Method and system for power equipment diagnosis based on windowed feature and Hilbert visualization

Abstract

A method and a system for power equipment diagnosis based on windowed feature and Hilbert visualization are provided, which belong to the field of power equipment fault diagnosis. The method includes: obtaining an original data set of monitoring data containing power equipment fault features; introducing windowed feature calculation considering logarithmic constraints to process data to obtain a feature sequence; using Hilbert visualization method for further processing to obtain a Hilbert image data set used to train and verify a convolutional neural network; and finally directly inputting newly obtained test sample data after windowed feature calculation and Hilbert visualization processing into the trained network for fault diagnosis and location. The disclosure uses windowed feature calculation and Hilbert visualization to process the monitoring data of a power equipment to fully extract fault features and effectively improve diagnostic accuracy, and uses the convolutional neural network for diagnosis to improve the intelligence of diagnosis.

Claims

1. A method for power equipment diagnosis based on windowed feature and Hilbert visualization, comprising: step 1 of obtaining a plurality of sets of monitoring data comprising different fault locations, fault types, and severity of a power equipment, analyzing each set of the monitoring data, obtaining a corresponding fault status label, and using each set of the monitoring data and the con.sup.-esponding fault status label to constitute a target data set, wherein the fault status label comprises the fault type and the fault location; step 2 of performing windowed feature calculation on the target data set considering logarithmic constraints, extracting fault features contained therein, and obtaining a feature sequence data set; step 3 of performing Hilbert visualization processing on each set of data in the feature sequence data set, and obtaining a Hilbert image data set; step 4 of constructing a fault diagnosis and location model, dividing the Hilbert image data set into a training set and a verification set, training the fault diagnosis and location model using the training set, and verifying the trained fault diagnosis and location model using the verification set; and step 5 of obtaining a final diagnosis result from the trained fault diagnosis and location model after performing the same windowed feature calculation and Hilbert visualization on diagnosis data.

2. The method according to claim 1, wherein the plurality of sets of the monitoring data of the power equipment is data.sub.1={a.sub.i,1, a.sub.i,2, . . . , a.sub.i,j, . . . , a.sub.i,N, s.sub.i}i∈[1, K], where K is K sets of the monitoring data of the power equipment, a.sub.i,j is a monitoring value corresponding to a j(j∈[1, N]) -th point in an i -th set of the monitoring data of the power equipment, s.sub.i is the fault status label of the power equipment corresponding to the i -th set of the monitoring data of the power equipment, and N is a number of monitoring data points.

3. The method according to claim 2, wherein step 2 comprises: taking a monitoring curve of the power equipment in a normal state as a reference curve, and performing the windowed feature calculation considering the logarithmic constraints based on the reference curve to obtain a feature sequence data set, wherein a feature value is a self-defined evaluation indicator or an evaluation indicator provided by different research work.

4. The method according to claim 3 , wherein step 2 comprises: step 2.1 of defining a window with a specified length W.sub.w, wherein the window scans a frequency band step by step with a step size s from a starting point x.sub.0 of a scanning frequency band range to an end x.sub.n, each time the window moves one step, the feature value is calculated within a frequency region of the length W.sub.w of the window, and a finally obtained feature sequence length is $N^{\prime }=\mathrm{floor}\left(\frac{\left(x_n-x_0\right)-W_w}{s}\right)+1,$ where s and W.sub.w, are positive integer multiples of a shortest interval between monitoring points, all monitoring data is to be scanned, thus W.sub.w≥2s , floor represents a round-down function, and returns to a largest previous integer; step 2.2 of determining an x-coordinate of a symmetry axis of a k(k=1, 2, . . . , N′)-th window from $x_k^c=x_0+\frac{W_w}{2}+\left(k-1\right).Math.s,$ determining an x-coordinate of a left boundary from $x_k^{\mathrm{left}}=x_k^c-\frac{W_w}{2},$ determining an x-coordinate of a right boundary from $x_k^{\mathrm{right}}=x_k^c+\frac{W_w}{2},$ and determining a number of points contained in each window from $n_w=\frac{W_w}{\Delta f}+1,$ where Δf is an interval between two monitoring points; and step 2.3 of obtaining a windowed feature value considering the logarithmic constraints from the x-coordinate of the symmetry axis, the x-coordinate of the left boundary, the x-coordinate of the right boundary, and the number of points contained in each window based on a selected feature value evaluation indicator, so as to obtain the feature sequence data set.

