Power transformer on-line monitoring method based on signal enhancement and graph domain transformation

Abstract

The invention discloses a power transformer on-line monitoring method based on signal enhancement and graph domain transformation, and belongs to the technical field of transformer monitoring, and the method comprises the following steps: S1, obtaining an original time domain signal of equipment; s2, purifying the equipment voiceprint signal; s3, establishing a graph domain database; and S4, carrying out identification and early warning on the equipment. The method provided by the invention overcomes the limitation in the conventional online monitoring method, can effectively sense the change degree of the working state of the transformer equipment, and has higher accuracy and reliability. Through the voiceprint signal enhancement technology, the pure voiceprint signal of the power transformer can be effectively extracted, the domain transformation processing is carried out, and the pure voiceprint signal is mapped into an image form for analysis. The method not only can accurately detect the fault state of the transformer, but also can identify and classify the fault type.

Description

Technical Field

[0001] The invention belongs to the technical field of transformer monitoring, and in particular relates to an online monitoring method for a power transformer based on signal enhancement and image domain transformation. Background Art

[0002] Power transformers are reliable power control and protection equipment. The reliability and safety of their operation directly affect the safety and stability of the entire power system. According to relevant statistics, winding failure is the main cause of transformer equipment accidents. Mechanical defects such as windings are often long-term and cumulative. Transformers with potential defects may have normal electrical performance in the early stages. Conventional methods (such as gas dissolution analysis in oil, infrared temperature detection, etc.) are usually difficult to achieve reliable monitoring, but potential mechanical defects always threaten the winding's short-circuit resistance and turn-to-turn insulation. If allowed to develop, it is very likely to cause other faults and eventually lead to serious accidents. Therefore, in order to improve the level of equipment operation and maintenance and ensure the safe and reliable operation of the power grid, it is extremely necessary to monitor the mechanical status of power transformers.

[0003] Due to the large number and wide distribution of power transformers, the detection data reflecting their operating status also presents massive and multimodal characteristics. The overall voiceprint data set has obvious nonlinear characteristics. Traditional signal analysis methods can only realize the extraction of acoustic signal features in complex environments based on certain known prior information, which is not conducive to intelligent data processing. In view of this, it is necessary to provide an online monitoring method for power transformers based on signal enhancement and image domain transformation. On the basis of improving the safety of equipment operation and maintenance detection based on non-contact voiceprint signal detection technology, the improved spectral subtraction method is used to obtain the pure operating sound signal of the monitored equipment, and a voiceprint image database of transformers in different states with high recognition is established based on the one-dimensional signal-two-dimensional image spectral domain transformation method. The accurate identification of the operating status of the power transformer is realized by relying on advanced classification and recognition algorithms, and finally the purpose of improving the stability of power system operation is achieved. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and to provide an online monitoring method for power transformers based on signal enhancement and image domain transformation, which solves the technical problem that the prior art cannot accurately perceive the operating status of the transformer and timely warn of potential defects.

[0005] The object of the present invention is achieved by: a method for online monitoring of a power transformer based on signal enhancement and image domain transformation, comprising the following steps:

[0006] S1. Obtain the original time domain signal of the device;

[0007] S2. Purify the device voiceprint signal;

[0008] S3, establish a graph domain database;

[0009] S4. Identify and warn the equipment.

[0010] Furthermore, the original time domain signal of the acquisition device is acquired through a voiceprint sensor, and the acquired original transformer signal is framed and windowed to obtain a time domain signal x(τ) of each frame.

[0011] Furthermore, the device voiceprint signal purification is achieved by spectral subtraction optimized by Hilbert transform; after obtaining the time domain signal x(τ) of each frame, the Hilbert transform is introduced to calculate the amplitude spectrum and phase spectrum of each frame signal, and the Hilbert transform is implemented as follows:

[0012]

[0013] In the formula, y(t) represents the transformed signal, t is time, and PV represents the Cauchy principal value, that is, the limit is taken after removing the neighborhood of the singular point t = τ in the integral;

[0014] For a real signal x(t), its analytical signal z(t) is defined as:

[0015] z(i)=x(i)+j·y(t)

[0016] In the formula, represents an imaginary unit;

[0017] Then there are calculation formulas for amplitude spectrum and phase spectrum:

[0018]

[0019]

[0020] Wherein, ∠(arctan) represents the inverse tangent function.

