Transformer fault diagnosis method based on chaos theory contrast learning under strong noise background

Abstract

The invention discloses a transformer fault diagnosis method based on chaos theory comparative learning under a strong noise background, and the method comprises the following steps: 1, obtaining a label-free data set and a label data set of vibration signals of a transformer box body, and carrying out the normalization processing of the collected signals; a second step of utilizing a dung beetle optimization algorithm to obtain an optimal parameter for ICEEMDAN decomposition of the vibration signal in each fault state, performing ICEEMDAN decomposition of the vibration signal to obtain a plurality of modal components IMF, calculating a multi-scale fuzzy entropy of each modal component, screening a component with a smaller entropy value and reconstructing the signal, and performing data cleaning and denoising; 3, performing random data enhancement on the de-noised label-free data set, defining positive and negative sample pairs, performing phase-space reconstruction, completing comparative learning of a pre-training model, and migrating trained encoder parameters to a downstream task; 4, performing phase-space reconstruction on the denoised labeled data set, and inputting the denoised labeled data set into a trained encoder after fine tuning; 5, performing preprocessing and phase-space reconstruction on the label-free test data, and inputting the preprocessed label-free test data into the trained encoder; sixthly, the output features are input into a classifier to be classified, and fault diagnosis is completed. According to the method, the problem of low diagnosis precision caused by a strong noise environment and lack of marked data in transformer fault diagnosis and the problems of misjudgment and missed judgment caused by nonlinear characteristics of vibration signals are effectively solved.

Description

[0001] The present invention relates to the research field of transformer fault diagnosis technology, specifically, a transformer fault diagnosis field based on chaos theory comparative learning under strong noise background. Background Art

[0002] As one of the important components of my country's power system, the health status of the transformer directly affects the stable operation of the power system. However, a large part of the transformers currently in use in my country have been in operation for a long time, and there are safety hazards such as loose transformer windings, loose cores, and winding deformation. Monitoring its health status is of great significance for the stable operation of the power system. In order to ensure the safe operation of the power system and prevent economic losses caused by power system shutdown, an efficient and accurate transformer fault diagnosis model must be established.

[0003] At present, transformer fault diagnosis methods are divided into traditional machine learning and deep learning. Traditional machine learning methods include support vector machine (SVM), artificial neural network (ANN) and other methods, and deep learning methods include convolutional neural network (CNN), long short-term memory network (LSTM) and other methods. However, although the above methods are flexible and have achieved good application results, the diagnosis process requires the use of a large amount of labeled data, and the acquisition of labels is overly dependent on expert knowledge and experience, which is costly and difficult to obtain high-quality labeled data. In the actual operation process, most of the acquired data is unlabeled data, which has a great impact on the diagnosis results. The model may misjudge or miss faults due to the nonlinear characteristics of the vibration signal at the fault location. In addition, many actual operation scenarios are outdoors. Due to the harsh working conditions of the transformer, the strong noise environment increases the difficulty of vibration signal collection and the difficulty of extracting effective sample features, resulting in poor model performance and reduced accuracy.

[0004] In view of the above problems, the present invention proposes a transformer fault diagnosis method based on chaos theory comparative learning under strong noise background, which can effectively solve the accurate identification of unlabeled sample faults in strong noise environment. Based on chaos theory, the phase space reconstruction method is adopted to restore the vibration signal to the chaotic space, analyze its power system characteristics, and effectively solve the fault misjudgment and missed judgment problems caused by the nonlinear characteristics of the vibration signal. Summary of the invention

[0005] The purpose of the present invention is to propose a transformer fault diagnosis method based on chaos theory comparative learning under strong noise background, which can effectively solve the problems of accurate identification of unlabeled sample faults in strong noise environment, and misjudgment and missed faults caused by nonlinear characteristics of vibration signals, and improve the diagnosis accuracy of the model.

[0006] The present invention is implemented by the following technical solution: a transformer fault diagnosis method based on chaos theory comparative learning under strong noise background, comprising the following steps:

[0007] The first step is to collect the vibration signal of the transformer box and normalize the vibration signal;

[0008] In the second step, the Dung Beetle Optimization (DBO) algorithm is used to optimize the optimal ICEEMDAN parameters of the vibration signal under each fault state, and the vibration signal is decomposed by ICEEMDAN to obtain multiple modal components IMF; the multi-scale fuzzy entropy of each modal component is calculated, the components with smaller entropy values ​​are screened and the signal is reconstructed to complete the cleaning and denoising of the transformer fault signal data;

[0009] The third step is to randomly perform two data enhancements on the denoised unlabeled dataset, define positive and negative sample pairs, reconstruct the phase space, complete the contrastive learning pre-training model, and migrate the trained encoder parameters to downstream tasks.

