CN115331696A - Multi-channel voiceprint signal blind source separation method for transformer abnormity diagnosis - Google Patents

Abstract

The invention provides a multi-channel voiceprint signal blind source separation method for transformer abnormity diagnosis. The method comprises the following steps: uniformly arranging a plurality of probes at different positions on the wall of an oil tank of a transformer to be subjected to fault diagnosis, combining vibration signals collected by the plurality of probes, inputting the combined vibration signals into a multichannel vibration signal blind source separation algorithm combining SDICA and SC, and calculating to obtain final separation signals; respectively constructing a sample data set by using winding and iron core vibration signals of the transformer in normal operation, and training the Gaussian mixture model by using the sample data set to obtain a trained Gaussian mixture model; and inputting the final separation signal into the trained mixed Gaussian model, and outputting a fault diagnosis result of the transformer by the trained mixed Gaussian model. The invention can obtain more accurate effect by using the mixed Gaussian model to diagnose the abnormal state of the transformer, and can realize the respective diagnosis of the winding and the iron core state after combining the multi-channel blind source separation algorithm.

Description

A Blind Source Separation Method of Multi-channel Voiceprint Signals for Abnormal Diagnosis of Transformers

technical field

The invention relates to the technical field of transformer fault diagnosis, in particular to a method for blind source separation of multi-channel voiceprint signals for abnormal diagnosis of transformers.

Background technique

With the continuous expansion of the grid scale, the requirements for the safe and stable operation of the power system have been greatly increased. Power transformers, as key power equipment, play an irreplaceable role, and their safety and reliability are crucial to the grid. Once the transformer or transformer fails, it will have an impact on the entire power grid, and in severe cases, it will cause a large-scale collapse of the power grid.

The vibration analysis method is different from the traditional method of abnormal diagnosis of transformers. At present, some scholars have used the vibration analysis method to carry out abnormal diagnosis of transformers. It uses vibration sensors to collect vibration signals on the surface of transformers, and performs feature extraction and abnormal detection algorithms on vibration signals. Determine the operating status of the transformer. However, since the collected vibration signal of the transformer is the vibration signal of the surface of the transformer oil tank, including the mixed signal of multiple internal vibration sources, the method of monitoring the condition of the transformer based on the vibration signal of the transformer oil tank surface cannot effectively identify the fault location and fault type of the transformer. .

In the "Research on Transformer Vibration Signal Separation Based on JADE Algorithm" published in "Modern Electric Power" (2012,29:42-49), some scholars used the JADE algorithm to simulate the vibration signal of the transformer using MATLAB and LabVIEW to realize the winding The separation from the core vibration signal, but it has not been applied to the actual transformer. Some scholars published "Transformer Winding and Core Vibration Based on RBF Neural Network" in "High Voltage Electrical Appliances" (2019,55(11):159-164) In "Signal Separation Research", the radial basis neural network is used to take the frequency domain characteristics of the mixed vibration signal as input to realize the separation of the core and winding vibration signals, and at the same time solve the problems of source signal sequence and amplitude uncertainty. In the "Application of Blind Source Separation Technology in Transformer Fault Detection by Vibration Method" published in "Journal of Electrotechnical Society" (2012,27(10):68-78), some scholars considered the vibration of transformer winding and core Frequency aliasing phenomenon of the signal, a subspace independent component separation method is proposed to separate the transformer vibration signals that are not completely independent.

However, for an actual transformer with a large volume and complex internal structure, the vibration signals at different positions on the surface of the oil tank are not exactly the same, especially when the transformer winding is deformed, for different deformation modes and degrees, when the installation position of the sensor is different At the same time, the separation effect is not exactly the same; in addition, when the traditional blind source separation algorithm separates the transformer acquisition signal, because the blind source separation algorithm is sensitive to the initial value, some even may not converge, and it is impossible to ensure that each The second separation effect is the best.

Contents of the invention

Embodiments of the present invention provide a method for blind source separation of multi-channel voiceprint signals for abnormal diagnosis of transformers, so as to effectively diagnose abnormal states of transformers.

In order to achieve the above object, the present invention adopts the following technical solutions.

A method for blind source separation of multi-channel voiceprint signals for abnormal diagnosis of transformers, comprising:

Evenly arranging a plurality of probes at different positions on the oil tank wall of the transformer to be fault diagnosed, and the plurality of probes respectively collect vibration signals of the transformer;

After the vibration signals collected by the plurality of probes are combined, they are input into a multi-channel vibration signal blind source separation algorithm combining spatial independent component separation method SDICA and spectral clustering SC, and the final separation signal is obtained through calculation;

Using the winding and iron core vibration signals of the transformer during normal operation to construct sample data sets respectively, using the sample data sets to train the mixed Gaussian model to obtain a trained mixed Gaussian model;

The final separated signal is input into the trained mixed Gaussian model, and the trained mixed Gaussian model outputs the fault diagnosis result of the transformer.

