Feature extraction method and device for converter transformer vibration signal multi-feature fusion

Abstract

The present invention discloses a feature extraction method and device for multi-feature fusion of converter transformer vibration signals. The method selects the on-site measured signal of the converter transformer as the analysis object, and performs a collective empirical mode decomposition on it to extract the intrinsic mode. Then, the time domain statistical characteristic components, wavelet entropy and fractal dimension of each intrinsic mode component are calculated to form a characteristic parameter matrix. Finally, the matrix is ​​subjected to data dimension reduction through principal component analysis to form a multi-feature fusion vector, and the feature vectors of different operating conditions of the converter transformer are classified using fuzzy C-means clustering. This method overcomes the limitations of a single feature and has good discrimination performance under different operating conditions.

Description

Technical Field

[0001] The present invention relates to the field of electric power systems, and in particular to a feature extraction method and device for fusing multiple features of a converter transformer vibration signal. Background Art

[0002] With the continuous growth of the national economy, my country's electricity demand continues to increase. The extremely unbalanced distribution of load power consumption and power generation resource reserves in my country makes ultra-long-distance and ultra-large-capacity DC transmission inevitable. The exploitable hydropower resources in Sichuan Province, Yunnan Province, and Tibet Autonomous Region account for about 2 / 3 of the national reserves, and the coal reserves in Shanxi Province, Shaanxi Province, and Inner Mongolia Autonomous Region account for about 2 / 3 of the national reserves. However, there is a serious shortage of power generation resources in the economically developed areas east of the Beijing-Guangzhou Railway and the eastern coastal areas, which makes large-scale cross-regional flow of energy and electricity inevitable. The high voltage, long distance, large capacity, low loss and other outstanding advantages of high-voltage direct current transmission technology provide an important way to solve the problem of how to reduce transmission line losses and save precious land resources.

[0003] Whether the converter transformer operates safely is directly related to whether the DC transmission system can operate normally. However, under the action of AC and DC electric fields, the converter transformer has problems such as too high withstand voltage level, coupling of electric, magnetic, mechanical and thermal multi-physical fields, complex insulation results, larger oil tank space and core cross-section than conventional transformers, and higher harmonic components of valve hall outflow voltage, which leads to much greater vibration and noise in the operation of the converter transformer than that of the power transformer. In recent years, the Southern Power Grid Company has experienced accidents such as explosion and fire of high-voltage bushings on the grid side of the converter transformer at Zhaoqing Converter Station and Xingren Converter Station, and cracking of the valve side bushing riser of the converter transformer of the Tianguang DC project, which pose a great threat to the safe operation of the DC transmission system. However, most of the existing detection methods are based on electrical signals and cannot directly detect the transformer under live operation conditions. Conventional dismantling and maintenance not only causes power outage losses but also has a great impact on the life of the equipment. Therefore, how to monitor the converter transformer online and quickly evaluate its operating status has become the main concern of the circuit department at present, and it is also a problem that needs to be solved urgently.

[0004] At present, the vibration method for evaluating the operating status of transformers and its online monitoring technology have received widespread attention. The surface vibration signal of the transformer contains a large number of rich states and is closely related to the working status of the transformer's own structural components (such as the core, windings and other external components). However, the processing of vibration signals often only focuses on the characteristic expression of one aspect (such as the frequency domain, time domain, etc. of the signal), which will lead to information loss and reduced evaluation accuracy. Summary of the invention

[0005] In order to solve the problem that the existing converter transformer vibration signal processing method causes information loss and reduced evaluation accuracy, an embodiment of the present invention provides a feature extraction method and device for converter transformer vibration signal multi-feature fusion. This method and device overcome the limitations of a single feature and have good discrimination performance under different working conditions.

