Zero abnormal sample fault diagnosis method and system for low-frequency power transformer

Abstract

The invention discloses a zero abnormal sample fault diagnosis method and system for a low-frequency power transformer, and belongs to the technical field of low-frequency power transformer operation state detection. According to the zero abnormal sample fault diagnosis method for the low-frequency power transformer, original voiceprint data are preprocessed by constructing a voiceprint fragment generation model, a voiceprint feature extraction model, a sample fault diagnosis model and a zero abnormal sample fault diagnosis model; dimensionality reduction compression and accurate and efficient feature extraction of the original voiceprint data are realized; meanwhile, the voiceprint multi-dimensional characteristic parameter matrix is used as model input, a sample fault diagnosis model is trained and tested, a zero abnormal sample fault diagnosis model is formed, efficient and accurate recognition of abnormal voiceprint samples of the low-frequency transformer can be achieved, therefore, under the condition of zero abnormal samples, sound samples which are not seen can be accurately diagnosed, and the fault diagnosis accuracy of the low-frequency transformer is improved. And the generalization capability is high, so that the accurate detection and evaluation of the operation state of the low-frequency power transformer can be realized.

Description

Technical Field

[0001] The invention relates to a zero-abnormal sample fault diagnosis method and system for a low-frequency power transformer, belonging to the technical field of low-frequency power transformer operating state detection. Background Art

[0002] Low Frequency Alternating Current (LFAC) is a new transmission technology that has the advantages of both DC transmission and conventional AC transmission. LFAC reduces 50Hz industrial frequency power to non-industrial frequency for transmission through AC-AC frequency conversion devices, or directly generates low-frequency power from renewable energy for transmission. On the one hand, it greatly reduces line reactance and charging reactive power, significantly improving line transmission capacity. On the other hand, AC transformer and circuit breaker technology can still be used, which has the advantages of AC easy voltage level change, easy fault disconnection, and easy networking.

[0003] As a key equipment in LFAC projects, the stability of the operating status of low-frequency power transformers is directly related to the reliability and safety of the power system. With the continuous development of the power industry, higher requirements are placed on the maintenance and status monitoring of transformers. During the operation of low-frequency transformers, the core, winding and other structures will vibrate mechanically and generate mechanical waves. The generated vibration is transmitted to the outside through the body, forming a soundprint signal that propagates to the external environment. The soundprint signal contains rich equipment status information; as an emerging transformer status monitoring method, the soundprint diagnosis technology can collect and analyze the soundprint signals generated by the transformer during operation, and can realize early warning and diagnosis of abnormal transformer conditions, thereby avoiding major accidents and ensuring the safe and stable operation of the power grid.

[0004] At present, the research on transformer voiceprint diagnosis technology mainly focuses on fault voiceprint detection and diagnosis methods, acoustic signal feature extraction, information processing technology, deep learning and pattern recognition, etc. These studies will have limitations due to the sound field reconstruction method, environmental noise interference processing, the generalization ability of the diagnosis algorithm and deep learning methods. They are easily interfered by environmental noise and have difficulties in locating the sound source. Some existing voiceprint diagnosis algorithms may perform well on specific data sets, but their generalization ability may be insufficient for unseen sound samples.

[0005] The information disclosed in this Background Art is only for understanding the background of the inventive concept and therefore it may include information that does not constitute the prior art. Summary of the invention

[0006] In response to the above problem or one of the above problems, an object of the present invention is to provide a zero-abnormal sample fault diagnosis method and system for low-frequency power transformers, which combines the abnormality detection network and the relationship network into a zero-abnormal sample fault diagnosis model, takes the voiceprint multi-dimensional feature parameter matrix as the model input, trains and tests the model, and can achieve efficient and accurate identification of abnormal voiceprint samples of low-frequency transformers, which can effectively improve fault detection and prevention capabilities.

[0007] In response to the above problem or one of the above problems, the second purpose of the present invention is to provide a zero-abnormal sample fault diagnosis method and system for low-frequency power transformers. By constructing a voiceprint fragment generation model, a voiceprint feature extraction model, a sample fault diagnosis model, and a zero-abnormal sample fault diagnosis model, the original voiceprint data is preprocessed to obtain feature parameter data that takes into account both time domain and frequency domain information, thereby achieving dimensionality reduction compression and accurate and efficient feature extraction of the original voiceprint data, which greatly reduces the data processing amount of the zero-abnormal sample fault diagnosis model; at the same time, the voiceprint multi-dimensional feature parameter matrix is ​​used as the model input to train and test the sample fault diagnosis model to form a zero-abnormal sample fault diagnosis model, which can achieve efficient and accurate identification of abnormal voiceprint samples of low-frequency transformers, thereby improving the accuracy of power equipment fault diagnosis and effectively improving fault detection and prevention capabilities. Therefore, in the case of zero abnormal samples, unseen sound samples can be accurately diagnosed, and its generalization ability is strong.

[0008] To achieve one of the above purposes, the first technical solution of the present invention is:

[0009] A zero-abnormal sample fault diagnosis method for a low-frequency power transformer comprises the following steps:

[0010] Step 1: split the collected low-frequency power transformer voiceprint audio through a pre-built voiceprint segment generation model to obtain a number of voiceprint segments;

[0011] Step 2: Using a pre-built voiceprint feature extraction model, extract the features of the voiceprint segment to obtain feature parameter data;

[0012] Step 3: Using a pre-built sample fault diagnosis model, based on the training data in the characteristic parameter data, the sample fault diagnosis model is trained to obtain a correlation score threshold to form a zero-abnormal sample fault diagnosis model;

[0013] Step 4: Input the voiceprint data to be tested into the zero-abnormal sample fault diagnosis model, and identify and detect the abnormal voiceprint audio based on the correlation score threshold.