5. The method according to claim 4, wherein a feature sequence data obtained after the windowed feature calculation is data′.sub.i={b.sub.i,1, b.sub.i,2, . . . , b.sub.i,j, . . . b.sub.i,N′, s.sub.i}i∈[1,K], where K is K sets of the feature sequence data, b.sub.i,j is a value corresponding to a j(j∈[1,N′])-th point in an i-th set of the feature sequence data, s.sub.i is a power equipment status corresponding to the i-th set of the feature sequence data, and N′ is a number of feature sequence points.

6. The method according to claim 5, wherein step 3 comprises: arranging sweep frequency response feature sequence data according to an extension direction of a Hilbert curve, expressing values of a feature sequence in different colors, and finally converting the feature sequence into an image, wherein a selected number of data points of the Hilbert curve must be greater than or equal to a number of data points of the feature sequence, when a number of Hilbert points is greater, subsequent points may be used to continue to draw a feature sequence once or set to zero, as long as an operation of an entire feature sequence data set is guaranteed to be the same.

7. The method according to claim 6, further comprising performing the Hilbert visualization processing on each set of the data in the feature sequence data set, and obtaining the Hilbert image data set, wherein the Hilbert image data set is expressed as data″.sub.i={c.sub.i,s.sub.i}i∈[1,K], where K is K sets of Hilbert image data, c.sub.i is a Hilbert image converted from an i-th set of data parameters in the feature sequence data set, and s.sub.i is a power equipment status corresponding to an i-th set of the Hilbert image data.

8. The method according to claim 2, wherein step 2 comprises: step 2.1 of defining a window with a specified length W.sub.w, wherein the window scans a frequency band step by step with a step size s from a starting point x.sub.0 of a scanning frequency band range to an end x.sub.n, each time the window moves one step, the feature value is calculated within a frequency region of the length W.sub.w of the window, and a finally obtained feature sequence length is $N^{\prime }=\mathrm{floor}\left(\frac{\left(x_n-x_0\right)-W_w}{s}\right)+1,$ where s and W.sub.w are positive integer multiples of a shortest interval between monitoring points, all monitoring data is to be scanned, thus W.sub.w≥2s , floor represents a round-down function, and returns to a largest previous integer; step 2.2 of determining an x-coordinate of a symmetry axis of a k(k=1, 2, . . . , N′)-th window from $x_k^c=x_0+\frac{W_w}{2}+\left(k-1\right).Math.s,$ determining an x-coordinate of a left boundary from $x_k^{\mathrm{left}}=x_k^c-\frac{W_w}{2},$ determining an x-coordinate of a right boundary from $x_k^{\mathrm{right}}=x_k^c+\frac{W_w}{2},$ and determining a number of points contained in each window from $n_w=\frac{W_w}{\Delta f}+1,$ where Δf is an interval between two monitoring points; and step 2.3 of obtaining a windowed feature value considering the logarithmic constraints from the x-coordinate of the symmetry axis, the x-coordinate of the left boundary, the x-coordinate of the right boundary, and the number of points contained in each window based on a selected feature value evaluation indicator, so as to obtain the feature sequence data set.

9. The method according to claim 8, wherein a feature sequence data obtained after the windowed feature calculation is data.sub.i={b.sub.i,1, b.sub.i,2, . . . , b.sub.i,j, . . . b.sub.i,N′, s.sub.i}i ∈[1, K], where K is K sets of the feature sequence data, b.sub.i,j is a value corresponding to a j(j∈[1, N′])-th point in an i-th set of the feature sequence data, s.sub.i is a power equipment status corresponding to the i-th set of the feature sequence data, and N′ is a number of feature sequence points.

