On-load tap-changer fault diagnosis method based on TVFEMD noise reduction-chaotic feature fusion-NGBoost

Abstract

The invention discloses an on-load tap-changer fault diagnosis method based on TVFEMD noise reduction-chaotic feature fusion-NGBoost, relates to the technical field of on-load tap-changer fault diagnosis, and is used for solving the problems that complex vibration signals are difficult to effectively analyze and single feature diagnosis precision is insufficient. Self-adaptive positioning of a vibration signal at the action moment of a change-over switch is realized by adopting a short-time energy-entropy ratio algorithm; signal noise reduction is carried out on the vibration signal by utilizing time-varying filtering empirical mode decomposition TVFEMD; feature extraction is carried out based on chaotic characteristics; delay time is determined through a mutual information method; and inputting the vibration signal into an NGBoost multi-classification fault diagnosis model for type identification. According to the method, the problems of instability, nonlinearity and high noise of OLTC vibration signals can be effectively solved, feature extraction is stable and reliable, and the fault diagnosis accuracy is high.

Description

Technical Field

[0001] This invention relates to the field of on-load tap changer fault diagnosis technology, specifically a method for on-load tap changer fault diagnosis based on TVFEMD noise reduction-chaotic feature fusion-NGBoost. Background Technology

[0002] During the operation of an on-load tap changer (OLTC), various components generate significant transient mechanical vibrations during switching operations, which are crucial for distinguishing between normal and abnormal operating conditions. The vibration signals contain key event characteristics related to contact opening and closing, switch switching, and impacts in the transmission mechanism, enabling early detection of potential faults such as loosening, wear, and misalignment in the mechanical structure. However, OLTC vibration signals typically exhibit non-stationarity, nonlinearity, and high noise characteristics, with significant differences between signal segments, short effective event durations, and complex superposition of impact components, posing a significant challenge to feature extraction. Extracting stable, sensitive, and repeatable features from complex vibration data is fundamental to achieving highly reliable fault diagnosis.

[0003] In existing technologies, traditional time-domain analysis methods characterize signal amplitude changes through mean, variance, peak value, and kurtosis, offering the advantage of simple implementation. However, they only reflect overall waveform characteristics and are not sensitive enough to typical transient shocks during OLTC switching, exhibiting poor stability in high-noise environments. Frequency-domain analysis methods, based on Fourier transform, can reflect the frequency components of vibration signals, but Fourier analysis assumes signal stationarity, making it difficult to accurately characterize short-term energy mutations during contact switching and insufficiently describing non-stationary behavior. Wavelet transform achieves multi-scale analysis through adjustable window functions, revealing instantaneous characteristics, but its performance is highly dependent on wavelet basis selection, lacking universal basis function selection criteria, and different basis functions significantly affect the decomposition results. EMD and its improved algorithms can adaptively decompose vibration signals into several intrinsic mode functions, suitable for processing non-stationary signals, but they are prone to mode aliasing and endpoint effects in high-noise environments, leading to insufficient decomposition stability. While VMD improves noise immunity through variational optimization, it requires pre-setting the number of decomposition levels and bandwidth parameters, making it difficult to adapt to different types of OLTC structures or diverse operating conditions. HHT, a combination of EMD and Hilbert transform, can obtain instantaneous frequency and energy spectrum that vary with time, exhibiting good time-frequency focusing capabilities. However, it is highly dependent on the quality of decomposition and still has certain limitations in complex vibration environments. Besides time-frequency analysis methods, there are also analyses of OLTC vibration signals from a nonlinear dynamics perspective. However, these methods often have low dimensionality, weak physical interpretability, and are easily affected by noise, making it difficult to directly form stable diagnostic criteria. Therefore, a single type of feature is often insufficient to comprehensively describe the complex mechanical dynamic processes of OLTCs.

[0004] In summary, as the core voltage regulating mechanism of power transformers, the on-load tap changer (OLTC) directly threatens the stability of the power system due to mechanical faults caused by its high-frequency operation, and vibration signals are a key basis for its condition diagnosis. However, OLTC vibration signals have complex characteristics of non-stationarity, nonlinearity, and high noise. Traditional time-domain and frequency-domain analysis methods are difficult to accurately capture transient impact characteristics. Existing time-frequency analysis methods such as wavelet transform, EMD, and VMD have limitations such as basis function dependence, difficulty in parameter determination, or mode aliasing. Chaotic feature analysis from the perspective of nonlinear dynamics faces problems such as low dimensionality, weak physical interpretability, and insufficient noise resistance. Single-type features cannot fully characterize the complex mechanical dynamic process of OLTC. Therefore, it is urgent to develop a vibration signal feature extraction and fault diagnosis scheme that integrates the advantages of multiple dimensions and has strong robustness to provide technical support for high-precision condition monitoring of OLTC. Summary of the Invention

[0005] The purpose of this invention is to provide an on-load tap changer fault diagnosis method based on TVFEMD noise reduction, chaotic feature fusion, and NGBoost, which can solve the problems of difficult effective analysis of complex vibration signals of OLTC and insufficient accuracy of single feature diagnosis.

[0006] The technical solution adopted by the present invention to solve its technical problem is: a fault diagnosis method for on-load tap changers based on TVFEMD noise reduction-chaotic feature fusion-NGBoost, which includes the following steps.

[0007] S1. The short-time energy entropy ratio algorithm is used to adaptively locate the vibration signal at the moment of switching action. By setting the frame length, frame shift and threshold parameters, the signal segment containing key fault information is accurately extracted, and the extraction duration is uniformly extracted to ensure timing consistency.

[0008] S2. Time-varying filtering empirical mode decomposition (TVFEMD) is used to denoise the vibration signal. The signal is reconstructed by rearranging the local cutoff frequencies and selecting the modal components based on the energy ratio.

[0009] S3. Based on the chaotic characteristics, feature extraction is carried out. KS entropy and average origin distance (AOD) are extracted as chaotic features and geometric features. At the same time, the phase space trajectory is mapped into a recursive graph. After grayscale preprocessing, three types of Tamura texture features, namely roughness, contrast and orientation, are extracted to form a 5-dimensional chaotic feature matrix.

[0010] S4. Construct an NGBoost multi-class fault diagnosis model, using the feature matrix as input, to identify the mechanical condition of the OLTC.

[0011] Furthermore, in step S1, the short-time energy entropy ratio algorithm process includes the following steps: S1.1, for a length of... time series Amplitude normalization is performed. ;right Apply windowing function Frame processing yields the first Frame vibration signal is , (1); where, Represents the numerical value of a frame; This is the frame shift length; Frame length; S1.2 represents the total number of frames after the signal is framed; Perform a Fourier transform to obtain the spectrum of the signal in each frame. Spectral line frequency components The energy spectrum is , No. Frequency components of the frame signal after Fourier transform Corresponding probability density function (2); No. Short-time spectral entropy of frame signal (3); No. Energy of frame signal (4); where, It is a constant; The amplitude of the nth sampling point of the i-th frame signal; The energy entropy ratio of the frame signal (5); S1.3, Calculate the normalized short-time energy entropy ratio (6); where R is a sequence of short-time energy entropy ratios of all frames; min(R) is the minimum value in the whole sequence; max(R) is the maximum value in the whole sequence.