[0010] The fourth step is to reconstruct the phase space of the denoised labeled dataset and fine-tune the trained encoder;

[0011] In the fifth step, the unlabeled test data is input into the trained encoder after preprocessing and phase space reconstruction;

[0012] In the sixth step, the output features are input into the classifier for classification to complete the transformer fault diagnosis.

[0013] Furthermore, the specific steps of the first step are as follows:

[0014] Step 1: Collect transformer box vibration signal sample data;

[0015] Specifically, unlabeled sample data and labeled sample data are collected respectively. The collected labeled sample data are transformer box vibration signals in three fault states, namely, loose transformer winding, loose core, and winding deformation, and a normal state.

[0016] Step 2: Normalize the collected signal.

[0017] Furthermore, the specific steps of the second step are as follows:

[0018] Step 1: Initialize the parameters of the dung beetle population, including the position of the individual dung beetle population vector [white noise amplitude weight (Nstd) and the number of noise additions (NE)], the number of iterations, etc., and divide the dung beetle population into rolling dung beetles, egg balls, small dung beetles, and stealing dung beetles in a ratio of 6:6:7:11;

[0019] Step 2: The rolling ball dung beetle simulates the rolling ball behavior and moves along the given direction in the entire search space. During the rolling process, the position of the rolling ball dung beetle is updated. When the dung beetle encounters an obstacle and cannot move forward, it needs to reposition itself by dancing in order to obtain a new route.

[0020] Step 3: Simulation The boundary selection strategy of the dung beetle's egg-laying area. Once the egg-laying area is determined, the female dung beetle will lay eggs, one egg per iteration.

[0021] Step 4: The grown-up dung beetles come out to forage for food. Establish the best dung beetle area and guide the dung beetles to forage for food.

[0022] Step 5: Thieves will steal food from other dung beetles and locate the best food source;

[0023] Step 6: Repeat the above process, update the position information of the dung beetle, determine whether the target agent is beyond the boundary, and update the current optimal solution and its fitness value until the maximum number of iterations is reached, and obtain the optimal parameter combination of the ICEEMDAN algorithm [white noise amplitude weight (Nstd) and number of noise additions (NE)];

[0024] Step 7: Bring the optimal parameter combination [white noise amplitude weight (Nstd) and noise addition times (NE)] into the ICEEMDAN algorithm to decompose the signal and obtain n modal component IMFs;

[0025] Step 8: Calculate the multi-scale fuzzy entropy (MFE) of each modal component IMF;

[0026] Step 9: Filter out the components with smaller entropy values ​​for signal reconstruction to obtain the reconstructed signal x(t n );

[0027] Furthermore, the specific steps of the third step are as follows:

[0028] Step 1: Construct positive and negative sample pairs from the denoised unlabeled one-dimensional vibration signal dataset, randomly extract a vibration signal, and randomly perform data enhancement twice. The data enhancement methods include but are not limited to random addition of noise, resampling, time reversal, local signal amplification, etc., to obtain two different but related samples x q and x k0 , define x q With x k0 are positive sample pairs, x q and other signals are negative sample pairs, x k0 It also forms a negative sample pair with other signals;

[0029] Step 2: Reconstruct the phase space of the obtained positive and negative samples, restore the one-dimensional vibration signal to the two-dimensional chaotic space, and transform it into a two-dimensional chaotic sequence signal. Use the autocorrelation coefficient method to calculate the extension time τ0, use the Cao method to calculate the embedding dimension m0, and calculate the maximum Lyapunov exponent at the same time;

[0030] Specifically, the maximum Lyapunov exponent is calculated to determine its chaotic characteristics. The formula is as follows:

[0031]

[0032] Among them, L i is the distance between two adjacent points in phase space reconstruction, L' i is the distance between two points at the next moment, and M is the total number of iterations.

