Converter transformer fault intelligent diagnosis method in strong noise environment
Through the method of combining cluster weighted envelope spectrum and VAE-CNN network, the accuracy problem of commutation rheology fault diagnosis in a strong noise environment is solved, and high-precision identification and diagnosis of commutation rheology faults are realized.
Patent Information
- Application Number
- CN202510412644.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-02
- Publication Date
- 2025-07-25
AI Technical Summary
The existing converter rheology fault diagnosis technology is difficult to effectively extract the fault information of vibration signals in a strong noise environment, resulting in misjudgment or misjudgment, especially in the case of small changes in complex fault types and frequency.
By combining cluster weighted envelope spectrum and VAE-CNN network, the cluster weighted envelope spectrum is constructed by identifying the potential fault mode frequency of the vibration signal, and combining the variational autoencoder and the convolutional neural network to achieve the extraction and diagnosis of fault features.
In a highly noise environment, the accuracy of commutation change fault diagnosis is improved, and the fault characteristics are clearly concentrated, which enhances the accuracy of high-dimensional time feature extraction and fault identification of vibration signals.
Smart Images
Figure CN120372382A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of converter transformer fault identification and diagnosis, and specifically to an intelligent diagnosis method for converter transformer faults in a strong noise environment. Background Art
[0002] As a core device in the power system, the converter transformer (abbreviated as converter transformer) plays a crucial role in ensuring the reliability and continuity of power supply. With the continuous expansion of the scale of the power system and the increasing complexity of the operating environment, the converter transformer faces various potential fault risks. Once a fault occurs, it may lead to large-scale power outages, causing huge economic losses and inconveniences to social production and people's lives.
[0003] Currently, the converter transformer fault diagnosis technologies mainly include oil sample analysis method, infrared imaging method, and vibration analysis method. Among them, the oil sample analysis method can accurately judge various internal fault types of the converter transformer by monitoring the components and contents of various gases dissolved in the converter transformer oil. For some early and latent faults, the oil sample analysis method has high sensitivity. However, the process of oil sample analysis is cumbersome, including multiple links such as sampling, transportation, and laboratory analysis. It often takes a long time, usually several days or even weeks, from sampling to obtaining the final analysis result. At the same time, it has the disadvantages of complex operation, easy deviation, and certain destructiveness. The infrared imaging method can detect the converter transformer at a distance without affecting its normal operation, avoiding the interference and safety risks that may be brought by physical contact with the converter transformer. However, it is greatly affected by the environment and the measurement accuracy will be reduced due to the differences in surface materials of different devices.
[0004] Vibration analysis technology has the advantages of non-invasiveness and being able to reflect the internal mechanical state of the converter transformer. By accurately collecting, processing, and deeply analyzing the vibration signals of the converter transformer, and excavating the hidden fault feature information, it can effectively diagnose and identify various fault types such as winding deformation, core loosening, and insulation aging of the converter transformer, improving the safety and economy of the power system operation, and having broad application prospects. However, the existing vibration signal analysis methods have the following deficiencies: 1) It is difficult to distinguish complex fault types when they occur in the converter transformer. For example, when the converter transformer winding is loose and accompanied by a multi-point grounding fault of the core, various fault features are intertwined and superimposed, resulting in an extremely complex signal composition; 2) There are a large number of electromagnetic interference, mechanical vibration interference and other noises around the converter transformer, which are likely to cover up the true fault feature signals, leading to false or missed fault judgments; 3) When the vibration frequency changes due to the winding deformation of the converter transformer, the frequency change range caused by a small deformation degree is relatively narrow, so it is impossible to distinguish the tiny frequency changes and it is difficult to detect early winding deformation faults; 4) For non-stationary signals, the existing analysis methods cannot well reflect the time-varying frequency characteristics, thus losing important fault information.
[0005] Therefore, there is an urgent need for an intelligent diagnosis method for converter transformer faults in a strong noise environment, which can effectively extract the fault information of the vibration signal of the converter transformer in strong noise and improve the diagnosis accuracy. Summary of the Invention
[0006] To solve the above problems, the object of the present invention is to provide an intelligent diagnosis method for converter transformer faults in a strong noise environment.
