A transformer online life prediction method under unbalanced multi-source small sample condition

Abstract

The application discloses a transformer online life prediction method under unbalanced multi-source small sample conditions, and steps are as follows: eliminating noise signal of collected transformer vibration signal fault sample data, expanding sample quantity under the premise of reserving minority class sample data characteristics; standardizing collected transformer fault sample data, dividing into a training set, a test set and a verification set, and respectively extracting characteristics; constructing a CNN transformer fault diagnosis model, obtaining fault diagnosis result 1 through the CNN model from the signal processed by RP; constructing a transformer fault diagnosis model, adjusting model hyperparameters by using the verification set, and obtaining fault diagnosis result 2; sending the obtained fault diagnosis result 1 and fault diagnosis result 2 into an FCNN model to adjust the FCNN model, and then obtaining a fault diagnosis result through the test set. The application can effectively solve the misjudgment and missed judgment problems of the minority class fault sample, and improve the diagnosis precision of the model.

Description

Technical Field

[0001] This invention relates to the field of transformer life prediction research, specifically to a transformer life prediction technology under unbalanced multi-source small sample conditions. Background Technology

[0002] As a crucial piece of equipment in power systems, the reliability of transformers directly impacts the stability and security of electricity. With increasing electricity demand and the extending service life of equipment, transformer aging is becoming an increasingly prominent issue. Therefore, online life prediction of transformers has become a research hotspot. Traditional transformer maintenance methods mainly rely on periodic inspections and maintenance, which often fail to detect potential faults in a timely manner, leading to unnecessary economic losses and power outage risks as problems are addressed only after they occur.

[0003] In recent years, with the advancement of sensor technology and data acquisition and processing technology, online monitoring systems have been gradually applied. These systems can collect transformer operating data in real time, including parameters such as temperature, pressure, current, and voltage. By analyzing this data, the health status of the transformer can be assessed. In particular, by monitoring the chemical properties of transformer oil, such as dissolved gas analysis (DGA) and moisture content, the deterioration of the transformer's insulation materials and oil can be effectively determined.

[0004] Furthermore, technologies based on big data and machine learning are increasingly being incorporated into transformer lifespan prediction. By analyzing historical data and building models, aging patterns and failure mechanisms of transformers can be identified, thereby improving prediction accuracy. Combined with IoT technology, transformer condition monitoring and data analysis enable remote management, providing decision support for power companies.

[0005] In summary, research on transformer online life prediction aims to improve the operational reliability and economy of transformers, reduce failure rates, and extend equipment lifespan through advanced monitoring and analysis technologies. With continuous technological advancements, transformer online life prediction will increasingly become a crucial component of power system management, providing strong support for the sustainable development of the power industry. Summary of the Invention

[0006] The purpose of this invention is to propose an online life prediction method for transformers under unbalanced multi-source small sample conditions, which can effectively improve the fault diagnosis accuracy of transformers.

[0007] This invention employs the following technical solution: a method for predicting the online lifetime of a transformer under unbalanced multi-source small sample conditions, comprising the following steps:

[0008] The first step is to use CEEMDAN-CCWK to remove noise signals from the collected transformer vibration signal fault sample data, and then use WGAN-GP to expand the number of samples while retaining the characteristics of minority class sample data.

[0009] The second step is to standardize the collected and generated data and divide them into training, testing, and validation sets.

[0010] The third step is to convert the vibration signals in the training set into threshold-free RP images and extract the vibration signal feature parameters using RQA.

[0011] The fourth step is to construct a CNN transformer fault diagnosis model. The CNN is used to automatically extract the nonlinear features of the RP image and learn them to obtain the preliminary fault diagnosis results.

[0012] The fifth step is to construct an ADACSO-SVM transformer fault diagnosis model, adjust the model hyperparameters using the validation set, and use the feature parameters extracted by RQA as input to the ADACSO-SVM model to obtain preliminary fault diagnosis results.

[0013] The sixth step is to feed the obtained fault diagnosis result 1 and fault diagnosis result 2 into the FCNN model for adjustment, and then obtain the fault diagnosis result through the test set.