[0033] Step 3: Reconstruct the signal x in phase space q The input encoder Q uses a residual network (ResNet) to extract the feature vector q and reconstruct the signal x in phase space. k0 、x k1 、x k2 、x k3 …x kn Input the momentum encoder and extract the feature vectors k0, k1, k2, k3…k n , and form a queue. Its momentum encoder K also uses a residual network (ResNet). Different from the parameter setting of encoder Q, the momentum update formula between encoder Q and momentum encoder K is as follows:

[0034] θ k ←mθ k +(1-m)θ q

[0035] where m∈[0,1),θ q is the parameter of encoder Q, θ k is the parameter of momentum encoder K;

[0036] Step 4: Extract the feature vectors q, k0, k1, k2, k3…k n Input into the MLP multilayer perceptron to obtain the projection vectors q', k0', k1', k2', k3'...k n ';

[0037] Step 5: Calculate the contrast loss between the positive and negative sample pairs. The contrast loss function selected in the present invention is InfoNCE, and its formula is as follows:

[0038]

[0039] In the formula, if there is only one k in the queue that can match q, it is recorded as k + , τ is recorded as the temperature hyperparameter

[0040] Step 6: Repeat the iteration until the contrast loss is minimized and the model is trained to the best. Then, transfer the parameters of the trained encoder Q to the downstream task.

[0041] Furthermore, the specific steps of the fourth step are as follows:

[0042] Step 1: Take the labeled one-dimensional vibration signal data set and restore the one-dimensional vibration signal to the two-dimensional chaotic space to become a two-dimensional chaotic sequence signal;

[0043] Step 2: Input the two-dimensional chaotic sequence signal into the encoder after parameter migration, and fine-tune the encoder parameters;

[0044] Furthermore, the specific steps of the fifth step are as follows:

[0045] Step 1: Normalize the unlabeled vibration signal test sample data;

[0046] Step 2: Decompose the normalized vibration signal into IMF components using DBO-ICEEMDAN, and select components with smaller entropy values ​​for reconstruction;

[0047] Step 3: After the data noise reduction and cleaning, the unlabeled test samples are restored to the two-dimensional chaotic space through phase space reconstruction, converted into a two-dimensional chaotic sequence signal, and input into the encoder that has been parameter migrated;

[0048] Furthermore, the specific steps of the sixth step are as follows:

[0049] Step 1: Input the output features into the Transformer classifier for classification to determine the fault type. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is an overall flow chart of the method involved in the present invention;

[0051] Figure 2 It is a schematic diagram of MLP multi-layer perceptron;

[0052] Figure 3 It is a schematic diagram of the residual block structure. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0055] The first step is to collect the vibration signal of the transformer box and normalize the vibration signal. The specific steps are as follows:

[0056] S101: Collecting transformer box vibration signals

[0057] Specifically, 500 groups of unlabeled sample data with a data length of 2048 data points and 100 groups of labeled sample data with a length of 2048 data points were collected. The collected labeled sample data were the vibration signals of the transformer box in three fault states: loose transformer winding, loose core, and winding deformation, and normal state.

[0058] S102: Normalize the collected signal and map the data value to [0, 1], effectively narrowing the value range, thereby speeding up the model running time. The specific formula is as follows:

[0059]

[0060] In the formula: x' is recorded as the normalized data value, x is recorded as the original data value, min(x) is recorded as the minimum data value, and max(x) is recorded as the maximum data value:

[0061] In the second step, the Dung Beetle Optimization (DBO) algorithm is used to optimize the optimal ICEEMDAN parameters of the vibration signal under each fault state, and the vibration signal is decomposed by ICEEMDAN to obtain multiple modal components IMF; the multi-scale fuzzy entropy of each modal component is calculated, the components with smaller entropy values ​​are screened and the signal is reconstructed to complete the cleaning and denoising of the transformer fault signal data;

[0062] S201: Initializing the population parameters of dung beetles, including the position of the individual population vector of dung beetles [white noise amplitude weight (Nstd) and the number of noise additions (NE)], the number of iterations, etc., and dividing the dung beetle population into rolling dung beetles, egg balls, small dung beetles, and stealing dung beetles in a ratio of 6:6:7:11;

[0063] S202: The rolling dung beetle simulates rolling behavior according to the light intensity and moves along a given direction in the entire search space. During the rolling process, the position of the rolling dung beetle is updated, and the formula is as follows:

[0064] x i (t+1)=x i (t)+α*k*x i (t-1)+b*Δx

[0065] Δx=|x i (t)-x w |

[0066] In the formula, t is the number of iterations, x i (t) is denoted as the position information of the i-th dung beetle at the t-th iteration, k∈(0, 0.2] is a constant, denoted as the deflection coefficient, b∈(0, 1) is a fixed value, α is 1 or -1, -1 means deviation from the original direction, 1 means no deviation from the original direction, x w Denoted as the global worst position, Δx simulates the light intensity change, where k is set to 0.1 and b is set to 0.3;

