Search

CN-122020297-A - Transformer core loosening mechanical fault diagnosis method based on acoustic signal ACGAN-transducer

CN122020297ACN 122020297 ACN122020297 ACN 122020297ACN-122020297-A

Abstract

Aiming at the technical problems that a traditional Transformer fault detection sensor is difficult to install, field fault data are difficult to obtain in a large amount, deep learning model training data are insufficient and fitting is easy to be excessively performed, and the like, the invention utilizes frequency domain gram angle field patterns and ACGAN to generate fault samples with high similarity with original sample characteristics through the Transformer operation voiceprint data, effectively solves the problem of unbalanced sample of the Transformer core fault data, avoids the defects of insufficient feature extraction, easy fitting and the like, adopts a transducer encoder as a classifier, effectively captures global time sequence features and deep relevant features of fault signals by means of an attention mechanism, obviously improves the generalization capability and the diagnosis precision of a diagnosis model compared with the traditional model such as CNN, has clear overall method flow, is applicable to the Transformer sample unbalanced acoustic fault diagnosis scene, and provides reliability for the safe and stable operation of a power system Transformer.

Inventors

  • ZHANG HONGLI
  • WANG NINI
  • MA PING
  • WANG CONG
  • LI XINKAI
  • MENG YUE

Assignees

  • 新疆大学

Dates

Publication Date
20260512
Application Date
20260127

Claims (10)

