Search

CN-121596055-B - Intelligent monitoring method, equipment and medium for partial discharge abnormality of electrical equipment

CN121596055BCN 121596055 BCN121596055 BCN 121596055BCN-121596055-B

Abstract

The invention discloses an intelligent monitoring method, equipment and medium for partial discharge abnormality of electrical equipment, which relate to the technical field of power system state monitoring and fault diagnosis, and effectively improve the accuracy and robustness of GIS equipment partial discharge abnormality monitoring by integrating a multi-scale feature extraction, an attention mechanism and an intelligent fusion frame; in the signal processing level, the multi-scale residual error network can capture local details and global modes at the same time, overcomes the defect of characterization of single-scale features in a complex discharge mode, combines a space-channel joint attention mechanism, dynamically focuses on key feature areas, remarkably enhances the separation capability of cross interference signals, and avoids mismatch problems caused by fixed parameters by coupling phase selectivity, amplitude threshold and neighborhood connectivity based on PRPD phase reference self-adaptive adjustment and inhibiting non-target noise while retaining effective signals.

Inventors

  • CHEN KAI
  • ZHAO JUAN

Assignees

  • 北京华晨浩智科技有限公司

Dates

Publication Date
20260508
Application Date
20251216

Claims (9)

  1. 1. An intelligent monitoring method for partial discharge abnormality of electrical equipment is used for a gas insulated switchgear GIS, and is characterized by comprising the following steps: S1, collecting, namely arranging at least two UHF sensors on a GIS shell and acquiring partial discharge electromagnetic signals, and optionally arranging at least one acoustic emission sensor to acquire acoustic signals; s2, synchronizing and calibrating, namely performing time synchronization and amplitude calibration on different sensor channels; S3, preprocessing, namely performing time domain/frequency domain or time-frequency domain transformation on the acquired signals based on PRPD phase references to obtain a feature map; s4, extracting the characteristics, namely inputting the characteristic map into a multi-scale residual characteristic extraction network to obtain multi-scale characteristics; S5, attention weighting, namely weighting the multi-scale feature by applying spatial-channel joint attention to obtain a weighted feature; S6, fusion and recognition, namely carrying out multi-sensor data fusion based on the weighted characteristics and outputting partial discharge type and position information; The method also comprises the steps of carrying out masking and noise suppression on the time frequency characteristics according to the PRPD phase reference, so that the characteristics inconsistent with the target discharge phase window are reduced in weight or shielded; The generation mode of the phase perception mask filtering is as follows: Step A1, after obtaining PRPD phase reference and time-frequency coefficient, for each time-frequency unit Calculate soft mask for weighting of step S5: , Wherein, the Representation of At the soft mask weights of the set, For the time index of the time index, For the frequency index to be used, As a function of the kernel of the phase window, For the moment of time The phase of the corresponding PRPD is determined, For the center of the phase of the target discharge, Is the half width of the phase window, For the time-frequency transform coefficients, For the amplitude value thereof, Is a frequency band Is provided with an adaptive amplitude threshold value of (a), Is the slope parameter of the amplitude threshold, As a measure of neighborhood consistency, As a slope parameter of the connectivity threshold, In order to communicate the reference value with the reference value, Is a natural exponential function; step A2, obtaining a frequency band threshold value based on sample statistics outside a phase window: , Wherein, the Is a frequency band Noise location statistics of samples outside the phase window, In order to correspond to the dispersion metric, Is the frequency band amplification factor; Step A3, using time-frequency neighborhood Counting threshold unit ratio: , Wherein, the Is that Is a set of time-frequency neighbors of (c), As the cardinality of the set, For the time and frequency index in the neighborhood, Taking 1 for the indication function, or taking 0 for the indication function; Step A4, carrying out peak tracking on the PRPD phase histogram by adopting a sliding time window, dynamically updating the phase center and the half width of the window, and refreshing background statistics by using samples outside the window when the load fluctuates so as to enable the threshold value and the phase window to cooperatively change; and step A5, carrying out point-by-point multiplicative weighting on the time-frequency diagram by using a soft mask, and then carrying out connected domain screening and small plaque inhibition to obtain a weighted characteristic with stable morphology for fusion and identification in step S6.
