CN-121982359-A - Quick identification method for partial discharge mode of power transformer

Abstract

The invention discloses a rapid identification method of a partial discharge mode of a power transformer, which comprises the steps of collecting partial discharge PD time domain signals, carrying out phase space reconstruction through an improved C-C parameter optimization algorithm, mapping one-dimensional time sequence data into three-dimensional point clouds with obvious topological structure differences, inputting the three-dimensional point clouds into a pre-constructed TopologyNet neural network model adapting to partial discharge signal characteristics to generate a continuous image, inputting the continuous image into a support vector machine SVM classifier, and finally classifying the topological characteristics extracted by TopologyNet. According to the invention, the phase space reconstruction parameters are optimized by improving the C-C algorithm, and the topological characteristics are extracted by combining TopologyNet high efficiency, so that the problems of insufficient characteristic representation and low efficiency of the traditional method are solved, and a practical and effective technical approach is provided for the partial discharge on-line detection of the transformer.

Inventors

• LI ZHONG
• ZHANG LINFENG
• ZHANG WEIHUA
• ZHANG KE
• WANG YU

Assignees

• 华北电力大学（保定）

Dates

Publication Date

20260505

Application Date

20251210

Claims (3)

1. 1. A method for rapidly identifying a partial discharge mode of a power transformer, the method comprising: Step 1, collecting partial discharge PD time domain signals, carrying out phase space reconstruction through an improved C-C parameter optimization algorithm, and mapping one-dimensional time sequence data into three-dimensional point clouds with obvious topological structure differences; Step 2, inputting the three-dimensional point cloud into a TopologyNet neural network model which is constructed in advance and is adapted to the partial discharge signal characteristics, and generating a continuous image; step 3, inputting the continuous image into a Support Vector Machine (SVM) classifier to realize final classification of the topology features extracted by TopologyNet; wherein the improved C-C parameter optimization algorithm is as follows: S101, consider a time sequence x= { X i |i=1, 2, the term, N }, where X i is the amplitude of the partial discharge time domain signal at the ith sampling moment, N is the total sampling point number of the original time domain signal, and the embedding dimension M is a phase space reconstruction parameter with the time delay τ, and perform phase space reconstruction on the time sequence to obtain a phase space point set x= { X i |i=1, 2, the term, M }, where X i is a single data point in the phase space, M is the total number of phase space valid data points, and at this time, the correlation integral of the embedding time sequence is: wherein r is a search radius (r > 0), d ij ＝||X i -X j ｜｜ ∞ represents an infinite norm distance between two phase points, θ is a Heaviside step function, and the following conditions are satisfied: Providing a unified topological feature quantization basis for subsequent statistic calculation; S102, calculating core statistics S 1 of a traditional C-C algorithm, wherein the core statistics show autocorrelation of a time sequence, and the expression is as follows: S 1 (m,N,r,τ)＝C(m,N,r,τ)-C m (1,N,r,θ) Wherein C m (1, N, r, τ) represents the result of performing the power of m operation on the correlation integral C (m, N, r, τ) of the one-dimensional time series; s103, adding statistics S 2 , and improving the operation speed by adopting a block average strategy, wherein the expression is as follows: Wherein t is the number of blocks, For the associated integration of the s-th block of data in m-dimensional phase space, For the association integration of the s-th block data in a one-dimensional space, local noise interference of the partial discharge signal is counteracted by block average, so that statistics can reflect global dynamic characteristics more stably; S104, for measuring the deviation of S 2 under different tau, selecting the maximum value and the minimum value corresponding to tau to define the difference, and the expression is as follows: ΔS(m,τ)=max{S(m,N,r k ,τ)}-min{S(m,N,r k ,τ)} Wherein, r k =kσ/2, σ is the standard deviation of the time sequence x= { x i }, the subjectivity of parameter value in the traditional algorithm is eliminated, the consistency of parameter estimation of different partial discharge signals is ensured, and the accuracy of deviation measurement is improved; S105, reasonably estimating m, N and r k according to BDS statistical conclusion, and detecting Determining the optimal time delay tau by the first local minimum point of the curve to The periodic node of (a) is used as an optimal embedding window tau w to reversely calculate the embedding dimension m, so that misjudgment caused by single statistic and single embedding dimension is avoided, experience deriving errors are eliminated, and decoupling accurate estimation of tau and m is realized.
2. 2. The method for quickly identifying the partial discharge mode of the power transformer according to claim 1, wherein the step 2 specifically comprises: s201, generating a continuous scatter diagram by constructing nested Rips complex shape record topology feature appearance and disappearance time, wherein Rips complex shape is connected with adjacent point cloud by setting radius threshold r, which is mathematically defined as Wherein sigma represents a non-empty subset of the phase space point set X and is a basic topological structure unit forming Rips complex, and xi ij represents Euclidean distance between points X i and X j , when r is gradually increased, a topological structure is generated, and the birth radius b and the death radius d of the continuous scatter diagram mark feature are generated; S202, converting the continuous scatter diagram into a 50×50 continuous image by using a Gaussian function f (x, y) with a standard deviation of 1×10 -2 , wherein the mathematical definition of the continuous image is as follows S203, inputting n multiplied by 3 dimensional point cloud tensors, extracting local features through a plurality of EdgeConv layers, and constructing a neighborhood graph association point and a neighborhood point to capture a geometric relationship; s204, extracting and maximally pooling the features by each layer EdgeConv, and sequentially generating n multiplied by 64, n multiplied by 128 and n multiplied by 256 intermediate vectors to enhance the robustness of the features; S205, splicing the intermediate vectors to form n multiplied by 448 dimensional features, mapping the features to 2500 dimensions through global maximum pooling and full connection layers, and corresponding to 50 multiplied by 50 continuous image pixel values; S206, outputting 50 multiplied by 50 continuous images, adopting a multi-task learning strategy to simultaneously predict topological characteristics of an H1 ring structure and an H2 cavity structure, sharing parameters in the first layers of the network, and respectively predicting in the last two layers; s207, defining a micro-topology loss function: wherein: the mean square error of the persistence image and the true persistence image is predicted for the loop structure, The mean square error of the continuous image and the real continuous image is predicted for the cavity structure, wherein lambda=1 is a super parameter; s208, training a model by adopting an Adam optimizer, setting the initial learning rate to be 1 multiplied by 10 < -3 >, attenuating the iterative learning rate to be 0.5 times of the original value every 20 rounds, training 50 rounds altogether, combining the advantages of momentum gradient decline and self-adaptive learning rate, enabling the Adam optimizer to self-adaptively adjust the learning rate of each parameter, accelerating the model convergence speed, setting the batch size to be 32, taking 1 by the weight lambda of the multi-task loss function, enabling the model to accurately learn the topological characteristics of point cloud through repeated iterative training, and generating a prediction result which is highly consistent with a real continuous image.
3. 3. The method for quickly identifying the partial discharge mode of the power transformer according to claim 1, wherein the step 3 specifically comprises: s301, inputting cloud samples to be measured into TopologyNet models, extracting 50 multiplied by 50 continuous image features, expanding the continuous image features into 2500-dimensional vectors, and then inputting SVM classifiers; s302, setting a penalty parameter C=10, an RBF kernel parameter gamma=0.1, and balancing fitting and generalization capabilities; s303, training by adopting batch processing of 128 samples, and improving efficiency and characteristic learning effects; S304, the data set is divided into a training set and a testing set according to the ratio of 8:2, and the average value is obtained by repeating the experiment for 3 times, so that the random error is reduced.

