A transformer operation fault detection method

Abstract

The present application belongs to the technical field of fault detection, and particularly relates to a transformer operation fault detection method. The present application first collects transformer vibration signals and performs pretreatment, adopts a double-feature joint criterion to accurately locate the mutation point and divide equal-length sub-sequences; obtains effective intrinsic mode function components by improving the variational mode decomposition, solving the end effect and modal aliasing problems of traditional decomposition methods; extracts the modal kinetic characteristics, cross-modal information flow characteristics and time sequence stability indexes of each component; inputs the multi-dimensional features and original sub-sequences into a fault detection model fused with multi-scale causal convolution and channel attention, and outputs the detection results. The present application effectively improves the signal quality and feature extraction reliability, and significantly enhances the detection accuracy and robustness of early weak faults of the transformer under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of fault detection application technology, and in particular relates to a method for detecting transformer operation faults. Background Technology

[0002] As a core and critical piece of equipment in the power system, the transformer undertakes the core functions of voltage transformation and power transmission. Its operational stability directly determines the reliability and operational safety of the power grid. During transformer operation, faults such as core loosening, winding deformation, and component aging can all cause abnormal vibration signals. These vibration signals contain a wealth of early fault characteristic information, serving as an important basis for fault detection. Existing transformer vibration fault detection technologies have significant shortcomings. For example, signal preprocessing often employs simple filtering methods, which are insufficient to effectively eliminate impulse noise. The positioning accuracy of abrupt signal changes is low, and critical fault information is easily lost. Mode decomposition often suffers from endpoint effects and mode aliasing problems. The number of modes is often manually preset, making it impossible to adaptively match signal spectral characteristics, resulting in poor reliability of decomposition results. These issues make it difficult to meet the high-precision fault detection requirements of transformers under complex operating conditions. Summary of the Invention

[0003] In view of the technical problems existing in the background art, the present invention proposes a method for detecting transformer operation faults.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] S1. Collect the vibration signal of the transformer and preprocess it. Divide the vibration signal into transformer vibration sub-sequences according to the preset time window.

[0006] S2. Perform improved variational mode decomposition on each transformer oscillation subsequence to obtain a preset number of eigenmode function components;

[0007] S3. Reconstruct the phase space for each intrinsic mode function component to obtain the reconstructed phase space trajectory of each intrinsic mode function component; construct a recursive graph based on the reconstructed phase space trajectory, extract recursive quantitative analysis features, obtain the modal dynamics features of each intrinsic mode function component, and obtain the modal dynamics feature vector of the transformer vibration subsequence based on the modal dynamics features of all intrinsic mode function components.

[0008] S4. Calculate the directional transfer entropy between the components of each intrinsic mode function to obtain the cross-modal information flow feature vector of the transformer oscillation subsequence;

[0009] S5. Obtain the dynamic time warping distance between the transformer oscillating subsequence and the corresponding intrinsic mode function components in the adjacent oscillating subsequences, and obtain the time series stability index of the transformer oscillating subsequence.

[0010] S6. Input the transformer oscillation subsequence, modal dynamics feature vector, cross-modal information flow feature vector, and time-series stability index into the transformer fault detection model, and output the transformer operation fault detection results from the transformer fault detection model.

[0011] Preferably, step S1, which involves acquiring and preprocessing the transformer's vibration signal and dividing it into transformer vibration sub-sequences according to a preset time window, is implemented as follows:

[0012] The collected vibration signals were processed using... The criteria identify and remove impulse noise points. The missing data points after removal are repaired by cubic spline interpolation of adjacent data points to obtain a preliminary denoised vibration signal.

[0013] The vibration signal was initially denoised by wavelet thresholding. The db4 wavelet basis was used for 5-level decomposition. The threshold of each level was dynamically calculated based on the energy distribution of the wavelet coefficients. The high-frequency coefficients were then soft-thresholded and the signal was reconstructed to obtain the pre-processed vibration signal.

[0014] For the preprocessed vibration signal, a sliding window with length D and step size D / 2 is used to calculate the short-time energy and short-time zero-crossing rate frame by frame to construct a dual-feature joint mutation criterion. When the increase in short-time energy exceeds 50% of the mean of the previous 10 frames within 3 consecutive frames, and the fluctuation of the short-time zero-crossing rate exceeds twice the standard deviation of the previous 10 frames, the frame is determined to be a mutation candidate point, and the coarse mutation localization is completed.

[0015] Based on the candidate mutation points, with two frames before and after the candidate point as reference segments, the cross-correlation coefficient between the reference segments and the current signal segment is calculated point by point, and the sampling point corresponding to the minimum value of the cross-correlation coefficient is taken as the precise mutation point; with the precise mutation point as the center, M sampling points are extended before and after it as the mutation boundary.

[0016] If a sudden change boundary is detected, the signal is divided into non-uniform initial segments using each sudden change boundary as a segmentation node. Then, each initial segment is uniformly windowed according to a preset basic window length to ensure that all subsequences have the same length. If no sudden change boundary is detected, uniform sliding windowing is directly performed using a preset basic time window. The transformer vibration subsequence is then output.