Preferably, the plurality of probes are evenly arranged at different positions on the oil tank wall of the transformer to be fault diagnosed, and the plurality of probes respectively collect vibration signals of the transformer, including:

Arrange n sensor probes evenly at different positions on the oil tank wall of the transformer to be fault diagnosed, select the vibration signals collected by i probes from the sensor probes at different positions, and respectively select the vibration signals from the sensor probes at n different positions Select the vibration signals collected by i probes, n∈[2,18], and the number J of observation signal groups generated by permutation and combination is:

In the formula, J _n is the number of calculations when n probes are taken, and k _i is the calculation ratio.

Preferably, after the vibration signals collected by the multiple probes are combined, they are input to a multi-channel vibration signal blind source separation algorithm combining spatial independent component separation method SDICA and spectral clustering SC, and the final separation is obtained by calculation Signals, including:

Input J groups of observation signals into SDICA for calculation to obtain the initial separation signal, where the initial separation signal after the jth operation is expressed as Y _j ”={y _j1 ”,y _j2 ”,...,y _jn ”} ^T ;

其中l_j为第j次计算后得到的有效分离信号个数，l_j≤n，经过J次运算，得到有效分离信号集合Y'＝{Y₁',Y₂',...,Y_J'}，Y为

维矩阵，总共

个有效信号；The initial separation signal is screened, and the separation signal whose amplitude is higher than the maximum separation signal amplitude of 1‰ is retained to obtain an effective separation signal. It is assumed that the effective separation signal after the jth retained amplitude is

Where l _j is the number of effective separation signals obtained after the jth calculation, l _{j ≤ n} , after J operations, the effective separation signal set Y'={Y ₁ ',Y ₂ ',...,Y _J '}, Y is

dimension matrix, the total

a valid signal;

Use the SC algorithm to cluster the effective separation signals. The effective separation signal corresponding to the point closest to the cluster center is the final separation signal. The final separation signal is used as the vibration source signal inside the transformer, which is recorded as Y={Y ₁ ,Y ₂ } .

Preferably, the J groups of observation signals are respectively input into SDICA for operation to obtain the initial separation signal, wherein the initial separation signal after the jth operation is expressed as Y _j ”={y _j1 ”,y _j2 ”,.. .,y _jn ”} ^T , including:

2-1) performing wavelet packet decomposition on the mixed J group of observation signals;

2-2) Calculate the mutual information MI value of each subspace, and select the first five subspaces with the strongest independence, and reconstruct these subspaces;

Among them, MI is the mutual information value of the observed signals X ₁ , X ₂ , X ₃ ,...X _L , and cum() represents the high-order statistics of the signals;

2-3) Demeaning and whitening operations are performed on the observed signal X;

即为分离出的近似源信号，G为全局矩阵，负熵最大化为算法的目标函数，其公式为：2-4) Set the number of source signals and the number of iterations, adopt positive definite blind source separation, that is, the number of source signals is equal to the number of observation signals, where the relationship between source signal S(t) and mixed signal X(t) is

That is, the separated approximate source signal, G is the global matrix, and the maximization of negative entropy is the objective function of the algorithm, and its formula is:

Where W is the separation matrix, X is the observed signal, E() represents the expectation of the data, G(.) is a non-quadratic function, usually G(y)=tanh(y), v is a Gaussian variable with zero mean unit variance , _ki is a normal number, and p is the number of functions that use non-quadratic functions to approximate the negative entropy;

Solve the separation matrix W so that the separation signal can maximize the objective function. When there is E{(wx) ² }=1, the objective function can be transformed into:

w is a row of the separation matrix, X _k is a certain observation signal, G(.) is a non-quadratic function, usually G(y)=tanh(y)

Initialize the weight phasor w(0);

2-5) Iterate over w:

w(n+1)=E{XG(w ^T (n)X)}-

E{G'(w ^T (n)X)}w(n)

w is a row of the separation matrix, G(.) is a non-quadratic function, usually G(y)=tanh(y), G'() is the first derivative of G(). n is the number of iterations.