[0014] After continuous exploration and experimentation, the present invention pre-processes the original voiceprint data by constructing a voiceprint fragment generation model, a voiceprint feature extraction model, a sample fault diagnosis model, and a zero-abnormal sample fault diagnosis model, and obtains feature parameter data that takes into account both time domain and frequency domain information, thereby achieving dimensionality reduction compression and accurate and efficient feature extraction of the original voiceprint data, which greatly reduces the data processing amount of the zero-abnormal sample fault diagnosis model; at the same time, the voiceprint multi-dimensional feature parameter matrix is ​​used as the model input to train and test the sample fault diagnosis model to form a zero-abnormal sample fault diagnosis model, which can achieve efficient and accurate identification of abnormal voiceprint samples of low-frequency transformers, thereby improving the accuracy of power equipment fault diagnosis, and effectively enhancing fault detection and prevention capabilities. Therefore, in the case of zero abnormal samples, unseen sound samples can be accurately diagnosed, and its generalization ability is strong, thereby achieving accurate detection and evaluation of the operating status of low-frequency power transformers.

[0015] Furthermore, in actual working conditions, it is difficult and costly to collect abnormal voiceprint samples. The present invention constructs a zero abnormal sample fault diagnosis model, learns the characteristics of normal voiceprint data, and combines the correlation score of the sample. It can accurately detect unknown abnormal voiceprint samples without abnormal samples, and realizes "zero abnormal sample" fault diagnosis. The fault detection cost is low, the solution is simple and practical, and it is feasible, especially suitable for low-frequency scenarios.

[0016] As the preferred technical measures:

[0017] Step 1: The method of splitting the collected low-frequency power transformer voiceprint audio by using the pre-built voiceprint segment generation model to obtain several voiceprint segments is as follows:

[0018] Get low-frequency power transformer soundprint audio;

[0019] Based on the characteristics of low-frequency power transformer voiceprint, the low-frequency power transformer voiceprint audio is subjected to noise reduction processing, and a low-pass filter with a frequency upper limit of 2000Hz is used to filter out the background noise to obtain the noise-reduced voiceprint signal;

[0020] The denoised voiceprint signal is split into several voiceprint segments, and fast Fourier transform is performed on each segment to achieve dimensionality reduction compression while retaining the time domain and frequency domain information in the voiceprint signal.

[0021] As the preferred technical measures:

[0022] Step 2: Using the pre-built voiceprint feature extraction model, the features of the voiceprint segment are extracted to obtain the feature parameter data as follows:

[0023] Process the voiceprint segment, extract the amplitude corresponding to the 20 Hz multiple frequency in its spectrum and take the square, and obtain the characteristic parameter sequence R(k) of the voiceprint segment, where k represents the 20 Hz multiple frequency number, k = 1, 2, 3, ..., 99, 100;

[0024] The characteristic parameter series R(k) of all voiceprint segments are combined into a voiceprint multi-dimensional characteristic parameter matrix R(k,t), where t represents the voiceprint segment number, to form characteristic parameter data.

[0025] As the preferred technical measures:

[0026] Step 3: The method of constructing a sample fault diagnosis model is as follows:

[0027] Create an anomaly detection network based on a generative adversarial algorithm to process feature parameter data and obtain reconstructed feature parameter data;

[0028] Create a relational network, which includes the embedding module F σ1 , Embedded Module F σ2 and a relation module G θ ;

[0029] Using embedded module F σ1 Extract the first feature of the input feature parameter data x and use the embedding module F σ2 Extract the second feature of the reconstructed feature parameter data x1;

[0030] The method for extracting the first feature is as follows:

[0031] The feature parameter data is taken as input and first input into the CNN network to extract the spatial features of the feature parameter data and output a vector sequence. Then, the vector sequence output by the CNN network is used as the input of the LSTM layer, and its gating mechanism is used to capture the time series features in the vector sequence and output the feature sequence, that is, to obtain the first feature.

[0032] The method for extracting the second feature is as follows:

[0033] The reconstructed feature parameter data is used as input and first input into the CNN network to extract the spatial features of the reconstructed feature parameter data and output a vector sequence. Then, the vector sequence output by the CNN network is used as the input of the LSTM layer, and its gating mechanism is used to capture the time series features in the vector sequence and output the feature sequence, that is, to obtain the second feature.

[0034] Using the relationship module G θ Compare the first feature and the second feature, give a correlation score, and complete the construction of the sample fault diagnosis model. The method is as follows:

[0035] First, the first feature and the second feature are concatenated to obtain a feature vector, which is then input into the relationship module G. θ In the first fully connected layer, the input feature vector is linearly combined through the connection weights between neurons, and then the nonlinear characteristics are introduced through the activation function, and then input into the relationship module G θ In the second fully connected layer, linear combination and nonlinear transformation are also performed to finally output a scalar as the correlation score of the two features.