[0017] Preferably, the improved variational mode decomposition in step S2 is specifically implemented as follows:

[0018] The transformer vibration subsequence is preprocessed by symmetric mirror extension, extending 20% ​​of the original length at both ends of the sequence. The extended part is a mirror symmetric copy of 20% of the data at the beginning and end of the original sequence.

[0019] The extended signal is subjected to a Fast Fourier Transform to obtain the power spectrum. All significant peaks in the power spectrum are detected, and the frequency interval between adjacent peaks is calculated. When the frequency interval between adjacent peaks is greater than the preset minimum frequency resolution, it is determined to be an independent mode peak. The number of independent mode peaks is counted as the initial number of modes. Then, by calculating the modal aliasing index of the decomposition results under different K values, the modal aliasing index with the smallest value and not less than 1 is selected. The K value is used as the final optimal number of modes;

[0020] Initialize the penalty factor of all modes to a uniform baseline value; in each round of variational mode decomposition iteration, calculate the instantaneous frequency variance of each current mode, and dynamically adjust the penalty factor of the mode according to the ratio of the instantaneous frequency variance to the preset threshold.

[0021] A constrained variational problem of variational mode decomposition is constructed. An augmented Lagrange function is constructed by introducing Lagrange multipliers and independent penalty factors for each mode. The alternating direction multiplier method is used for iterative solution. When the energy change rate of all modes is less than the preset convergence threshold in two consecutive iterations, the iteration is stopped and the components of each intrinsic mode function are output.

[0022] The output intrinsic mode function components are filtered for validity. The cross-correlation coefficient between each intrinsic mode function component and the original vibrational subsequence is calculated. Mode components with cross-correlation coefficients less than a preset correlation threshold are removed, and valid mode components are retained as the final decomposition result.

[0023] Preferably, the method for reconstructing the phase space of each intrinsic mode function component in step S3 and obtaining the reconstructed phase space trajectory of each intrinsic mode function component is as follows:

[0024] The time delay of each intrinsic mode function component is determined by the mutual information method, and the time corresponding to the first minimum point of the mutual information function is taken as the time delay.

[0025] Determine the embedding dimension of each intrinsic mode function component; based on time delay. and embedding dimension For a sequence of eigenmode function components of length L Perform time delay embedding reconstruction to obtain the included The m-dimensional reconstructed phase space trajectory of n state vectors, where each state vector is represented as... .

[0026] Preferably, the method for constructing a recursive graph based on the reconstructed phase space trajectory and extracting recursive quantitative analysis features in step S3 is as follows:

[0027] A recursive graph is constructed based on the reconstructed phase space trajectory, and the recursion matrix of the recursive graph is... Defined as: ,in, For Heaviside step function, Let i and j be the i-th and j-th state vectors in the reconstructed phase space trajectory, respectively. The preset recursion threshold is determined by taking 15% of the standard deviation of the Euclidean distance between all state vectors in the reconstructed phase space trajectory as the preset recursion threshold.

[0028] Extract recursive quantitative analysis characteristic parameters from the recursion graph, including recursion rate (RR), deterministic DET, and average diagonal length. The diagonal length entropy (ENTR) and stacking property (LAM) are calculated as follows: recursion rate (RR) is the proportion of elements with a value of 1 in the recursion matrix; determinism (DET) is the proportion of recursion points forming a diagonal structure in the recursion graph; and average diagonal length is... is the mean length of all diagonal structures in the recursive graph; ENTR is the information entropy of the diagonal length distribution; LAM is the proportion of recursive points that constitute vertical line structures to non-diagonal recursive points in the recursive graph.

[0029] The eigenvector composed of recursion rate, determinism, average diagonal length, diagonal length entropy, and stacking property is used as the modal dynamics feature of the intrinsic mode function component.

[0030] Preferably, the method for calculating the directional transfer entropy between the intrinsic mode function components and obtaining the cross-modal information flow feature vector of the transformer oscillation subsequence in step S4 is as follows:

[0031] The amplitude sequence of each intrinsic mode function component is discretized by equal-interval binning to obtain the symbolic sequence of each intrinsic mode function component;

[0032] For each pair of intrinsic mode function components, calculate the directional transfer entropy from the p-th intrinsic mode function component to the q-th intrinsic mode function component. Defined as: ,in, Let b be the b-th order history vector of the symbolic sequence of the q-th eigenmode function component at time t. The symbolic sequence of the p-th intrinsic mode function component at time t The history vector of order b and The embedding delay order;

[0033] The directional transit entropy between all pairs of intrinsic mode function components is used to form the cross-modal information flow feature vector.

