CN-121479286-B - Method and system for enhancing data of discharge impact signals in oil based on mixed time-frequency convolution

CN121479286BCN 121479286 BCN121479286 BCN 121479286BCN-121479286-B

Abstract

The invention belongs to the technical field of high-voltage equipment monitoring and processing, and discloses a method and a system for enhancing data of discharge impact signals in oil based on mixed time-frequency convolution, wherein the method carries out two-dimensional processing on collected discharge impact signals in oil to obtain a two-dimensional time-frequency diagram; training a generated countermeasure network by using a two-dimensional time-frequency diagram, wherein the generated countermeasure network comprises an encoder, a decoder and a discriminator, and acquiring the generated time-frequency diagram by using the trained decoder as a generator to realize data enhancement of the discharge impact signal in oil. The invention can enhance the data of the discharge impact signal in the oil, obviously improve the fidelity of the time-frequency characteristics of the generated sample, and can realize the fault identification of the discharge type.

Inventors

Xiang Yadong
HU QING
LIU RONGHUI
LI YUAN
ZHANG CHENGJIE
SHI YI
JIANG HUAN

Assignees

国网四川省电力公司宜宾供电公司

Dates

Publication Date: 20260505
Application Date: 20260109

Claims (8)

1. The method for enhancing the data of the discharge impact signal in the oil based on the mixed time-frequency convolution is characterized by comprising the following steps of: s1, carrying out two-dimensional processing on the collected oil discharge impact signals to obtain a two-dimensional time-frequency diagram; s2, training a generated countermeasure network by utilizing a two-dimensional time-frequency diagram, wherein the generated countermeasure network comprises an encoder, a decoder and a discriminator; The encoder takes a two-dimensional time-frequency diagram as input to acquire corresponding potential space distribution, the encoder is formed by cascading one mixed time-frequency convolution module or more than two mixed time-frequency convolution modules, the mixed time-frequency convolution module comprises a first filling unit, a convolution unit, a first fusion unit and a downsampling unit, the first filling unit carries out time-frequency axis heterogeneous filling on the two-dimensional time-frequency diagram by adopting corresponding filling strategies along a time axis and a frequency axis respectively, the first filling unit specifically adopts zero value filling along the time axis direction of the two-dimensional time-frequency diagram and adopts interpolation filling along the frequency axis direction of the two-dimensional time-frequency diagram, the convolution unit comprises a group of parallel heterogeneous convolution kernels, carries out convolution processing on the two-dimensional time-frequency diagram after heterogeneous filling by adopting transverse convolution kernels and longitudinal convolution kernels with different sizes respectively, and the first fusion unit carries out weighted fusion on output characteristics of each convolution kernel to obtain fused time-frequency characteristics; The decoder takes output of an encoder and noise as input to obtain a reconstructed time-frequency diagram and generate a time-frequency diagram, is symmetrically arranged relative to the encoder and is formed by cascading one mixed time-frequency deconvolution module or more than two mixed time-frequency deconvolution modules, wherein the mixed time-frequency deconvolution module comprises a second filling unit, a deconvolution unit, a second fusion unit and an up-sampling unit, the second filling unit carries out time-frequency axis heterogeneous filling on noise input into the decoder by adopting corresponding filling strategies along a time axis and a frequency axis respectively, the deconvolution unit comprises a group of parallel heterogeneous deconvolution cores, the heterogeneous filled noise is reconstructed by the transverse deconvolution cores and the longitudinal deconvolution cores with different sizes respectively, the second fusion unit carries out weighted fusion on output characteristics of each deconvolution core to obtain fusion reconstruction characteristics, and the up-sampling unit is used for carrying out up-sampling processing on the fusion reconstruction characteristics and has the same filling strategy as the first filling unit; The encoder and the decoder are formed based on mixed time-frequency convolution; And S3, acquiring a generated time-frequency diagram by using the trained decoder as a generator, and realizing data enhancement of the discharge impact signal in the oil.
2. The method for enhancing the data of the oil discharge impact signals based on the mixed time-frequency convolution according to claim 1 is characterized by comprising the step S1 of carrying out two-dimensional processing on the collected oil discharge impact signals by utilizing short-time Fourier transform, wherein the short-time Fourier transform is continuous Fourier transform or discrete Fourier transform.
3. The method for enhancing discharge-in-oil impact signal data based on mixed time-frequency convolution according to claim 1, wherein the final output of the encoder is the mean value of the potential spatial variable distribution of the two-dimensional time-frequency diagram And logarithmic variance Thereby mapping the input time-frequency diagram into a continuous and smooth potential distribution : ; Wherein, the Representing an encoder; Representing a two-dimensional time-frequency diagram; From potential distribution Sampling seed noise z, as input to the decoder, noise satisfies: ; Wherein I represents a unit diagonal matrix; the noise input to the decoder includes: ; ; wherein the noise For obtaining reconstructed time-frequency map by decoder, noise The decoder is used for obtaining the generated time-frequency diagram; Representing standard deviation; indicating a standard normal distribution noise, and by which is meant an element-wise multiplication.
4. A method of enhancing discharge in oil shock signal data based on mixed time-frequency convolution as claimed in claim 3, wherein the method is directed to input noise And Reconstructing a time-frequency diagram by the decoder output And generating a time-frequency diagram : ; ; Wherein, the Representing the decoder.
5. The method for enhancing the data of the discharge shock signal in oil based on the mixed time-frequency convolution as claimed in claim 4, wherein the discriminator comprises a feature extraction module and a discrimination output module, and the input sample of the feature extraction module is a true input two-dimensional time-frequency diagram Reconstructing a time-frequency diagram And generating a time-frequency diagram The feature extraction module adopts four third convolution units which are sequentially arranged, wherein each third convolution unit comprises a convolution layer, a normalization layer and an activation function, and the discrimination output module comprises a first full-connection layer and is used for outputting true and false scores representing input samples.
6. The method for enhancing the data of the discharge impact signal in oil based on the mixed time-frequency convolution according to claim 5, wherein the discrimination output module further comprises a second full connection layer for outputting the prediction classification type.
7. The method for enhancing the data of the discharge shock signal in oil based on the mixed time-frequency convolution according to claim 6, wherein the generation of the countermeasure network is trained, and a multi-objective joint loss function is set Including varying lower bound loss Countering losses Auxiliary classification loss Three parts: ; ; ; ; Wherein, the Representing mathematical expectations of the variables; a specific gravity coefficient representing a KL divergence constraint term; Representing the calculation of a KL divergence function; Representing prior distribution of potential variable z, min G representing minimized solution by taking parameters of a decoder as optimization variables, and max D representing maximized solution by taking parameters of a discriminator as optimization variables; representing the process of obtaining the true and false score of the input sample by the discriminator; The representation arbiter obtains a predictive classification process, Representing the type of prediction classification, Representing the input sample real type label; Is that Or (b) ; Representing a pair of slave real data distributions A sample obtained by sampling and a mathematical expectation calculated by a corresponding label thereof; representing a priori distributions of pairs of slave latent variables A mathematical expectation calculated from the sampled latent variables and type tags in the real data distribution; , , Is a super parameter.
8. A system for implementing the method for enhancing discharge shock signal data in oil based on mixed time-frequency convolution as defined in any one of claims 1 to 7, comprising: the pretreatment module is used for carrying out two-dimensional treatment on the collected oil discharge impact signals to obtain a two-dimensional time-frequency diagram; The method comprises the steps of generating a countermeasure network, wherein the countermeasure network comprises an encoder, a decoder and a discriminator, the encoder takes a two-dimensional time-frequency diagram as input to acquire corresponding potential space distribution, the decoder takes output and noise of the encoder as input to acquire a reconstructed time-frequency diagram and generate a time-frequency diagram, the discriminator is used for discriminating authenticity of an input image, and the encoder and the decoder are formed based on mixed time-frequency convolution.