[0014] The seventh step is to input the collected vibration signals from the online test into the meta-classifier to achieve online fault diagnosis of the transformer.

[0015] The eighth step is to normalize the collected dissolved gas data in the oil from the online test and then establish a hidden Markov model to determine the aging state of the oil-paper insulation.

[0016] The ninth step involves combining the aging state of the oil-paper insulation with the Weibull risk model to determine the reliability function, predict the remaining life of the oil-paper, and inputting the prediction into the meta-classifier to achieve the prediction of the remaining life of the transformer.

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

[0018] Step 1: Perform CEEMDAN on the transformer vibration signal to decompose the signal into multiple IMFs. Define CCWK below the threshold VThea as weakly correlated spurious components and noise, remove them, and then reconstruct the sensitive IMFs.

[0019] Step 2: Use the WGAN-GP algorithm to expand the sample size;

[0020] Furthermore, the specific steps of the second step are as follows:

[0021] Step 1: After performing MAF filtering on the collected fault sample data, the vibration signal is standardized.

[0022] Step 2: Divide the standardized sample data into training set, test set, and validation set;

[0023] Furthermore, the specific steps of the third step are as follows: after converting the vibration signals of the training set into threshold-free RP images, the vibration signal feature parameters are extracted using RQA;

[0024] Furthermore, the specific steps of the fourth step are as follows:

[0025] Step 1: Define the network structure by stacking convolutional and pooling layers alternately, followed by fully connected layers and an output layer;

[0026] Step 2: Initialize the grid parameters by randomly assigning values ​​to the kernel weights, biases, and filter indices of the convolutional layer;

[0027] Step 3: Use the unstructured data without threshold RP as the input vector of the CNN network, and extract features from the input vector through one-dimensional convolution;

[0028] Step 4: Train the CNN network, including passing the input data through the convolutional layers, pooling layers and fully connected layers in the CNN to obtain the output results, calculating the error between the output results and the true labels, and using the backpropagation algorithm to calculate the gradient of each weight, thereby updating the weights and biases.

[0029] Step 5: Obtain fault diagnosis result 1 through the test set;

[0030] Furthermore, the specific steps of the fifth step are as follows:

[0031] Step 1: Use the Booststrap sampling method to extract samples of the same size from the training set N to generate a training subset;

[0032] Step 2: Construct an SVM model, using radial basis functions as the SVM kernel function, and determine the optimal classification surface decision function;

[0033] Step 3: Determine the fitness function, initialize the parameters memory pool, change domain, number of changes, maximum number of iterations, and the position of each individual cat, and use the maximum value of the combination of the penalty factor and kernel function of the SVM model as the optimization condition;

[0034] Step 4: Begin iteration and update the individual's position. Compare the fitness value of the new position with the fitness value of the original position. If the fitness value of the new position is smaller, update the individual's position; otherwise, keep the original position unchanged.

[0035] Step 5: Stop iterating after reaching the maximum number of iterations and output the optimal solution; otherwise, return to step 3 and iterate to search. If the termination condition is met, output the optimal individual, which is the optimal solution found by the algorithm; otherwise, return to step 4.

[0036] Step 6: Construct the ADACSO-SVM transformer fault diagnosis model, adjust the model structure and hyperparameters using the validation set, and then save the diagnosis model;

[0037] Step 7: Obtain fault diagnosis result 2 through the test set;

[0038] Furthermore, the specific steps of the sixth step are as follows:

[0039] Step 1: Define the network structure by stacking convolutional and pooling layers alternately, followed by fully connected layers and an output layer;

[0040] Step 2: Initialize the grid parameters by randomly assigning values ​​to the kernel weights, biases, and filter indices of the convolutional layer;

[0041] Step 3: Use fault diagnosis result 1 and fault diagnosis result 2 as input vectors for the FCNN network, and extract features from the input vectors;

[0042] Step 4: Train the FCNN network, including passing the input data through the convolutional layers, pooling layers and fully connected layers in FCNN to obtain the output results, calculating the error between the output results and the true labels, and using the backpropagation algorithm to calculate the gradient of each weight, thereby updating the weights and biases.