10. The method according to claim 9, wherein step 3 comprises: arranging sweep frequency response feature sequence data according to an extension direction of a Hilbert curve, expressing values of a feature sequence in different colors, and finally converting the feature sequence into an image, wherein a selected number of data points of the Hilbert curve must be greater than or equal to a number of data points of the feature sequence, when a number of Hilbert points is greater, subsequent points may be used to continue to draw a feature sequence once or set to zero, as long as an operation of an entire feature sequence data set is guaranteed to be the same.

11. The method according to claim 10, further comprising performing the Hilbert visualization processing on each set of the data in the feature sequence data set, and obtaining the Hilbert image data set, wherein the Hilbert image data set is expressed as data″.sub.i={c.sub.i, s.sub.i}i∈[1, K], where K is K sets of Hilbert image data, c.sub.i is a Hilbert image converted from an i-th set of data parameters in the feature sequence data set, and s.sub.i is a power equipment status corresponding to an i-th set of the Hilbert image data.

12. The method according to claim 1, wherein step 4 comprises: constructing a basic neural network model when a number of the Hilbert image data sets is greater than or equal to a preset number threshold, and training and verifying using the Hilbert image data set; and importing a neural network model pre-trained using a large image database when the number of the Hilbert image data sets is less than the preset number threshold, replacing a last layer of the imported neural network model, so that a number of outputs thereof is equal to a number of diagnostic label categories, then freezing a part of parameters of the pre-trained neural network model, and inputting the Hilbert image data set into different pre-trained neural network models for transfer training and verification.

13. A system for power equipment diagnosis based on windowed feature and Hilbert visualization, comprising a processor configured to: obtaining a plurality of sets of monitoring data containing different fault locations, fault types, and severity of a power equipment, analyzing each set of the monitoring data, obtaining a corresponding fault status label, and using each set of the monitoring data and the corresponding failure status label to constitute a target data set, wherein the fault status label comprises the fault type and the fault location; performing windowed feature calculation on each set of data in the target data set considering logarithmic constraints, extracting fault features contained therein, and obtaining a feature sequence data set; performing Hilbert visualization processing on each set of data in the feature sequence data set, and obtaining a Hilbert image data set; constructing a fault diagnosis and location model, divide the Hilbert image data set into a training set and a verification set, training the fault diagnosis and location model using the training set, and verifying the trained fault diagnosis and location model using the verification set; and performing the same windowed feature calculation and Hilbert visualization on diagnosis data in a subsequent diagnosis process, and then obtaining a final diagnosis result from the trained fault diagnosis and location model.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 is a schematic diagram of a flowchart of a transformer depth vision fault location method considering windowed feature calculation and Hilbert visualization according to an embodiment of the disclosure.

(2) FIG. 2 is a schematic diagram of division of transformer fault locations according to an embodiment of the disclosure.

(3) FIG. 3 is a schematic diagram of an equivalent circuit of a transformer winding according to an embodiment of the disclosure.

(4) FIG. 4 is a schematic diagram of a process of the windowed feature calculation and the Hilbert visualization according to an embodiment of the disclosure.

(5) FIG. 5 is a schematic diagram of a system structure according to an embodiment of the disclosure.

DETAILED DESCRIPTION OF DISCLOSED EMBODIMENTS

(6) In order for the objectives, technical solutions, and advantages of the disclosure to be more comprehensible, the disclosure is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the disclosure, but not to limit the disclosure. In addition, the technical features involved in the various embodiments of the disclosure described below may be combined with each other as long as there is no conflict therebetween.

(7) The disclosure provides a method for power equipment diagnosis based on windowed feature and Hilbert visualization, and takes fault location of sweep frequency response analysis of a transformer as specific embodiments for description. However, the disclosure is not only applicable to the fault location of the sweep frequency response analysis of a transformer winding, but may also be extended to other fields of diagnosis and location.