[0067] When a dung beetle encounters an obstacle and cannot move forward, it needs to reposition itself by dancing in order to obtain a new route. The formula is as follows:

[0068] x i (t+1)=x i (t)+tanθ|x i (t)-x i (t-1)|

[0069] In the formula, θ∈[0,π] is recorded as the deflection angle of the direction. When θ=0, When θ = π, the dung beetle does not update its position;

[0070] S203: Simulation The boundary selection strategy for the dung beetle's egg-laying area. Once the egg-laying area is determined, the female dung beetle will lay eggs, and lay one egg per iteration. The formula is as follows:

[0071] R=1-t / Tmax

[0072] Lb*=max(X*×(1-R), Lb)

[0073] Ub*=min(X*×(1+R),Ub)

[0074] B i (t+1)=X * +b1×(B i (t)-Lb * )+b2*(B i (t)-Ub * )

[0075] In the formula, X * is recorded as the global optimal position, Lb is recorded as the lower bound of the optimization problem, Ub is recorded as the upper bound of the optimization problem, and Lb * Recorded as the lower bound of the spawning area, Ub * The upper limit of the spawning area is B i (t) is denoted as the position information of the i-th egg ball at the t-th iteration, b1 and b2 are denoted as 1×D independent random vectors, and D is denoted as the dimension of the optimization problem;

[0076] S204: The grown-up dung beetles come out to forage for food. An optimal dung beetle area is established to guide the dung beetles to forage for food. The formula is as follows:

[0077] Lb b =max(X b ×(1-R), Lb)

[0078] Ub b =min(X b ×(1+R),Ub)

[0079] x i (t+1)=x i (t)+C1×(x i (t)-Lb b )+C2*(x i (t)-Ub b )

[0080] In the formula, x i (t) is recorded as the position information of the i-th dung beetle at the t-th iteration, Lb b Denote as the lower bound of the optimal feeding area, Ub b Recorded as the upper bound of the optimal feeding area;

[0081] S205: Stealing dung beetles will steal food from other dung beetles. The formula for determining the location of the optimal food source is as follows:

[0082] x i (t+1)=x b +S×g×(|x i (t)-X * |+|x i (t)-X b )

[0083] x i (t) is denoted as the position information of the i-th stealing dung beetle at the t-th iteration, g is denoted as a random vector of size 1×D, serving the normal distribution, and S is a constant value.

[0084] S206: Repeat the above process, update the position information of the dung beetle, determine whether the target agent exceeds the boundary, update the current optimal solution and its fitness value, until the maximum number of iterations is reached, and obtain the optimal parameter combination of the ICEEMDAN algorithm [white noise amplitude weight (Nstd) and noise addition times (NE)];

[0085] S207: Bring the optimal parameter combination [white noise amplitude weight (Nstd) and noise addition times (NE)] into the ICEEMDAN algorithm to decompose the signal and obtain n modal component IMFs;

[0086] S208: Calculate the multi-scale fuzzy entropy (MFE) of each modal component IMF;

[0087] S209: Filter out components with smaller entropy values ​​for signal reconstruction, obtain the reconstructed signal of the unlabeled sample data set and the reconstructed signal of the labeled sample data set, and complete data cleaning and denoising;

[0088] Specifically, after the collected vibration signal is decomposed by DBO-ICEEMDAN, multiple IMF components are obtained, including noisy IMF components dominated by noise and non-noise signals dominated by real signals. The multi-scale fuzzy entropy is calculated. When the component has a smaller entropy value, it means that the signal is more orderly or regular at this scale, while a larger entropy value means that the signal is more chaotic or complex at this scale, with more noise. The threshold value P is set to 0.6, and the IMF components with a threshold value less than 0.6 are screened out, and the signal is reconstructed to complete data cleaning and denoising.

[0089] In the third step, the denoised unlabeled dataset is randomly augmented twice, and positive and negative sample pairs are defined to reconstruct the phase space, complete the contrastive learning pre-training model, and migrate the trained encoder parameters to downstream tasks;

[0090] S301: Based on the MoCo contrastive learning model, the denoised unlabeled one-dimensional vibration signal dataset is improved to construct positive and negative sample pairs, a vibration signal is randomly extracted, and data enhancement is performed twice randomly, wherein the data enhancement method includes but is not limited to random addition of noise, resampling, time reversal, local signal expansion and other methods, to obtain two different but related samples xq and xk0, and define xq and xk0 as positive sample pairs, xq and other signals as negative sample pairs, and xk0 and other signals as negative sample pairs;

