Mutual inductance equipment risk assessment method
Patent Information
- Application Number
- CN202411791587.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-06
- Publication Date
- 2025-05-13
Smart Images
Figure CN119989071A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of mutual inductance equipment monitoring and diagnosis, in particular to a mutual inductance equipment risk assessment method based on variational mode decomposition-converter. Background Art
[0002] With the rapid development of intelligent technology, mutual inductance equipment, as a key equipment, has received increasing attention for its risk assessment technology. When the mutual inductance equipment is closed without load, it may generate an excitation surge current several times the rated current, causing waveform distortion, which not only produces mechanical and thermal stress on the winding of the mutual inductance equipment, but also may cause current mutual inductance saturation, thus affecting the normal operation of the protection device.
[0003] Traditional methods for identifying magnetizing inrush currents rely mainly on the analysis of current waveform characteristics, but these methods have limited performance under complex disturbance conditions. In recent years, deep learning techniques, especially converter networks combined with attention mechanisms, have shown high diagnostic accuracy in complex fault scenarios. However, existing risk assessment methods based on converter networks still lack attention to frequency characteristics, resulting in poor diagnostic results in some cases.
[0004] Variational mode decomposition is an adaptive signal decomposition method that can effectively obtain key frequency information in the signal. Combining mode with deep learning technology has achieved remarkable results in mechanical risk assessment. However, the application of mode in risk assessment of mutual inductance equipment is still insufficient, and its diagnostic reliability needs to be further improved. Summary of the invention
[0005] The present invention provides a mutual inductance equipment risk assessment method, which solves the problems of insufficient attention to key frequency information and low diagnostic accuracy in the prior art based on variational modal decomposition and entropy converter network. The method uses modal technology to decompose the excitation inrush current signal, extracts key modes, and enhances the expression ability of signal features through entropy calculation, and finally combines the converter network to achieve accurate diagnosis of mutual inductance equipment failure risks.
[0006] The technical solution of the present invention is as follows:
[0007] A mutual induction equipment risk assessment method, the method comprising the following steps:
[0008] S1: Collect excitation inrush current signal from mutual inductance equipment;
[0009] S2: Establish a feature set, including the intrinsic mode function of the excitation inrush current signal and its corresponding key frequency;
[0010] S3: Perform modal component, spatial and feature encoding, segment the obtained modal components in time order, and map these segments to high-dimensional space through trainable linear transformation to form pattern encoding, embedding learnable spatial encoding and feature encoding in the modal components;
[0011] S4: Use the pattern encoder to convert the feature set into the input of the transformer model;
[0012] S5: Use self-attention mechanism to enhance the modal part that is most relevant to the feature;
[0013] S6: Introduce the soft maximum function and construct a risk assessment model for mutual inductance equipment based on variational mode-converter network;
[0014] S7: Return risk assessment information to the system user interface. The interface displays information including the excitation inrush current signal, modal decomposition results, fault type and its probability. If an abnormal fault type is detected, the system automatically issues a risk warning.
[0015] Preferably, in step S2, a feature set is constructed as an input of the converter network, the feature set including modal components, key frequencies, and complex information entropy for evaluating waveform complexity, and the specific steps are as follows:
[0016] S2.1: Decompose the input waveform using the variational modal algorithm to obtain the modal components and the corresponding key frequencies:
[0017]
[0018] In the formula, Represents the derivative with respect to time t, the complex exponential term is a phasor, k is the predefined mode number of decomposition, f is the original signal to be decomposed, δ is the Dirac distribution, * represents the convolution operation, β is the quadratic penalty factor, and λ represents the Lagrange multiplier; in general, the excitation inrush current signal f is decomposed into modal components {α1, α2, ..., α k} and the corresponding key frequencies {θ1,θ2,...,θ k}, at this time the intrinsic mode function can be regarded as a pure harmonic signal with amplitude and instantaneous frequency;
[0019] S2.2: Accumulate the complex information entropy of the intrinsic mode function components to evaluate the complex characteristics of the signal waveform SumTE:
[0020]
[0021] where the sequence X = {x(1), x(2), ..., x(N)} is an m-dimensional sequence; define d[X(i), X(j)], (i≠j) as H m H m+1(r) The maximum value of the difference between the corresponding elements of the two, then count the number of d < r, where r is the similarity tolerance. Generally, r = (0.1 - 0.25)std, and std is the standard deviation of the sequence. Here, r = 0.2std.
