CN115438685A - Transformer fault sound feature identification method based on neural network - Google Patents

Abstract

The invention discloses a transformer fault sound feature identification method based on a neural network, which is characterized in that a noise signal and a fault sound source signal are separated by adopting a rapid independent component analysis algorithm according to the frequency characteristics of noise and fault sound, the energy distribution of sound is used as a feature vector to calculate the energy distribution of each frequency band of the fault sound, the neural network calculation is carried out on the fault feature vector according to the energy distribution fault feature vector, and the fault type of the transformer is determined according to an output value. The invention realizes the non-contact acquisition of equipment state information and the perspective of the closed equipment state, achieves the visualization and the perception of the latent defect of the transformer, can reduce the calculated amount of fault identification and improve the identification speed.

Description

Transformer fault sound feature identification method based on neural network

Technical Field

The invention relates to the technical field of transformer fault identification, in particular to a transformer fault sound feature identification method based on a neural network.

Background

In recent years, the economic and industrial development makes our country demand for electric power continuously increase, and the ultrahigh voltage large power grid becomes a new trend of electric power development. After years of operation, the failure probability of the transformer is increased continuously, and the risk of various failures such as insulation aging, component looseness and the like exists. As an electric power apparatus for electric energy conversion in an electric power system, the number of transformers is large, and the operation time is long, so that the number of transformers which have a failure is also large. Transformers are important devices in power systems and it is important to ensure their safe operation. Therefore, the running state of the transformer is analyzed, the hidden danger of the transformer is eliminated in time, and the method has important significance for the development of a power system.

However, the transformer faults are in various forms, such as the situation that a butterfly valve bolt is loosened, abraded, broken and separated, even the transformer faults caused by the fact that the broken bolt and the oil flow indicator are abraded and impurities enter the coil, the insulation faults caused by the fact that foreign matters and impurities in an oil conservator rush into an oil tank of a transformer body due to the fact that the flow speed is too high, and the like. Such failures without discharge characteristics cannot be detected by common means such as infrared, ultraviolet and radio frequency inspection.

The main method for detecting the mechanical fault of the transformer is a vibration signal analysis method, and at present, research on the vibration test method mainly focuses on development of a vibration test system and an acoustic test system and selection of vibration and acoustic test point positions. The iean university of transportation measures the vibration of the transformer core and the winding by using the piezoelectric vibration acceleration sensor at the earliest time, and the vibration acceleration sensor with the permanent magnet suction seat is widely concerned due to the advantage of easy adhesion and installation; the university of Chongqing has established an iron core vibration test system using an optical fiber Fabry-Perot sensor, and has conducted intensive research on the test system from the aspects of sensor arrangement, signal demodulation technology, light path design and the like. However, when the vibration signal analysis method is applied, the sensor needs to be adhered to the surface of the equipment, the structure and characteristics of the equipment need to be known very much when the measurement points are selected, and the measurement results can be directly influenced by the selection of different measurement points; secondly, the vibration signal analysis method can diagnose the mechanical defect, but is difficult to position, if the positioning is needed, a large number of sensors are used, and the sensors with large number are not practical to be adhered to the surface of the equipment.

The transformer fault diagnosis method has the advantages that the transformer fault generates vibration and radiates abnormal sound under electromagnetic stress, and sound frequency bands generated by different faults are different, so that another direction is provided for diagnosing the fault through the abnormal sound. However, the coincidence degree of the sound frequency band of the mechanical fault and the noise frequency band of the transformer body is high, the probability of sound confusion is increased, and the difficulty of realizing fault identification by identifying the sound is improved. The abnormal sound is difficult to accurately position by naked eyes and ears, the potential transformation facility close to the electrified potential transformation facility can also generate potential safety hazards to field personnel, if power failure detection is carried out, the abnormal sound can disappear once power failure occurs, and the troubleshooting treatment is difficult.

How to separate the noise in the mixed sound from the fault sound is a problem to be solved for realizing the transformer non-contact online fault diagnosis technology.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides a transformer fault sound feature identification method based on a neural network, which is used for solving the problem of separation of noise and fault sound in mixed sound in the prior art.

In order to achieve the purpose, the technical scheme of the invention is as follows:

a transformer fault sound feature identification method based on a neural network comprises the following steps:

collecting mixed sound, namely collecting the sound of the transformer in a power-on state through a microphone to form a mixed audio signal;

separating mixed sound, analyzing and operating the mixed audio signal, and separating a noise signal and a fault sound source signal;

extracting sound signal characteristics, calculating energy distribution of each frequency band in the fault sound signal, and taking a wavelet packet energy distribution vector and a Mel logarithmic spectrum as the characteristics of the fault sound source signal;

and identifying the fault type, namely identifying the fault type corresponding to the fault sound source signal by adopting a neural network model.

Further, in the mixed sound separation, a fast independent component analysis algorithm is adopted to optimize through an established objective function and gradually approach the fault sound source signal:

1) Constructing an objective function based on the entropy maximization algorithm as

(3-3) wherein k is _i Is a normal number, v is a normalized Gaussian variable, y is a random variable with unit variance and zero mean, G _i Is an arbitrary, actual, non-quadratic function;

2) Selecting a function G which grows slowly with independent variables according to the advantages and disadvantages of robustness _i Said G is _i Selected among the following four formulas:

G ₂ (y)＝-exp(-y ² /2)(3-5)，

wherein, constant a ₁ The value range of (a) is more than or equal to 1 ₁ Less than or equal to 2; function G when the fault sound source signal is sub-Gaussian and super-Gaussian _i Selecting formula (3-4); when all fault sound source signals are super-Gaussian signals, G _i Selecting formula (3-5); when all the fault sound source signals are sub-Gaussian signals, G _i Selecting formula (3-6); when all the fault sound source signals are skewed distribution signals, G _i Selecting formula (3-7);

3) Constructing the relation between the mixed audio signal and the fault sound source signal:

y＝W ^T X (3-8)，

(3-8) wherein y is one independent component of the fault sound source signal, W is one row of a separation matrix W, and X is a matrix of a one-to-one mixed signal;

substituting formula (3-8) for formula (3-3), and letting p =1 give:

J _G (W){E[G(W ^T X)]-E[G(V)]} ² (3-9)；

4) Solving the function J in the energy-enabling formula (3-9) _G (W) the separation matrix W to which the maximum is reached is formed by E { (W) ^T X) ² =1, the objective function is:

converting the conditional problem into an unconditional problem according to the Kuhn-Tucker condition, the objective function becomes:

F(w)＝E[G(W ^T X)]+C(||W|| ² -1) (3-11)，

W ⁺ ＝E{Xg(W ^T X)}-E{g(W ^T X)}w (3-12)，

(3-11) wherein C is a constant;

5) The method is obtained by an iterative formula of Newton method:

F(w)＝E[G(W ^T X)]+C(||w|| ² -1) (3-14)。

further, in the sound signal feature extraction, the energy distribution of each frequency band of the fault sound source signal is calculated through a wavelet packet algorithm:

1) Decomposing the fault sound source signal to frequency bands with equal width by wavelet packet transformation algorithm, decomposing n layers of the fault sound source signal, and obtaining 2 in the nth layer ⁿ And (3) carrying out n-1 layer wavelet packet decomposition on the signal in each frequency band, wherein the decomposition formula is as follows:

(3-15) and (3-16) wherein,

low frequency coefficients for the nth layer wavelet packet decomposition,

high frequency coefficient of wavelet packet decomposition for the nth layer, H _k-21 Low-pass filter coefficients for wavelet packet decomposition, G _k-21 High-pass filter coefficients for wavelet packet decomposition, P _n-1，0 (t) is the wavelet packet decomposition coefficient of the (n-1) th layer;

2) Reconstructing the wavelet packet coefficient of each frequency band, wherein the reconstruction formula of the wavelet packet of the nth layer is as follows:

(3-17) and (3-18) wherein P is _n，0 (t) low frequency signal reconstructed for n-th layer wavelet packet, P _n，1 (t) is the reconstructed high frequency signal of the nth layer wavelet packet, H _1-2k Low-pass filter coefficient, g, for wavelet packet reconstruction _1-2k The high-pass filter coefficients reconstructed for the wavelet packets,

for the low-frequency reconstruction coefficients, the coefficients,

is a high frequency reconstruction coefficient;

3) Calculating the energy of each frequency band: calculation 2 ⁿ Individual frequency band signal S _n，j Energy E of ₁ 、E ₂ … Ei, the formula is:

(3-19) formula (I), wherein S _i (t) is the original signal, P _i As discrete point amplitudes, E _i The energy of the jth frequency band is n, and the number of sampling points of the jth frequency band is n;

4) Normalizing the energy of each frequency band to construct a characteristic vector T:

further, training and parameter determination are carried out through a BP neural network algorithm in fault type identification:

1) After the structure of the BP neural network is determined, training and learning are carried out by continuously adjusting the weight value and the deviation value until the network finishes the specified mapping relation, and the activation function of the BP neural network is selected:

the input layer selects the linear activating function, the output y of the node i _i Comprises the following steps:

y _i ＝P _i (3-21)，

(3-21) formula (I), wherein P _i Selecting logarithmic S-shaped function for the activation function of the hidden layer for the input of the i node, and outputting y _i Comprises the following steps:

in the formula, n _j -the total input of node j;

the S-shaped function is selected by the output layer, the output y of the node k _k Comprises the following steps:

randomly setting network weight value delta W _kj ,Dw _ji The threshold value and the initial value of the learning factor n are input, the input sample calculates the output of the node k according to the selected activation function, the difference value between the actual output value of the calculated node k and the expected value is the error, and the total error E of the k nodes is as follows:

the network continues iterative computation according to the direction of the negative gradient of the total error E along E

Adjusting the network weight value delta w until the error E is more than the specified error E ₀ Small, that is:

(3-27) and (3-28) wherein η is a learning rate;

calculating the error delta of the output layer and the hidden layer _k And delta _j ：

Wherein, weight value correction formula W _kj 、W _ji ：

w _kj (t+1)＝w _kj (t)+ηδ _k y _j (3-31)，

w _ji (t+1)＝w _ji (t)+ηδ _j y _i (3-32)；

2) Determining the node number of an input layer and an output layer of the BP neural network, and determining the node number of a hidden layer according to the following three empirical formulas:

m＝log ₂ n (3-34)，

(3-33) - (3-35) wherein m is the number of hidden layer nodes, n is the number of input layer nodes, l is the number of output layer nodes, and a is a constant between 1 and 10;

and setting different hidden node numbers in sequence to substitute the hidden node numbers into the neural network model, and selecting the node number corresponding to the highest recognition rate of the model as the hidden node number.

Further, in the fault type identification, sound features are converted into images through a Mel logarithmic spectrogram, and then the Mel logarithmic spectrogram is analyzed by utilizing the identification capability of a convolution neural network on the images, so that the identification of abnormal sounds of the transformer is realized:

1) When the convolution neural network carries out convolution operation on an image, the pixels of the image are regarded as a matrix form, the image to be identified and a specified convolution kernel are subjected to convolution operation to obtain a characteristic diagram, and the calculation formula is as follows:

(3-41) wherein I (I, j) is input two-dimensional image data, and W (m, n) is a two-dimensional convolution kernel;

2) The convolutional neural network samples the frequency spectrogram by adopting a maximum pooling method for the characteristic map;

3) The convolutional neural network adopts a full connection process of a Softmax classification function:

y ⁱ ＝f(w ⁱ x ^i-1 +b ⁱ ) (3-42)，

(3-42) in the formula, i is the number of the network layer, y ⁱ For the output of the full connection layer, X ^i-I For input of fully connected layers, w ⁱ Is a weight coefficient, b ⁱ For the bias term, F (x) is the classification function.

Further, the fault sound source signal is decomposed by adopting a 5-layer wavelet packet algorithm, if the identification result of the neural network fault type is that energy distribution is mainly concentrated below 1000Hz, the frequency band below 1500Hz of the fault sound source signal is continuously subdivided, 9-layer wavelet packet decomposition is carried out, and the obtained sound characteristic vector is input into the neural network model for secondary identification.

Further, the method also comprises denoising processing of the noise signal, wherein an unbiased likelihood estimation of a coefficient is adopted, and a nonlinear wavelet transform threshold method of a unified threshold is determined according to the minimum variance of the denoising signal:

1) The method comprises the following steps of adopting a linear wavelet threshold method for noise signals with known noise characteristics, applying an empirical formula to determine the size of a threshold, and adopting a default threshold to determine a model, wherein the threshold is obtained by the following formula:

(3-43) where m is the length of the signal, wavelet packet transformation is used, and the threshold is given by the following equation:

(3-43) and (3-44) wherein C _k Sorting wavelet packet decomposition coefficients to obtain a first large coefficient, wherein n is the total number of the coefficients, delta is the signal noise intensity, and a is an empirical coefficient;

the processed signals are regarded as an estimation formula similar to an unknown regression function, and the extreme value estimation realizes the minimization of the maximum mean square error in a given function in a centralized way;

2) And determining the size of a threshold value by adopting a soft threshold estimation method for a noise signal with unknown noise characteristics: the maximum minimum criterion method is to determine an intermediate fixed threshold value first, and then generate an extreme value of minimum mean square error (minimax), and the calculation method is as follows:

regarding the processed signals as estimation formulas similar to unknown regression functions, and intensively realizing maximum mean square error minimization in the given functions by extreme value estimation;

3) And reserving a useful signal on a lower scale order, and eliminating a noise signal on a maximum scale order, wherein the calculation formula is as follows:

(3-46) in the formula, N is preset noise power, J is the scale magnitude, and constant 2,A is the maximum extreme point amplitude;

and setting a threshold value which is optimally matched with each level of scale magnitude for noise reduction, and reconstructing a signal according to the modulus maximum value point reserved on each level of scale after noise reduction.

The method comprises the steps of obtaining sound emitted by a transformer, adopting a rapid independent component analysis algorithm to separate a noise signal and a fault sound source signal according to the frequency characteristics of the noise and the fault sound, taking the energy distribution of the sound as a characteristic vector, calculating the energy distribution of each frequency band of the fault sound, carrying out neural network calculation on the fault characteristic vector according to the energy distribution fault characteristic vector, and determining the fault type of the transformer according to an output value. And the BP neural network is adopted to identify the sound, the low-frequency part of the sound is further subdivided under the condition of insufficient primary identification, and the sound energy distribution of the frequency band below 1500Hz is calculated for secondary identification. The invention realizes the non-contact acquisition of equipment state information and the state perspective of closed equipment, achieves the visualization and the sense of latent defects of the transformer, can reduce the calculated amount of fault identification, improves the identification speed, is matched with auxiliary operation and maintenance personnel for inspection and abnormal control, can save the labor cost, and effectively avoids the problems of wrong inspection and missed inspection possibly brought by the traditional inspection operation and maintenance mode.

