings: First, the sensor types are limited, mostly relying on only one or two discharge signals for detection, failing to integrate multiple physical quantities—for example, TEV, ultrasound, and UHF sense the same discharge event from different physical principles, complementing each other. Meanwhile, infrared temperature signals can capture the thermal effect of discharge, and environmental temperature and humidity signals reflect important external causes of insulation defects; existing technologies lack the systematic integration and utilization of these multi-source heterogeneous signals. Second, feature extraction is mostly limited to manually designed statistical features or single phase-resolved spectral features, failing to systematically construct feature representations from multiple dimensions such as the time domain, frequency domain, phase domain, and time-frequency domain, resulting in insufficient characterization capabilities under complex working conditions. Third, existing methods mostly only perform binary classification judgments of "normal / abnormal" or identify a few types of defects, unable to achieve multi-granular progressive diagnosis from anomaly detection to defect type identification and defect location, and lacking information transfer between different levels of classification tasks, failing to utilize the classification results of the previous layer to assist the decision-making of the next layer. Fourth, existing denoising strategies are mostly fixed, lacking a closed-loop mechanism that adaptively selects denoising methods based on noise type and optimizes based on classification results. In addition, existing methods generally ignore the influence of ambient temperature and humidity on insulation status and do not include it in the input parameters of the detection model, resulting in significant differences in false alarm rates under different seasons or climate conditions. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by providing a method and system for detecting insulation defects in ring main units based on support vector machines, effectively solving the problems existing in the prior art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for detecting insulation defects in ring main units based on support vector machines, comprising: The partial discharge associated signal and environmental auxiliary signal of the ring main unit are collected synchronously and preprocessed accordingly. The partial discharge associated signal includes transient voltage to ground signal, ultrasonic signal and ultra-high frequency electromagnetic wave signal. The environmental auxiliary signal includes infrared temperature signal reflecting thermal effect and ambient temperature and humidity signal reflecting operating environment. Feature extraction was performed on the preprocessed partial discharge correlation signal and environmental auxiliary signal to obtain the feature vector representation of each signal; The feature vector representations of each signal are fused in three progressive layers to obtain fused features; wherein, the three progressive layers of fusion include: first, the feature vector representations corresponding to the partial discharge associated signal are fused within a layer, then the feature vector representations corresponding to the infrared temperature signal are incorporated, and finally the feature vector representations corresponding to the ambient temperature and humidity signal are incorporated. Based on the fusion features, anomaly detection, defect type identification, and defect location localization are sequentially performed using a pre-built cascaded support vector machine model to obtain defect detection results.

[0006] Furthermore, in the three-layer progressive fusion: the first layer performs kernel canonical correlation analysis on the feature vectors of the transient ground voltage signal, ultrasonic signal and ultra-high frequency electromagnetic wave signal in pairs, and splices the three pairs of analysis results into discharge fusion features; The second layer performs kernel canonical correlation analysis on the discharge fusion characteristics and the feature vector of the infrared temperature signal to obtain the physical fusion characteristics. The third layer performs kernel canonical correlation analysis on the physical fusion features and the feature vectors of the environmental temperature and humidity signals to obtain the final fusion features.

[0007] Furthermore, the three-layer progressive fusion is implemented through kernel canonical correlation analysis, including: The kernel function maps the feature vectors to be fused to the kernel space to obtain the corresponding kernel matrix. The kernel function uses a multi-kernel combination strategy to construct a candidate set of kernel functions. The candidate set of kernel functions includes at least Gaussian radial basis kernel functions of different scales, polynomial kernel functions of different orders, and linear kernel functions. The weight of each kernel function is adaptively determined by the kernel target alignment criterion. After centering the kernel matrix, the corresponding generalized eigenvalue problem is solved, and the eigenvectors corresponding to the first r largest eigenvalues ​​are selected as the kernel canonical correlation projection directions, where r is the preset number of projection directions; The original feature vector is projected along each projection direction to obtain the corresponding canonical variables. The canonical variables of each projection direction are then concatenated to form the fusion feature of the layer.

[0008] Furthermore, in the cascaded support vector machine model, after the previous support vector machine completes its classification, its decision function value is used as an auxiliary feature and concatenated into the fused feature as the input of the next support vector machine.

[0009] Furthermore, the preprocessing includes: identifying the noise type of the acquired signal, wherein the noise type identification specifically involves: extracting three discrimination indicators—spectral flatness, peak prominence, and peak-to-average power ratio—from the acquired signal, and, based on the discrimination indicators, classifying the noise type into one of four categories according to a preset priority: predominantly narrowband interference, predominantly pulse interference, predominantly white noise, or mixed noise. Based on the determined noise type, a corresponding denoising strategy is determined, and denoising is performed based on the denoising strategy to obtain preprocessed signal data.