[0022] Preferably, in step S3, a learnable spatial encoding and feature encoding are embedded into the modal component. The spatial encoding is used to mark the key frequencies, and the feature encoding is used to mark the complex information entropy. The specific steps are as follows:
[0023] S3.1: α k In the time domain, it is divided into N segments with a length of L t segments, Then, these segments are mapped to the space through a trainable linear transformation:
[0024]
[0025] where d t is the length of a single segmented vector, and the parameter is trainable;
[0026] S3.2: Embed the learnable spatial encoding E pos and feature encoding E se .
[0027] Preferably, in step S4, a learned class label is introduced. This label follows the spatial encoding and feature encoding and is used as part of the model output. The specific steps are as follows:
[0028] S4.1: Introduce a learnable class label at the beginning of the modal sequence. This label follows the spatial encoding and is used as part of the model output;
[0029] S4.2: The final output can be expressed in the following form:
[0030]
[0031] Preferably, in step S5, the self-attention mechanism is used to strengthen the modal part most relevant to the fault features. The specific steps are as follows:
[0032] S5.1: The Transformer network encoder mainly consists of multi-head attention, a feed-forward network, a normalization layer, and a residual connection to accelerate convergence and the learning speed v. The multi-head self-attention mechanism can be expressed as:
[0033]
[0034] In the formula, the matrix is the input of the pattern encoder, Attention(·) represents single-head self-attention, which can be expressed as:
[0035]
[0036] S5.2: Take a multilayer perceptron with one hidden layer and represent it as:
[0037] MLP(X) = GeLU(XΠ 1 +b 1 )Π 2 +b 2 (9)
[0038] In the formula, Π 1 ,Π 2 ,b 1 and b 2 Represents the weights and biases of the two layers;
[0039] S5.3: GELU is used as the activation function, and the function is expressed as follows:
[0040]
[0041] S5.4: In the pattern encoder, the key frequency encoding and waveform features are the embedding of positions and features, respectively, which enhances the model's sensitivity to different parameter patterns, expressed as follows:
[0042]
[0043] Among them, n class Indicates the class label, E f represents the encoding key frequency, E SumTE Indicates waveform feature encoding;
[0044] S5.5: The output is then embedded into the input converter encoder, which is represented by the following equation:
[0045]
[0046] Where LN(·) represents the normalization layer and i represents the number of encoder layers.
[0047] Preferably, in step S6, a normalized exponential function is introduced to construct a mutual inductance equipment risk assessment model based on a variational mode-converter network, and the specific steps are as follows:
[0048] S6.1: After extracting features using the pattern encoder and transformer network encoder, a classifier function consisting of a fully connected layer and a normalized exponential function is used to associate the extracted features with the fault type;
[0049] S6.2: The output from the converter network is input into the final classifier, which utilizes the global characteristics of the signal and calculates the probability distribution of various mutual inductance device fault types through a normalized exponential function to achieve the identification of the excitation inrush current signal and the diagnosis of the fault mode.
[0050] Compared with the prior art, the present invention has the following beneficial effects:
[0051] 1. The present invention combines variational modal decomposition technology with a converter network to extract key modal features from complex excitation inrush current signals, enhances the expressiveness of signal features through entropy calculation, and pays more attention to key frequency information in the signal, effectively solving the problem of insufficient attention to frequency features in traditional methods. Using the self-attention mechanism, the model can automatically focus on the signal part that is most relevant to the mutual inductance equipment failure, improving the accuracy of risk assessment. Through entropy coding, the model's sensitivity to the complexity of the signal waveform is enhanced, improving the robustness and reliability of risk assessment. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1 is a risk assessment flow chart of the present invention;
[0053] Figure 2 It is a phase excitation surge current signal waveform and an intrinsic mode function reconstructed signal waveform diagram of the present invention;
[0054] Figure 3 It is the spectrum diagram of the intrinsic mode function of the present invention;
[0055] Figure 4 It is the waveform diagram of the intrinsic mode function of the present invention. DETAILED DESCRIPTION
[0056] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.
