CN115600088A - Distribution transformer fault diagnosis method based on vibration signals - Google Patents
Distribution transformer fault diagnosis method based on vibration signals Download PDFInfo
- Publication number
- CN115600088A CN115600088A CN202211266435.1A CN202211266435A CN115600088A CN 115600088 A CN115600088 A CN 115600088A CN 202211266435 A CN202211266435 A CN 202211266435A CN 115600088 A CN115600088 A CN 115600088A
- Authority
- CN
- China
- Prior art keywords
- model
- fault
- distribution transformer
- layer
- factor
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Mathematical Physics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Computational Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Mathematical Analysis (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Complex Calculations (AREA)
Abstract
The invention relates to a distribution transformer fault diagnosis method based on vibration signals, which comprises the following steps: acquiring a distribution transformer vibration signal, processing the distribution transformer vibration signal by adopting the combination of self-adaptive noise complete set empirical mode decomposition and Hilbert transform, and respectively solving marginal spectrum structure characteristic vectors of different frequency bands; constructing a non-directional complete graph weighted by a Gaussian function for the feature vector matrix, solving an adjacent matrix, and constructing a multi-channel multi-connected graph convolution neural network model for excavating deep features and classifying faults; in a graph convolution neural network model, a perturbation factor with a sine function is used for improving a Huilus optimization algorithm to optimize a Gaussian kernel bandwidth to obtain an optimal diagnosis model; and carrying out fault identification on the object to be identified through the obtained optimal diagnosis model. The method is beneficial to improving the diagnosis precision and identifying the unknown type of fault.
Description
Technical Field
The invention belongs to the technical field of power equipment fault diagnosis, and particularly relates to a distribution transformer fault diagnosis method based on vibration signals.
Background
In recent years, the grid pattern of China is greatly changed, and the initial stage of low power, low voltage and small scale is changed into the modernization development stage of large capacity, large unit, ultrahigh voltage and large scale. The development scale of the power system is continuously enlarged, the voltage level of the power equipment is continuously improved, the capacity of the power equipment is continuously enlarged, and the probability of the fault of the power equipment is also increased, so that higher requirements on the overall safety of the power system and the reliability and stability of the operation of the power equipment are necessarily provided. Distribution transformers are core devices in a distribution network, and sudden failures can cause serious safety accidents and economic losses. According to statistics, accidents caused by transformer faults are not counted in each year. Therefore, in order to ensure the normal operation of the distribution network system, the monitoring of the operation state of the distribution transformer is very necessary.
The transformer fault detection technology based on the vibration signal analysis method is a current research hotspot. The vibration method is to install displacement, speed or acceleration sensors on the supporting part, the side of the winding or the surface of the oil tank of the transformer, and analyze the running state of the transformer through vibration signals, even identify the fault type. The method has the advantages of convenient and flexible installation, no electrical connection and the like.
The existing distribution transformer fault diagnosis algorithms are various and mainly comprise two categories of artificial extraction features and artificial intelligence methods. The manual characteristic extraction is to decompose the vibration signal of the transformer, extract and replace the surface characteristic to train a strong classifier or extract some diagnostic indexes, and classify in a threshold value mode. With the development of digital intelligent technology, the deep learning method has become one of the most popular research directions at present. The artificial intelligence method directly inputs the vibration signals into the network, omits the step of feature extraction, and directly analyzes and identifies the data. In the actual operation process of the distribution transformer, the operation working condition is not invariable, and the vibration signal difference is larger due to the fluctuation of the working condition, and the distinguishing effect of the characteristics provided by manual experience is poor. In addition, some unknown types of faults (namely newly-appeared faults) often exist in the operation process of the distribution transformer, and the existing fault diagnosis algorithm has poor identification effect on the unknown types of faults.
Disclosure of Invention
The invention aims to provide a distribution transformer fault diagnosis method based on vibration signals, which is beneficial to improving the diagnosis precision and identifying unknown types of faults.
In order to realize the purpose, the invention adopts the technical scheme that: a distribution transformer fault diagnosis method based on vibration signals comprises the following steps:
the method comprises the following steps of S1, collecting distribution transformer vibration signals, processing the distribution transformer vibration signals by combining adaptive noise complete set empirical mode decomposition and Hilbert transformation, and respectively solving marginal spectrum structure characteristic vectors of different frequency bands;
s2, constructing a multidirectional complete graph weighted by a Gaussian function for the feature vector matrix, solving an adjacent matrix, and constructing a multi-channel multi-connected graph convolution neural network model for excavating deep features and classifying faults;
s3, optimizing the Gaussian kernel bandwidth by using a disturbance factor with a sine function to improve a Huilu optimization algorithm in the graph convolution neural network model to obtain an optimal diagnosis model;
and S4, carrying out fault identification on the object to be identified through the obtained optimal diagnosis model.