[0006] To achieve the above object, the technical solution of the present invention is:

[0007] In a first aspect, an embodiment of the present invention provides a feature extraction method for multi-feature fusion of a converter transformer vibration signal, comprising:

[0008] Obtaining the original vibration signal of the converter transformer, and decomposing the original signal into several basic mode components;

[0009] Calculate the dimensionless indexes of the first A basic mode components respectively, and obtain the time domain statistical characteristics of the corresponding basic mode components;

[0010] Perform continuous wavelet transform on the first A basic mode components respectively to obtain the wavelet coefficient matrix corresponding to the basic mode components, and calculate the entropy value of the wavelet coefficient matrix to obtain the wavelet entropy feature corresponding to the basic mode components;

[0011] Extract the fractal dimension features of the first A basic mode components respectively to obtain the fractal dimension features of the corresponding basic mode components;

[0012] The obtained time domain statistics features, wavelet entropy features, and fractal dimension features of each basic mode component are combined to form a feature matrix;

[0013] The principal component analysis method is used to reduce the dimension of the feature matrix and obtain the fused feature vector to realize the feature extraction of the converter transformer vibration signal.

[0014] Perform fuzzy C-means clustering analysis on the fused feature vector, compare the real label with the classification label to judge the effect of feature extraction, and realize the detection of the operating status of the converter transformer;

[0015] Wherein, A is a positive integer.

[0016] In a second aspect, an embodiment of the present invention provides a feature extraction device for a converter transformer vibration signal by fusing multiple features, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described above when executing the computer program.

[0017] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method described above are implemented.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] This method selects the measured signals of converter transformers as the analysis object, and extracts the inherent modes by performing ensemble empirical mode decomposition. Then, the time-domain statistical characteristic components, wavelet entropy and fractal dimension of each inherent mode component are calculated to form a characteristic parameter matrix. Finally, the matrix is ​​reduced in dimension by principal component analysis to form a multi-feature fusion vector, and the characteristic vectors of different working conditions of converter transformers are classified by fuzzy C-means clustering. This method overcomes the limitations of single features and has good discrimination performance under different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flow chart of a feature extraction method for multi-feature fusion of converter transformer vibration signals provided in Example 1 of the present invention;

[0021] Figure 2 This is a schematic diagram of the measurement point arrangement for the converter transformer;

[0022] Figure 3 It is the empirical mode decomposition diagram of load signal;

[0023] Figure 4 It is the empirical mode decomposition diagram of no-load signal;

[0024] Figure 5 is the ideal cluster distribution diagram;

[0025] Figure 6 is the actual cluster distribution diagram;

[0026] Figure 7 A schematic diagram of the composition of a feature extraction device for fusing multiple features of a converter transformer vibration signal provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0028] Embodiment 1:

[0029] See also Figure 1 As shown, the feature extraction method for multiple feature fusion of converter transformer vibration signal provided in this embodiment includes the following steps:

[0030] Empirical mode decomposition steps:

[0031] The original vibration signal of the converter transformer is obtained and decomposed into several basic mode components; the details are as follows:

[0032] Take the original signal x(t) as the signal to be processed, calculate all local extreme points of the signal, use the cubic spline interpolation method to fit all the maximum and minimum points respectively and record them as the upper and lower envelopes of the signal, calculate the mean of the upper and lower envelopes to get the mean sequence m(t),

[0033] Subtract the mean sequence m(t) from the original signal x(t) to obtain:

[0034] h1(t)=x(t)-m(t)

[0035] Check whether h1(t) satisfies the following two conditions:

[0036] (1) The number of extreme points in the entire data sequence is the same as the zero-crossing point or the difference is no more than 1

[0037] (2) The mean of the upper and lower envelopes determined by the local extreme value of the signal is 0

[0038] If it is not satisfied, h1(t) is used as the signal to be processed and the above operation is repeated until the condition is satisfied. At this time, the mean sequence h n (t) is recorded as a basic mode component c1(t) (imf component)

[0039] h n (t) = c1(t)

[0040] Decompose the first basic mode component from the original signal to obtain the residual sequence r1(t)

[0041] r1(t)=x(t)-c1(t)

[0042] Repeating the above operation with the remaining sequence r1(t) as the signal to be processed can obtain the second, third, and Ath basic mode components in sequence until the remaining sequence meets the stopping criterion. In this way, the original signal is decomposed into several basic mode components and a residual component:

[0043]

[0044] The stopping criterion is set to limit the standard deviation S between two consecutive processing results. d

[0045]

[0046] Where T represents the time span of the signal, h (k-1) (t),h k (t) is the time series of two consecutive processing results in the process of screening basic components, S d The threshold is set to 0.2.