[0010] Furthermore, the feature extraction includes: extracting time-domain features, frequency-domain features, phase-domain features, and time-frequency-domain features from each of the denoised signals; wherein, the time-domain features include peak value, mean value, RMS value, waveform factor, impulse factor, margin factor, kurtosis, skewness, pulse rise time, and discharge repetition rate; The frequency domain features include centroid frequency, frequency variance, root mean square frequency, and energy percentage of each frequency band. The phase domain features are extracted based on the phase-resolved spectrum of partial discharge, including the skewness of the positive and negative half-cycles, the kurtosis of the positive and negative half-cycles, the number of discharges in the positive and negative half-cycles, the cross-correlation factor of the positive and negative half-cycles, the discharge asymmetry, the starting phase of the positive and negative half-cycles, and the phase width of the positive and negative half-cycles. The time-frequency domain features include the energy proportion of each frequency band, wavelet packet energy entropy, and standard deviation of wavelet packet energy spectrum based on wavelet packet decomposition.

[0011] Furthermore, both defect type identification and defect location localization adopt a one-to-one multi-classification strategy. Specifically, a binary support vector machine is constructed for each pair of classes. During detection, each binary classifier predicts and votes on the two classes it distinguishes, and the class with the most votes is used as the output result.

[0012] A ring main unit insulation defect detection system based on support vector machine includes: The data acquisition module is configured to simultaneously acquire partial discharge associated signals and environmental auxiliary signals of the ring main unit and perform corresponding preprocessing. The partial discharge associated signals include transient voltage signals to ground, ultrasonic signals and ultra-high frequency electromagnetic wave signals. The environmental auxiliary signals include infrared temperature signals reflecting thermal effects and ambient temperature and humidity signals reflecting the operating environment. The feature extraction module is configured to extract features from the preprocessed partial discharge associated signal and the environmental auxiliary signal to obtain the feature vector representation of each signal. The feature fusion module is configured to perform a three-layer progressive fusion of the feature vector representations of each signal to obtain fused features; wherein, the three-layer progressive fusion includes: first, performing intra-layer fusion of the feature vector representations corresponding to the partial discharge associated signal, then incorporating the feature vector representations corresponding to the infrared temperature signal, and finally incorporating the feature vector representations corresponding to the ambient temperature and humidity signal. The defect detection module is configured to: based on the fused features, sequentially perform anomaly detection, defect type identification, and defect location localization through a pre-built cascaded support vector machine model to obtain defect detection results.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are as follows: (1) This invention provides a method and system for detecting insulation defects in ring main units based on support vector machines. The scheme achieves a progressive diagnosis of insulation defects in ring main units from "whether there is a defect → what type of defect → where the defect is located" by organically combining multi-source heterogeneous signal acquisition, multi-domain feature extraction, three-layer progressive fusion and cascaded support vector machine classification. The scheme effectively improves the integrity and complementarity of detection information through the systematic fusion of multi-source heterogeneous signals. Secondly, the original high-dimensional features are compressed into compact fused features through three-layer progressive fusion, preserving multi-source complementary information and environmental adaptability. At the same time, through the cascaded support vector machine, the decision function values ​​are passed between layers, enabling the classifier of the later layer to utilize the classification prior of the previous layer, further improving the accuracy of classification at each level.

[0014] (2) The solution described in this invention integrates three types of partial discharge correlation signals, namely TEV, ultrasound and UHF, with two types of environmental auxiliary signals, namely infrared temperature and ambient temperature and humidity, which breaks through the limitation of the single sensor type in the prior art. At the same time, feature fusion is performed through three-layer progressive nuclear canonical correlation analysis, namely: the first layer fuses the discharge signal, the second layer incorporates thermal effect information, and the third layer incorporates environmental parameters, so that the fused features have the ability to complement multiple physical fields and adapt to ambient temperature and humidity, effectively reducing the false alarm rate caused by seasonal climate change.

[0015] (3) The solution described in this invention completes anomaly detection, defect type identification and defect location location sequentially through the constructed three-layer cascaded support vector machine model, which solves the limitation that most existing technologies can only perform binary classification or identification of a few classes. The decision function values ​​are passed between each layer, so that the classifier of the later layer can use the classification results of the previous layer as auxiliary features, which not only improves the accuracy of classification at each level, but also realizes a complete diagnostic link from "detection" to "identification" and then to "location".

[0016] (4) The present invention automatically identifies the noise type by three indicators: spectral flatness, frequency peak prominence and time-domain peak-to-average ratio, and matches the corresponding denoising strategy according to priority, thus avoiding the problem of large differences in the effect of fixed denoising methods under different noise conditions in the prior art; at the same time, through the confidence feedback of the subsequent classifier output, a closed loop of "denoising → identification → feedback → strategy change" is formed, which further ensures the detection reliability under low quality signals. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below: Figure 1 This is a flowchart of a ring main unit insulation defect detection method based on support vector machine as described in an embodiment of the present invention; Figure 2 This is a flowchart of the adaptive data preprocessing process described in the embodiments of the present invention; Figure 3 This is a flowchart illustrating the three-layer progressive fusion process described in this embodiment of the invention; Figure 4 This is a schematic diagram of a ring main unit insulation defect detection system based on support vector machine, as described in an embodiment of the present invention. Detailed Implementation

[0018] To better understand the above-mentioned objectives, features and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0019] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0020] Example 1 The following describes in detail, with reference to the accompanying drawings, a method for detecting insulation defects in ring main units based on support vector machines, as described in Embodiment 1.

