Transformer fault diagnosis method based on empirical wavelet and Riemannian cutting spatial characteristics
By combining empirical wavelet and Riemann chevron space features, and using twin support vector machines and bat algorithm optimization, the problem of low accuracy in transformer fault diagnosis was solved, achieving higher diagnostic accuracy and robustness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-13
- Publication Date
- 2026-04-03
AI Technical Summary
Existing transformer fault diagnosis methods lack accuracy and reliability under complex operating conditions. Traditional methods rely on human experience, resulting in poor feature extraction and difficulty in fully reflecting the transformer's operating status.
Empirical wavelet transform (EWT) is used to adaptively decompose transformer fault signals. Combined with Riemann chevron space feature extraction, and optimized using twin support vector machine and bat algorithm, feature fusion and classifier training are performed to achieve high-dimensional feature representation.
This improves the accuracy and robustness of transformer fault diagnosis, providing a strong guarantee for the safe operation of the power system.
Smart Images
Figure CN121786598A_ABST
Abstract
Description
[0001] This invention relates to the field of transformer fault diagnosis technology research, specifically to a transformer fault diagnosis method based on empirical wavelet and Riemann tangent space features. Background Technology
[0002] In power systems, the normal operation of transformers is crucial for ensuring the stability and reliability of the power grid. However, during long-term operation, transformers inevitably experience various faults, such as winding short circuits, insulation aging, and abnormal gas content in the oil. If these faults are not diagnosed and addressed in a timely manner, they will pose a serious threat to the power system. Existing transformer fault diagnosis methods, such as those based on dissolved gas analysis in oil and vibration signal analysis, can identify transformer faults to a certain extent, but their diagnostic accuracy and reliability still need improvement under complex operating conditions.
[0003] Traditional fault diagnosis methods often rely on manual experience to extract features. This approach is ineffective when dealing with the complexity and nonlinearity of transformer fault signals, leading to inaccurate diagnostic results. Furthermore, due to the diversity and uncertainty of transformer fault signals, a single feature extraction method cannot comprehensively reflect the transformer's operating state. In recent years, the application of Empirical Wavelet Transform (EWT) and Riemannian geometry in signal processing has gained increasing attention, demonstrating significant advantages in processing nonlinear and non-stationary signals. Summary of the Invention
[0004] To address the problems existing in current technologies, this invention proposes a transformer fault diagnosis method based on empirical wavelet transform and Riemann chevron space features. This method utilizes empirical wavelet transform to adaptively decompose the transformer fault signal, obtaining intrinsic mode functions (IMFs) with time-frequency localization characteristics. Combined with Riemann chevron space feature extraction, the fault signal is characterized in high dimension. This approach effectively improves the accuracy and robustness of transformer fault diagnosis, providing strong support for the safe operation of power systems.
[0005] The technical solution of this invention is as follows: A transformer fault diagnosis method based on empirical wavelet and Riemann tangent space features, comprising the following steps:
[0006] The first step is to install sensors at key parts of the transformer to collect transformer tank vibration signals and partial discharge detection signals, and then decentralize the transformer tank vibration signals and partial discharge detection signals.
[0007] The second step involves using Empirical Wavelet Decomposition (EWT) to perform mode decomposition on the two decentralized transformer signal datasets. Qualified mode components are then selected by calculating the fuzzy entropy value of each mode. The qualified mode components are used to reconstruct the two transformer signal matrices, thereby further filtering and denoising the transformer signal datasets.
[0008] The third step involves using the Riemann tangent space to extract features from the two types of transformer signal datasets after filtering and denoising.
[0009] The fourth step involves fusing the Riemann chevron space features of the two signal datasets and inputting them into a twin support vector machine. The parameters of the twin support vector machine optimized using the bat algorithm are then optimized to complete the training of the classifier.
