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

Abstract

The invention discloses a transformer online life prediction method under an unbalanced multi-source small sample condition, which comprises the following steps of: eliminating noise signals from collected transformer vibration signal fault sample data, and expanding the number of samples on the premise of retaining minority class sample data characteristics; performing standardization processing on the acquired transformer fault sample data, dividing the data into a training set, a test set and a verification set, and respectively extracting features; a CNN transformer fault diagnosis model is constructed, and a fault diagnosis result 1 is obtained through the CNN model of the signals processed by the RP; constructing a transformer fault diagnosis model, and adjusting model hyper-parameters by using the verification set to obtain a fault diagnosis result 2; and sending the obtained fault diagnosis result 1 and the fault diagnosis result 2 into an FCNN model to adjust the FCNN model, and then obtaining a fault diagnosis result through the test set. According to the method, the problems of misjudgment and missed judgment of minority class fault samples can be effectively solved, and the diagnosis precision of the model is improved.

Description

Technical Field

[0001] The present invention relates to the research field of transformer life prediction, and specifically to the technical field of transformer life prediction under unbalanced multi-source small sample conditions. Background Art

[0002] As an important equipment in the power system, the reliability of the transformer directly affects the stability and safety of the power. With the growth of power demand and the extension of equipment service life, the aging problem of transformers has become increasingly prominent. Therefore, online life prediction of transformers has become a hot topic of research. Traditional transformer maintenance methods mainly rely on regular inspections and overhauls, which often fail to detect potential faults in time, resulting in equipment being handled only after the fault occurs, causing unnecessary economic losses and power outage risks.

[0003] In recent years, with the advancement of sensor technology, data acquisition and processing technology, online monitoring systems have gradually been applied. These systems can collect real-time transformer operating data, including parameters such as temperature, pressure, current and voltage. By analyzing these data, the health status of the transformer can be evaluated. 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 judged.

[0004] In addition, technologies based on big data and machine learning are gradually being introduced into transformer life prediction. By analyzing historical data and building models, the aging mode and failure mechanism of transformers can be identified, thereby improving the accuracy of predictions. Combined with IoT technology, transformer status monitoring and data analysis can achieve remote management and provide decision support for power companies.

[0005] In general, the research on online transformer life prediction aims to improve the operating reliability and economy of transformers, reduce the failure rate and extend the service life of equipment through advanced monitoring and analysis technology. With the continuous development of technology, online transformer life prediction will become an increasingly important part of power system management, providing a strong guarantee for the sustainable development of the power industry. Summary of the invention

[0006] The purpose of the present invention is to propose a transformer online life prediction method under unbalanced multi-source small sample conditions, so as to effectively improve the transformer fault diagnosis accuracy.

[0007] The present invention is implemented by the following technical solution: a transformer online life prediction method under unbalanced multi-source small sample conditions, comprising the following steps:

[0008] In the first step, the collected transformer vibration signal fault sample data is used to eliminate the noise signal in the sample using CEEMDAN-CCWK, and then WGAN-GP is used to expand the number of samples while retaining the characteristics of the minority class sample data;

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

[0010] In the third step, the vibration signals of the training set are converted into non-threshold RP images and the vibration signal feature parameters are extracted using RQA;

[0011] The fourth step is to build a CNN transformer fault diagnosis model, use CNN to automatically extract the nonlinear features of the RP image and obtain the preliminary fault diagnosis results after learning 1;

[0012] Step 5: Construct the ADACSO-SVM transformer fault diagnosis model, use the validation set to adjust the model hyperparameters, and use the feature parameters extracted by RQA as the input of the ADACSO-SVM model to obtain the preliminary fault diagnosis results 2;

[0013] Step 6: Send the obtained fault diagnosis result 1 and fault diagnosis result 2 into the FCNN model to adjust the FCNN model, and then obtain the fault diagnosis result through the test set;

[0014] Step 7: Input the collected vibration signals in the online test set into the meta-classifier to realize online fault diagnosis of transformer;

[0015] In the eighth step, after normalizing the collected data of dissolved gas in oil in the online test set, a hidden Markov model is established to determine the aging state of oil-paper insulation;

[0016] In the ninth step, the aging state of the oil-paper insulation is combined with the Weibull ratio risk model to determine the reliability function, predict the remaining life of the oil-paper insulation, and input it into the meta-classifier to realize the remaining life prediction 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 false components and noise, remove them and reconstruct sensitive IMFs.