2-6) Normalization processing: w(n+1)=w(n+1)/||w(n+1)||

2-7) If the signal does not converge, return to step (2-5), if converge, obtain the separation matrix W;

2-8) The mixed signal is multiplied to the left by the separation matrix W ^T , and the initial separation signal of the separated approximate source signal is obtained

Preferably, the SC algorithm is used to cluster the effective separation signals, and the separation effective signal corresponding to the point closest to the cluster center is the final separation signal, and the final separation signal is used as the internal vibration source signal of the transformer, which is recorded as Y={ Y ₁ ,Y ₂ }, including:

4-1) Construct an undirected weight graph g=(V, E) for the sample data set V, where V represents the point set of the graph, and E represents the edge set;

4-2) Construct the adjacency matrix W by using the full connection method based on Gaussian distance;

The weight of the edge connecting any two nodes v _i and v _j in the graph is w _ij , and the adjacency matrix W∈R ^n×n is constructed;

4-3) Define the degree matrix D of the node as an N-order square matrix, and the expression of element d _uv in D is:

Among them, the diagonal elements represent the sum of the weights of all edges connected to the node v _u , and the Laplacian matrix is defined as Ls=DW;

4-4) Define the indicator vector:

h _y ＝{h ₁ ,h ₂ ,...,h _x ,...h _N }, y=1,2,...,Y ₀ , where x represents the subscript of the node, and y represents the subscript of the subset , h _xy is the indication of node x to subset y, expressed as:

In the formula, |A _y | is the number of points in the subset A _y of V, which means that each subset A _y corresponds to an indicator vector h _y , and there are N elements in h _y , representing the indication of N nodes result;

The goal of graph cutting is to divide V into Y ₀ subgraphs, so that the weight of the edges of each subgraph is minimized, then the objective function is the trace optimization problem of the matrix:

H is the characteristic matrix, Ls is the Laplacian matrix, and Tr() represents the sum of the diagonal elements of the matrix

4-5) Perform K-means clustering on the feature matrix U after the feature matrix H is unitized, and divide V into subsets A ₁ , A ₂ ,...,A _y ,..., A _Y .

Preferably, using the winding and core vibration signals of the transformer during normal operation to construct sample data sets respectively, using the sample data sets to train the mixed Gaussian model to obtain a trained mixed Gaussian model, including:

Use the vibration signals collected during the normal operation of the transformer to perform multi-channel blind source separation calculations to obtain the vibration signals of the winding and iron core, and use the vibration signals of the winding and iron core as the training set for mixed Gaussian modeling, and select the transformer under different operating conditions The vibration signals of the winding and iron core are used as the test set;

Using the linear filter bank and principal component analysis method to linearly reduce the training set signal to obtain the reduced training set;

Iteratively optimize the parameters of the mixed Gaussian model, initialize the parameters of M Gaussian models, use the EM algorithm to iterate the parameters, and output the mixed Gaussian model parameters after meeting the convergence conditions, and obtain the trained mixed Gaussian model parameters.

Preferably, the input of the final separation signal into the trained mixed Gaussian model, and the trained mixed Gaussian model outputs the fault diagnosis result of the transformer, including:

Substituting the final separation signal A of the transformer to be fault diagnosed into the trained mixed Gaussian model, the mixed Gaussian model calculates the abnormality of the final separation signal A, and the calculation formula is:

F(A)＝1-GMM _A (X _normal )

In the formula, F(A) is the abnormal degree of the final separation signal A, X _normal is the vibration signal of the transformer under normal state, GMM _A (X _normal ) is the probability that the final separation signal A belongs to the GMM model, if the final separation signal A The greater the probability of belonging to the GMM model, that is, the more similar the final separation signal A is to the vibration signal during normal operation of the transformer, the lower the abnormality of the transformer to be fault diagnosed.

From the technical solutions provided by the above embodiments of the present invention, it can be seen that the embodiment of the present invention uses a mixed Gaussian model to diagnose the abnormal state of the transformer to obtain a more accurate effect, and after combining the multi-channel blind source separation algorithm, it can realize the diagnosis of Separate diagnosis of winding and core condition.

Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description, or may be learned by practice of the invention.

Description of drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For Those of ordinary skill in the art can also obtain other drawings based on these drawings without making creative efforts.

Fig. 1 is an implementation schematic diagram of a method for blind source separation of multi-channel voiceprint signals for transformer abnormality diagnosis provided by an embodiment of the present invention;

Fig. 2 is a schematic diagram of data collected by 18 vibration sensors provided by an embodiment of the present invention;

FIG. 3 is a schematic diagram of a separation signal provided by an embodiment of the present invention;

4 is a schematic diagram of a winding training set fitting GMM model provided by an embodiment of the present invention;

FIG. 5 is a schematic diagram of a GMM model fitted to an iron core training set provided by an embodiment of the present invention;

Fig. 6 is a schematic diagram of a three-phase voltage and current harmonic content change curve provided by an embodiment of the present invention.

Detailed ways

Embodiments of the present invention are described in detail below, examples of which are shown in the drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the figures are exemplary only for explaining the present invention and should not be construed as limiting the present invention.