[0012] Further, in step S1, the adaptive localization step of the vibration signal includes: S1.4, selecting a window function and setting the window length and frame length, performing frame-by-frame processing on the vibration signal, and calculating the short-time energy entropy ratio R′ of each frame; S1.5, selecting a time threshold. and short-time energy entropy ratio threshold S1.6. Compare the short-time energy entropy ratio R′ of each frame, when And the signal exceeds the frame duration of the short-time entropy ratio threshold. When this frame is defined as the start position of the signal; S1.7, when When the signal is less than the short-time entropy ratio threshold, the frame duration is greater than... When this happens, the frame is defined as the end position of the signal.

[0013] Furthermore, in step S2, TVFEMD includes two processes: S2.1, local cutoff frequency rearrangement, and S2.2, time-varying filter-based screening.

[0014] Furthermore, in step S2, the specific steps for rearranging the local cutoff frequencies include: S2.1.1, for the real-valued signal... The analytic signal is obtained through Hilbert transform. (7); where, Let be the vibration amplitude at the i-th sampling point; , and All are amplitude values. , and All are phases; (8); j is the imaginary unit; S2.1.2, when hour, It is a local maximum; hour, It is a local minimum; for All local maxima Minimum point Interpolation was performed separately to obtain the curves. and Furthermore, we obtain: (9); S2.1.3, on All local maxima Interpolation is performed to obtain ,right All local minima Interpolation is performed to obtain and order (10); among which for The derivative of for The derivative of is used to define the local cutoff frequency. (11); S2.1.4, Define signal The time series of the maximum points is , (12); where ρ=0.25; if formula (12) holds, then Let be a point of discontinuity, let , Represents a sequence of discontinuities; if Then it is called for The rising edge, Consider it as the minimum value; if Then it is called for The falling edge, Consider it as the minimum value; The remaining portion is considered as the peak value; S2.1.5, interpolation is performed between all peak values ​​to obtain the rearranged local cutoff frequency. .

[0015] Further, in step S2, the filtering step based on time-varying filtering includes: S2.2.1, reconstructing the signal according to the rearranged cutoff frequency: (13); will The extreme points are taken as nodes, Divide the function into n segments, each with a step size of m; where n is the order of the B-spline function; perform B-spline interpolation approximation between the extreme points, and denote the approximation result as... This represents the local mean curve; S2.2.2, Determine the residual signal. Does it meet the stopping criteria? ;in, Given a bandwidth threshold; if satisfied, then Let IMF be the modal component obtained from TVFEMD decomposition; if not satisfied, let In the formula, x (k) (t) represents the signal in the k-th screening iteration, m (k) (t) is the local mean curve for the k-th iteration, until the stopping criterion is met; (14); where, For Loughlin instantaneous bandwidth, (15); For weighted average instantaneous frequency, (16); S2.2.3, Filter the modal components IMF, assuming that the K modal components of the signal obtained by TVFEMD decomposition are ;in, The energy of each modal component is defined as: (16); among which, Signal length; energy percentage of the modes (18); Cumulative energy percentage (19); Based on the set threshold, select the cumulative energy percentage that meets the requirements. The first m modes with values ​​less than the threshold are reconstructed to obtain the denoised signal. (20).

[0016] Furthermore, in step S3, the step of reconstructing the phase space based on chaotic characteristics is as follows: the OLTC vibration signal is considered as a set of signals with a length of... time series , (21); where, This is the starting time point of the vibration signal; For time intervals; divide the time series Reconstructed indivual A dimensional vector, the reconstructed phase space is represented as (twenty two); (23); where, To reconstruct the phase point coordinates in the phase space; To delay time, The embedding dimension.

[0017] Furthermore, in step S3, the mutual information method is selected to determine the delay time. Original vibration signal After a delay time τ, it becomes The mutual information value between S and Q (24); where, (25) The information entropy of signal S; The information entropy of signal Q; Let S be the joint information entropy of Q. The median value of the sampled points is The probability of a point appearing within the entire data segment; The median value of the sampled points is The probability of a point appearing within the entire data segment; The sampled values ​​in the original signal And the delayed signal sampling value is The probability of taking mutual information; The time corresponding to the first local minimum point is taken as the delay time τ.

[0018] Furthermore, in step S3, the CAO algorithm is used to calculate the embedding dimension m: for the first... Reconstructed vectors When the embedding dimension is m, it is represented as , To and nearest neighbor, (26); Define parameters (27); where, Represents Euclidean distance; This is the i-th reconstructed vector when the embedding dimension is m+1; To achieve the same result as when the embedding dimension is m+1 The nearest point; define parameters average (28) (29); through Parameters from arrive Changes over time; when Greater than a certain value hour, If change stops, then This is the minimum embedding dimension m that we are ultimately looking for.

[0019] Further, in step S3, a reference quantity is defined. (30); (31);Use The following steps were taken to improve the CAO algorithm: First, based on Determining the threshold based on fluctuations The second step is to calculate. mean and absolute value of deviation , (32); (33); The third step is to compare the absolute values ​​of the deviations. With threshold Size, if the maximum absolute value of the deviation Then the first one satisfies corresponding That is, the embedding dimension ;like Then find the minimum value of the absolute value of the deviation. middle Corresponding subscript ,by Repeat steps one and two for the new starting point until an embedding dimension m that satisfies an acceptable level of bias is found.

[0020] Furthermore, in step S3, the KS entropy is calculated using the correlation integral method, and the calculation formula is as follows: (34); where K is the value of KS entropy; For the first The correlation integral function of dimension; Let K be the correlation integral function in the (m+1)th dimension; choose the stable K value as the estimate of the KS entropy; define the mean of the sum of the distances from each state point in phase space to the origin of the phase space coordinate system as the mean origin moment AOD, calculated by the following formula: (35); where, The number of state points. These are the coordinates of a point in phase space.

[0021] Furthermore, in step S3, the method for mapping the phase space trajectory to a recursive graph is as follows: construct a recursive matrix based on the reconstructed signal. , (36); (37); where, , The number of vectors after the phase space is reconstructed; For the Heaviside function; It is an L2 norm; The recursive threshold constant is given in advance; combined with the recursion matrix ,by x-axis Create a 2D raster chart for the vertical axis; if In a two-dimensional dot matrix diagram Add a black dot at the location; if If the point is not drawn on the graph, then a point is obtained. A two-dimensional recursive graph.

[0022] Furthermore, in step S3, the grayscale preprocessing method is as follows: scan the entire image using a small matrix, and perform weighted calculations on the region surrounding each pixel to obtain the local black dot density. , (38); where, These are the weights of the convolution kernel; It is a recursive matrix; The kernel size is [size missing]. The image region density values ​​are normalized and mapped to a 5-bit grayscale range, with values ​​in the range [0, 32]. The normalization calculation formula is as follows: (39), where min(d) is The minimum value, max(d) is The maximum value.

[0023] Furthermore, in step S3, the roughness extraction method is as follows: First, calculate the roughness contained in the image. Average gray value of pixels in the active window (pixels) (40); where, Indicates that it is located at The pixel grayscale value at that location; Then, the average intensity difference between adjacent non-overlapping windows in the horizontal direction for each pixel is calculated. and the average intensity difference between adjacent non-overlapping windows in the vertical direction. . (41) The optimal activity window is , at this time Substitute the value into equation (41). The value reaches its maximum; finally, in the entire image The average value is the roughness. The calculation formula is: (42).