[0021] Furthermore, the amplitude spectrum |N(k)| of the noise signal is estimated by the averaging method, and the amplitude spectrum |X(k)| calculated by the device voiceprint signal of each frame and the noise amplitude spectrum |N(k)| are spectrally subtracted to obtain the noise-reduced speech amplitude spectrum |Y(k)|:

[0022] |Y(k)|=max{|X(k)|-α|N(k)|,β|X(k)|}

[0023] In the formula, α is the over-reduction factor, which is used to control the amount of noise reduction, and β is the residual factor, which is used to prevent excessive subtraction of the effective voiceprint.

[0024] Furthermore, α is between 0.5 and 2, and β is between 0.01 and 0.1; preferably, α=0.5, β=0.03.

[0025] Furthermore, the Hilbert transform domain signal |Y(k)| of the denoised voiceprint signal is constructed using the amplitude spectrum |Y(k)| of the denoised voiceprint signal and the phase spectrum ∠|X(k)| of the original voiceprint signal:

[0026] Y(k)=|Y(k)|e j∠X(k)

[0027] The Hilbert transform domain signal of the denoised voiceprint signal is inversely Hilbert transformed to obtain the denoised speech time domain signal y(n), which is then overlapped and windowed to obtain the final pure transformer voiceprint signal.

[0028] Furthermore, the establishment of the graph domain database includes realizing accurate mapping of the one-dimensional signal of the pure transformer to the two-dimensional graph spectrum based on the graph spectrum domain transformation method of deep mining of the time domain signal state information, and defining the graph model as:

[0029] G = {P, E, W}

[0030] In the formula, P represents the set of vertices, which is used to describe different sampling points; E represents the set of connections, which is used to describe the connections between different nodes; W represents the weighted adjacency matrix; Based on the adjacency matrix, the graph Laplacian matrix L is obtained, which can better reflect the information parameters of each vertex in the graph model: P :

[0031] L P =D ∑ -W

[0032] Where D ∑ Represents the diagonal matrix of the graph model, whose diagonal elements are the number of edges connecting each vertex;

[0033] The graph transformation is expressed as:

[0034]

[0035] in, It is the result of Fourier transform of the graph, f is the original graph signal, U is the matrix composed of the eigenvectors of the graph Laplacian matrix, the superscript "T" represents the transpose, and then the one-dimensional result of the Fourier transform of the graph is mapped to a two-dimensional matrix to form a graph spectrum.

[0036] Furthermore, the identification and early warning of the equipment includes using a learning deep autoencoder to realize the spectrum identification and classification of different working conditions of the transformer equipment, and the implementation formula is:

[0037]

[0038] In the formula, x is the input data, W1 and W2 are weight matrices, b1 and b2 are bias vectors, g1 and g2 are activation functions, which are used to add nonlinear factors so that the self-encoder has nonlinear mapping learning ability and is used to learn features; with the help of deep automatic encoders with strong computational training robustness, accurate identification and early warning of the operating status of transformer equipment can be achieved.

[0039] Beneficial effects of the present invention: The present invention proposes an online monitoring method for power transformers based on signal enhancement and image domain transformation, which overcomes the limitations of conventional online monitoring methods, can effectively perceive the degree of change in the working state of the transformer equipment, and has higher accuracy and reliability. Through the voiceprint signal enhancement technology, the pure voiceprint signal of the power transformer can be effectively extracted, and the image domain transformation processing is performed, and the pure voiceprint signal is mapped into an image form for analysis. This method can not only accurately detect the fault state of the transformer, but also identify and classify the fault type. At the same time, the present invention also provides corresponding storage media and electronic devices to facilitate the implementation of the online monitoring method. The storage medium and device have the advantages of strong real-time performance and simple operation, are suitable for a wide range of power transformer monitoring fields, and have high application value and economic benefits. The present invention has efficient calculation and accurate results. The obtained two-dimensional map contains rich transformer equipment status information, which provides reliable data and technical support for the status identification and defect warning of the transformer equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the device framework of the present invention;