[0034] Preferably, step S5, which obtains the dynamic time warping distance between the transformer oscillating subsequence and the corresponding intrinsic mode function components in adjacent oscillating subsequences, and specifically implements the time-series stability index of the transformer oscillating subsequence, is as follows:

[0035] The six closest oscillation subsequences to the transformer oscillation subsequence in the time series are denoted as adjacent oscillation subsequences;

[0036] For each adjacent oscillating subsequence, perform dynamic time warping and alignment with the corresponding intrinsic mode function components in the current transformer oscillating subsequence, and calculate the dynamic time warping distance between the k-th intrinsic mode function components. ;

[0037] Calculate the energy percentage of each intrinsic mode function component in the current vibrational subsequence, where the energy percentage is the ratio of the energy of that mode function component to the total energy of all effective mode functions; using the energy percentage as a weight, apply the energy percentage to all corresponding eigenmode functions of the adjacent subsequence. By performing a weighted summation, the comprehensive dynamic time warp distance between the adjacent subsequence and the current vibrational subsequence is obtained;

[0038] Calculate the arithmetic mean of the comprehensive dynamic time warping distances of all adjacent oscillator subsequence sets, and use the reciprocal of this arithmetic mean as the temporal stability index of the current transformer oscillator subsequence.

[0039] Preferably, step S6, which involves inputting the transformer vibration subsequence, modal dynamics feature vector, cross-modal information flow feature vector, and temporal stability index into the transformer fault detection model, and having the transformer fault detection model output the transformer operation fault detection result, is specifically implemented as follows:

[0040] The transformer fault detection model includes a multi-scale time feature extraction module, a channel attention fusion module, and a classification output module;

[0041] The transformer vibration subsequence is input into the multi-scale temporal feature extraction module, which consists of multiple cascaded temporal causal convolutional blocks. Each temporal causal convolutional block contains three parallel branches. Each branch uses temporal causal convolutional kernels with dilation coefficients of 1, 2, and 4 for convolution operations. The convolutional outputs of each branch are concatenated, processed by batch normalization and ReLU activation function, and then added to the input of the layer through residual connection to output the feature representation of the layer. After multiple cascades, the multi-scale temporal features are output.

[0042] Multi-scale temporal features are concatenated with modal dynamics feature vectors, cross-modal information flow feature vectors, and temporal stability indices along the feature dimension and input into the channel attention fusion module. The channel attention fusion module performs global average pooling on the concatenated features and inputs the pooling result into two fully connected layers and a sigmoid activation function to generate attention weight vectors for each feature channel. The attention weight vectors are then multiplied with the concatenated features channel by channel to obtain the enhanced feature vectors.

[0043] The enhanced feature vector is input into the classification output module, which consists of two fully connected layers and a Softmax activation function, and outputs the transformer operation fault detection results.

[0044] Compared with existing technologies, the advantages and positive effects of this invention are as follows: Addressing the shortcomings of existing transformer vibration fault detection technologies, this invention proposes a high-precision detection scheme based on multi-dimensional feature fusion. In the preprocessing stage, precise impulse noise removal, adaptive wavelet denoising, and dual-feature joint abrupt change localization significantly improve signal quality and fault abrupt change point localization accuracy. Improved variational mode decomposition effectively solves the problems of endpoint effects and mode aliasing through mirror continuation, adaptive mode number, and dynamic penalty factors. Simultaneously, by integrating multi-dimensional features such as modal dynamics, cross-modal information flow, and temporal stability, combined with multi-scale causal convolution and channel attention models, the detection accuracy and robustness of early, weak faults under complex operating conditions are greatly improved. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating a method for detecting transformer operational faults. Detailed Implementation

[0047] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0048] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0049] This invention provides a method for detecting transformer operational faults, specifically implemented as follows: Figure 1 As shown.

[0050] First, the vibration signal of the transformer is collected and preprocessed. The vibration signal is then divided into transformer vibration sub-sequences according to a preset time window. This is achieved by first processing the collected vibration signal... The criteria identify and remove impulse noise points. Missing data points are then repaired using cubic spline interpolation of adjacent data points to obtain a preliminary denoised vibration signal. Specifically, the acquired raw vibration signal is segmented, with each segment containing 1024 sampling points. The arithmetic mean and standard deviation of each segment are calculated. For each sampling point within a segment, if the following conditions are met... If a sampling point is identified as an impulse noise point, it is removed. This method effectively removes spike noise with abnormal amplitude while preserving normal vibration signal components. For missing data points resulting from impulse noise removal, cubic spline interpolation of five adjacent data points is used for repair. Cubic spline interpolation constructs piecewise cubic polynomial functions between adjacent data points, ensuring that the interpolation curve has continuous first and second derivatives throughout the entire interval, thus smoothly fitting the changing trend of the vibration signal. Specifically, a cubic spline function is constructed using two valid data points before and after the missing point as interpolation nodes, and the interpolation result at the missing point is calculated. This method has higher accuracy than linear interpolation, better preserves the high-frequency details of the vibration signal, and avoids introducing additional noise. After repair, a preliminary denoised vibration signal is obtained, laying the foundation for subsequent wavelet threshold denoising.