[0010] Preferably, the specific steps of the first step are as follows:
[0011] Step 1: Install sensors at key locations on the transformer to collect sample data matrix of transformer tank vibration signals. and partial discharge detection signal
[0012] X and Z are signal matrices constructed from the original sampled signals. Each row in the matrix represents the acquisition signal of one sensor, and each column represents a sampling point. That is, m is the number of sensors and n is the number of sampling points.
[0013] Step 2: Specifically, collect transformer tank vibration signals and partial discharge detection signals under three fault states: loose transformer windings, loose iron core, and winding deformation, as well as under normal operating conditions.
[0014] Step 3: Decentralize the transformer tank vibration signal using a common-average reference to obtain a decentralized matrix. and
[0015] Preferably, the specific steps of the second step are as follows:
[0016] Step 1: Analyze the single-channel vibration signal y of the transformer tank. i (t) Perform a Fourier transform and normalize to the [0,π] interval to obtain the frequency domain signal Y. i (ω), where i = 1, 2, ..., m
[0017] Step 2: Use the scale-space method to measure Y i (ω)(ω∈[0,π]) is adaptively divided into Q sub-bands, each sub-band defined as Λ q =[ω q-1 ,ω q ], where q = 1, 2, ..., Q;
[0018] Step 3: Construct wavelet filters for each sub-band based on Meyer wavelet theory:
[0019]
[0020] In the formula, Φ q (ω) and Ψ q (ω) represents the scaling function and wavelet function of the empirical wavelet, respectively, and β(y) is a function defined between [0,1]. Generally, β(y) is taken as y = y. 4 (35-84y+70y 2 -20y 3 ),
[0021] For the single-channel vibration signal y of the transformer housing i (t), the detailed parameters of its wavelet function and detailed parameters of the scaling function They are defined as follows:
[0022]
[0023] In the formula, Y i (ω) and These are the discretized transformer enclosure vibration signal, wavelet function value, and scaling function value, respectively.
[0024] Finally, the y i (t) After empirical wavelet decomposition, the following empirical modes are obtained:
[0025]
[0026] Step 4: Modal components of the vibration signal from the transformer housing of each sensor channel. The fuzzy entropy algorithm is used for filtering. First, each modal component is defined. Given the phase space embedding dimension w (w < l-1) and similarity tolerance r, reconstruct the phase space:
[0027]
[0028] Then, calculate the window vector. and Maximum distance:
[0029]
[0030] calculate Fuzzy membership degree And seek all mean
[0031]
[0032] Calculate all window vectors Mean:
[0033]
[0034] Finally, repeat the above process to calculate. And based on this, we obtain The fuzzy entropy is:
[0035]
[0036] Step 5: Select the one with the largest fuzzy entropy value The vibration signal y of the transformer housing for each sensor channel was reconstructed using empirical wavelet inverse transform. i (t):
[0037]
[0038] in, and Ψ max (t) represent the values with the maximum fuzzy entropy. The corresponding wavelet function values and wavelet function detail parameters, the reconstructed transformer tank vibration signal are as follows:
[0039] Step 6: Repeat steps 1 to 5 to reconstruct the transformer partial discharge detection signal z for each sensor channel. i (t), the reconstructed transformer partial discharge detection signal is
[0040] Preferably, the specific steps of the third step are as follows:
[0041] Step 1: Use the covariance matrix of the transformer vibration signal as the statistical feature of the data. The covariance matrix of the transformer vibration signal for a single test is:
[0042]
[0043] Step 2: Concatenate the covariance matrices of the j experiments {P1...P j The Riemann mean of j experiments is calculated using the following formula. That is, the Riemann Center
[0044]
[0045] Step 3: Using the AIRM metric, calculate the m-th order Riemann tangent space eigenvector using the following formula, to characterize the deviation of a single covariance matrix from its Riemann center:
[0046]
[0047] Step 4: Similarly, calculate the m-th order Riemann tangent space eigenvector of the transformer partial discharge detection signal as follows:
[0048] Preferably, the specific steps of the fourth step are as follows:
[0049] Step 1: Construct a joint feature vector. Combine the Riemann tangent space features of the transformer vibration signal with the Riemann tangent space features of the transformer partial discharge detection signal to construct the joint feature vector shown below:
[0050]
[0051] The joint feature vector f represents data from the same trial.