[0024] Further, in step S3, the contrast ratio calculation formula is as follows: (43); where, ; It is a fourth-order moment; It is variance.

[0025] Furthermore, in step S3, the orientation calculation process is as follows: First, calculate the gradient vector at each pixel. , (44); where, (45), Direction angle (46); where, This represents the amount of change in the gradient vector in the horizontal direction; This represents the change in the gradient vector in the vertical direction; then, a histogram of the vector direction is constructed. , (47); where, It is a quantification level of direction and angle; It corresponds to the direction angle. The total number of edge pixels on the surface; For when , Number of pixels at time The threshold for the magnitude of the gradient vector; finally, the directionality. (48); where, p is a histogram The peak value; This represents the number of peaks in the histogram. For quantification range; for The quantization value in the maximum histogram.

[0026] Furthermore, in step S4, the method for fault type identification using the NGBoost multi-class fault diagnosis model is as follows: for the sample feature vector The corresponding class probability vector It satisfies the probability normalization constraint: (49); Category probability (50); where, This is the output function for the model corresponding to the k-th class; Let J be the model output function corresponding to class j; the model is trained using the negative log-likelihood function as the loss function, expressed as: (51); Introduce a natural gradient optimization mechanism during parameter update: (52); among which, This represents the ordinary gradient. The Fisher information matrix is ​​used; in the model implementation, features are first extracted from the collected vibration signals to construct a fixed-length feature vector for the samples. (53) is used as input to the NGBoost multi-class classification model and is standardized before being input into the model; the model uses a regression tree as the base learner and gradually approximates the class probability distribution through gradient boosting, and the parameters are updated as follows: (54); where, For learning rate, This is a regression tree model trained based on natural gradients; after training, the class with the highest probability value is selected as the final predicted class. (55).

[0027] The beneficial effects of this invention are as follows: It achieves adaptive localization of key signal segments through short-time entropy ratio, solving the problem of inconsistent signal timing; it effectively reduces noise using the TVFEMD algorithm while preserving core mechanical impact features; it extracts KS entropy, AOD, and Tamura texture features based on chaos theory, constructing a feature system with both anti-interference and discriminative power; and finally, it achieves high-precision fault identification under multiple operating conditions through the NGBoost model. This invention effectively handles the non-stationarity, nonlinearity, and high noise problems of OLTC vibration signals, with stable and reliable feature extraction and high fault diagnosis accuracy. It provides a new technical path for OLTC online condition monitoring and has significant engineering application value for improving transformer operation and maintenance levels and reducing power outage risks in power systems. Further development can expand the operating condition coverage, optimize feature extraction efficiency and model generalization ability, and promote the large-scale application of the technology in actual power grid scenarios. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the vibration signal waveform of an on-load tap changer.

[0029] Figure 2 The waveform of the vibration signal extracted at the moment of operation of the OLTC switching device under normal working conditions;

[0030] Figure 3 The vibration signal waveform was extracted at the moment of operation of the OLTC switching device under the condition of loose contacts and arc plate.

[0031] Figure 4 The vibration signal waveform was extracted at the moment the OLTC switch was activated under the condition of loose drive shaft screws.

[0032] Figure 5 The vibration signal waveform was extracted at the moment of operation of the OLTC switching device under the condition of gear jamming.

[0033] Figure 6 Phase space reconstruction curves of vibration signals from switching devices under different operating conditions;

[0034] Figure 7 Comparison results of eigenvalues ​​under different working conditions;

[0035] Figure 8 A recursive diagram mapping vibration signals of a switching device under different operating conditions;

[0036] Figure 9Gray-scale recursive diagram of vibration signal of switch under different operating conditions;

[0037] Figure 10 One-dimensional distribution map of Tamura texture features under different working conditions. Detailed Implementation

[0038] This invention employs a signal localization-denoising-feature extraction-fault identification technical solution. The steps include: S1. Using a short-time energy entropy ratio algorithm to adaptively locate the vibration signal at the moment of switching action. By setting frame length, frame shift, and threshold parameters, signal segments containing key fault information are accurately extracted, and the extraction duration is uniformly maintained to ensure temporal consistency. S2. Time-varying filtering empirical mode decomposition (TVFEMD) is used to denoise the vibration signal. Local cutoff frequency rearrangement solves the mode aliasing problem, and the signal is reconstructed by filtering modal components based on energy proportion. The denoising effect is verified to be good by SNR and RMSE. S3. Feature extraction is performed based on chaotic characteristics. The delay time is determined by mutual information, and the embedding dimension is optimized by improving the CAO algorithm to complete phase space reconstruction. KS entropy and average origin distance (AOD) are then extracted as chaotic and geometric features. Simultaneously, the phase space trajectory is mapped to a recursive graph. After grayscale preprocessing, three types of Tamura texture features—roughness, contrast, and directionality—are extracted to form a 5-dimensional chaotic feature matrix. S4. An NGBoost multi-class fault diagnosis model is constructed. Using the feature matrix as input, the model stability is improved through natural gradient optimization. In sample tests under four operating conditions, the overall classification accuracy reaches 96.875%, achieving efficient and accurate identification of the mechanical condition of the OLTC. The four operating conditions refer to normal, loose contacts and arc plates, loose drive shaft screws, and jammed drive gears. The operation steps of this invention are described in detail below with reference to the accompanying drawings.

[0039] S1, adaptive positioning based on vibration signal at the moment the switch is activated.

[0040] The acquisition of vibration signals from an on-load tap changer (OLTC) typically begins with the start of the drive motor and continues until it stops. This process involves multiple operational phases, such as... Figures 1 to 5 As shown, the acquired vibration signals exhibit significant non-stationary characteristics. Among them, the vibration signals during the switching operation phase contain the most critical mechanical state information and are the primary research object for fault analysis and diagnosis. This invention uses short-time energy entropy ratio to locate and extract the vibration signals at the moment of OLTC switching operation.

[0041] The short-time energy entropy ratio calculation process is as follows: S1.1, for a length of... Vibration signal time series Amplitude normalization is performed. ,right Apply windowing function Frame processing yields the first Frame vibration signal is , (1). In the formula, Represents the numerical value of a frame; Frame length; This is the frame shift length; This represents the total number of frames after the signal is framed. S1.2, For a single frame of data... Perform a Fourier transform to obtain the spectrum of the signal in each frame. Spectral line frequency components The energy spectrum is Then the first The first frame signal after Fourier transform Frequency components Corresponding probability density function (2). No. Short-time spectral entropy of frame signal (3). No. Energy of frame signal (4). In the formula, This is a constant used to adjust the degree of energy change and helps distinguish signal abrupt changes. i (n) represents the amplitude of the nth sampling point of the i-th frame signal; The energy entropy ratio of the frame signal (5). S1.3 Calculate the normalized short-time energy entropy ratio (6). In the formula, R is a sequence of short-time energy entropy ratios of all frames; min(R) is the minimum value in the whole sequence; max(R) is the maximum value in the whole sequence.