Those skilled in the art will understand that unless otherwise stated, the singular forms "a", "an", "said" and "the" used herein may also include plural forms. It should be further understood that the word "comprising" used in the description of the present invention refers to the presence of said features, integers, steps, operations, elements and/or components, but does not exclude the presence or addition of one or more other features, Integers, steps, operations, elements, components, and/or groups thereof. It will be understood that when an element is referred to as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements may also be present. Additionally, "connected" or "coupled" as used herein may include wirelessly connected or coupled. As used herein, the term "and/or" includes any and all combinations of one or more of the associated listed items.

Those skilled in the art can understand that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. It should also be understood that terms such as those defined in commonly used dictionaries should be understood to have a meaning consistent with the meaning in the context of the prior art, and unless defined as herein, are not to be interpreted in an idealized or overly formal sense explain.

In order to facilitate the understanding of the embodiments of the present invention, several specific embodiments will be taken as examples for further explanation below in conjunction with the accompanying drawings, and each embodiment does not constitute a limitation to the embodiments of the present invention.

In order to realize the abnormal diagnosis of different structures inside the transformer and reduce the influence of the installation position of the vibration sensor on the effect of blind source separation, this invention proposes a blind source separation algorithm for transformer multi-channel voiceprint signals. Vibration signals on the surface are subjected to multiple SDICA (Sub-band decomposition independent component analysis) using vibration signals of different channel positions and numbers, and then the separated signals are separated by SC (Spectral clustering, spectral clustering) clustering) algorithm to obtain the vibration source signal inside the transformer. Finally, the winding and core vibration signals of the transformer during normal operation are used as the training set to model the winding and GMM (Gaussian mixed model, core mixed Gaussian model), which realizes the diagnosis of the abnormal state of the winding and core, and provides a basis for the transformer. Fault diagnosis provides a new method.

In the present invention, 18 probes are evenly arranged on the wall of the transformer oil tank for fault diagnosis to collect vibration signals. In order to make full use of the vibration signals at various positions of the transformer, the internal vibration source signals that exist stably for a long time can be obtained. The abnormal diagnosis of the structure can reduce the influence of the installation position of the vibration sensor on the result of blind source separation, and improve the effect of blind source separation. A multi-channel vibration signal blind source separation algorithm combining SDICA and SC is proposed, which is the invention It is innovative and adopts Gaussian mixed model (GMM) for abnormal state diagnosis. First, the combination of observation signals with different positions and numbers is selected for SDICA separation, and then the calculation results of multiple separations are analyzed by SC, and the signal closest to the cluster center is extracted as the final separated winding and core vibration signals. Finally, the vibration signals of windings and cores of transformers during normal operation are used to construct sample data sets, which are used as training sets for GMM modeling of windings and cores to realize the diagnosis of abnormal states of windings and cores.

An implementation principle diagram of a method for blind source separation of multi-channel voiceprint signals for transformer abnormality diagnosis provided by an embodiment of the present invention is shown in FIG. 1 . Specific steps are as follows:

Step (1) Generate multiple sets of observation signals of the transformer to be diagnosed abnormally

The vibration signals collected by i probes are respectively selected from n sensor probes at different positions, and the collected vibration signals are used as the observation signals input into the SDICA algorithm. The SDICA algorithm outputs n initial separation signals, where n∈[2,18], Under the same number of probes, due to different positions, the probe combinations are also different. Therefore, the number J of observation signal groups generated by permutation and combination is:

表示从18个振动传感器探头中选择n个探头的所有组合方法。In the formula, J _n is the number of calculations when n probes are taken, k _i is the calculation ratio, which is used to reduce the calculation amount,

Indicates all combination methods of selecting n probes from 18 vibration sensor probes.

The time-domain and frequency-domain waveforms of data collected by 18 vibration sensors provided by the present invention are shown in FIG. 2 .

Step (2) Input the observation signal, and use the SDICA algorithm to obtain the initial separation signal

Input J groups of observation signals into SDICA for calculation, where the initial separation signal after the jth operation can be expressed as Y _j ”={y _j1 ”,y _j2 ”,...,y _jn ”} ^T . The specific process of this step includes:

2-1) Perform wavelet packet decomposition on the mixed signal.

2-2) Calculate the MI (Mutual Information, mutual information value) value of each subspace, and select the first five subspaces with the strongest independence, and reconstruct these subspaces.

Among them, MI is the mutual information value of the observed signals X ₁ , X ₂ , X ₃ ,...X _L , and cum() represents the high-order statistics of the signals.

2-3) Demeaning and whitening operations are performed on the observed signal X.

即为分离出的近似源信号，G为全局矩阵，理论情况下为单位阵。负熵最大化为算法的目标函数，其公式为：2-4) Set the number of source signals and the number of iterations. The present invention adopts positive definite blind source separation, that is, the number of source signals is equal to the number of observation signals, and the relationship between source signal S(t) and mixed signal X(t) is

That is, the separated approximate source signal, G is a global matrix, and in theoretical cases is an identity matrix. Negative entropy maximization is the objective function of the algorithm, and its formula is:

Where W is the separation matrix, X is the observed signal, E() represents the expectation of the data, G(.) is a non-quadratic function, usually G(y)=tanh(y), v is a Gaussian variable with zero mean unit variance , _ki is a normal number, and p is the number of functions that use non-quadratic functions to approximate the negative entropy.

