Transient signal-based power grid fault analysis method

Abstract

The invention provides a transient signal-based power grid fault analysis method, which comprises the following steps of: acquiring an original signal of a power grid, and preprocessing the original signal to obtain a corresponding transient signal; wherein the preprocessing at least comprises signal filtering, signal enhancement, signal enhancement and consistency correction; decomposing the transient signal into a plurality of quasi-stationary sub-signals through a preset separation model; performing continuous wavelet transform, time domain path feature extraction and frequency domain path feature extraction on the stationary sub-signals in sequence, and extracting corresponding time domain features and frequency domain features; and fusing the time domain features and the frequency domain features, determining a corresponding fault type and a confidence probability according to a fusion result, and determining a power grid fault through the fault type and the confidence probability. According to the method, the multiple characteristic quantities are comprehensively analyzed and mined, so that the cause of the fault and the specific position of the fault are determined, fault analysis is more comprehensive and deep, and power grid fault processing and prevention are assisted.

Description

Technical Field

[0001] The present invention relates to the technical field of transient signal processing, and in particular to a method for analyzing power grid faults based on transient signals. Background Art

[0002] Comprehensive analysis of power grid faults typically involves acquiring grid fault data, preprocessing the data, obtaining multiple fault characteristic data sets based on the preprocessed data, and identifying the fault type based on the fault characteristic data to obtain comprehensive analysis results. However, these methods often suffer from shortcomings such as low modeling capabilities for time series signals when faced with complex waveforms; interference resistance, ranging from "fragile threshold" to "robust modeling"; and difficulty in distinguishing noise from true transient characteristics using filtering algorithms, resulting in increased false alarm rates. Summary of the Invention

[0003] The purpose of the present invention is to propose a method for analyzing power grid faults based on transient signals, and to solve the technical problem of how to realize fault mode identification through multiple fault characteristic quantities.

[0004] On the one hand, a method for analyzing power grid faults based on transient signals is provided, comprising:

[0005] Obtaining the original signal of the power grid and preprocessing the original signal to obtain the corresponding transient signal; wherein the preprocessing includes at least signal filtering, signal enhancement, signal enhancement and consistency correction;

[0006] Decompose the transient signal into multiple quasi-stationary sub-signals through a preset separation model;

[0007] The quasi-stationary sub-signal is subjected to continuous wavelet transform, time domain path feature extraction and frequency domain path feature extraction in sequence to extract the corresponding time domain features and frequency domain features. The time domain features and frequency domain features are fused, and the corresponding fault type and confidence probability are determined based on the fusion results. The power grid fault is determined based on the fault type and confidence probability.

[0008] Preferably, the preprocessing of the original signal includes performing at least five layers of multi-scale wavelet decomposition on the original signal using a preset wavelet basis function to obtain coefficients of each scale;

[0009] An adjustment threshold for dynamic adaptation is estimated based on noise energy at each scale, and a threshold function of a preset wavelet threshold filter is updated according to the adjustment threshold and each scale coefficient;

[0010] The original signal is filtered using the updated threshold function.

[0011] Preferably, the preprocessing of the original signal further comprises performing a short-time Fourier transform on the original signal after signal filtering using a preset time-frequency matrix to generate a time-frequency graph.

[0012] Preferably, the preprocessing of the original signal also includes taking the time-frequency graph as input features and inputting it into a preset deep neural network for denoising and enhancement; wherein the preset deep neural network includes at least a CNN branch, an LSTM branch, a spatial attention module, a feature fusion layer and a fully connected layer; multiple layers of convolution are set in the CNN branch to extract local time-frequency features, and the spatial attention module is used to enhance the weight of the high-frequency area; the LSTM branch expands the time-frequency graph into a sequence according to time slices to capture the long-term dependency of the non-stationary signal and obtain time series features; the feature fusion layer splices the spatial features output by the CNN branch with the time series features output by the LSTM, and generates a denoised time-frequency graph through the fully connected layer.

[0013] Preferably, the preprocessing of the original signal further includes performing a short-time Fourier transform on the amplitude of the time-frequency graph output by the neural network, and obtaining a final restored time-frequency graph when the number of iterations reaches a preset threshold;

[0014] The gain is applied to the transient interval according to the final restored time-frequency diagram to perform consistency correction.

[0015] Preferably, the separation model includes,

[0016]

[0017] Among them, u k (t) is the kth quasi-stationary sub-signal, ω k is the center frequency, is the time derivative of the signal, δ(t)+j / (πt) is the impulse response of the analytical signal, which is used to convert the real signal u k (t) is converted into a complex signal, K is the maximum value of k, and e is the base of the natural logarithm in the wavelet basis function.