Further, collecting vibration signals of the surface of the distribution transformer box body, and processing the vibration signals of the distribution transformer by combining adaptive noise complete set empirical mode decomposition and Hilbert transformation, wherein the method comprises the following steps:
s11, acquiring a vibration signal on the surface of a distribution transformer box body by using a vibration signal acquisition device;
step S12, processing the vibration signal of the distribution transformer by adopting adaptive noise complete set empirical mode decomposition (CEEMDAN), wherein the specific process is as follows:
define operator E j The method comprises the steps of obtaining a j-order IMF component through EMD; n is i White gaussian noise with mean 0 and variance 1; m (.) represents a local mean operator; std (.) represents the standard deviation; x is the original signal; epsilon 0 Is a coefficient for controlling the signal-to-noise ratio of the auxiliary noise and the original signal, and generates an adaptive coefficient beta when calculating the k mode component k-1 Controlling the magnitude of the noise added to the primary allowance; when k =1, β 0 ＝ε 0 std(x)/std(E 1 (n i ) When k is not less than 2, beta k-1 ＝ε 0 std(r k )；
By calculating x i ＝x+β 0 E 1 (n i ) Obtaining a first margin r 1 ：
Wherein i =1,2., N is the number of times of noise addition;
in phase 1, i.e. k =1, the 1 st modal component is calculated:
IMF 1 ＝x-r 1 (2)
2 nd IMF 2 Expressed as:
adding adaptive noise signal r to the residue of stage 1 1 +β 1 E 2 (n i )；
In the same way, the kth IMF is obtained k Where k =3,4,.., m, m is the total number of IMF components:
IMF k ＝r k-1 -r k (4)
repeating the steps until the residual r meets the residual component termination condition; finally, the original signal x is decomposed into:
step S13, solving marginal spectrum information of an m-order IMF component obtained by decomposing a distribution transformer vibration signal CEEMDAN by using Hilbert transform, wherein the calculation process is shown as formula (6) -formula (11):
in the formula, a k (t) represents the instantaneous amplitude function of the kth modal component; phi is a unit of k (t) represents the corresponding instantaneous phase function; omega k (t) represents the corresponding instantaneous frequency; h (ω, t) represents the Hilbert spectrum; b (omega) represents a Hilbert marginal spectrum which represents the amplitude distribution condition of the signal at each instantaneous frequency;
and S14, forming a feature vector by Hilbert marginal spectrums of different frequency bands.
Further, an undirected complete graph weighted by a Gaussian function is constructed for the feature vector matrix, an adjacent matrix is solved, and a graph convolution neural network model is constructed for excavating deep features and fault classification, and the method comprises the following steps:
s21, constructing an undirected complete graph weighted by a Gaussian function for a characteristic vector matrix formed by marginal spectrum information, and solving an adjacent matrix;
taking each sample as a vertex, assuming that all the vertices have edge connections but the weights of the edges are different, the weights of the edges are calculated by a Gaussian function, and the specific values are as follows:
in the formula, A pq ＝A qp Representing the weight of the connection between two vertices, eta represents the Gaussian kernel bandwidth, X p ,X q Representing the characteristic vector of the p and q samples in the characteristic vector matrix X;
s22, constructing a multi-channel and multi-connected graph convolution neural network model, namely an improved GCN model, wherein the improved GCN model uses a plurality of independent graph convolution layers gc1, gc2 and gc4 to extract the characteristics of each channel, each channel is connected with a graph convolution layer gc5 after the characteristics of each channel are fused, and an output layer is connected with a classifier; adding a graph convolution layer gc3 in a forward propagation network, wherein the graph convolution layer gc3 is used for extracting marginal spectrum information from different scales and increasing diversity of gc4 layer node characteristics; finally, calculating a loss value by adopting a cross entropy loss function, and updating model parameters by using an Adam optimizer; GCN forward propagation is shown as equation (13) -equation (18):
model input layer:
H (1) ＝σ[D -1/2 (A+I)D -1/2 XW (1) ] (13)
model hidden layer:
H (2) ＝σ[D -1/2 (A+I)D -1/2 (H (1) W (2) +XW (3) )] (14)
H (4) ＝σ[D -1/2 (A+I)D -1/2 H (2) W (4) ] (15)
feature fusion layers for each channel:
H (5) ＝[H 1(4) ,H 2(4) ,...,H h(4) ] (16)
a model output layer:
y＝[D -1/2 (A+I)D -1/2 H (5) W (5) ] (17)
lg _ softmax classification:
in the formula, A + I is an adjacent matrix added with a self-loop; i is an identity matrix; d is a corresponding degree matrix of A + I; w (l) Represents the weight of the l-th layer; h (5) Is the feature of each channel feature fusion; h represents the number of channels; σ () represents the activation function, and ReLU () = max (0,); y = [ y) 1 ,y 2 ,...y n ]For output layer features, the dimension is equal to the number of categories n; y = [ Y = 1 ,Y 2 ,…Y n ]A probability value is output for the classifier.