[0057] Example:
[0058] like Figure 1-4 As shown, in this embodiment, the mutual inductance device takes the current transformer in the transformer as an example, firstly, the excitation inrush current signal is collected from the current transformer (CT) of the power transformer, and the sampling frequency is set to 2kHz to ensure high-resolution capture of the signal. The collected excitation inrush current signal data includes signals under various disturbance scenarios, such as no-load closing, turn-to-turn fault, CT saturation and other complex working conditions.
[0059] Table 1 Parameters of the excitation inrush current circuit model
[0060]
[0061] In order to verify the performance risk assessment of the proposed method in the field of transformer faults, the reference (Research on fast identification algorithm of excitation inrush current based on complex waveform characteristics, authors: Shen Chuncheng, Yan Baiping, Huang Dazhuo, etc.) takes residual magnetism and closing angle as the main influencing factors of excitation inrush current changes, and establishes an excitation inrush current simulation data set under multi-scenario disturbances.
[0062] (1) Scenario 1: Closing the circuit breaker without load
[0063] The scenario simulates the no-load closing of the transformer under the conditions of no residual magnetism, balanced residual magnetism, and unbalanced residual magnetism. The closing angle ranges from 0 to 360 degrees, and the residual magnetism values are 0, ±0.2, ±0.5, and ±0.8 pu, respectively. This scenario has a total of 160 different working conditions.
[0064] (2) Scenario 2: Closing the circuit breaker with no load due to turn-to-turn fault
[0065] This scenario includes 5% inter-turn short circuit and inter-turn ground fault, and the phase angle when closing the circuit breaker also covers from 0 to 360 degrees. This scenario has a total of 30 working conditions.
[0066] (3) Scenario 3: Closing the circuit breaker without load when CT is saturated
[0067] Similar to scenario 1, this scenario simulates no-load closing under the conditions of no residual magnetism, balanced residual magnetism, and unbalanced residual magnetism. In addition, the value range of residual magnetism ra and the setting of closing angle are the same as those in scenario 1, with a total of 160 working conditions.
[0068] (4) Scenario 4: Transformer internal fault
[0069] The scenarios consider turn-to-turn short circuit and turn-to-ground fault, with turn ratios of 5%, 8%, 10% and 20%. The phase angle at closing ranges from 0 to 360 degrees. There are 82 different fault conditions in total.
[0070] (5) Scenario 5: Fault in transformer area
[0071] This scenario involves single-phase grounding, phase-to-phase, phase-to-phase grounding, and three-phase short circuit faults. The phase angle when the fault occurs is also from 0 to 360 degrees. This scenario involves a total of 100 working conditions.
[0072] The risk assessment process of the mutual induction equipment of the present invention is as follows:
[0073] S1: Collect the excitation inrush current signal from the current transformer (CT) of the transformer;
[0074] S2: Establish a feature set, including the intrinsic mode function of the excitation inrush current signal and its corresponding key frequency;
[0075] S2.1: Decompose the input waveform using the variational modal algorithm to obtain the modal components and the corresponding key frequencies:
[0076]
[0077] In the formula, represents the derivative with respect to time t, and the complex exponential term is a phasor, k is the predefined number of modes for decomposition, f is the original signal to be decomposed, δ is the Dirac distribution, * represents the convolution operation, β is the quadratic penalty factor, and λ represents the Lagrange multiplier; generally speaking, the inrush current signal f is decomposed into modal components {α1, α2,..., α k} and the corresponding key frequencies {θ1, θ2,..., θ k}, and at this time the intrinsic mode function can be regarded as a pure harmonic signal with amplitude and instantaneous frequency;
[0078] S2.2: Accumulate the complex information entropy of the intrinsic mode function components to evaluate the complex characteristics of the signal waveform:
[0079]
[0080] where the sequence X = {x(1), x(2),..., x(N)} is an m-dimensional sequence; define d[X(i), X(j)], (i≠j) as the maximum value of the difference between the corresponding elements of H m H m+1 (r), then count the number of d < r, r is the similarity tolerance, generally r = (0.1 - 0.25)std, std is the standard deviation of the sequence, and here r = 0.2std.