Drawings

Fig. 1 is a block diagram of a transformer fault diagnosis method of the present invention.

Fig. 2 is a flowchart of wavelet packet decomposition according to the present invention.

Fig. 3 is a flow chart of the improved wavelet packet decomposition of the present invention.

FIG. 4 is a diagram of the convolution operation of the present invention.

Fig. 5 is a time domain waveform of a normal operation of a transformer.

Fig. 6 is a full-band frequency domain waveform of a transformer in normal operation.

Fig. 7 is a frequency domain waveform of normal operation of a transformer.

Fig. 8 is a frequency domain waveform of a transformer fault sound.

FIG. 9 is a time domain diagram of the FastICA algorithm separating transformer body noise from mechanical fault sounds.

FIG. 10 is a time domain diagram of the FastICA algorithm separating transformer body noise from discharge fault sounds.

FIG. 11 is a frequency domain plot comparison of the sound separated by the FastICA algorithm with the original sound.

Fig. 12 is a transformer fault sound and normal operation sound energy distribution diagram.

Fig. 13 is a disturbance sound energy distribution diagram.

FIG. 14 is a BP neural network model confusion matrix.

FIG. 15 is an energy distribution plot below 1500Hz for transformer fault sounds and normal operating sounds.

FIG. 16 is an energy distribution diagram of a disturbing sound below 1500 Hz.

FIG. 17 is a BP neural network model confusion matrix.

FIG. 18 is a graph of transformer fault sound and normal operation sound time-frequency spectra.

FIG. 19 is an interference sound time-frequency spectrum.

FIG. 20 is a convolutional neural network confusion matrix.

Detailed Description

Technical solutions in the embodiments of the present disclosure will be described clearly and completely with reference to the accompanying drawings in the embodiments of the present disclosure, and it is obvious that the described embodiments are only some embodiments of the present disclosure, not all embodiments. The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the disclosure, its application, or uses. All other embodiments, which can be derived by a person skilled in the art from the embodiments disclosed herein without making any creative effort, shall fall within the scope of the present disclosure.

The invention provides a transformer fault sound feature identification method based on a neural network, which comprises the following steps:

collecting mixed sound, namely collecting the sound of the transformer in a power-on state through a microphone to form a mixed audio signal;

separating mixed sound, analyzing and operating the mixed audio signal, and separating a noise signal and a fault sound source signal;

extracting sound signal characteristics, calculating energy distribution of each frequency band in the fault sound signal, and taking a wavelet packet energy distribution vector and a Mel logarithmic spectrum as the characteristics of the fault sound source signal;

and identifying the fault type, namely identifying the fault type corresponding to the fault sound source signal by adopting a neural network model.

The mechanical fault and the discharge fault of the transformer are two common fault types, the sound frequency of the transformer is mainly concentrated below 1000Hz, the sound discovery of the mechanical fault and the discharge fault simulated by an analysis experiment is carried out, the sound energy of the mechanical fault is mainly concentrated within 1000Hz, and a small amount of high-frequency components exist; the discharge sound has a wide frequency distribution in the audible sound range and has an energy distribution within 15 kHz.

The sound collected when the actual transformer fails is aliasing sound mixed with noise and fault sound, the noise is mainly noise of the transformer body, the fault sound is mainly mechanical fault sound, and the requirement for fault identification precision cannot be met by simply removing the noise. Depending on the frequency range of audible sounds and the sampling theorem, the sampling frequency must be above 40kHz in order not to lose sound information. However, this means that at least 4 ten thousand pieces of data need to be stored only for the is audio, and in order to fully utilize the failure feature information included in the voiceprint signal, reduce the amount of calculation for failure recognition, and increase the recognition speed, it is necessary to extract the voice feature first.

Therefore, the invention provides a fault diagnosis method based on voiceprints, which can be divided into three parts, namely mixed sound acquisition and separation, sound signal characteristic extraction and fault type identification, and is shown in figure 1. Firstly, collecting sound near a transformer by using a microphone, and then separating the noise of the transformer body from fault sound by using a Fast independent component Analysis (Fast independent Analysis FastICA) algorithm; and finally, extracting the characteristics of the sound signals and identifying different sounds by adopting a neural network model. Respectively using the wavelet packet energy distribution vector and the Mel logarithm spectrum as the characteristics of the sound, and adopting BP neural network and convolution neural network algorithms to identify the characteristics of the sound.

The fault occurring in the actual operation process of the transformer is unknown, so that the separation of aliasing signals is a difficult problem in the field of voiceprint recognition. Signal separation algorithms such as empirical mode decomposition and principal component analysis cannot separate independent signals but only uncorrelated signals. To solve such problems, the blind source separation algorithm employed in the present invention has significant advantages. The blind source separation algorithm is a method for extracting an unobserved original signal from mixed signals recorded by a plurality of microphones under the condition that the original signal and a signal mixing method are unknown.

In the mixed sound separation, the invention adopts a quick independent component analysis algorithm in a blind source separation algorithm to optimize through the established target function and gradually approximate to the fault sound source signal. The independence of the components is measured by an objective function. Firstly, an objective function J (W) reflecting the mutual independence degree of the components is established, and when a certain value enables the objective function to reach an extreme point, the W is the optimal solution. And the optimization algorithm is an effective algorithm for solving W. Usually, the statistical characteristics of the algorithm, such as robustness, consistency, etc., are determined by an objective function, and the properties of the algorithm, such as memory requirement, convergence speed, stability, etc., are determined by an optimization algorithm. The specific process is as follows:

1) Constructing an objective function based on the entropy maximization algorithm as

(3-3) wherein k is _i Is a normal number, v is a normalized Gaussian variable, y is a random variable with unit variance and zero mean, G _i Is an arbitrary, actual, non-quadratic function;

2) Selecting a function G which grows slowly with independent variables according to the advantages and disadvantages of robustness _i Said G is _i Selected among the following four formulas:

G ₂ (y)＝-exp(-y ² /2) (3-5)，

wherein, constant a ₁ The value range of (a) is more than or equal to 1 ₁ Less than or equal to 2; function G when the fault sound source signal is sub-Gaussian and super-Gaussian _i Selecting formula (3-4); when all fault sound source signals are super-Gaussian signals, G _i Selecting formula (3-5); when all of the fault sound source signals are sub-Gaussian signals, G _i Selecting formula (3-6); when all the fault sound source signals are skewed distribution signals, G _i Selecting formula (3-7);

3) Constructing the relation between the mixed audio signal and the fault sound source signal:

y＝W ^T X (3-8)，

(3-8) wherein y is one independent component of the fault sound source signal, W is one row of a separation matrix W, and X is a matrix of a one-to-one mixed signal;

substituting formula (3-8) for formula (3-3), and letting p =1 give:

J _G (W){E[G(W ^T X)]-E[G(V)]} ² (3-9)；

4) Solving the function J in the energy-enabling formula (3-9) _G (W) the separation matrix W to which the maximum is reached is formed by E { (W) ^T X) ² =1, the objective function is:

converting the conditional problem into an unconditional problem according to the Kuhn-Tucker condition, the objective function becomes:

F(w)＝E[G(W ^T X)]+C(||W|| ² -1) (3-11)，

W ⁺ ＝E{Xg(W ^T X)}-E{g(W ^T X)}w (3-12)，

(3-11) wherein C is a constant;

5) From the iterative formula of newton's method:

F(w)＝E[G(W ^T X)]+C(||w|| ² -1) (3-14)。

in the sound signal feature extraction, the energy distribution of each frequency band of the fault sound source signal is calculated through a wavelet packet algorithm. Different kinds of sounds tend to have different energy distributions in the respective frequency bands. Therefore, the fault characteristic vector can be constructed by adopting the wavelet packet frequency band energy parameter. The sound is decomposed into frequency bands of equal width by a wavelet packet transform algorithm. Decomposing the n layers of the signal to obtain 2 in the n-th layer ⁿ And (4) frequency bands. And reconstructing the wavelet packet coefficient of each frequency band, and then calculating the energy of each frequency band, namely the energy is the characteristic parameter representing different fault sound signals of the transformer. In order to make the characteristic parameters of the same order of magnitude, the different characteristic parameters being numerically comparable, these fault characteristic parameters are normalized. For example, when level n is chosen to be 5, the feature vector of the last sound consists of 32 relative energy values.