[0018] Preferably, performing continuous wavelet transform on the quasi-stationary sub-signal includes:

[0019] Continuous wavelet transform is performed on each quasi-stationary sub-signal to extract the corresponding time-frequency features; and the time-frequency diagram of each quasi-stationary sub-signal is determined according to the time-frequency features.

[0020] Preferably, extracting time domain path features from the quasi-stationary sub-signal includes:

[0021] Convolutional coding is performed through independent convolution of each channel in the preset Transformer model to extract the local pattern of the time-frequency graph of each quasi-stationary sub-signal, and the global relationship across time and channels is determined through the preset attention model to obtain the corresponding time domain features.

[0022] Preferably, performing frequency domain path feature extraction on the quasi-stationary sub-signal includes:

[0023] Through the cross-scale time dimension convolution in the preset Transformer model, the frequency domain features of the time-frequency graph of each quasi-stationary sub-signal are extracted, and the maximum spatial pooling is performed along the scale dimension to obtain the final frequency domain features.

[0024] Preferably, the fusing of time domain features and frequency domain features includes:

[0025] Concatenate the time domain features and frequency domain features along the channel dimension to form a unified time-aligned feature matrix:

[0026] The concatenated time-aligned features are input into the encoder of the Transformer model. The time-aligned feature matrix is ​​transposed into a sequence format, and each time point is mapped to a one-dimensional feature vector. The features of each time point are updated to fuse global information and compressed into a global feature vector along the time dimension. The output is the fault type.

[0027] The probability distribution of the fault category is calculated through the output layer in the Transformer model to obtain the confidence probability.

[0028] In summary, the implementation of the embodiments of the present invention has the following beneficial effects:

[0029] The transient signal-based power grid fault analysis method provided by the present invention identifies fault patterns through multiple fault characteristic quantities. It not only focuses on a single characteristic quantity, but also comprehensively analyzes and mines multiple characteristic quantities to determine the cause of the fault and the specific location of the fault, making the fault analysis more comprehensive and in-depth, thereby assisting in power grid fault handling and prevention. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, without paying any creative work, other drawings obtained based on these drawings still fall within the scope of the present invention.

[0031] Figure 1The figure is a schematic diagram of the main process of a method for analyzing power grid faults based on transient signals in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be described in further detail below with reference to the accompanying drawings.

[0033] like Figure 1 FIG. 1 is a schematic diagram of an embodiment of a method for analyzing power grid faults based on transient signals provided by the present invention. In this embodiment, the method includes the following steps:

[0034] Step S1: Obtain the original signal of the power grid and preprocess it to obtain the corresponding transient signal; wherein the preprocessing includes at least signal filtering, signal enhancement, signal enhancement, and consistency correction. Noise can be removed as much as possible while preserving the transient characteristics.

[0035] In a specific embodiment, the preprocessing of the original signal includes performing at least five layers of multi-scale wavelet decomposition on the original signal using a preset wavelet basis function to obtain scale coefficients; estimating a dynamic adaptive threshold value based on the noise energy at each scale, and updating a preset threshold function for wavelet threshold filtering based on the adjusted threshold value and the scale coefficients; and filtering the original signal using the updated threshold function. Specifically, a complex Morlet wavelet basis function is selected to perform a five-layer multi-scale wavelet decomposition on the transient signal x(t). The wavelet basis function is:

[0036] w=5(center frequency)

[0037] Get the scale coefficient W j (k), where j is the number of decomposition levels and k is the time index.

[0038]

[0039] Based on the noise energy estimation at each scale, the threshold is adjusted dynamically and adaptively:

[0040]

[0041] where N j is the coefficient length of the jth layer, σ j is the noise standard deviation estimate. Then, a continuously differentiable Sigmoid threshold function is used to improve the threshold function of the wavelet threshold filter to avoid the discontinuity of the hard threshold and the constant deviation of the soft threshold:

[0042]

[0043] Furthermore, the original signal after signal filtering is subjected to a short-time Fourier transform (STFT) using a preset time-frequency matrix to generate a time-frequency graph. Specifically, the time-frequency matrix is ​​constructed and the preprocessed signal is subjected to a short-time Fourier transform (STFT) to generate a time-frequency graph S(m,n), which serves as the input feature of the deep learning network.

[0044]

[0045] Where ω is the Hanning window, L is the window length, and N is the number of FFT points.