[0036] As the preferred technical measures:

[0037] Step 3: Use the pre-built sample fault diagnosis model to train the sample fault diagnosis model based on the training data in the feature parameter data to obtain the correlation score threshold to form a zero-abnormal sample fault diagnosis model as follows:

[0038] The feature parameter data is divided into feature parameter data for training and feature parameter data for testing, with the ratio of the two being 5:1;

[0039] According to the characteristic parameter data for training, i.e., the training data, the anomaly detection network and the relationship network based on the generative adversarial algorithm are trained alternately at a ratio of 2:1 until a balance point is reached; the method for training the anomaly detection network and the relationship network is as follows:

[0040] Using the anomaly detection network, the feature parameter data used for training is processed to obtain reconstructed feature parameter data;

[0041] Using embedded module F σ1 Extract the first feature of the input training feature parameter data x, and use the embedding module F σ2 Extract the second feature of the reconstructed feature parameter data x1;

[0042] Using the relationship module G θ Compare the first feature and the second feature and give a correlation score;

[0043] For all feature parameter data, calculate the correlation scores one by one to obtain a correlation score array;

[0044] According to the correlation score array, select the minimum correlation score as the correlation score threshold;

[0045] Based on the correlation score threshold, a zero-abnormal sample fault diagnosis model is formed.

[0046] As the preferred technical measures:

[0047] Using the anomaly detection network, the characteristic parameter data is processed to obtain the method of reconstructing the characteristic parameter data as follows:

[0048] The feature parameter data x passes through the encoder G E (x) is encoded and mapped to a low-dimensional latent vector z as follows:

[0049] The feature parameter data x is taken as input, firstly, it is extracted through the convolution layer, these features are reduced in dimension through the pooling layer, then normalized through the batch normalization layer, and finally nonlinear characteristics are introduced through the activation function layer, and finally the input data is compressed into a low-dimensional potential vector z;

[0050] The latent vector z passes through the decoder G D (z), and obtain the reconstructed feature parameter data x1 of the feature parameter data x, the method is as follows:

[0051] The low-dimensional latent vector z is taken as input, and the latent vector z is upsampled through the transposed convolution layer to gradually restore the dimension of the data; in this process, the input of each layer needs to be normalized through the batch normalization layer, and then the nonlinear characteristics are introduced through the activation function layer; after a series of upsampling, normalization and activation operations, the decoder finally outputs the reconstructed feature parameter data x1 with the same dimension as the feature parameter data x;

[0052] The reconstruction encoder E(x1) then re-encodes the reconstructed feature parameter data x1 to obtain the latent vector z1 of the reconstructed feature parameter data x1.

[0053] As the preferred technical measures:

[0054] A semi-supervised generative adversarial training method is used to minimize the distance between these feature parameter data and the latent vector during training, so that the abnormal network model can learn the feature parameter data distribution of normal samples;

[0055] After mapping, samples that are far away from the normal feature parameter data distribution are considered abnormal samples;

[0056] During the training phase, the loss function L of the anomaly detection network consists of two parts: reconstruction loss L rec and the reconstruction loss L enc , and its calculation formula is as follows:

[0057] L rec =||x-x1||1

[0058] L enc =||z-z1||2

[0059] L=w rec L rec +w enc L enc

[0060] During the training phase, the anomaly detection network is trained using normal samples;

[0061] In the test phase, when a normal sample is input, it passes through the encoder G E (x) The difference between the encoded latent vector z and the latent variable z1 obtained by the reconstructed encoder is small;

[0062] But when the abnormal sample is input, the decoder G D (z) does not have the ability to reconstruct abnormal samples, and the reconstructed feature parameter data x1 will lose some abnormal information. Then the latent vector z1 obtained by reconstructing the feature parameter data x1 through the reconstruction encoder E(x1) will be significantly different from the latent vector z obtained by the first encoding.

[0063] As the preferred technical measures:

[0064] Embed Module F σ1 , Embedded Module F σ2 It is created based on the convolutional neural network (CNN) and the long short-term memory (LSTM) network to form a combined structure that uses a small batch cross entropy loss function as an evaluation indicator for fault identification to capture the multi-dimensional features of the voiceprint in the feature parameter matrix;

[0065] Relationship Module G θ Setting two fully connected layers, which include an input fully connected layer and an output fully connected layer;

[0066] The input fully connected layer uses the ReLU function as the activation function;

[0067] The output fully connected layer uses the Sigmoid function as the activation function of the output fully connected layer.

[0068] To achieve one of the above purposes, the second technical solution of the present invention is:

[0069] A zero-abnormal sample fault diagnosis method for a low-frequency power transformer includes the following contents:

[0070] The collected low-frequency power transformer voiceprint audio is denoised using a low-pass filter and then split into several voiceprint segments;

[0071] Extract the characteristic parameter sequence R(k) of the voiceprint segment, and combine the characteristic parameter sequence R(k) of all voiceprint segments into a voiceprint multi-dimensional characteristic parameter matrix R(k,t);

[0072] Divide the voiceprint multi-dimensional feature parameter matrix R(k,t) into a training set and a test set;

[0073] The training set is input into the zero-anomaly sample fault diagnosis model, and the anomaly detection network and the relationship network based on the generative adversarial algorithm are alternately trained until convergence. The minimum value of the correlation score output by the relationship module in the zero-anomaly sample fault diagnosis model is selected as the threshold γ for testing the zero-anomaly sample fault diagnosis model.

[0074] The test set is input into the trained zero-abnormal sample fault diagnosis model to identify and detect abnormal samples.