Further, in a graph convolution neural network model, a perturbation factor with a sine function is used for improving a Husky optimization algorithm to optimize the Gaussian kernel bandwidth, and an optimal diagnosis model is obtained; the method comprises the following steps:
s31, improving a wolf optimization algorithm by using a disturbance factor with a sine function, and performing mathematical modeling on the predation behavior of the wolf as shown in a formula (19) -a formula (25);
the grey wolf optimization algorithm defines the first three wolfs with the best fitness in a wolf group as alpha, beta and delta respectively according to a grey wolf social grade system, and the rest are defined as alpha, beta and delta
The optimal three solutions in each generation of wolf colony guide to complete target search and position update;
F＝2γr 1 -γ (21)
C＝2r 2 (22)
wherein d represents the distance between the individual and the target;
representing an update to the gray wolf location;
representing a target vector position; t is the current iteration number;
representing a gray wolf location vector; f and C are vector coefficients; r is 1 、r 2 Is [0,1]A random number; gamma is a disturbance factor with a sine function, and is characterized in that the attenuation speed of the disturbance factor is slowed down at the initial stage of algorithm execution to improve the global search capability; in the later stage of the algorithm, the attenuation speed of the disturbance factor is increased and a smaller value is obtained to avoid the optimal solution fluctuation and accelerate the convergence of the algorithm;
s32, optimizing the bandwidth of a Gaussian function kernel by using an improved Huilus optimization algorithm, minimizing a loss value of a model verification set into a target function, and setting a search space to be (0.1-5);
and S33, iterating for multiple times until an iteration stop condition is met, namely the loss value of the model verification set is not changed or reaches the set maximum iteration time, and finishing the optimization to obtain a better kernel bandwidth value.
Further, based on the obtained optimal diagnosis model, fault recognition is carried out on the object to be recognized by adopting a two-stage classification method: the first-stage classification is carried out, the peak value factor is calculated by utilizing the last-stage output result of the optimal diagnosis model, if the peak value factor exceeds a threshold value, the fault is judged to be an unknown type fault, and if the peak value factor does not exceed the threshold value, the second-stage classification is carried out; in the second class, specific faults are identified using lg _ softmax; the method specifically comprises the following steps:
s41, calculating a peak value factor of a characteristic vector output by a known type fault sample (training set) through the optimal model by using the optimal diagnosis model obtained in the S3, and setting an upper limit and a lower limit of the peak value factor;
s42, when each unknown fault sample is tested, the optimal diagnosis model is used for obtaining the last layer of output characteristic vector y, and the peak value factor f of the last layer of output characteristic vector y is calculated;
step S43, in the first-level classification, judging whether the peak value factor f obtained in the step S42 is within the upper and lower limit range of the peak value factor of the fault of the known type, if not, identifying the fault as the fault of the unknown type; otherwise, entering a second-stage classification;
and S44, in the second-level classification, calculating lg _ softmax of the last-layer output feature vector y of the model, and determining a specific fault classification.
Compared with the prior art, the invention has the following beneficial effects: the method solves the problems that the existing distribution transformer fault diagnosis method is low in diagnosis precision, difficult in feature extraction and incapable of identifying unknown faults, adopts the combination of self-adaptive noise complete set empirical mode decomposition and Hilbert transformation to process the vibration signals of the distribution transformer to obtain the structural feature vector of the marginal spectrum, combines the Hilbert marginal spectrum with the graph convolution neural network, and then uses the wolf optimization algorithm to perform optimization to obtain the optimal diagnosis model, compared with the manual feature extraction and end-to-end deep learning technology, the extracted features can be clearer, so that the diagnosis precision of the model is improved; on the basis, the peak value factor of the characteristic vector of the output layer of the model is calculated, and the unknown type fault is identified by a threshold value method, so that the problem that the unknown type fault cannot be identified by the traditional algorithm and is usually misjudged as the most similar type of the model is solved. Therefore, the invention has strong practicability and wide application prospect.
Drawings
FIG. 1 is a flow chart of a method implementation of an embodiment of the present invention;
FIG. 2 is a flow chart of distribution transformer vibration signal processing in an embodiment of the present invention;
FIG. 3 is a topological structure diagram of an undirected weighted complete graph in an embodiment of the present invention;
FIG. 4 is a schematic diagram of an improved graph-convolution neural network model structure in an embodiment of the present invention;
FIG. 5 is a flowchart illustrating an implementation of optimizing Gaussian kernel bandwidth by using a Huilus optimization algorithm according to an embodiment of the present invention;
FIG. 6 is a flow chart of unknown type fault prediction in an embodiment of the present invention;
FIG. 7 is a graph of the original waveform and spectrum of each state in an embodiment of the present invention;
FIG. 8 is a graph of diagnostic results in an embodiment of the present invention;
fig. 9 is a peak factor distribution graph in an embodiment of the invention.
Detailed Description
The invention is further explained below with reference to the drawings and the embodiments.
It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the disclosure. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments according to the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.
As shown in fig. 1, the present embodiment provides a distribution transformer fault diagnosis method based on vibration signals, including the following steps:
s1, acquiring a distribution transformer box body vibration signal, processing the distribution transformer vibration signal by adopting the combination of self-adaptive noise complete set empirical mode decomposition and Hilbert transformation, and respectively solving marginal spectrum structure characteristic vectors of different frequency bands. As shown in fig. 2, step S1 specifically includes the following steps:
and S11, acquiring a vibration signal of the surface of the distribution transformer by using a vibration signal acquisition device, wherein the sampling frequency is set to be 3kHz, and the sampling time is set to be 0.1S.