[0081] Table 2 Complex information entropy and maximum value Smax at different fault turn ratios during internal faults
[0082]
[0083] S3: Perform modal component, spatial, and feature encoding. Segment the obtained modal components in chronological order and map these segments to a high-dimensional space through a trainable linear transformation to form pattern encoding, and embed learnable spatial encoding and feature encoding in the modal components;
[0084] S3.1: α k is divided into N segments with a length of L t in the time domain, then, these segments are mapped to the space through a trainable linear transformation:
[0085]
[0086] where d t is the length of a single segmented vector, and the parameter It is trainable;
[0087] S3.2: Embedding learnable spatial encodings E into modal components pos and feature encoding E se .
[0088] S4: Use the pattern encoder to convert the feature set into the input of the transformer model;
[0089] S4.1: Introducing a learnable category marker at the beginning of the modal sequence This tag follows the spatial encoding and is used as part of the model output;
[0090] S4.2: The final output can be expressed as follows:
[0091]
[0092] S5: Use the self-attention mechanism to enhance the modal part that is most relevant to the feature. The specific steps are as follows:
[0093] S5.1: The transformer network encoder is mainly composed of multi-head attention as well as a feedforward network, a normalization layer, and a residual connection to accelerate convergence and learning speed v, where the multi-head self-attention mechanism can be expressed as:
[0094]
[0095] In the formula, the matrix is the input of the pattern encoder, Attention(·) represents single-head self-attention, which can be expressed as:
[0096]
[0097] S5.2: Take a multilayer perceptron with one hidden layer and represent it as:
[0098] MLP(X) = GeLU(XΠ 1 +b 1 )Π 2 +b 2 (9)
[0099] In the formula, Π 1 ,Π 2 ,b 1 and b 2 Represents the weights and biases of the two layers;
[0100] S5.3: GELU is used as the activation function, and the function is expressed as follows:
[0101]
[0102] S5.4: In the pattern encoder, the key frequency encoding and waveform features are the embedding of positions and features, respectively, which enhances the model's sensitivity to different parameter patterns, expressed as follows:
[0103]
[0104] Among them, n class Indicates the class label, E f represents the encoding key frequency, E SumTE Indicates waveform feature encoding;
[0105] S5.5: The output is then embedded into the input converter encoder, which is represented by the following equation:
[0106]
[0107] Where LN(·) represents the normalization layer and i represents the number of encoder layers.
[0108] S6: Introduce the soft maximum function and construct a risk assessment model for mutual inductance equipment based on variational mode-converter network;
[0109] S6.1: After extracting features using the pattern encoder and transformer network encoder, a classifier function consisting of a fully connected layer and a normalized exponential function is used to associate the extracted features with the fault type;
[0110] S6.2: The output from the converter network is input into the final classifier, which utilizes the global characteristics of the signal and calculates the probability distribution of various mutual inductance device fault types through a normalized exponential function to achieve the identification of the excitation inrush current signal and the diagnosis of the fault mode.
[0111] S7: Return risk assessment information to the system user interface. The interface displays information including the excitation inrush current signal, modal decomposition results, fault type and its probability. If an abnormal fault type is detected, the system automatically issues a risk warning.
[0112] The embodiments of the present invention are provided for the purpose of illustration and description. Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations of the present invention. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present invention.
Claims
1. A mutual inductance equipment risk assessment method, characterized in that: The method comprises the following steps: S1: Collect excitation inrush current signal from mutual inductance equipment; S2: Establish a feature set, including the modal components of the excitation inrush current signal, key frequencies, and complex information entropy for evaluating the complexity of the waveform; S3: Perform modal component, spatial and feature encoding, segment the obtained modal components in time order, and map these segments to high-dimensional space through trainable linear transformation to form pattern encoding, embedding learnable spatial encoding and feature encoding in the modal components; S4: Use the pattern encoder to convert the feature set into the input of the transformer model; S5: Use self-attention mechanism to enhance the modal part that is most relevant to the feature; S6: Introduce the soft maximum function and construct a risk assessment model for mutual inductance equipment based on variational mode-converter network; S7: Return risk assessment information to the system user interface. The interface displays information including the excitation inrush current signal, modal decomposition results, fault type and its probability. If an abnormal fault type is detected, the system automatically issues a risk warning.