[0047] Thus, by performing the above steps, the original signal can be accurately and effectively decomposed into several basic mode components.

[0048] Steps for extracting time domain statistics:

[0049] The waveform index, peak index, margin index, pulse index and kurtosis index of each basic mode component are calculated respectively to obtain the dimensionless statistical component parameters (i.e., time-domain statistical characteristics) of the time domain waveform of the corresponding basic mode component. Since the dimensionless parameters are only related to the signal probability density p(x), and these parameters are sensitive enough to the operating status, faults and defects of the machine, they have a high evaluation value and are defined as:

[0050]

[0051] For a signal sequence x(n), its probability density p(x) is defined as T is the total observation time.

[0052] Waveform index l=2,m=1:

[0053]

[0054] is the root mean square value of the signal

[0055] is the absolute mean amplitude,

[0056] Peak index l→∞, m=1:

[0057]

[0058] x p =max{x(n)} is the signal peak value

[0059] Margin index l→∞, m=1 / 2:

[0060]

[0061] is the square root amplitude of the signal

[0062] Pulse index l→∞, m=1:

[0063]

[0064] Kurtosis index

[0065]

[0066] Among them, α4 is the fourth-order moment, σ xis the signal standard deviation,

[0067] Wavelet entropy feature extraction steps:

[0068] Perform continuous wavelet transform on each basic mode component to obtain the wavelet coefficient matrix of the corresponding basic mode component, and calculate the entropy value of the wavelet coefficient matrix to obtain the wavelet entropy feature of the corresponding basic mode component. b,a (t) is inner-producted with the signal to find the similar components between the original signal and the wavelet function. The wavelet transform of the signal is recorded as:

[0069]

[0070] The wavelet basis function is Morle wavelet, a and b are scale factors and translation factors respectively.

[0071] Assume that the signal x(n) is decomposed into high-frequency signal components D j (n) and low-frequency signal component A j (n), A m (n) is represented by D m+1 (n):

[0072]

[0073] The sliding window is defined on the wavelet reconstruction coefficients as:

[0074] W(m;w,δ)={d j (n),n=1+mδ,...,w+mδ}

[0075] Where m=0,1,...,M, and M=(Nw) / δ

[0076] Therefore, the total energy expression is:

[0077]

[0078] According to Shannon entropy theory, wavelet entropy can be defined as:

[0079]

[0080] in, is a relative wavelet.

[0081] Through the above steps, wavelet entropy features can be accurately and effectively extracted

[0082] Fractal dimension feature extraction steps:

[0083] Four fractal dimension extraction methods based on the time domain signal are used to extract characteristic parameters for each mode component to construct the characteristic matrix:

[0084] Katz Method:

[0085] The length of the curve for calculating the fundamental mode component is denoted by L

[0086] L = sum(dist(i,i+1))

[0087] Calculate the distance between the first point of the basic mode component and its farthest point and record it as d

[0088] d=max(dist(1,i))

[0089] Calculating Katz fractal dimension

[0090]

[0091] Where N' is the total number of discrete points of the fundamental mode component.