[0091] S302: Reconstruct the phase space of the obtained positive and negative samples, restore the one-dimensional vibration signal to the two-dimensional chaotic space, and transform it into a two-dimensional chaotic sequence signal, analyze its dynamic system characteristics, and prevent the reduction of diagnostic accuracy caused by misjudgment and missed judgment. Use the autocorrelation coefficient method to calculate the extension time τ0, use the Cao method to calculate the embedding dimension m0, and calculate the maximum Lyapunov exponent at the same time;

[0092] Specifically, the one-dimensional vibration signal {x a |a=1,2,3,4,...,n} is embedded in m0-dimensional space to obtain the phase point Y0, whose formula is as follows:

[0093]

[0094] In the formula, τ0 is the extension time, and m0 is the embedding dimension. In the phase space reconstruction, the embedding dimension (m) and the delay time (T) are two key parameters that determine the characteristics of the chaotic attractor. If the selected embedding dimension is too small, the attractor may not be able to show its complete structure, resulting in overlapping features. On the contrary, if the embedding dimension is too large, the calculation time will increase and noise will be introduced, which will fail to accurately reflect the dynamic characteristics of the system. For the delay time τ, if the selected value is too small, there will be a strong correlation between the coordinates and the independent characteristics will be lost. On the contrary, if the selected τ value is too large, the relationship between the current moment and the next moment will become more complicated. The autocorrelation coefficient method is used to calculate the extension time τ0, and the Cao method is used to calculate the embedding dimension m0;

[0095] At the same time, the maximum Lyapunov exponent is calculated to determine its chaotic characteristics. The formula is as follows:

[0096]

[0097] Among them, L i is the distance between two adjacent points in phase space reconstruction, L′ i is the distance between two points at the next moment, and M is the total number of iterations.

[0098] S303: Reconstruct the signal x in phase space q The encoder Q is input. The encoder Q uses a residual network (ResNet) and introduces a residual block to prevent gradient explosion and gradient disappearance. First, the encoder Q and momentum encoder K parameters are initialized, and the feature vector q is extracted. At the same time, the signal x reconstructed in the phase space is k0 、x k1 、x k2 、x k3 ...x kn Input momentum encoder K and extract feature vectors k0, k1, k2, k3...k n , and form a queue to form a dynamic dictionary. Its momentum encoder K also uses a residual network (ResNet), which is different from the parameter setting of encoder Q. The momentum update formula between encoder Q and momentum encoder K is as follows:

[0099] θ k ←mθ k +(1-m)θ q

[0100] where m is set to 0.99, θ q is the parameter of encoder Q, θ k is the parameter of momentum encoder K;

[0101] S304: Extract the feature vectors q, k0, k1, k2, k3, ...kn Input into the MLP multilayer perceptron g(.), the MLP multilayer perceptron g(.) is as follows Figure 2 As shown, it contains two fully connected layers and one activation function layer. The activation function is the ReLu function, and the projection vectors q', k0', k1', k2', k3'...k are obtained. n ', the MLP multi-layer perceptron formula is as follows:

[0102] Z i =g(h i )=W (2) σ(W (1) h i )

[0103] Among them, σ is the activation function ReLu function, h i is the feature vector extracted by the encoder,

[0104] h i ∈(q,k0,k1,k2,k3...kn);

[0105] S305: The obtained projection vectors q', k0', k1', k2', k3'...k n ', splice them together, calculate the contrast loss between the positive and negative sample pairs, make the distance between samples with large similarity closer and closer, and the distance between samples with small similarity farther and farther, so as to increase the difference between samples of different categories. The contrast loss function selected by the present invention is InfoNCE, and its formula is as follows:

[0106]

[0107] In the formula, if there is only one k in the queue that can match q, it is recorded as k+, and τ is recorded as the temperature hyperparameter;

[0108] S305: The obtained contrast loss function value and gradient are passed back to the encoder Q. At the same time, the encoder Q and the encoder K are updated with momentum, and 0.01% of the parameters are updated to the momentum encoder K. The new negative sample replaces the old negative sample, and the iteration is repeated until the contrast loss is minimized and the model training is optimal. The loss function adaptively updates the target domain network parameters to improve the generalization ability of the model after migration. The parameters of the trained encoder Q are migrated to the downstream tasks;

[0109] The fourth step is to reconstruct the phase space of the denoised labeled dataset and fine-tune the trained encoder;

[0110] S401: restore the labeled one-dimensional vibration signal data set to a two-dimensional chaotic space and convert it into a two-dimensional chaotic sequence signal.