[0046] Furthermore, the time-frequency map is used as input features and fed into a pre-set deep neural network for denoising and enhancement. The pre-set deep neural network comprises at least a CNN branch, an LSTM branch, a spatial attention module, a feature fusion layer, and a fully connected layer. The CNN branch employs multiple convolution layers to extract local time-frequency features, and the spatial attention module enhances the weights of high-frequency regions. The LSTM branch expands the time-frequency map into sequences by time slice to capture the long-term dependencies of non-stationary signals and obtain temporal features. The feature fusion layer concatenates the spatial features output by the CNN branch with the temporal features output by the LSTM, and generates a denoised time-frequency map via a fully connected layer. Specifically, a hybrid CNN-LSTM model is used for denoising and enhancement. The model structure is as follows: the CNN branch employs three convolution layers (3×3 kernel, 32 / 64 / 128 channels) to extract local time-frequency features, and the spatial attention module enhances the weights of high-frequency regions. The LSTM branch expands the time-frequency map into sequences by time slice and feeds them into a bidirectional LSTM (256 hidden units) to capture the long-term dependencies of non-stationary signals. The feature fusion layer combines the spatial features output by CNN with the temporal features output by LSTM, and generates a denoised time-frequency graph through the fully connected layer.

[0047] The loss function of the neural network is designed based on multi-objective joint optimization. The time-frequency domain amplitude loss function is:

[0048]

[0049] The time domain waveform similarity loss function is:

[0050]

[0051] The loss function of the high-frequency component penalty term (in order to protect transient characteristics) is:

[0052]

[0053] Therefore, it can be proposed that the total loss function is:

[0054] L total =λ1Lspec +λ2L time +λ3L high .

[0055] Furthermore, the amplitude of the time-frequency graph output by the neural network is transformed from short-time Fourier transform. When the number of iterations reaches the preset threshold, the final restored time-frequency graph is obtained. According to the final restored time-frequency graph, the transient interval is subjected to gain correction for consistency. Specifically, the amplitude of the time-frequency graph output by the neural network is The Griffin-Lim algorithm is iterated to recover the phase:

[0056]

[0057] After 5 iterations, the final restored time-frequency diagram can be obtained.

[0058] After signal reconstruction, a gain is applied to the transient interval (detected by the Teager energy operator) to enhance the transient component:

[0059] x final (t) = x denoised (t)·[1+0.5·E Teager (t)]

[0060] The Teager energy operator is:

[0061] E Teager (t) = x 2 (t)-x(t-1)x(t+1)

[0062] In summary, through adaptive time-frequency preprocessing, the improved wavelet threshold function can take into account both noise suppression and transient feature retention; the cross-domain joint learning of the CNN-LSTM hybrid network can simultaneously capture the local time-frequency features and the global evolution law; at the same time, by adding a high-frequency penalty term to the physical constraint loss function, the excessive smoothing of transient components by traditional methods is avoided; finally, a lightweight design is adopted to reduce the number of network parameters to <1M, which can be deployed in embedded edge devices.

[0063] Step S2, decompose the transient signal into multiple quasi-stationary sub-signals through a preset separation model; after the above preprocessing, a non-stationary transient signal x with feature preservation and noise suppression can be obtained. final (t), decompose the non-stationary signal into K quasi-stationary sub-signals (IMF) to realize the analysis of fault characteristics and noise.

[0064] In a specific embodiment, the separation model includes:

[0065]

[0066] Among them, u k(t) is the kth quasi-stationary sub-signal, ω k is the center frequency, is the time derivative of the signal, δ(t)+j / (πt) is the impulse response of the analytical signal, which is used to convert the real signal u k (t) is converted to a complex signal, K is the maximum value of k, and e is the base of the natural logarithm in the wavelet basis function. Generally, K is usually set to 3 to 5 based on spectrum peak estimation and experimental experience.

[0067] Performing continuous wavelet transform on the quasi-stationary sub-signal includes performing continuous wavelet transform on each quasi-stationary sub-signal to extract corresponding time-frequency features; and determining a time-frequency graph of each quasi-stationary sub-signal according to the time-frequency features.

[0068] Perform continuous wavelet transform (CWT) on each IMF to extract time-frequency features:

[0069]

[0070] In order to take into account both time and frequency resolution, ψ(t) is a complex Morlet wavelet basis:

[0071] ω0=6

[0072] According to the above steps, the time-frequency diagram of each IMF can be obtained (Scale A, time point T).

