Fault classification detection method and device based on transformer vibration characteristics and time-frequency diagram

Abstract

The invention provides a fault classification detection method and device based on transformer vibration characteristics and a time-frequency diagram, and belongs to the technology of transformer fault classification detection, and the method comprises the steps: collecting the real-time vibration data of each part of a transformer; establishing a real-time vibration signal matrix according to the real-time vibration data; extracting frequency characteristics of the real-time vibration signal matrix by adopting a variational mode decomposition model to obtain a real-time frequency characteristic curve; transforming the real-time frequency characteristic curve by adopting Hilbert transform to obtain a time-frequency diagram of the real-time vibration signal; and inputting the time-frequency diagram of the real-time vibration signal into a pre-trained support vector machine fault diagnosis model to obtain a fault classification result, extracting an amplitude value by adopting the time-frequency diagram, calculating a fault similarity coefficient, and further determining whether the fault classification result is correct or not. According to the invention, rapid, accurate and reliable diagnosis of the transformer fault is realized.

Description

Technical Field

[0001] The present invention belongs to the technology of transformer fault classification detection, and specifically relates to a fault classification detection method and device based on transformer vibration characteristics and time-frequency diagram. Background Art

[0002] As one of the indispensable and important equipment in the power system, the transformer plays a vital role in the power transmission and distribution process. However, due to long-term operation and the influence of external environmental factors, the transformer may have various faults, such as core damage, winding breakage, oil tank leakage, cooler failure, etc. These faults will not only affect the normal operation of the transformer, but may also cause further damage to the equipment or even accidents. Therefore, it is very important to detect and diagnose the transformer faults in a timely and accurate manner.

[0003] Traditional transformer fault detection methods mainly rely on manual inspections and offline testing, which have disadvantages such as low detection accuracy, long cycle and high cost. With the development of Internet of Things, big data and artificial intelligence technology, fault classification detection methods based on transformer vibration signals and time-frequency images have gradually attracted widespread attention. As one of the important indicators reflecting the operating status of equipment, vibration signals can intuitively reflect the internal operating conditions of the equipment, so they are widely used in the fields of fault detection and health monitoring. Time-frequency diagram is an image representation method that combines time domain and frequency domain information. It can intuitively show the changes of signals over time and frequency, so it is widely used in the analysis and processing of vibration signals. However, the existing technical methods have a large amount of calculation, which is difficult to meet the needs of real-time transformer detection, and the accuracy is relatively low. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present application proposes a fault classification detection method and device based on transformer vibration characteristics and time-frequency diagrams.

[0005] In the first aspect, the present application proposes a fault classification detection method based on transformer vibration characteristics and time-frequency diagrams, comprising:

[0006] Step S1: collecting real-time vibration data of each part of the transformer;

[0007] Step S2: establishing a real-time vibration signal matrix according to the real-time vibration data;

[0008] Step S3: extracting the frequency characteristics of the real-time vibration signal matrix using a variational mode decomposition model to obtain a real-time frequency characteristic curve;

[0009] Step S4: transforming the real-time frequency characteristic curve using Hilbert transform to obtain a time-frequency diagram of the real-time vibration signal;

[0010] Step S5: inputting the time-frequency diagram of the real-time vibration signal into a pre-trained support vector machine fault diagnosis model to obtain a fault classification result and a predicted fault time-frequency diagram;

[0011] Step S6: Obtain a historical fault time-frequency diagram according to the fault classification result, and extract n first amplitude values ​​from the historical fault time-frequency diagram;

[0012] Step S7: extracting n second amplitude values ​​from the predicted fault time-frequency diagram;

[0013] Step S8: calculating a fault similarity coefficient according to the n first amplitude values ​​and the n second amplitude values;

[0014] Step S9: If the fault similarity coefficient is greater than or equal to the similarity threshold of the fault corresponding to the fault classification result, it is determined that the transformer has a fault corresponding to the fault classification result;

[0015] Step S10: If the fault similarity coefficient is less than the similarity threshold of the fault corresponding to the fault classification result, it is determined that the transformer does not have the fault corresponding to the fault classification result, and the process returns to step S1 to continue detection.

[0016] The real-time vibration data of each part of the transformer is collected, including:

[0017] Collect real-time vibration data of the core;

[0018] Collect real-time vibration data of windings;

[0019] Collect real-time vibration data of the fuel tank;

[0020] Collect real-time vibration data of chillers.

[0021] The real-time vibration data of the iron core is collected, including collecting real-time vibration data at the following positions: the intersection of the iron core column corresponding to the A-phase high and low voltage winding and the upper iron yoke, the intersection of the iron core column corresponding to the A-phase high and low voltage winding and the lower iron yoke, the intersection of the iron core column corresponding to the B-phase high and low voltage winding and the upper iron yoke, the intersection of the iron core column corresponding to the B-phase high and low voltage winding and the lower iron yoke, the intersection of the iron core column corresponding to the C-phase high and low voltage winding and the upper iron yoke, and the intersection of the iron core column corresponding to the C-phase high and low voltage winding and the lower iron yoke;

[0022] The real-time vibration data of the windings are collected, including collecting real-time vibration data at the following positions: one-half and five-sixths of the A-phase high-voltage winding, one-half and five-sixths of the A-phase low-voltage winding, one-half and five-sixths of the B-phase high-voltage winding, one-half and five-sixths of the B-phase low-voltage winding, one-half and five-sixths of the C-phase high-voltage winding, and one-half and five-sixths of the C-phase low-voltage winding;

[0023] The collecting of the real-time vibration data of the oil tank includes collecting the real-time vibration data at the following positions: the top center position of the oil tank and the bottom center position of the oil tank;

[0024] The collecting of the real-time vibration data of the cooler includes collecting the real-time vibration data at the following positions: the connection between the cooler support and the transformer oil tank, and at the cooling fan of the cooler.

[0025] The fault similarity coefficient is calculated as follows:

[0026]

[0027] Among them, C p The fault similarity coefficient corresponding to the pth fault, n is the total number of the first amplitude value or the second amplitude value, A x is the xth first amplitude value, B x is the x-th second amplitude value.

[0028] The n first amplitude values ​​characterize the frequency distribution of the historical fault transformer signal and the amplitude variation characteristics in different time periods, and draw a scatter plot of the historical fault amplitude values, wherein the scatter plot of the historical fault amplitude values ​​is defined as the scatter plot characteristics of the amplitude values ​​of the time-frequency diagram of different fault types when the transformer is in historical normal operation, and the scatter plot characteristics include time, frequency and amplitude values;

[0029] The n second amplitude values ​​characterize the frequency distribution of the predicted fault transformer signal and the amplitude variation characteristics in different time periods, and draw a scatter plot of the predicted fault amplitude values, wherein the scatter plot of the predicted fault amplitude values ​​is defined as the scatter plot characteristics of the amplitude values ​​of the time-frequency diagram of different fault types when the transformer operates normally in real time, and the scatter plot characteristics include time, frequency and amplitude values.