[0043] Furthermore, the specific steps of the seventh step are as follows: input the collected online test set into the optimal classifier to realize online fault diagnosis of the transformer;

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

[0045] Step 1: Normalize the collected dissolved gas data in the oil;

[0046] Step 2: Before building the Hidden Markov Model, the Bayesian information criterion is used to determine the number of hidden states in the model;

[0047] Step 3: Given the number of hidden states, extract the different features of the observed sequence and perform clustering using the k-means clustering algorithm;

[0048] Step 4: Use the Viterbi algorithm to identify the state of the observed sequence and thus determine the aging state of the oil paper insulation;

[0049] Furthermore, the specific steps of the ninth step are as follows:

[0050] Step 1: Determine the weights of characteristic parameters of dissolved gases in oil using the entropy weight method;

[0051] Step 2: Use the maximum likelihood estimation method to estimate the parameters of the Weibull proportional hazards model, which yields the shape parameter β, scale parameter η, and regression parameter γ, and determines the reliability function.

[0052] Step 3: Based on the aging state of the oil-paper insulation and the reliability function, predict the remaining life of the oil-paper insulation. Attached Figure Description

[0053] Figure 1 This is a flowchart of the method involved in this invention;

[0054] Figure 2 This is a schematic diagram showing the lifespan of the oil-paper insulation under different aging conditions according to the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The specific embodiments described herein are merely illustrative and are not intended to limit the invention.

[0056] 363 transformer fault samples were collected, representing the row numbers in the DGA sample matrix. These include 82 normal faults, 71 high-temperature overheating faults, 67 high-energy discharge faults, 52 medium-low temperature overheating faults, 39 low-energy discharge faults, 30 partial discharge faults, and 22 cases of discharge combined with overheating. Dissolved gases in the oil include H2, CH4, C2H4, C2J2, and C2H6, representing the column numbers in the DGA sample matrix. Based on whether a transformer fault occurred and what type of fault it was, transformers can be classified into normal, high-temperature overheating, high-energy discharge, medium-low temperature overheating, low-energy discharge, partial discharge, and discharge combined with overheating, and labeled from 1 to 7.

[0057] The first step is to use CEEMDAN-CCWK to remove noise signals from the collected transformer vibration signal fault sample data, and then use WGAN-GP to expand the number of samples while retaining the characteristics of minority class sample data.

[0058] S101: Perform CEEMDAN analysis on the transformer vibration signal, decomposing the signal into multiple IMFs.

[0059] S102: Calculate the IMF (value is Y) separately. IMF The correlation coefficient ρ, kurtosis K, and CCWK (with a value of M) CCWK )

[0060]

[0061] M CCWK =ρK

[0062] In the formula: x is the original signal of transformer vibration; Cov(*) is the signal covariance; E(*) is the signal expectation; D(*) is the signal variance.

[0063] S103: Define CCWK as being below the threshold V Thea Weakly correlated spurious components and noise were removed before reconstructing the sensitive IMF.

[0064]

[0065] Where: M CCWKk The CCWK value of the k-th IMF; The mean CCWK of IMF

[0066] S104: Extract the reconstructed signal, and measure the signal-to-noise ratio (value E). SNR ) and root mean square error (the magnitude of which is E) RMSE The overall denoising effect is evaluated. SNR is the ratio of the power of the denoised signal to the power of the noise, and RMSE is the square root of the ratio of the noise power to the number of observations.

[0067]

[0068] Where: h l The pure signal from the l-th observation; x l is the noisy signal from the l-th observation; T is the number of observations.

[0069] S105: The WGAN-GP algorithm is used to expand the number of samples;

[0070] The objective function of WGAN-GP can be expressed as:

[0071] Where D is the discriminator model; G is the generator model; p data (x) is the true data distribution; p z (z) is the noise distribution; z is a vector sampled from the noise distribution; G(z) is the generator output that maps the noise vector z to the data sample; λ is a hyperparameter in the added gradient penalty term; It is a linear interpolation point between real data and generated data; It is its probability distribution.