The energy characteristic of the sound is calculated by using a wavelet packet algorithm and comprises the following four steps of wavelet packet decomposition, wavelet packet reconstruction, energy calculation and characteristic vector construction:

1) Decomposing the fault sound source signal to frequency bands with equal width by wavelet packet transformation algorithm, decomposing n layers of the fault sound source signal, and obtaining 2 in the nth layer ⁿ And (3) carrying out n-1 layer wavelet packet decomposition on the signal in each frequency band, wherein the decomposition formula is as follows:

(3-15) and (3-16) wherein,

the low frequency coefficients for the nth layer wavelet packet decomposition,

high frequency coefficient, H, for the n-th layer wavelet packet decomposition _k-21 Low-pass filter coefficients for wavelet packet decomposition, G _k-21 High-pass filter coefficients for wavelet packet decomposition, P _n-1，0 (t) is the wavelet packet decomposition coefficient of the (n-1) th layer;

2) Reconstructing the wavelet packet coefficient of each frequency band, wherein the reconstruction formula of the wavelet packet of the nth layer is as follows:

(3-17) and (3-18) wherein P is _n，0 (t) low frequency signal reconstructed for the nth layer of wavelet packets, P _n，1 (t) is the reconstructed high frequency signal of the nth layer wavelet packet, H _1-2k Low-pass filter coefficient, g, for wavelet packet reconstruction _1-2k The high-pass filter coefficients reconstructed for the wavelet packets,

for the low-frequency reconstruction coefficients, the coefficients,

is a high frequency reconstruction coefficient;

3) Calculating each frequency binEnergy: calculation 2 ⁿ Individual frequency band signal S _n，j Energy E of ₁ 、E ₂ … Ei, the formula is:

(3-19) formula (I), wherein S _i (t) is the original signal, P _i As discrete point amplitudes, E _i The energy of the jth frequency band is n, and the number of sampling points of the jth frequency band is n;

4) Normalizing the energy of each frequency band to construct a characteristic vector T:

the BP neural network has the advantages of good self-adaption performance, high fault tolerance, strong self-learning capability, strong nonlinear capability and the like, and the invention trains and determines parameters through a BP neural network algorithm in fault type identification:

1) After the structure of the BP neural network is determined, training and learning are carried out by continuously adjusting the weight value and the deviation value until the network finishes the specified mapping relation, and the activation function of the BP neural network is selected:

the input layer selects the linear activation function, then the output y of the node i _i Comprises the following steps:

y _i ＝P _i (3-21)，

(3-21) formula (I), wherein P _i Selecting logarithmic S-shaped function for the activation function of the hidden layer for the input of the i node, and outputting y _i Comprises the following steps:

in the formula, n _j Of node jTotal input;

the S-shaped function is selected by the output layer, the output y of the node k _k Comprises the following steps:

randomly setting network weight value delta W _kj ,Dw _ji The threshold value and the initial value of the learning factor n are input, the input sample calculates the output of the node k according to the selected activation function, the difference value between the actual output value of the calculated node k and the expected value is the error, and the total error E of the k nodes is as follows:

the network continues iterative computation according to the direction of the negative gradient of the total error E along E

Adjusting the network weight value delta w until the error E is more than the specified error E ₀ Small, that is:

(3-27) and (3-28) wherein η is a learning rate;

calculating the error delta of the output layer and the hidden layer _k And delta _j ：

Wherein, weight value correction formula W _kj 、W _ji ：

w _kj (t+1)＝w _kj (t)+ηδ _k y _j (3-31)，

w _ji (t+1)＝w _ji (t)+ηδ _j y _i (3-32)；

2) The number of the feature vectors of the sound determines the number of network input layers, and the number of nodes of an output layer is determined by the number of transformer fault types. Thus, when the number n of layers is 5, the number of nodes of the input layer of the neural network is 32, and the number of nodes of the output layer is 8. Determining the node number of an input layer and an output layer of the BP neural network, and determining the node number of a hidden layer according to the following three empirical formulas:

m＝log ₂ n (3-34)，

(3-33) - (3-35) wherein m is the number of hidden layer nodes, n is the number of input layer nodes, l is the number of output layer nodes, and a is a constant between 1 and 10;

and setting different hidden node numbers in sequence to substitute the hidden node numbers into the neural network model, and selecting the node number corresponding to the highest recognition rate of the model as the hidden node number.

Compared with a spectrogram, the Mel logarithmic spectrogram is designed according to the characteristics of the human ears for receiving sound, and the Mel logarithmic spectrogram is smaller than the spectrogram while containing sound frequency domain information, so that more detail information can be reserved.

Therefore, in the fault type identification, the sound characteristics are converted into the image through the Mel logarithmic spectrogram, and the Mel logarithmic spectrogram is analyzed by utilizing the identification capability of the convolutional neural network on the image, so that the identification of the abnormal sound of the transformer is realized. The convolutional neural network mainly comprises 5 parts, namely an input layer, a convolutional layer, a pooling layer, a full-link layer and an output layer.

1) Convolutional layer

The convolution operation is a key step of the convolutional neural network. Signal features may be extracted from the image by a convolution operation, enhancing certain features of the signal and reducing noise. When an image is subjected to convolution operation, pixels of the image are regarded as a matrix form, and therefore the pixels are discontinuous. And carrying out convolution operation on the image to be identified and the specified convolution kernel. The calculation formula is as follows:

in the formula (3-41), I (I, j) is input two-dimensional image data, and W (m, n) is a two-dimensional convolution kernel.

Fig. 4 is a schematic diagram of the convolution operation. The pixel values of an SXS dimension are only 0 and 1 images, and a value is obtained after computation of a convolution kernel of 3X 3, wherein the convolution kernel is also called a filter. And performing sliding calculation on the convolution kernel in the input matrix by a certain step length, and calculating to obtain a number every time one step is slid, thereby finally obtaining a characteristic diagram. The selected convolution kernels are different, and the finally obtained feature maps are different.