[0030] The variational mode decomposition model, the construction process includes:

[0031] Collect historical vibration data of each part of the transformer;

[0032] Establishing a historical vibration signal matrix according to the historical vibration data;

[0033] The variational modal decomposition algorithm is used to decompose the historical vibration signal matrix into multiple modal components, and the multiple modal components are arranged from low frequency to high frequency to obtain the intrinsic mode function;

[0034] Constructing a constraint model of the variational mode decomposition model based on the intrinsic mode functions;

[0035] Lagrange multipliers and penalty factors are introduced to solve the constraint model of the variational mode decomposition model to obtain an optimal solution, which includes: Wiener filtering, the power center of gravity of the intrinsic mode function and Lagrange multipliers. The optimal solution is used as the final variational mode decomposition model.

[0036] The frequency characteristics of the real-time vibration signal matrix are extracted by using a variational modal decomposition model to obtain a real-time frequency characteristic curve, including:

[0037] Step S3.1: Initialize the Wiener filter, the power center of gravity of the intrinsic mode function and the Lagrange multiplier to obtain initialization data;

[0038] Step S3.2: updating the initialization data according to the variational mode decomposition model;

[0039] Step S3.3: If the convergence condition is met, the corresponding k-th decomposed intrinsic mode function is output;

[0040] Step S3.4: If the convergence condition is not met, the number of iterations is increased by 1, and the process returns to step S3.2 to continue updating;

[0041] Step S3.5: taking the k decomposed intrinsic mode functions as the frequency characteristics of the real-time vibration signal matrix, and drawing a corresponding real-time frequency characteristic curve according to the frequency characteristics of the real-time vibration signal matrix.

[0042] The real-time frequency characteristic curve is transformed by Hilbert transform to obtain a time-frequency diagram of the real-time vibration signal, including:

[0043] Perform Hilbert transform on the k decomposed intrinsic mode functions to obtain the Hilbert transform result;

[0044] Performing an inverse Hilbert transform on the Hilbert transform result to obtain an inverse Hilbert transform result;

[0045] Analyze the inverse Hilbert transform result to obtain the instantaneous amplitude, instantaneous phase and instantaneous frequency of the vibration signal;

[0046] A time-frequency diagram of the real-time vibration signal is obtained according to the instantaneous amplitude, instantaneous phase and instantaneous frequency of the vibration signal.

[0047] The pre-trained support vector machine fault diagnosis model construction process includes:

[0048] A variational mode decomposition model is used to extract the frequency characteristics of the historical vibration signal matrix to obtain a historical frequency characteristic curve;

[0049] The historical frequency characteristic curve is transformed by using Hilbert transform to obtain a time-frequency diagram of the historical vibration signal;

[0050] Taking the time-frequency diagram of historical vibration signals as input and the fault classification results as output, the classification function of the support vector machine and the constraint condition equation with slack variables are constructed;

[0051] According to Lagrange's theorem, the constraint equation with slack variables is transformed into Lagrangian form;

[0052] Solving the Lagrangian form to obtain a dual form of the constraint equation with slack variables;

[0053] The classification function of the support vector machine is updated according to the dual form to obtain an optimal classification decision function, and the optimal classification decision function is used as a pre-trained support vector machine fault diagnosis model.

[0054] In the second aspect, the present application proposes a fault classification detection device based on transformer vibration characteristics and time-frequency diagrams, comprising:

[0055] The vibration signal acquisition module is connected to the transformer and the data acquisition card respectively, and is used to collect the real-time vibration data and historical vibration data of each part of the transformer, and send the real-time vibration data and historical vibration data to the data acquisition card.

[0056] A data acquisition card, used for storing real-time vibration data and historical vibration data, and sending the real-time vibration data and historical vibration data to a fault detection module;

[0057] A preliminary fault detection module is used to establish a real-time vibration signal matrix according to the real-time vibration data; extract the frequency characteristics of the real-time vibration signal matrix using a variational mode decomposition model to obtain a real-time frequency characteristic curve; transform the real-time frequency characteristic curve using a Hilbert transform to obtain a time-frequency diagram of the real-time vibration signal; input the time-frequency diagram of the real-time vibration signal into a pre-trained support vector machine fault diagnosis model to obtain a fault classification result and a predicted fault time-frequency diagram;

[0058] A fault determination module is used to obtain a historical fault time-frequency diagram corresponding to the fault classification result, extract n first amplitude values ​​from the historical fault time-frequency diagram; extract n second amplitude values ​​from the predicted fault time-frequency diagram; calculate a fault similarity coefficient based on the n first amplitude values ​​and the n second amplitude values; if the fault similarity coefficient is greater than or equal to a similarity threshold of the fault corresponding to the fault classification result, determine that the transformer has a fault corresponding to the fault classification result; if the fault similarity coefficient is less than a similarity threshold of the fault corresponding to the fault classification result, determine that the transformer has not a fault corresponding to the fault classification result, and return to the vibration signal acquisition module to continue detection.

[0059] Beneficial effects:

[0060] This application proposes a fault classification detection method and device based on transformer vibration characteristics and time-frequency diagrams to solve the problems of low detection accuracy, long cycle, high cost, etc. of traditional transformer fault detection methods. By obtaining transformer structural vibration data, extracting frequency characteristics of vibration signals, decomposing vibration signals, extracting time-frequency diagrams, and establishing SVM fault diagnosis models, a fast, accurate, and reliable diagnosis of transformer faults can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Flow chart of the fault classification detection method based on transformer vibration characteristics and time-frequency diagram in the embodiment of the present application;

[0062] Figure 2 A schematic flow chart of a fault classification detection method based on transformer vibration characteristics and time-frequency diagrams according to an embodiment of the present application;

[0063] Figure 3 Schematic diagram of the core vibration measurement points in the embodiment of the present application;

[0064] Figure 4 Schematic diagram of high voltage winding measurement points in an embodiment of the present application;

[0065] Figure 5 Schematic diagram of low voltage winding measurement points in an embodiment of the present application;

[0066] Figure 6 Original signal diagram of transformer short circuit fault operation in the embodiment of the present application;

[0067] Figure 7 Original signal diagram of the transformer load operation in the embodiment of the present application;

[0068] Figure 8 The original signal diagram of the transformer no-load operation in the embodiment of the present application;