[0072] The second step involves standardizing the collected and generated data, and dividing it into training, testing, and validation sets. The specific steps are as follows:

[0073] S201: Using MAF filtering to filter signal power using finite impulse response, the calculation formula is as follows:

[0074]

[0075] In the formula, P and Let P be the signal power before and after filtering, respectively. Calculate P and... The ratio of these values ​​results in a dimensionless signal with values ​​near 1. Shifting the overall signal by 1 in the negative direction yields a standardized signal near 0.

[0076] S202: Divide the collected and generated data into training, testing, and validation sets in a ratio of 7:2:1.

[0077] The third step involves converting the vibration signals in the training set into threshold-free RP images and extracting vibration signal feature parameters using RQA. The specific steps are as follows:

[0078] S301: Convert the vibration signals in the training set into threshold-free RP images.

[0079] Suppose we have a sequence t = {t1, t2, ..., tn} of length n. n According to Takens embedding theory and time delay methods, this sequence can be reconstructed in phase space to obtain a series of new phase space trajectories T:

[0080] T(n-(m-1)τ)={t n-(m-1)τ ,t n-(m-2)τ ,……,t nτ )

[0081] Where: m is the embedding dimension; τ is the time delay; T(i) represents the i-th reconstructed phase space trajectory of t.

[0082]

[0083] Where: ||·|| represents the Euclidean norm; Θ(·) represents the Heaviside function; ε is the recursive threshold parameter; R(i,j) is a simulation of length n-(m-1)τ. The resulting graph after plotting R is an RP image.

[0084] S302: Perform RQA analysis on the RP image to obtain structured feature data, specifically:

[0085]

[0086] Where: n-(m-1)τ

[0087]

[0088] Where: l represents the length of the diagonal; l minP(l) represents the minimum diagonal length; P(l) represents the frequency of a diagonal of length l. A higher DET value indicates a more stable system, meaning it is not a random system.

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095] Where: l min ,v min ,w min These are the minimum values ​​of the diagonal length, vertical line length, and white vertical line length, respectively; p(l), p(v), and p(w) represent the proportions of the diagonal line of length l, the vertical line of length v, and the white vertical line of length w, respectively.

[0096]

[0097] Among them, the intermittency of LAN and RP points is related.

[0098]

[0099] Where TT is the average length of the vertical line, which estimates the average time it takes for an RP state to be captured.

[0100] The fourth step is to construct a CNN transformer fault diagnosis model. The CNN is used to automatically extract the nonlinear features of the RP image and learn them to obtain the preliminary fault diagnosis result 1. The specific steps are as follows:

[0101] S401: Construct a CNN-based transformer fault diagnosis model. Use CNN to automatically extract nonlinear features from the RP image and learn them to obtain preliminary fault diagnosis results.

[0102] Specifically, the CNN network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer;

[0103] The convolution process expression of the convolutional layer is:

[0104]

[0105] In the formula, σ(·) represents the f-th feature map of the e-th convolutional layer; σ(·) is the non-linear activation function, and * denotes the convolution operation; Let f be the weight matrix of the e-th convolutional layer; This represents the f-th feature map of the (e-1)-th convolutional layer. This is the f-th bias vector of the (e-1)-th convolutional layer;

[0106] The pooling process expression for the pooling layer is as follows:

[0107]

[0108] in, For the feature output of the pooling layer, max(·) represents the downsampling function for the maximum value. is the feature vector output by the convolutional layer, and s is the pooling size;

[0109] The activation function layer uses the ReLU activation function, whose expression is:

[0110] σ r (x) = max{0,x}

[0111] The fully connected layer is used to perform feature classification and learn the global features of the input data.

[0112] The output layer uses SVM to classify risk categories and generate the final risk assessment result, the expression of which is:

[0113]

[0114] Where w is the weight vector, x χ Let y be the feature vector corresponding to the x-th sample. χ Let ξ be the label of the x-th sample, b be the threshold, ε be the number of samples, and C be the regularization parameter used to control the degree of penalty for misclassification; χ is the slack variable for the x-th sample, used to allow samples to be misclassified.