[0052] Step 2: Input the feature vectors of all samples as the training set into the Siamese Support Vector Machine to train the classifier, and use the Bat Algorithm to optimize the parameters of the Siamese Support Vector Machine, including penalty factor c1, penalty factor c2 and kernel parameter λ, and establish a classification model with the optimized c1, c2 and λ.
[0053] Step 3: Set relevant parameters, including bat population size N, maximum number of iterations M, and foraging space dimension dim;
[0054] Step 4: Randomly generate the position x of bat i. i and velocity v i bat's position x i The three parameters representing TWSVM—penalty factor c1, penalty factor c2, and kernel parameter λ—are used to evaluate the individual fitness of bats and find the optimal solution x at the current time step. * The bat individual with the highest fitness is the global optimal solution. This invention uses the penalty factor c and kernel parameter λ, along with the recognition accuracy obtained from cross-validation, as the standard for evaluating fitness, defined as:
[0055]
[0056] Step 5: Substitute the location of the bat population into the evaluation fitness function, continuously update the location corresponding to the best bat individual until the iteration termination condition is met, record the optimal parameter value, and substitute the globally optimal parameters into the Siamese support vector machine to complete the training of the classifier model.
[0057] Step 6: Input the obtained fused feature vector into the trained classifier to identify multiple types of fault signals. Attached Figure Description
[0058] Figure 1 This is a flowchart of the method of the present invention;
[0059] Figure 2 Flowchart of the optimization algorithm for bats. Detailed Implementation
[0060] 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.
[0061] Example 1
[0062] like Figure 1 As shown, a transformer fault diagnosis method based on empirical wavelet and Riemann tangent space features comprises the following steps:
[0063] The first step is to install sensors at key parts of the transformer to collect vibration signals and partial discharge detection signals from the transformer tank, and then decentralize these signals.
[0064] S101: Install sensors at key locations on the transformer to collect sample data matrix of transformer tank vibration signals. and partial discharge detection signal
[0065] Specifically, six sensors are installed at key parts of the transformer to collect 25 sets of transformer tank vibration signal data for each of three fault states: loose winding, loose core, and winding deformation, as well as normal operating state, with a total data length of 2048 data points.
[0066] S102: The transformer tank vibration signal and partial discharge detection signal are decentered using a common-average reference. The specific formula is as follows:
[0067]
[0068] In the formula, Y i (t) represents the original transformer vibration signal from the i-th sensor, Z i (t) represents the partial discharge detection signal of the i-th sensor.
[0069] The second step involves using Empirical Wavelet Decomposition (EWT) to perform mode decomposition on the two decentralized transformer signal datasets. Qualified mode components are then selected by calculating the fuzzy entropy value of each mode. The qualified mode components are used to reconstruct the two transformer signal matrices, thereby further filtering and denoising the transformer signal datasets.
[0070] S201: Single-channel vibration signal y from transformer enclosure i (t) Perform a Fourier transform and normalize to the [0,π] interval to obtain the frequency domain signal Y. i (ω), where i = 1, 2, ..., m;
[0071] S202: Using the scale-space method to measure Y i (ω)(ω∈[0,π]) is adaptively divided into Q sub-bands, each sub-band defined as Λ q =[ω q-1 ,ω q ], where q = 1, 2, ..., Q:
[0072] S203: Construct wavelet filters for each sub-band based on Meyer wavelet theory:
[0073]
[0074] In the formula, Φ q (ω) and Ψ q (ω) represents the scaling function and wavelet function of the empirical wavelet, respectively, and β(y) is a function defined between [0,1]. Generally, β(y) is taken as y = y. 4 (35-84y+70y 2 -20y 3 ),
[0075] For the single-channel vibration signal y of the transformer housing i (t), the detailed parameters of its wavelet function and detailed parameters of the scaling function They are defined as follows:
[0076]
[0077] In the formula, Y i (ω) and These are the discretized transformer enclosure vibration signal, wavelet function value, and scaling function value, respectively.