  2. 2. An electrical equipment partial discharge anomaly intelligent monitoring method as claimed in claim 1, wherein the preprocessing comprises performing short-time fourier transform STFT or continuous wavelet transform CWT on the acquired signals to obtain an amplitude-phase time-frequency diagram, and generating a PRPD diagram by power frequency voltage or equivalent phase reference.
  3. 3. The intelligent monitoring method for partial discharge abnormality of electrical equipment according to claim 1, wherein the multi-scale residual feature extraction network comprises residual blocks with at least three different convolution kernel sizes, wherein the residual blocks are connected in a jumping manner and comprise a convolution layer and a batch normalization layer, and the convolution layer and the batch normalization layer are used for simultaneously extracting local textures and global patterns.
  4. 4. A method for intelligent monitoring of partial discharge anomalies in electrical equipment according to claim 1, wherein the attention weighting includes non-local or self-attention to model correlation between locations of the signature, and wherein weights are assigned on multiple scales in conjunction with the feature pyramid.
  5. 5. The intelligent monitoring method for partial discharge abnormality of electrical equipment according to claim 1, wherein the multi-sensor data fusion is realized by adopting a graph neural network, wherein each sensor is taken as a graph node, edges are constructed and side weights are set based on the geometric distance and cross-channel cross-correlation of the sensors, and multi-channel weighting characteristics are propagated and aggregated on the graph; the side weight is constructed and normalized by combining a physical distance and cross-channel cross-correlation, and the method comprises the following steps: Step B1, on the graph with each sensor as a graph node, for any pair of nodes Original joint weights are calculated: , Wherein, the As the joint edge weights are not normalized, For the sensor index to be used, And (3) with Is a non-negative weighting coefficient and , As a function of the distance decay, Is a node And node Is used for the geometric distance of (a), As a function of the correlation map, Normalizing the cross-correlation metric for the cross-channel; step B2, the distance component adopts adjustable exponential decay, and is assisted by distance gating: , Wherein, the As a parameter of the distance scale, In order to attenuate the shape parameter of the mold, As a distance component after the distance gating, To indicate a function, the condition satisfies a1, otherwise a 0, Is the upper threshold of the distance; Step B3, sliding time window With time delay search sets The inter-channel normalization of the inner estimation correlates and takes peak values, and the threshold is smoothed by an S-shaped function: , Wherein, the And (3) with Respectively nodes 、 Is provided with a time domain signal of (a), And (3) with For corresponding time window The average value of the inner part of the frame, For the cross-correlation time delay, A time delay set for searching; , Wherein, the As a function of the logical S-shape, For the relevant mapping slope parameter(s), Is the relevant soft threshold; Step B4, based on the threshold and The neighbors generate sparse adjacencies, and then symmetric normalization is performed to give two types of graph Laplacians: , Wherein, the For the joint edge weights after the sparsification, For the associated hard-threshold value, Is pressed by Ordered node A kind of electronic device A set of neighbors that are close together, Is an adjacency matrix; , , Wherein, the In the form of a degree matrix, For a symmetrical normalized adjacency, And (3) with Respectively symmetric and random walk-type drawing laplace, Is in combination with Identity matrix of the same order; In the step B5, the step of setting the position of the base plate, And (3) with At the position of Internal values are selected and satisfy the convex combination relation, recognition loss selection is minimized on a verification set, and the device layout is improved when sparse Increased when geometry topology dominates 。
  6. 6. An intelligent monitoring method for partial discharge anomalies in electrical equipment according to claim 1, characterized in that the attention weights are dynamically adjusted according to the complexity of the input signal, said complexity being estimated from spectral entropy and/or time domain variance.
  7. 7. An electrical equipment partial discharge anomaly intelligent monitoring method according to claim 1, wherein the position information of the identification output is given in a GIS shell coordinate system based on the time difference of arrival TDoA positioning and/or energy imaging, and the confidence evaluation is carried out on the positioning result.
  8. 8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and the method is characterized in that the processor implements the steps of the method for monitoring abnormal partial discharge of an electrical device according to any one of claims 1 to 7 when executing the computer program.
  9. 9. A computer readable storage medium having stored thereon a computer program, wherein the computer program when executed by a processor implements the steps of a method for intelligent monitoring of partial discharge anomalies in an electrical installation according to any one of claims 1 to 7.