(8) First, an original data set of sweep frequency response containing fault features of a transformer is obtained. Then, windowed feature calculation considering logarithmic constraints is introduced to process original data, and a feature sequence is obtained. Graphical features of the feature sequence are further highlighted using a Hilbert visualization method. A Hilbert image data set is obtained, and is used for transfer training and verification of a deep convolutional neural network. Finally, newly obtained sweep frequency response data may be directly input into the trained network for diagnosis after performing the windowed feature calculation and Hilbert visualization processing. A location result is obtained. The disclosure converts one-dimensional frequency response data into a two-dimensional Hilbert image, which provides a new way of expression for a sweep frequency response fingerprint. In addition, the logarithmic constraints are introduced when performing the feature calculation, which reduces the influence of extreme values on the image and highlights the graphical features of the feature sequence. Finally, the model may obtain higher accuracy.

(9) As shown in FIG. 3, the fault location of a power transformer winding is taken as an example. FIG. 1 shows a schematic diagram of a flowchart of the method for power equipment diagnosis based on windowed feature and Hilbert visualization according to the embodiment of the disclosure, which includes the following steps.

(10) S1: Several sets of monitoring data of each transformer sweep frequency response are obtained. Each set of the monitoring data is analyzed. A corresponding fault status label is obtained. Each set of the monitoring data and the corresponding fault status label are used to constitute a target data set. The fault status label includes the fault type and the fault location.

(11) In the embodiment of the disclosure, the data may be taken from FR curve data containing complex fault features obtained based on PSpice simulation and the corresponding transformer fault status label thereof.

(12) In the embodiment of the disclosure, the several sets of the monitoring data of the transformer sweep frequency response in Step S1 may be expressed as:
data.sub.i={a.sub.i,1,a.sub.i,2, . . . ,a.sub.i,j, . . . ,a.sub.i,N,s.sub.i}i∈[1,K]

(13) where K is K sets of sweep frequency response monitoring data, a.sub.i,j is the response value corresponding to a j(j∈[1, N])-th point frequency in an i-th set of dissolved gas monitoring data, s.sub.i is the transformer fault status label (including the fault type and the fault location) corresponding to the i-th set of the sweep frequency response monitoring data, and N is the number of scanning frequency points.

(14) In the embodiment of the disclosure, as shown in FIG. 2, transformer fault locations (statuses) are divided into 8 types, which are respectively the normal state and faults occurring in seven types of different locations, corresponding to labels 0 to 7. Different faults may occur at the same location, but the corresponding labels are the same. Specific data is as shown in Table

(15) TABLE-US-00001 TABLE 1 Data set of sweep frequency response analysis of transformer winding Synchronous Simultaneous fault of asynchronous inductance L and Capacitor Capacitor fault of Label mutual inductance Cd fault Cg fault the three Sum 0 20 20 20 20 80 1 20 20 20 20 80 2 20 20 20 20 80 3 20 20 20 20 80 4 20 20 20 20 80 5 20 20 20 20 80 6 20 20 20 20 80 7 20 20 20 20 80 Total 160 160 160 160 640

(16) S2: Windowed feature calculation is performed on the target data set considering logarithmic constraints to obtain a feature sequence data set.

(17) In Step S2, as shown in FIG. 4, a sweep frequency response curve (such as a FR fingerprint curve) of the transformer in the normal state needs to be taken as a reference, and the calculation of a feature value is performed based thereon. The feature value may be an evaluation indicator of the FR curve provided by different research work, such as a correlation coefficient (CC), a root mean square error (RMSE), a mean absolute percentage error (MAPE), etc. For example, the calculation formula of MAPE after adding the logarithmic constraints is:

(18) ${\mathrm{MAPE}}^{*}=\lg \frac{\underset{k=1}{\overset{n_w}{.Math.}}.Math.\frac{y_k^o-y_k^m}{y_k^o}.Math.\times 100}{n_w},$
where MAPE* refers to MAPE introduced with the logarithmic constraints, y.sub.k.sup.o is the y-coordinate of a point on the FR fingerprint curve in a k-th window obtained, y.sub.k.sup.m is the x-coordinate of a point on other FR curves, and n.sub.w is the number of points in each window.