[0078] Finally, the y i (t) After empirical wavelet decomposition, the following empirical modes are obtained:
[0079]
[0080] S204: Modal components of the vibration signal from the transformer housing of each sensor channel. The fuzzy entropy algorithm is used for filtering. First, each modal component is defined. Given the phase space embedding dimension w (w < l-1) and similarity tolerance r, reconstruct the phase space:
[0081]
[0082] Then, calculate the window vector. and Maximum distance:
[0083]
[0084] calculate Fuzzy membership degree And seek all mean
[0085]
[0086] Calculate all window vectors Mean:
[0087]
[0088] Finally, repeat the above process to calculate. And based on this, we obtain The fuzzy entropy is:
[0089]
[0090] S205: Select the value with the largest fuzzy entropy. The vibration signal y of the transformer housing for each sensor channel was reconstructed using empirical wavelet inverse transform. i (t):
[0091]
[0092] in, and Ψ max (t) represent the values with the maximum fuzzy entropy. The corresponding wavelet function values and wavelet function detail parameters, the reconstructed transformer tank vibration signal are as follows:
[0093] S206: Repeat steps S201-S205 to reconstruct the transformer partial discharge detection signal for each sensor channel. i(t), the reconstructed transformer partial discharge detection signal is
[0094] The third step involves using the Riemann tangent space to extract features from the two types of transformer signal datasets after filtering and denoising.
[0095] S301: The covariance matrix of the transformer vibration signal is used as a statistical feature of the data. The covariance matrix of the transformer vibration signal for a single test is:
[0096]
[0097] S302: Concatenate the covariance matrices of the 25 experiments {P1...P 25 The Riemann mean of 25 experiments was calculated using the following formula. That is, the Riemann Center
[0098]
[0099] S303: Using the AIRM metric, the 6th-order Riemann tangent space eigenvector is calculated using the following formula to characterize the deviation of a single covariance matrix from its Riemann center:
[0100]
[0101] S304: Similarly, the m-th order Riemann tangent space eigenvector of the transformer partial discharge detection signal is calculated as follows:
[0102] The fourth step involves fusing the Riemann chevron space features of the two signal datasets and inputting them into a twin support vector machine. The parameters of the twin support vector machine optimized using the bat algorithm are then optimized to complete the training of the classifier.
[0103] S401: Construct a joint feature vector by combining the Riemann tangent space features of the transformer vibration signal and the Riemann tangent space features of the transformer partial discharge detection signal, as shown below:
[0104]
[0105] Joint eigenvectors It is the same sample data.
[0106] S402: Input the feature vectors of all samples as the training set into the twin support vector machine to train the classifier, and use the bat algorithm to optimize the parameters of the twin support vector machine, such as penalty factor c1, penalty factor c2 and kernel parameter λ, and establish a classification model with the optimized c1, c2 and λ.
[0107] S403: Set relevant parameters, including bat population size N, maximum number of iterations M, and foraging space dimension dim;
[0108] S404: Randomly generate the position x of bat i i and velocity v i bat's position x i The three parameters representing TWSVM—penalty factor c1, penalty factor c2, and kernel parameter λ—are used to evaluate the individual fitness of bats and find the optimal solution x at the current time step. * The bat individual with the highest fitness is the global optimal solution. This invention uses the penalty factor c and kernel parameter λ, along with the recognition accuracy obtained from cross-validation, as the standard for evaluating fitness, defined as:
[0109]
[0110] S405: Substitute the location of the bat population into the evaluation fitness function, continuously update the location corresponding to the best bat individual until the iteration termination condition is met, record the optimal parameter value, and substitute the globally optimal parameters into the Siamese support vector machine to complete the training of the classifier model.
[0111] S406: Input the obtained fused feature vector into the trained classifier to identify multiple types of fault signals.