[0021] In one or more embodiments, such as Figure 1 As shown, a method for detecting insulation defects in ring main units based on support vector machines includes: The partial discharge associated signal and environmental auxiliary signal of the ring main unit are collected synchronously and preprocessed accordingly. The partial discharge associated signal includes transient voltage to ground signal, ultrasonic signal and ultra-high frequency electromagnetic wave signal. The environmental auxiliary signal includes infrared temperature signal reflecting thermal effect and ambient temperature and humidity signal reflecting operating environment. Feature extraction was performed on the preprocessed partial discharge correlation signal and environmental auxiliary signal to obtain the feature vector representation of each signal; The feature vector representations of each signal are fused in three progressive layers to obtain fused features; wherein, the three progressive layers of fusion include: first, the feature vector representations corresponding to the partial discharge associated signal are fused within a layer, then the feature vector representations corresponding to the infrared temperature signal are incorporated, and finally the feature vector representations corresponding to the ambient temperature and humidity signal are incorporated. Based on the fusion features, anomaly detection, defect type identification, and defect location localization are sequentially performed using a pre-built cascaded support vector machine model to obtain defect detection results.

[0022] In one or more embodiments, the acquisition of partial discharge associated signals and environmental auxiliary signals of the ring main unit specifically includes: (1) Sensor deployment The acquisition of partial discharge-related signals includes: Transient Voltage to Ground (TEV) Sensor: Utilizing capacitive coupling, this sensor is attached to the outer surface of the ring main unit's metal casing. When a partial discharge occurs inside the ring main unit, the discharge pulse induces a transient voltage to ground on the metal casing. The TEV sensor outputs this voltage signal through the coupling capacitor. Its detection frequency range is 3MHz to 100MHz, and its amplitude range is 0 to 10V. The advantage of the TEV sensor is its non-intrusive installation, requiring no power outage modifications to the ring main unit.

[0023] Ultrasonic sensors: Utilizing the piezoelectric acoustic emission detection principle, these sensors are installed inside each compartment of the ring main unit. The thermal expansion effect generated by partial discharge excites ultrasonic waves, with frequencies concentrated between 20kHz and 200kHz. This embodiment selects a piezoelectric ceramic sensor with a resonant frequency of 40kHz and a sensitivity of no less than -70dB (reference standard 1V / μbar). During installation, it is attached to the inner wall of the compartment using silicone grease coupling agent to ensure good acoustic impedance matching.

[0024] Ultra-high frequency (UHF) sensors: These employ broadband microstrip antennas or Archimedes spiral antennas and are installed inside a compartment or observation window within the ring main unit, with a detection frequency range of 300MHz to 3GHz. When partial discharge occurs, the rise time of the discharge pulse can reach sub-nanosecond levels, radiating abundant UHF electromagnetic wave components. UHF sensors receive signals in this frequency band, effectively avoiding conventional corona interference and mobile communication interference present in the substation environment.

[0025] The acquisition of environmental auxiliary signals includes: Infrared temperature sensor: Employing an uncooled focal plane infrared detector, it is positioned through the ring main unit's observation window or a pre-set infrared window at key heat-generating locations such as cable joints and busbar connections. The temperature measurement range is -20℃ to 300℃, with a thermal sensitivity of no less than 0.05℃. This sensor is used to detect localized abnormal temperature rises caused by insulation defects or poor contact.

[0026] Ambient temperature and humidity sensors: Digital integrated temperature and humidity sensors are installed both inside and outside the ring main unit to synchronously collect ambient temperature and relative humidity. The temperature measurement range is -40℃ to 85℃, with a measurement error of no more than ±0.5℃; the relative humidity measurement range is 0 to 100%RH, with a measurement error of no more than ±3%RH.

[0027] (2) Signal modulation and synchronous acquisition The analog signals output by each sensor are amplified, filtered, and impedance matched by the signal conditioning circuit before being connected to a high-speed multi-channel synchronous acquisition card. The acquisition card has a sampling rate of no less than 500MS / s, an analog bandwidth of no less than 200MHz, a vertical resolution of no less than 14bit, and supports synchronous sampling of no less than 8 channels.

[0028] Meanwhile, to achieve precise time alignment of multi-source signals, the IEEE 1588 (PTP: Precision Time Protocol) precision clock synchronization protocol is uniformly adopted to synchronize the sampling clocks of each sensor node, with a synchronization accuracy better than 1. Secondly, using GPS / BeiDou timing pulses as the global time reference, hardware triggering is applied to each acquisition channel to ensure that the sampling time deviation between channels is no greater than 100ns. This design ensures that the phase calculation error of the subsequent Phase Resolved Partial Discharge (PRPD) spectrum is less than 0.02°, meeting the time accuracy requirements for insulation defect analysis.

[0029] (3) Data output The data from each channel after synchronous acquisition are framed in units of fixed duration (e.g., 20ms, corresponding to one power frequency cycle). Each frame of data contains a timestamp, channel identifier, and sampled data segment, and is output to the subsequent preprocessing module in the form of structured data packets.