[0019] Step 2: Use the WGAN-GP algorithm to expand the number of samples;

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

[0021] Step 1: After the collected fault sample data is subjected to MAF filtering, 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: converting the vibration signal of the training set into a threshold-free RP image and then using RQA to extract the characteristic parameters of the vibration signal;

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

[0025] Step 1: Define the network structure, alternately stack convolutional layers and pooling layers, and connect to the fully connected layer and output layer;

[0026] Step 2: Initialize the grid parameters and randomly assign values ​​to the convolution kernel weights, biases, and filter index values ​​of the convolution layer;

[0027] Step 3: Use the unstructured data threshold-free 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 layer, pooling layer, and fully connected layer in the CNN to obtain the output result, calculate the error between the output result and the true label, and use the back propagation algorithm to calculate the gradient of each weight, thereby updating the weight and bias;

[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 the SVM model, use the radial basis function as the SVM kernel function, and determine the optimal classification surface decision function;

[0033] Step 3: Determine the fitness function, initialize the parameter memory pool, change domain, change number, maximum number of iterations, and the position of each cat individual, 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: Start iteration and update individual positions. The fitness value of the new position is compared with the fitness value of the original position. If the fitness value of the new position is smaller, the individual position is updated, otherwise the original position remains unchanged. ;

[0035] Step 5: After reaching the maximum number of iterations, stop the iteration and output the optimal solution. Otherwise, return to step 3 and iterate the search. If the termination condition is reached, the optimal individual is output, that is, the optimal solution found by the algorithm; otherwise, return to step 4;

[0036] Step 6: Build the ADACSO-SVM transformer fault diagnosis model, use the validation set to adjust the model structure and hyperparameters, 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, alternately stack convolutional layers and pooling layers, and connect to the fully connected layer and output layer;

[0040] Step 2: Initialize the grid parameters and randomly assign values ​​to the convolution kernel weights, biases, and filter index values ​​of the convolution layer;

[0041] Step 3: Use fault diagnosis result 1 and fault diagnosis result 2 as input vectors of the FCNN network and perform feature extraction on the input vectors;

[0042] Step 4: Train the FCNN network, including passing the input data through the convolutional layer, pooling layer, and fully connected layer in the FCNN to obtain the output result, calculate the error between the output result and the true label, and use the back propagation algorithm to calculate the gradient of each weight, thereby updating the weight and bias.

[0043] Furthermore, the specific steps of the seventh step are as follows: inputting 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 in oil data;

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

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

[0048] Step 4: Use the Viterbi algorithm to identify the state of the observation sequence and 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 characteristic parameter weights of dissolved gas in oil by entropy weight method;

[0051] Step 2: Use the maximum likelihood estimation method to estimate the parameters of the Weibull proportional hazard model, obtain the shape parameter β, scale parameter η and regression parameter γ, and determine the reliability function;

[0052] Step 3: Predict the remaining life of the oil-paper insulation based on the aging state of the oil-paper insulation and the reliability function. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flowchart of the method involved in the present invention;

[0054] Figure 2 It is a schematic diagram of the oil-paper insulation life under different aging conditions of the present invention. DETAILED DESCRIPTION

[0055] 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.

[0056] 363 transformer fault samples were collected, representing the number of rows in the DGA sample matrix. Among them, 82 were normal, 71 were high-temperature overheating, 67 were high-energy discharge, 52 were medium-low temperature overheating, 39 were low-energy discharge, 30 were partial discharge, and 22 were discharge and overheating. The dissolved gases in the oil include H2, CH4, C2H4, C2J2, and C2H6, representing the number of columns in the DGA sample matrix. According to whether the transformer has a fault and what kind of fault exists, the transformer can be divided into normal, high-temperature overheating, high-energy discharge, medium-low temperature overheating, low-energy discharge, partial discharge, and discharge and overheating, and labeled 1 to 7.