[0069] Fig. 9 Original frequency characteristic diagram of the transformer in the embodiment of the present application when it is in short circuit fault operation;

[0070] Fig.10 Frequency characteristic diagram of transformer load operation in an embodiment of the present application;

[0071] Fig.11 Frequency characteristic diagram of the transformer in no-load operation according to the embodiment of the present application;

[0072] Fig.12 Time-frequency diagram of transformer short-circuit fault operation in the embodiment of the present application;

[0073] Fig.13 The time-frequency diagram of the transformer winding fault operation in the embodiment of the present application;

[0074] Fig.14 Time-frequency diagram of the transformer running at no load in the embodiment of the present application

[0075] Fig.15 Amplitude value characteristic diagram of transformer short circuit fault operation in the embodiment of the present application;

[0076] Fig.16 A characteristic diagram of the amplitude value when the transformer is under load in the embodiment of the present application;

[0077] Fig.17 A characteristic diagram of the amplitude value of the transformer when it is running at no load according to the embodiment of the present application;

[0078] Fig.18 Schematic diagram of a fault classification detection device based on transformer vibration characteristics and time-frequency diagram according to an embodiment of the present application;

[0079] Fig.19 A block diagram of the internal principles of the vibration signal acquisition module in an embodiment of the present application. DETAILED DESCRIPTION

[0080] The specific implementation of the present application is further described in detail below in conjunction with the drawings and examples.

[0081] Traditional transformer fault detection methods have problems such as low detection accuracy, long cycle, and high cost. In order to solve the above problems, this application proposes a fault classification detection method and device based on transformer vibration characteristics and time-frequency diagrams. By performing time-frequency analysis on the transformer vibration signal, the frequency distribution characteristics and time domain variation law of the signal can be effectively extracted, which provides an important basis for fault detection and diagnosis. The variational mode decomposition algorithm (VMD) is a method based on signal adaptive decomposition, which can decompose the signal into multiple intrinsic mode functions (IMFs), each IMF represents the vibration characteristics of the signal in different frequency ranges. Through the application of the VMD algorithm, various frequency components in the vibration signal can be captured more accurately, and the accuracy and reliability of fault diagnosis can be improved. Support vector machine (SVM) is a supervised learning model based on statistical learning theory, which can effectively process high-dimensional data and nonlinear classification problems, and has been widely used in the field of pattern recognition and data mining. By training the SVM classifier, the time-frequency diagram features can be used to classify the vibration signal and identify the category of the transformer fault, which provides important technical support for the timely diagnosis of the transformer fault.

[0082] In summary, a fault classification detection method and device based on transformer vibration characteristics and time-frequency diagram is an advanced fault detection method that combines time-frequency analysis of vibration signals with machine learning technology. It has the advantages of high detection accuracy, fast response speed, and low cost. It is of great significance to improve the accuracy and reliability of transformer fault detection and diagnosis.

[0083] Example 1

[0084] This embodiment proposes a fault classification detection method based on transformer vibration characteristics and time-frequency diagrams, such as Figure 1 , Figure 2 As shown, including:

[0085] Step S1: collecting real-time vibration data of each part of the transformer;

[0086] Step S2: establishing a real-time vibration signal matrix according to the real-time vibration data;

[0087] Step S3: extracting the frequency characteristics of the real-time vibration signal matrix using a variational mode decomposition model to obtain a real-time frequency characteristic curve;

[0088] Step S4: transforming the real-time frequency characteristic curve using Hilbert transform to obtain a time-frequency diagram of the real-time vibration signal;

[0089] Step S5: inputting the time-frequency diagram of the real-time vibration signal into a pre-trained support vector machine fault diagnosis model to obtain a fault classification result and a predicted fault time-frequency diagram;

[0090] In this embodiment, the pre-trained support vector machine fault diagnosis model is trained using historical frequency characteristic curves and time-frequency diagrams of historical vibration signals.

[0091] Step S6: obtaining a fault time-frequency diagram according to the fault classification result, and extracting n first amplitude values ​​from the fault time-frequency diagram;

[0092] Step S7: extracting n second amplitude values ​​from the predicted fault time-frequency diagram;

[0093] Step S8: calculating a fault similarity coefficient according to the n first amplitude values ​​and the n second amplitude values;

[0094] Step S9: If the fault similarity coefficient is greater than or equal to the similarity threshold of the fault corresponding to the fault classification result, it is determined that the transformer has a fault corresponding to the fault classification result;

[0095] Step S10: If the fault similarity coefficient is less than the similarity threshold of the fault corresponding to the fault classification result, it is determined that the transformer does not have the fault corresponding to the fault classification result, and the process returns to step S1 to continue detection.

[0096] In step S1, the real-time vibration data of each part of the transformer is collected, including:

[0097] Collect real-time vibration data of the core;

[0098] Collect real-time vibration data of windings;

[0099] Collect real-time vibration data of the fuel tank;

[0100] Collect real-time vibration data of chillers.

[0101] In this embodiment, the core vibration mainly reflects the state and changes of the transformer core, the winding vibration mainly reflects the characteristics of the current load, insulation state and winding structure, the oil tank vibration mainly reflects the overall characteristics of the transformer, the cooler vibration mainly reflects the heat dissipation of the transformer, and the insulation vibration mainly reflects the state and changes of the insulating material.

[0102] The real-time vibration data of the iron core is collected, including collecting real-time vibration data at the following positions: the intersection of the iron core column corresponding to the A-phase high and low voltage winding and the upper iron yoke, the intersection of the iron core column corresponding to the A-phase high and low voltage winding and the lower iron yoke, the intersection of the iron core column corresponding to the B-phase high and low voltage winding and the upper iron yoke, the intersection of the iron core column corresponding to the B-phase high and low voltage winding and the lower iron yoke, the intersection of the iron core column corresponding to the C-phase high and low voltage winding and the upper iron yoke, and the intersection of the iron core column corresponding to the C-phase high and low voltage winding and the lower iron yoke, such as Figure 3 shown.

[0103] The real-time vibration data of the windings are collected, including collecting real-time vibration data at the following positions: one-half and five-sixths of the A-phase high-voltage winding, one-half and five-sixths of the A-phase low-voltage winding, one-half and five-sixths of the B-phase high-voltage winding, one-half and five-sixths of the B-phase low-voltage winding, one-half and five-sixths of the C-phase high-voltage winding, and one-half and five-sixths of the C-phase low-voltage winding. Figure 4 and Figure 5 shown.

[0104] The collecting of the real-time vibration data of the oil tank includes collecting the real-time vibration data at the following positions: the top center position of the oil tank and the bottom center position of the oil tank.