[0115] The fifth step involves constructing an ADACSO-SVM transformer fault diagnosis model, adjusting the model hyperparameters using the validation set, and inputting the feature parameters extracted by RQA into the ADACSO-SVM model to obtain preliminary fault diagnosis results. The specific steps are as follows:

[0116] S501: Using the Booststrap sampling method, samples of the same size are extracted from the training set N to generate a training subset;

[0117] S502: Construct an SVM model, using radial basis functions as the SVM kernel function. The optimal classification surface decision function is:

[0118]

[0119] S503: Determine the fitness function, initialize parameters such as memory pool, domain of variation, number of variations, maximum number of iterations, and the position of each individual cat. Use the maximum value of the combination of the SVM model's penalty factor and kernel function as the optimization condition, specifically:

[0120]

[0121] S504: Calculate the fitness value for each cat, record the optimal fitness value and the corresponding position vector, and then update the optimal solution;

[0122] S505: If the termination condition is met, output the optimal individual, i.e., the optimal solution found by the algorithm; otherwise, return to step four.

[0123] S506: Construct an ADACSO-SVM transformer fault diagnosis model, adjust the model structure and hyperparameters using the validation set, and then save the diagnosis model;

[0124] S507: Obtain fault diagnosis results 2 through the test set;

[0125] The sixth step involves feeding the obtained fault diagnosis result 1 and fault diagnosis result 2 into the FCNN model for adjustment, and then obtaining the fault diagnosis result through the test set. The specific steps are as follows:

[0126] S601: Combine the prediction results of the two base classifiers, CNN and ADACSO, into a new dataset;

[0127] S602: Construct the FCNN model. The FCNN network includes an input layer, hidden layers, and an output layer, as detailed below:

[0128] The input layer uses the newly generated dataset as input;

[0129] The hidden layers use the ReLU activation function, and the output layer uses the tanh activation function.

[0130]

[0131] In the formula: x is the input value of the neuron.

[0132] The mean absolute error (MAE) is selected as the loss function of the model to reflect the difference between the predicted label value and the actual label value of the vibration signal:

[0133]

[0134] In the formula: θ i This represents the actual value of the i-th label; Let be the predicted value for the i-th label.

[0135] S603: Using the newly generated dataset as input to FCNN, the preliminary prediction results are further optimized to obtain the final diagnosis result.

[0136] The seventh step is to input the collected online test vibration signals into the meta-classifier to realize online fault diagnosis of the transformer. Specifically, the collected online test vibration signals are input into the meta-classifier to realize online fault diagnosis of the transformer.

[0137] Step 8: After normalizing the collected dissolved gas data in the oil from the online test, a hidden Markov model is established to determine the aging state of the oil-paper insulation. The specific steps are as follows:

[0138] S801: Normalize the collected dissolved gas data in the oil. Dissolved gases in the oil include H2, CO, CO2, and total alkynes (the sum of CH4, C2H2, C2H4, and C2H6), specifically:

[0139]

[0140] In the formula, x' represents the normalized data of the variable, x represents the original value of the variable, and x' represents the normalized data of the variable. min This represents the minimum value of the variable, x. max This indicates the maximum value of the variable.

[0141] S802: The hidden states of the Hidden Markov Model are divided into state 1 (normal state), state 2 (early aging state), state 3 (aging state), and state 4 (severe aging state). The BIC criterion is used to determine the optimal Hidden Markov Model. Then, the k-means algorithm is used to divide the feature parameter matrix into 4 classes, and the dissolved gas content in the oil is classified into states according to the IEEE standard.

[0142] S803: Set the initial state of the Hidden Markov Model to the normal state, use the Baum-Welch algorithm to train the model based on the observation sequence, and use the Viterbi algorithm to identify the state of the observation sequence, thereby obtaining the log-likelihood function value corresponding to each state.

[0143] S804: Determine the proportion of each characteristic parameter in the whole using the entropy weight method.