[0057] In the first step, the collected transformer vibration signal fault sample data is used to eliminate the noise signal in the sample using CEEMDAN-CCWK, and then WGAN-GP is used to expand the number of samples while retaining the characteristics of the minority class sample data;

[0058] S101: Perform CEEMDAN on the transformer vibration signal and decompose the signal into multiple IMFs

[0059] S102: Calculate IMF (value is Y IMF )'s correlation coefficient ρ, kurtosis K, CCWK (value is M CCWK )

[0060]

[0061] M CCWK =ρK

[0062] Where: 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 is lower than the threshold V Thea The weakly correlated false components and noise are removed and the sensitive IMF is reconstructed

[0064]

[0065] Where: M CCWKk is the CCWK value of the kth I MF; is the CCWK mean of I MF

[0066] S104: Extract the reconstructed signal and calculate the signal-to-noise ratio (value is E SNR ) and the root mean square error (value is E RMSE ) Comprehensively evaluate the denoising effect. SNR is the power ratio of the denoised signal to 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 is the pure signal of the lth observation; x l is the noisy signal observed for the lth time; T is the number of observations.

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

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

[0071] Among them, D is the discriminator model; G is the generator model; p data (x) is the real data distribution; p z (z) is the noise distribution; z is the vector sampled from the noise distribution; G(z) is the output of the generator mapping the noise vector z to data samples; λ is the hyperparameter in the added gradient penalty term; is the linear interpolation point between the real data and the generated data; is its probability distribution.

[0072] The second step is to standardize the collected and generated data and divide them into training sets, test sets, and validation sets. The specific steps are as follows:

[0073] S201: Use MAF filtering to filter the signal power with a finite impulse response, and the calculation formula is:

[0074]

[0075] In the formula, P and are the signal powers before and after filtering, respectively. Calculate P and The result is a dimensionless signal with a value near 1. After the overall shift in the negative direction by 1, a standardized signal near 0 can be obtained.

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

[0077] In the third step, the vibration signals of the training set are converted into non-threshold RP images and the vibration signal feature parameters are extracted using RQA. The specific steps are as follows:

[0078] S301: Convert the training set vibration signal into a threshold-free RP image

[0079] Assume a sequence of length n t={t1,t2,……,t n According to Takens embedding theory and time delay method, the sequence can be reconstructed through 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 with a length of n-(m-1)τ. After plotting R, the resulting graph 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 minis the minimum length of the diagonal; P(l) represents the frequency of diagonals with a length of l. The higher the value of DET, the more stable the system is, and it is neither a random system nor a

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095] Among them: min ,v min ,w min 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 with a length of l, the vertical line with a length of v, and the white vertical line with a length of w, respectively.

[0096]

[0097] Among them, the intermittent nature 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 build a CNN transformer fault diagnosis model, use CNN to automatically extract the nonlinear features of the RP image and obtain the preliminary fault diagnosis result 1 after learning. The specific steps are as follows;

[0101] S401: Construct a CNN transformer fault diagnosis model, use CNN to automatically extract the nonlinear features of the RP image and obtain preliminary fault diagnosis results after learning 1

[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 convolution layer is:

[0104]

[0105] In the formula, is the f-th feature map of the e-th convolutional layer; σ(·) is a nonlinear activation function, and * represents a convolution operation; is the fth weight matrix of the eth convolutional layer; is the f-th feature map of the e-1-th convolutional layer; is the fth bias vector of the e-1th convolutional layer;

[0106] The pooling process expression of the pooling layer is:

[0107]

[0108] in, is the feature output of the pooling layer, max(·) represents the downsampling function of 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, and its expression is:

[0110] σ r (x)=max{0,x}

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

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

[0113]

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

[0115] The fifth step is to build the ADACSO-SVM transformer fault diagnosis model, use the validation set to adjust the model hyperparameters, and use the feature parameters extracted by RQA as the input of the ADACSO-SVM model to obtain the preliminary fault diagnosis results 2. The specific steps are as follows:

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

[0117] S502: Construct an SVM model, use the radial basis function as the SVM kernel function, and the optimal classification surface decision function is:

[0118]

[0119] S503: Determine the fitness function, initialize the parameter memory pool, change domain, change number, maximum number of iterations, and the position of each cat individual, and use the maximum value of the combination of the penalty factor and kernel function of the SVM model as the optimization condition, specifically:

[0120]

[0121] S504: Calculate the corresponding fitness value of 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, the optimal individual is output, that is, the optimal solution found by the algorithm; otherwise, return to step 4;

[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 result 2 through the test set;

[0125] In the sixth step, the obtained fault diagnosis results 1 and 2 are sent to the FCNN model to adjust the FCNN model, and then the fault diagnosis results are obtained through the test set. The specific steps are as follows:

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

[0127] S602: Construct an FCNN model. The FCNN network includes an input layer, a hidden layer, and an output layer, as follows:

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

[0129] The activation function in the hidden layer uses the RELU activation function, and the activation function in the output layer uses the tanh activation function:

[0130]

[0131] Where: 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 of the vibration signal and the actual label value of the vibration signal:

[0133]

[0134] Where: θ i is the actual value of the i-th label; Predict the value for the i-th label.

[0135] S603: Use the newly generated data set as the input of FCNN, further optimize the preliminary prediction results, and obtain the final diagnosis results

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

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

[0138] S801: Normalize the collected data of dissolved gases in oil. The dissolved gases in 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 min Indicates the minimum value of the variable, x max 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 characteristic parameter matrix into four categories, and the state of the dissolved gas content in oil is divided according to the IEEE standard.

[0142] S803: The initial state of the hidden Markov model is set to a normal state, the Baum-Welch algorithm is used to train the model according to the observation sequence, and the Viterbi algorithm is used to identify the state of the observation sequence, so as to obtain the log-likelihood function value corresponding to each state.

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

[0144] In the ninth step, the oil-paper insulation aging state is combined with the Weibull ratio risk model to determine the reliability function, predict the remaining life of the oil-paper insulation, and input it into the meta-classifier to realize the remaining life prediction of the transformer. The specific steps are as follows:

[0145] S901: The Weibull proportional hazard model is combined with the hidden Markov model to establish the reliability function, and the aging state is used as a covariate to affect the remaining life of the transformer oil-paper insulation. The specific expression is as follows:

[0146]

[0147] Where: 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 is only related to time t and obeys the Weibull distribution; β is the shape parameter of the Weibull distribution; η is the scale parameter; Z k is a covariate, which is only related to the aging state of the transformer oil-paper insulation; γ is a regression parameter, which reflects the influence of the aging state on the failure rate

[0148] The reliability function is:

[0149]

[0150] The equipment operating life at time t is:

[0151]

[0152] S902: Use the maximum likelihood estimation method to estimate the parameters of the Weibull proportional hazard model, and obtain the shape parameter β, scale parameter η and regression parameter γ

[0153] S903: Calculate the remaining life of the oil paper and input it into the meta-classifier, and output it when there is no sudden failure.

[0154] Example:

[0155] First, 355 transformer fault samples were collected, including 80 normal cases, 70 high-temperature overheating cases, 66 high-energy discharge cases, 51 medium-low temperature overheating cases, 38 low-energy discharge cases, 29 partial discharge cases and 21 discharge and overheating cases. Then, CEEMDAN-CCWK was used to eliminate the noise signal in the samples. Then, WGAN-GP was used to expand the number of samples while retaining the data characteristics of minority class samples. 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 sets, test sets, and validation sets.

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

[0158]

[0159] Finally, the vibration signal is converted into a non-threshold RP image and features are extracted using RQA.