[0030] It should be noted that the five sensors used in this embodiment are not simply superimposed, but rather systematically selected based on the multi-physics coupling characteristics of the ring main unit's insulation defects: TEV, ultrasound, and UHF sensors perceive the same partial discharge event from three independent physical dimensions—electromagnetic transients, acoustic emission, and ultra-high frequency radiation—forming a redundant and complementary discharge sensing system; infrared temperature sensors supplement thermal anomalies that cannot be directly captured by discharge signals; and environmental temperature and humidity sensors provide environmental condition references for subsequent classifiers to suppress false alarms introduced by seasonal climate changes. These five types of signals, through the aforementioned precise synchronous acquisition mechanism, provide a time-consistent and physically complementary data foundation for subsequent multi-source feature fusion.

[0031] In one or more embodiments, the step of extracting features from the preprocessed partial discharge correlation signal and the environmental auxiliary signal to obtain feature vector representations of each signal specifically includes: (1) Adaptive data preprocessing Considering the complex electromagnetic environment at the ring main unit's operating site, various noises inevitably get mixed into the sensor-acquired signals. This embodiment does not employ a fixed denoising strategy, but instead first identifies the noise type, then matches the corresponding denoising method, and simultaneously optimizes the denoising strategy based on feedback from a subsequent classifier. Figure 2 As shown, it includes: 1) Noise type identification For each frame of signal Its spectrum is obtained by performing a fast Fourier transform. ,in, =0, 1, ..., N-1 are the temporal sampling point indices, and N is the number of sampling points per frame. f Based on frequency, three discriminant indicators are extracted: Spectral flatness Defined as: in, The number of effective frequency points in the spectrum. The frequency corresponding to the i-th frequency point. For frequency Spectral amplitude at that location, The value range of is [0,1]. The spectrum of white noise is approximately flat. Approaching 1. Peak prominence is defined as: in, The frequency corresponding to the peak value of the spectrum. The peak neighborhood width, To exclude the number of frequency points after excluding the peak neighborhood, narrowband interference has 1 to 3 prominent frequency peaks in the spectrum. The value is significantly high.

[0032] Time-domain peak-to-average ratio Defined as: in, Let n be the signal amplitude at the nth sampling point. The peak value is the signal value, and the denominator is the root mean square value of the signal. Pulse-type interference manifests as isolated spikes. It is much higher than the continuous discharge signal.

[0033] Based on the comparison results of the above three indicators and their corresponding thresholds, the noise type of the current frame signal is determined to be one of four categories according to the following priority: If peak prominence If the interference exceeds the first threshold, it is primarily determined to be narrowband interference. Otherwise, if the peak-to-average power ratio in the time domain is higher than that in the standard domain... If the interference exceeds the second threshold, it is determined to be primarily pulse-type interference. Otherwise, if the spectral flatness If the noise exceeds the third threshold, it is determined to be predominantly white noise. Otherwise, it is judged as mixed noise.

[0034] 2) Adaptive denoising strategy matching If the noise is determined to be predominantly white noise, wavelet thresholding is used for noise reduction. If the interference is determined to be mainly narrowband interference, an adaptive notch filter is used to filter out the interference frequencies. If the interference is determined to be primarily pulse-type, morphological filtering should be used. If the noise is determined to be mixed, variational mode decomposition is used.

[0035] 3) Noise reduction quality feedback optimization The denoised first i After the frame signal is input to the subsequent classifier (i.e., the cascaded support vector machine model), it receives the output of the first layer anomaly detection support vector machine from the cascaded support vector machine model. i Frame signal classification confidence Feedback, when confidence level continuously below the preset threshold At that time, that is: , appearing consecutively Frame time; This All frame signals are marked as samples with questionable denoising quality. They are then re-denoised using a variational mode decomposition strategy and fed back into the classifier, forming a closed loop of "denoising - recognition - feedback - strategy change".

[0036] Understandably, in some extreme cases, even after changing the strategy and re-denoising, the confidence level may still remain consistently lower than [a certain value]. At this point, noise reduction optimization is abandoned, these frame signals are marked as unknown modes, and manual review is triggered.

[0037] (2) Feature extraction For each source signal after denoising, features are extracted from four dimensions: time domain, frequency domain, phase domain, and time-frequency domain, to form the feature vector representation of each signal. For ease of understanding, the TEV signal is used as an example to illustrate the calculation method of each domain feature. The calculation method of each domain feature of other signals is the same.

[0038] 1) Temporal characteristics For each frame of denoised signal n = 0, 1, ..., N-1, where N is the number of sampling points per frame. Extract the following 10-dimensional temporal features: peak : ; mean : ; effective value : ; Waveform factor : ; in, The effective value of the signal. The mean of the signal; Pulse factor : ; margin factor : ; Kudo : ; in, The mean of the signal. The standard deviation of the signal; Skewness : ; Pulse rise time Defined as the average time required for a discharge pulse to rise from 10% to 90% of its peak amplitude, assuming a total of [number] signals are detected in one frame. The discharge pulse, the first pulse rise time for: ; in, and The first The moments when the pulse amplitude reaches 10% and 90% of its peak value; Discharge repetition rate : ; in, If the amplitude of a signal within a single frame exceeds a preset discharge pulse detection threshold. The number of pulses, The duration is the frame length.