[0051] Next, wavelet thresholding is applied to the initially denoised vibration signal. A 5-level decomposition is performed using the db4 wavelet basis. The threshold for each level is dynamically calculated based on the energy distribution of the wavelet coefficients. Soft thresholding is applied to the high-frequency coefficients before reconstructing the signal, resulting in the preprocessed vibration signal. Specifically, a 5-level decomposition of the initially denoised vibration signal is performed using the db4 wavelet basis. The db4 wavelet basis possesses tight support, orthogonality, and good time-frequency localization characteristics, making it very suitable for processing non-stationary transformer vibration signals. Wavelet decomposition divides the original signal into one low-frequency approximation coefficient and five high-frequency detail coefficients. The low-frequency coefficients contain the main energy and trend information of the signal, while the high-frequency coefficients contain the details and noise information. The threshold for each level is dynamically calculated based on the energy distribution of the wavelet coefficients. The threshold calculation formula is as follows: ,in Let be the standard deviation of the wavelet coefficients of the j-th layer. Let be the length of the wavelet coefficients in the j-th layer. This threshold calculation method adaptively adjusts the threshold size for each layer, setting a larger threshold for high-energy high-frequency layers and a smaller threshold for low-energy high-frequency layers, thereby effectively removing noise while retaining more useful signal features. Soft thresholding is applied to the high-frequency coefficients of each layer; the soft thresholding function is defined as... ,when Otherwise, y=0. Soft thresholding offers better continuity compared to hard thresholding, avoiding pseudo-Gibbs phenomenon in the reconstructed signal. After processing, the processed high-frequency and low-frequency coefficients are subjected to inverse wavelet transform to reconstruct the preprocessed vibration signal. This method effectively removes Gaussian white noise from the signal and improves the signal-to-noise ratio.

[0052] For the preprocessed vibration signal, a sliding window of length D and step size D / 2 is used to calculate the short-time energy and short-time zero-crossing rate frame by frame, constructing a dual-feature joint abrupt change criterion. When the increase in short-time energy exceeds 50% of the mean of the previous 10 frames within 3 consecutive frames, and the fluctuation of the short-time zero-crossing rate exceeds twice the standard deviation of the previous 10 frames, the frame is determined as a candidate point for abrupt change, completing the coarse localization of the abrupt change. Specifically, in this step, the sliding window processes the preprocessed vibration signal frame by frame, calculating the two feature parameters of short-time energy and short-time zero-crossing rate for each frame. Short-time energy reflects the energy magnitude of the signal within a frame, and the calculation formula is as follows: ,in Let m be the m-th sampling point of the n-th frame signal, and N be the frame length. The short-time zero-crossing rate reflects the number of times the signal crosses a zero point within a frame, and is calculated using the following formula: A dual-feature joint mutation criterion is constructed to monitor the changes in short-time energy and short-time zero-crossing rate of the signal in each frame in real time. A sliding window of length 10 frames is maintained, and the mean of the short-time energy within the window is calculated. and the standard deviation of the short-time zero crossing rate When the short-term energy increase within three consecutive frames exceeds 50%, and at the same time, the short-term zero-crossing rate fluctuation exceeds When the value is twice that of the target frame, the frame is considered a candidate for mutation.

[0053] Based on candidate mutation points, using two frames before and after each candidate point as reference segments, the cross-correlation coefficient between the reference segments and the current signal segment is calculated point by point. The sampling point corresponding to the minimum cross-correlation coefficient is taken as the precise mutation point. With the precise mutation point as the center, M sampling points are extended before and after it as mutation boundaries. If a mutation boundary is detected, the signal is divided into non-uniform initial segments using each mutation boundary as a segmentation node. Then, each initial segment is uniformly windowed according to a preset basic window length to ensure that all subsequences have the same length. If no mutation boundary is detected, uniform sliding windowing is directly performed using a preset basic time window. The transformer vibration subsequence is output.

[0054] Next, an improved variational mode decomposition is performed on each transformer oscillation subsequence to obtain a predetermined number of eigenmode function components. Specifically, the transformer oscillation subsequence undergoes symmetric mirror extension preprocessing to suppress endpoint effects during variational mode decomposition; let the original transformer oscillation subsequence be... n=0,1,2,...,N-1, where N is the sequence length, and the extended signal for:

[0055] That is, each end of the sequence is extended by 20% of the original length, and the extended part is a mirror symmetric copy of the first and last 20% of the original sequence data.

[0056] The optimal number of modes K is adaptively determined based on the spectral characteristics of the transformer vibration signal, i.e., for the extended signal. The power spectrum is obtained by performing a fast Fourier transform. Detect all amplitude values ​​in the power spectrum that are greater than a preset peak threshold. The local maximum point is denoted as: Sort the peaks by frequency from smallest to largest; calculate the frequency interval between adjacent peaks: ,when At that time, the judgment and For independent modal peak values, where To preset the minimum frequency resolution, the number of independent modal peaks is counted as the initial number of modes. ; Calculate the mode mixing index (MAI) of the decomposition results under different K values. The mode mixing index is defined as: ,in, The Fourier transforms of the i-th and j-th intrinsic mode function components are taken respectively, and the mode mixing exponent is selected with the smallest exponent and not less than 1. The K value is used as the final optimal number of modes.