[0111] S402: Inputting the two-dimensional chaotic sequence signal into the encoder after parameter migration, and fine-tuning the encoder parameters to reduce the number of training parameters.

[0112] In the fifth step, the unlabeled test data is input into the trained encoder after preprocessing and phase space reconstruction;

[0113] S501: normalize the unlabeled test data, decompose and reconstruct it through DBO-ICCEMDAN, and perform data cleaning and noise reduction;

[0114] S502: After the processed unlabeled test data is reconstructed in phase space, the one-dimensional vibration signal is converted into a two-dimensional chaotic sequence signal and input into a trained encoder to extract surface features;

[0115] The sixth step is to input the output features into the classifier for classification to complete the transformer fault diagnosis, such as Figure 1 As shown, the multi-classifier selects the Transformer classifier. Transformer is a neural network architecture based on the attention mechanism. In Transformer, for the data at the corresponding position in the input sequence, positional encoding (PositionalEncoding) is first added so that the model can capture the relative position information of each element in the sequence. The purpose of this is to enable the model to better understand the input sequence without clear order information. Next, through a learnable linear mapping, the position-encoded data is mapped to three sets of vectors: query, key, and value. This linear mapping is completed through the learned weight matrix, and its purpose is to map the original input data to the different representation spaces required by the attention mechanism. Finally, the relationship between these three sets of vectors is mapped and represented through the self-attention mechanism. The calculation formula of the self-attention mechanism can be expressed as follows:

[0116]

[0117] in: Denoted as the dimension of the query vector or key vector, the purpose is to scale the attention weight;

[0118] Q represents the query vector (query), K represents the key vector (key), and V represents the value vector (value).

[0119] Example:

[0120] The present invention first collects 500 groups of unlabeled sample data with a data length of 2048 data points, and 100 groups of four types of labeled sample data, each with a length of 2048 data points, and normalizes the collected data, effectively narrowing the value range and reducing the model learning time.

[0121] The present invention uses a dung beetle optimization algorithm to optimize the best ICEEMDAN parameters of the vibration signal under each fault state. The initialization parameter of the dung beetle optimization algorithm is set to 30, wherein the dung beetle population is divided into rolling ball dung beetles, egg balls, small dung beetles, and stealing dung beetles in a ratio of 6:6:7:11, wherein the optimization problem dimension D is set to 2, the maximum number of iterations is 20, and the white noise amplitude weight (Nstd) is 0.2 and the number of noise additions (NE) is 50. The white noise is brought into ICEEMDAN for modal decomposition, and finally 10 IMF components are obtained. The multi-scale fuzzy entropy (MFE) of the 10 IMF components is calculated respectively, and the threshold p is set to 0.6. The IMF components with a threshold value less than 0.6 are screened out, and finally 6 IMF components that meet the requirements are screened to complete the reconstruction of the signal, and data cleaning and denoising are performed;

[0122] The present invention is based on the idea of ​​MoCo contrast learning model and combines chaos theory to analyze the dynamic characteristics of vibration signals. Specifically, the unlabeled sample data set after noise reduction is randomly extracted, and the dimension is 1*2048. Two data enhancements are randomly performed, wherein the data enhancement method includes but is not limited to random addition of noise, resampling, time series reversal, local signal expansion and other methods, to obtain two different but related samples xq and xk0, and define xq and xk0 as positive sample pairs, xq and other signals as negative sample pairs, and xk0 as negative sample pairs with other signals. The obtained positive and negative samples are reconstructed in phase space and restored to two-dimensional chaotic space, and the extended time τ0 is calculated by the autocorrelation coefficient method, and the embedding dimension m0 is calculated by the Cao method. At the same time, the maximum Lyapunov exponent is calculated to prove that the sample has chaotic characteristics, wherein the signal dimension is 2*2048;

[0123] The two-dimensional chaotic sequence signal is input into the encoder to extract features. The encoder Q and the momentum encoder K use the residual network (ResNet-50), but the parameters of the two encoders are different. The encoder Q participates in the gradient update, and the momentum encoder K does not participate in the gradient update. The momentum update is used between the encoder Q and the momentum encoder K. After passing through the encoder, the final output is a 128-dimensional feature vector, in which the negative samples form a queue to form a dynamic dictionary. The output vector is input into the multi-layer perceptron (MLP), and the activation function is the ReLu function to obtain the projection vectors q', k0', k1', k2', k3'...k n', concatenate the projection vectors together to calculate the contrast loss. Here, the contrast loss function uses the InfoNCE function to maximize the difference between samples. Similar samples are closer and closer, while dissimilar samples are farther and farther away. The contrast loss is used to update the parameters of the encoder Q. At the same time, the momentum is updated between the encoder Q and the momentum encoder K until the contrast loss is minimized and the model is optimal, completing the contrast learning pre-training.