[0039] 2) Frequency domain characteristics For each frame of signal Perform Fourier transform to obtain the spectrum Calculate the power spectrum Extract the following 8-dimensional frequency domain features: Center of gravity frequency : ; Frequency variance : ; Root mean square frequency : ; Frequency band energy proportion to : To make the signal spectrum within the effective bandwidth The internal frequency is divided into 5 consecutive bands. ,in, Given the sampling rate, calculate the proportion of energy in the m-th frequency band to the total energy: ; in, This represents the frequency range of the m-th equally divided frequency band.

[0040] 3) Phase domain characteristics For three types of discharge-related signals—TEV, ultrasound, and UHF—partial discharge phase-resolved (PRPD) spectra were constructed, with the power frequency voltage phase as the reference. The horizontal axis ( ), with discharge pulse amplitude Using the vertical axis, the distribution of discharge pulses within M power frequency cycles (M=50 in this embodiment) is statistically analyzed.

[0041] Divide the phase axis into equal-width panes, let the first... i The center phase of each phase pane is The average amplitude of the discharge pulse within the window pane is The total number of discharges is The positive half-cycle corresponds to The negative half-cycle corresponds to Based on this, the following 12-dimensional features are extracted: Positive half-circle skew and negative half-cycle skew : in, and These are the weighted average values ​​of the discharge phases during the positive and negative half-cycles, respectively. and This corresponds to the phase standard deviation; positive half-circumference kurtosis and negative half-cycle kurtosis : positive half-cycle discharge number and negative half-cycle discharge number : Positive and negative half-cycle cross-correlation factors : in, and These are the discharge amplitudes with the same phase bias during the positive and negative half-cycles, respectively. Discharge asymmetry : Positive half-cycle starting phase and the starting phase of the negative half-cycle : Positive half-cycle phase width and negative half-cycle phase width : It should be noted here that for infrared temperature signals and ambient temperature signals, since the discharge phase is not involved, their phase domain characteristics can be replaced by statistical quantities (segmented mean, segmented variance, etc.).

[0042] 4) Time-frequency domain characteristics Each frame of signal is generated using the db8 wavelet basis. Perform 5-level wavelet packet decomposition to obtain 32 terminal frequency band nodes. Let the th... j The wavelet packet coefficients of each terminal frequency band node are , , The number of coefficients in this frequency band, and the energy of this frequency band. for: ; Total Energy for: Energy proportion of each frequency band for: Based on the above indicators, the following 10-dimensional time-frequency domain features are extracted: Wavelet packet energy percentage of the top 8 highest energy frequency bands to All frequency bands are ordered by energy. Sort in descending order and take the energy percentage values ​​of the first 8 frequency bands.

[0043] Wavelet packet energy entropy : Among them, it is agreed that when hour, .

[0044] Standard deviation of wavelet packet energy spectrum : in, This is the average energy across 32 frequency bands.

[0045] 5) Feature Output The feature dimensions extracted from each signal and their summary are shown in Table 1: Table 1 Feature Dimensions For each frame of data, five sets of 40-dimensional feature vectors are output, which are then encapsulated according to the corresponding signal channel representation and passed into subsequent processes for feature fusion.

[0046] In one or more embodiments, the three-layer progressive fusion of the feature vector representations of each signal to obtain fused features specifically includes the following processing steps: The three-layer progressive fusion includes: first, performing intra-layer fusion on the feature vector representations corresponding to the partial discharge associated signals; then, incorporating the feature vector representations corresponding to the infrared temperature signals; and finally, incorporating the feature vector representations corresponding to the environmental temperature and humidity signals. In this three-layer progressive fusion, the fusion of at least one layer (in this embodiment, all three layers use kernel canonical correlation analysis) is obtained through the following kernel canonical correlation analysis: determining the maximum correlation projection direction of each feature vector to be fused in the kernel space, projecting each feature vector along the projection direction, and then stitching the projection results together to form the fusion feature of that layer. Specifically: (1) Basic principles of nuclear canonical correlation analysis Canonical correlation analysis (CCA) is a statistical method used to find the direction of maximum correlation between two sets of variables. Traditional CCA can only capture linear correlations. The scheme described in this embodiment uses kernel canonical correlation analysis, which maps features to a high-dimensional reproducing kernel Hilbert space (RKHS) through a kernel function, and finds the direction of maximum nonlinear correlation projection between two sets of features in the RKHS.

[0047] Let the two sets of feature matrices to be fused be respectively and Where n is the number of samples (i.e., the number of frames), and p and q are the dimensions of the two sets of features, respectively. This is achieved through a kernel function. and Map the two sets of features to their corresponding RKHS values ​​to obtain the kernel matrix. and .

[0048] In practice, the elements of the kernel matrix are calculated using a kernel function. Used to measure two feature vectors and Inner product in kernel space: in, Let be the mapping function from the original feature space to the kernel space. This indicates the inner product operation.

[0049] The scheme described in this embodiment adopts a multi-core combination strategy, and its kernel function pool includes the following candidate kernel functions: Gaussian radial basis function (RBF) kernel (3 different scales): in, The squared Euclidean distance between two eigenvectors is given. , , This is the kernel width parameter, used to control the local scope of the kernel function.