Step S12, processing the vibration signal of the distribution transformer by adopting a self-adaptive noise complete set empirical mode decomposition (CEEMDAN), wherein the specific process is as follows:
define operator E j Representing the acquisition of the IMF component of the j order by EMD; n is a radical of an alkyl radical i White gaussian noise with mean 0 and variance 1; m (.) represents a local mean operator; std (.) represents the standard deviation; x is the original signal; epsilon 0 Is a coefficient for controlling the signal-to-noise ratio of the auxiliary noise and the original signal, and generates a new adaptive coefficient beta when calculating the k mode component k-1 Controlling the magnitude of the noise added to the upper margin; when k =1, β 0 ＝ε 0 std(x)/std(E 1 (n i ) When k is not less than 2, beta k-1 ＝ε 0 std(r k )；
1) By calculating x i ＝x+β 0 E 1 (n i ) Obtaining a first margin r 1 ：
Where i =1,2,.., N is the number of times of noise addition.
2) In phase 1, i.e. k =1, the 1 st modal component is calculated:
IMF 1 ＝x-r 1 (2)
3) 2 nd IMF 2 Expressed as:
adding adaptive noise signal r to the residue of stage 1 1 +β 1 E 2 (n i )。
4) In the same way, the kth IMF is obtained k Where k =3,4,.., m, m is the total number of IMF components:
IMF k ＝r k-1 -r k (4)
5) Repeating the steps until the residual r meets the residual component termination condition; finally, the original signal x is decomposed into:
step S13, solving marginal spectrum information of an m-order IMF component obtained by decomposing a distribution transformer vibration signal CEEMDAN by using Hilbert transform (Hilbert), wherein the calculation process is shown as formula (6) -formula (11):
in the formula, a k (t) represents the instantaneous amplitude function of the kth modal component; phi is a unit of k (t) represents the corresponding instantaneous phase function; omega k (t) represents the corresponding instantaneous frequency; h (ω, t) represents the Hilbert spectrum; b (ω) represents the Hilbert margin spectrum, which characterizes the amplitude distribution of the signal at each instantaneous frequency.
And S14, forming a feature vector by Hilbert marginal spectrums of different frequency bands.
And S2, constructing a nondirectional complete graph weighted by the Gaussian function for the feature vector matrix, solving the adjacent matrix, and constructing a graph convolution neural network model for excavating deep features and classifying faults. The step S2 specifically includes the following steps:
and S21, constructing a non-directional complete graph weighted by a Gaussian function for the feature vector matrix formed by the marginal spectrum information, and solving an adjacent matrix.
In the graph G = (V, E, a) model, V and E represent sets of vertices and edges, respectively, and a represents the adjacency matrix of the graph. The method can be divided into an undirected graph and a directed graph according to the connection relation of edges; the weighted graph and the unweighted graph can be divided according to the weight relationship of the edges. The topological structure of the undirected complete graph does not change along with the change of parameters, the direction of edge connection between vertexes does not need to be considered, and the construction process is simple. The method adopts an undirected weighting complete graph, takes each sample as a vertex, assumes that all the vertices have edge connection but the weights of the edges are different, and the weights of the edge connection are calculated by a Gaussian function, and is concretely as follows:
in the formula, A pq ＝A qp Representing the weight of the connection between two vertices, eta represents the Gaussian kernel bandwidth, X p ,X q Representing the feature vectors of the p and q samples in the feature vector matrix X.
And S22, constructing a multi-channel and multi-connected graph convolution neural network (improved GCN) model by using a pytrol frame in spyder software so as to realize feature mining of different scales and multi-channel information connection and improve the quality of extracted information. In order to mine effective information reflecting fault characteristics from marginal spectrums of different frequencies, the improved GCN model constructed by the present embodiment is shown in fig. 4. The constructed improved GCN model uses a plurality of independent graph convolution layers gc1, gc2 and gc4 to extract the characteristics of each channel, each channel is connected with a graph convolution layer gc5 after the characteristics of each channel are fused, and an output layer is connected with a classifier; adding a graph convolution layer gc3 in the forward propagation network, wherein the graph convolution layer gc3 is used for extracting marginal spectrum information from different scales and increasing the diversity of the characteristics of the gc4 layer nodes; finally, calculating a loss value by adopting a cross entropy loss function, and updating model parameters by using an Adam optimizer; GCN forward propagation is shown as equation (13) -equation (18):
model input layer:
H (1) ＝σ[D -1/2 (A+I)D -1/2 XW (1) ] (13)
model hidden layer:
H (2) ＝σ[D -1/2 (A+I)D -1/2 (H (1) W (2) +XW (3) )] (14)
H (4) ＝σ[D -1/2 (A+I)D -1/2 H (2) W (4) ] (15)
feature fusion layers for each channel:
H (5) ＝[H 1(4) ,H 2(4) ,...,H h(4) ] (16)
model output layer:
y＝[D -1/2 (A+I)D -1/2 H (5) W (5) ] (17)
lg _ softmax classification:
in the formula, A + I is an adjacent matrix added with a self-loop; i is an identity matrix; d is a corresponding degree matrix of A + I; w (l) Represents the weight of the l-th layer; h (5) Is the feature of each channel feature fusion; h represents the number of channels; σ () represents the activation function, and ReLU () = max (0,); y = [ y 1 ,y 2 ,...y n ]For output layer features, the dimensionality is equal to the class number n; y = [ Y = 1 ,Y 2 ,…Y n ]A probability value is output for the classifier.