[0007] To achieve the above object, the present invention is realized through the following technical solutions:
[0008] An intelligent diagnosis method for converter transformer faults in a strong noise environment, first collects vibration data under different faults of the converter transformer, then extracts the fault characteristics of the converter transformer by using the clustering weighted envelope spectrum, and establishes a VAE-CNN network diagnosis model, so as to realize the fault identification and diagnosis of the converter transformer.
[0009] An intelligent diagnosis method for converter transformer faults in a strong noise environment includes the following steps:
[0010] ① After collecting the vibration data of typical converter transformer faults, preprocess it and divide the training set and the test set;
[0011] ② Identify the mode and calculate the weight: Without any prior knowledge, identify the potential fault mode frequencies of each frequency slice in the vibration signal spectrum, and generate the labels and weights of each frequency slice by using the potential fault mode frequencies;
[0012] ③ Construct the clustering weighted envelope spectrum: First, cluster the frequency slices with similar labels, and then sort them based on the weights; Finally, construct the clustering weighted envelope spectrum by using the weighted average value of the SFS in the selected cluster to realize feature extraction;
[0013] ④ Build and train the VAE-CNN diagnosis model: Combine the variational autoencoder and the CNN, and use the Monte Carlo sampling method to simplify the calculation process to complete the construction and training of the converter transformer fault diagnosis model.
[0014] Preferably, it further includes inputting the extracted operation characteristics of the converter transformer into the established diagnosis model for verification.
[0015] Preferably, in step ①, the vibration data of typical converter transformer faults includes vibration data in four states: normal state, bushing looseness, winding looseness, and combined looseness of winding and bushing.
[0016] Preferably, in step ①, the sensor acquisition frequency is 10 kHz.
[0017] Preferably, 75-125 signals are collected for each state, and each signal is expressed as x = (x1, x2, L, xN ),x i represents the i-th sampling point of the vibration signal x, where 1 ≤ i ≤ N and N is the total number of sampling points of the signal. The collected signal is randomly divided into a training set and a test set according to a certain proportion for training and evaluation of the subsequent fault diagnosis model.
[0018] Further preferably, the division ratio of the training set to the test set is 8 - 7:2 - 3.
[0019] Preferably, the specific steps of step ② include:
[0020] Calculate the two-dimensional Fourier transform of the instantaneous autocorrelation function of the vibration signal x, which is defined as follows:
[0021] S x (α, f) = ∫∫R x (t, τ)e -j2π(αt+fτ) dtdτ
[0022] R x (t, τ) = E{x(t)x(t - τ) *}
[0023] where R x (t, τ) is the instantaneous autocorrelation function of the vibration signal x, τ is the time delay, E{·} is the expected conjugate operator, and (·) * is the complex conjugate operator. α represents the multiple frequency of the fault characteristic frequency, f is the spectral frequency corresponding to the carrier frequency; t is time; i is the imaginary unit;
[0024] Normalize the above obtained result to get:
[0025]
[0026] where α m = f s / L, f n = f s / N ω , f s is the signal sampling frequency, L is the length of the signal, and N ω is the length of the window; α m represents the m-th cyclic frequency; (α m and f n together are used as variables for the two-dimensional Fourier transform to calculate the spectral correlation function of the signal)
[0027] For some spectral frequency slices SFSs, multiple groups of candidate fault modal frequencies FMF are obtained. Therefore, the index average noise frequency ratio RMFN is calculated by the following formula to measure the signal-to-noise ratio of each group of candidate FMF:
[0028]
[0029] Among them, is the floor function; p k is the k-th candidate fault mode frequency, and fn is the n-th spectral frequency;
[0030] Select the group with the largest RMFN as the potential fault mode frequencies PFMFs, and the other groups are regarded as noise and do not require further analysis.