[0105] The collecting of the real-time vibration data of the cooler includes collecting the real-time vibration data at the following positions: the connection between the cooler support and the transformer oil tank, and at the cooling fan of the cooler.

[0106] In step S2, a real-time vibration signal matrix is ​​established according to the real-time vibration data;

[0107] In this embodiment, a vibration signal calculation matrix is ​​established, a real-time vibration signal matrix is ​​established for real-time vibration data, and a historical vibration signal matrix is ​​established for historical vibration data, and an original image of the vibration signal is drawn. The vibration data in step S1 is written in the form of a matrix, specifically:

[0108] E i =[e1,e2,e3…e y ],i=[1,2,3…n] (1)

[0109] Among them, E i represents the vibration signal matrix (in a specific implementation, it can represent the real-time vibration signal matrix or the historical vibration signal), the subscript i indicates that there are n vibration signal matrices in total, e y is the yth type of data in the vibration signal matrix. By setting the sampling frequency and the number of sampling points, the data in the vibration signal calculation matrix is ​​read and the original image of the vibration signal is drawn.

[0110] In step S3, before the frequency characteristics of the real-time vibration signal matrix are extracted by using the variational mode decomposition model, it is necessary to pre-establish the variational mode decomposition model based on the collected historical vibration data, and extract the frequency characteristics of the vibration signal calculation matrix through the variational mode decomposition algorithm, and draw it in the form of a two-dimensional image. The vibration signal calculation matrix established in step S3 will be further processed by the variational mode decomposition algorithm (Variational Mode Decomposition, VMD algorithm). The VMD algorithm has the characteristics of an adaptive decomposition method, which decomposes the vibration signal into multiple intrinsic mode functions (IMFs), each of which represents the vibration characteristics of the signal in different frequency ranges. Through the application of the VMD algorithm, various frequency components in the vibration signal can be captured more accurately.

[0111] The construction process of the variational mode decomposition model includes:

[0112] Step S100: collecting historical vibration data of each part of the transformer;

[0113] Step S101: establishing a historical vibration signal matrix according to the historical vibration data;

[0114] The process of step S100 and step S101 in this embodiment is consistent with step S1 and step S2, except that the historical vibration data stored in the data acquisition card is collected at this time. This embodiment will not be described in detail.

[0115] Step S102: Decomposing the historical vibration signal matrix into multiple modal components using a variational modal decomposition algorithm, and arranging the multiple modal components from low frequency to high frequency to obtain an intrinsic mode function;

[0116] In this embodiment, the obtained vibration signal calculation matrix is ​​decomposed into multiple modal components by an adaptive decomposition method, and arranged from low frequency to high frequency to obtain the intrinsic mode function (IMF):

[0117] u k (t) = A k (t)cos(φ k (t)) (2)

[0118] Among them, A k is the amplitude of the acquired signal, that is, the amplitude of the kth vibration signal in the historical vibration signal matrix, t is the time variable, φ k is the phase function, u k (t) is the amplitude modulation-frequency modulation function; at the same time, the phase function φ k With instantaneous frequency ω k (t) The following relationship exists:

[0119]

[0120] Step S103: constructing a constraint model of the variational mode decomposition model according to the intrinsic mode function;

[0121] In this embodiment, the constraint model of the VMD model is obtained as follows:

[0122]

[0123] Among them, f(t) is the original vibration signal of the transformer, {u k}={u1,...u K} is the k IMFs obtained after decomposing the original vibration signal of the transformer, {ω k}={ω1,...ω K} is the center frequency corresponding to each IMF; * is the convolution operation, and δ(t) is the impulse function.

[0124] Step S104: Introduce Lagrange multipliers and penalty factors, solve the constraint model of the variational mode decomposition model, and obtain an optimal solution. The optimal solution includes: Wiener filtering, the power center of gravity of the intrinsic mode function and Lagrange multipliers. The optimal solution is used as the final variational mode decomposition model.

[0125] In this embodiment, the constraint model is solved, and the Lagrange multiplier λ(t) and the penalty factor α are introduced to obtain the following solution:

[0126]

[0127] Among them, δ(t) is the impulse function, and the penalty factor α is used to ensure the accuracy of signal reconstruction during signal reconstruction. Through the above solution process, the optimal solution of the constraint model formula (4) is obtained. Among them, for the solution of the optimization problem, the center frequency and λ n+1 The operator is expressed as follows:

[0128]

[0129] In the above formula, The Wiener filter, is the power spectrum centroid of the corresponding intrinsic mode function (IMF); τ is the noise tolerance, m is the number of iterations, and λ m+1 is the operator of the Lagrange multiplication for the m+1th iteration.

[0130] The frequency characteristics of the real-time vibration signal matrix are extracted by using a variational modal decomposition model to obtain a real-time frequency characteristic curve, including:

[0131] Step S3.1: Initialize the Wiener filter, the power center of gravity of the intrinsic mode function and the Lagrange multiplier to obtain initialization data;

[0132] Step S3.2: updating the initialization data according to the variational mode decomposition model;

[0133] Step S3.3: If the convergence condition is met, the corresponding k-th decomposed intrinsic mode function is output;

[0134] Step S3.4: If the convergence condition is not met, m=m+1, and return to step S3.2 to continue updating;

[0135] Step S3.5: taking the k decomposed intrinsic mode functions as the frequency characteristics of the real-time vibration signal matrix, and drawing a corresponding real-time frequency characteristic curve according to the frequency characteristics of the real-time vibration signal matrix.

[0136] In this embodiment, the frequency characteristics of the transformer vibration signal calculation matrix are obtained according to the variational mode decomposition algorithm (VMD algorithm), and the corresponding specific steps are as follows:

[0137] (1) Initialization λ 1 ;

[0138] (2) According to formula (6), formula (7), and formula (8), initialize the data λ 1 Make updates;

[0139] (3) Repeat the above steps (1) and (2). During the iteration, m=m+1. When the following convergence condition is reached, the loop ends.

[0140]

[0141] Among them, ε is the judgment accuracy given in advance. The VMD algorithm is mainly implemented through the above three steps, and finally obtains k decomposed IMFs of the vibration signal, that is, obtains the frequency characteristics of the transformer vibration signal, and draws the frequency characteristic curve of the corresponding vibration signal calculation matrix after decomposition.