Description

Quick identification method for partial discharge mode of power transformer Technical Field The invention belongs to the technical field of power equipment fault diagnosis, and particularly relates to a method for rapidly identifying a partial discharge mode of a power transformer. Background The power transformer is a key basic device in the power grid, and the safe and stable operation of the power transformer directly affects the reliability of the power system. Partial discharge (PARTIALDISCHARGE, PD) is an important sign of transformer insulation degradation, and timely and accurately identifying a partial discharge mode is important for equipment state evaluation and fault early warning. However, partial discharge signals have strong non-linearity, time-varying characteristics, and high similarity between different types of discharge signals, which presents challenges for pattern recognition. The traditional partial discharge pattern recognition method mainly relies on manually extracting features such as statistical features, phase distribution features and the like, and then classifying the features by combining machine learning models such as a Support Vector Machine (SVM), an artificial neural network and the like. However, the manual feature extraction often depends on expert experience, and has the problems of insufficient feature characterization capability and poor generalization, and has low calculation efficiency, so that the real-time requirement of online monitoring is difficult to meet. In recent years, topology data analysis (TopologicalDataAnalysis, TDA) has received attention as a new method of analyzing nonlinear dynamic systems. The core idea of the TDA is that a one-dimensional time sequence signal is mapped into a high-dimensional point cloud through phase space reconstruction, and then the topological structure of the TDA is analyzed to capture a dynamic evolution rule which is difficult to describe by a traditional method. Literature studies have shown that TDA is used for partial discharge point cloud feature extraction to better describe the nonlinear behavior of the discharge than statistical features. However, the conventional phase space reconstruction method has the difficulty in determining parameters such as delay time, embedding dimension and the like, and for strong nonlinear signals such as partial discharge, reconstruction deviation is easy to generate by adopting a conventional parameter selection method such as a small window method, a C-C algorithm and the like, so that the reconstructed point cloud cannot accurately reflect the dynamics characteristics of the original signals, and the physical meaning and classification precision of the subsequent topological characteristics are affected. In addition, the TDA method has high calculation complexity and long time consumption in the feature extraction process, and is difficult to meet the real-time requirement of online monitoring. For example, the traditional TDA method may take tens of minutes to extract topological features from each sample, and cannot accommodate the requirement of rapid diagnosis in the field. Accordingly, in view of the above-described shortcomings of the prior art, there is a need to propose an improved method for identifying partial discharge patterns of a power transformer to improve identification accuracy and efficiency. Disclosure of Invention The invention aims to provide a rapid identification method for a partial discharge mode of a power transformer, which aims to solve the problems of insufficient extraction of partial discharge characteristics, low identification precision and low efficiency in the prior art. In order to achieve the above purpose, the invention provides a method for rapidly identifying a partial discharge mode of a power transformer, which comprises the following steps: Step 1, collecting partial discharge PD time domain signals, carrying out phase space reconstruction through an improved C-C parameter optimization algorithm, and mapping one-dimensional time sequence data into three-dimensional point clouds with obvious topological structure differences; Step 2, inputting the three-dimensional point cloud into a TopologyNet neural network model which is constructed in advance and is adapted to the partial discharge signal characteristics, and generating a continuous image; step 3, inputting the continuous image into a Support Vector Machine (SVM) classifier to realize final classification of the topology features extracted by TopologyNet; wherein the improved C-C parameter optimization algorithm is as follows: S101, consider a time sequence x= { X i |i=1, 2, the term, N }, where X i is the amplitude of the partial discharge time domain signal at the ith sampling moment, N is the total sampling point number of the original time domain signal, and the embedding dimension M is a phase space reconstruction parameter with the time delay τ, and perform phase space reconst