(19) When performing the windowed feature calculation considering the logarithmic constraints, a window with a specified length W.sub.w is first defined. The window scans a frequency band step by step with a step size s from a starting point x.sub.0 of a scanning frequency band range to an end x.sub.n. Each time the window moves one step, the feature value is calculated within a frequency region of the length W.sub.w of the window. The finally obtained feature sequence length is

(20) $N^{\prime }=\mathrm{floor}\left(\frac{\left(x_n-x_0\right)-W_w}{s}\right)+1,$
where s and W.sub.w are positive integer multiples of an FR scanning frequency interval. The entire FR frequency band is to be scanned, thus W.sub.w≥2s. floor represents a round-down function, and returns to the largest previous integer. An x-coordinate of a symmetry axis of a k(k=1, 2, . . . , N′)-th window is

(21) $x_k^c=x_0+\frac{W_w}{2}+\left(k-1\right).Math.s.$
An x-coordinate of a left boundary is

(22) 0$x_k^{\mathrm{left}}=x_k^c-\frac{W_w}{2}.$
An x-coordinate of a right boundary is

(23) $x_k^{\mathrm{right}}=x_k^c+\frac{W_w}{2}.$
A number of points contained in each window is

(24) $n_w=\frac{W_w}{\Delta f}+1,$
where Δf is the scanning frequency interval.

(25) the feature sequence data obtained after the windowed feature calculation is data.sub.i′={b.sub.i,1, b.sub.i,2, . . . , b.sub.i,j, . . . , b.sub.i,N, s.sub.i} i∈[1, K], where K is K sets of sweep frequency response feature sequence data, b.sub.i,j is the value corresponding to a j(j∈[1, N′])-th point in an i-th set of the sweep frequency response feature sequence data, s.sub.i is the transformer status corresponding to the i-th set of the sweep frequency response feature sequence data, and N′ is the number of feature sequence points.

(26) S3: The Hilbert visualization processing is performed on each set of data in the feature sequence data set, and a Hilbert image data set is obtained.

(27) In Step S3, as shown in FIG. 4, a 6th order Hilbert curve is selected for data visualization. Since the selected number of data points of the Hilbert curve must be greater than or equal to the number of data points of the feature sequence in order to fully present the feature sequence in the image, when the number of Hilbert points is greater, the subsequent points may be used to continue to draw a feature sequence once or set to zero, as long as the operation of the entire feature sequence data set is guaranteed to be the same. In the embodiment, the method of continuing to draw the feature sequence once is adopted.

(28) S4: A convolutional neural network (CNN) fault location model is constructed. The Hilbert image data set is divided into a training set and a verification set. The CNN fault location model is trained using the training set. A trained CNN diagnosis model is verified using the verification set.

(29) As an optional implementation, when the number of Hilbert image data sets is greater than or equal to a preset number threshold, it indicates that the amount of data is large enough to construct the most basic CNN diagnosis model, and the CNN diagnosis model is trained and verified using big data. When the number of Hilbert image data sets is less than or equal to the preset number threshold, it indicates that the amount of data is smaller, transfer learning may be used to perform transfer training and verification on classic models, such as ResNet50. The training set and the verification set are divided into 8:2.

(30) The preset number threshold may be determined according to actual requirements, which is not uniquely limited by the embodiments of the disclosure.

(31) In the embodiment of the disclosure, other models may also be chosen as the CNN diagnosis model, which is not uniquely limited by the embodiments of the disclosure.