[0124] The present invention combines the idea of ​​transfer learning, transfers the trained encoder parameters to downstream tasks, and inputs the labeled sample data after noise reduction processing into the encoder with parameter transfer after phase space reconstruction, fine-tunes the encoder, and reduces the model training time; at the same time, the unlabeled test data is pre-processed by data cleaning and denoising, and phase space reconstruction is performed, and input into the fine-tuned encoder to extract features, and the extracted feature vector is input into the Transformer classifier to complete transformer fault diagnosis.

[0125] Based on the background of harsh working conditions of transformers, the patent of the present invention proposes a transformer fault diagnosis method based on contrast learning of chaos theory under strong noise background, and combines noise reduction algorithm with deep learning to perform transformer fault diagnosis. Compared with the prior art, the advantage of this patent is that it introduces chaos theory and contrast learning framework based on MoCo model, utilizes the recognition ability of unlabeled samples of self-supervised learning tasks, extracts chaotic features by deep learning, and performs contrast learning to analyze the dynamic characteristics of the system, so as to complete fault diagnosis, and achieves the effect of solving the problem of low diagnostic accuracy caused by strong noise environment and lack of labeled data in transformer fault diagnosis, as well as the problem of misjudgment and missed judgment of faults by transformer fault diagnosis model caused by nonlinear characteristics of vibration signals.

[0126] 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 aforementioned embodiments, it should be understood by those skilled in the art that the technical solutions described in the aforementioned embodiments can still be modified, or some or all of the technical features therein can be replaced by equivalents. Therefore, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A transformer fault diagnosis method based on chaos theory comparative learning under strong noise background, characterized in that: The following steps are involved: The first step is to collect the vibration signal of the transformer box and normalize the vibration signal; In the second step, the Dung Beetle Optimization (DBO) algorithm is used to optimize the optimal ICEEMDAN parameters of the vibration signal under each fault state, and the vibration signal is decomposed by ICEEMDAN to obtain multiple modal components IMF; the multi-scale fuzzy entropy of each modal component is calculated, the component with a larger entropy value is selected and the signal is reconstructed to complete the cleaning and denoising of the transformer fault signal data; The third step is to randomly perform two data enhancements on the denoised unlabeled dataset, define positive and negative sample pairs, reconstruct the phase space, complete the contrastive learning pre-training model, and migrate the trained encoder parameters to downstream tasks. The fourth step is to reconstruct the phase space of the denoised labeled dataset and fine-tune the trained encoder; In the fifth step, the unlabeled test data is input into the trained encoder after preprocessing and phase space reconstruction; In the sixth step, the output features are input into the classifier for classification to complete the transformer fault diagnosis.

2. According to claim 1, a transformer fault diagnosis method based on chaos theory contrast learning under strong noise background is characterized in that: The specific steps of the first step are as follows: Step 1: Collect transformer box vibration signal sample data; Specifically, unlabeled sample data and labeled sample data are collected respectively. The collected labeled sample data are transformer box vibration signals in three fault states, namely, loose transformer winding, loose core, and winding deformation, and in a normal state; Step 2: Normalize the collected signal.

3. The transformer fault diagnosis method based on chaos theory contrast learning under strong noise background according to claim 1 is characterized in that: The specific steps of the second step are as follows: Step 1: Initialize the parameters of the dung beetle population, including the position of the individual dung beetle population vector [white noise amplitude weight (Nstd) and the number of noise additions (NE)], the number of iterations, etc., and divide the dung beetle population into rolling dung beetles, egg balls, small dung beetles, and stealing dung beetles in a ratio of 6:6:7:11; Step 2: The rolling ball dung beetle simulates the rolling ball behavior and moves along the given direction in the entire search space. During the rolling process, the position of the rolling ball dung beetle is updated. When the dung beetle encounters an obstacle and cannot move forward, it needs to reposition itself by dancing in order to obtain a new route. Step 3: Simulation The boundary selection strategy of the dung beetle's egg-laying area. Once the egg-laying area is determined, the female dung beetle will lay eggs, one egg per iteration. Step 4: The grown-up dung beetles come out to forage for food. Establish the best dung beetle area and guide the dung beetles to forage for food. Step 5: Thieves will steal food from other dung beetles and locate the best food source; Step 6: Repeat the above process, update the position information of the dung beetle, determine whether the target agent is beyond the boundary, and update the current optimal solution and its fitness value until the maximum number of iterations is reached, and obtain the optimal parameter combination of the ICEEMDAN algorithm [white noise amplitude weight (Nstd) and number of noise additions (NE)]; Step 7: Bring the optimal parameter combination [white noise amplitude weight (Nstd) and noise addition times (NE)] into the ICEEMDAN algorithm to decompose the signal and obtain n modal component IMFs; Step 8: Calculate the multi-scale fuzzy entropy (MFE) of each modal component IMF; Step 9: Filter out the components with smaller entropy values ​​for signal reconstruction to obtain the reconstructed signal x(t n ).