Therefore, the problem is transformed into solving the separation matrix W so that the separation signal can maximize the objective function. When there is E{(wx) ² }=1, the objective function can be transformed into:

w is a row of the separation matrix, E{.} represents the mathematical expectation, X _k is a certain observation signal, G(.) is a non-quadratic function, usually G(y)=tanh(y)

Initialize the weight phasor w(0).

2-5) Iterate over w:

w(n+1)=E{XG(w ^T (n)X)}-

E{G'(w ^T (n)X)}w(n)

w is a row of the separation matrix, and E{.} represents the mathematical expectation. G(.) is a non-quadratic function, usually G(y)=tanh(y), G'() is the first derivative of G(). n is the number of iterations.

2-6) Normalization processing: w(n+1)=w(n+1)/||w(n+1)||

2-7) If the signal does not converge, return to step (5). If converged, the separation matrix W is obtained.

2-8) The mixed signal is multiplied to the left by the separation matrix W ^T , and the initial separation signal of the separated approximate source signal is obtained

Step (3) screening signals to obtain effective separation signals.

维矩阵，即总共

个有效信号。本发明提供的一种有效分离信号如图3所示。Since the positions of the selected probes are random, it is possible that the selected probes are all concentrated in the same area, so that the source signal components received in other areas are small, resulting in too much separation signal error and no reference value. Therefore, the maximum separation signal amplitude is used as a reference. , keep the separation signal whose amplitude is higher than the maximum separation signal amplitude of 1‰, and set the effective separation signal after the j-th reserved amplitude as Y _j '={y _j1 ',y _j2 ',...,y _jlj '} ^T , where l _j is the number of effective separation signals obtained after the jth calculation, l _{j ≤ n} . After J operations, the effective separation signal set Y'={Y ₁ ', Y ₂ ',...,Y _J '} can be obtained, and Y is

dimension matrix, that is, the total

a valid signal. An effective separation signal provided by the present invention is shown in FIG. 3 .

Step (4) Use the SC algorithm to cluster the effectively separated signals

Input the effective separation signal set Y into the clustering algorithm for clustering, and the separation effective signal corresponding to the point closest to the cluster center is the final separation signal. So far, the internal vibration source signal of the transformer is obtained, which is recorded as Y={Y ₁ ,Y ₂ }. Specifically, the steps of the spectral clustering algorithm mainly include five parts: constructing an undirected weight graph, constructing a similarity matrix, constructing a Laplacian matrix, cutting a graph, and K-means clustering. The specific steps are as follows:

4-1) Construct an undirected weight graph g=(V, E) for the sample data set V, where V represents the point set of the graph, and E represents the edge set.

4-2) The adjacency matrix W is constructed using the full connection method based on Gaussian distance.

The weight of the edge connecting any two nodes v _i and v _j in the graph is w _ij , and the adjacency matrix W∈R ^n×n can be constructed.

4-3) Define the degree matrix D of the node as an N-order square matrix, and the expression of element d _uv in D is:

where the diagonal elements represent the sum of the weights of all edges connected to node v _u . Define the Laplacian matrix as Ls=DW.

4-4) Define the indicator vector:

h _y ＝{h ₁ ,h ₂ ,...,h _x ,...h _N }, y=1,2,...,Y ₀ , where x represents the subscript of the node, and y represents the subscript of the subset , h _xy is the indication of node x to subset y, which can be expressed as:

In the formula, |A _y | is the number of points in the subset A _y of V, which means that each subset A _y corresponds to an indicator vector h _y , and there are N elements in h _y , representing the indication of N nodes result.

The goal of graph cutting is to divide V into Y ₀ subgraphs, so that the weight of the edges of each subgraph is minimized, then the objective function is the trace optimization problem of the matrix:

H is the characteristic matrix, Ls is the Laplacian matrix, Tr() represents the sum of diagonal elements of the matrix, and I is the identity matrix.

4-5) Perform K-means clustering on the feature matrix U after the feature matrix H is unitized, and then divide V into subsets A ₁ , A ₂ ,...,A _y ,... ., A _Y .

Step (5) Input the final separation signal into GMM for abnormal diagnosis

Using the vibration signals collected during the normal operation of the transformer to perform multi-channel blind source separation calculations, the vibration signals of the winding and iron core are obtained, and the vibration signals of the winding and iron core are used as training sets for mixed Gaussian modeling, as shown in Figures 4 and 5. As shown in Fig. 6 and Fig. 7, the vibration signals of the winding and core under different operating states of the transformer are selected as the test set. The specific algorithm steps are as follows:

5-1) Construct the training sets of transformer winding and core vibration signals under normal operating conditions respectively.