[0144] Step 9: Combine the aging state of the oil-paper insulation with the Weibull risk model to determine the reliability function, predict the remaining life of the oil-paper, and input it into the meta-classifier to achieve the prediction of the transformer's remaining life. The specific steps are as follows:

[0145] S901: A reliability function is established using a Weibull proportional hazards model combined with a hidden Markov model, taking aging status as a covariate affecting the remaining life of the transformer's oil-paper insulation. The specific expression is as follows:

[0146]

[0147] In the formula: k = 0, 1, 2, ..., kΔ ≤ t ≤ (k+1)Δ; h(t, Z) k ) is the failure rate function; h0(t) is the basic failure rate that depends only on time t and follows a Weibull distribution; β is the shape parameter of the Weibull distribution; η is the scale parameter; Z k γ is a covariate, related only to the aging state of the transformer oil-paper insulation; γ is a regression parameter, reflecting the degree of influence of aging state on the failure rate.

[0148] The reliability function is:

[0149]

[0150] The equipment's operating life at time t is:

[0151]

[0152] S902: Using the maximum likelihood estimation method, the parameters of the Weibull proportional hazards model are estimated, yielding the shape parameter β, the scale parameter η, and the regression parameter γ.

[0153] S903: Calculates the remaining lifespan of the oil paper and inputs it into the meta-classifier; outputs it when there is no sudden failure.

[0154] Example:

[0155] First, 355 transformer fault samples were collected, including 80 normal cases, 70 cases of high-temperature overheating, 66 cases of high-energy discharge, 51 cases of medium-low temperature overheating, 38 cases of low-energy discharge, 29 cases of partial discharge, and 21 cases of discharge combined with overheating. Then, CEEMDAN-CCWK was used to eliminate noise signals in the samples. Then, WGAN-GP was used to expand the number of samples while retaining the data characteristics of minority classes. The distribution of the number of samples before and after processing is shown in Table 1.

[0156] Secondly, the collected and generated data are standardized and divided into training, testing, and validation sets.

[0157] Table 1 Comparison of training samples before and after preprocessing

[0158]

[0159] Finally, the vibration signal was converted into a threshold-free RP image and features were extracted using RQA.

[0160] This embodiment uses CNN and SVM models as base classifiers and FCNN models as meta classifiers. The ADACSO algorithm is employed to optimize the number and depth of decision trees in the SVM model. The optimization range for the number of decision trees is [0, 200], the optimization range for the decision tree depth is [0, 70], the population size is 10, and the maximum number of iterations is 100. The threshold-free RP image is input into the CNN classifier, which automatically extracts features and diagnoses the fault, obtaining diagnosis result 1. The features extracted by RQA are input into the SVM classifier to diagnose the ADACSO-SVM transformer fault diagnosis model, obtaining diagnosis result 2. Both fault diagnosis results 1 and 2 are then input into the FCNN model to obtain the transformer diagnosis result.

[0161] To verify the superiority of this integrated fault diagnosis method, it was compared with five other models: ADACSO-SVM, SO-SVM, CNN, SVM, and ELM. All models used the same input data. The superiority of this method was evaluated using recall, precision, and F1 score, as shown in Table 2.

[0162] Table 2 Comparison of Multi-Model Diagnostic Accuracy

[0163]

[0164] As can be seen from Table 2, this method has a high accuracy rate in diagnosing faults under unbalanced small sample conditions.

[0165] To verify the reliability of this life prediction method, life prediction was performed based on the collected DGA data corresponding to the above transformer vibration signal samples. Some dissolved gas data in the oil are shown in Table 3. Then, the BIC criterion was used to determine the optimal hidden Markov model, and the hidden states were divided into four states: state 1 (normal state), state 2 (early aging state), state 3 (aging state), and state 4 (severe aging state). After the division was completed, the hidden Markov model was trained.

[0166] Table 3. Partial data on dissolved gases in oil.