Description

Intelligent monitoring method, equipment and medium for partial discharge abnormality of electrical equipment Technical Field The invention relates to the technical field of power system state monitoring and fault diagnosis, in particular to an intelligent monitoring method, equipment and medium for partial discharge abnormality of electrical equipment. Background With the popularization of GIS on-line state monitoring, partial discharge identification is developed by combining ultra-high frequency (UHF) electromagnetic measurement with PRPD/time-frequency spectrum, and the application is expanded under IECTS62478 and IEC60270 frames. However, in the actual operation environment of the transformer substation, various discharges such as metal particles, floating electrodes, surface pollution/defects of insulators and the like can coexist, electromagnetic radiation is overlapped into mixed waveforms after being transmitted and reflected in a shell, so that mode boundary blurring and misjudgment risk are increased, in order to improve robustness under a noise and multisource mixed working condition, in the prior art, on one hand, UHF arrays are arranged at basin-type insulators and combined with acoustic/optical channels to obtain complementary characteristics, and on the other hand, a deep learning framework based on attention and characteristic pyramids is introduced to extract judgment modes from a PRPD/time-frequency diagram. However, when the phase/spectrum characteristics of different defects are converged or the sampling rate of the cross-sensor is not synchronous with the clock, the fusion information is easy to distort, and meanwhile, the field EMI and the operation noise further increase the difficulty of source separation and positioning, and influence the identification stability. Disclosure of Invention The present invention has been made in view of the above-described problems occurring in the prior art. The invention provides an intelligent monitoring method, equipment and medium for partial discharge abnormality of electrical equipment, which solve the problems of reduced identification accuracy caused by cross interference of multisource partial discharge signals in GIS equipment, similar phase characteristics, multi-sensor data fusion and limitation in complex noise environments of the existing method. In order to solve the technical problems, the invention provides the following technical scheme: In a first aspect, an embodiment of the present invention provides a method for intelligently monitoring abnormal partial discharge of an electrical device, including: S1, collecting, namely arranging at least two UHF sensors on a GIS shell and acquiring partial discharge electromagnetic signals, and optionally arranging at least one acoustic emission sensor to acquire acoustic signals; s2, synchronizing and calibrating, namely performing time synchronization and amplitude calibration on different sensor channels; S3, preprocessing, namely performing time domain/frequency domain or time-frequency domain transformation on the acquired signals based on PRPD phase references to obtain a feature map; s4, extracting the characteristics, namely inputting the characteristic map into a multi-scale residual characteristic extraction network to obtain multi-scale characteristics; S5, attention weighting, namely weighting the multi-scale feature by applying spatial-channel joint attention to obtain a weighted feature; S6, fusion and identification, namely, carrying out multi-sensor data fusion based on the weighted characteristics and outputting partial discharge type and position information. The invention relates to an intelligent monitoring method for partial discharge abnormality of electrical equipment, which comprises the following steps of preprocessing an acquired signal, performing short-time Fourier transform (STFT) or Continuous Wavelet Transform (CWT) to obtain an amplitude-phase time-frequency diagram, and generating a PRPD diagram through power frequency voltage or equivalent phase reference. The multi-scale residual feature extraction network comprises residual blocks with at least three different convolution kernel sizes, wherein the residual blocks are connected in a jumping manner and comprise a convolution layer and a batch normalization layer, and the convolution layer and the batch normalization layer are used for simultaneously extracting local textures and global modes. As a preferable scheme of the intelligent monitoring method for the partial discharge abnormality of the electrical equipment, the attention weighting comprises non-local attention or self-attention so as to model the correlation among the positions of the feature map, and the weighting is distributed on multiple scales by combining a feature pyramid. The invention also comprises a phase sensing mask filtering step, wherein the phase sensing mask filtering step is used for masking and suppressing noise on time frequency c