2) Pooling layer

The main purpose of the pooling layer is to delete unnecessary details in the picture, reduce the size of the feature map and realize secondary extraction of the signal features. Common pooling operations are maximum, mean and random pooling. The maximum pooling is the most common pooling mode, namely the maximum value of the output characteristic value in the adjacent area, and the characteristics of the pictures can be well reserved; the average value pooling is to output the average value of the characteristic value in the adjacent areas, so that the background of the picture can be well reserved, but the picture is easy to blur; and random pooling is to output elements in the characteristic pictures at random, and the method is simple and has strong generalization capability. And selecting a maximum value pooling method to sample the spectrogram. The convolutional neural network samples the frequency spectrogram by adopting a maximum pooling method for the characteristic map;

3) Full connection layer

The fully-connected layer serves as a classification and is often located at the end of the network. The fully connected layer needs to train a classification function. The convolutional neural network adopts a full connection process of a Softmax classification function:

y ⁱ ＝f(w ⁱ x ^i-1 +b ⁱ ) (3-42)，

(3-42) in the formula, i is the network layer number, y ⁱ For the output of the full connection layer, X ^i-I For input of fully connected layers, w ⁱ Is a weight coefficient, b ⁱ For the bias term, F (x) is the classification function.

Wavelet packet decomposition is to pass the signal through high and low pass filters simultaneously and the signal is decomposed into high and low frequency signals as shown in fig. 2. However, as can be seen from analyzing the sound characteristics of the transformer, the mechanical fault and the interference, the frequency of the noise of the transformer body, some mechanical fault sounds and interference sounds are mainly distributed below 1000 Hz. For the sounds, if 5 layers of wavelet packet decomposition are carried out, energy is mainly distributed in the 1 st frequency band and the 2 nd frequency band, and the sounds are difficult to distinguish; if the number of wavelet packet decomposition layers is increased, such as 9-layer wavelet decomposition, 512 bands are generated. The frequency range is too many, which not only increases the difficulty of voice recognition, but also the high frequency part almost has energy of 0 and does not contain the characteristic information of voice. If wavelet transform is used, only the low frequency part is decomposed finely, and since the discharge sound contains energy distribution in the high frequency part, much information may be lost by the discharge sound. Therefore, the whole frequency interval is equally divided, and the requirement of the accuracy of the voice recognition is difficult to meet. Thereby, an improved algorithm of the wavelet packet algorithm is provided. Firstly, 5-layer wavelet packet decomposition is carried out on the sound, and the type of the sound is identified through a neural network. If the recognition result is that the energy distribution is mainly concentrated on the sounds with the frequency bands below 1000Hz, such as transformer body noise, mechanical fault sounds and the like, the frequency bands below 1 Hz of the sounds are continuously subdivided, 9 layers of wavelet packet decomposition are carried out, the obtained sound feature vectors are input into the neural network model for secondary recognition, and the recognition accuracy is improved. A schematic exploded view is shown in fig. 3.

The device sound signal has discontinuous singular points, and the noise has the same property. With regard to noise cancellation processing of such signals, if conventional fourier transform is used or fixed threshold selection is used, the processed signal coefficients become sparse.

Therefore, the invention also includes the noise elimination processing of the noise signal, which adopts the unbiased likelihood estimation of the coefficient first, and the nonlinear wavelet transform threshold method of the unified threshold is determined according to the minimum variance of the noise elimination signal:

1) The method comprises the following steps of adopting a linear wavelet threshold method for noise signals with known noise characteristics, applying an empirical formula to determine the size of a threshold, and adopting a default threshold to determine a model, wherein the threshold is obtained by the following formula:

(3-43) where m is the length of the signal, wavelet packet transformation is used, and the threshold is given by the following equation:

(3-43) and (3-44) wherein C _k The first largest coefficient after sorting wavelet packet decomposition coefficients, n is the total number of coefficients, δ is the signal noise strength, and a is an empirical coefficient.

The processed signal is regarded as an estimation formula similar to an unknown regression function, and the extreme value estimation realizes the maximum mean square error minimization in a given function in a centralized mode.

2) Determining the size of a threshold value by adopting a soft threshold estimation method for a noise signal with unknown noise characteristics: the maximum minimum criterion method is to determine an intermediate fixed threshold value first, and then generate an extreme value of minimum mean square error (minimax), and the calculation method is as follows:

regarding the processed signals as estimation formulas similar to unknown regression functions, and intensively realizing maximum mean square error minimization in the given functions by extreme value estimation;

3) And reserving a useful signal on a lower scale order, and eliminating a noise signal on a maximum scale order, wherein the calculation formula is as follows:

(3-46) in the formula, N is preset noise power, J is the scale magnitude, and constant 2,A is the maximum extreme point amplitude;

and setting a threshold value which is optimally matched with each level of scale magnitude for noise reduction, and reconstructing a signal according to the modulus maximum value point reserved on each level of scale after noise reduction.

For the method, the transformer fault simulation and experiment are further used for verification.

In order to restore the real scene of transformer failure, a transformer test is set up in a laboratory to simulate the failure sound which may occur in the transformer. Models of mechanical faults and discharge faults are placed in the transformer to simulate different types of transformer faults.

1) Mechanical failure

When the small metal parts fall into the transformer, the small metal parts and the transformer winding collide with each other under the action of an electric field to generate abnormal sound. The metal parts are put at different positions in the oil tank, 50Hz alternating current is applied to the coil in the oil tank, and the metal parts and the coil collide with each other under the action of electromagnetic force to generate a 'humming' sound. Simulating the sound of metal parts hitting at three different positions in the fuel tank:

a) The sound generated by friction between the small metal parts and the bottom of the coil in the oil tank;

b) The sound generated by friction between the small metal parts and the clamping piece;

c) The sound produced by the friction of the small metal parts and the pressure plate.

2) Discharge failure

The transformer has a complex insulation structure, and according to experience, the common discharge types in the transformer comprise the discharge of an oil channel at the end part of a winding; discharging between the insulating baffle and the oil; partial discharges in oilpaper insulation; discharging gaps at the joint of the insulating paper and the lead; creeping discharge of the insulating paper; breakdown of turn-to-turn insulation; discharge between coils, etc

The type of discharge that may appear in the transformer interior is analyzed, 3 kinds of discharge models are designed, respectively are flat plate electrode discharge, corona discharge and creeping discharge model.

3) Fault sound collection

The sound sensor is an important device in the sound fault detection process, and a microphone with poor performance loses the details of sound, so that the sound signal is distorted, and the accuracy of sound identification is reduced. The selection of sensors requires reference to the following properties:

1) Has good frequency response characteristic in the audible sound range. The frequency response reflects the relationship between the sound pressure phase and the sound frequency. The flatter the frequency response curve, the less the acoustic distortion, and the higher the performance of the sensor.

2) Higher signal-to-noise ratio. The signal-to-noise ratio is the ratio of signal to noise in the sensor. The higher the signal-to-noise ratio is, the smaller the proportion of noise doped in the representative sound signal is, the better the tone quality of the secondary playback sound is, and the higher the accuracy of sound identification is. The international electrotechnical commission. The performance of the combined amplifier with the signal-to-noise ratio larger than 90dB is better.

3) And (4) high sensitivity. Higher sensitivity can collect finer sounds.

4) The cost performance is high.

The microphone has the advantages of simplicity in operation, convenience in use and the like, and is selected as the sound collection device. The working principle of the capacitor microphone is that sound waves are received through the metal film, so that the capacitor generates ripple output current, and sound signals are acquired through signal amplification. Since the diaphragm of the condenser microphone is thin and light compared to the dynamic coil microphone, it can more accurately track the change of sound waves and reproduce sound in more detail and accurately. Condenser microphones have the best transient response and the widest frequency response range in all types of microphones, capture a rich range of sound details, and have higher output and lower noise than dynamic coil microphones. The condenser microphone has the advantages of high directivity, high sensitivity and the like. Therefore, a condenser microphone is selected as the sound collection device.