[0075] The present invention preprocesses the original voiceprint data and obtains a multi-dimensional feature parameter matrix of the voiceprint that takes into account both time domain and frequency domain information, thereby achieving dimensionality reduction compression and accurate and efficient feature extraction of the original voiceprint data, which greatly reduces the data processing amount of the zero-abnormal sample fault diagnosis model.

[0076] Furthermore, the present invention combines the anomaly detection network and the relationship network into a zero-abnormal sample fault diagnosis model, takes the voiceprint multi-dimensional feature parameter matrix as the model input, trains and tests the network model, and can achieve efficient and accurate identification of abnormal voiceprint samples of low-frequency transformers, which can effectively improve fault detection and prevention capabilities.

[0077] To achieve one of the above purposes, the third technical solution of the present invention is:

[0078] A zero abnormal sample fault diagnosis system for a low-frequency power transformer, comprising:

[0079] one or more processors;

[0080] A storage device for storing one or more programs;

[0081] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned zero-abnormal sample fault diagnosis method for a low-frequency power transformer.

[0082] Compared with the prior art solutions, the present invention has the following beneficial effects:

[0083] After continuous exploration and experimentation, the present invention pre-processes the original voiceprint data by constructing a voiceprint fragment generation model, a voiceprint feature extraction model, a sample fault diagnosis model, and a zero-abnormal sample fault diagnosis model, and obtains feature parameter data that takes into account both time domain and frequency domain information, thereby achieving dimensionality reduction compression and accurate and efficient feature extraction of the original voiceprint data, which greatly reduces the data processing amount of the zero-abnormal sample fault diagnosis model; at the same time, the voiceprint multi-dimensional feature parameter matrix is ​​used as the model input to train and test the sample fault diagnosis model to form a zero-abnormal sample fault diagnosis model, which can achieve efficient and accurate identification of abnormal voiceprint samples of low-frequency transformers, thereby improving the accuracy of fault diagnosis of power equipment, and effectively enhancing fault detection and prevention capabilities. Therefore, in the case of zero abnormal samples, unseen sound samples can be accurately diagnosed, and its generalization ability is strong.

[0084] Furthermore, the present invention combines the anomaly detection network and the relationship network into a zero-abnormal sample fault diagnosis model, uses the voiceprint multi-dimensional feature parameter matrix as the model input, trains and tests the model, and can achieve efficient and accurate recognition of abnormal voiceprint samples of low-frequency transformers, which can effectively improve fault detection and prevention capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 A schematic diagram of a flow chart of the zero abnormal sample fault diagnosis method of the present invention;

[0086] Figure 2 A structural schematic diagram of a zero abnormal sample fault diagnosis model of the present invention;

[0087] Figure 3 Another flowchart of the zero abnormal sample fault diagnosis method of the present invention is shown in FIG. DETAILED DESCRIPTION

[0088] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0089] On the contrary, the present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention as defined by the claims. Further, in order to make the public have a better understanding of the present invention, some specific details are described in detail in the following detailed description of the present invention. Those skilled in the art can fully understand the present invention without the description of these details.

[0090] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0091] like Figure 1 As shown, the first specific embodiment of the zero abnormal sample fault diagnosis method for low-frequency power transformer of the present invention is:

[0092] A zero-abnormal sample fault diagnosis method for a low-frequency power transformer comprises the following steps:

[0093] Step 1: split the collected low-frequency power transformer voiceprint audio through a pre-built voiceprint segment generation model to obtain a number of voiceprint segments;

[0094] Step 2: Using a pre-built voiceprint feature extraction model, extract the features of the voiceprint segment to obtain feature parameter data;

[0095] Step 3: Using a pre-built sample fault diagnosis model, based on the training data in the characteristic parameter data, the sample fault diagnosis model is trained to obtain a correlation score threshold to form a zero-abnormal sample fault diagnosis model;

[0096] Step 4: Input the voiceprint data to be tested into the zero-abnormal sample fault diagnosis model, and identify and detect the abnormal voiceprint audio based on the correlation score threshold.

[0097] A second specific embodiment of the zero abnormal sample fault diagnosis method for low-frequency power transformer of the present invention:

[0098] A zero-abnormal sample fault diagnosis method for a low-frequency power transformer includes the following contents:

[0099] The collected low-frequency power transformer voiceprint audio is denoised using a low-pass filter and then split into several voiceprint segments;

[0100] Extract the characteristic parameter sequence R(k) of the voiceprint segment, and combine the characteristic parameter sequence R(k) of all voiceprint segments into a voiceprint multi-dimensional characteristic parameter matrix R(k,t);

[0101] Divide the voiceprint multi-dimensional feature parameter matrix R(k,t) into a training set and a test set;

[0102] The training set is input into the zero-abnormal sample fault diagnosis model, and the anomaly detection network and the relational network based on the generative adversarial algorithm are alternately trained until convergence. The relational module G in the zero-abnormal sample fault diagnosis model is selected. θ The minimum value of the output correlation score is used as the threshold γ for testing the zero-abnormal sample fault diagnosis model;

[0103] The test set is input into the trained zero-abnormal sample fault diagnosis model to identify and detect abnormal samples, thus completing the optimization of the zero-abnormal sample fault diagnosis model.

[0104] Based on the optimized zero-abnormal sample fault diagnosis model, zero-abnormal sample fault diagnosis of the operating status of the low-frequency transformer is realized.