[0041] Figure 2 is a schematic diagram of the third and fourth aspects of the present invention;

[0042] Figure 3 It is a schematic flow chart of the method of the present invention;

[0043] Figure 4 It is a schematic diagram of the denoising effect of the original voiceprint data and the purified data of the present invention;

[0044] Figure 5 It is the normal transformer and mechanical abnormal diagram transformation mapping spectrum comparison of the present invention;

[0045] Figure 6 It is a confusion matrix diagram of the present invention. DETAILED DESCRIPTION

[0046] The present invention is further described in detail below in conjunction with the accompanying drawings. It should be noted that all directional words such as up, down, front, back, left, and right that appear in the present invention do not limit the present invention, but are only for a clearer description and explanation of the present invention. The following are only preferred specific embodiments of the present invention, and the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, shall be equivalently replaced or changed according to the technical solution and its conception of the present invention, and all shall be covered within the protection scope of the present invention.

[0047] like Figure 1-6 As shown, the present invention discloses an online monitoring method for a power transformer based on signal enhancement and image domain transformation, a storage medium and an online monitoring system device.

[0048] The online monitoring method of power transformers based on signal enhancement and domain transformation includes the following steps: S1, obtaining the original time domain signal of the equipment; S2, purifying the voiceprint signal of the equipment; S3, establishing a domain database; S4, identifying and warning the equipment. First, the original time domain signal of the power transformer equipment is collected by a high-precision voiceprint sensor, and then the transformer voiceprint signal is purified under a complex environment by means of spectral subtraction based on Hilbert transform optimization; secondly, the domain transformation method is introduced to realize the precise mapping of the one-dimensional time domain signal to the two-dimensional spectrum, and then the domain transformation two-dimensional spectrum of the monitored equipment under different states is obtained, and a domain database is established to provide data support for subsequent state identification; finally, the accurate identification and reliable warning of the operating status of the transformer equipment are realized by means of a deep autoencoder with high computational efficiency and strong training robustness.

[0049] The specific steps are as follows:

[0050] First, the collected original transformer signal is framed and windowed to obtain the time domain signal x(τ) of each frame, and the Hilbert transform with high calculation accuracy is introduced to realize the calculation of the amplitude spectrum and phase spectrum of each frame signal. The Hilbert transform is implemented as follows:

[0051]

[0052] Where y(t) represents the transformed signal, t is time, and PV represents the Cauchy principal value, that is, the limit is taken after removing the neighborhood of the singular point t = τ in the integral.

[0053] For a real signal x(t), its analytical signal z(t) is defined as:

[0054] z(i)=x(i)+j·y(t)

[0055] In the formula, Represents an imaginary unit.

[0056] Then there are calculation formulas for amplitude spectrum and phase spectrum:

[0057] Amplitude spectrum:

[0058]

[0059] Phase spectrum:

[0060]

[0061] Wherein, ∠(arctan) represents the inverse tangent function.

[0062] For better results, the amplitude spectrum |N(k)| of the noise signal is estimated by the averaging method, and the amplitude spectrum |X(k)| calculated from the device voiceprint signal of each frame and the noise amplitude spectrum |N(k)| are spectrally subtracted to obtain the noise-reduced speech amplitude spectrum |Y(k)|:

[0063] |Y(k)|=max{|X(k)|-α|N(k)|,β|X(k)|}

[0064] In the formula, α is an over-reduction factor, which is used to control the amount of noise reduction, and β is a residual factor, which is used to prevent excessive reduction of the effective soundprint. Generally, α is between 0.5 and 2, and β is between 0.01 and 0.1. According to the characteristics of the on-site environment, α=0.5 and β=0.03 in the present invention.