  1. 1. The Transformer core loosening mechanical fault diagnosis method based on the acoustic signal ACGAN-transducer is characterized by comprising the following specific steps of: S1, acquiring and preprocessing data, namely acquiring operation state voiceprint data of a transformer core under different working conditions when the transformer core is loose and fails, wherein the operation state comprises a state that a clamping piece of the transformer core is normally compressed, and a state that no load, rated capacity, B-phase unbalanced load and 3 rd harmonic are contained when 50% of compression force fails; s2, sample feature transformation, namely constructing a frequency domain gram difference angle field (FrequencyDomainGramianangulardifferentfield, FGADF) two-dimensional feature map for the transformer voiceprint operation data obtained in the step S1; s3, constructing a ACGAN sample expansion unit, namely constructing a ACGAN model, wherein the ACGAN model comprises a generator G and a discriminator D, inputting the sample characteristic map obtained in the S2 and a corresponding class label into the ACGAN model, training the generator G to learn the data distribution characteristics of the sample, training the discriminator D to judge the authenticity and class accuracy of the generated sample, and finally outputting a new sample similar to the original sample data characteristics; S4, constructing a classification model based on a transducer, namely constructing a transducer classification model, wherein the model consists of a Patch embedded layer, a class encoder, a position encoder, multi-layer stacked encoder blocks and a classification layer, each encoder block comprises a multi-head self-attention layer and a feedforward neural network layer, and residual error connection and layer normalization processing are adopted; S5, a Transformer fault diagnosis step, namely dividing the extended training set obtained in the S3 into a training set, a verification set and a test set according to the ratio of 4:4:2, inputting the training set into a transducer encoder classification model for training, adjusting model parameters through the verification set, stopping training when the diagnosis accuracy of the model on the verification set is more than 90%, obtaining a trained fault diagnosis model, and adopting the fault diagnosis model to test on the test set and outputting a diagnosis result.
  2. 2. The method for diagnosing the mechanical failure of the Transformer core loosening based on the acoustic signal ACGAN-transducer according to claim 1, wherein in the step S1, the specific process of data acquisition and segmentation processing is that the data acquisition frequency is 65536Hz, the acquired data length in each state is 787200, the original data in each state is divided into 600 non-overlapping subsamples according to the 2-time period length of the signal, 200 samples are selected from the B-phase unbalanced load and 600 original samples containing 3 rd harmonics at random, and a few-sample unbalanced data set is constructed together with 600 samples of other three types of samples.
  3. 3. The Transformer core loosening mechanical fault diagnosis method based on the acoustic signal ACGAN-transducer according to claim 1 is characterized in that in the step S2, the specific step of constructing FGADF is that after 2.1 pre-emphasizes a one-dimensional voiceprint sample, the one-dimensional voiceprint sample is converted into a frequency domain range by utilizing Fourier transformation, 2.2 normalized compression of [0,1] intervals is carried out on frequency domain data, 2.3 the frequency domain data is represented by angle and amplitude variables by utilizing polar coordinate transformation, and 2.4 matrix element values are obtained by calculating a trigonometric function to form a two-dimensional characteristic diagram.
  4. 4. The method for diagnosing a mechanical failure of a Transformer core loosening based on an acoustic signal ACGAN-transducer according to claim 1 or 3, wherein after normalized compression is performed on the frequency domain data in step S2, the method further comprises the step of performing validity check on the compressed frequency domain data, dividing the frequency domain data into a plurality of continuous data segments according to a preset rule, respectively checking whether a numerical fluctuation range of each data segment is within a preset reasonable interval, removing the data segments with fluctuation ranges exceeding the reasonable interval, and completing the data segments by adopting a mean value of adjacent valid data segments.
  5. 5. The method for diagnosing mechanical failure of Transformer core loosening based on acoustic signal ACGAN-transducer according to claim 1, wherein in step S3, generator G of ACGAN model is composed of convolution layer matching batch normalization and Relu activation function, input is target data label and random gaussian noise, output is generated as sample, and discriminator D is composed of convolution layer matching batch normalization and LeakyRelu activation function, and each convolution layer introduces spectral normalization while global average pooling layer is set, and output of discriminator includes scalar probability of input sample from real dataset and class probability distribution of sample.
  6. 6. The method for diagnosing a mechanical failure of a Transformer core loosening based on an acoustic signal ACGAN-transducer as claimed in claim 1, wherein in the step S3, the objective function of ACGAN is a conditional maximum and minimum game, including a loss with a true and false discrimination and a loss based on class classification, and the total loss is a weighted sum of two losses: wherein For the expectations computed under the true sample distribution, To identify the probability of a true sample to be true for the arbiter, For the expectations computed under a random gaussian noise distribution, The probability that the generated sample will be identified as true for the arbiter is the loss based on category classification: wherein Currently identifying the class of the real sample for the arbiter as Is a function of the probability of (1), Generating a sample for the current identification of the arbiter of the class The objective of the arbiter is to maximize both losses, both true and false and classification, so the total losses are: the overall goal of the generator is to minimize the corresponding loss to let the generated samples be judged true and fit the target class, so the overall loss is: in the following The weighting coefficients are lost.
  7. 7. The method for diagnosing the mechanical fault of the Transformer core loosening based on the acoustic signal ACGAN-transducer according to claim 1, wherein in the step S4, the Patch embedding layer in the transducer classification model is implemented by a convolution layer with a convolution kernel size of Patchsize, so that an input image with a size of [3, long, wide ] is converted into a sequence form, the sequence length is the number of images split according to Patchsize, the dimension of each sequence element is a word dimension, the class encoder adds a group of 1-dimensional vectors to represent class information on the basis of the output of the Patch embedding layer, and the position encoder represents the position information in a matrix form.
  8. 8. The method for diagnosing a loosening mechanical fault of a Transformer core based on an acoustic signal ACGAN-transducer according to claim 1, wherein in the step 4, each encoder in the transducer classification model can be understood as two sub-layer structures, the first sub-layer structure is a multi-head attention mechanism layer and normalization process and one residual connection, and the second sub-layer structure is a feedforward neural network layer and normalization process and one residual connection. The processing procedure of the multi-head attention mechanism comprises the steps of firstly carrying out weight mapping on input features to obtain inquiry Key and key Value of Three kinds of characteristics, and then pass through the formula Calculation of wherein 、 、 Respectively representing weight matrix Q, K, V, respectively dividing three types of features into several sub-feature sets, and making them pass through the formula Calculating the attention output of each sub-feature set, and finally passing through the formula Connecting and mapping the attention outputs of all sub-feature sets to obtain a final multi-head attention result, wherein The number of heads is indicated and the number of heads is indicated, Representing the matrix of output weights and, For connecting together the outputs of multiple heads, Is a function of attention. Each multi-head attention machine sublayer and the feedforward neural network sublayer carry out residual connection on input, then carry out layer normalization processing on the input and output the input, and the first layer normalization processing is carried out The output of the sublayer structure of the multi-head attention machine is First, the The output of the individual feed-forward neural network sublayer structures is Wherein , Indicating the number of encoder stacks, Representing a layer normalization process, FFN represents a feed forward neural network process, Represent the first Layer encoder output, final class output is formulated Extracting the feature corresponding to the class mark of the last encoder module, wherein The category of the last base module is marked.
  9. 9. The method for diagnosing the mechanical fault of the Transformer core loosening based on the acoustic signal ACGAN-transducer according to claim 1 or 8, wherein the step S4 is characterized by further comprising an optimization step of optimizing the attention weight when the multi-head self-attention layer performs the feature processing, namely, after the initial attention weight is calculated, identifying and suppressing abnormal associated features with excessively low weight values, strengthening the weight ratio corresponding to similar fault features, and performing feature fusion based on the optimized attention weight.
  10. 10. The method for diagnosing a loosening mechanical failure of a Transformer core based on an acoustic signal ACGAN-transducer as claimed in claim 1, wherein in said step S4, the classification layer is formulated by the formula Implementing classification in which Is a weight matrix of the category classification layer; is a bias vector; the function is used to normalize the input vector so that each element in the output vector is in the [0,1] range and the sum of all elements is 1.