(32) As an optional implementation, Step S4 uses transfer learning to perform transfer training and verification on the classic model ResNet50. Firstly, the last fully connected layer of ResNet50 is replaced, so that the number of outputs thereof is equal to the number of diagnostic label categories, which is 8 in the embodiment. Then, the parameters of the first 50% layers of the pre-trained network are frozen. Finally, hyperparameters of the transfer training based on the ResNet50 model are set as shown in Table 3. The training set is input train the unfrozen part of the network, and the verification set is used to verify.

(33) TABLE-US-00002 TABLE 3 Hyperparameter setting of CNN diagnostic model Hyperparameter name Setting Initial gradient 0.00006 Minimum batch 100 Number of iterations 150 Learning plan Segmented Starting drop location of learning rate 100 Drop factor of learning rate 0.2 Optimizer ‘adam’

(34) Finally, the location accuracy of the fault location model of the transformer winding according to the embodiment is 94.53%.

(35) FIG. 5 shows a schematic diagram of a structure of a transformer depth vision fault location system considering windowed feature calculation and Hilbert visualization according to an embodiment of the disclosure, which includes the following.

(36) A data processing module 201 is configured to obtain several sets of monitoring data containing different fault locations, fault types, and severity of a power equipment, analyze each set of the monitoring data, obtain a corresponding fault status label, and use each set of the monitoring data and the corresponding failure status label to constitute a target data set. The fault status label includes the fault type and the fault location.

(37) A feature extraction module 202 is configured to perform windowed feature calculation on each set of data in the target data set considering logarithmic constraints, extract fault features contained therein, and obtain a feature sequence data set.

(38) A visualization module 203 is configured to perform Hilbert visualization processing on each set of data in the feature sequence data set, and obtain a Hilbert image data set.

(39) A network training module 204 is configured to construct a fault diagnosis and location model, divide the Hilbert image data set into a training set and a verification set, train the fault diagnosis and location model using the training set, and verify the trained fault diagnosis and location model using the verification set.

(40) A diagnosis module 205 is configured to directly input newly obtained sweep frequency response data through the feature extraction module and the visualization module into a trained network for diagnosis in a subsequent diagnosis process, so as to obtain a final fault location result.

(41) For the specific implementation of each module, reference may be made to the description of the foregoing embodiment for the method, which will not be reiterated in the embodiment of the disclosure.

(42) In another embodiment of the disclosure, a computer-readable storage medium stored with a computer program is also provided. The computer program implements the method for power equipment diagnosis based on windowed feature and Hilbert visualization in the foregoing embodiment when executed by a processor.

(43) It should be pointed out that according to implementation requirements, each step/component described in the disclosure may be split into more steps/components, or two or more steps/components or partial operations of steps/components may be combined into a new step/component to implement the objective of the disclosure.

(44) The above method according to the disclosure may be implemented in hardware and firmware, implemented as software or computer codes that may be stored in a recording medium (such as compact disc read-only memory (CD ROM), random access memory (RAM), floppy disk, hard disk, or magneto-optical disk), or implemented as computer codes that are originally stored in a remote recording medium or a non-transitory machine-readable medium downloaded through the network and will be stored in a local recording medium, so that the method may be processed by such software stored on a recording medium using a general-purpose computer, a special-purpose processor, or a programmable or special-purpose hardware (such as application specific integrated circuit (ASIC) or field programmable gate array (FPGA)). It may be understood that a computer, a processor, a microprocessor, a controller, or a programmable hardware includes a storage component (for example, RAM, ROM, flash memory, etc.) that may store or receive the software or computer codes. When the software or computer codes are accessed and executed by a computer, a processor, or hardware, the processing method is implemented. Furthermore, when a general-purpose computer accesses codes for implementing the processing, the execution of the codes converts the general-purpose computer into a special-purpose computer for executing the processing.

(45) Persons skilled in the art may easily understand that the above descriptions are only preferred embodiments of the disclosure and are not intended to limit the disclosure. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the disclosure should be contained within the protection scope of the disclosure.