[0092] Castiglioni Method:

[0093] The maximum amplitude difference of the fundamental mode component is calculated and recorded as d

[0094] d=max(x i )-min(x i )

[0095] The total length of the basic mode component is calculated and recorded as L

[0096]

[0097] Calculating Castiglioni fractal dimension

[0098]

[0099] Sevcik Method:

[0100] Normalize the fundamental mode components

[0101]

[0102] Calculate the length of the curve represented by the fundamental mode component

[0103] L = sum(dist(i,i+1))

[0104] Calculating the Sevcik fractal dimension

[0105]

[0106] Higuchi Method

[0107] Using the delayed reconstruction method on the basic mode components, a new matrix is ​​obtained

[0108]

[0109] Calculate each The length of the curve is denoted as L m (k)

[0110]

[0111] The average of the total sequence curve length is estimated using the matrix curve lengths generated by k delays

[0112]

[0113] Calculating Higuchi fractal dimension

[0114]

[0115] Steps for principal component analysis dimensionality reduction of feature matrix:

[0116] After empirical mode decomposition, the original signal is transformed into several basic mode components. Take each component r1, r2, ..., r n The time domain statistics, wavelet entropy and fractal dimension features of each basic mode component are extracted respectively to form a feature matrix, and each row of the matrix represents all the features of a basic mode component.

[0117] Although multi-feature extraction can avoid the one-sidedness and inaccuracy of a single feature index, it will increase the parameter dimension. Since each feature index is directly or indirectly correlated, directly analyzing the feature matrix will lead to overfitting. In order to reduce the computational complexity and avoid data redundancy, the principal component analysis method is used to reduce the dimension of the feature to obtain a one-dimensional feature vector.

[0118] For the feature matrix X m×N The dimensionality reduction process is as follows:

[0119] Normalization

[0120]

[0121]

[0122] Where x(i) is the signal sequence, x'(i) is the normalized signal sequence, mean is the mean of the signal, N is the number of discrete points in the sequence, and σ is the variance.

[0123] Sample feature covariance matrix calculation:

[0124]

[0125] Solve the reduced-dimensional matrix according to the singular value decomposition:

[0126] [U,S,V]=svd[Cov]

[0127] Where svd represents the singular value decomposition of the matrix, and the obtained U matrix is ​​called the dimensionality reduction matrix. U is obtained by arranging all the eigenvectors of Cov in descending order.

[0128] Dimensionality reduction calculation:

[0129]

[0130] k is the number of rows of the dimension reduction matrix selected to control the dimension of the output.

[0131] Fuzzy C-means clustering analysis steps:

[0132] Perform fuzzy C-means clustering analysis on the fused feature vector, compare the real label with the classification label to judge the effect of feature extraction, and realize the detection of the operating status of the converter transformer

[0133] Fuzzy C-means integrates the essence of fuzzy theory and assigns a membership weight to each object and each cluster in the sample set to characterize the degree to which the object belongs to the cluster. The closer the object is to the cluster center, the higher the membership.

[0134] First, set the minimization objective function J m

[0135]

[0136] The membership matrix u is processed by random initialization method ij ;

[0137] Calculate the centroid:

[0138]

[0139] Iterative update of membership:

[0140]

[0141] If||U k+1 -U k ||<ε, then stop the iteration and calculate the minimum objective function, otherwise return to step 2 to update the cluster center and membership. The algorithm is terminated when the objective function is less than the set value.

[0142] Where m is the number of clusters; i, j are class labels; u ij Represents sample x i The degree of membership to class j; cj is the center of cluster j; the || || operation represents the distance measure.

[0143] The following is a further verification and explanation of this method with an application scenario example:

[0144] The vibration sensors on the converter transformer box surface are arranged as follows: Figure 2 As shown, the numbers 1-12 are the sensor numbers. Figure 3 , 4 The empirical mode decomposition of 422 sets of vibration data at one measuring point is shown (243 sets of vibration signals under 1.05pu load operation and 179 sets of vibration signals under rated voltage no-load operation). Figure 2 , Figure 3 It is not difficult to see that the complexity and sparsity of the basic mode components of each order obtained after empirical mode decomposition of vibration signals under different working conditions vary greatly. Subsequently, multi-feature extraction is performed based on the time domain statistics, wavelet entropy, and fractal dimension of the signal, and the fused feature vector is obtained through principal component analysis. Each fused feature vector is labeled, with the label 1 when running under load and the label 2 when running without load. These multidimensional feature vectors are classified using fuzzy C-means clustering analysis, and the classification accuracy is calculated by comparing the self-set labels with the cluster labels to determine the effect of feature extraction. The predetermined number of clusters is 2; the weighted index is 2; and the Gaussian kernel parameter is 10. Figure 5 , Figure 6 This is a comparison diagram of the actual situation and the classification situation. Except for the reference mark "1", all other numbers in the figure are "2".