2. A mutual induction equipment risk assessment method according to claim 1, characterized in that: In step S2, a feature set is constructed as the input of the converter network. The feature set includes modal components, key frequencies, and complex information entropy for evaluating waveform complexity. The specific steps are as follows: S2.1: Decompose the input waveform using the variational modal algorithm to obtain the modal component α k and the corresponding key frequency θ k : In the formula, Represents the derivative with respect to time t, the complex exponential term is a phasor, K is the predefined modal number of decomposition, f is the original signal to be decomposed, δ is the Dirac distribution, * represents the convolution operation, β is the quadratic penalty factor, and λ represents the Lagrange multiplier; in general, the excitation inrush current signal f is decomposed into modal components {α1, α2, ..., α k } and the corresponding key frequencies {θ1,θ2,...,θ k }, at this time the intrinsic mode function can be regarded as a pure harmonic signal with amplitude and instantaneous frequency; S2.2: Accumulate the complex information entropy SumTE of the intrinsic mode function components to evaluate the complex characteristics of the signal waveform: Among them, the sequence X = {x(1), x(2),..., x(N)} is an m-dimensional sequence; define d[X(i), X(j)], (i ≠ j) as H m H m+1 (r) the maximum value of the difference between the corresponding elements of the two. Then count the number of d < r, where r is the similarity tolerance. Generally, r = (0.1 - 0.25)std, std is the standard deviation of the sequence, and here r = 0.2std.
3. A mutual induction equipment risk assessment method according to claim 1, characterized in that: In step S3, the modal component is embedded with learnable spatial encoding and feature encoding, where the spatial encoding is used to mark the key frequency θ k , feature encoding is used to mark complex information entropy SumTE, the specific steps are as follows: S3.1: α k In the time domain, it is divided into t N segment, Next, a trainable linear transformation is used to map these segments to In the space of: Among them, d t is the length of a single segment vector, and the parameter It is trainable; S3.2: Embedding learnable spatial encodings E into modal components pos and feature encoding E se .
4. A mutual induction equipment risk assessment method according to claim 1, characterized in that: In step S4, the learned category label is introduced, which follows the spatial encoding and feature encoding and is used as part of the model output. The specific steps are as follows: S4.1: Introducing a learnable category marker at the beginning of the modal sequence This tag follows the spatial encoding and is used as part of the model output; S4.2: The final output can be expressed as follows:
5. A mutual induction equipment risk assessment method according to claim 1, characterized in that: In step S5, the self-attention mechanism is used to strengthen the modal part that is most relevant to the fault characteristics. The specific steps are as follows: S5.1: The transformer network encoder is mainly composed of multi-head attention as well as a feedforward network, a normalization layer, and a residual connection to accelerate convergence and learning speed v, where the multi-head self-attention mechanism can be expressed as: In the formula, the matrix is the input of the pattern encoder, Attention(·) represents single-head self-attention, which can be expressed as: S5.2: Take a multilayer perceptron with one hidden layer and represent it as: MLP(X)=GeLU(XΠ 1 +b 1 )Π 2 +b2 In the formula, Π 1 ,Π 2 ,b 1 and b 2 Represents the weights and biases of the two layers; S5.3: GELU is used as the activation function, and the function is expressed as follows: S5.4: In the pattern encoder, the key frequency encoding and waveform features are the embedding of positions and features, respectively, which enhances the model's sensitivity to different parameter patterns, expressed as follows: Among them, m class Indicates the class label, E f represents the encoding key frequency, E SumTE Indicates waveform feature encoding; S5.5: The output is then embedded into the input converter encoder, which is represented by the following equation: Where LN(·) represents the normalization layer and i represents the number of encoder layers.
6. A mutual induction equipment risk assessment method according to claim 1, characterized in that: In step S6, a normalized exponential function is introduced to construct a mutual inductance equipment risk assessment model based on a variational mode-converter network. The specific steps are as follows: S6.1: After extracting features using the pattern encoder and transformer network encoder, a classifier function consisting of a fully connected layer and a normalized exponential function is used to associate the extracted features with the fault type; S6.2: The output from the converter network is input into the final classifier, which utilizes the global characteristics of the signal and calculates the probability distribution of various mutual inductance device fault types through a normalized exponential function to achieve the identification of the excitation inrush current signal and the diagnosis of the fault mode.