[0142] In step S5, the real-time frequency characteristic curve is transformed by Hilbert transform to obtain a time-frequency diagram of the real-time vibration signal, including:

[0143] Step S5.1: performing Hilbert transform on the k decomposed intrinsic mode functions to obtain a Hilbert transform result;

[0144] Step S5.2: performing an inverse Hilbert transform on the Hilbert transform result to obtain an inverse Hilbert transform result;

[0145] Step S5.3: Analyze the Hilbert inverse transform result to obtain the instantaneous amplitude, instantaneous phase and instantaneous frequency of the vibration signal;

[0146] Step S5.4: Obtain a time-frequency diagram of the real-time vibration signal according to the instantaneous amplitude, instantaneous phase and instantaneous frequency of the vibration signal.

[0147] In this embodiment, the frequency characteristic curve obtained in step S4 is Hilbert transformed by the Hilbert transform method to obtain the time-frequency diagram of the vibration signal. For the K decomposed intrinsic mode functions (IMFs) obtained in step S4, and the characteristic frequency curve after the vibration signal calculation matrix decomposition, the Hilbert transform method is used to represent each IMF as a time-frequency image of instantaneous frequency and instantaneous amplitude. This process converts the signal from the time domain to the time-frequency domain by performing a Hilbert transform on each IMF to visualize the characteristics of the frequency change of the vibration signal over time.

[0148] The AM-FM function u in step S4 is k (t), and perform Hilbert transform on it to obtain:

[0149]

[0150] Among them, Y(t) is the Hilbert transform result.

[0151] Performing the Hilbert inverse transform on the above formula yields:

[0152]

[0153] Finally, we get the amplitude modulation-frequency modulation function u k (t) The signal after analysis is:

[0154] Z(t)=u k (t)+iY(t)=a(t)e iθ(t) (12)

[0155] Among them, Z(t) is the analytical result, and a(t) represents the instantaneous amplitude of the vibration signal, which can be expressed as follows:

[0156]

[0157] θ(t) represents the instantaneous phase and can be expressed as follows:

[0158]

[0159] The instantaneous frequency is calculated as follows:

[0160]

[0161] After the above calculations, the Hilbert transform time-frequency diagram of the vibration signal is obtained.

[0162] In step S5, the time-frequency graph of the real-time vibration signal is input into the pre-trained support vector machine fault diagnosis model to obtain the fault classification result, establish the SVM fault diagnosis model, make a preliminary diagnosis of the transformer fault, and determine the fault category of the transformer. A support vector machine (SVM) fault diagnosis model is established to preliminarily classify and diagnose possible faults of the transformer. The vibration signal is classified by training the SVM classifier using the time-frequency graph features to identify the category of the transformer fault. The model is based on the time-frequency graph features of the vibration signal and uses a machine learning algorithm to perform fault diagnosis. By judging the similarity between the frequency characteristic curve and the time-frequency graph features of the transformer vibration signal and the known fault types, possible fault types are identified and classified. The classification method of the algorithm is: find a hyperplane that can separate the training data so that the interval between the samples of the two categories is maximized. First of all, the support vector machine fault diagnosis model needs to be pre-trained. The construction process of the pre-trained support vector machine fault diagnosis model includes:

[0163] Step S200: extracting the frequency characteristics of the historical vibration signal matrix using a variational mode decomposition model to obtain a historical frequency characteristic curve;

[0164] Step S201: transforming the historical frequency characteristic curve using Hilbert transform to obtain a time-frequency diagram of the historical vibration signal;

[0165] In this embodiment, step S200 and step S201 are consistent with step S3 and step S4, except that the processing object is the historical vibration signal matrix, which will not be repeated in this embodiment.

[0166] Step S202: using the time-frequency diagram of the historical vibration signal as input and the fault classification result as output, constructing a classification function of a support vector machine and a constraint condition equation with slack variables;

[0167] In this embodiment, the image is transformed from nonlinear to linear to linear by using a kernel function, and its classification function is:

[0168]

[0169] Among them, w T is the normal vector of the hyperplane, b is the intercept; the constraint equation is:

[0170]

[0171] ξ i is a slack variable that can satisfy any hyperplane. i The size of is controlled and expressed by the following constraint equation:

[0172]

[0173] Among them, C is the value that guarantees the optimal hyperplane is found. i Keep to a minimum.

[0174] Step S203: transforming the constraint condition equation with slack variables into Lagrangian form according to Lagrangian theorem;

[0175] In this embodiment, according to Lagrange's theorem, it is expressed by the following function:

[0176]

[0177] Step S204: Solve the Lagrangian form to obtain the dual form of the constraint condition equation with slack variables; solve the above equation to obtain:

[0178]

[0179] Where λ and α are Lagrange multipliers. Substituting equations (20), (21) and (22) into equation (19), we can obtain the dual form of equation (18):

[0180]

[0181] in, is the kernel function.

[0182] Step S205: improving the classification function of the support vector machine according to the dual form to obtain an optimal classification decision function, and using the optimal classification decision function as a pre-trained support vector machine fault diagnosis model.

[0183] After the above process, the optimal classification decision function of formula (16) is finally obtained as:

[0184]

[0185] According to the above SVM algorithm classification principle, the frequency characteristic curve and time-frequency diagram obtained by variational mode decomposition and Hilbert transform in steps four and five are used as the input of the SVM algorithm in step six for classification calculation.

[0186] In step S5, finally, the time-frequency diagram of the real-time vibration signal is input into the pre-trained support vector machine fault diagnosis model to obtain the fault classification result and the predicted fault time-frequency diagram. In the prior art, if the support vector machine fault diagnosis model is directly used, no matter how the support vector machine fault diagnosis model is adjusted, the final classification result still cannot achieve the expected effect, and the misjudgment rate is high, which affects the application in actual engineering. The present application adds a fault similarity coefficient judgment process to the support vector machine fault diagnosis model, so that the classification accuracy is significantly increased. In addition, if the fault similarity coefficient judgment process is used alone, it is necessary to calculate the fault similarity coefficients corresponding to all fault types, then the amount of calculation is too large, and the time consumption is too long, which cannot meet the real-time requirements in actual engineering. Therefore, the present embodiment first uses the support vector machine fault diagnosis model to preliminarily judge the fault classification result, and performs the fault similarity coefficient judgment process based on the fault classification result. Only one fault similarity coefficient calculation is required, which not only meets the real-time requirements but also improves the classification accuracy.

[0187] The support vector machine fault diagnosis model is used to make a preliminary prediction of the classification results, and the real category and predicted category of each real-time vibration signal time-frequency graph are output and compared with the real label. At the same time, an evaluation report is obtained, and the classification_report function is used to generate an evaluation report of the classifier to evaluate the performance of the model. It can be preliminarily determined that the real-time fault of the transformer is one of the four types of transformer faults: core fault, winding fault, oil tank fault, and cooler fault.