[0057] Initialize the penalty factor for all modes to a uniform baseline value. In each iteration of variational mode decomposition, the k-th mode component is... The analytic signal is obtained by performing a Hilbert transform: Where k is the index of the intrinsic mode function (IMF) component, used to identify each independent vibrational mode component generated during variational mode decomposition, and to calculate its instantaneous frequency: , Calculate the instantaneous frequency variance: ,in, Let T be the average instantaneous frequency of the k-th mode, and T be the signal duration. The penalty factor for this mode is dynamically adjusted based on the instantaneous frequency variance. ,in, To preset the frequency variance threshold, To prevent extremely small constants with a denominator of zero; the larger the instantaneous frequency variance, the greater the modal frequency fluctuation, and the smaller the corresponding penalty factor, and vice versa.

[0058] An improved method of alternating direction multipliers is used for iterative solution: A constrained variational problem for variational mode decomposition is constructed. , Introducing Lagrange multipliers and modality-independent penalty factors Construct the augmented Lagrangian function: ,in, For the first-order partial derivative operator with respect to time t, For the Dirac function, This is the inner product operator.

[0059] An iterative solution is performed using the alternating direction multiplier method, updating sequentially in each iteration. The modal energy convergence criterion is introduced to replace the fixed iteration number termination condition. When the energy change rate of all modes in two consecutive iterations is less than the preset convergence threshold, the iteration stops and each intrinsic mode function component is output. The output intrinsic mode function components are then screened for effectiveness: the cross-correlation coefficient between each intrinsic mode function component and the original vibrational subsequence is calculated, noise mode components with cross-correlation coefficients less than the preset correlation threshold are removed, and the effective mode components are retained as the final decomposition result.

[0060] Then, phase space reconstruction is performed on each intrinsic mode function component to obtain the reconstructed phase space trajectory of each intrinsic mode function component; a recursive graph is constructed based on the reconstructed phase space trajectory, recursive quantitative analysis features are extracted, modal dynamic features of each intrinsic mode function component are obtained, and modal dynamic feature vector of transformer oscillator subsequence is obtained based on the modal dynamic features of all intrinsic mode function components.

[0061] The method for reconstructing the phase space of each intrinsic mode function (IMF) component and obtaining the reconstructed phase space trajectory of each IMF component is as follows: The time delay of each IMF component is determined using the mutual information method; the time corresponding to the first minimum point of the mutual information function is taken as the time delay; zero-mean normalization is performed on each IMF component sequence to eliminate the interference of amplitude fluctuations on the mutual information calculation; and the preset time delay value range is traversed. , To calculate the maximum value for each delay, the mutual information value is determined. The Sturges rule is used to determine the number of bins, and the original and delayed sequences are discretized into equally spaced bins. The number of samples within each bin is counted, and the marginal and joint probability distributions are calculated. These are then substituted into the mutual information formula to calculate the mutual information value for different delays. The mutual information value is used to plot the random number matrix based on the obtained mutual information value. For the changing curve, the neighborhood comparison method is used to find the first minimum point to determine the value. The optimal time delay for this IMF component is determined, and the embedding dimension *m* of each intrinsic mode function component is determined using the Cao method. Based on the time delay... and embedding dimension For a sequence of eigenmode function components of length L Perform time delay embedding reconstruction to obtain the included The m-dimensional reconstructed phase space trajectory of n state vectors, where each state vector is represented as... .

[0062] Furthermore, the method for extracting recursive quantitative analysis features based on the reconstructed phase space trajectory to build a recursive graph is as follows: The recursive graph is constructed based on the reconstructed phase space trajectory, and the recursion matrix of the recursive graph... Defined as: ,in, For Heaviside step function, Let i and j be the i-th and j-th state vectors in the reconstructed phase space trajectory, respectively. A preset recursion threshold is set; the preset recursion threshold is determined by taking 15% of the standard deviation of the Euclidean distance between all state vectors in the reconstructed phase space trajectory as the preset recursion threshold; recursive quantitative analysis feature parameters are extracted from the recursion graph, including recursion rate RR, deterministic DET, and average diagonal length. The diagonal length entropy (ENTR) and stacking property (LAM) are calculated as follows: recursion rate (RR) is the proportion of elements with a value of 1 in the recursion matrix; determinism (DET) is the proportion of recursion points forming a diagonal structure in the recursion graph; and average diagonal length is... The mean length of all diagonal structures in the recursive graph is given by ; the diagonal length entropy (ENTR) is the information entropy of the diagonal length distribution; the stacking property (LAM) is the proportion of recursive points forming vertical line structures in the recursive graph to the non-diagonal recursive points; the eigenvector composed of recursion rate, determinism, mean diagonal length, diagonal length entropy, and stacking property is used as the modal dynamics characteristic of this intrinsic modal function component.