[0112] Example 2
[0113] The invention first installs six sensors on key parts of the sample transformer to collect 25 sets of transformer tank vibration signals and partial discharge detection signals for three fault states: loose winding, loose core, and winding deformation, as well as normal operating conditions. The total data length is 2048 data points.
[0114] This invention employs Empirical Wavelet Decomposition (EWT) to perform mode decomposition on two decentralized transformer signal datasets. Qualified mode components are selected by calculating the fuzzy entropy value of each mode. These qualified mode components are then used to reconstruct the two transformer signal matrices, thereby further filtering and denoising the transformer signal datasets. Next, feature extraction is performed on the filtered and denoised transformer signal datasets using the Riemann tangent space. Finally, the Riemann tangent space features of the two signal datasets are fused and input into a Siamese Support Vector Machine (SVM). The parameters of the Siamese SVM optimized using the Bat Algorithm are optimized to complete the training of the classifier.
[0115] This invention patent proposes a transformer fault diagnosis method based on empirical wavelet and Riemann chevron space features, under the background of severe transformer operating conditions. It combines Riemann chevron space features with bat optimization (the algorithm of which is as follows). Figure 2The present invention combines a twin support vector machine (as shown) for transformer fault diagnosis. Compared with the existing technology, the advantage of this patent is that it introduces an optimization algorithm based on Riemann tangent space features and combines it with a machine learning framework to complete the fault diagnosis. This achieves the effect of solving the problem of low accuracy in transformer fault diagnosis and the problem of misjudgment and omission of faults caused by the nonlinear characteristics of vibration signals in the transformer fault diagnosis model.
[0116] 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 transformer fault diagnosis method based on empirical wavelet and Riemann tangent space features, characterized in that, Includes the following steps: The first step is to install sensors at key parts of the transformer to collect transformer tank vibration signals and partial discharge detection signals, and then decentralize the transformer tank vibration signals and partial discharge detection signals. The second step involves using Empirical Wavelet Decomposition (EWT) to perform mode decomposition on the two decentralized transformer signal datasets. Qualified mode components are then selected by calculating the fuzzy entropy value of each mode. The qualified mode components are used to reconstruct the two transformer signal matrices, thereby further filtering and denoising the transformer signal datasets. The third step involves using the Riemann tangent space to extract features from the two types of transformer signal datasets after filtering and denoising. The fourth step involves fusing the Riemann chevron space features of the two signal datasets and inputting them into a twin support vector machine. The parameters of the twin support vector machine optimized using the bat algorithm are then optimized to complete the training of the classifier.
2. The transformer fault diagnosis method based on empirical wavelet and Riemann tangent space features according to claim 1, characterized in that, The specific steps of the first step are as follows: Step 1: Install sensors at key locations on the transformer to collect sample data matrix of transformer tank vibration signals. and partial discharge detection signal X and Z are signal matrices constructed from the original sampled signals. Each row in the matrix represents the acquisition signal of one sensor, and each column represents a sampling point. That is, m is the number of sensors and n is the number of sampling points. Step 2: Specifically, collect transformer tank vibration signals and partial discharge detection signals under three fault states: loose transformer windings, loose iron core, and winding deformation, as well as under normal operating conditions. Step 3: Decentralize the transformer tank vibration signal using a common-average reference to obtain a decentralized matrix. and 3. The transformer fault diagnosis method based on empirical wavelet and Riemann tangent space features according to claim 1, characterized in that, The specific steps of the second step are as follows: Step 1: Analyze the single-channel vibration signal y of the transformer tank. i (t) Perform a Fourier transform and normalize to the [0,π] interval to obtain the frequency domain signal Y. i (ω), where i = 1, 2, ..., m Step 2: Use the scale-space method to measure Y i (ω)(ω∈[0,π]) is adaptively divided into Q sub-bands, each sub-band defined as Λ q =[ω q-1 ,ω q ], where q = 1, 2, ..., Q; Step 3: Construct wavelet filters for each sub-band based on Meyer wavelet theory: In the formula, Φ q (ω) and Ψ q (ω) represents the scaling function and wavelet function of the empirical wavelet, respectively, and β(y) is a function defined between [0,1]. Generally, β(y) is taken as y = y. 4 (35-84y+70y 2 -20y 3 ), For the single-channel vibration signal y of the transformer housing i (t), the detailed parameters of its wavelet function and detailed parameters of the scaling function They are defined as follows: In the formula, Y i (ω) and These are the discretized transformer enclosure vibration signal, wavelet function value, and scaling function value, respectively. Finally, the y i (t) After empirical wavelet decomposition, the following empirical modes are obtained: Step 4: Modal components of the vibration signal from the transformer housing of each sensor channel. The fuzzy entropy algorithm is used for filtering. First, each modal component is defined. Given the phase space embedding dimension w (w < l-1) and similarity tolerance r, reconstruct the phase space: Then, calculate the window vector. and Maximum distance: calculate Fuzzy membership degree And seek all mean Calculate all window vectors Mean: Finally, repeat the above process to calculate. And based on this, we obtain The fuzzy entropy is: Step 5: Select the one with the largest fuzzy entropy value The vibration signal y of the transformer housing for each sensor channel was reconstructed using empirical wavelet inverse transform. i (t): in, and Ψ max (t) represent the values with the maximum fuzzy entropy. The corresponding wavelet function values and wavelet function detail parameters, the reconstructed transformer tank vibration signal are as follows: Step 6: Repeat steps 1 to 5 to reconstruct the transformer partial discharge detection signal z for each sensor channel. i (t), the reconstructed transformer partial discharge detection signal is 4. The transformer fault diagnosis method based on empirical wavelet and Riemann tangent space features according to claim 1, characterized in that, The specific steps of the third step are as follows: Step 1: Use the covariance matrix of the transformer vibration signal as the statistical feature of the data. The covariance matrix of the transformer vibration signal for a single test is: Step 2: Concatenate the covariance matrices of the j experiments {P1...P j The Riemann mean of j experiments is calculated using the following formula. That is, the Riemann Center Step 3: Using the AIRM metric, calculate the m-th order Riemann tangent space eigenvector using the following formula, to characterize the deviation of a single covariance matrix from its Riemann center: Step 4: Similarly, calculate the m-th order Riemann tangent space eigenvector of the transformer partial discharge detection signal as follows:
5. The transformer fault diagnosis method based on empirical wavelet and Riemann tangent space features according to claim 1, characterized in that, The specific steps of the fourth step are as follows: Step 1: Construct a joint feature vector. Combine the Riemann tangent space features of the transformer vibration signal with the Riemann tangent space features of the transformer partial discharge detection signal to construct the joint feature vector shown below: The joint feature vector f represents data from the same trial. Step 2: Input the feature vectors of all samples as the training set into the Siamese Support Vector Machine to train the classifier, and use the Bat Algorithm to optimize the parameters of the Siamese Support Vector Machine, including penalty factor c1, penalty factor c2 and kernel parameter λ, and establish a classification model with the optimized c1, c2 and λ. Step 3: Set relevant parameters, including bat population size N, maximum number of iterations M, and foraging space dimension dim; Step 4: Randomly generate the position x of bat i. i and speed v i bat's position x i The three parameters representing TWSVM—penalty factor c1, penalty factor c2, and kernel parameter λ—are used to evaluate the individual fitness of bats and find the optimal solution x at the current time step. * The bat individual with the highest fitness is the global optimal solution. This invention uses the penalty factor c and kernel parameter λ, along with the recognition accuracy obtained from cross-validation, as the standard for evaluating fitness, defined as: Step 5: Substitute the location of the bat population into the evaluation fitness function, continuously update the location corresponding to the best bat individual until the iteration termination condition is met, record the optimal parameter value, and substitute the globally optimal parameters into the Siamese support vector machine to complete the training of the classifier model. Step 6: Input the obtained fused feature vector into the trained classifier to identify multiple types of fault signals.