[0188] According to the fault classification results, the historical fault time-frequency diagram is obtained, and n first amplitude values ​​A are extracted from the historical fault time-frequency diagram. x[x=1,2,3.......n], this feature characterizes the frequency distribution of the historical fault transformer signal and the amplitude variation characteristics in different time periods, and draws a scatter plot of the historical fault amplitude values, wherein the scatter plot of the historical fault amplitude values ​​is defined as the scatter plot characteristics of the amplitude values ​​of the time-frequency diagram of different fault types when the transformer is in normal operation in history, and the scatter plot characteristics include time, frequency and amplitude values;

[0189] Extract n second amplitude values ​​B from the predicted fault frequency graph x [x=1,2,3.......n], this feature characterizes the frequency distribution of the predicted fault transformer signal and the amplitude variation characteristics in different time periods, and draws a scatter plot of the predicted fault amplitude value, wherein the scatter plot of the predicted fault amplitude value is defined as the scatter plot characteristics of the amplitude values ​​of the time-frequency diagram of different fault types when the transformer is operating normally in real time, and the scatter plot characteristics include time, frequency and amplitude values.

[0190] Define the fault similarity coefficient C p [p=1,2,3,4], the fault similarity coefficient represents the similarity between the predicted fault graph and the historical fault graph, p represents four fault types, where:

[0191]

[0192] Among them, C p The fault similarity coefficient corresponding to the pth fault, n is the total number of the first amplitude value or the second amplitude value, A x is the xth first amplitude value, B x is the x-th second amplitude value.

[0193] Compare the scatter plot of historical fault amplitude values ​​with the scatter plot of predicted fault amplitude values, and calculate the fault similarity coefficient C p , when the fault similarity coefficient C p Greater than or equal to the similarity threshold G of each fault type p When the fault similarity coefficient C p Less than the fault similarity threshold G p If the fault is not found in the transformer, the fault should be eliminated and the prediction should be repeated.

[0194] The following is a specific example to describe the implementation process of the fault classification detection method based on transformer vibration characteristics and time-frequency diagram:

[0195] (1) Data preparation stage:

[0196] By using the vibration signal acquisition module of the transformer, the vibration signals of its different structural parts are measured, and a fault classification detection method based on the transformer vibration characteristics and time-frequency diagram is established. The characteristic frequency curve of the transformer vibration signal is obtained and converted into a time-frequency diagram. An SVM fault diagnosis model is established to preliminarily diagnose the fault type of the transformer. Then, the fault type of the transformer is monitored by extracting the amplitude value characteristics in the time-frequency diagram and matching the frequency characteristic curve diagram.

[0197] In this embodiment, through step S1, the transformer vibration signal is collected, the obtained transformer vibration signal data is sorted by calculation formula (1), stored in an Excel file, and its original image is drawn. Due to the large number of data images, only the original images of the transformer when it is running under short-circuit fault, when the transformer is running under load, and when the transformer is running under no-load are displayed here, such as Figure 6 , Figure 7 , Figure 8 shown.

[0198] The frequency characteristic curve graph and the time-frequency graph in step S3 and S3 are stored in a certain path, and the data is read by reading the image in the specified path. And divide the training set and the test set: Use the train_test_split function to divide the loaded image data into a training set and a test set, of which 20% is used as a test set and the rest is used as a training set.

[0199] (2) Feature extraction stage:

[0200] The image pixel histogram in the time-frequency image is extracted as the SVM classification feature, the pixel histogram of the image is calculated by the function in the Open CV library, and the characteristics of the image are described by counting the number of each pixel value.

[0201] The transformer vibration signal is processed by calculating formula (2)-formula (9), and the frequency characteristics of the vibration signal calculation matrix are extracted by variational mode decomposition algorithm, and then drawn into a two-dimensional image, as shown in Fig. 9 , Fig.10 , Fig.11 shown.

[0202] (3) Model training phase:

[0203] Use the SVC class in the Scikit-learn library to create an SVM classifier. At the same time, use the linear kernel function to train the input time-frequency image and match it with the corresponding frequency feature curve to obtain a trained SVM model.

[0204] By calculating the Hilbert transform method of formula (10)-formula (15), the frequency characteristic curve obtained in step S4 is Hilbert transformed to obtain the time-frequency diagram of its vibration signal, such as Fig.12 , Fig.13 , Fig.14 shown.

[0205] (4) Model prediction stage:

[0206] Use the trained SVM model to classify and predict the test set data. Call the predict method, pass in the test set feature vector, and get the predicted label of the test set.

[0207] The SVM fault diagnosis model established in step S5 is used to preliminarily diagnose the transformer fault and determine the fault type of the transformer. A support vector machine (SVM) fault diagnosis model is established to preliminarily classify and diagnose possible faults of the transformer. By training the SVM classifier and using the time-frequency graph features to classify the vibration signal, the transformer fault type is identified as core fault, winding fault, oil tank fault, or cooler fault.

[0208] Then, the time-frequency graph after preliminary classification is subjected to feature extraction, the amplitude value features in the time-frequency graph are extracted, and a scatter plot of the amplitude value is drawn based on the time-frequency graph, such as Fig.15 , Fig.16 , Fig.17 shown.

[0209] The scatter plot features of the amplitude values ​​of the time-frequency graphs of different fault types when the transformer is operating normally are defined, including three features: time, frequency and maximum amplitude. The time-frequency image output by the SVM classifier is compared with it. If the similarity of the time-frequency image is greater than or equal to 80%, it is preliminarily determined to be the fault type.

[0210] (5) Result evaluation stage:

[0211] The output results are predicted, and the true category and predicted category of each test time-frequency image are output and compared with the true label. At the same time, an evaluation report is obtained, and the classification_report function is used to generate an evaluation report of the classifier to evaluate the performance of the model. Through the above operations, the transformer fault is classified and it is determined whether it belongs to the transformer primary fault: core fault, winding fault, oil tank fault, or cooler fault.

[0212] The predicted fault scatter plot extracted from the time-frequency image output by the SVM classifier is compared with the historical fault comparison scatter plot, and the similarity between the predicted fault map and the historical fault map is calculated according to the fault similarity coefficient calculation formula. p, when the fault similarity coefficient C p Greater than or equal to the fault similarity threshold G p When the fault similarity coefficient C p Less than the fault similarity threshold G p If the fault is not found in the transformer, the fault should be eliminated and the prediction should be repeated.