[0063] In addition, the directional transfer entropy between each intrinsic mode function component is calculated to obtain the cross-modal information flow feature vector of the transformer oscillation subsequence. Specifically, the amplitude sequence of each intrinsic mode function component is discretized by equally spaced bins to obtain the symbolic sequence of each intrinsic mode function component; for each pair of intrinsic mode function components, the directional transfer entropy from the p-th intrinsic mode function component to the q-th intrinsic mode function component is calculated. Defined as: ,in, Let b be the b-th order history vector of the symbolic sequence of the q-th eigenmode function component at time t. The symbolic sequence of the p-th intrinsic mode function component at time t The history vector of order b and To embed the delay order; the directional transfer entropy between all pairs of intrinsic mode function components is used to form a cross-modal information flow feature vector.

[0064] Next, the dynamic time warping distance between the transformer oscillating subsequence and the corresponding intrinsic mode function components in adjacent oscillating subsequences is obtained to acquire the temporal stability index of the transformer oscillating subsequence. Specifically, the six oscillating subsequences that are closest to the transformer oscillating subsequence in time series are denoted as adjacent oscillating subsequences; for each adjacent oscillating subsequence, dynamic time warping is performed on its corresponding intrinsic mode function component in the current transformer oscillating subsequence, and the dynamic time warping distance between the k-th intrinsic mode function component is calculated. Calculate the energy percentage of each intrinsic mode function component in the current vibrational subsequence, where the energy percentage is the ratio of the energy of that mode function component to the total energy of all effective mode functions; using the energy percentage as a weight, apply the energy percentage to all corresponding eigenmode functions of the adjacent subsequence. A weighted summation is performed to obtain the comprehensive dynamic time warping distance between the adjacent subsequence and the current oscillating subsequence; the arithmetic mean of all comprehensive dynamic time warping distances in the set of adjacent oscillating subsequences is calculated, and the reciprocal of the arithmetic mean is used as the temporal stability index of the current transformer oscillating subsequence.

[0065] Finally, the transformer oscillation subsequence, modal dynamics feature vector, cross-modal information flow feature vector, and temporal stability index are input into the transformer fault detection model, which then outputs the transformer operation fault detection results. Specifically, the transformer fault detection model includes a multi-scale temporal feature extraction module, a channel attention fusion module, and a classification output module.

[0066] The transformer vibration subsequence is input into the multi-scale temporal feature extraction module, which consists of multiple cascaded temporal causal convolutional blocks. Each temporal causal convolutional block contains three parallel branches. Each branch uses temporal causal convolutional kernels with dilation coefficients of 1, 2, and 4 for convolution operations. The convolutional outputs of each branch are concatenated, processed by batch normalization and ReLU activation function, and added to the input of the layer through residual connection to output the feature representation of the layer. After multiple cascades, the multi-scale temporal features are output.

[0067] Multi-scale temporal features are concatenated with modal dynamics feature vectors, cross-modal information flow feature vectors, and temporal stability indices along the feature dimension and input into the channel attention fusion module. The channel attention fusion module performs global average pooling on the concatenated features, and inputs the pooling result into two fully connected layers and a sigmoid activation function to generate attention weight vectors for each feature channel. These attention weight vectors are then multiplied channel-by-channel with the concatenated features to obtain enhanced feature vectors. This process is equivalent to weighting each feature channel; a larger weight indicates that the feature of that channel is more important for fault detection. Through this adaptive fusion method, the model can automatically learn the importance of different feature types.

[0068] The enhanced feature vector is input into the classification output module, which consists of two fully connected layers and a Softmax activation function. The first fully connected layer uses the ReLU activation function to further perform nonlinear transformation and feature extraction on the enhanced features, mapping them to a 64-dimensional hidden feature space. The number of neurons in the second fully connected layer corresponds to the normal and fault operation states of the transformer. The output of the second fully connected layer is processed by the Softmax activation function, converting the output value into a probability distribution. The sum of the probabilities of all states is 1, and the state with the highest probability is taken as the final fault detection result, outputting the transformer operation fault detection result.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for detecting transformer operational faults, characterized in that, Includes the following steps: S1. Collect the vibration signal of the transformer and preprocess it. Divide the vibration signal into transformer vibration sub-sequences according to the preset time window. S2. Perform improved variational mode decomposition on each transformer oscillation subsequence to obtain a preset number of eigenmode function components; S3. Reconstruct the phase space for each intrinsic mode function component to obtain the reconstructed phase space trajectory of each intrinsic mode function component; Based on the reconstructed phase space trajectory, a recursive graph is constructed, recursive quantitative analysis features are extracted, and modal dynamic features of each intrinsic mode function component are obtained. Based on the modal dynamic features of all intrinsic mode function components, the modal dynamic feature vector of the transformer vibration subsequence is obtained. S4. Calculate the directional transfer entropy between the components of each intrinsic mode function to obtain the cross-modal information flow feature vector of the transformer oscillation subsequence; S5. Obtain the dynamic time warping distance between the transformer oscillating subsequence and the corresponding intrinsic mode function components in the adjacent oscillating subsequences, and obtain the time series stability index of the transformer oscillating subsequence. S6. Input the transformer oscillation subsequence, modal dynamics feature vector, cross-modal information flow feature vector, and time-series stability index into the transformer fault detection model, and output the transformer operation fault detection results from the transformer fault detection model.