[0050] Polynomial kernel functions (two different orders): in, For vector inner product, the exponents are set to 2 and 3, respectively, corresponding to second-order and third-order polynomial mappings.

[0051] Linear kernel function (1): In practical implementation, multi-core combined kernel matrix It is constructed by weighted summation of the candidate kernel matrices: in, m =6 represents the total number of candidate kernel functions. For the first The kernel matrix corresponding to each candidate kernel function. For the first The weights of the kernel functions satisfy: , ; ; Kernel function weights The kernel target alignment (KTA) criterion is adaptively determined. KTA measures the consistency between the kernel matrix and the ideal target matrix, and is defined as follows: in, The sample label matrix, The KTA value is the Frobenius inner product. The larger the KTA value, the more consistent the kernel matrix is ​​with the label structure. The choice of kernel weight maximizes the KTA.

[0052] The scheme described in this embodiment uses Kernel Canonical Correlation Analysis (KCCA) as the core method for fusion of each layer, ultimately obtaining a compact and highly discriminative fusion feature. Within the KCCA framework, the projection direction is found. and The problem is transformed into solving the following generalized eigenvalue problem. First, the kernel matrix is ​​centered: in, and The kernel matrix after centralization. For a centered matrix, , for n An identity matrix of order 1. It is a column vector of all 1s, that is , where n is the number of samples.

[0053] The projection direction can be obtained by solving the following eigenvalue problem: in, The eigenvalue represents the degree of correlation between the two sets of features in the projection direction; This is a regularization parameter used to prevent overfitting and ensure numerical stability. It is an identity matrix.

[0054] Take before The eigenvector pairs corresponding to the largest eigenvalues ,in, , The preset maximum number of projection directions (the scheme described in this embodiment takes...) ),this The projection direction is the direction of maximum correlation projection of the two sets of feature vectors in the kernel space.

[0055] Projecting the features along each projection direction yields the corresponding canonical variables. and : , ; in, and All n The dimensional column vectors represent the two sets of features at the th dimension. Projection results along the most relevant projection direction.

[0056] Will The typical variables are concatenated to form the fusion feature matrix of this layer. : in, This indicates that column vectors are concatenated horizontally.

[0057] (2) Three-layer progressive converged architecture The scheme described in this embodiment divides the five sets of feature vectors into three levels according to their physical meaning, and achieves feature dimensionality reduction and information refinement through layer-by-layer KCCA fusion. For example... Figure 3 As shown, its specific fusion architecture is as follows: First layer: Fusion within the discharge signal layer The feature vectors of three types of discharge-related signals—transient voltage to ground (TEV), ultrasonic waves, and ultra-high frequency (UHF) radiation—are fused. These three sensors detect the same partial discharge event from three different physical principles: electromagnetic transients, acoustic emission, and UHF radiation, and there is a nonlinear complementary relationship between them. In the specific fusion process, the features of the three sensors are paired and KCCA is performed—KCCA is applied to the three pairs of features (TEV, ultrasonic), (TEV, UHF), and (ultrasonic, UHF), and the first five projection directions of each pair are taken (i.e.,... =5), each obtained 10-dimensional canonical variables, totaling 30 dimensions in three pairs, which were spliced ​​together to form the discharge fusion characteristics. The reason for using pairwise pairing is that KCCA is naturally suitable for correlation analysis between two sets of variables. Pairwise pairing can preserve the independent complementary information between each sensor pair, and then the information can be synthesized by splicing.

[0058] Second layer: Cross-layer fusion of physical information The discharge fusion characteristics of the first layer output The feature vectors of the infrared temperature signal are fused. Since the infrared temperature signal reflects the thermal effect of partial discharge or leakage current and has a physical causal relationship with the discharge signal, KCCA is used to extract the maximum correlation projection direction between the discharge features and the thermal features (taking...). =10), and the physical fusion feature is obtained by concatenating the two sets of 10-dimensional canonical variables. .

[0059] Third layer: Cross-layer fusion of environmental parameters Physical fusion features output from the second layer The feature vectors of the ambient temperature and humidity signals are fused. Since ambient temperature and humidity are important external inducing factors for insulation defects in ring main units—under high temperature and high humidity conditions, even normal insulation may exhibit similar defect characteristics due to surface condensation—KCCA is used to learn the nonlinear coupling relationship between the physical fusion features and environmental parameters (taking...). =10), enabling the final fused features to adapt to the environment, thus obtaining the final fused features. .

[0060] Through the above three-layer progressive fusion, the original five groups of features, totaling 200 dimensions, are compressed to 20 dimensions. While retaining multi-source complementary information and environmental adaptability, redundant information is effectively removed, providing a compact and highly discriminative input for the subsequent classifier.

[0061] In one or more embodiments, the step of obtaining defect detection results by sequentially performing anomaly detection, defect type identification, and defect location localization based on the fused features and a pre-built cascaded support vector machine model specifically includes: (1) Construction of cascaded support vector machine model The scheme described in this embodiment constructs a three-layer cascaded support vector machine model, wherein: The first layer is an anomaly detection support vector machine, used to distinguish between normal and abnormal states; The second layer is a defect type identification support vector machine, which is used to identify defect types, including five types: surface discharge, floating discharge, tip discharge, internal air gap discharge, and moisture condensation. The third layer is a support vector machine for defect location, used to locate the component where the defect is located, including cable joints, busbar compartments, insulators, CT / PT compartments, circuit breaker compartments and six other types.