5-2) Linear dimensionality reduction is performed on the training set signal by using a linear filter bank and principal component analysis (PCA), to obtain a dimensionality-reduced training set.

5-3) Iterative optimization of GMM model parameters. Initialize the parameters of M Gaussian models, use the EM algorithm to iterate the parameters, and output the GMM parameters after meeting the convergence conditions to obtain the trained GMM parameters.

5-4) Calculate according to the Bayesian Criterion (BIC), and select the parameters of the appropriate model to avoid over-fitting.

5-5) Substituting the final separation signal A of the above-mentioned transformer to be fault diagnosed into the trained GMM model, the GMM model calculates the degree of abnormality of the test sample A, and the calculation formula is:

F(A)＝1-GMM _A (X _normal )

In the formula, F(A) is the abnormal degree of the test sample, X _normal is the vibration signal of the transformer under normal state, and GMM _A (X _normal ) is the probability that the test sample A belongs to the GMM model. If the probability that the test sample A belongs to the GMM model is greater, that is, the more similar the vibration signal of A to the transformer during normal operation, the lower the abnormality of the transformer to be diagnosed is, so the above can be judged by F(A). The degree of abnormality of the transformer to be fault diagnosed.

The present invention carries out a multi-channel voiceprint signal blind source separation experiment in a 220kV substation in operation. A total of 18 vibration sensors are installed on the wall of the oil tank on the high-voltage side of the transformer to collect vibration signals. The time-frequency domain diagram of the vibration signals is shown in Figure 2. Each of the 6 channels corresponds to the three phases of windings A, B, and C, and the sampling frequency is set to 15625Hz. At the same time, a FLUKE power quality analyzer is installed to collect the electrical quantity data synchronously.

Intercept the 0.2s vibration data, apply the multi-channel blind source separation algorithm to the 18 vibration signals, set the clustering category to 2 in the clustering, and select the cluster center signal, which is the final separated winding and iron Core vibration signals: Y ₁ and Y ₂ , and their time-domain and frequency-domain waveforms are drawn as shown in Figure 3. Combined with theoretical analysis, it can be determined that the signal Y ₁ represents the vibration signal of the winding in the transformer, and Y ₂ represents the vibration signal of the iron core.

The present invention selects the separation signal of the transformer multi-channel vibration signal collected at 5:00-7:00 on a certain day as the training set, selects the sample duration of each vibration signal as 1s, and then passes the winding and iron core data generated in this time period Each set has 7200 samples, which are used as the training set of the normal state, and the mixed Gaussian modeling is carried out respectively. As shown in FIGS. 4 and 5 , FIG. 6 is a three-phase voltage and current harmonic content intercepted within a certain period of time for 100 s provided by an embodiment of the present invention.

The present invention respectively selects the vibration separation signals of the 10th, 25th, and 70th s as test samples A, B, and C. It can be seen that the voltage and current harmonic content are at the normal level at the 10th s, and the voltage and current harmonic content are at the normal level at the 25th s. Much higher than the normal value, the voltage harmonic content returned to close to the normal value at the 70th second, and the average value of the three-phase current harmonic content was 5.93%, which was still slightly higher than the normal maximum value of 4%, so A was taken as a normal sample, B and C are taken as abnormal samples. Verify the diagnostic performance of the algorithm. Send the test sets A, B, and C into the GMM model of the winding and iron core respectively, and calculate their subordination probability, and the results are shown in the following table.

Table 2 Abnormal degree of test samples

It can be seen from the above table that the abnormal degree of sample A in the GMM model of winding and iron core is close to 0, indicating that the winding and iron core signal of sample A are normal signals; the abnormal degree of sample B in the GMM model of winding and iron core is close to 1 , indicating that the winding and core signals of sample B are abnormal signals; sample C has a higher degree of abnormality in the winding GMM model, and a lower degree of abnormality in the iron core GMM model, indicating that the probability of sample C's winding signal abnormality is greater , the probability of abnormal core signal of sample C is small. The voltage harmonic content of sample C at the corresponding moment is close to the normal value, and the current harmonic content is relatively large. Since the winding vibration is mainly excited by the current, and the iron core vibration is mainly excited by the voltage, the winding vibration signal of sample C should be compared with the normal state. Abnormal, the algorithm diagnosis results are consistent with the theoretical analysis.

In summary, the embodiment of the present invention can separate the winding and core vibration signals inside the transformer, and based on the separated signals, can realize the abnormal diagnosis of the transformer winding and core. Using the mixed Gaussian model to diagnose the abnormal state of the transformer can obtain more accurate results, and combined with the multi-channel blind source separation algorithm, the separate diagnosis of the winding and core states can be realized.