S3, optimizing the Gaussian kernel bandwidth by using a disturbance factor with a sine function to improve a Huidou optimization algorithm in the graph convolution neural network model to obtain an optimal diagnosis model; as shown in fig. 5, step S3 specifically includes the following steps:
step S31, a grey wolf optimization algorithm is improved by using disturbance factors with sine functions, and mathematical modeling is carried out on the actions of the grey wolf predation as shown in the formula (19) -formula (25).
The grey wolf optimization algorithm defines the first three wolfs with the best fitness in the wolf group as alpha, beta and delta respectively according to the grey wolf social grade system, and the rest are defined as alpha, beta and delta
And the target search and the position update are guided and completed by the optimal three solutions in each generation of wolf colony.
F＝2γr 1 -γ (21)
C＝2r 2 (22)
Wherein d represents the distance between the individual and the target;
representing an update to the gray wolf location;
representing a target vector position; t is the current iteration number;
representing a gray wolf location vector; f and C are vector coefficients; r is 1 、r 2 Is [0,1]A random number; gamma is a disturbance factor with a sine function, and is characterized in that the attenuation speed of the disturbance factor is slowed down at the initial stage of algorithm execution to improve the global search capability; and in the later stage of the algorithm, the attenuation speed of the disturbance factor is increased and a smaller value is obtained to avoid the optimal solution fluctuation and accelerate the convergence of the algorithm.
And S32, optimizing the bandwidth of a Gaussian function kernel by using an improved Huulen optimization algorithm, minimizing a loss value of a model verification set into an objective function, and setting a search space to be (0.1-5).
And S33, iterating for multiple times until an iteration stop condition is met, namely the loss value of the model verification set does not change or reaches the set maximum iteration time, and ending the optimization to obtain a better core bandwidth value.
S4, for each object to be identified with unknown faults, fault identification is carried out on the object to be identified by adopting a two-stage classification method based on the obtained optimal diagnosis model: the first-stage classification is carried out, the peak value factor is calculated by utilizing the last-stage output result of the optimal diagnosis model, if the peak value factor exceeds a threshold value, the fault is judged to be an unknown type fault, and if the peak value factor does not exceed the threshold value, the second-stage classification is carried out; in the second class, specific faults are identified using lg _ softmax. As shown in fig. 6, step S4 specifically includes the following steps:
s41, obtaining a trained optimal diagnosis model by using the S3, calculating a peak factor of a characteristic vector output by each sample of the known type fault samples (training sets) of the training sets through the optimal model, and setting upper and lower limits of the peak factor;
s42, when each unknown fault sample is tested, the optimal diagnosis model is used for obtaining the last layer of output characteristic vector y, and the peak value factor f of the last layer of output characteristic vector y is calculated;
step S43, in the first-level classification, judging whether the peak value factor f obtained in the step S42 is within the upper and lower limit range of the peak value factor of the fault of the known type, if not, identifying the fault as the fault of the unknown type; otherwise, entering a second-stage classification;
and S44, in the second-level classification, calculating lg _ softmax of the last-layer output feature vector y of the model, and determining a specific fault classification.
In the embodiment, the acquired data samples are all from an oil-immersed transformer with the model number of S11-M-315/10. Fault simulation is carried out on the distribution transformer, deformation of a distribution transformer winding, a loose state and a normal state of the winding, and vibration signals of four fault states of two-point grounding of an iron core and loosening of the iron core are recorded as states 1-5 below under short-circuit experiments and no-load experiments respectively, and server parameters for calculation are as follows: I5-10200H, and 16G.
The original waveform and spectrum for each state is shown in fig. 7. The diagnostic results of the method are shown in fig. 8. And dividing the samples under the rated working condition into 1200 groups of training sets and 300 groups of verification sets for model training. The test set contains five states, each of which has 3 different operating conditions, for a total of 3000 samples. And drawing a confusion matrix for the diagnosis result, wherein 6 represents an unknown type fault, and the method can accurately identify five states of normal condition of the transformer, winding deformation, winding looseness, two-point grounding of the iron core and iron core looseness, and the accuracy rate reaches 97.73%. The diagnostic results for the different operating conditions for each state are shown in tables 1 and 2. The loose and normal winding and the two-point grounding state of the iron core are less influenced by the fluctuation of the operation condition of the transformer. Under no-load voltage fluctuation, the accuracy rate of diagnosing the loosening fault of the iron core is lower to 94.33% -95%; the accuracy of the two-point grounding fault diagnosis of the iron core is high. Through observation, a sample with misjudgment of the iron core loosening fault is misjudged as an iron core two-point grounding fault and an unknown type fault. Under the change of the load current, the diagnosis accuracy rate of the winding deformation fault fluctuates greatly, and all identification error samples are misjudged as winding looseness states and unknown faults. Therefore, the method has certain precision in diagnosis of different working conditions.