In order to comprehensively research the sound signal characteristics of the transformer, seven transformers with different voltage levels of different transformer substations are selected to collect sound, and a time domain graph with the time duration of 0.12 second is drawn, as shown in fig. 5. The horizontal axis represents the level signal amplitude obtained by converting the sound signal through the microphone, and the unit is mV, which represents the loudness of the sound. From the time domain diagram we can see the situation of the transformer waveform over time. Time domain graphs of different types of voltage class transformers are greatly different, but whether the transformers have faults or not is difficult to distinguish by only the time domain graphs. In order to compare the difference between the transformers conveniently, the noise waveform of the transformer is subjected to Fourier transform, and the frequency domain characteristics of the noise waveform are analyzed.

Fig. 6 is a frequency domain waveform diagram of a transformer in normal operation. It can be seen from the figure that the frequency domain distribution of the transformer is mainly below 1000 Hz. In order to observe the details of the frequency domain diagram of the sound of the normal operation of the transformer, only the frequency domain diagram of the sound below 1000Hz of the transformer is analyzed, as shown in FIG. 7.

Fig. 8 is a frequency domain plot of mechanical fault sounds and electrical discharge fault sounds that may occur inside a transformer. The graphs (a), (b) and (c) are frequency domain graphs of mechanical failure sounds emitted by different positions of the metal parts in the transformer. The frequency domain diagrams of three kinds of discharge sounds that may occur in the transformer are shown in (d), (e) and (f). It can be seen from the figure that the sound of friction between the metal part and the bottom of the coil in the oil tank in fig. 8 (a) is similar to the frequency domain waveform of the normal operation sound of the transformer in fig. 6, and the sound frequency is concentrated below 1000Hz, which is the difficulty of identifying the fault of the transformer. The sound loudness of corona discharge is low, and sound capability is distributed in each frequency range.

And verifying the effect of separating the body noise of the transformer and the abnormal fault sound by the FastICA algorithm. Fig. 9 is a time domain diagram of the separation of transformer body noise from mechanical fault sounds using the FastICA algorithm. The sampling rate for microphone sound acquisition is 48kHz. Respectively intercepting the transformer body noise and the mechanical fault sound of is, and drawing a time domain diagram of the sound, as shown in fig. 9 (a), wherein the abscissa is the number of sampling points of the source signal, and the ordinate is the amplitude of the source signal. S is the noise of the transformer body, and Sz is the sound of friction between the metal part and the bottom of the coil in the oil tank. The observation signals Xi, X are mixed sounds of the source signals Si, S, as shown in FIG. 9 (b). The fault signal is submerged in the noise of the transformer body and is difficult to identify. The graph (c) is obtained by separating the observation signals Xi, X by using the FastICA algorithm, comparing the obtained predicted source signals Isi, isao with the predicted source signals Isi, isz and the source signals s }, s2, and successfully extracting fault signals from the sound of the transformer.

FIG. 10 is a time domain diagram demonstrating the separation of transformer body noise and discharge sound effects by the FastICA algorithm. S1 is the transformer body noise, and S2 is the sound of the flat plate electrode discharge. Comparing fig. 10 (a), (c), the FastICA algorithm can successfully separate the discharging sound from the transformer sound.

In order to further compare the separation effect of the FastICA algorithm, the frequency domain waveforms of the discharge sound, the mechanical failure sound, and the transformer body noise extracted by the FastICA algorithm are compared with the frequency domain waveform of the original sound, as shown in fig. 11. By contrast, it is found that the mixed sound separated by the FastICA algorithm changes the amplitude of the sound frequency domain waveform, but the distribution of the frequency waveform hardly changes. The energy distribution characteristics of the sound cannot be changed after the sound extraction characteristic parameters are normalized, and the separation effect is good.

And carrying out wavelet packet transformation on the extracted fault sound, and extracting the energy distribution characteristics of the sound. The bandwidth of the sound signal is 24kHz, the sound is decomposed into 32 parts, the frequency range of each frequency band is 750Hz, the energy of each frequency band of the noise of the transformer body and the fault sound is calculated, a characteristic vector is formed, and the energy distribution diagram shown in figure 12 is drawn. It can be known from the figure that the energy of the normal operation sound of the transformer, the sound of the friction between the metal part and the bottom of the coil in the oil tank, and the sound of the friction between the metal part and the clamping piece are mainly concentrated in the first frequency band, 750Hz frequency range. The energy distribution frequency of the discharge sound is wide. The sound of the flat electrode discharge and the creeping discharge is within 11 frequency bands, and the energy is distributed within 8250Hz range. The energy distribution range of corona discharge sound is widest within 25 frequency bands, and energy distribution exists within a range of 18750 Hz.

Some transformers are placed outdoors, a lot of noise interference exists around the transformers, and in order to prevent the fault recognition system from misjudging the interference noise as that the transformer has a fault, the common noise around the transformers is collected to be used as a noise interference library. Fig. 13 is a diagram of interference noise energy distribution. Comparing 12,4-11, it can be known that most of the energy of the sound of the normal operation of the transformer, the sound of the friction between the metal part and the bottom of the coil in the oil tank, the sound of the friction between the metal part and the clamping piece, the wind sound and the footstep sound is concentrated in the first frequency band, and below 750Hz, the sound characteristic is not obvious, which may cause misjudgment in the sound recognition process.

The BP neural network model takes the energy distribution of sound as a feature vector to identify different sounds. A total of 8 sounds were identified:

1. the sound of friction between the metal part and the bottom of the coil in the oil tank;

2. the sound of friction between the metal part and the clamping piece;

3. the sound of the friction of the metal part and the pressure plate;

4. flat electrode discharge sound;

5. corona discharge sound;

6. creeping discharge sound;

7. noise of a transformer body;

8. disturbing the sound.

Wherein, the interference sound library consists of 6 sounds, namely wind sound, footstep sound, bird calling sound, automobile sound, bird repeller sound and human sound. The audio library for each sound contains 450 samples for a total of 3600 samples. The audio library was divided into 3 parts, with a 70. And training the neural network model by using the samples of the training set, and judging whether the model is over-fitted or not by using the samples of the verification set. The effect of the fault sound is identified using a sample inspection model of the test set. The accuracy of the BP neural network model to identify different sounds is plotted as a confusion matrix, as shown in fig. 14. The horizontal and vertical axis numbers correspond to the eight sounds numbers mentioned above, representing eight different sounds. The horizontal axis represents the actual type of sound signal and the vertical axis represents the type of sound predicted by the neural network model. The green area on the diagonal represents the correct audio frequency for the neural network to identify. For example, the total 450 samples of the sounds of the first sound metal part rubbing against the bottom of the coil in the oil tank are all recognized as the first sounds, and the recognition accuracy is 100%; 449 samples of the second sound metal part rubbing against the clip were recognized as the second sound, and one sample was recognized as the fourth flat electrode discharge sound with an erroneous recognition rate of 99.8%. The recognition rate of the entire model was 93.9%.

Analysis of the confusion matrix revealed that the recognition rate of the disturbing sounds was low, and 145 tones were erroneously recognized as sounds of the first metal part rubbing against the bottom of the coil in the tank. As can be seen from 12,4-11, the wind noise and footstep noise in the disturbing sound are very similar to the sound energy distribution of the friction between the metal component and the bottom of the coil in the oil tank, most of the energy is concentrated in the first frequency band, and the sound recognition is difficult.