Description

Transformer core loosening mechanical fault diagnosis method based on acoustic signal ACGAN-transducer Technical Field The invention relates to the technical field of power equipment fault diagnosis, in particular to a Transformer core loosening mechanical fault diagnosis method based on an acoustic signal ACGAN-transducer. Background The power transformer is used as core junction equipment of a power system, the key tasks of power transmission and voltage conversion are borne, the stable operation of the power transformer is directly related to the safety and reliability of a power grid and the continuity of energy supply, the iron core is used as a core framework of the transformer and is a core carrier for magnetic flux conversion and current communication, the structural integrity and the operation state of the power transformer are critical to the working performance of the transformer, once the iron core has loose faults, the operational noise of the transformer is increased, the energy consumption is increased, the interlocking problems such as iron core overheat and winding deformation can be caused, even the transformer is stopped or the power grid is disassembled when serious, and the important hidden trouble is buried for the safe operation of the power system, so that the research of the accurate diagnosis technology of the transformer iron core loose faults is developed, and the power transformer has important engineering significance for early warning faults, reducing the operation and maintenance cost and guaranteeing the stability of the power grid. In the field of transformer fault diagnosis, the existing diagnosis methods can be divided into two types of internal signal detection and external signal detection according to data sources. The internal signal detection needs to install a sensor in the transformer, has the problems of complex installation flow, large transformation difficulty, complicated data statistics and analysis and the like, and possibly influences the original running state of the transformer, and in the external signal, the vibration signal and the ultrasonic image detection mostly adopt a contact type measurement mode, are easily interfered by equipment running vibration and site environment, so that the accuracy of a diagnosis result is limited, and the infrared image detection equipment has higher cost and is difficult to popularize and apply on a large scale. As a non-contact external signal, the voiceprint signal has the advantages of convenience in acquisition, high economy, no need of shutdown reconstruction and the like, and becomes a research hotspot for transformer fault diagnosis. However, the traditional diagnosis technology based on the voiceprint signals still has the obvious defects that firstly, fault data acquisition is difficult, the design and manufacture standard of a transformer is high, the fault probability is low, the fault data base number is small, the fault data is dependent on post analysis, the acquisition is required to be stopped and wiring is not repeated, faults such as winding deformation and the like are slow in development and strong in burst property, the fault process data are difficult to capture, a diagnosis model based on data driving faces unbalanced dilemma of fault samples, secondly, the traditional deep learning model has weak capturing capability of global time sequence characteristics and deep correlation characteristics of the fault signals, generalization performance is poor and fitting is easy under the unbalanced sample scene, and high-precision diagnosis requirements are difficult to meet, and thirdly, part of the method needs manual intervention characteristic selection, is low in efficiency and strong in subjectivity, and is difficult to adapt to the diagnosis requirements of complex field working conditions. In summary, the existing Transformer core loosening fault diagnosis technology has the problems of difficult data acquisition, weak model generalization capability, manual feature selection dependence, limited detection mode and the like, and cannot effectively meet the actual requirements of a power system on high-precision and convenient fault diagnosis under a sample imbalance scene, so that the development of a Transformer core loosening mechanical fault diagnosis method based on acoustic signals ACGAN-transducer is particularly important. Disclosure of Invention The invention aims to make up the defects of the prior art, and provides a Transformer core loosening mechanical fault diagnosis method based on acoustic signals ACGAN-transducer, which can collect the operation acoustic signals of a Transformer as detection information through an acoustic sensor to carry out fault diagnosis; the method comprises the steps of constructing FGADF two-dimensional time-frequency spectrum by utilizing frequency domain Graham differential angle field, enriching signal characteristics, constructing sample imbalance fault d