2. The method for detecting transformer operation faults according to claim 1, characterized in that, The specific implementation of step S1, which involves acquiring the vibration signal of the transformer and preprocessing it, and dividing the vibration signal into transformer vibration sub-sequences according to a preset time window, is as follows: The collected vibration signals were processed using... The criteria identify and remove impulse noise points. The missing data points after removal are repaired by cubic spline interpolation of adjacent data points to obtain a preliminary denoised vibration signal. The vibration signal was initially denoised by wavelet thresholding. The db4 wavelet basis was used for 5-level decomposition. The threshold of each level was dynamically calculated based on the energy distribution of the wavelet coefficients. The high-frequency coefficients were then soft-thresholded and the signal was reconstructed to obtain the pre-processed vibration signal. For the preprocessed vibration signal, a sliding window with length D and step size D / 2 is used to calculate the short-time energy and short-time zero-crossing rate frame by frame to construct a dual-feature joint mutation criterion. When the increase in short-time energy exceeds 50% of the mean of the previous 10 frames within 3 consecutive frames, and the fluctuation of the short-time zero-crossing rate exceeds twice the standard deviation of the previous 10 frames, the frame is determined to be a mutation candidate point, and the coarse mutation localization is completed. Based on the candidate mutation points, with two frames before and after the candidate point as reference segments, the cross-correlation coefficient between the reference segments and the current signal segment is calculated point by point, and the sampling point corresponding to the minimum value of the cross-correlation coefficient is taken as the precise mutation point; with the precise mutation point as the center, M sampling points are extended before and after it as the mutation boundary. If a sudden change boundary is detected, the signal is divided into non-uniform initial segments using each sudden change boundary as a segmentation node. Then, each initial segment is uniformly windowed according to a preset basic window length to ensure that all subsequences have the same length. If no sudden change boundary is detected, uniform sliding windowing is directly performed using a preset basic time window. The transformer vibration subsequence is then output.

3. The method for detecting transformer operation faults according to claim 1, characterized in that, The specific implementation of the improved variational mode decomposition in step S2 is as follows: The transformer vibration subsequence is preprocessed by symmetric mirror extension, extending 20% ​​of the original length at both ends of the sequence. The extended part is a mirror symmetric copy of 20% of the data at the beginning and end of the original sequence. The extended signal is subjected to a Fast Fourier Transform to obtain the power spectrum. All significant peaks in the power spectrum are detected, and the frequency interval between adjacent peaks is calculated. When the frequency interval between adjacent peaks is greater than the preset minimum frequency resolution, it is determined to be an independent mode peak. The number of independent mode peaks is counted as the initial number of modes. Then, by calculating the modal aliasing index of the decomposition results under different K values, the modal aliasing index with the smallest value and not less than 1 is selected. The K value is used as the final optimal number of modes; Initialize the penalty factor of all modes to a uniform baseline value; in each round of variational mode decomposition iteration, calculate the instantaneous frequency variance of each current mode, and dynamically adjust the penalty factor of the mode according to the ratio of the instantaneous frequency variance to the preset threshold. A constrained variational problem of variational mode decomposition is constructed. An augmented Lagrange function is constructed by introducing Lagrange multipliers and independent penalty factors for each mode. The alternating direction multiplier method is used for iterative solution. When the energy change rate of all modes is less than the preset convergence threshold in two consecutive iterations, the iteration is stopped and the components of each intrinsic mode function are output. The output intrinsic mode function components are filtered for validity. The cross-correlation coefficient between each intrinsic mode function component and the original vibrational subsequence is calculated. Mode components with cross-correlation coefficients less than a preset correlation threshold are removed, and valid mode components are retained as the final decomposition result.

4. The method for detecting transformer operational faults according to claim 1, characterized in that, In step S3, the phase space is reconstructed for each intrinsic mode function component, and the method for obtaining the reconstructed phase space trajectory of each intrinsic mode function component is as follows: The time delay of each intrinsic mode function component is determined by the mutual information method, and the time corresponding to the first minimum point of the mutual information function is taken as the time delay. Determine the embedding dimension of each intrinsic mode function component; based on time delay. and embedding dimension For a sequence of eigenmode function components of length L Perform time delay embedding reconstruction to obtain the included The m-dimensional reconstructed phase space trajectory of n state vectors, where each state vector is represented as... .