[0042] The vibration signal at the moment of OLTC switching action is separated by calculating the short-time energy entropy ratio of the OLTC vibration waveform. The specific steps are as follows: S1.4, Select an appropriate window function and set an appropriate window length and frame length to perform frame processing on the vibration signal, and calculate the short-time energy entropy ratio R′ of each frame. S1.5, Select an appropriate time threshold. and short-time energy entropy ratio threshold S1.6. Compare the short-time energy entropy ratio R′ of each frame. And the signal exceeds the frame duration of the short-time entropy ratio threshold by more than [a certain number of frames]. When this frame is defined, the starting position of the vibration signal is determined. S1.7, when At that time, and the signal is less than the frame duration of the short-time entropy ratio threshold is greater than... When this happens, the frame is defined as the end position of the vibration signal.

[0043] This invention selects a frame length of 10ms and a frame shift of 5ms, and converts it into points according to the sampling frequency. Three frames of data are continuously compared, i.e., when... Furthermore, when the signal duration exceeds 15ms, the current frame is determined as the signal start position, and the end position is determined similarly. Based on the above parameter settings, the vibration signal at the moment of contact collision of the switching switch is extracted from the vibration signal of the entire on-load tap changer operation process under four operating conditions. The start time of OLTC switching switch operation varies under different operating conditions, but the length of the vibration signal at the moment of switching switch operation extracted is relatively similar. To ensure the consistency of the timing of vibration signals under different operating conditions in subsequent analysis, this invention extracts 0.15s of data backward from the start time of switching switch operation located by the short-time energy entropy ratio for each operating condition. Thus, the waveforms of the extracted vibration signals under different operating conditions are synchronized in the time domain, which can effectively reduce the error caused by human factors and reduce the impact of human intervention on subsequent feature extraction and fault diagnosis results. At the same time, the subsequent fault diagnosis model training requires a large amount of data, and the adaptive localization extraction method can significantly improve data processing efficiency.

[0044] S2, Signal noise reduction based on TVFEMD.

[0045] To effectively suppress high-frequency noise in OLTC vibration signals while preserving mechanical impact characteristics, a vibration signal denoising method based on time-varying empirical mode decomposition (TVFEMD) is proposed. TVFEMD essentially constructs a low-pass filter with a cutoff frequency that varies with time to perform iterative mean removal during EMD decomposition. It also uses a local narrowband signal to replace the intrinsic mode function as the iteration stopping condition, thus addressing the intermittency and mode aliasing problems that occur when processing nonlinear and non-stationary signals.

[0046] TVFEMD consists of two processes: S2.1, local cutoff frequency rearrangement, and S2.2, time-varying filter-based screening.

[0047] Local cutoff frequency rearrangement is used to eliminate mode aliasing. The specific steps are as follows: S2.1.1, for any real-valued signal... Analytical signal obtained through Hilbert transform Represented as a complex two-component signal: (7). Among them , and Represents amplitude. , and Represents phase, and (8), Let be the vibration amplitude at the i-th sampling point. S2.1.2, when... hour, It is a local maximum; hour, This is a local minimum. For All local maxima Minimum point Interpolation was performed separately to obtain the curves. and Furthermore, we obtain: (9). S2.1.3, regarding All local maxima Minimum point Interpolation was performed separately to obtain and and order (10). In the formula, for The derivative of for The derivative of is used to define the local cutoff frequency. (11). S2.1.4, Define the signal The time series of the maximum points is , (12). In the formula, ρ=0.25. If formula (12) holds, then Let be a point of discontinuity, let , This represents a sequence of discontinuities. If... Then it is called for The rising edge, Consider it as the minimum value; if Then it is called for The falling edge, Consider it as the minimum value. The remaining portion is considered as the peak value. S2.1.5. Interpolation is performed between all peak values ​​to obtain the rearranged local cutoff frequency. .

[0048] The filtering process based on time-varying filtering includes the following steps: S2.2.1, reconstructing the signal based on the rearranged cutoff frequency: (13). [The following appears to be a separate, unrelated sentence:] Will The extreme points are taken as nodes, Divide the function into n segments, each with a step size of m, where n is the order of the B-spline function. Perform B-spline interpolation approximation between the extreme points, and denote the approximation result as... This represents the local mean curve. S2.2.2, Determining the residual signal. Does it meet the stopping criteria? ;in, Given a bandwidth threshold. If satisfied, then Let IMF be the modal component obtained from TVFEMD decomposition; if not satisfied, let Repeat steps S2.2.1 and S2.2.2 above until the stopping criterion is met. Residual signal The calculation formula is: (14); where, For Loughlin instantaneous bandwidth, (15); For weighted average instantaneous frequency, (16).

[0049] To effectively remove noise from the vibration signal while preserving the mechanical impact characteristics of OLTC, the modal components (IMFs) obtained from TVFEMD decomposition are screened. Specific steps include: (1) Let the k modal components of the signal obtained from TVFEMD decomposition be... , Energy of each modal component (17); where, Signal length. Mode energy percentage. (18). Cumulative energy percentage (19).

[0050] Based on the set threshold, select the cumulative energy percentage that meets the requirements. The first m modes with values ​​less than the threshold are reconstructed to obtain the denoised signal. (20). The signal-to-noise ratio (SNR) and root mean square error (RMSE) of the denoised signal were evaluated, and the denoising effect was good.

[0051] S3. Extraction of chaotic features from vibration signals.

[0052] Starting from the chaotic characteristics of vibration signals, high-dimensional phase space reconstruction is performed on vibration signals under various operating conditions, and relevant chaotic and geometric features are extracted from the reconstructed chaotic attractors. Simultaneously, the reconstructed phase space curves are converted into two-dimensional recursive graphs, and Tamura texture features are extracted from the recursive graphs to depict the intrinsic variation laws of the dynamic system. Compared with traditional time-frequency domain features, chaotic features have stronger anti-interference capabilities and are less susceptible to interference from external factors such as noise, thus fully reflecting the mechanical characteristics of the OLTC under various operating conditions.

[0053] S3.1. Starting from the chaotic characteristics of vibration signals, high-dimensional phase space reconstruction is performed on vibration signals under various working conditions.

[0054] Specifically, the steps for high-dimensional phase space reconstruction of the vibration signal are as follows: The OLTC vibration signal is considered as a set of signals with a length of... time series , (21). Among them, This is the starting time point of the vibration signal; Time interval; time series Reconstructed indivual The reconstructed phase space is a dimensional vector. (twenty two); (23). Among them, To reconstruct the phase point coordinates in the phase space, The vectors together form the phase space of the reconstructed signal; To delay time, Let m be the embedding dimension. According to the embedding theorem, for a noise-free and infinitely long ideal time series, the embedding dimension m and the delay time are... The embedding dimension *m* and the delay time can be chosen arbitrarily without affecting the reconstruction quality. However, in actual working conditions, because the acquired vibration signal has a finite length and contains a lot of noise, the embedding dimension *m* and the delay time... The choice of embedding dimension m and delay time has a significant impact on the quality of the reconstructed signal. Appropriate selection of these factors is crucial. This is the foundation for phase space reconstruction.