TABLE 1 diagnosis results at different load currents
TABLE 2 diagnosis results at different No-load voltages
In this embodiment, to verify the effect of the method on predicting the unknown type fault, 300 BC short-circuit fault samples are collected on the S11-M-315/10 transformer as the unknown type fault, and the two-stage classification method is used for identification. The diagnosis result is 92.76%, the peak factor distribution of the feature vector of the output layer of the optimal diagnosis model is shown in fig. 9, and the difference of 1200 known fault types (model training set samples) and 300 unknown fault types (newly-appeared fault types) on the peak factor can be known from the graph. The peak value factors obtained after the samples of the known fault types pass through the optimal diagnosis model are relatively concentrated, and the calculation results of the unknown fault types are scattered to two ends. That is, as long as a suitable threshold is selected, an unknown type of fault can be accurately identified. Their boundaries can be calculated according to statistical principles.
In summary, the distribution transformer fault diagnosis method based on vibration signals provided by the invention achieves the following technical effects:
(1) The Hilbert marginal spectrum is combined with the graph convolution neural network, compared with an artificial feature extraction and end-to-end deep learning technology, the method can enable the extracted features to be clearer, and therefore the diagnosis precision of the model is improved.
(2) Aiming at the problem that the convergence factor of the traditional grey wolf optimization algorithm is inconsistent with the actual grey wolf predation behavior, a nonlinear disturbance factor with a sine function is provided, so that the value of the convergence factor is improved at the initial stage of algorithm execution, enough disturbance is obtained to increase the global search capability of the algorithm, and the value of the convergence factor is reduced at the later stage to avoid the fluctuation of the optimal solution and accelerate the convergence. Meanwhile, the structure of the GCN model is improved, a layer of graph volume layers with different sizes are added in a forward propagation network to achieve multi-scale feature mining, and meanwhile, a plurality of parallel channels are designed to achieve multi-source information fusion.
(3) Providing a two-stage classification method to realize the identification of the unknown type faults, wherein in the first-stage classification, the peak value factors of the feature vectors of the output layers of the models are calculated, and the identification of the unknown type faults is realized by a threshold value method; in the second class, known type faults are identified by utilizing lg _ softmax, and the problem that the unknown type faults cannot be identified by a traditional algorithm and are usually judged as the class with the most similar model by mistake is solved.
The foregoing is directed to preferred embodiments of the present invention, other and further embodiments of the invention may be devised without departing from the basic scope thereof, and the scope thereof is determined by the claims that follow. However, any simple modification, equivalent change and modification of the above embodiments according to the technical essence of the present invention are within the protection scope of the technical solution of the present invention.
Claims (5)
1. A distribution transformer fault diagnosis method based on vibration signals is characterized by comprising the following steps:
the method comprises the following steps of S1, collecting distribution transformer vibration signals, processing the distribution transformer vibration signals by combining adaptive noise complete set empirical mode decomposition and Hilbert transformation, and respectively solving marginal spectrum structure characteristic vectors of different frequency bands;
s2, constructing a multidirectional complete graph weighted by a Gaussian function for the feature vector matrix, solving an adjacent matrix, and constructing a multi-channel multi-connected graph convolution neural network model for excavating deep features and classifying faults;
s3, optimizing the Gaussian kernel bandwidth by using a disturbance factor with a sine function to improve a Huilu optimization algorithm in the graph convolution neural network model to obtain an optimal diagnosis model;
and S4, carrying out fault identification on the object to be identified through the obtained optimal diagnosis model.
2. The method of claim 1, wherein the vibration signals of the surface of the box body of the distribution transformer are collected and processed by combining adaptive noise complete set empirical mode decomposition and Hilbert transform, and the method comprises the following steps:
s11, acquiring a vibration signal on the surface of a distribution transformer box body by using a vibration signal acquisition device;
step S12, processing the vibration signal of the distribution transformer by adopting adaptive noise complete set empirical mode decomposition (CEEMDAN), wherein the specific process is as follows:
define operator E j Representing the acquisition of the IMF component of the j order by EMD; n is a radical of an alkyl radical i White gaussian noise with mean 0 and variance 1; m (.) represents a local mean operator; std (. Lamda.) represents a standard deviation; x is the original signal; epsilon 0 Is a coefficient for controlling the signal-to-noise ratio of the auxiliary noise and the original signal, and generates an adaptive coefficient beta when calculating the k mode component k-1 Controlling the magnitude of the noise added to the primary allowance; when k =1, β 0 ＝ε 0 std(x)/std(E 1 (n i ) When k is not less than 2, beta k-1 ＝ε 0 std(r k )；
By calculating x i ＝x+β 0 E 1 (n i ) Obtaining a first margin r 1 ：
Wherein i =1,2., N is the number of times of noise addition;
in phase 1, i.e. k =1, the 1 st modal component is calculated:
IMF 1 ＝x-r 1 (2)
2 nd IMF 2 Expressed as:
adding adaptive noise signal r to the residue of stage 1 1 +β 1 E 2 (n i )；
In the same way, the kth IMF is obtained k Where k =3,4,.., m, m is the total number of IMF components:
IMF k ＝r k-1 -r k (4)
repeating the steps until the residual r meets the residual component termination condition; finally, the original signal x is decomposed into:
step S13, solving marginal spectrum information of an m-order IMF component obtained by decomposing a distribution transformer vibration signal CEEMDAN by using Hilbert transform, wherein the calculation process is shown as formula (6) -formula (11):
in the formula, a k (t) represents the instantaneous amplitude function of the kth modal component; phi is a unit of k (t) represents the corresponding instantaneous phase function; omega k (t) represents the corresponding instantaneous frequency; h (ω, t) represents the Hilbert spectrum; b (omega) represents a Hilbert marginal spectrum which represents the amplitude distribution condition of the signal at each instantaneous frequency;
and S14, forming a feature vector by Hilbert marginal spectrums of different frequency bands.