[0073] The time domain path feature extraction of the quasi-stationary sub-signal includes performing convolution coding through independent convolution of each channel in the preset Transformer model, extracting the local pattern of the time-frequency graph of each quasi-stationary sub-signal, and determining the global relationship between time and channels through the preset attention model to obtain the corresponding time domain features. Input the IMF time domain signal {u1(t),u2(t),...,u K (t)}. 1D convolutional coding with independent convolution per channel is used to extract local patterns:

[0074]

[0075] Adding learnable positional encodings

[0076] Use a lightweight Transformer model and introduce a local attention mechanism (limiting the window size to L = 32 to minimize the amount of computation) to capture global relationships across time and channels:

[0077]

[0078] Where M is a local mask matrix, and only adjacent L points are allowed to interact.

[0079] The frequency domain path feature extraction of the quasi-stationary sub-signal includes extracting the frequency domain features of the time-frequency graph of each quasi-stationary sub-signal through the cross-scale time dimension convolution in the preset Transformer model, and performing maximum spatial pooling along the scale dimension to obtain the final frequency domain features. Input the time-frequency graph of each IMF {W1, W2, ... W K}, using 2D convolutional coding with cross-scale-time dimension convolution to extract frequency domain features:

[0080]

[0081] Doing maximum spatial pooling along the scale dimension yields

[0082] In step S3, the quasi-stationary subsignal undergoes continuous wavelet transform, time-domain path feature extraction, and frequency-domain path feature extraction in sequence to extract the corresponding time-domain and frequency-domain features. These features are then fused, and the corresponding fault type and confidence probability are determined based on the fusion results. The power grid fault is then identified based on the fault type and confidence probability. Feature fusion is the core component of the hybrid architecture, aiming to organically combine the multimodal features extracted from the time and frequency domain branches. Time-domain features capture transient details, while frequency-domain features reveal energy distribution. The combined features cover the multi-dimensional characteristics of the signal, transforming the feature extraction method from "manual experience" to "automated learning." VMD decomposition is used to separate noise into specific IMFs (such as high-frequency IMF1). The Transformer attention mechanism then dynamically suppresses low-contributing components, ultimately achieving adaptive noise suppression. The Transformer and self-attention mechanisms automatically learn the correlation weights between time-frequency and IMFs, eliminating the need for manually setting fusion rules. Global attention also maintains a manageable computational load within a set timeframe, making it suitable for embedded deployment.

[0083] In a specific embodiment, the fusion of time domain features and frequency domain features includes: splicing the time domain features and frequency domain features along the channel dimension to form a unified time-aligned feature matrix: inputting the spliced ​​time-aligned features into the encoder in the Transformer model, transposing the time-aligned feature matrix into a sequence format, and corresponding each time point to a one-dimensional feature vector, updating the features of each time point to fused global information, and compressing them into a global feature vector along the time dimension, and outputting the fault type; calculating the probability distribution of the fault category through the output layer in the Transformer model to obtain the confidence probability. It is understandable that the two branches need to be aligned through the branch output of the time domain-frequency domain. After lightweight Transformer encoding, the time domain feature dimension of each IMF (K is the number of IMFs) is (Optional number of channels D = 64, time point T' = 128); After 2D CNN and scale pooling, the frequency domain feature dimension of each IMF is (Optional number of channels C = 64, time point T' = 128) Splice the time domain and frequency domain features along the channel dimension to form a unified time-aligned feature matrix:

[0084]

[0085] Then, the concatenated features are input into the global Transformer encoder, and the time-aligned feature matrix is ​​transposed and converted into a sequence format, namely:

[0086]

[0087] Each time point corresponds to a (C+D)K-dimensional feature vector as the input token of the Transformer. Add a learnable positional encoding Preserve timing information, i.e.

[0088]

[0089] In the encoding layer of Transformer, cross-modal dependencies are modeled through global self-attention.

[0090]

[0091] The features of each time point are updated to be expressed as fusion global information:

[0092] Next, Z is averaged along the time dimension T' and compressed into a global feature vector:

[0093]

[0094] In this way, the interference of time series fluctuations can be suppressed and the steady-state features most relevant to classification can be retained.

[0095] Finally, Softmax normalization is used in the output layer to obtain the probability distribution of the fault category:

[0096] y=Softmax(W·Z Global +b)

[0097] in, N is the number of fault categories.

[0098] After the above calculations and analysis, the model will give the fault type label and confidence probability corresponding to the current input waveform, thereby directly guiding the operation and maintenance personnel to take corresponding treatment measures (such as tripping, isolating the fault area).

[0099] In summary, the implementation of the embodiments of the present invention has the following beneficial effects:

[0100] The transient signal-based power grid fault analysis method provided by the present invention identifies fault patterns through multiple fault characteristic quantities. It not only focuses on a single characteristic quantity, but also comprehensively analyzes and mines multiple characteristic quantities to determine the cause of the fault and the specific location of the fault, making the fault analysis more comprehensive and in-depth, thereby assisting in power grid fault handling and prevention.