Aiming at the problem, an improved algorithm is provided, the frequency below 1500Hz is further subdivided, 9 layers of wavelet packet decomposition are carried out, 32 frequency bands obtained by decomposition are calculated to obtain energy distribution, and the energy distribution is input into a BP neural network model for pattern recognition. The frequency range of each band after quadratic decomposition is 46.875Hz.

Fig. 15 is a graph of the energy distribution of the transformer fault sound and the normal operation sound in the frequency range below 1500 Hz. Fig. 16 is an energy distribution diagram of 6 different interfering sounds below 1500 Hz. Through comparison, the energy distribution of the normal operation sound of the transformer, the friction sound of the metal part and the bottom of the coil in the oil tank, the friction sound of the metal part and the clamping piece, the wind sound and the footstep sound is greatly different in a frequency range below 1500Hz, and the energy distribution can be used as a feature vector of sound identification.

The accuracy of the new network model for identifying different sounds is plotted as a confusion matrix as shown in fig. 17, with the energy distribution vector below 1500Hz as input. As can be seen from the graph, the neural network fault diagnosis model based on the improved feature vectors has high recognition rate, and the overall recognition rate is 99.6%. The sound recognition effect of discharging on the fourth sound flat plate electrode is poor, 6 samples are wrongly recognized as the sound of friction between a metal part and the bottom of a coil in the oil tank, 5 samples are wrongly recognized as the sound of corona discharge, and the recognition rate also reaches 97.6%.

Merr log spectral features are extracted from the speech signal and then input into a convolutional neural network model for training. FIG. 18 is a graph of transformer fault sound and normal operation sound time-frequency spectra. Fig. 19 is a time-frequency spectrum diagram of 6 kinds of interference noise. And the time-frequency spectrogram differences of transformer fault sounds, normal operation sounds and interference noises are found by comparing the pictures. Therefore, a logarithmic Mel frequency spectrum with time-frequency characteristics is adopted as the sound characteristics. The logarithmic Mel frequency spectrum is obtained by taking logarithm after Mel filtering when the frequency spectrum is linear, and the logarithmic Mel frequency spectrum is more in line with the characteristic of human ear perception sound.

Fig. 20 is a confusion matrix of a convolutional neural network model, the abscissa represents the type of sound predicted by the convolutional neural network model, and the ordinate represents the actual type of sound signal. Wherein:

MI is the sound of friction between a metal part and the bottom of a coil in the oil tank;

m2, the sound generated by friction between the metal parts and the clamping piece;

m3, sound generated by friction between the metal parts and the pressing plate;

pdl one-by-one plate electrode discharge sound;

pd2 corona discharge sound one by one;

pd3 creeping discharge sound one by one;

t is the noise of the transformer body;

background one by one interferes with noise.

The recognition rate of the voice recognition method based on the convolutional neural network to the interference noise is the lowest, and is only 87%. 10 samples were erroneously identified as sounds of rubbing of the M3 metal part against the platen, and 3 samples were erroneously identified as sounds of pd2 corona discharge. The recognition rate of the sound of M1 metal part rubbing with the bottom of a coil in an oil tank, the sound of M2 metal part rubbing with a clamping piece, pd2 corona discharge sound, pd3 creeping discharge sound and T transformer body noise is 100%.

The recognition rate of the convolutional neural network model is lower than that of the convolutional neural network model based on the improved wavelet packet-BP. However, the feature vector of the BP network model only contains frequency domain information of the sound, the feature of the sound is few, and the recognition rate is higher only when the frequency range is optimized for the recognized sound. And the Mel logarithm spectrum of the convolutional neural network model contains more sound characteristics, so that the identification effect on strange sounds is better. If only one algorithm is selected to identify the abnormal sound of the transformer, when other detection equipment such as a partial discharge detector and the like is arranged near the transformer, the abnormal sound of the transformer can be identified by using an improved wavelet packet-BP neural network algorithm, and the method is higher in calculation speed and identification rate; if no other auxiliary detection device exists and the condition of the transformer is monitored only by the sound detection device, a Mel logarithm frequency spectrum convolution neural network algorithm is adopted, and the generalization performance of the method is good.

Experimental results show that the method can better distinguish discharge, mechanical faults and transformer body noise, and can be used for fault diagnosis of the transformer in the transformer substation.

The technical solutions provided by the embodiments of the present invention are described in detail above, and the principles and embodiments of the present invention are explained herein by using specific examples, and the descriptions of the embodiments are only used to help understanding the principles of the embodiments of the present invention; meanwhile, for a person skilled in the art, according to the embodiments of the present invention, there may be variations in the specific implementation manners and application ranges, and in summary, the content of the present description should not be construed as a limitation to the present invention.

Claims (7)

1. A transformer fault sound feature identification method based on a neural network is characterized by comprising the following steps:

collecting mixed sound, namely collecting sound of the transformer in a power-on state through a microphone to form a mixed audio signal;

separating mixed sound, analyzing and operating the mixed audio signal, and separating a noise signal and a fault sound source signal;

extracting sound signal characteristics, calculating energy distribution of each frequency band in the fault sound signal, and taking a wavelet packet energy distribution vector and a Mel logarithmic spectrum as the characteristics of the fault sound source signal;

and identifying the fault type, namely identifying the fault type corresponding to the fault sound source signal by adopting a neural network model.

2. The transformer fault sound feature identification method based on the neural network as claimed in claim 1, wherein in the mixed sound separation, a fast independent component analysis algorithm is adopted to optimize through an established objective function, and gradually approximate fault sound source signals:

1) Constructing an objective function based on the entropy maximization algorithm as

(3-3) wherein k is _i Is a normal number, v is a normalized Gaussian variable, y is a random variable with unit variance and zero mean, G _i Is an arbitrary, actual, non-quadratic function;

2) Selecting a function G which grows slowly with independent variables according to the advantages and disadvantages of robustness _i Said G is _i Selected among the following four formulas:

G ₂ (y)＝-exp(-y ² /2) (3-5)，

wherein, constant a ₁ A is in a value range of 1 to a ₁ Less than or equal to 2; function G when the fault sound source signal is sub-Gaussian and super-Gaussian _i Selecting formula (3-4); when all fault sound source signals are super-Gaussian signals, G _i Selecting formula (3-5); when all the fault sound source signals are sub-Gaussian signals, G _i Selecting formula (3-6); when all the fault sound source signals are skewed distribution signals, G _i Selecting formula (3-7);

3) Constructing the relation between the mixed audio signal and the fault sound source signal:

y＝W ^T X (3-8)，

(3-8) wherein y is one independent component of the fault sound source signal, W is one row of a separation matrix W, and X is a matrix of a one-to-one mixed signal;

substituting formula (3-8) for formula (3-3), and letting p =1 give:

J _G (W){E[G(W ^T X)]-E[G(V)]} ² (3-9)；

4) Solving the function J in the energy-enabling formula (3-9) _G (W) the separation matrix W to which the maximum is reached is formed by E { (W) ^T X) ² =1, the objective function is:

converting the conditional problem into an unconditional problem according to Kuhn-Tucker conditions, the objective function becomes:

F(w)＝E[G(W ^T X)]+C(||W|| ² -1) (3-11)，

W ⁺ ＝E{Xg(W ^T X)}-E{g(W ^T X)}w (3-12)，

(3-11) wherein C is a constant;

5) From the iterative formula of newton's method:

F(w)＝E[G(W ^T X)]+C(||w|| ² -1) (3-14)。

3. the method for identifying the sound characteristics of the fault of the transformer based on the neural network as claimed in claim 2, wherein in the sound signal characteristic extraction, the energy distribution of each frequency band of the fault sound source signal is calculated by a wavelet packet algorithm:

1) Decomposing the fault sound source signal to frequency bands with equal width through a wavelet packet transformation algorithm, decomposing n layers of the fault sound source signal, and obtaining 2 in the nth layer ⁿ And (3) carrying out n-1 layer wavelet packet decomposition on the signal in each frequency band, wherein the decomposition formula is as follows:

(3-15) and (3-16) wherein,

low frequency coefficients for the nth layer wavelet packet decomposition,

high frequency coefficient of wavelet packet decomposition for the nth layer, H _k-21 Low-pass filter coefficients for wavelet packet decomposition, G _k-21 High-pass filter coefficient, P, for wavelet packet decomposition _n-1，0 (t) wavelet packet decomposition coefficient of the (n-1) th layer;

2) Reconstructing the wavelet packet coefficient of each frequency band, wherein the reconstruction formula of the wavelet packet of the nth layer is as follows:

(3-17) and (3-18) wherein P is _n，0 (t) low frequency signal reconstructed for n-th layer wavelet packet, P _n，1 (t) is the reconstructed high frequency signal of the nth layer wavelet packet, H _1-2k Low-pass filter coefficient, g, for wavelet packet reconstruction _1-2k The high-pass filter coefficients reconstructed for the wavelet packets,

for the low-frequency reconstruction coefficients, the coefficients,

is a high frequency reconstruction coefficient;

3) Calculating the energy of each frequency band: calculation 2 ⁿ Individual frequency band signal S _n，j Energy E of ₁ 、E ₂ … Ei, the formula is:

(3-19) formula (I), wherein S _i (t) is the original signal, P _i As discrete point amplitudes, E _i The energy of the jth frequency band is n, and the number of sampling points of the jth frequency band is n;

4) Normalizing the energy of each frequency band to construct a characteristic vector T:

4. the method for recognizing the sound characteristics of the fault of the transformer based on the neural network as claimed in claim 3, wherein in the fault type recognition, training and parameter determination are carried out through a BP neural network algorithm:

1) After the structure of the BP neural network is determined, training and learning are carried out by continuously adjusting the weight value and the deviation value until the network finishes the specified mapping relation, and the activation function of the BP neural network is selected:

the input layer selects the linear activating function, the output y of the node i _i Comprises the following steps:

y _i ＝P _i (3-21)，

(3-21) formula (I), wherein P _i Selecting logarithmic S-shaped function for the activation function of the hidden layer for the input of the i node, and outputting y _i Comprises the following steps:

in the formula, n _j -the total input of node j;

the S-shaped function is selected by the output layer, the output y of the node k _k Comprises the following steps:

randomly setting network weight value delta W _kj ,Dw _ji The threshold value, the initial value of the learning factor n, the output of the input sample calculation node k according to the selected activation function, and the difference between the actual output value of the calculation node k and the expected value is an error, so that the total error E of the k nodes is as follows:

the network continues iterative computation according to the direction of the negative gradient of the total error E along E

Adjusting the network weight value delta w until the error E is more than the specified error E ₀ Small, that is:

(3-27) and (3-28) wherein η is a learning rate;

calculating the error delta of the output layer and the hidden layer _k And delta _j ：

Wherein, weight value correction formula W _kj 、W _ji ：

w _kj (t+1)＝w _kj (t)+ηδ _k y _j (3-31)，

w _ji (t+1)＝w _ji (t)+ηδ _j y _i (3-32)；

2) Determining the node number of an input layer and an output layer of the BP neural network, and determining the node number of a hidden layer according to the following three empirical formulas:

m＝log ₂ n (3-34)，

(3-33) - (3-35) wherein m is the number of hidden layer nodes, n is the number of input layer nodes, l is the number of output layer nodes, and a is a constant between 1 and 10;

and setting different hidden node numbers in sequence to be substituted into the neural network model, and selecting the node number corresponding to the highest recognition rate of the model as the hidden node number.

5. The transformer fault sound feature identification method based on the neural network as claimed in claim 4, wherein in the fault type identification, the sound feature is converted into an image through a Mel logarithmic spectrogram, and then the Mel logarithmic spectrogram is analyzed by utilizing the identification capability of the convolutional neural network on the image, so that the identification of the abnormal sound of the transformer is realized:

1) When the convolution neural network carries out convolution operation on the image, the pixels of the image are regarded as a matrix form, the image to be identified and a specified convolution kernel are subjected to convolution operation to obtain a characteristic diagram, and the calculation formula is as follows:

(3-41) wherein I (I, j) is input two-dimensional image data, and W (m, n) is a two-dimensional convolution kernel;

2) The convolutional neural network samples the frequency spectrogram by adopting a maximum pooling method for the characteristic map;

3) The convolutional neural network adopts a full connection process of a Softmax classification function:

y ⁱ ＝f(w ⁱ x ^i-1 +b ⁱ ) (3-42)，

(3-42) in the formula, i is the number of the network layer, y ⁱ For the output of the full connection layer, X ^i-I For input of fully connected layers, w ⁱ Is a weight coefficient, b ⁱ For the bias term, F (x) is the classification function.

6. The transformer fault sound feature identification method based on the neural network as claimed in claim 5, wherein the fault sound source signal is decomposed by a 5-layer wavelet packet algorithm, if the identification result of the neural network fault type is that energy distribution is mainly concentrated below 1000Hz, the frequency band below 1500Hz of the fault sound source signal is continuously subdivided, 9-layer wavelet packet decomposition is performed, and the obtained sound feature vector is input into the neural network model for secondary identification.

7. The neural network-based transformer fault sound feature identification method as claimed in any one of claims 1 to 6, characterized by further comprising denoising processing on the noise signal, wherein the denoising processing adopts unbiased likelihood estimation on coefficients, and a nonlinear wavelet transform threshold method of a uniform threshold is determined according to the smallest variance of the denoising signal:

1) The method comprises the following steps of adopting a linear wavelet threshold method for noise signals with known noise characteristics, applying an empirical formula to determine the size of a threshold, and adopting a default threshold determination model, wherein the threshold is obtained by the following formula:

(3-43) where m is the length of the signal, wavelet packet transformation is used, and the threshold is given by the following equation:

(3-43) and (3-44) wherein C _k Sorting wavelet packet decomposition coefficients to obtain a first large coefficient, wherein n is the total number of the coefficients, delta is the signal noise intensity, and a is an empirical coefficient;

the processed signals are regarded as an estimation formula similar to an unknown regression function, and the extreme value estimation realizes the minimization of the maximum mean square error in a given function in a centralized way;

2) And determining the size of a threshold value by adopting a soft threshold estimation method for a noise signal with unknown noise characteristics: the maximum minimum criterion method is to determine an intermediate fixed threshold value first, and then generate an extreme value of minimum mean square error (minimax), and the calculation method is as follows:

regarding the processed signals as estimation formulas similar to unknown regression functions, and intensively realizing maximum mean square error minimization in the given functions by extreme value estimation;

3) And reserving a useful signal on a lower scale order, and eliminating a noise signal on a maximum scale order, wherein the calculation formula is as follows:

(3-46) wherein N is preset noise power, J is the scale magnitude, and a constant 2,A is the maximum extreme point amplitude;

and setting a threshold value which is optimally matched with each level of scale magnitude for noise reduction, and reconstructing a signal according to the modulus maximum value point reserved on each level of scale after noise reduction.

Images