[0065] For better results, the Hilbert transform domain signal |Y(k)| of the denoised voiceprint signal is constructed using the amplitude spectrum of the denoised voiceprint signal |Y(k)| and the phase spectrum ∠|X(k)| of the original voiceprint signal:

[0066] Y(k)=|Y(k)|e j∠X ( k)

[0067] Then, the Hilbert transform domain signal of the denoised voiceprint signal is inversely Hilbert transformed to obtain the denoised speech time domain signal y(n), which is overlapped and windowed to obtain the final pure transformer voiceprint signal.

[0068] Secondly, based on the spectral domain transformation method of deep mining of time domain signal state information, the precise mapping of the one-dimensional signal of the pure transformer to the two-dimensional spectrum is realized. First, a graph model is defined:

[0069] G = {P, E, W}

[0070] In the formula, P represents the set of vertices, which is used to describe different sampling points; E represents the set of connections, which is used to describe the connections between different nodes; W represents the weighted adjacency matrix; Based on the adjacency matrix, the graph Laplacian matrix L is obtained, which can better reflect the information parameters of each vertex in the graph model: P :

[0071] L P =D ∑ -W

[0072] Where D ∑ Represents the diagonal matrix of the graph model, whose diagonal elements are the number of edges connecting each vertex.

[0073] The spectrum transformation can be expressed as follows:

[0074]

[0075] in, It is the result of Fourier transform of the graph, f is the original graph signal, U is the matrix composed of the eigenvectors of the graph Laplacian matrix, the superscript "T" represents the transpose, and then the one-dimensional result of the Fourier transform of the graph is mapped to a two-dimensional matrix to form a graph spectrum.

[0076] Finally, a deep autoencoder with excellent learning performance is used to realize the spectrum recognition and classification of different working conditions of transformer equipment. The implementation formula is as follows:

[0077]

[0078] In the formula, x is the input data, W1 and W2 are weight matrices, b1 and b2 are bias vectors, and g1 and g2 are activation functions, which are used to add nonlinear factors so that the self-encoder has nonlinear mapping learning capabilities and can learn more complex features. With the help of deep autoencoders with high computational efficiency and strong training robustness, accurate identification and reliable early warning of the operating status of transformer equipment can be achieved.

[0079] like Figure 1-6 As shown, the present invention discloses an online monitoring method for a power transformer based on signal enhancement and image domain transformation, a storage medium and an online monitoring system device.

[0080] The invention discloses an online monitoring system device for power transformer based on signal enhancement and image domain transformation, which specifically comprises: a data perception module, a data acquisition module, a data processing module and a result display module.

[0081] In this embodiment, the data perception module is used to collect the acoustic signals of the target device, and the data acquisition module is used to process the original acoustic signal data, specifically including collecting the acoustic signals obtained by the data perception module and amplifying the signals, ADC acquisition and signal storage, and transmitting the data to the data analysis module.

[0082] In this embodiment, the online monitoring of the overheating fault of the power transformer is based on a high-performance infrared thermal imager, which monitors the transformer in all directions, locates the target, amplifies it, and conducts comprehensive observation and analysis; a microcomputer is used as the monitoring core, and a set of infrared image analysis and processing software is developed in the monitoring center, and the noise on the infrared image is eliminated by using the theory of image processing. The processed image is used by the host computer fault diagnosis system, combined with the temperature information in the infrared red image database, and the transformer state is diagnosed and predicted through comprehensive analysis. Among them, the selection of the shooting position of the infrared thermal imager is based on the principle of covering as many monitoring points or areas as possible, but it cannot be too far away from the power transformer, otherwise it will affect the temperature measurement accuracy. A reasonable monitoring distance should be selected without affecting safety. In the online monitoring of overheating faults, all signals and data are transmitted through the network. First, the infrared image signals from each monitoring point in the substation are converted into digital images with 25 frames / second by the network video server, and MPEG-4 image compression is completed. The compressed video data stream is sent to the image monitoring center through the network. The monitoring computer in the monitoring center decompresses the image data received from the front end and displays it on the monitoring computer.