[0031] Preferably, the specific steps of step ③ include:
[0032] The SFSs of the same cyclostationary component have similar labels. Therefore, for different cyclostationary components, a clustering algorithm is used for clustering to aggregate similar features in order to remove irrelevant interferences;
[0033] Let f c represent the fault mode frequency of the cyclostationary component. According to the fault frequency mechanism, the final clustering assignment will be similar to [f c , 2f c , 3f c , 4f c or [2f c , 4f c , 6f c , 8f c ;
[0034] The clustering clusters calculated using the RMFN index are then sorted in descending order, and retaining the top six clusters can ensure the complete extraction of fault information while removing interference information;
[0035] Calculate the weighted average of the SFSs in the clustering clusters obtained by the above screening to construct a clustering weighted envelope spectrum, as follows:
[0036]
[0037] Among them, CWES h is the CWES constructed for the h-th clustering cluster, I h represents the number of SESs in the h-th clustering cluster, ω(·) is the weight function, and this spectrogram is used as the input of the subsequent VAE-CNN network model to complete the training and classification of the converter transformer fault diagnosis model.
[0038] Preferably, step ④ includes the following steps:
[0039] Model the input data a:
[0040]
[0041] Among them, P represents the set of parameters, and g(a, h, P) represents the combined probability distribution of the sample and the hidden variable h;
[0042] The objective function for optimization in variational inference is:
[0043] L(r) = E r {ln[g(a|h, P)] - KL[r(h|a, P)||g(h, P)]};
[0044] Among them, the KL divergence is used to measure the distance between two distributions, and KL[r(h|a, P)||g(h, P)] represents transforming the input variable a into the hidden variable h;
[0045] The Markov chain Monte Carlo sampling method is used to estimate the likelihood function, and the network optimization function is modified to:
[0046] L(r) = E r {ln[g(a|h, P) = μ(a) + σ 1 / 2 (a)×e] - KL[r(h|a, P)||g(h, P)]}.
[0047] The present invention has the following advantages compared with the prior art:
[0048] The intelligent diagnosis method for converter transformer faults under strong noise environment of the present invention can effectively extract the fault information of the vibration signal of the converter transformer in the noise and improve the diagnosis accuracy. First, the clustering weighted envelope spectrum can extract potential fault feature components, cluster and sort the frequency slices of similar fault modes, classify the signals with similar features into one category, remove the messy and irrelevant frequency components, make the fault features in the envelope spectrum clearer and more concentrated, and improve the fault diagnosis accuracy. Second, by using CNN to construct a time series dependent on the vibration signal, the high-dimensional time features in the vibration signal are deeply extracted; while VAE is a new generative model, which can compress the input data into a low-dimensional spatial representation and generate new data samples therefrom, while maintaining the features of the original data, and improve the accuracy of converter transformer fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 It is a schematic diagram of the label recognition and weight calculation process;
[0050] Figure 2 It is a schematic diagram of the clustering and sorting process;
[0051] Figure 3 It is a schematic diagram of the obtained clustering weighted envelope spectrum;
[0052] Figure 4 It is a structural diagram of the VAE-CNN network;
[0053] Figure 5 is the vibration signal of the converter transformer;
[0054] Figure 6 is the constructed clustering weighted envelope spectrum;
[0055] Figure 7 is the model comparison result. Specific implementation manners
[0056] The object of the present invention is to provide an intelligent diagnosis method for converter transformer faults in a strong noise environment. The following further describes the present invention in conjunction with specific embodiments.
[0057] Embodiment 1
[0058] An intelligent diagnosis method for converter transformer faults in a strong noise environment. First, collect vibration data of the converter transformer under different faults, and then use the clustering weighted envelope spectrum to extract the fault characteristics of the converter transformer, and establish a VAE-CNN network diagnosis model, so as to realize the fault identification and diagnosis of the converter transformer.
[0059] Specifically, it includes the following steps:
[0060] ① After collecting the vibration data of typical converter transformer faults, perform preprocessing and divide the training set and the test set;
[0061] The vibration data of typical converter transformer faults includes vibration data in four states: normal state, bushing looseness, winding looseness, and combined looseness of winding and bushing. The sensor acquisition frequency is 10 kHz. 75 - 125 signals are collected for each state, and each signal is expressed as x = (x1, x2, L, x N ), x i represents the i-th sampling point of the vibration signal x, 1 ≤ i ≤ N, N is the total number of sampling points of the signal. The collected signals are randomly divided into the training set and the test set according to a ratio, and are used to train and evaluate the subsequent fault diagnosis model. For example, N can be set to 300, and 500 collected signals are randomly divided into the training set and the test set according to a ratio of 8:2, and are used to train and evaluate the subsequent fault diagnosis model.