[0062] In practice, information is transferred between different layers. Specifically, after the previous layer's support vector machine completes its classification, its decision function value is concatenated as an additional one-dimensional auxiliary feature into the fused features and provided to the next layer. For example, the decision function value of the first layer's SVM is... The second layer i The decision function value of a one-to-one binary classifier is Then the input features of the second layer are [ , The input features of the third layer are [ , , , …, ],in,[ ] indicates vector concatenation.

[0063] (2) Defect detection Based on the above cascaded support vector machine model, the defect detection process is as follows: Final fusion features Input the first-layer anomaly detection support vector machine, and output a binary classification result (normal state or anomaly state) and classification confidence score. Specifically: when classified as a normal state and the confidence score is greater than or equal to a preset threshold... If the condition is met, output "Insulation normal" and end the detection process; otherwise, pass the fusion feature to the second layer.

[0064] The second-layer defect type identification support vector machine classifies the input fused features and outputs the specific defect type. The second layer employs a "one-to-one" multi-classification strategy: two classes are randomly selected from five defect categories to form a class pair. A separate binary classifier is set up for each class pair to distinguish between the two defect types, resulting in a total of 10 binary classifiers, each corresponding to a unique pair of defect categories. During detection, the input features are fed into these 10 binary classifiers. Each classifier determines which category the input sample belongs to and votes. By counting the number of votes for each of the five defect types, the category with the most votes is taken as the defect type output by the second layer. If multiple categories have the same number of votes, the sum of the absolute values ​​of the decision function values ​​for the input sample in their respective binary classifiers is compared, and the category with the largest sum is taken as the final output.

[0065] The third-layer defect location support vector machine classifies the input fused features and outputs the location of the defect in the component. Like the second layer, the third layer employs a "one-to-one" multi-classification strategy: two classes are randomly selected from six possible locations to form a class pair. A binary classifier is set up for each class pair, resulting in a total of 15 binary classifiers. During detection, each binary classifier makes predictions and votes. The class with the most votes is selected as the defect location output by the third layer. If multiple classes have the same number of votes, the absolute values ​​of the decision function values ​​for that input sample in their respective binary classifiers are compared, and the class with the largest sum is taken as the final output.

[0066] The final output of the defect detection results includes whether a defect exists, the type of defect, and the location of the defect.

[0067] Example 2 In one or more embodiments, corresponding to the method described in Embodiment 1, such as Figure 4 As shown, this embodiment provides a ring main unit insulation defect detection system based on support vector machine, including: The data acquisition module is configured to simultaneously acquire partial discharge associated signals and environmental auxiliary signals of the ring main unit and perform corresponding preprocessing. The partial discharge associated signals include transient voltage signals to ground, ultrasonic signals and ultra-high frequency electromagnetic wave signals. The environmental auxiliary signals include infrared temperature signals reflecting thermal effects and ambient temperature and humidity signals reflecting the operating environment. The feature extraction module is configured to extract features from the preprocessed partial discharge associated signal and the environmental auxiliary signal to obtain the feature vector representation of each signal. The feature fusion module is configured to perform a three-layer progressive fusion of the feature vector representations of each signal to obtain fused features; wherein, the three-layer progressive fusion includes: first, performing intra-layer fusion of the feature vector representations corresponding to the partial discharge associated signal, then incorporating the feature vector representations corresponding to the infrared temperature signal, and finally incorporating the feature vector representations corresponding to the ambient temperature and humidity signal. The defect detection module is configured to: based on the fused features, sequentially perform anomaly detection, defect type identification, and defect location localization through a pre-built cascaded support vector machine model to obtain defect detection results.

[0068] It is understood that the system described in this embodiment corresponds to the method described in Embodiment 1, and its technical details have been described in detail in Embodiment 1, so they will not be repeated here.

[0069] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for detecting insulation defects in ring main units based on support vector machines, characterized in that, include: The partial discharge associated signal and environmental auxiliary signal of the ring main unit are collected synchronously and preprocessed accordingly. The partial discharge associated signal includes transient voltage to ground signal, ultrasonic signal and ultra-high frequency electromagnetic wave signal. The environmental auxiliary signal includes infrared temperature signal reflecting thermal effect and ambient temperature and humidity signal reflecting operating environment. Feature extraction was performed on the preprocessed partial discharge correlation signal and environmental auxiliary signal to obtain the feature vector representation of each signal; The feature vector representations of each signal are fused in three progressive layers to obtain fused features; wherein, the three progressive layers of fusion include: first, the feature vector representations corresponding to the partial discharge associated signal are fused within a layer, then the feature vector representations corresponding to the infrared temperature signal are incorporated, and finally the feature vector representations corresponding to the ambient temperature and humidity signal are incorporated. Based on the fusion features, anomaly detection, defect type identification, and defect location localization are sequentially performed using a pre-built cascaded support vector machine model to obtain defect detection results.