[0083] Infrared image denoising: In the process of acquiring infrared images, they will inevitably be affected by the shooting instrument and the surrounding environment, causing the acquired images to be distorted or contain various noises, making it difficult to distinguish the target. In order to have a more accurate basis for subsequent target recognition, it is necessary to filter and denoise the acquired infrared images in the early stage.

[0084] Mean filtering is used for denoising; mean filtering is a method that uses the average grayscale value of all pixels in a pixel neighborhood to replace the pixel value. This method is also called neighborhood averaging and is a linear filtering method.

[0085] Suppose the input image is f(x, y), which is an M*N matrix, h(x, y) is the low-pass impulse response, which is an L*L matrix, and the low-pass output result after convolution processing is an M*N matrix.

[0086]

[0087] The low-pass impulse response h(x, y) can select templates as needed: four-neighborhood template, eight-neighborhood template, 5x5 or 7×7 neighborhood template. Considering the problem of operation speed. Here, the selected Gauss template can be processed into a horizontal template and a vertical template multiplied. All pixels in the image are averaged with the surrounding 8 points, so that the pixels with more drastic changes can be removed and the noise can be eliminated.

[0088] Median filtering can also be used for denoising. Median filtering is a nonlinear filtering technology based on sorting statistics theory. Its basic principle is to sort the neighborhood pixels according to the size of the grayscale level and take the median value instead of the original grayscale value of the pixel to be processed. After the image is processed, the edge and detail information of the image are taken into account while eliminating noise. The output of the two-dimensional median filter is:

[0089] g(x, y)=median{f(xk, yl), (k, l)∈W}

[0090] Where f(x, y) and g(x, y) are the original image and the processed image respectively. W is a window, which contains an odd number of pixels, usually a 3×3, 5×5 area. If its size is an even number, the median can be taken as the average of the two middle pixels.

[0091] Steps of median filtering;

[0092] 1) Make the template roam on the image and make its center coincide with a certain pixel in the image;

[0093] 2) Read the grayscale values ​​of all pixels covered by the template:

[0094] 3) Arrange the grayscale values ​​read in the previous step from small to large;

[0095] 4) Find the median of all gray values;

[0096] 5) Assign the median value in the previous step to the pixels in the image that coincide with the center of the template.

[0097] The processing object of median filtering is those pixels whose grayscale values ​​are quite different from those of the surrounding pixels, so that they take the median value of the pixels in their own neighborhood, which is very effective in eliminating isolated noise in the image. The degree of noise elimination of median filtering image processing depends on two factors: the size of the filter (which determines the neighborhood range) and the number of pixels involved in finding the median of the pixel. In general, objects that are bright or dark and have an area smaller than half of the filter can be filtered out, while objects with a larger area larger than half of the filter will be left behind. Therefore, taking into account the filtering effect, the selection of the template size should be based on actual needs.

[0098] In this embodiment, the data processing module is used to process and calculate acoustic data. First, the spectral subtraction optimized by Hilbert transform is introduced to purify the original transformer collected voiceprint signal. Then, a graph model is constructed and the graph Laplace matrix is ​​defined. On the basis of the above, the graph Fourier transform method is used to realize the graph signal conversion of the one-dimensional pure transformer signal, and it is mapped into a two-dimensional graph spectrum to establish an abnormal state warning graph spectrum set for transformer equipment; finally, based on a deep encoder with high computational efficiency and strong training robustness, accurate identification and reliable warning of the operating status of the transformer equipment are realized.

[0099] In this embodiment, the storage medium unit is used to store a computer program, and the program is used by the data processing unit to execute the relevant steps of the online monitoring method proposed in this article. The result display module is used to display the graph transformation matrix, graph transformation spectrum, and state identification and warning results of the current state of the transformer equipment.

[0100] Secondly, the online monitoring method of power transformers based on signal enhancement and graph domain transformation uses the spectral subtraction method optimized by Hilbert transform to purify the original transformer collected voiceprint signal; constructs a graph model, defines the graph Laplace matrix, and uses the graph Fourier transform method to realize the graph signal conversion of one-dimensional pure transformer signals, and maps it into a two-dimensional graph spectrum; based on a deep encoder with high computational efficiency and strong training robustness, accurate identification and reliable early warning of the operating status of transformer equipment are achieved.