[0167]

[0168] After determining the initial conditions of the model, the Baum-Welch algorithm is used to train the model based on the observation sequence. A corresponding Hidden Markov Model is trained for each hidden state. After training, the Viterbi algorithm is used to identify the state of the observation sequence, i.e., calculating the probability P(O|λi), i=1,2,3,4, to obtain the log-likelihood function value for each state. The state corresponding to the maximum log-likelihood function value represents the current state of the observation sequence.

[0169] Finally, the characteristic parameters are processed using the entropy weight method, yielding the weights P = [0.270, 0.292, 0.262, 0.176] for H2, CO, CO2, and total hydrocarbons. The maximum likelihood estimation method is then used to estimate the parameters of the Weibull proportional hazards model, obtaining the shape parameter β, scale parameter η, and regression parameter γ. This yields the insulation life of the oil-paper under different aging states, as shown in the figure. Figure 2 As shown.

[0170] Depend on Figure 2 It can be seen that when the reliability of the paper insulation is 0.9, the remaining life under normal conditions is 10.45 years; when the reliability is 0.8, the remaining life under normal conditions is 15.58 years; and when the reliability is 0.7, the remaining life under normal conditions is 20.02 years. This is basically consistent with the actual operating life of paper insulation, which is about 20 years. This proves that the method has a high accuracy in lifetime prediction under unbalanced multi-source small sample conditions.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; therefore, these modifications or substitutions 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 method for predicting the online lifespan of a transformer under unbalanced multi-source small sample conditions, characterized in that, Includes the following steps: The first step is to use CEEMDAN-CCWK to remove noise signals from the collected transformer vibration signal fault sample data, and then use WGAN-GP to expand the number of samples while retaining the characteristics of minority class sample data. The second step is to standardize the collected and generated data and divide them into training, testing, and validation sets. The third step is to convert the vibration signals in the training set into threshold-free RP images and then use RQA to extract the feature parameters of the vibration signals. The fourth step is to construct a CNN transformer fault diagnosis model. The CNN is used to automatically extract the nonlinear features of the RP image and learn them to obtain the preliminary fault diagnosis results. The fifth step is to construct an ADACSO-SVM transformer fault diagnosis model, adjust the model hyperparameters using the validation set, and use the feature parameters extracted by RQA as input to the ADACSO-SVM model to obtain preliminary fault diagnosis results. The sixth step is to feed the obtained fault diagnosis result 1 and fault diagnosis result 2 into the FCNN model for adjustment, and then obtain the fault diagnosis result through the test set. The seventh step is to input the collected vibration signals from the online test into the meta-classifier to achieve online fault diagnosis of the transformer. The eighth step is to normalize the collected dissolved gas data in the oil from the online test and then establish a hidden Markov model to determine the aging state of the oil-paper insulation. The ninth step involves combining the aging state of the oil-paper insulation with the Weibull risk model to determine the reliability function, predict the remaining life of the oil-paper, and inputting the prediction into the meta-classifier to achieve the prediction of the remaining life of the transformer. The specific steps of the fifth step are as follows: Step 1: Use the Booststrap sampling method to extract samples of the same size from the training set N to generate a training subset; Step 2: Construct an SVM model, using radial basis functions as the SVM kernel function, and determine the optimal classification surface decision function; Step 3: Determine the fitness function, initialize the parameters memory pool, change domain, number of changes, maximum number of iterations, and the position of each individual cat, and use the maximum value of the combination of the penalty factor and kernel function of the SVM model as the optimization condition; Step 4: Start the iteration, update the individual's position, compare the fitness value of the new position with the fitness value of the original position, if the fitness value of the new position is smaller, update the individual's position, otherwise keep the original position unchanged; Step 5: Stop iterating after reaching the maximum number of iterations and output the optimal solution; otherwise, return to step 3 and iterate to search. If the termination condition is met, output the optimal individual, which is the optimal solution found by the algorithm; otherwise, return to step 4. Step 6: Construct the ADACSO-SVM transformer fault diagnosis model, adjust the model structure and hyperparameters using the validation set, and then save the diagnosis model; Step 7: Obtain fault diagnosis result 2 through the test set.