2. The method for detecting insulation defects in ring main units based on support vector machines as described in claim 1, characterized in that, In the three-layer progressive fusion: the first layer performs kernel canonical correlation analysis on the feature vectors of the transient ground voltage signal, ultrasonic signal and ultra-high frequency electromagnetic wave signal in pairs, and splices the three pairs of analysis results into discharge fusion features; The second layer performs kernel canonical correlation analysis on the discharge fusion characteristics and the feature vector of the infrared temperature signal to obtain the physical fusion characteristics. The third layer performs kernel canonical correlation analysis on the physical fusion features and the feature vectors of the environmental temperature and humidity signals to obtain the final fusion features.

3. The method for detecting insulation defects in ring main units based on support vector machines as described in claim 1, characterized in that, The three-layer progressive fusion is achieved through kernel canonical correlation analysis, including: The kernel function maps the feature vectors to be fused to the kernel space to obtain the corresponding kernel matrix. The kernel function uses a multi-kernel combination strategy to construct a candidate set of kernel functions. The candidate set of kernel functions includes at least Gaussian radial basis kernel functions of different scales, polynomial kernel functions of different orders, and linear kernel functions. The weight of each kernel function is adaptively determined by the kernel target alignment criterion. After centering the kernel matrix, the corresponding generalized eigenvalue problem is solved, and the eigenvectors corresponding to the first r largest eigenvalues ​​are selected as the kernel canonical correlation projection directions, where r is the preset number of projection directions; The original feature vector is projected along each projection direction to obtain the corresponding canonical variables. The canonical variables of each projection direction are then concatenated to form the fusion feature of the layer.

4. The method for detecting insulation defects in ring main units based on support vector machines as described in claim 1, characterized in that, In the cascaded support vector machine model, after the previous support vector machine completes its classification, its decision function value is used as an auxiliary feature and concatenated into the fused feature as the input of the next support vector machine.

5. The method for detecting insulation defects in ring main units based on support vector machines as described in claim 1, characterized in that, The preprocessing includes: identifying the noise type of the acquired signal, wherein the noise type identification specifically involves: extracting three discrimination indicators—spectral flatness, peak prominence, and peak-to-average power ratio—from the acquired signal, and, based on the discrimination indicators, classifying the noise type into one of four categories according to a preset priority: predominantly narrowband interference, predominantly pulse interference, predominantly white noise, or mixed noise. Based on the determined noise type, a corresponding denoising strategy is determined, and denoising is performed based on the denoising strategy to obtain preprocessed signal data.

6. The method for detecting insulation defects in ring main units based on support vector machines as described in claim 1, characterized in that, The feature extraction includes: extracting time-domain features, frequency-domain features, phase-domain features, and time-frequency-domain features from each denoised signal; wherein, the time-domain features include peak value, mean value, RMS value, waveform factor, impulse factor, margin factor, kurtosis, skewness, pulse rise time, and discharge repetition rate; The frequency domain features include centroid frequency, frequency variance, root mean square frequency, and energy percentage of each frequency band. The phase domain features are extracted based on the phase-resolved spectrum of partial discharge, including the skewness of the positive and negative half-cycles, the kurtosis of the positive and negative half-cycles, the number of discharges in the positive and negative half-cycles, the cross-correlation factor of the positive and negative half-cycles, the discharge asymmetry, the starting phase of the positive and negative half-cycles, and the phase width of the positive and negative half-cycles. The time-frequency domain features include the energy proportion of each frequency band, wavelet packet energy entropy, and standard deviation of wavelet packet energy spectrum based on wavelet packet decomposition.

7. The method for detecting insulation defects in ring main units based on support vector machines as described in claim 1, characterized in that, Both defect type identification and defect location localization adopt a one-to-one multi-classification strategy. Specifically, a binary support vector machine is constructed for each pair of classes. During detection, each binary classifier predicts and votes on the two classes it distinguishes, and the class with the most votes is used as the output result.

8. A ring main unit insulation defect detection system based on support vector machine, characterized in that, include: The data acquisition module is configured to simultaneously acquire partial discharge associated signals and environmental auxiliary signals of the ring main unit and perform corresponding preprocessing. The partial discharge associated signals include transient voltage signals to ground, ultrasonic signals and ultra-high frequency electromagnetic wave signals. The environmental auxiliary signals include infrared temperature signals reflecting thermal effects and ambient temperature and humidity signals reflecting the operating environment. The feature extraction module is configured to extract features from the preprocessed partial discharge associated signal and the environmental auxiliary signal to obtain the feature vector representation of each signal. The feature fusion module is configured to perform a three-layer progressive fusion of the feature vector representations of each signal to obtain fused features; wherein, the three-layer progressive fusion includes: first, performing intra-layer fusion of the feature vector representations corresponding to the partial discharge associated signal, then incorporating the feature vector representations corresponding to the infrared temperature signal, and finally incorporating the feature vector representations corresponding to the ambient temperature and humidity signal. The defect detection module is configured to: based on the fused features, sequentially perform anomaly detection, defect type identification, and defect location localization through a pre-built cascaded support vector machine model to obtain defect detection results.

Images