[0213] Embodiment 2:

[0214] This embodiment proposes a fault classification detection device based on transformer vibration characteristics and time-frequency diagrams, such as Fig.18 As shown, it includes: a vibration signal acquisition module, a data acquisition card, a preliminary fault detection module and a fault determination module, the vibration signal acquisition module is connected to the data acquisition card, the data acquisition card is connected to the preliminary fault detection module, the preliminary fault detection module is connected to the fault determination module, the fault determination module is connected to the vibration signal acquisition module, and the preliminary fault detection module and the fault determination module are arranged in a host computer;

[0215] The vibration signal acquisition module is connected to the transformer and the data acquisition card respectively, and is used to collect the real-time vibration data and historical vibration data of each part of the transformer, and send the real-time vibration data and historical vibration data to the data acquisition card.

[0216] A data acquisition card, used for storing real-time vibration data and historical vibration data, and sending the real-time vibration data and historical vibration data to a fault detection module;

[0217] A preliminary fault detection module is used to establish a real-time vibration signal matrix according to the real-time vibration data; extract the frequency characteristics of the real-time vibration signal matrix using a variational mode decomposition model to obtain a real-time frequency characteristic curve; transform the real-time frequency characteristic curve using a Hilbert transform to obtain a time-frequency diagram of the real-time vibration signal; input the time-frequency diagram of the real-time vibration signal into a pre-trained support vector machine fault diagnosis model to obtain a fault classification result and a predicted fault time-frequency diagram;

[0218] A fault determination module is used to obtain a historical fault time-frequency diagram corresponding to the fault classification result, extract n first amplitude value characteristic values ​​from the historical fault time-frequency diagram; extract n second amplitude values ​​from the predicted fault time-frequency diagram; calculate a fault similarity coefficient based on the n first amplitude values ​​and the n second amplitude values; if the fault similarity coefficient is greater than or equal to a similarity threshold of the fault corresponding to the fault classification result, determine that the transformer has a fault corresponding to the fault classification result; if the fault similarity coefficient is less than a similarity threshold of the fault corresponding to the fault classification result, determine that the transformer has not a fault corresponding to the fault classification result, and return to the vibration signal acquisition module to continue detection.

[0219] The vibration signal acquisition module is composed of six parts: sensor module, signal conditioning module, analog-to-digital converter module, communication module, microcontroller module and power module. Fig.19 shown.

[0220] Among them, the sensor module is mainly composed of FA02-A optical fiber vibration sensor, which is used to measure the vibration data of different parts of the transformer and is also the core component for collecting transformer vibration data. It is responsible for converting mechanical vibration into optical signal output and collecting it through the DH5922n data acquisition card, usually in the form of analog signals; the signal conditioning module plays the role of signal amplification, filtering and conditioning. By amplifying the signal, the signal-to-noise ratio can be improved, filtering can remove unnecessary frequency components, and conditioning can make the signal meet the input requirements of subsequent processing devices. Ensure that the signal quality obtained from the FA02-A optical fiber vibration sensor is good and suitable for subsequent digital processing; the analog-to-digital converter module is responsible for converting analog signals into digital signals for computer processing and analysis; the communication module is used to transmit the collected vibration data to the computer for further analysis and processing. Common communication methods include serial communication, Ethernet communication, Wi-Fi or Bluetooth, etc. In this embodiment, Ethernet is used to transmit the signal; the microcontroller module is responsible for controlling the operation of the entire acquisition module, including the control of the sensor and ADC, data processing and communication functions; the power module is responsible for providing a stable power supply for the acquisition module and managing the battery power.

[0221] The various embodiments in the present application are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0222] The protection scope of the present application is not limited to the above-mentioned embodiments. Obviously, those skilled in the art can make various changes and modifications to the present disclosure without departing from the scope and spirit of the present disclosure. If these changes and modifications fall within the scope of the claims of the present disclosure and their equivalents, the intention of the present disclosure also includes these changes and modifications.

Claims

1. A fault classification detection method based on transformer vibration characteristics and time-frequency diagram, characterized in that: include: Step S1: collecting real-time vibration data of each part of the transformer; Step S2: establishing a real-time vibration signal matrix according to the real-time vibration data; Step S3: extracting the frequency characteristics of the real-time vibration signal matrix using a variational mode decomposition model to obtain a real-time frequency characteristic curve; Step S4: transforming the real-time frequency characteristic curve using Hilbert transform to obtain a time-frequency diagram of the real-time vibration signal; Step S5: inputting the time-frequency diagram of the real-time vibration signal into a pre-trained support vector machine fault diagnosis model to obtain a fault classification result and a predicted fault time-frequency diagram; Step S6: Obtain a historical fault time-frequency diagram according to the fault classification result, and extract n first amplitude values ​​from the historical fault time-frequency diagram; Step S7: extracting n second amplitude values ​​from the predicted fault time-frequency diagram; Step S8: calculating a fault similarity coefficient according to the n first amplitude values ​​and the n second amplitude values; Step S9: If the fault similarity coefficient is greater than or equal to the similarity threshold of the fault corresponding to the fault classification result, it is determined that the transformer has a fault corresponding to the fault classification result; Step S10: If the fault similarity coefficient is less than the similarity threshold of the fault corresponding to the fault classification result, it is determined that the transformer does not have the fault corresponding to the fault classification result, and the process returns to step S1 to continue detection.

2. The fault classification detection method based on transformer vibration characteristics and time-frequency diagram according to claim 1 is characterized in that: The real-time vibration data of each part of the transformer is collected, including: Collect real-time vibration data of the core; Collect real-time vibration data of windings; Collect real-time vibration data of the fuel tank; Collect real-time vibration data of chillers.

3. The fault classification detection method based on transformer vibration characteristics and time-frequency diagram according to claim 2 is characterized in that: The real-time vibration data of the iron core is collected, including collecting real-time vibration data at the following positions: the intersection of the iron core column corresponding to the A-phase high and low voltage winding and the upper iron yoke, the intersection of the iron core column corresponding to the A-phase high and low voltage winding and the lower iron yoke, the intersection of the iron core column corresponding to the B-phase high and low voltage winding and the upper iron yoke, the intersection of the iron core column corresponding to the B-phase high and low voltage winding and the lower iron yoke, the intersection of the iron core column corresponding to the C-phase high and low voltage winding and the upper iron yoke, and the intersection of the iron core column corresponding to the C-phase high and low voltage winding and the lower iron yoke; The real-time vibration data of the windings are collected, including collecting real-time vibration data at the following positions: one-half and five-sixths of the A-phase high-voltage winding, one-half and five-sixths of the A-phase low-voltage winding, one-half and five-sixths of the B-phase high-voltage winding, one-half and five-sixths of the B-phase low-voltage winding, one-half and five-sixths of the C-phase high-voltage winding, and one-half and five-sixths of the C-phase low-voltage winding; The collecting of the real-time vibration data of the oil tank includes collecting the real-time vibration data at the following positions: the top center position of the oil tank and the bottom center position of the oil tank; The collecting of the real-time vibration data of the cooler includes collecting the real-time vibration data at the following positions: the connection between the cooler support and the transformer oil tank, and at the cooling fan of the cooler.