Description

Method and system for enhancing data of discharge impact signals in oil based on mixed time-frequency convolution Technical Field The invention belongs to the technical field of high-voltage equipment monitoring and processing, relates to enhancement of discharge impact signal data in transformer oil, and particularly relates to a method and a system for enhancing discharge impact signal data in oil based on mixed time-frequency convolution. Background The oil immersed transformer is used as key equipment of a modern power system, plays important roles of voltage conversion and electric energy transmission, and the safety and stability performance of the oil immersed transformer directly determines the overall operation reliability of the power system. As the power demand continues to increase, the load carried by the transformer increases increasingly, and the influence of strong coupling factors such as high voltage, strong electric field, high temperature, strong vibration and the like in the transformer on the insulating property of the transformer becomes more remarkable, so that arc faults in the transformer frequently occur. When an arc discharge fault occurs, rapid injection of external energy causes rapid expansion and temperature rise of an arc discharge channel, and a huge pressure difference formed in a short time causes the pressure level in the oil tank to rise suddenly. If the local stress exceeds the limit of the strength of the tank material, the tank will break and the tank structure will be destroyed. The discharge accident in the transformer oil has extremely strong contingency and instantaneity, and a sufficient quantity of complete discharge impact signal samples in the oil are difficult to collect aiming at different fault reasons and working conditions. This "small sample" problem severely constrains the development of deep learning based fault diagnosis and early warning models that typically rely on extensive amounts of high quality data to adequately train, avoid overfitting, and ensure their generalization capability. At present, a research method for solving the problem of small samples of the discharge impact signals in oil mainly comprises traditional data enhancement and generation type models. Although the traditional method such as adding noise, time sequence distortion and the like is simple and easy to implement, the generated signals are often subjected to simple transformation only on a time domain level, and abundant time-frequency characteristic changes in a discharge physical process are difficult to effectively simulate, so that insufficient data diversity is caused, and the model performance is limited. Based on the method for generating the depth generation model such as the countermeasure network, the used simple convolution mode cannot be used for extracting or generating independent characteristics according to the inherent difference of the time domain and the frequency domain characteristics, the training process is unstable, mode collapse or over fitting is easy to occur under the condition of extremely deficient data volume, the generated data authenticity and diversity are difficult to ensure, and deep fault characteristic information contained in a small sample cannot be fully mined. In addition, the existing data enhancement method cannot fully simulate the inherent mode of the impact signal on time-frequency distribution in the arc discharge process, cannot effectively and pertinently extract the differentiation characteristics of the time domain and the frequency domain of the discharge impact signal, causes deviation between enhanced data and the physical mechanism of the real discharge signal, and is difficult to be used for constructing a diagnosis model with high robustness. Disclosure of Invention Aiming at the problem that sufficient and complete signal characteristics of arc fault discharge impact signals in transformer oil are difficult to collect under different typical faults and working conditions, and the use requirement of a large amount of high-quality data in fault identification deep learning research cannot be met, the invention aims to provide a method and a system for enhancing discharge impact signal data in oil based on mixed time-frequency convolution, which give consideration to the time domain-frequency domain independent characteristics of the discharge impact signals, reconstruct the time domain characteristics and the frequency domain characteristics of a time-frequency diagram simultaneously in the data generation process through the time-frequency mixed convolution, and enhance data through generation of an countermeasure network. The invention adopts the technical idea that a variational self-coding generation countermeasure Network (Variational Autoencoder GENERATIVE ADVERSARIAL Network, VAE-GAN) is taken as a basic model, and a mixed time-frequency convolution structure is introduced to extract and generate time domain a