[0101] In a third aspect, a computer device is provided, comprising a storage module and a processing module, wherein the storage module stores a computer program, and the processing module implements the transformer device status monitoring method described in any one of the second aspects when executing the computer program.

[0102] In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processing module, the transformer device status monitoring method described in any one of the second aspects is implemented. Figure 2 .

[0103] In order to highlight the superiority of the method of the present invention, a group of transformer measured data was subjected to denoising. Figure 4 It can be seen that the spectral energy of the transformer's normal operation voiceprint signal is mainly concentrated within 500 Hz, and the proposed optimization method achieves the effective purification of the transformer's operation voiceprint signal.

[0104] like Figure 5 The figure shows the graph transformation mapping spectrum under the two states of the transformer. Figure 5It can be seen that the energy distribution of the transformer diagram transformation mapping spectrum under normal conditions is relatively uniform and symmetrical, while the spectrum under mechanical abnormal conditions shows thin and extremely uneven energy distribution, with energy mainly concentrated on the left. Obviously, the diagram can realize reliable identification and distinction between normal and abnormal conditions of transformers.

[0105] Introduce confusion matrix to quantitatively evaluate classification effect;

[0106] like Figure 6 The table shows the confusion matrix calculation results obtained using the proposed method. Figure 6 The diagonal lines of the confusion matrix in respectively represent the number of correctly identified samples for each state using the proposed method. Figure 6 From the table, we can see that only 2 of the 600 test sets were misclassified, and the classification accuracy of the test sample set was 99.67%. From the table, we can see that the average state recognition accuracy, recall rate and F1 value of the proposed method are 0.99, 0.99 and 0.99 respectively. This result shows that the proposed method has good classification performance, proving the effectiveness and reliability of the proposed method.

[0107] Parameters / Tags 1 2 average value Accuracy 99.34％ 100％ 99.67％ Recall 100％ 99.33％ 99.67％ F1 value 99.67％ 99.67％ 99.67％

[0108] Compared with the existing conventional technologies, the present invention provides an online monitoring method for power transformers based on signal enhancement and image domain transformation, a storage medium and an online monitoring system device. The main advantages include: the online monitoring method for power transformers based on signal enhancement and image domain transformation proposed by the present invention overcomes the limitations of conventional online monitoring methods, can effectively perceive the degree of change in the working state of transformer equipment, and has higher accuracy and reliability. Through the voiceprint signal enhancement technology, the pure voiceprint signal of the power transformer can be effectively extracted, and the image domain transformation processing is performed, and the pure voiceprint signal is mapped into an image form for analysis. This method can not only accurately detect the fault state of the transformer, but also identify and classify the fault type. At the same time, the present invention also provides a corresponding storage medium and an electronic device to facilitate the implementation of the online monitoring method. The technology has the advantages of strong real-time performance and simple operation, is applicable to a wide range of power transformer monitoring fields, and has high application value and economic benefits. In summary, the present invention has efficient calculation and accurate results, and the obtained two-dimensional map contains rich transformer equipment status information, which provides reliable data and technical support for the status identification and defect warning of transformer equipment.

Claims

1. A method for online monitoring of power transformers based on signal enhancement and image domain transformation, characterized in that: The following steps are involved: S1. Obtain the original time domain signal of the device; S2. Purify the device voiceprint signal; S3, establish a graph domain database; S4. Identify and warn the equipment.

2. The method for online monitoring of power transformers based on signal enhancement and image domain transformation according to claim 1 is characterized in that: The original time domain signal of the acquisition device is acquired through a voiceprint sensor, and the acquired original transformer signal is framed and windowed to obtain a time domain signal x(τ) of each frame.