[0062] ② Identify the mode and calculate the weight: Without any prior knowledge, identify the potential fault mode frequencies of each frequency slice in the vibration signal spectrum, and use the potential fault mode frequencies to generate the labels and weights of each frequency slice. The specific process is as Figure 1 shown. Without any prior knowledge of faults, identify the potential fault mode frequencies (PFMFs) of each frequency slice (SFS), and use the PFMF to generate the labels and weights of each SFS, specifically as follows:
[0063] Calculate the two-dimensional Fourier transform of the instantaneous autocorrelation function of the vibration signal x, defined as follows:
[0064] S x (α, f) = ∫∫R x (t, τ)e -j2π(αt+fτ) dtdτ
[0065] R x (t, τ) = E{x(t)x(t - τ) *}
[0066] where R x (t, τ) is the instantaneous autocorrelation function of the vibration signal x, τ is the time delay, E{·} is the expectation conjugate operator, (·) * is the complex conjugate operator, α represents the multiple frequency of the fault characteristic frequency, f is the spectral frequency corresponding to the carrier frequency; t is time; i is the imaginary unit;
[0067] Normalize the above obtained result to get:
[0068]
[0069] where α m = f s / L, f n = f s / N ω , f s is the signal sampling frequency, L is the length of the signal, N ω is the length of the window; α m represents the m-th cyclic frequency; (α m and f n together are used as the variables of the two-dimensional Fourier transform for calculating the spectral correlation function of the signal).
[0070] For some spectral frequency slices SFSs, multiple groups of candidate fault modal frequencies FMF are obtained. Therefore, calculate the index average noise frequency ratio RMFN through the following formula to measure the signal-to-noise ratio of each group of candidate FMF:
[0071]
[0072] where is the floor function; p k is the k-th candidate fault modal frequency, fn is the n-th spectral frequency;
[0073] Select the group with the largest RMFN as the potential fault modal frequencies PFMFs, and the other groups are regarded as noise and do not need further analysis.
[0074] ③Construct the clustering weighted envelope spectrum: First, cluster the frequency slices with similar labels, and then sort them based on weights; finally, construct the clustering weighted envelope spectrum using the weighted average of the SFSs in the selected clusters to achieve feature extraction; as Figure 2 shown
[0075] Although the PFMFs obtained by the above screening are very likely to contain fault information, there will inevitably be SFSs related to noise. The SFSs of the same cyclic stationary component have similar labels. Therefore, for different cyclic stationary components, a clustering algorithm is used for clustering to aggregate similar features in order to remove irrelevant interferences;
[0076] Let f c represent the fault mode frequency of the cyclic stationary component. According to the fault frequency mechanism, the final clustering assignment will be similar to [f c , 2f c , 3f c , 4f c or [2f c , 4f c , 6f c , 8f c form;
[0077] Use the RMFN index to calculate the clustering clusters, and then sort them in descending order. Retaining the first six clusters can ensure the complete extraction of fault information while removing interference information;
[0078] Calculate the weighted average of the SFSs in the clustering clusters obtained by the above screening for constructing the clustering weighted envelope spectrum, as Figure 3 shown, specifically as follows:
[0079]
[0080] Among them, CWES h is the CWES constructed by the h-th clustering cluster, I h represents the number of SESs in the h-th clustering cluster, ω(·) is the weight function, and this spectrogram is used as the input of the subsequent VAE-CNN network model to complete the training and classification of the converter transformer fault diagnosis model.
[0081] ④VAE-CNN diagnostic model construction and training: Combine the variational autoencoder and CNN, and use the Monte Carlo sampling method to simplify the calculation process to complete the construction and training of the converter transformer fault diagnosis model.
[0082] VAE is an unsupervised learning method that can decode high-dimensional data and map it to a low-dimensional latent space. Compared with traditional autoencoders, it provides enhanced interpretability and superior ability to capture the complexity of data distributions. Combining with the CNN model can better enhance the expressiveness of signal data and solve problems such as gradient vanishing and explosion.