[0055] Specifically, delay time The correlation between the coordinate components of each phase point after reconstruction was characterized, and the mutual information method was selected to determine the delay time. Assuming the original vibration signal After a delay time τ, the signal becomes The mutual information value between sequences S and Q is... (twenty four); (25). In the formula: The information entropy of signal S; The information entropy of signal Q; Let S be the joint information entropy of Q. The median value of the sampled points is The probability of a point appearing within the entire data segment; The median value of the sampled points is The probability of a point appearing within the entire data segment; The sampled values ​​in the original signal And the delayed signal sampling value is The probability of that. Minimum. This approach achieves minimal redundancy and allows for the reconstruction of the best attractor. Therefore, mutual information is considered when searching for the optimal delay parameter. The first local minimum point generated is taken as the optimal choice.

[0056] Specifically, the CAO algorithm is used to calculate the embedding dimension m: for the Reconstructed vectors When the embedding dimension is m, it is represented as , To and nearest neighbor, (26); Define parameters (27). Here, Represents Euclidean distance; This represents the i-th reconstructed vector; To achieve the same result as when the embedding dimension is m+1 The nearest point. Define parameters. average (28). At this time, the parameter The size depends only on the embedding dimension With delay time ,definition (29). Through Parameters can be obtained from arrive Changes over time. When Greater than a certain value hour, If change stops, then This is the minimum embedding dimension m that we are ultimately looking for.

[0057] When time series data is completely unordered, as the embedding dimension increases... The increase, indicators It is difficult to reach a true saturation value. When When it is already quite large, it is difficult to judge. Is it still growing slowly or has it stabilized? Due to the limited sample size of observable data, even if the sequence itself is random, In some dimensions, there may be situations where change appears to have stopped. To address this issue, the present invention adds a reference value. , (30); where, (31). Regarding the reference quantity The following steps are added to improve the CAO algorithm, including: (1) According to Determining the threshold based on fluctuations (2) Calculate mean and absolute value of deviation : (32); (33). (3) Compare the absolute values ​​of the deviations. With threshold Size. If the absolute value of the deviation maximum value Then the first one satisfies corresponding That is, the embedding dimension ;like Then find the absolute value of the deviation. minimum value middle Corresponding subscript ,by Repeat steps (2)-(3) for the new starting point until an embedding dimension m that satisfies an acceptable level of bias is found.

[0058] Figure 6 Phase space reconstruction curves of OLTC switch vibration signals under different operating conditions are shown. The phase space trajectory shapes of the OLTC switch vibration signals differ significantly under different operating conditions. Under normal operating conditions, the phase trajectory distribution is relatively concentrated and the overall shape is small. When a fault occurs, the phase space trajectory diagram is larger and the phase trajectory points are more dispersed. Therefore, the phase trajectory characteristics of the vibration signals can effectively identify mechanical defects in the OLTC.

[0059] S3.2 After the phase space reconstruction process is completed, the signal is represented in the form of reconstructed points in the high-dimensional space. To describe the differences in phase space reconstruction curves under different operating conditions, KS entropy is introduced to describe the chaotic characteristics of the reconstructed signal, and the average origin distance (AOD) is introduced to describe the geometric characteristics of the reconstructed signal.

[0060] Specifically, the KS entropy is solved using the correlation integral method, and the calculation formula is as follows: (34). In the formula, For the first The correlation integral function of dimension, Let be the correlation integral function of the (m+1)th dimension. The selection is no longer random. The stable value of K, which changes with the time of change, is used as an estimate of the KS entropy. If the KS entropy of a dynamical system is positive, it means that the uncertainty of the system will continuously increase with the evolution of time, which is an important characteristic of chaotic systems. Figure 6 It can be seen that the trajectory of the attractor after reconstruction of the OLTC vibration signal deviates from the origin of the phase space coordinate system to varying degrees under different operating conditions. The mean of the sum of the distances from each state point in phase space to the origin of the phase space coordinate system is defined as the average origin moment AOD, and its calculation formula is as follows: (35). In the formula, The number of state points. These are the coordinates of a point in phase space.

[0061] like Figure 7 The figure shows a comparison of KS entropy and AOD characteristic values ​​for 10 groups of samples under different operating conditions. As can be seen from the results, there are significant differences in the KS entropy values ​​and the average distance from the origin (AOD) values ​​under different operating conditions. Figure 6 As shown in Figure (a), the KS entropy is greater than zero under all operating conditions, indicating that the OLTC vibration signal has strong chaotic characteristics, and the KS entropy of normal and fault states is well distinguished. The KS entropy under normal operating conditions is concentrated between 30 and 36; the KS entropy for contact and arc plate loosening faults is concentrated below 25; the KS entropy for drive shaft screw loosening faults is concentrated above 35; and the KS entropy for transmission gear jamming faults is between 25 and 30. Figure 6 As shown in Figure (b), the AOD characteristic values ​​are distributed in different ranges under the four operating conditions. Among them, the contact and arc plate loosening fault and the transmission gear jamming fault are significantly different from the normal operating conditions. The larger AOD values ​​of these two operating conditions indicate that the chaotic attractor trajectory deviates from the origin of the phase space coordinates. The AOD value of the transmission shaft screw loosening fault is closer to the normal operating condition but can still be effectively distinguished.

[0062] Introducing KS entropy and average origin distance (AOD) can effectively describe the chaotic and geometric features of the reconstructed signal, and demonstrate good discriminative power under different operating conditions, making them suitable as fault feature values ​​for OLTC fault diagnosis. However, these feature parameters are highly sensitive to noise. Under low signal-to-noise ratio conditions, these features are easily interfered with, leading to misjudgments and consequently affecting the accuracy of the identification.

[0063] S3.3. A recursive graph mapping of vibration signals is introduced to further extract the chaotic features of the OLTC vibration signals. As a nonlinear feature extraction method, the recursive graph has the advantages of requiring less data and having strong noise resistance, thus solving the problem of noise sensitivity. Compared with traditional chaotic features based on Lyapunov exponents and correlation dimensions, the recursive graph can display the dynamic behavior of high-dimensional chaotic attractor trajectory diagrams in a two-dimensional graph, thereby depicting the intrinsic variation law of the dynamic system and showing the changes in system structure under different operating conditions in more detail.

[0064] Specifically, based on the reconstructed signal , Constructing a recursive matrix , (36); (37). In the formula, , The number of vectors after the phase space is reconstructed; For the Heaviside function; It is an L2 norm; This is a pre-defined recursive threshold constant. A small recursive threshold constant leads to relatively sparse recursive points, affecting the representation of phase space trajectory features. A large threshold constant leads to too many recursive points, hiding the true characteristics of the trajectory. Therefore, the standard deviation of the time series is used. 15%.

[0065] Combining recursive matrices ,by x-axis Create a 2D raster chart for the vertical axis. If... In a two-dimensional dot matrix diagram Add a black dot at the location; if If the point is not drawn on the graph, then it will not be plotted. Based on the above process, a... A two-dimensional recursive graph.

[0066] like Figure 8 The diagram shows recursion plots for four operating conditions. The recursion points in the vibration signal recursion plots are vertically distributed, concentrated in the 0-1000, 1500, and 2000-3000 ranges, exhibiting a predominantly layered structure. Combined with the vibration signal time-series waveforms in the recursion plots, it can be seen that each vibration shock wave in the OLTC vibration signal corresponds to a white area in the recursion plot. The stronger the vibration amplitude, the more prominent the white area. Therefore, the recursion plot can fully demonstrate the dynamic characteristics of the OLTC. Meanwhile, comparing the recursion plots of the OLTC vibration signals under different operating conditions reveals significant differences. Under normal operating conditions, the layered structure of the recursion plot is relatively sparse, with clear distinctions between different white areas, corresponding to the collision processes of different contacts. Under fault conditions, the transition between the black and white areas in the recursion plot becomes less smooth, indicating that the contact collision process of the OLTC changes. The amplitude and time of the OLTC vibration signal under fault conditions differ from those under normal operating conditions, consistent with the characteristics of the recursion plot. Therefore, the mechanical state changes of OLTC under different operating conditions can be reflected on the recursion graph, and it is feasible to further extract the chaotic characteristics of OLTC vibration signals based on the recursion graph.