Those skilled in the art can understand that the accompanying drawing is only a schematic diagram of an embodiment, and the modules or processes in the accompanying drawing are not necessarily necessary for implementing the present invention.

It can be seen from the above description of the implementation manners that those skilled in the art can clearly understand that the present invention can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the essence of the technical solution of the present invention or the part that contributes to the prior art can be embodied in the form of software products, and the computer software products can be stored in storage media, such as ROM/RAM, disk , CD, etc., including several instructions to make a computer device (which may be a personal computer, server, or network device, etc.) execute the methods described in various embodiments or some parts of the embodiments of the present invention.

Each embodiment in this specification is described in a progressive manner, the same and similar parts of each embodiment can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and for relevant parts, refer to part of the description of the method embodiments. The device and system embodiments described above are only illustrative, and the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, It can be located in one place, or it can be distributed to multiple network elements. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. It can be understood and implemented by those skilled in the art without creative effort.

The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present invention can easily think of changes or Replacement should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the protection scope of the claims.

Claims (7)

1. A method for blind source separation of multi-channel voiceprint signals for abnormal diagnosis of transformers, characterized in that it comprises: Evenly arranging a plurality of probes at different positions on the oil tank wall of the transformer to be fault diagnosed, and the plurality of probes respectively collect vibration signals of the transformer; After the vibration signals collected by the plurality of probes are combined, they are input into a multi-channel vibration signal blind source separation algorithm combining spatial independent component separation method SDICA and spectral clustering SC, and the final separation signal is obtained through calculation; Using the winding and iron core vibration signals of the transformer during normal operation to construct sample data sets respectively, using the sample data sets to train the mixed Gaussian model to obtain a trained mixed Gaussian model; The final separated signal is input into the trained mixed Gaussian model, and the trained mixed Gaussian model outputs the fault diagnosis result of the transformer. 2. The method according to claim 1, characterized in that, the multiple probes are evenly arranged at different positions on the oil tank wall of the transformer to be fault diagnosed, and the multiple probes collect the vibration of the transformer respectively Signals, including: Arrange n sensor probes evenly at different positions on the oil tank wall of the transformer to be fault diagnosed, select the vibration signals collected by i probes from the sensor probes at different positions, and respectively select the vibration signals from the sensor probes at n different positions Select the vibration signals collected by i probes, n∈[2,18], and the number J of observation signal groups generated by permutation and combination is:

In the formula, J _n is the number of calculations when n probes are taken, and k _i is the calculation ratio. 3. The method according to claim 1, characterized in that, after the vibration signals collected by the plurality of probes are combined, they are input to the multi-channel combination of the spatial independent component separation method SDICA and spectral clustering SC. Vibration signal blind source separation algorithm, through calculation to obtain the final separation signal, including: Input J groups of observation signals into SDICA for calculation to obtain the initial separation signal, where the initial separation signal after the jth operation is expressed as Y _j ”={y _j1 ”,y _j2 ”,...,y _jn ”} ^T ; The initial separation signal is screened, and the separation signal whose amplitude is higher than the maximum separation signal amplitude of 1‰ is retained to obtain an effective separation signal. It is assumed that the effective separation signal after the jth retained amplitude is

Where l _j is the number of effective separation signals obtained after the jth calculation, l _{j ≤ n} , after J operations, the effective separation signal set Y'={Y ₁ ',Y ₂ ',...,Y _J '}, Y is

dimension matrix, the total

a valid signal; Use the SC algorithm to cluster the effective separation signals. The effective separation signal corresponding to the point closest to the cluster center is the final separation signal. The final separation signal is used as the vibration source signal inside the transformer, which is recorded as Y={Y ₁ ,Y ₂ } . 4. The method according to claim 3, characterized in that, the J groups of observation signals are respectively input into SDICA for calculation to obtain an initial separation signal, wherein the initial separation signal after the jth operation is expressed as Y _j " ＝{y _j1 ”,y _j2 ”,...,y _jn ”} ^T , including: 2-1) performing wavelet packet decomposition on the mixed J group of observation signals; 2-2) Calculate the mutual information MI value of each subspace, and select the first five subspaces with the strongest independence, and reconstruct these subspaces;

Among them, MI is the mutual information value of the observed signals X ₁ , X ₂ , X ₃ ,...X _L , and cum() represents the high-order statistics of the signals; 2-3) Demeaning and whitening operations are performed on the observed signal X; 2-4) Set the number of source signals and the number of iterations, adopt positive definite blind source separation, that is, the number of source signals is equal to the number of observation signals, where the relationship between source signal S(t) and mixed signal X(t) is

That is, the separated approximate source signal, G is the global matrix, and the maximization of negative entropy is the objective function of the algorithm, and its formula is:

Where W is the separation matrix, X is the observed signal, E() represents the expectation of the data, G(.) is a non-quadratic function, usually G(y)=tanh(y), v is a Gaussian variable with zero mean unit variance , _ki is a normal number, and p is the number of functions that use non-quadratic functions to approximate the negative entropy; Solve the separation matrix W so that the separation signal can maximize the objective function. When there is E{(wx) ² }=1, the objective function can be transformed into:

w is a row of the separation matrix, X _k is a certain observation signal, G(.) is a non-quadratic function, usually G(y)=tanh(y) Initialize the weight phasor w(0); 2-5) Iterate over w: w(n+1)=E{XG(w ^T (n)X)}-E{G'(w ^T (n)X)}w(n) w is a row of the separation matrix, G(.) is a non-quadratic function, usually G(y)=tanh(y), G'() is the first derivative of G(). n is the number of iterations. 2-6) Normalization processing: w(n+1)=w(n+1)/||w(n+1)|| 2-7) If the signal does not converge, return to step (2-5), if converge, obtain the separation matrix W; 2-8) The mixed signal is multiplied to the left by the separation matrix W ^T , and the initial separation signal of the separated approximate source signal is obtained

5. The method according to claim 3, wherein the effective separation signal is clustered using the SC algorithm, and the separation effective signal corresponding to the point closest to the cluster center is the final separation signal, and the final separation signal As the internal vibration source signal of the transformer, it is recorded as Y={Y ₁ ,Y ₂ }, including: 4-1) Construct an undirected weight graph g=(V, E) for the sample data set V, where V represents the point set of the graph, and E represents the edge set; 4-2) Construct the adjacency matrix W by using the full connection method based on Gaussian distance;

The weight of the edge connecting any two nodes v _i and v _j in the graph is w _ij , and the adjacency matrix W∈R ^n×n is constructed; 4-3) Define the degree matrix D of the node as an N-order square matrix, and the expression of element d _uv in D is:

Among them, the diagonal elements represent the sum of the weights of all edges connected to the node v _u , and the Laplacian matrix is defined as Ls=DW; 4-4) Define the indicator vector: h _y ＝{h ₁ ,h ₂ ,...,h _x ,...h _N }, y=1,2,...,Y ₀ , where x represents the subscript of the node, and y represents the subscript of the subset , h _xy is the indication of node x to subset y, expressed as:

In the formula, |A _y | is the number of points in the subset A _y of V, which means that each subset A _y corresponds to an indicator vector h _y , and there are N elements in h _y , representing the indication of N nodes result; The goal of graph cutting is to divide V into Y ₀ subgraphs, so that the weight of the edges of each subgraph is minimized, then the objective function is the trace optimization problem of the matrix:

H is the characteristic matrix, Ls is the Laplacian matrix, and Tr() represents the sum of the diagonal elements of the matrix 4-5) Carry out K-means clustering on the feature matrix U after the feature matrix H is unitized, and divide V into subsets A ₁ , A ₂ ,...,A _y ,..., A _Y . 6. according to the described method of claim 3,4 or 5, it is characterized in that, when described using the winding of transformer and iron core vibration signal during normal operation, construct sample data set respectively, utilize described sample data set to mix Gaussian model Perform training to obtain a trained mixed Gaussian model, including: Use the vibration signals collected during the normal operation of the transformer to perform multi-channel blind source separation calculations to obtain the vibration signals of the winding and iron core, and use the vibration signals of the winding and iron core as the training set for mixed Gaussian modeling, and select the transformer under different operating conditions The vibration signals of the winding and iron core are used as the test set; Using the linear filter bank and principal component analysis method to linearly reduce the training set signal to obtain the reduced training set; Iteratively optimize the parameters of the mixed Gaussian model, initialize the parameters of M Gaussian models, use the EM algorithm to iterate the parameters, and output the mixed Gaussian model parameters after meeting the convergence conditions, and obtain the trained mixed Gaussian model parameters. 7. The method according to claim 6, characterized in that, the described final separation signal is input to the trained mixed Gaussian model, and the trained mixed Gaussian model outputs the fault diagnosis result of the transformer, include: Substituting the final separation signal A of the transformer to be fault diagnosed into the trained mixed Gaussian model, the mixed Gaussian model calculates the abnormality of the final separation signal A, and the calculation formula is: F(A)＝1-GMM _A (X _normal ) In the formula, F(A) is the abnormal degree of the final separation signal A, X _normal is the vibration signal in the normal state of the transformer, GMM _A (X _normal ) is the probability that the final separation signal A belongs to the GMM model, if the final separation signal A The greater the probability of belonging to the GMM model, that is, the more similar the final separation signal A is to the vibration signal during normal operation of the transformer, the lower the abnormality of the transformer to be fault diagnosed.

Images