[0105] A third specific embodiment of the zero abnormal sample fault diagnosis method for low-frequency power transformer of the present invention:

[0106] In view of the shortcomings of the existing transformer voiceprint diagnosis algorithms and the lack of abnormal voiceprint samples of low-frequency power transformers, the present invention proposes a zero-abnormal sample fault diagnosis method for the operating status of low-frequency power transformers based on voiceprint information.

[0107] Firstly, the voiceprint data of low-frequency transformer is preprocessed and feature extracted to extract the multi-dimensional feature parameter matrix of voiceprint. Then, based on the anomaly detection network and relationship network of generative adversarial algorithm (GAN), a zero-abnormal sample fault diagnosis model is constructed to realize zero-abnormal sample fault diagnosis of the operating status of low-frequency transformer.

[0108] In this embodiment, the method of data preprocessing and feature extraction is as follows:

[0109] In view of the voiceprint characteristics of low-frequency power transformers, this application preprocesses the voiceprint samples, reduces the dimensionality of the original voiceprint data, and retains the time domain and frequency domain information in the voiceprint, while ensuring that the multi-dimensional characteristics of the voiceprint signal are retained and the data volume is compressed at the same time. The processing steps are as follows:

[0110] Step 1: Perform noise reduction processing on the collected voiceprint signal, and select a low-pass filter with a frequency upper limit of 2000Hz to filter out background noise;

[0111] Step 2, split the denoised voiceprint signal into several segments. Taking a 10-second voiceprint signal as an example, split the signal into voiceprint segments with a length of 0.1s. Divide it into 100 voiceprint segments in total and perform fast Fourier transform on each of them.

[0112] Step 3, the low-frequency power transformer acoustic signal is mainly 20Hz multiple frequency. Taking a single voiceprint segment as an example, the present application extracts the amplitude corresponding to the 20Hz multiple frequency (the upper frequency limit is 2000Hz) in the spectrum and takes the square to obtain the characteristic parameter sequence R(k) of the voiceprint segment (where k represents the 20Hz multiple frequency number, k=1,2,3,...,99,100).

[0113] Step 4, combine the feature parameter series R(k) of all voiceprint segments into a voiceprint multi-dimensional feature parameter matrix R(k,t) (where t represents the voiceprint segment number, t=1, 2, 3, ..., 99, 100).

[0114] In this embodiment, the zero abnormal sample fault diagnosis model includes the following contents:

[0115] The zero-abnormal sample fault diagnosis model mainly consists of two parts: anomaly detection network based on the generative adversarial algorithm (GAN) and relationship network. Figure 2 .

[0116] Anomaly detection network based on Generative Adversarial Network (GAN). The network consists of an encoder G E (x), decoder G D (z) and reconstruction encoder E(x1). The network first passes the input data x through the encoder G E (x) is encoded and mapped to a low-dimensional latent vector z, which is passed through the decoder G D (z) can get the reconstructed data x1 of x, and then the reconstruction encoder E(x1) re-encodes x1 to get the latent vector z1 of the reconstructed data x1. This network adopts semi-supervised generative adversarial training to minimize the distance between these data and the latent vector during training, so that the network model can learn the data distribution of normal samples. Therefore, samples that are far away from the normal data distribution after mapping can be regarded as abnormal samples.

[0117] The encoding process is as follows:

[0118] The feature parameter matrix is ​​used as input, and first passes through the convolution layer for feature extraction. These features are reduced in dimension by the pooling layer, and then normalized by the batch normalization layer. Finally, the nonlinear characteristics are introduced by the activation function layer, and the input data is finally compressed into a low-dimensional latent vector (this latent vector corresponds one-to-one with the feature parameter matrix). The decoding process is as follows:

[0119] The low-dimensional latent vector is used as input, and the latent vector is upsampled through the transposed convolution layer to gradually restore the dimension of the data. In this process, the input of each layer needs to be normalized through the batch normalization layer, and then the nonlinear characteristics are introduced through the activation function layer. After a series of upsampling, normalization and activation operations, the decoder finally outputs reconstructed data with the same dimension as the original input data.

[0120] The re-encoding process is the same as the above encoding process, except that the input data is different. The input data of the re-encoding process is the reconstructed data.

[0121] During the training phase, the network loss function consists of two parts: reconstruction loss L recand the reconstruction loss L enc , as shown below:

[0122] L rec =||x-x1||1 (3)

[0123] L enc =||z-z1||2 (4)

[0124] L=w rec L rec +w enc L enc (5)

[0125] The training status of the model can be observed intuitively through these two losses. During the model training process, the reconstruction loss L is set when the model completes the training. rec and the reconstruction loss L enc The maximum value of the reconstruction loss L during model training rec and the reconstruction loss L enc When it is less than the maximum value, the model training is considered complete and has converged.

[0126] The training method used by this network is the same as that of conventional GAN. During the training phase, the network is trained using normal samples, i.e., the encoder G E (x), decoder G D (z) and reconstruction encoder E(x1) are both suitable for normal samples. In the test phase, when a normal sample is input, it passes through the encoder G E (x) The difference between the encoded latent vector z and the latent variable z1 obtained by the reconstructed encoder is small; however, when an abnormal sample is input, the decoder G D (z) does not have the ability to reconstruct abnormal samples. The reconstructed data x1 will lose some abnormal information. Then the latent vector z1 obtained by the reconstruction encoder E(x1) of the reconstructed data x1 will inevitably be significantly different from the latent vector z obtained by the first encoding. Then the reconstruction loss Lrec and the reconstruction loss Lenc are calculated to judge whether the sample is normal or not.