[0145] It has been verified that the classification accuracy of the feature extraction vector based on the method of the present invention reaches 91.25%, which can effectively distinguish the no-load and load operating states of the converter transformer, overcoming the inaccuracy and limitations of a single feature.

[0146] In summary, the method of the present invention can perform fuzzy C-means clustering analysis on the fused feature vector, compare the real label with the classification label to judge the quality of feature extraction, and achieve the purpose of detecting the operating status of the converter transformer through the vibration signal.

[0147] Embodiment 2:

[0148] See also Figure 7 As shown, the feature extraction device for converter transformer vibration signal multi-feature fusion provided in this embodiment includes a processor 71, a memory 72, and a computer program 73 stored in the memory 72 and executable on the processor 71, such as a feature extraction program for converter transformer vibration signal multi-feature fusion. When the processor 71 executes the computer program 73, the steps of the above embodiment 1 are implemented, such as Figure 1 Steps shown.

[0149] Exemplarily, the computer program 73 may be divided into one or more modules / units, which are stored in the memory 72 and executed by the processor 71 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program 73 in the feature extraction of the multi-feature fusion of the converter transformer vibration signal.

[0150] The feature extraction device for fusion of multiple features of the converter transformer vibration signal can be a computing device such as a desktop computer, a notebook, a PDA, and a cloud server. The feature extraction device for fusion of multiple features of the converter transformer vibration signal can include, but is not limited to, a processor 71 and a memory 72. Those skilled in the art can understand that Figure 7 It is only an example of a feature extraction device for fusion of multiple features of converter transformer vibration signals, and does not constitute a limitation of the feature extraction device for fusion of multiple features of converter transformer vibration signals. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the feature extraction device for fusion of multiple features of converter transformer vibration signals may also include input and output devices, network access equipment, buses, etc.

[0151] The processor 71 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0152] The memory 72 may be an internal storage unit of the feature extraction device for the multi-feature fusion of the converter transformer vibration signal, such as a hard disk or memory of the feature extraction device for the multi-feature fusion of the converter transformer vibration signal. The memory 72 may also be an external storage device of the feature extraction device for the multi-feature fusion of the converter transformer vibration signal, such as a plug-in hard disk, a smart memory card (SmartMedia Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the feature extraction device for the multi-feature fusion of the converter transformer vibration signal. Further, the memory 72 may also include both an internal storage unit and an external storage device of the feature extraction device for the multi-feature fusion of the converter transformer vibration signal. The memory 72 is used to store the computer program and other programs and data required by the feature extraction device for the multi-feature fusion of the converter transformer vibration signal. The memory 72 may also be used to temporarily store data that has been output or is to be output.

[0153] Embodiment 3:

[0154] This embodiment provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the method described in Embodiment 1 are implemented.

[0155] The computer-readable medium shown can be any device that can contain, store, communicate, propagate or transmit the program for use in an instruction execution system, device or apparatus or in conjunction with these instruction execution systems, devices or apparatus. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection portion with one or more wirings (electronic devices), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and editable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium can even be paper or other suitable media on which the program can be printed, such as by optically scanning the paper or other medium, then editing, interpreting or processing in other suitable ways as necessary to obtain the program electronically, and then storing it in the computer memory.

[0156] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable ordinary technicians in the field to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made based on the essence of the content of the present invention should be included in the protection scope of the present invention.