[0160] In this embodiment, CNN and SVM models are used as base classifiers, FCNN model is used as meta-classifier, and ADACSO algorithm is used to optimize the number and depth of decision trees in SVM model. The number of decision trees is optimized in the range of [0,200], the depth of decision trees is optimized in the range of [0,70], the population is 10, and the maximum number of iterations is 100. The unthresholded RP image is input into the CNN classifier, and features are automatically extracted and diagnosed to obtain diagnosis result 1; the features extracted by RQA are input into the SVM classifier, and the ADACSO-SVM transformer fault diagnosis model is diagnosed to obtain diagnosis result 2, and the obtained fault diagnosis result 1 and fault diagnosis result 2 are input into the FCNN model to obtain the transformer diagnosis result.

[0161] In order to verify the superiority of this integrated fault diagnosis method, a comparative analysis was carried out with five models, namely ADACSO-SVM, SO-SVM, CNN, SVM, ELM, etc. The input data of all models were the same. The recall rate, precision rate and F1 were used to evaluate the superiority of this method, as shown in Table 2.

[0162] Table 2 Comparison of diagnostic accuracy of multiple models

[0163]

[0164] It can be seen from Table 2 that this method has a higher accuracy in diagnosing faults under unbalanced small sample conditions.

[0165] In order to verify the reliability of this life prediction method, life prediction is performed based on the collected DGA data corresponding to the above transformer vibration signal samples. Some dissolved gas data in oil are shown in Table 3. Then the BIC criterion is used to determine the optimal hidden Markov model, and the hidden state is 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 is completed, the hidden Markov model is trained.

[0166] Table 3 Data of dissolved gases in some oils

[0167]

[0168] After determining the initial conditions of the model, the Baum-Welch algorithm is used to train the model according to the observation sequence. Each hidden state must train the corresponding hidden Markov model. After training, the Viterbi algorithm is used to identify the state of the observation sequence, that is, to calculate the probability of the observation sequence P(O|λi), i=1,2,3,4 to obtain the log-likelihood function value corresponding to each state. The state corresponding to the maximum log-likelihood function value indicates the state of the current observation sequence.

[0169] Finally, the characteristic parameters are processed by entropy weight method, and the weights of H2, CO, CO2 and total hydrocarbons P = [0.2700.292 0.262 0.176] are obtained. The parameter estimation of Weibull proportional risk model is carried out by maximum likelihood estimation method, and the shape parameter β, scale parameter η and regression parameter γ are obtained. The life of oil-paper insulation under different aging conditions is obtained. Figure 2 shown.

[0170] Depend on Figure 2 It can be seen that when the reliability of oil-paper insulation is 0.9, the remaining life in normal state is 10.45 years; when the reliability is 0.8, the remaining life in normal state is 15.58 years; when the reliability is 0.7, the remaining life in normal state is 20.02 years, which is basically consistent with the actual life of oil-paper insulation of about 20 years. It can be proved that this method has a high accuracy in life 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, 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 online life prediction method under unbalanced multi-source small sample conditions, characterized in that: The following steps are involved: In the first step, the collected transformer vibration signal fault sample data is used to eliminate the noise signal in the sample using CEEMDAN-CCWK, and then WGAN-GP is used to expand the number of samples while retaining the characteristics of the minority class sample data; The second step is to standardize the collected and generated data and divide them into training sets, test sets, and validation sets; The third step is to convert the vibration signals of the training set into non-threshold RP images and then use RQA to extract the vibration signal feature parameters; The fourth step is to build a CNN transformer fault diagnosis model, use CNN to automatically extract the nonlinear features of the RP image and obtain the preliminary fault diagnosis results after learning 1; Step 5: Construct the ADACSO-SVM transformer fault diagnosis model, use the validation set to adjust the model hyperparameters, and use the feature parameters extracted by RQA as the input of the ADACSO-SVM model to obtain the preliminary fault diagnosis results 2; Step 6: Send the obtained fault diagnosis result 1 and fault diagnosis result 2 into the FCNN model to adjust the FCNN model, and then obtain the fault diagnosis result through the test set; Step 7: Input the collected vibration signals in the online test set into the meta-classifier to realize online fault diagnosis of transformer; In the eighth step, after normalizing the collected data of dissolved gas in oil in the online test set, a hidden Markov model is established to determine the aging state of oil-paper insulation; In the ninth step, the aging state of the oil-paper insulation is combined with the Weibull ratio risk model to determine the reliability function, predict the remaining life of the oil-paper insulation, and input it into the meta-classifier to realize the remaining life prediction of the transformer.