[0127] Normal samples are voiceprint samples of power equipment without any abnormality in actual working conditions. The voiceprint samples used in the present invention are all normal samples, and there are no abnormal samples, that is, "zero abnormal samples".

[0128] In actual working conditions, it is difficult and costly to collect abnormal voiceprint samples. The present invention aims to preliminarily construct an abnormality detection model. Although the model in the present invention only learns the characteristics of normal voiceprint data, it still has a high accuracy in detecting unknown abnormal voiceprint samples in the future.

[0129] The relational network in the zero-abnormal sample fault diagnosis model of this embodiment mainly consists of two embedding modules Fσ1 、F σ2 and a relation module G θ Composition. σ1 Used to extract the features of input data x, F σ2 Used to extract the features of the reconstructed data x1, G θ Used to compare two types of features and give a correlation score.

[0130] The method of extracting data features is as follows:

[0131] The aforementioned feature parameter matrix is ​​first input into the CNN network as input to extract the spatial features of the data and output a vector sequence. Then, the feature sequence output by CNN is used as the input of LSTM, and its gating mechanism is used to capture the time series features in the feature sequence and output the feature sequence.

[0132] The method to compare two types of features and give a correlation score is as follows:

[0133] First, the feature sequence is concatenated and then input into G θ In the first fully connected layer, the input feature vector is linearly combined through the connection weights between neurons, and then the nonlinear characteristics are introduced through the activation function (such as ReLU), and then input to G θ In the second fully connected layer, linear combination and nonlinear transformation are also performed to finally output a scalar as the correlation score of the two features.

[0134] Embedding module F for extracting voiceprint features σ1 、F σ2 The combined structure of convolutional neural network (CNN) and long short-term memory (LSTM) network can capture the features in the multi-dimensional feature parameter matrix of voiceprint, so that the model can more accurately identify and diagnose faults. The small batch cross entropy loss function is used as the evaluation index of fault identification. The extracted features output by the LSTM layer are input into the comparison relationship network G θ In. Relationship module G θ Two fully connected layers are used, Sigmoid is used as the activation function of the output fully connected layer, and ReLU is used as the activation function of the other fully connected layer.

[0135] like Figure 3 As shown, the fourth specific embodiment of the zero abnormal sample fault diagnosis method for low-frequency power transformer of the present invention is as follows:

[0136] A zero-abnormal sample fault diagnosis method for a low-frequency power transformer is a zero-abnormal sample fault diagnosis method for a low-frequency power transformer operating state based on voiceprint information, and the process is as follows:

[0137] In the first step, the low-frequency power transformer voiceprint audio is denoised using a low-pass filter, and then split into several voiceprint segments, extracting the characteristic parameter sequence R(k) of the voiceprint segments, combining the characteristic parameter sequence R(k) of all voiceprint segments into a voiceprint multi-dimensional feature parameter matrix R(k, t), and dividing the matrix R(k, t). The ratio of training samples to test samples is 5:1.

[0138] In the second step, the relationship network is pre-trained, and then the anomaly detection network based on the generative adversarial algorithm (GAN) and the relationship network are trained alternately at a ratio of 2:1 until the equilibrium point is reached, and the relationship module G is selected. θ The minimum value of the output correlation score is used as the threshold γ for subsequent model testing.

[0139] The balance point is the training convergence point. During the model training process, the maximum values ​​of the reconstruction loss Lrec and the reconstruction loss Lenc are set when the model completes the training. When the reconstruction loss Lrec and the reconstruction loss Lenc during the model training process are less than the maximum values, the model training is considered to be completed and has converged.

[0140] The threshold γ is used for model testing, which includes the following:

[0141] When the model is tested, if the correlation score of a sample is less than or equal to the threshold γ, the sample is considered normal; if it is greater than the threshold γ, the sample is considered abnormal.

[0142] The third step is to input the test set into the trained zero-abnormal sample fault diagnosis model to identify and detect abnormal samples, and then evaluate the model detection effect.

[0143] This application preprocesses the original voiceprint data to obtain a multi-dimensional feature parameter matrix of the voiceprint that takes into account both time domain and frequency domain information, achieving dimensionality reduction compression of the original voiceprint data and accurate and efficient feature extraction. At the same time, this application extracts a multi-dimensional feature parameter matrix from the collected low-frequency transformer voiceprint information, which greatly reduces the data processing volume of the zero-abnormal sample fault diagnosis model, while taking into account both time domain and frequency domain information, achieving accurate and efficient extraction of voiceprint features.

[0144] Furthermore, the present invention combines the anomaly detection network and the relationship network into a zero-abnormal sample fault diagnosis model, uses the voiceprint multi-dimensional feature parameter matrix as the model input, trains and tests the network model, and can achieve efficient and accurate recognition of abnormal voiceprint samples of low-frequency transformers. At the same time, the zero-abnormal sample fault diagnosis method of the present application provides a novel method for processing voiceprint data, especially in the low-frequency range, which may be an unexplored area in previous technologies.

[0145] Furthermore, this application combines voiceprint data processing technology with zero-abnormal sample fault diagnosis technology to realize a new state assessment technology and improve the accuracy of power equipment fault diagnosis. This method shows higher accuracy in identifying and classifying low-frequency voiceprint signals, which can effectively improve fault detection and prevention capabilities.