[0083] Model the input data a:
[0084]
[0085] Where P represents the set of parameters, and g(a, h, P) represents the combined probability distribution of the sample and the hidden variable h;
[0086] The objective function used for optimization in variational inference is:
[0087] L(r) = E r {ln[g(a|h, P)] - KL[r(h|a, P)||g(h, P)]};
[0088] Where the KL divergence is used to measure the distance between two distributions, and KL[r(h|a, P)||g(h, P)] represents transforming the input variable a into the hidden variable h;
[0089] The integration of the posterior probability is difficult to handle for high-dimensional or complex distributions, and sampling methods need to be used to estimate the likelihood function. Using the Markov chain Monte Carlo sampling method to estimate the likelihood function, the network optimization function is modified to:
[0090] L(r) = E r {ln[g(a|h, P) = μ(a) + σ 1 / 2 (a)×e] - KL[r(h|a, P)||g(h, P)]}.
[0091] As Figure 4 shown, it can be seen that the network structure of the variational autoencoder mainly includes three parts: an encoder, a hidden layer, and a decoder. Among them, the encoder learns to capture the important features of the input data in this distribution, transforms the data from the original space to the latent space, and provides a basis for generating new data later; the hidden layer processes between the encoder and the decoder and is used to store the latent representation of the data; the decoder maps the latent representation of the hidden layer back to the original space and restores the data.
[0092] Example 2
[0093] Adopt the intelligent fault diagnosis method of converter transformer under strong noise environment in Example 1. After collecting the typical fault vibration data of the converter transformer in step ① and preprocessing it, divide the training set and the test set; and in step ②, identify the mode and calculate the weight, and process the vibration signal of the converter transformer, asFigure 5 As shown, it can be seen that the vibration signal of the converter transformer after being processed by the steps of the present invention can remove other interference information such as noise and retain the effective features. After performing the processing of step ③ to construct the clustering weighted envelope spectrum, the result is as Figure 6 , and it can be seen that the fault characteristic frequency of the transformer can be effectively amplified.
[0094] This technology also includes inputting the extracted operation characteristics of the converter transformer into the established diagnostic model for verification. A VAE-CNN diagnostic model is constructed, and the same features are input into the model and compared with the diagnostic models of support vector machine (SVM), extreme learning machine (ELM), and convolutional neural network (CNN). The result is as Figure 7 shown, and it can be seen that the present invention has the highest accuracy for transformer fault identification compared with other comparison methods.
Claims
1. An intelligent diagnosis method for converter transformer faults in a strong noise environment, characterized in that: First, collect the vibration data of the converter transformer under different faults, then extract the fault characteristics of the converter transformer using the clustering weighted envelope spectrum, and establish a VAE-CNN network diagnosis model to achieve the fault identification and diagnosis of the converter transformer.
2. The intelligent diagnosis method for converter transformer faults in a strong noise environment according to claim 1, characterized in that: It includes the following steps: ① After collecting the vibration data of typical faults of the converter transformer, perform preprocessing and divide the training set and the test set; ② Identify the mode and calculate the weight: Without any prior knowledge, identify the potential fault mode frequencies of each frequency slice in the vibration signal spectrum, and use the potential fault mode frequencies to generate the labels and weights of each frequency slice; ③ Construct the clustering weighted envelope spectrum: First, cluster the frequency slices with similar labels, then sort them based on the weights; Finally, construct the clustering weighted envelope spectrum using the weighted average of the SFSs in the selected clusters to achieve feature extraction; ④ Build and train the VAE-CNN diagnosis model: Combine the variational autoencoder and the CNN, and use the Monte Carlo sampling method to simplify the calculation process to complete the construction and training of the converter transformer fault diagnosis model.
3. The intelligent diagnosis method for converter transformer faults in a strong noise environment according to claim 2, wherein: It also includes inputting the extracted operating characteristics of the converter transformer into the established diagnosis model for verification.
4. The intelligent diagnosis method for converter transformer faults in a strong noise environment according to claim 2, characterized in that: In step ①, the vibration data of typical faults of the converter transformer includes the vibration data in four states: normal state, bushing looseness, winding looseness, and combined looseness of winding and bushing.