2. The transformer online life prediction method under unbalanced multi-source small sample conditions according to claim 1 is 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 false components and noise, remove them and reconstruct sensitive IMFs. Step 2: Use the WGAN-GP algorithm to expand the number of samples.

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

4. The transformer online life prediction method under unbalanced multi-source small sample conditions according to claim 1 is characterized in that: The specific steps of the third step are as follows: converting the vibration signal of the training set into a threshold-free RP image and then using RQA to extract the characteristic parameters of the vibration signal.

5. The method for online transformer life prediction under unbalanced multi-source small sample conditions according to claim 1 is characterized in that: The specific steps of the fourth step are as follows: Step 1: Define the network structure, alternately stack convolutional layers and pooling layers, and connect to the fully connected layer and output layer; Step 2: Initialize the grid parameters and randomly assign values ​​to the convolution kernel weights, biases, and filter index values ​​of the convolution layer; Step 3: Use the unstructured data threshold-free 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 layer, pooling layer, and fully connected layer in the CNN to obtain the output result, calculate the error between the output result and the true label, and use the back propagation algorithm to calculate the gradient of each weight, thereby updating the weight and bias; Step 5: Obtain fault diagnosis result 1 through the test set.

6. The method for online transformer life prediction under unbalanced multi-source small sample conditions according to claim 1 is characterized in that: 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 the SVM model, use the radial basis function as the SVM kernel function, and determine the optimal classification surface decision function; Step 3: Determine the fitness function, initialize the parameter memory pool, change domain, change number, maximum number of iterations, and the position of each cat individual, 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 iteration, update individual positions, 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 position, otherwise keep the original position unchanged; Step 5: After reaching the maximum number of iterations, stop the iteration and output the optimal solution. Otherwise, return to step 3 and iterate the search. If the termination condition is reached, output the optimal individual, that is, the optimal solution found by the algorithm; otherwise, return to step 4; Step 6: Build the ADACSO-SVM transformer fault diagnosis model, use the validation set to adjust the model structure and hyperparameters, and then save the diagnosis model; Step 7: Obtain fault diagnosis result 2 through the test set.

7. The method for online transformer life prediction under unbalanced multi-source small sample conditions according to claim 1 is characterized in that: The specific steps of the sixth step are as follows: Step 1: Define the network structure, alternately stack convolutional layers and pooling layers, and connect to the fully connected layer and output layer; Step 2: Initialize the grid parameters and randomly assign values ​​to the convolution kernel weights, biases, and filter index values ​​of the convolution layer; Step 3: Use fault diagnosis result 1 and fault diagnosis result 2 as input vectors of the FCNN network and perform feature extraction on the input vectors; Step 4: Train the FCNN network, including passing the input data through the convolutional layer, pooling layer, and fully connected layer in the FCNN to obtain the output result, calculate the error between the output result and the true label, and use the back propagation algorithm to calculate the gradient of each weight, thereby updating the weight and bias.

8. The method for online transformer life prediction under unbalanced multi-source small sample conditions according to claim 1 is characterized in that: The specific steps of the seventh step are as follows: inputting the collected online test set into the optimal classifier to realize online fault diagnosis of the transformer.

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

10. The transformer online life prediction method under unbalanced multi-source small sample conditions according to claim 1 is characterized in that: The specific steps of the ninth step are as follows: Step 1: Determine the characteristic parameter weights of dissolved gas in oil by entropy weight method; Step 2: Use the maximum likelihood estimation method to estimate the parameters of the Weibull proportional hazard model, obtain the shape parameter β, scale parameter η and regression parameter γ, and determine the reliability function; Step 3: Predict the remaining life of the oil-paper insulation based on the aging state of the oil-paper insulation and the reliability function.

Images