[0146] An embodiment of a device applying the method of the present invention:

[0147] An electronic device comprising:

[0148] one or more processors;

[0149] A storage device for storing one or more programs;

[0150] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned zero-abnormal sample fault diagnosis method for a low-frequency power transformer.

[0151] A computer medium embodiment of the method of the present invention is applied:

[0152] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned zero-abnormal sample fault diagnosis method for a low-frequency power transformer.

[0153] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, and computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, optical storage, etc.) containing computer-usable program codes.

[0154] The present application is described by flowcharts or / and block diagrams of the methods, devices (systems), and computer program products of the embodiments of the present application. It should be understood that each process or / and block in the flowchart or / and block diagram and the combination of the processes or / and blocks in the flowchart or / and block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart or / and block diagram. Figure 1 Process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0155] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 Process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0156] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 Process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0157] The model in this application is an object that objectively describes the morphological structure with the help of physical or virtual representation. The object is not equal to the object and is not limited to physical and virtual. It can be a data processing function, software program, processing mode, usage method, operation method, workflow, application process, electronic hardware, circuit module, processing system, system imitation or simulation object.

[0158] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field can still modify or replace the specific implementation methods of the present invention with equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A zero-abnormal sample fault diagnosis method for a low-frequency power transformer, characterized in that: The following steps are involved: Step 1: split the collected low-frequency power transformer voiceprint audio through a pre-built voiceprint segment generation model to obtain a number of voiceprint segments; Step 2: Using a pre-built voiceprint feature extraction model, extract the features of the voiceprint segment to obtain feature parameter data; Step 3: Using a pre-built sample fault diagnosis model, based on the training data in the characteristic parameter data, the sample fault diagnosis model is trained to obtain a correlation score threshold to form a zero-abnormal sample fault diagnosis model; Step 4: Input the voiceprint data to be tested into the zero-abnormal sample fault diagnosis model, and identify and detect the abnormal voiceprint audio based on the correlation score threshold.

2. A zero-abnormal sample fault diagnosis method for a low-frequency power transformer as claimed in claim 1, characterized in that: Step 1: The method of splitting the collected low-frequency power transformer voiceprint audio by using the pre-built voiceprint segment generation model to obtain several voiceprint segments is as follows: Get low-frequency power transformer soundprint audio; Based on the characteristics of low-frequency power transformer voiceprint, the low-frequency power transformer voiceprint audio is subjected to noise reduction processing, and a low-pass filter with a frequency upper limit of 2000Hz is used to filter out the background noise to obtain the noise-reduced voiceprint signal; The denoised voiceprint signal is split into several voiceprint segments, and fast Fourier transform is performed on each segment to achieve dimensionality reduction compression while retaining the time domain and frequency domain information in the voiceprint signal.

3. A zero abnormal sample fault diagnosis method for a low frequency power transformer as claimed in claim 1, characterized in that: Step 2: Using the pre-built voiceprint feature extraction model, the features of the voiceprint segment are extracted to obtain the feature parameter data as follows: Process the voiceprint segment, extract the amplitude corresponding to the 20 Hz multiple frequency in its spectrum and take the square, and obtain the characteristic parameter sequence R(k) of the voiceprint segment, where k represents the 20 Hz multiple frequency number, k = 1, 2, 3, ..., 99, 100; The characteristic parameter series R(k) of all voiceprint segments are combined into a voiceprint multi-dimensional characteristic parameter matrix R(k,t), where t represents the voiceprint segment number, to form characteristic parameter data.

4. A zero abnormal sample fault diagnosis method for a low frequency power transformer as claimed in claim 1, characterized in that: Step 3: The method of constructing a sample fault diagnosis model is as follows: Create an anomaly detection network based on a generative adversarial algorithm to process feature parameter data and obtain reconstructed feature parameter data; Create a relational network, which includes the embedding module F σ1 , Embedded Module F σ2 and a relation module G θ ; Using embedded module F σ1 Extract the first feature of the input feature parameter data x and use the embedding module F σ2 Extract the second feature of the reconstructed feature parameter data x1; The method for extracting the first feature is as follows: The feature parameter data is taken as input and first input into the CNN network to extract the spatial features of the feature parameter data and output a vector sequence. Then, the vector sequence output by the CNN network is used as the input of the LSTM layer, and its gating mechanism is used to capture the time series features in the vector sequence and output the feature sequence, that is, to obtain the first feature. The method for extracting the second feature is as follows: The reconstructed feature parameter data is used as input and first input into the CNN network to extract the spatial features of the reconstructed feature parameter data and output a vector sequence. Then, the vector sequence output by the CNN network is used as the input of the LSTM layer, and its gating mechanism is used to capture the time series features in the vector sequence and output the feature sequence, that is, to obtain the second feature. Using the relationship module G θ Compare the first feature and the second feature, give a correlation score, and complete the construction of the sample fault diagnosis model. The method is as follows: First, the first feature and the second feature are concatenated to obtain a feature vector, which is then input into the relationship module G. θ In the first fully connected layer, the input feature vector is linearly combined through the connection weights between neurons, and then the nonlinear characteristics are introduced through the activation function, and then input into the relationship module G θ In the second fully connected layer, linear combination and nonlinear transformation are also performed to finally output a scalar as the correlation score of the two features.