3. The method for online monitoring of power transformers based on signal enhancement and image domain transformation according to claim 2 is characterized in that: The device voiceprint signal purification is achieved by spectral subtraction optimized by Hilbert transform; after obtaining the time domain signal x(τ) of each frame, the Hilbert transform is introduced to calculate the amplitude spectrum and phase spectrum of each frame signal. The Hilbert transform is implemented as follows: In the formula, y(t) represents the transformed signal, t is time, and PV represents the Cauchy principal value, that is, the limit is taken after removing the neighborhood of the singular point t = τ in the integral; For a real signal x(t), its analytical signal z(t) is defined as: z(t)=x(t)+j·y(t) In the formula, represents an imaginary unit; Then there are calculation formulas for amplitude spectrum and phase spectrum: Wherein, ∠(arctan) represents the inverse tangent function.

4. The method for online monitoring of power transformers based on signal enhancement and image domain transformation according to claim 1 is characterized in that: The amplitude spectrum |N(k)| of the noise signal is estimated by the averaging method, and the amplitude spectrum |X(k)| calculated from the device voiceprint signal of each frame and the noise amplitude spectrum |N(k)| are spectrally subtracted to obtain the noise-reduced speech amplitude spectrum |Y(k)|: |Y(k)|=max{|X(k)|-α|N(k)|,β|X(k)|} In the formula, α is the over-reduction factor, which is used to control the amount of noise reduction, and β is the residual factor, which is used to prevent excessive subtraction of the effective voiceprint.

5. The method for online monitoring of power transformers based on signal enhancement and image domain transformation according to claim 4 is characterized in that: α is between 0.5 and 2, and β is between 0.01 and 0.

1.

6. The method for online monitoring of power transformers based on signal enhancement and image domain transformation according to claim 5 is characterized in that: α＝0.5，β＝0.03。 7. The method for online monitoring of power transformers based on signal enhancement and image domain transformation according to claim 6 is characterized in that: The Hilbert transform domain signal |Y(k)| of the denoised voiceprint signal is constructed using the amplitude spectrum |Y(k)| of the denoised voiceprint signal and the phase spectrum ∠|X(k)| of the original voiceprint signal: Y(k)=|Y(k)|e j∠X(k) The Hilbert transform domain signal of the denoised voiceprint signal is inversely Hilbert transformed to obtain the denoised speech time domain signal y(n), which is then overlapped and windowed to obtain the final pure transformer voiceprint signal.

8. The method for online monitoring of power transformers based on signal enhancement and image domain transformation according to claim 1 is characterized in that: The establishment of the graph domain database includes realizing accurate mapping of the one-dimensional signal of the pure transformer to the two-dimensional graph spectrum based on the graph spectrum domain transformation method of deep mining of the time domain signal state information, and defining the graph model as: G={P,E,W} In the formula, P represents the set of vertices, which is used to describe different sampling points; E represents the set of connections, which is used to describe the connections between different nodes; W represents the weighted adjacency matrix; Based on the adjacency matrix, the graph Laplacian matrix L is obtained, which can better reflect the information parameters of each vertex in the graph model: P : L P =D Σ -W Where D Σ Represents the diagonal matrix of the graph model, whose diagonal elements are the number of edges connecting each vertex; The spectral transformation is expressed as: in, It is the result of Fourier transform of the graph, f is the original graph signal, U is the matrix composed of the eigenvectors of the graph Laplacian matrix, the superscript "T" represents the transpose, and then the one-dimensional result of the Fourier transform of the graph is mapped to a two-dimensional matrix to form a graph spectrum.

9. The method for online monitoring of power transformers based on signal enhancement and image domain transformation according to claim 1 is characterized in that: The identification and early warning of the equipment includes using a learning deep autoencoder to realize the spectrum identification and classification of different working conditions of the transformer equipment, and the implementation formula is: In the formula, x is the input data, W1 and W2 are weight matrices, b1 and b2 are bias vectors, g1 and g2 are activation functions, which are used to add nonlinear factors so that the self-encoder has nonlinear mapping learning ability for learning features; With the help of a deep autoencoder with strong computational training robustness, accurate identification and early warning of the operating status of transformer equipment can be achieved.

Images