[0067] The texture structure of a recursive graph can depict the intrinsic dynamic mechanism of OLTC and can be used for analysis. As seen in the recursive graph acquisition process, a recursive graph is a raster image composed only of black and white dots. Each pixel has only two states: black or white. The transitions between pixels are relatively sharp, lacking variation in level. Directly extracting image texture features from it will not yield effective features. Therefore, grayscale preprocessing is performed before extracting texture features from the recursive graph. Although the recursive graph itself consists of only black and white states, its visual appearance exhibits different grayscale effects due to varying black dot density. When the black dot density is high, the image area appears darker, while when the black dot density is low, the image area appears lighter. Based on this characteristic, by calculating the local black dot density, the original binary recursive graph is converted into a grayscale recursive graph. The higher the local black dot density, the darker the grayscale value, thus more intuitively reflecting the local structural features of the recursive graph.

[0068] like Figure 9As shown, a small matrix is ​​used to scan the entire image, and the local black dot density is obtained by weighting the region around each pixel. (38). In the formula, These are the weights of the convolution kernel; It is a recursive matrix; The kernel size is [5×5], and the kernel size is selected as the [5×5] matrix in this invention.

[0069] During image convolution, the convolution kernel cannot cover edge pixels. This invention primarily calculates the overall texture features of the image, as edge pixels have a relatively small impact on feature extraction. Therefore, only pixels completely covered by the convolution kernel are calculated, ignoring boundary regions. The convolution kernel moves pixel-by-pixel across the image, calculating the local black point density at each location until the entire image is covered. The convolution kernel in this invention is a [5×5] matrix, with its region density values ​​ranging from 0 to 25. To map the original region density values ​​to a grayscale range, the image region density values ​​are normalized and mapped to a 5-bit grayscale range, with values ​​ranging from [0, 32]. This ensures complete coverage of the image region density values ​​without data loss, effectively controls image data storage and computational complexity, and still effectively represents the features of the image region. At this point, a grayscale value of 0 represents pure black, and 32 represents pure white. The normalization calculation formula is as follows: (39); where min(d) is The minimum value, max(d) is The maximum value.

[0070] like Figure 9 As shown, after converting the mapping recursive graph to grayscale, a grayscale recursive graph is obtained. The black and white pixels have certain gradient characteristics, and the density information of the recursive points can smoothly present the changing trend, thus describing the state information of the on-load tap changer switching process in more detail. The texture features of the grayscale recursive graph are extracted and analyzed below.

[0071] Tamura texture is a texture feature extraction method that includes six attributes: roughness, contrast, orientation, granularity, regularity, and coarseness. Among these, contrast, roughness, and orientation are particularly important for image recognition. This invention calculates the roughness, contrast, and orientation of the grayscale recursive map of OLTC vibration signals to describe its texture features.

[0072] Specifically, roughness is used to describe the granularity of texture. The larger the granularity of the texture, the rougher the image; conversely, the smaller the granularity, the more refined the image appears. Roughness can reflect the global structural characteristics of a signal; high roughness indicates the existence of a large-scale structure in the system, while low roughness indicates the existence of a highly complex short-time structure, typically corresponding to non-stationary signals. Roughness calculation involves the following steps: First, calculate the roughness contained in the image... Average gray value of pixels in the active window (pixels) (40). In the formula, Indicates that it is located at The pixel grayscale value at that location; Then, the average intensity difference between adjacent non-overlapping windows in the horizontal direction for each pixel is calculated. and the average intensity difference between adjacent non-overlapping windows in the vertical direction. . (41). The optimal activity window is At this time Substitute the value into equation (41). The value reaches its maximum. Finally, in the entire image... The average value is the roughness. , (42).

[0073] Specifically, contrast reflects the depth of image texture and the clarity of the image; the deeper the texture, the greater the contrast, and the clearer the visual effect. Reflected in vibration signals, contrast can indicate the intensity of signal energy distribution; the greater the contrast, the stronger the signal impact. The formula for calculating contrast is... (43). In the formula, , It is a fourth-order moment; It is variance.

[0074] Specifically, orientation is used to characterize the degree of concentration or distribution of texture in a specific direction. The more irregular the shape of an image, the greater its orientation. The orientation calculation process is as follows: First, calculate the gradient vector at each pixel. (44); (45); Direction angle (46). In the formula, The gradient vector at each pixel; This represents the amount of change in the gradient vector in the horizontal direction; This represents the change in the gradient vector in the vertical direction. Then, a histogram of the vector directions is constructed. , (47). In the formula, It is a quantification level of direction and angle; It corresponds to the direction angle. The total number of edge pixels on the surface; For when , The number of pixels at time, of which This is the threshold for the magnitude of the gradient vector. Finally, the directionality... (48). In the formula, p is a histogram The peak value; This represents the number of peaks in the histogram. For quantification range; for The quantization value in the maximum histogram.

[0075] like Figure 10 One-dimensional distribution map of Tamura texture features under different working conditions. Figure 9 (a) shows the roughness results of the recursion plot under different working conditions. The roughness is greatest under normal working conditions and smallest under gear jamming faults. This is because the recursion points are relatively sparse under these working conditions, which is reflected in the rougher image features. Figure 9 As shown in (b), the contrast between the loose contact and arc plate faults and the jammed transmission gear faults is relatively large, indicating that the texture of the recursion graph is relatively deep and the visual effect is obvious; Figure 9 As shown in (c), the orientation is smallest under normal operating conditions, indicating that the recurrence graph shape is relatively regular under normal operating conditions. However, the vibration signal noise is greater under fault conditions, therefore, the recurrence graph under fault conditions is relatively irregular, i.e., the orientation is greater. Overall, the Tamura texture features under different operating conditions can distinguish the different operating conditions of OLTC, and the extracted features are strongly correlated with its operating conditions.

[0076] S4. Introduce natural gradient boosting NGBoost to construct a multi-class fault diagnosis model.

[0077] In multi-class classification tasks, sample categories It is considered a random variable that follows a categorical distribution. For a given sample feature vector... The model outputs its corresponding class probability vector. And satisfy the probability normalization constraint: (49). Among them, the category probability Parameterized representation using the Softmax function: (50). In the formula, This is the output function for the model corresponding to the k-th class; Let be the model output function corresponding to the j-th class. Model training uses the negative log-likelihood function as the loss function, expressed as: (51). To improve the optimization stability during the probability distribution modeling process, NGBoost introduces a natural gradient optimization mechanism during parameter update, the basic form of which is: (52). Among them, This represents the ordinary gradient; This is the Fisher information matrix. Through natural gradient correction, the model can iteratively update along a more reasonable descent direction in the probability distribution space. In the model implementation, features are first extracted from the collected vibration signals to construct a fixed-length feature vector for the samples: (53). The input model is then standardized. The aforementioned feature vectors serve as the input to the NGBoost multi-class model, which in this invention is the obtained [1×5] chaotic feature matrix, used to learn the mapping relationship between the feature space and the class probability distribution. The model uses a regression tree as the base learner and gradually approximates the class probability distribution through gradient boosting. Its parameter update form is expressed as: (54). Among them, For learning rate, This is a regression tree model trained based on natural gradients. In this invention, the number of base learners is set to 300, the maximum depth of the regression tree is 3, the learning rate is set to 0.05, and the minimum number of leaf node samples is set to 10. After the model training is completed, for the test samples... The category with the highest probability value is selected as the final predicted category. (55).