5. A zero abnormal sample fault diagnosis method for a low frequency power transformer as claimed in claim 4, characterized in that: Step 3: Use the pre-built sample fault diagnosis model to train the sample fault diagnosis model based on the training data in the feature parameter data to obtain the correlation score threshold to form a zero-abnormal sample fault diagnosis model as follows: The feature parameter data is divided into feature parameter data for training and feature parameter data for testing, with the ratio of the two being 5:1; According to the characteristic parameter data for training, i.e., the training data, the anomaly detection network and the relationship network based on the generative adversarial algorithm are trained alternately at a ratio of 2:1 until a balance point is reached; the method for training the anomaly detection network and the relationship network is as follows: Using the anomaly detection network, the feature parameter data used for training is processed to obtain reconstructed feature parameter data; Using embedded module F σ1 Extract the first feature of the input training feature parameter data x, and use the embedding module F σ2 Extract the second feature of the reconstructed feature parameter data x1; Using the relationship module G θ Compare the first feature and the second feature and give a correlation score; For all feature parameter data, calculate the correlation scores one by one to obtain a correlation score array; According to the correlation score array, select the minimum correlation score as the correlation score threshold; Based on the correlation score threshold, a zero-abnormal sample fault diagnosis model is formed.

6. A zero-abnormal sample fault diagnosis method for a low-frequency power transformer as claimed in claim 5, characterized in that: Using the anomaly detection network, the characteristic parameter data is processed to obtain the method of reconstructing the characteristic parameter data as follows: The feature parameter data x passes through the encoder G E (x) is encoded and mapped to a low-dimensional latent vector z as follows: The feature parameter data x is taken as input, firstly, it is extracted through the convolution layer, these features are reduced in dimension through the pooling layer, then normalized through the batch normalization layer, and finally nonlinear characteristics are introduced through the activation function layer, and finally the input data is compressed into a low-dimensional potential vector z; The latent vector z passes through the decoder G D (z), and obtain the reconstructed feature parameter data x1 of the feature parameter data x, the method is as follows: The low-dimensional latent vector z is taken as input, and the latent vector z is upsampled through the transposed convolution layer to gradually restore the dimension of the data; in this process, the input of each layer needs to be normalized through the batch normalization layer, and then the nonlinear characteristics are introduced through the activation function layer; after a series of upsampling, normalization and activation operations, the decoder finally outputs the reconstructed feature parameter data x1 with the same dimension as the feature parameter data x; The reconstruction encoder E(x1) then re-encodes the reconstructed feature parameter data x1 to obtain the latent vector z1 of the reconstructed feature parameter data x1.

7. A zero-abnormal sample fault diagnosis method for a low-frequency power transformer as claimed in claim 6, characterized in that: A semi-supervised generative adversarial training method is used to minimize the distance between these feature parameter data and the latent vector during training, so that the abnormal network model can learn the feature parameter data distribution of normal samples; After mapping, samples that are far away from the normal feature parameter data distribution are considered abnormal samples; During the training phase, the loss function L of the anomaly detection network consists of two parts: reconstruction loss L rec and the reconstruction loss L enc , and its calculation formula is as follows: L rec =||x-x1||1 L enc =||z-z1||2 L=w rec L rec +w enc L enc During the training phase, the anomaly detection network is trained using normal samples; In the test phase, when a normal sample is input, it passes through the encoder G E (x) The difference between the encoded latent vector z and the latent variable z1 obtained by the reconstructed encoder is small; But when the abnormal sample is input, the decoder G D (z) does not have the ability to reconstruct abnormal samples, and the reconstructed feature parameter data x1 will lose some abnormal information. Then the latent vector z1 obtained by reconstructing the feature parameter data x1 through the reconstruction encoder E(x1) will be significantly different from the latent vector z obtained by the first encoding.

8. A zero-abnormal sample fault diagnosis method for a low-frequency power transformer as claimed in claim 4, characterized in that: Embed Module F σ1 , Embedded Module F σ2 It is created based on the convolutional neural network (CNN) and the long short-term memory (LSTM) network to form a combined structure that uses a small batch cross entropy loss function as an evaluation indicator for fault identification to capture the multi-dimensional features of the voiceprint in the feature parameter matrix; Relationship Module G θ Setting two fully connected layers, which include an input fully connected layer and an output fully connected layer; The input fully connected layer uses the ReLU function as the activation function; The output fully connected layer uses the Sigmoid function as the activation function of the output fully connected layer.

9. A zero-abnormal sample fault diagnosis method for a low-frequency power transformer, characterized in that: Includes the following: The collected low-frequency power transformer voiceprint audio is denoised using a low-pass filter and then split into several voiceprint segments; Extract the characteristic parameter sequence R(k) of the voiceprint segment, and combine the characteristic parameter sequence R(k) of all voiceprint segments into a voiceprint multi-dimensional characteristic parameter matrix R(k,t); Divide the voiceprint multi-dimensional feature parameter matrix R(k,t) into a training set and a test set; The training set is input into the zero-anomaly sample fault diagnosis model, and the anomaly detection network and the relationship network based on the generative adversarial algorithm are alternately trained until convergence. The minimum value of the correlation score output by the relationship module in the zero-anomaly sample fault diagnosis model is selected as the threshold γ for testing the zero-anomaly sample fault diagnosis model. The test set is input into the trained zero-abnormal sample fault diagnosis model to identify and detect abnormal samples.

10. A zero abnormal sample fault diagnosis system for low frequency power transformer, characterized in that: It includes: one or more processors; A storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement a zero-abnormal sample fault diagnosis method for a low-frequency power transformer as described in any one of claims 1-9.

Images