Claims

1. A feature extraction method for multi-feature fusion of converter transformer vibration signal, characterized in that: include: Obtaining the original vibration signal of the converter transformer, and decomposing the original signal into several basic mode components; Calculate the dimensionless indexes of the first A basic mode components respectively, and obtain the time domain statistical characteristics of the corresponding basic mode components; Perform continuous wavelet transform on the first A basic mode components respectively to obtain the wavelet coefficient matrix corresponding to the basic mode components, and calculate the entropy value of the wavelet coefficient matrix to obtain the wavelet entropy feature corresponding to the basic mode components; Extract the fractal dimension features of the first A basic mode components respectively to obtain the fractal dimension features of the corresponding basic mode components; The obtained time domain statistics features, wavelet entropy features, and fractal dimension features of each basic mode component are combined to form a feature matrix; The principal component analysis method is used to reduce the dimension of the feature matrix and obtain the fused feature vector to realize the feature extraction of the converter transformer vibration signal. Perform fuzzy C-means clustering analysis on the fused feature vector, compare the real label with the classification label to judge the effect of feature extraction, and realize the detection of the operating status of the converter transformer; Wherein, A is a positive integer; Decomposing the original signal into a plurality of basic mode components comprises: Take the original signal x(t) as the signal to be processed, calculate all local extreme points of the signal, use the cubic spline interpolation method to fit all the maximum and minimum points respectively and record them as the upper and lower envelopes of the signal, calculate the average of the upper and lower envelopes to get the mean sequence m(t); Subtract the mean sequence m(t) from the original signal x(t) to obtain: h1(t)=x(t)-m(t) Check whether h1(t) meets the set conditions; If it is not satisfied, h1(t) is used as the signal to be processed and the above operation is repeated until the condition is satisfied. At this time, the mean sequence h n (t) is recorded as a basic mode component c1(t) h n (t)=c1(t) Decompose the first basic mode component from the original signal to obtain the residual sequence r1(t) r1(t)=x(t)-c1(t) By repeating the above operation with the remaining sequence r1(t) as the signal to be processed, the second, third, and finally Ath basic mode components can be obtained in sequence until the remaining sequence meets the stopping criterion, thereby decomposing the original signal into several basic mode components and one residual component; n is a positive integer.

2. The feature extraction method for multi-feature fusion of converter transformer vibration signal according to claim 1, characterized in that: The set conditions include: (1) The number of extreme points in the entire data sequence is the same as the zero-crossing point or the difference is no more than 1; (2) The mean of the upper and lower envelopes determined by the local extreme values ​​of the signal is 0.

3. The feature extraction method for multi-feature fusion of converter transformer vibration signal according to claim 1, characterized in that: The stopping criterion is set to limit the standard deviation S between two consecutive processing results. d Where T represents the time span of the signal, h (k-1) (t),h k (t) is the time series of two consecutive processing results in the process of screening basic components.

4. The feature extraction method for multi-feature fusion of converter transformer vibration signal according to claim 1, characterized in that: Said N is a positive integer five.

5. The feature extraction method for converter transformer vibration signal multi-feature fusion according to claim 1, characterized in that: The dimensionless index includes a waveform index, a peak index, a margin index, a pulse index and a kurtosis index.

6. The feature extraction method for converter transformer vibration signal multi-feature fusion according to claim 4, characterized in that: Four fractal dimension calculation algorithms are used to extract the fractal dimension features of the first five basic mode components respectively, and the Katz, Castiglioni, Sevcik and Higuchi fractal dimension features of the sequence are calculated.

7. The feature extraction method for converter transformer vibration signal multi-feature fusion according to claim 1, characterized in that: The principal component analysis method is used to reduce the dimension of the feature matrix, which includes: Normalization Where x(i) is the signal sequence, x'(i) is the normalized signal sequence, mean is the mean of the signal, N is the number of discrete points in the sequence, and σ is the variance; Sample feature covariance matrix calculation: Solve the reduced-dimensional matrix according to the singular value decomposition: [U,S,V]=svd[Cov] Where svd represents the singular value decomposition of the matrix, and the obtained U matrix is ​​called the dimension reduction matrix. U is obtained by arranging all the eigenvectors of Cov in descending order; Dimensionality reduction calculation: k is the number of rows of the dimension reduction matrix selected to control the dimension of the output.

8. A feature extraction device for multi-feature fusion of converter transformer vibration signals, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images