4. The transformer fault diagnosis method based on chaos theory contrast learning under strong noise background according to claim 1 is characterized in that: The specific steps of the third step are as follows: Step 1: Construct positive and negative sample pairs from the denoised unlabeled one-dimensional vibration signal dataset, randomly extract a vibration signal, and randomly perform data enhancement twice. The data enhancement methods include but are not limited to random addition of noise, resampling, time reversal, local signal amplification, etc., to obtain two different but related samples x q and x k0 , define x q With x k0 are positive sample pairs, x q and other signals are negative sample pairs, x k0 It is also a negative sample pair with other signals; Step 2: Reconstruct the phase space of the obtained positive and negative samples, restore the one-dimensional vibration signal to the two-dimensional chaotic space, and transform it into a two-dimensional chaotic sequence signal. Use the autocorrelation coefficient method to calculate the extension time τ0, use the Cao method to calculate the embedding dimension m0, and calculate the maximum Lyapunov exponent at the same time; Specifically, the maximum Lyapunov exponent is calculated to determine its chaotic characteristics. The formula is as follows: Among them, L i is the distance between two adjacent points in phase space reconstruction, L' i is the distance between two points at the next moment, and M is the total number of iterations; Step 3: Reconstruct the signal x in phase space q The input encoder Q uses a residual network (ResNet) to extract the feature vector q and reconstruct the signal x in phase space. k0 、x k1 、x k2 、x k3 …x kn Input the momentum encoder and extract the feature vectors k0, k1, k2, k3…k n , and form a queue. Its momentum encoder K also uses a convolutional neural network (CNN). The parameter setting is different from that of the encoder Q. The momentum update formula between the encoder Q and the momentum encoder K is as follows: i k ←mθ k +(1-m)θ q where m∈[0,1),θ q is the parameter of encoder Q, θ k is the parameter of momentum encoder K; Step 4: Extract the feature vectors q, k0, k1, k2, k3…k n Input into the MLP multilayer perceptron to obtain the projection vectors q', k0', k1', k2', k3'...k n '; Step 5: Calculate the contrast loss between the positive and negative sample pairs. The contrast loss function selected in the present invention is InfoNCE, and its formula is as follows: In the formula, if there is only one k in the queue that can match q, it is recorded as k + , τ is recorded as the temperature hyperparameter Step 6: Repeat the iteration until the contrast loss is minimized and the model is trained to the best. Then, transfer the parameters of the trained encoder Q to the downstream task.

5. The transformer fault diagnosis method based on chaos theory contrast learning under strong noise background according to claim 1 is characterized in that: The specific steps of the fourth step are as follows: Step 1: Take the labeled one-dimensional vibration signal data set and restore the one-dimensional vibration signal to the two-dimensional chaotic space to become a two-dimensional chaotic sequence signal; Step 2: Input the two-dimensional chaotic sequence signal into the encoder after parameter migration and fine-tune the encoder parameters.

6. The transformer fault diagnosis method based on chaos theory contrast learning under strong noise background according to claim 1 is characterized in that: The specific steps of the fifth step are as follows: Step 1: Normalize the unlabeled vibration signal test sample data; Step 2: Decompose the normalized vibration signal into IMF components using DBO-ICEEMDAN, and select components with smaller entropy values ​​for reconstruction; Step 3: After the data denoising and cleaning, the unlabeled test samples are restored to the two-dimensional chaotic space through phase space reconstruction, converted into a two-dimensional chaotic sequence signal, and input into the encoder that has performed parameter migration.

7. The transformer fault diagnosis method based on chaos theory contrast learning under strong noise background according to claim 1 is characterized in that: The specific steps of the sixth step are as follows: Step 1: Input the output features into the Transformer classifier for classification to determine the fault type.

Images