5. The method for detecting transformer operational faults according to claim 1, characterized in that, The method for constructing a recursive graph based on the reconstructed phase space trajectory and extracting recursive quantitative analysis features in step S3 is as follows: A recursive graph is constructed based on the reconstructed phase space trajectory, and the recursion matrix of the recursive graph is... Defined as: ,in, For Heaviside step function, Let i and j be the i-th and j-th state vectors in the reconstructed phase space trajectory, respectively. The preset recursion threshold is determined by taking 15% of the standard deviation of the Euclidean distance between all state vectors in the reconstructed phase space trajectory as the preset recursion threshold. Extract recursive quantitative analysis characteristic parameters from the recursion graph, including recursion rate (RR), deterministic DET, and average diagonal length. The diagonal length entropy (ENTR) and stacking property (LAM) are calculated as follows: recursion rate (RR) is the proportion of elements with a value of 1 in the recursion matrix; determinism (DET) is the proportion of recursion points forming a diagonal structure in the recursion graph; and average diagonal length is... is the mean length of all diagonal structures in the recursive graph; ENTR is the information entropy of the diagonal length distribution; LAM is the proportion of recursive points that constitute vertical line structures to non-diagonal recursive points in the recursive graph. The eigenvector composed of recursion rate, determinism, average diagonal length, diagonal length entropy, and stacking property is used as the modal dynamics feature of the intrinsic mode function component.

6. The method for detecting transformer operational faults according to claim 1, characterized in that, The method for calculating the directional transfer entropy between each intrinsic mode function component in step S4 to obtain the cross-modal information flow feature vector of the transformer oscillation subsequence is as follows: The amplitude sequence of each intrinsic mode function component is discretized by equal-interval binning to obtain the symbolic sequence of each intrinsic mode function component; For each pair of intrinsic mode function components, calculate the directional transfer entropy from the p-th intrinsic mode function component to the q-th intrinsic mode function component. Defined as: ,in, Let b be the b-th order history vector of the symbolic sequence of the q-th eigenmode function component at time t. The symbolic sequence of the p-th intrinsic mode function component at time t The history vector of order b and The embedding delay order; The directional transit entropy between all pairs of intrinsic mode function components is used to form the cross-modal information flow feature vector.

7. The method for detecting transformer operational faults according to claim 1, characterized in that, Step S5, which obtains the dynamic time warping distance between the transformer oscillating subsequence and the corresponding intrinsic mode function components in adjacent oscillating subsequences, and specifically obtains the time-series stability index of the transformer oscillating subsequence, is implemented as follows: The six closest oscillation subsequences to the transformer oscillation subsequence in the time series are denoted as adjacent oscillation subsequences; For each adjacent oscillating subsequence, perform dynamic time warping and alignment with the corresponding intrinsic mode function components in the current transformer oscillating subsequence, and calculate the dynamic time warping distance between the k-th intrinsic mode function components. ; Calculate the energy percentage of each intrinsic mode function component in the current vibrational subsequence, where the energy percentage is the ratio of the energy of that mode component to the total energy of all effective mode components; Using energy percentage as weight, all corresponding to that adjacent subsequence By performing a weighted summation, the comprehensive dynamic time warp distance between the adjacent subsequence and the current vibrational subsequence is obtained; Calculate the arithmetic mean of the comprehensive dynamic time warping distances of all adjacent oscillator subsequence sets, and use the reciprocal of this arithmetic mean as the temporal stability index of the current transformer oscillator subsequence.

8. The method for detecting transformer operation faults according to claim 1, characterized in that, The specific implementation of step S6, which involves inputting the transformer vibration subsequence, modal dynamics feature vector, cross-modal information flow feature vector, and temporal stability index into the transformer fault detection model, and outputting the transformer operation fault detection result from the transformer fault detection model, is as follows: The transformer fault detection model includes a multi-scale time feature extraction module, a channel attention fusion module, and a classification output module. The transformer vibration subsequence is input into the multi-scale temporal feature extraction module, which consists of multiple cascaded temporal causal convolutional blocks. Each temporal causal convolutional block contains three parallel branches. Each branch uses temporal causal convolutional kernels with dilation coefficients of 1, 2, and 4 for convolution operations. The convolutional outputs of each branch are concatenated, processed by batch normalization and ReLU activation function, and then added to the input of the layer through residual connection to output the feature representation of the layer. After multiple cascades, the multi-scale temporal features are output. Multi-scale temporal features are concatenated with modal dynamics feature vectors, cross-modal information flow feature vectors, and temporal stability indices along the feature dimension and input into the channel attention fusion module. The channel attention fusion module performs global average pooling on the concatenated features and inputs the pooling result into two fully connected layers and a sigmoid activation function to generate attention weight vectors for each feature channel. The attention weight vectors are then multiplied with the concatenated features channel by channel to obtain the enhanced feature vectors. The enhanced feature vector is input into the classification output module, which consists of two fully connected layers and a Softmax activation function, and outputs the transformer operation fault detection results.

Images