4. The fault classification detection method based on transformer vibration characteristics and time-frequency diagram according to claim 1 is characterized in that: The fault similarity coefficient is calculated as follows: Among them, C p The fault similarity coefficient corresponding to the pth fault, n is the total number of the first amplitude value or the second amplitude value, A x is the xth first amplitude value, B x is the x-th second amplitude value.

5. The fault classification detection method based on transformer vibration characteristics and time-frequency diagram according to claim 1 is characterized in that: The n first amplitude values ​​characterize the frequency distribution of the historical fault transformer signal and the amplitude variation characteristics in different time periods, and draw a scatter plot of the historical fault amplitude values, wherein the scatter plot of the historical fault amplitude values ​​is defined as the scatter plot characteristics of the amplitude values ​​of the time-frequency diagram of different fault types when the transformer is in historical normal operation, and the scatter plot characteristics include time, frequency and amplitude values; The n second amplitude values ​​characterize the frequency distribution of the predicted fault transformer signal and the amplitude variation characteristics in different time periods, and draw a scatter plot of the predicted fault amplitude values, wherein the scatter plot of the predicted fault amplitude values ​​is defined as the scatter plot characteristics of the amplitude values ​​of the time-frequency diagram of different fault types when the transformer operates normally in real time, and the scatter plot characteristics include time, frequency and amplitude values.

6. The fault classification detection method based on transformer vibration characteristics and time-frequency diagram according to claim 1 is characterized in that: The variational mode decomposition model, the construction process includes: Collect historical vibration data of each part of the transformer; Establishing a historical vibration signal matrix according to the historical vibration data; The variational modal decomposition algorithm is used to decompose the historical vibration signal matrix into multiple modal components, and the multiple modal components are arranged from low frequency to high frequency to obtain the intrinsic mode function; Constructing a constraint model of the variational mode decomposition model based on the intrinsic mode functions; Lagrange multipliers and penalty factors are introduced to solve the constraint model of the variational mode decomposition model to obtain an optimal solution, which includes: Wiener filtering, the power center of gravity of the intrinsic mode function and Lagrange multipliers. The optimal solution is used as the final variational mode decomposition model.

7. The fault classification detection method based on transformer vibration characteristics and time-frequency diagram according to claim 6 is characterized in that: The frequency characteristics of the real-time vibration signal matrix are extracted by using a variational modal decomposition model to obtain a real-time frequency characteristic curve, including: Step S3.1: Initialize the Wiener filter, the power center of gravity of the intrinsic mode function and the Lagrange multiplier to obtain initialization data; Step S3.2: updating the initialization data according to the variational mode decomposition model; Step S3.3: If the convergence condition is met, the corresponding k-th decomposed intrinsic mode function is output; Step S3.4: If the convergence condition is not met, the number of iterations is increased by 1, and the process returns to step S3.2 to continue updating; Step S3.5: taking the k decomposed intrinsic mode functions as the frequency characteristics of the real-time vibration signal matrix, and drawing a corresponding real-time frequency characteristic curve according to the frequency characteristics of the real-time vibration signal matrix.

8. The fault classification detection method based on transformer vibration characteristics and time-frequency diagram according to claim 7 is characterized in that: The real-time frequency characteristic curve is transformed by Hilbert transform to obtain a time-frequency diagram of the real-time vibration signal, including: Perform Hilbert transform on the k decomposed intrinsic mode functions to obtain the Hilbert transform result; Performing an inverse Hilbert transform on the Hilbert transform result to obtain an inverse Hilbert transform result; Analyze the inverse Hilbert transform result to obtain the instantaneous amplitude, instantaneous phase and instantaneous frequency of the vibration signal; A time-frequency diagram of the real-time vibration signal is obtained according to the instantaneous amplitude, instantaneous phase and instantaneous frequency of the vibration signal.

9. The fault classification detection method based on transformer vibration characteristics and time-frequency diagram according to claim 1 is characterized in that: The pre-trained support vector machine fault diagnosis model construction process includes: A variational mode decomposition model is used to extract the frequency characteristics of the historical vibration signal matrix to obtain a historical frequency characteristic curve; The historical frequency characteristic curve is transformed by using Hilbert transform to obtain a time-frequency diagram of the historical vibration signal; Taking the time-frequency diagram of historical vibration signals as input and the fault classification results as output, the classification function of the support vector machine and the constraint condition equation with slack variables are constructed; According to Lagrange's theorem, the constraint equation with slack variables is transformed into Lagrangian form; Solving the Lagrangian form to obtain a dual form of the constraint equation with slack variables; The classification function of the support vector machine is updated according to the dual form to obtain an optimal classification decision function, and the optimal classification decision function is used as a pre-trained support vector machine fault diagnosis model.

10. A fault classification detection device based on transformer vibration characteristics and time-frequency diagram, characterized in that: include: The vibration signal acquisition module is connected to the transformer and the data acquisition card respectively, and is used to collect the real-time vibration data and historical vibration data of each part of the transformer, and send the real-time vibration data and historical vibration data to the data acquisition card. A data acquisition card, used for storing real-time vibration data and historical vibration data, and sending the real-time vibration data and historical vibration data to a fault detection module; A preliminary fault detection module is used to establish a real-time vibration signal matrix according to the real-time vibration data; extract the frequency characteristics of the real-time vibration signal matrix using a variational mode decomposition model to obtain a real-time frequency characteristic curve; transform the real-time frequency characteristic curve using a Hilbert transform to obtain a time-frequency diagram of the real-time vibration signal; input the time-frequency diagram of the real-time vibration signal into a pre-trained support vector machine fault diagnosis model to obtain a fault classification result and a predicted fault time-frequency diagram; A fault determination module is used to obtain a historical fault time-frequency diagram according to the fault classification result, and extract n first amplitude values ​​from the historical fault time-frequency diagram; Extracting n second amplitude values ​​from the predicted fault time-frequency diagram; calculating the fault similarity coefficient according to the n first amplitude values ​​and the n second amplitude values; If the fault similarity coefficient is greater than or equal to the similarity threshold of the fault corresponding to the fault classification result, it is determined that the transformer has the fault corresponding to the fault classification result; if the fault similarity coefficient is less than the similarity threshold of the fault corresponding to the fault classification result, it is determined that the transformer has not the fault corresponding to the fault classification result, and returns to the vibration signal acquisition module to continue detection.

Images