[0078] This invention achieves adaptive localization of key signal segments through short-time entropy ratio, solving the problem of signal timing inconsistency; it effectively reduces noise using the TVFEMD algorithm while preserving core mechanical impact features; it extracts KS entropy, AOD, and Tamura texture features based on chaos theory, constructing a feature system with both anti-interference and discriminative power; and finally, it achieves high-precision fault identification under multiple operating conditions through the NGBoost model. This invention effectively handles the non-stationarity, nonlinearity, and high noise issues of OLTC vibration signals, with stable and reliable feature extraction and high fault diagnosis accuracy. It provides a new technical path for OLTC online condition monitoring and has significant engineering application value for improving transformer operation and maintenance levels and reducing power outage risks in power systems. Future development can further expand the operating condition coverage, optimize feature extraction efficiency and model generalization ability, and promote the large-scale application of this technology in actual power grid scenarios.

Claims

1. A load tap changer fault diagnosis method based on TVFEMD denoising-chaos feature fusion-NGBoost, characterized in that, The method comprises the following steps: S1, using short-time energy entropy ratio algorithm to adaptively locate the vibration signal at the switching switch action time, by setting frame length, frame shift and threshold parameters, accurately intercepting the signal segment containing key fault information, and uniformly extracting the time length to ensure the time sequence consistency; S2, using time-varying filter empirical mode decomposition (TVFEMD) to denoise the vibration signal, rearranging the local cutoff frequency and selecting the mode component to reconstruct the signal according to the energy proportion; S3, based on the chaotic characteristics, extracting K-S entropy and average origin distance (AOD) as chaotic features and geometric features, and mapping the phase space trajectory into a recurrence plot, extracting roughness, contrast and direction degree as three types of Tamura texture features after gray preprocessing, and forming a 5-dimensional chaotic feature matrix; S4, constructing an NGBoost multi-classification fault diagnosis model, taking the feature matrix as the input, and realizing the recognition of the OLTC mechanical state.

2. The TVFEMD de-noising-chaos feature fusion-NGBoost-based on-load tap changer fault diagnosis method according to claim 1, characterized in that, In step S1, the short-time energy entropy ratio algorithm process includes the following steps: S1.1, for a length of... time series Amplitude normalization is performed. ;right Apply windowing function Frame processing yields the first Frame vibration signal is , (1); where, Represents the numerical value of a frame; This is the frame shift length; Frame length; S1.2 represents the total number of frames after the signal is framed; Perform a Fourier transform to obtain the spectrum of the signal in each frame. Spectral line frequency components The energy spectrum is , No. Frequency components of the frame signal after Fourier transform Corresponding probability density function (2); No. Short-time spectral entropy of frame signal (3); No. Energy of frame signal (4); where, It is a constant; The amplitude of the nth sampling point of the i-th frame signal; The energy entropy ratio of the frame signal (5); S1.3, Calculate the normalized short-time energy entropy ratio (6); where R is a sequence of short-time energy entropy ratios of all frames; min(R) is the minimum value in the whole sequence; max(R) is the maximum value in the whole sequence.

3. The TVFEMD de-noising-chaos feature fusion-NGBoost-based on-load tap changer fault diagnosis method according to claim 2, characterized in that, In step S1, the adaptive positioning step of the vibration signal comprises: S1.4, selecting a window function and setting a window length and a frame length, performing frame processing on the vibration signal, and calculating a short-time energy entropy ratio R' of each frame; S1.5, selecting a time threshold value and a short-time energy entropy ratio threshold value ; S1.6, comparing the short-time energy entropy ratio R' of each frame, when and the frame number of the signal exceeding the short-time energy entropy ratio threshold value is longer than , the frame is defined as the starting position of the signal; S1.7, when and the frame number of the signal being less than the short-time energy entropy ratio threshold value is longer than , the frame is defined as the ending position of the signal.

4. The TVFEMD de-noising-chaos feature fusion-NGBoost-based on-load tap changer fault diagnosis method according to claim 3, characterized in that, In step S2, TVFEMD includes two processes of S2.1, local cutoff frequency rearrangement and S2.2, screening based on time-varying filter.

5. The TVFEMD de-noising-chaos feature fusion-NGBoost-based on-load tap changer fault diagnosis method according to claim 4, characterized in that, The specific steps of the local cutoff frequency rearrangement in step S2 include: S2.1.1, for the real value signal The analytic signal is obtained by Hilbert transform (7); wherein, is the vibration amplitude of the i th sampling point; , and are amplitudes, , and are phases; (8); j is the imaginary unit; S2.1.2, when , is a local maximum; , is a local minimum; for all maximum points , minimum points , interpolation is performed to obtain curves and , and further obtain: (9); S2.1.3, for all maximum points , interpolation is performed to obtain , for all minimum points , interpolation is performed to obtain , and (10); wherein is the derivative of , is the derivative of , and the local cutoff frequency (11) is defined; S2.1.4, the time sequence of the maximum points of the signal is , (12); wherein, ρ=0.25; if formula (12) is established, then is an interruption point, and , represents the interruption point sequence; if , then is the rising edge of , and is regarded as the minimum value; if , then is the falling edge of , and is regarded as the minimum value; the remaining part is regarded as the peak value; S2.1.5, interpolation is performed between all peak values to obtain the rearranged local cutoff frequency .

6. The TVFEMD de-noising-chaos feature fusion-NGBoost-based on-load tap changer fault diagnosis method according to claim 5, characterized in that In step S2, the filtering step based on time-varying filtering includes: S2.2.1, reconstructing the signal based on the rearranged cutoff frequency: (13); will The extreme points are taken as nodes, Divide the function into n segments, each with a step size of m; where n is the order of the B-spline function; perform B-spline interpolation approximation between the extreme points, and denote the approximation result as... This represents the local mean curve; S2.2.2, Determine the residual signal. Does it meet the stopping criteria? ;in, Given a bandwidth threshold; if satisfied, then Let IMF be the modal component obtained from TVFEMD decomposition; if not satisfied, let In the formula, x (k) (t) represents the signal in the k-th screening iteration, m (k) (t) is the local mean curve for the k-th iteration, until the stopping criterion is met; (14); where, For Loughlin instantaneous bandwidth, (15); For weighted average instantaneous frequency, (16); S2.2.3, Filter the modal components IMF, assuming that the K modal components of the signal obtained by TVFEMD decomposition are ;in, The energy of each modal component is defined as: (16); among which, Signal length; energy percentage of the modes (18); Cumulative energy percentage (19); Based on the set threshold, select the cumulative energy percentage that meets the requirements. The first m modes with values ​​less than the threshold are reconstructed to obtain the denoised signal. (20).