2. The method for predicting the online lifespan of a transformer under unbalanced multi-source small sample conditions according to claim 1, characterized in that, The specific steps of the first step are as follows: Step 1: Perform CEEMDAN on the transformer vibration signal to decompose the signal into multiple IMFs. Define CCWK below the threshold VThea as weakly correlated spurious components and noise, remove them, and then reconstruct the sensitive IMFs. Step 2: Use the WGAN-GP algorithm to expand the number of samples.

3. The method for predicting the online lifespan of a transformer under unbalanced multi-source small sample conditions according to claim 1, characterized in that, The specific steps of the second step are as follows: Step 1: After performing MAF filtering on the collected fault sample data, the vibration signal is standardized. Step 2: Divide the standardized sample data into training set, test set and validation set.

4. The method for predicting the online lifespan of a transformer under unbalanced multi-source small sample conditions according to claim 1, characterized in that, The specific steps of the third step are as follows: After converting the vibration signal of the training set into a threshold-free RP image, the vibration signal feature parameters are extracted using RQA.

5. The method for predicting the online lifespan of a transformer under unbalanced multi-source small sample conditions according to claim 1, characterized in that, The specific steps of the fourth step are as follows: Step 1: Define the network structure by stacking convolutional and pooling layers alternately, followed by fully connected layers and an output layer; Step 2: Initialize the grid parameters by randomly assigning values ​​to the kernel weights, biases, and filter indices of the convolutional layer; Step 3: Use the unstructured data without threshold RP as the input vector of the CNN network, and extract features from the input vector through one-dimensional convolution; Step 4: Train the CNN network, including passing the input data through the convolutional layers, pooling layers and fully connected layers in the CNN to obtain the output results, calculating the error between the output results and the true labels, and using the backpropagation algorithm to calculate the gradient of each weight, thereby updating the weights and biases. Step 5: Obtain the fault diagnosis result 1 through the test set.

6. The method for predicting the online lifespan of a transformer under unbalanced multi-source small sample conditions according to claim 1, characterized in that, The specific steps of the sixth step are as follows: Step 1: Define the network structure by stacking convolutional and pooling layers alternately, followed by fully connected layers and an output layer; Step 2: Initialize the grid parameters by randomly assigning values ​​to the kernel weights, biases, and filter indices of the convolutional layer; Step 3: Use fault diagnosis result 1 and fault diagnosis result 2 as input vectors for the FCNN network, and extract features from the input vectors; Step 4: Train the FCNN network, including passing the input data through the convolutional layers, pooling layers and fully connected layers in FCNN to obtain the output results, calculating the error between the output results and the true labels, and using the backpropagation algorithm to calculate the gradient of each weight, thereby updating the weights and biases.

7. The method for predicting the online lifespan of a transformer under unbalanced multi-source small sample conditions according to claim 1, characterized in that, The specific steps of the seventh step are as follows: input the collected online test set into the optimal classifier to realize online fault diagnosis of transformer.

8. The method for predicting the online lifespan of a transformer under unbalanced multi-source small sample conditions according to claim 1, characterized in that, The specific steps of the eighth step are as follows: Step 1: Normalize the collected dissolved gas data in the oil; Step 2: Before building the Hidden Markov Model, the Bayesian information criterion is used to determine the number of hidden states in the model; Step 3: Given the number of hidden states, extract the different features of the observed sequence and perform clustering using the k-means clustering algorithm; Step 4: Use the Viterbi algorithm to identify the state of the observed sequence and then determine the aging state of the oil paper insulation.

9. The method for predicting the online lifespan of a transformer under unbalanced multi-source small sample conditions according to claim 1, characterized in that, The specific steps of the ninth step are as follows: Step 1: Determine the weights of characteristic parameters of dissolved gases in oil using the entropy weight method; Step 2: Use the maximum likelihood estimation method to estimate the parameters of the Weibull proportional hazards model, which yields the shape parameter β, scale parameter η, and regression parameter γ, and determines the reliability function; Step 3: Based on the aging state of the oil-paper insulation and the reliability function, predict the remaining life of the oil-paper insulation.

Images