3. The distribution transformer fault diagnosis method based on vibration signals as claimed in claim 1, wherein a non-directional complete graph weighted by Gaussian functions is constructed for the eigenvector matrix, a adjacency matrix is obtained, and a graph convolution neural network model is constructed for mining deep features and fault classification, comprising the following steps:
s21, constructing a non-directional complete graph weighted by a Gaussian function for a feature vector matrix formed by marginal spectrum information, and solving an adjacent matrix;
taking each sample as a vertex, assuming that all the vertices have edge connections but the weights of the edges are different, the weights of the edges are calculated by a gaussian function, which is as follows:
in the formula, A pq ＝A qp Representing the weight of the connection between two vertices, η represents the Gaussian kernel bandwidth, X p ,X q Representing the characteristic vector of the p and q samples in the characteristic vector matrix X;
s22, constructing a multi-channel and multi-connected graph convolution neural network model, namely an improved GCN model, wherein the improved GCN model uses a plurality of independent graph convolution layers gc1, gc2 and gc4 to extract the characteristics of each channel, the characteristics of each channel are fused and then connected with a graph convolution layer gc5, and an output layer is connected with a classifier; adding a graph convolution layer gc3 in a forward propagation network, wherein the graph convolution layer gc3 is used for extracting marginal spectrum information from different scales and increasing diversity of gc4 layer node characteristics; finally, calculating a loss value by adopting a cross entropy loss function, and updating model parameters by using an Adam optimizer; GCN forward propagation is shown as equation (13) -equation (18):
model input layer:
H (1) ＝σ[D -1/2 (A+I)D -1/2 XW (1) ] (13)
model hidden layer:
H (2) ＝σ[D -1/2 (A+I)D -1/2 (H (1) W (2) +XW (3) )] (14)
H (4) ＝σ[D -1/2 (A+I)D -1/2 H (2) W (4) ] (15)
feature fusion layers for each channel:
H (5) ＝[H 1(4) ,H 2(4) ,...,H h(4) ] (16)
a model output layer:
y＝[D -1/2 (A+I)D -1/2 H (5) W (5) ] (17)
lg _ softmax classification:
in the formula, A + I is an adjacent matrix added with a self-loop; i is an identity matrix; d is a corresponding degree matrix of A + I; w is a group of (l) Represents the weight of the l-th layer; h (5) Is the feature of each channel feature fusion; h represents the number of channels; σ () represents the activation function, and ReLU () = max (0,); y = [ y 1 ,y 2 ,...y n ]For output layer features, the dimensionality is equal to the class number n; y = [ Y = 1 ,Y 2 ,…Y n ]A probability value is output for the classifier.
4. The distribution transformer fault diagnosis method based on the vibration signals as claimed in claim 1, characterized in that in the graph convolution neural network model, a gaussian kernel bandwidth is optimized by using a disturbance factor improved grayish wolf optimization algorithm with a sine function to obtain an optimal diagnosis model; the method comprises the following steps:
s31, improving a wolf optimization algorithm by using a disturbance factor with a sine function, and performing mathematical modeling on the predation behavior of the wolf as shown in a formula (19) -a formula (25);
the grey wolf optimization algorithm defines the first three wolfs with the best fitness in a wolf group as alpha, beta and delta respectively according to a grey wolf social grade system, and the rest are defined as alpha, beta and delta
The optimal three solutions in each generation of wolf colony guide to complete target search and position update;
F＝2γr 1 -γ (21)
C＝2r 2 (22)
wherein d represents the distance between the individual and the target;
representing an update to the gray wolf location;
representing a target vector position; t is the current iteration number;
representing a gray wolf location vector; f and C are vector coefficients; r is 1 、r 2 Is [0,1]A random number; gamma is a disturbance factor with a sine function, and is characterized in that the attenuation speed of the disturbance factor is slowed down at the initial stage of algorithm execution to improve the global search capability; in the later stage of the algorithm, the attenuation speed of the disturbance factor is increased and a smaller value is obtained to avoid the optimal solution fluctuation and accelerate the convergence of the algorithm;
s32, optimizing the bandwidth of a Gaussian function kernel by using an improved Huilus optimization algorithm, minimizing a loss value of a model verification set into a target function, and setting a search space to be (0.1-5);
and S33, iterating for multiple times until an iteration stop condition is met, namely the loss value of the model verification set is not changed or reaches the set maximum iteration time, and finishing the optimization to obtain a better kernel bandwidth value.