7. The TVFEMD de-noising-chaos feature fusion-NGBoost-based on-load tap changer fault diagnosis method according to claim 6, characterized in that, In step S3, the step of reconstructing the phase space based on chaotic characteristics is as follows: the OLTC vibration signal is regarded as a set of lengths... time series , (21); where, This is the starting time point of the vibration signal; For time intervals; divide the time series Reconstructed indivual A dimensional vector, the reconstructed phase space is represented as (twenty two); (23); where, To reconstruct the phase point coordinates in the phase space; To delay time, The embedding dimension.

8. The TVFEMD de-noising-chaos feature fusion-NGBoost-based on-load tap changer fault diagnosis method according to claim 7, characterized in that, In step S3, the delay time is determined by using the mutual information method : original vibration signal After delay time τ, it becomes The mutual information value between S and Q (24); In the formula, (25), The information entropy of signal S; The information entropy of signal Q; The joint information entropy of S and Q; The probability of the point with the median value of in the sampling point appearing in the entire data segment; The probability of the point with the median value of in the sampling point appearing in the entire data segment; The probability of the sampling value of in the original signal and the sampling value of in the delay signal; Take the mutual information The first local minimum point generated by the mutual information corresponds to the delay time τ.

9. The TVFEMD de-noising-chaos feature fusion-NGBoost-based on-load tap changer fault diagnosis method according to claim 8, characterized in that, In step S3, the CAO algorithm is used to calculate the embedding dimension m: for the Reconstructed vectors When the embedding dimension is m, it is represented as , To and nearest neighbor, (26); Define parameters (27); where, Represents Euclidean distance; This is the i-th reconstructed vector when the embedding dimension is m+1; To achieve the same result as when the embedding dimension is m+1 The nearest point; define parameters average (28) (29); through Parameters from arrive Changes over time; when Greater than a certain value hour, If change stops, then This is the minimum embedding dimension m that we are ultimately looking for.

10. The TVFEMD de-noising-chaos feature fusion-NGBoost-based on-load tap changer fault diagnosis method according to claim 9, characterized in that, In step S3, a reference quantity is defined. (30); (31);Use The following steps were taken to improve the CAO algorithm: First, based on Determining the threshold based on fluctuations The second step is to calculate. mean and absolute value of deviation , (32); (33); The third step is to compare the absolute values ​​of the deviations. With threshold Size, if the maximum absolute value of the deviation Then the first one satisfies corresponding That is, the embedding dimension ;like Then find the minimum value of the absolute value of the deviation. middle Corresponding subscript ,by Repeat steps one and two for the new starting point until an embedding dimension m that satisfies an acceptable level of bias is found.

11. The TVFEMD de-noising-chaos feature fusion-NGBoost-based fault diagnosis method for on-load tap changers according to claim 10, characterized in that, In step S3, the KS entropy is calculated using the correlation integral method, and the calculation formula is as follows: (34); where K is the value of KS entropy; For the first The correlation integral function of dimension; Let K be the correlation integral function in the (m+1)th dimension; choose the stable K value as the estimate of the KS entropy; define the mean of the sum of the distances from each state point in phase space to the origin of the phase space coordinate system as the mean origin moment AOD, calculated by the following formula: (35); where, The number of state points. These are the coordinates of a point in phase space.

12. The TVFEMD de-noising-chaos feature fusion-NGBoost-based on-load tap changer fault diagnosis method according to claim 11, characterized in that, In step S3, the method for mapping the phase space trajectory to a recursive graph is as follows: construct a recursive matrix based on the reconstructed signal. , (36); (37); where, , The number of vectors after the phase space is reconstructed; For the Heaviside function; It is an L2 norm; The recursive threshold constant is given in advance; combined with the recursion matrix ,by x-axis Create a 2D raster chart for the vertical axis; if In a two-dimensional dot matrix diagram Add a black dot at the location; if If the point is not drawn on the graph, then a point is obtained. A two-dimensional recursive graph.

13. The TVFEMD de-noising-chaos feature fusion-NGBoost-based on-load tap changer fault diagnosis method according to claim 12, characterized in that, In step S3, the grayscale preprocessing method is as follows: scan the entire image using a small matrix, and perform weighted calculations on the area surrounding each pixel to obtain the local black dot density. , (38); where, These are the weights of the convolution kernel; It is a recursive matrix; The kernel size is [size missing]. The image region density values ​​are normalized and mapped to a 5-bit grayscale range, with values ​​in the range [0, 32]. The normalization calculation formula is as follows: (39), where min(d) is The minimum value, max(d) is The maximum value.

14. The TVFEMD de-noising-chaos feature fusion-NGBoost-based fault diagnosis method for on-load tap changers according to claim 13, characterized in that, In step S3, the roughness is extracted by the following method: first, the average gray value of the pixels in the active window of each pixel in the image is calculated (40); where represents the gray value of the pixel at ; ; then, the average intensity difference between adjacent non-overlapping windows in the horizontal direction and the average intensity difference between adjacent non-overlapping windows in the vertical direction are calculated; (41); the optimal active window is ; the value of at this time is substituted into equation (41), and the value of reaches the maximum; finally, the average value of in the entire image is the roughness , and the calculation formula is (42).​ 15. The TVFEMD denoising-chaos feature fusion-NGBoost-based on-load tap changer fault diagnosis method according to claim 14, characterized in that, In step S3, the contrast calculation formula is (43); wherein, ; is the fourth moment; is the variance; the direction degree calculation process is as follows: first, the gradient vector at each pixel is calculated , (44); wherein, (45), the direction angle (46); wherein, represents the change amount of the gradient vector in the horizontal direction; represents the change amount of the gradient vector in the vertical direction; Then, a histogram of the vector direction is constructed , (47); where, is the quantization level of the direction angle; is the total number of edge pixels in the corresponding direction angle ; is the number of pixels when , ; is the threshold of the gradient vector module; and finally, the direction degree (48); where, ; p is the peak value of the histogram ; is the number of histogram peaks; is the quantization range; is the maximum histogram in .

16. The TVFEMD de-noising-chaos feature fusion-NGBoost-based fault diagnosis method for on-load tap changers according to claim 15, characterized in that, In step S4, the method for fault type identification using the NGBoost multi-classification fault diagnosis model is: for the sample feature vector , the corresponding category probability vector satisfies the probability normalization constraint: (49); the category probability (50); in the formula, is the model output function corresponding to the kth category; is the model output function corresponding to the jth category; the model training takes the negative log-likelihood function as the loss function, and the expression is: (51); in the parameter updating process, the natural gradient optimization mechanism is introduced: (52); wherein, represents the ordinary gradient, is the Fisher information matrix; in the model implementation, firstly, the collected vibration signal is subjected to feature extraction, a fixed-length feature vector (53) of the sample is constructed as the input of the NGBoost multi-classification model, and standardization processing is performed before the input model; the model uses the regression tree as the base learner, and gradually approaches the category probability distribution through the gradient boosting method, and the parameter updating is: (54); in the formula, is the learning rate, is the regression tree model obtained based on the natural gradient training; after the model training is completed, the category with the maximum category probability value is selected as the final prediction category, (55).

Images