[0101] The above disclosure is merely a preferred embodiment of the present invention and certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for analyzing power grid faults based on transient signals, characterized in that: include: Obtaining the original signal of the power grid and preprocessing the original signal to obtain the corresponding transient signal; wherein the preprocessing includes at least signal filtering, signal enhancement, signal enhancement and consistency correction; Decompose the transient signal into multiple quasi-stationary sub-signals through a preset separation model; The quasi-stationary sub-signal is subjected to continuous wavelet transform, time domain path feature extraction and frequency domain path feature extraction in sequence to extract the corresponding time domain features and frequency domain features. The time domain features and frequency domain features are fused, and the corresponding fault type and confidence probability are determined based on the fusion results. The power grid fault is determined based on the fault type and confidence probability.

2. The method according to claim 1, wherein The preprocessing of the original signal includes performing at least five layers of multi-scale wavelet decomposition on the original signal using a preset wavelet basis function to obtain coefficients of each scale; An adjustment threshold for dynamic adaptation is estimated based on noise energy at each scale, and a threshold function of a preset wavelet threshold filter is updated according to the adjustment threshold and each scale coefficient; The original signal is filtered using the updated threshold function.

3. The method according to claim 2, wherein The preprocessing of the original signal further includes performing a short-time Fourier transform on the original signal after signal filtering using a preset time-frequency matrix to generate a time-frequency graph.

4. The method according to claim 3, wherein The preprocessing of the original signal also includes taking the time-frequency graph as input features and inputting it into a preset deep neural network for denoising and enhancement; wherein the preset deep neural network includes at least a CNN branch, an LSTM branch, a spatial attention module, a feature fusion layer and a fully connected layer; multiple layers of convolution are set in the CNN branch to extract local time-frequency features, and the spatial attention module is used to enhance the weight of the high-frequency area; the LSTM branch expands the time-frequency graph into a sequence according to time slices to capture the long-term dependency of the non-stationary signal and obtain time series features; the feature fusion layer splices the spatial features output by the CNN branch with the time series features output by the LSTM, and generates a denoised time-frequency graph through the fully connected layer.

5. The method according to claim 4, wherein The preprocessing of the original signal further includes performing a short-time Fourier transform on the amplitude of the time-frequency graph output by the neural network, and obtaining a final restored time-frequency graph when the number of iterations reaches a preset threshold; The gain is applied to the transient interval according to the final restored time-frequency diagram to perform consistency correction.

6. The method according to claim 5, wherein The separation model includes, Among them, u k (t) is the kth quasi-stationary sub-signal, ω k is the center frequency, is the time derivative of the signal, δ(t)+j / (πt) is the impulse response of the analytical signal, which is used to convert the real signal u k (t) is converted into a complex signal, K is the maximum value of k, and e is the base of the natural logarithm in the wavelet basis function.

7. The method according to claim 6, wherein The continuous wavelet transform of the quasi-stationary sub-signal includes: Continuous wavelet transform is performed on each quasi-stationary sub-signal to extract the corresponding time-frequency features; and the time-frequency diagram of each quasi-stationary sub-signal is determined according to the time-frequency features.

8. The method according to claim 7, wherein The time domain path feature extraction of quasi-stationary sub-signals includes: Convolutional coding is performed through independent convolution of each channel in the preset Transformer model to extract the local pattern of the time-frequency graph of each quasi-stationary sub-signal, and the global relationship across time and channels is determined through the preset attention model to obtain the corresponding time domain features.

9. The method according to claim 8, wherein Frequency domain path feature extraction of quasi-stationary sub-signals includes: Through the cross-scale time dimension convolution in the preset Transformer model, the frequency domain features of the time-frequency graph of each quasi-stationary sub-signal are extracted, and the maximum spatial pooling is performed along the scale dimension to obtain the final frequency domain features.

10. The method according to claim 9, wherein The fusion of time domain features and frequency domain features includes: Concatenate the time domain features and frequency domain features along the channel dimension to form a unified time-aligned feature matrix: The concatenated time-aligned features are input into the encoder of the Transformer model. The time-aligned feature matrix is ​​transposed into a sequence format, and each time point is mapped to a one-dimensional feature vector. The features of each time point are updated to fuse global information and compressed into a global feature vector along the time dimension. The output is the fault type. The probability distribution of the fault category is calculated through the output layer in the Transformer model to obtain the confidence probability.

Images