5. The intelligent diagnosis method for converter transformer faults in a strong noise environment according to claim 2, characterized in that: In step ①, the sensor acquisition frequency is 10 kHz.
6. The intelligent diagnosis method for converter transformer faults in a strong noise environment according to claim 4, characterized in that: Collect 75 to 125 signals for each state, and each signal is represented as x = (x1, x2, …, x N ), where x i represents the i-th sampling point of the vibration signal x, 1 ≤ i ≤ N, and N is the total number of sampling points of the signal. The collected signals are randomly divided into a training set and a test set according to a ratio to train and evaluate the subsequent fault diagnosis model.
7. An intelligent diagnosis method for converter transformer faults in a strong noise environment according to claim 6, characterized in that: The division ratio of the training set and the test set is 8 - 7:2 - 3.
8. The intelligent diagnosis method for converter transformer faults in a strong noise environment according to claim 2, characterized in that: The specific steps of step ② include: Calculate the two-dimensional Fourier transform of the instantaneous autocorrelation function of the vibration signal x, defined as follows: S x (α,f) = ∫∫R x (t,τ)e -j2π(αt+fτ) dtdτ; R x R(t,τ) = E{x(t)x(t - τ) *}; where R x (t,τ) is the instantaneous autocorrelation function of the vibration signal x, τ is the time delay, E{·} is the expected conjugate operator, (·) * is the complex conjugate operator, α represents the multiple frequency of the fault characteristic frequency, f is the spectral frequency corresponding to the carrier frequency; t is the time; j is the imaginary unit; Normalize the above obtained result to get: where α m = f s / L, f n = f s / N ω , fs is the signal sampling frequency, L is the length of the signal, N ω is the length of the window; α m represents the m-th cyclic frequency; For some spectral frequency slices SFSs, obtain multiple groups of candidate fault mode frequencies FMF. Therefore, calculate the index average noise frequency ratio RMFN through the following formula to measure the signal-to-noise ratio of each group of candidate FMF: Among them, is the floor function; p k is the k-th candidate fault mode frequency, and fn is the n-th spectrum frequency; Select the group with the largest RMFN as the potential fault mode frequencies PFMFs, and the other groups are regarded as noise and do not require further analysis.
9. The intelligent diagnosis method for converter transformer faults in a strong noise environment according to claim 2, characterized in that: The specific steps of step ③ include: SFSs with the same cyclic stationary component have similar labels. Therefore, for different cyclic stationary components, use the clustering algorithm for clustering to gather similar features in order to remove irrelevant interference; Let f c represent the fault modal frequency of the cyclostationary component. According to the fault frequency mechanism, the final clustering assignment will be similar to [f c , 2f c , 3f c , 4f c or [2f c , 4f c , 6f c , 8f c ; Use the clustering clusters calculated by the RMFN index, and then sort them in descending order. Retaining the first six clusters can ensure the complete extraction of fault information while removing interference information; Calculate the weighted average of the SFSs in the clustering clusters selected above to construct the clustering weighted envelope spectrum, specifically as follows: Among them, CWES h is the CWES constructed for the h-th clustering cluster, and I h represents the number of SESs in the h-th clustering cluster. ω(·) is a weight function. Using this spectrogram as the input of the subsequent VAE-CNN network model, the training and classification of the converter transformer fault diagnosis model are completed.
10. The intelligent diagnosis method for converter transformer faults in a strong noise environment according to claim 2, wherein: Step ④ includes the following steps: Model the input data a: Among them, P represents the parameter set, and g(a, h, P) represents the combined probability distribution of the sample and the hidden variable h; The objective function for optimization in variational inference is: L(r) = E r {ln[g(a|h,P)] - KL[r(h|a,P) || g(h,P)]}; Among them, the KL divergence is used to measure the distance between two distributions, and KL[r(h|a, P)||g(h, P)] represents transforming the input variable a into the hidden variable h; Adopt the Markov chain Monte Carlo sampling method to estimate the likelihood function, and the network optimization function is modified to: L(r) = E r {ln[g(a|h,P) = μ(a) + σ 1 / 2 (a)×e] - KL[r(h|a,P)||g(h,P)]}。