5. The distribution transformer fault diagnosis method based on vibration signals as claimed in claim 1, characterized in that, based on the obtained optimal diagnosis model, a two-stage classification method is adopted to perform fault recognition on an object to be recognized: the first-stage classification is carried out, the peak value factor is calculated by utilizing the output result of the last stage of the optimal diagnosis model, if the peak value factor exceeds a threshold value, the fault is judged to be an unknown type fault, otherwise, the second-stage classification is carried out; in the second class, specific faults are identified using lg _ softmax; the method specifically comprises the following steps:
s41, calculating a peak value factor of a characteristic vector output by a known type fault sample (training set) through the optimal model by using the optimal diagnosis model obtained in the S3, and setting an upper limit and a lower limit of the peak value factor;
s42, when each unknown fault sample is tested, the optimal diagnosis model is used for obtaining the last layer of output characteristic vector y, and the peak value factor f of the last layer of output characteristic vector y is calculated;
step S43, in the first-level classification, judging whether the peak value factor f obtained in the step S42 is within the upper and lower limit range of the peak value factor of the fault of the known type, if not, identifying the fault as the fault of the unknown type; otherwise, entering a second-stage classification;
and S44, in the second-level classification, calculating lg _ softmax of the last-layer output feature vector y of the model, and determining a specific fault classification.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211266435.1A CN115600088A (en) | 2022-10-15 | 2022-10-15 | Distribution transformer fault diagnosis method based on vibration signals |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211266435.1A CN115600088A (en) | 2022-10-15 | 2022-10-15 | Distribution transformer fault diagnosis method based on vibration signals |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115600088A true CN115600088A (en) | 2023-01-13 |
Family
ID=84847054
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211266435.1A Pending CN115600088A (en) | 2022-10-15 | 2022-10-15 | Distribution transformer fault diagnosis method based on vibration signals |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115600088A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116358871A (en) * | 2023-03-29 | 2023-06-30 | 哈尔滨理工大学 | Rolling bearing weak signal composite fault diagnosis method based on graph rolling network |
| CN116680556A (en) * | 2023-08-02 | 2023-09-01 | 昆明理工大学 | Method for extracting vibration signal characteristics and identifying state of water pump unit |
-
2022
- 2022-10-15 CN CN202211266435.1A patent/CN115600088A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116358871A (en) * | 2023-03-29 | 2023-06-30 | 哈尔滨理工大学 | Rolling bearing weak signal composite fault diagnosis method based on graph rolling network |
| CN116358871B (en) * | 2023-03-29 | 2024-01-23 | 哈尔滨理工大学 | Rolling bearing weak signal composite fault diagnosis method based on graph rolling network |
| CN116680556A (en) * | 2023-08-02 | 2023-09-01 | 昆明理工大学 | Method for extracting vibration signal characteristics and identifying state of water pump unit |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109271975A (en) | A kind of electrical energy power quality disturbance recognition methods based on big data multi-feature extraction synergetic classification | |
| CN115600088A (en) | Distribution transformer fault diagnosis method based on vibration signals | |
| CN107909118B (en) | Power distribution network working condition wave recording classification method based on deep neural network | |
| CN109633368A (en) | The method of duration power quality disturbances containing distributed power distribution network based on VMD and DFA | |
| CN104155585A (en) | GIS partial discharge type identification method based on GK fuzzy clustering | |
| CN108761287B (en) | Transformer partial discharge type identification method | |
| CN109597401A (en) | A kind of equipment fault diagnosis method based on data-driven | |
| CN104966161B (en) | A kind of power quality recorder data calculation and analysis methods based on gauss hybrid models | |
| CN110082082B (en) | GIS state identification method based on vibration signal principal component analysis method | |
| CN109389325B (en) | Method for evaluating state of electronic transformer of transformer substation based on wavelet neural network | |
| CN109284672A (en) | A kind of Mechanical Failure of HV Circuit Breaker diagnostic method based on PSO-Kmeans algorithm | |
| CN103558519A (en) | GIS partial discharge ultrasonic signal identification method | |
| CN110726898B (en) | Power distribution network fault type identification method | |
| CN111680875A (en) | Unmanned aerial vehicle state risk fuzzy comprehensive evaluation method based on probability baseline model | |
| CN112557826A (en) | Ship electric power system fault diagnosis method | |
| CN110737976A (en) | mechanical equipment health assessment method based on multi-dimensional information fusion | |
| CN110766313A (en) | Cable tunnel comprehensive state evaluation method based on operation and maintenance system | |
| CN110738232A (en) | grid voltage out-of-limit cause diagnosis method based on data mining technology | |
| Gu et al. | Identification of concurrent control chart patterns with singular spectrum analysis and learning vector quantization | |
| CN115331696A (en) | Multi-channel voiceprint signal blind source separation method for transformer abnormity diagnosis | |
| CN114091549A (en) | Equipment fault diagnosis method based on deep residual error network | |
| CN115587309A (en) | Method, device and equipment for extracting key features of short-circuit resistance of transformer | |
| CN112085062A (en) | Wavelet neural network-based abnormal energy consumption positioning method | |
| CN108508297A (en) | A kind of fault arc detection method based on change coefficient and SVM | |
| CN112636328B (en) | Medium-voltage distribution network equipment load state identification method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |