CN112395959B - Power transformer fault prediction and diagnosis method and system based on audio frequency characteristics - Google Patents

Abstract

The invention provides a power transformer fault prediction and diagnosis method based on audio characteristics, which specifically includes: S1. Detecting the effective signal of the audio data of the power transformer in the noise background based on the chaotic oscillator; S2. Calculating the pair on the nonlinear Mel scale Digital energy spectrum, as the characteristic quantity of the power transformer audio signal; S3. Use the principal component analysis method to calculate the principal component component of the power transformer audio signal characteristic quantity; S4. Use the quantum particle swarm algorithm to optimize the optimal hyperparameters of the vector machine algorithm to train electric power Transformer fault prediction model; S5. If the power transformer is in a fault state, the 1/3 octave algorithm is used to extract the fault characteristic frequency range amplitude, and compared with the expert experience rule base, the fault type that occurs in the power transformer is predicted/obtained. The present invention can improve the identification accuracy of fault prediction of power transformer operating status and reduce the amount of calculation; it can also perform fault judgment based on other measurement data or infrared images.

Description

Power transformer fault prediction and diagnosis method and system based on audio characteristics

Technical field

The invention belongs to the technical field of power transformer fault diagnosis, and specifically relates to a power transformer fault prediction and diagnosis method and system based on audio characteristics.

Background technique

In the power system, power transformers play important roles in voltage conversion, power distribution, voltage regulation, isolation, etc., which are related to the safety, stability, reliability and economic operation of the power system. The power transformer has a complex structure and is mainly composed of core, winding, oil tank, oil pillow, insulating bushing, tap and other parts. It is generally installed outdoors and has a harsh working environment. As the operating time increases, faults will inevitably occur, and the faults involve Windings, main insulation, leads, tap changers, bushings, etc. According to DL/T573-95 "Guidelines for Maintenance of Power Transformers" and the operating environment of power transformers, the electric power department follows the principles of "1) Power transformers newly put into operation shall be overhauled within 5 years, and thereafter every 10 years; 2) Every year Carry out at least one minor repair;" standard arranges the annual maintenance plan of the power transformer and the operation mode of the power system. Regular maintenance of power transformers can easily reduce the effective utilization rate of power transformers, waste a lot of manpower and material resources, and even cause maintenance failures, resulting in "over-repair or under-repair".

During the operation of the power transformer, mechanical deformation occurs under the action of internal current and magnetic field, which is manifested as a vibration signal through its own conduction. The vibration signal propagates through the air medium and becomes a sound signal when the power transformer is in operation. The human ear can hear sounds of 20 to 20 kHz. Experienced operation and maintenance personnel can use their ears to listen to the sound of a running power transformer to determine whether the equipment is in a fault state. However, this fault diagnosis method, which relies on the experience and subjective judgment of operation and maintenance personnel, has strong uncertainties. Based on the sound sensor device, the sound signal of the power transformer in the operating state is acquired in real time, and the audio characteristics of the power transformer in the operating state are extracted. , can detect various faults of power transformers in time and prevent them before they happen. In addition, it can also provide convenience for the operation and management of power transformers, provide a basis for maintenance, and reduce the waste of manpower and material resources. Therefore, developing a power transformer fault prediction and diagnosis system based on audio characteristics has great social significance and economic benefits.

Research on power transformer faults generally includes two directions: fault prediction and fault diagnosis. Fault diagnosis is to accurately diagnose the cause of equipment failure and provide specific treatment methods after equipment failure, but it is only applicable after equipment failure. For important equipment with high intensity and long-term operation, fault prediction can reduce maintenance costs, predict the operating status of the equipment in advance, reduce economic losses, and carry out planned maintenance to ensure normal production. Therefore, fault diagnosis and prediction of power transformers are essential, so that losses can be stopped in time and troubles can be nipped in the bud.

The operating sound of a power transformer can reflect the current health status of the equipment. Literature 1 (Du Yiming. Research on Transformer Fault Diagnosis System Based on Acoustic Signals [D]. Huazhong University of Science and Technology, 2013.) studied the sound generation mechanism of power transformers under operating conditions, and pointed out that whether the power transformer is in normal operation or fault state , its sound frequency is basically within 1000Hz. When selecting a sensor, you should focus on the sensitivity of the sensor in the low frequency range and the flatness of the frequency response curve; Document 2 (Wu Song. Research on transformer fault diagnosis based on acoustic characteristics [D]. Huazhong Technology University, 2012.) demonstrated that the characteristic frequency of the transformer discharge sound signal is 250Hz, and comparatively analyzed the effects of spectrum analysis, wavelet algorithm, Hilbert-Huang transform and other algorithms in signal extraction; Document 3 (Zhang Ruiqi. Transformer based on acoustic signals Research on discharge fault diagnosis methods [D]. Huazhong University of Science and Technology, 2018.) Wavelet packets were used to analyze the energy distribution of the sound signal in the normal operating state and spark discharge state of the power transformer. The analysis showed that 90% of the sound energy was distributed in the normal operating state. In the 0-1280Hz frequency band, the audio characteristic frequency band of pin-pin spark discharge is 5000-6000Hz, the audio characteristic frequency band of pin-plate spark discharge is 6000-8000Hz, and the audio characteristic frequency band of suspension discharge is 1000-3000Hz. Document 4 (Hou Zenqi. Research on classification and location methods of substation equipment faults based on sound characteristics [D]. North China Electric Power University, 2018.) The two-dimensional principal component analysis method is used to extract the main spectral components in the operating state of the power transformer, and the MUSIC algorithm is used Locate the fault interval; Document 5 (Deng Kai. Research on extraction and identification of characteristic parameters of transformer oil discharge acoustic signal [D]. Huazhong University of Science and Technology, 2016.) uses the wavelet threshold noise reduction method to denoise the transformer oil discharge acoustic signal, and uses The multi-scale feature entropy method extracts the characteristic quantities of discharge fault sounds; Document 6 (Jin Xiao. Research on distribution transformer fault diagnosis method based on acoustic signals [D]. Wuhan University, 2017.) uses a fast independent analysis algorithm based on negative entropy to separate For the target sound source, perform a complete set empirical mode decomposition algorithm on the target sound source to extract the singular spectral entropy that reflects the complexity and irregularity of the signal, the complete set empirical modal band energy entropy that reflects the signal energy characteristics, and the time-frequency characteristics of the signal. The marginal spectral entropy and center of gravity frequency are used as characteristic quantities.

Through summary analysis, it can be concluded that the current research on power transformer fault prediction and diagnosis methods based on sound signals faces the following shortcomings:

(1) Most of the extraction of operating sound features of power transformers uses frequency domain analysis methods, or uses wavelet transform, Hilbert-Huang transform, etc. to extract certain components that reflect faults, without drawing on the sound recognition principle of the human ear. Spectrum, Mel cepstral angle for further analysis;

(2) The structure of the power transformer is complex and the operating environment is harsh. The background noise cannot be ignored. Especially when the background noise energy is large, the sound component reflecting the fault characteristics of the power transformer may be drowned. However, the existing transformer fault prediction based on sound signals Most diagnostic methods only use simple filtering algorithms to filter out noise;

(3) Due to the limitations of existing sound signal feature extraction methods, artificial intelligence methods such as neural networks and Support Vector Machine (SVM) are not yet common in power transformer fault prediction based on sound signals.

Contents of the invention

The purpose of the present invention is to address the above problems and provide a power transformer fault prediction and diagnosis method and system based on audio characteristics, which can solve the problems in the existing technology and monitor, predict and diagnose equipment faults online in real time based on the audio characteristics. It has higher prediction accuracy when the features are submerged in the noise and are not obvious; it reduces the waste of manpower and material resources and improves the reliability of power supply; it can also support other measurement data of power transformers.

In order to achieve the above objects, the present invention adopts the following technical solutions:

A power transformer fault prediction and diagnosis method based on audio characteristics. The specific steps include:

S1. Based on chaotic oscillator detection of effective signals of power transformer audio data under strong noise background;

S2. For the effective audio signal extracted in step S1, use Mel frequency cepstrum technology to calculate the logarithmic energy spectrum on the nonlinear Mel scale as the characteristic quantity of the power transformer audio signal; during the extraction process Perform Mel scale transformation. This nonlinear transformation makes the sound signal have higher anti-noise performance;

S3. Use the PCA principal component analysis method to calculate the principal component components of the audio signal features of the power transformer;

S4. Use the QPSO quantum particle swarm algorithm to optimize the optimal hyperparameters of the SVM vector machine algorithm to train the power transformer fault prediction model;

S5. If the power transformer is in a fault state, use the 1/3 octave band algorithm to extract the fault characteristic frequency band amplitude, and calculate the T2 and SEP statistics of the power transformer audio signal based on the 1/3 octave band amplitude. Find out the frequency range corresponding to the octave that contributes the most to the T2 or SEP statistics when the T2 or SEP statistics exceeds the threshold, as the audio characteristics of the power transformer fault, compare it with the expert experience rule base, and predict/find the occurrence of the power transformer type of fault.

Further, step S1 uses the power transformer audio signal as a weak periodic or quasi-periodic disturbance signal of the chaotic system, using the perturbation and sensitivity of the chaotic system to parameters to cause essential changes in the phase state of the chaotic system, and through calculation and identification, Detect the audio signal from the power transformer. According to the fact that when a nonlinear dynamic system is in a chaotic state and a large-scale periodic state, the corresponding system phase is completely different from the system state, the transition of the system from the chaotic phase to the large-scale periodic phase is used to detect weak periodic or quasi-periodic signals. in accordance with. The chaotic system is based on Duffing-Holmes, and the specific method is:

S101. The Duffing-Holmes oscillator equation is

Among them, k is the damping ratio, -x+x ³ is the nonlinear restoring force term, rcos(ωt) is the periodic driving force or excitation signal, r is the periodic driving force amplitude, Poincare map

There is chaos in the sense of Smale;

S102. Calculate the threshold of chaos through the Melnikov method, r/k satisfies the inequality r/k>R ^m

Among them, R ^m is the bifurcation value of the m-th harmonic orbit of parameter r/k. When r/k crosses the bifurcation value R ^m , the topological properties of the vector field of the chaotic system in the phase space will undergo a jump; when r/k When r/k is less than R ^m , the system is in a chaotic state and there is no periodic closed orbit; when r/k is greater than R ^m , the system is in a large periodic state and there is a cluster of periodic closed orbits.

Further, the specific calculation method of the Mel frequency cepstral technology in step S2 includes the following steps:

S201. Use a high-pass filter to pre-emphasize audio clips to solve the problem of high-frequency component signals losing more than low-frequency signals during transmission:

H(z)＝1-bz ^-1

Among them, b is 0.98;

S202. In order to avoid mutations at the two endpoints during discrete FFT transformation, Hamming windows are used to window the audio clips:

w[n]＝0.54-0.46cos(2nπ/L)

Among them, n represents the sampling point number, and L represents the window length;

S203. Perform N-point discrete Fourier transform DFT on each frame of time domain signal, N≥L:

S204. Set several bandpass filters H _m (k) within the sound spectrum range, 0≤m<M, where M is the number of filters;

The filter has triangular filtering characteristics, the center frequency is f(m), and the filter has equal bandwidth in the Mel frequency range; the actual frequency domain is converted to the Mel frequency domain using the following formula

m(f)＝1125ln(1+f/700)

Map the limited actual frequency range [fl, fh] to the Mel frequency range [Fmell, Fmelh] and divide it equally into M+1 parts to obtain M Mel center frequencies, and then map the obtained M Mel center frequencies to the actual frequency domain , get the center frequencies of M triangular filters in the actual frequency domain;

According to the center frequency of the actual frequency domain, determine the transfer function of the Mel filter bank, where m is the filter number:

Calculate the logarithmic energy of the frequency domain signal X(k) output by the m-th Mel filter. The expression is:

Among them, E(m) is the logarithmic energy, H _m (k) is the Mel filter bank, and X(k) is the frequency domain signal;

S205. Obtain the Mel cepstral coefficient through discrete cosine transform DCT, the expression is:

Among them: C(n) is the nth frequency cepstral coefficient, E(m) is the logarithmic energy, and M is the number of Mel filters, that is, the output dimension.

Further, step S3 uses the principal component analysis method to extract the principal component of the audio Mel cepstrum feature, using X to represent the Mel cepstrum feature matrix of the audio signal, and R to represent the correlation coefficient matrix of X. According to the characteristic equation |R-λE|= 0 calculate its eigenvalue, that is, solve the characteristic equation

r _n λ ^m +r _n-1 λ ^m-1 +…+r ₁ λ+r ₀ =0

Obtain the eigenvalues λ ₁ , λ ₂ ,…, λ _m , and arrange the eigenvalues in order from large to small, that is, λ ₁ ≥λ ₂ ≥…≥λ _m ≥0, according to the set cumulative contribution rate threshold Determine the main component; among them, the contribution rate of a single feature is The cumulative contribution of the first k eigenvalues is/> Take the first, second, ..., and p (p≤m) principal components corresponding to the eigenvalues whose cumulative contribution reaches 85 to 95%.

Further, the method of using the vector machine algorithm to train the prediction model in step S4 is:

The given training sample set is {( _xi ,y _i )|i=1,2,...,l}, where x _i ∈R ⁿ represents the input vector, y _i ∈R represents the output result, and nonlinear mapping is used. Map the input vector x _i to a higher-dimensional feature space R ^k (k>n), and construct an optimal hyperplane in this space/> Minimize the "total deviation" of all sample points from the hyperplane,

Among them, ω represents the weight coefficient vector; b is the bias constant;

Using the ε-insensitive loss function, the deviation of the sample point from the optimal hyperplane can be expressed as

c(x,y,f(x))=max(0,|y-f(x)|-ε)

Among them, ε represents the allowable error;

Add relaxation factor ξ _i , When there is an error in the division, ξ _i ,/> are all greater than 0; when there is no error in the division, ξ _i ,/> Take all 0 and convert it into a problem of minimizing the optimization objective function:

Among them, C represents the penalty factor, which is converted into a convex quadratic optimization problem; the Lagrange coefficient method is used to solve the constrained optimal problem according to the optimization condition KKT

Among them, α _i , represents the Lagrange coefficient, and SV represents the support vector.

Further, in step S4, the quantum particle swarm algorithm is used to optimize the optimal hyperparameters of SVM. Let the particle population size be m and the particle dimension be D. The position of each particle represents the current solution of the particle. The i-th particle has the following Attributes:

Current position: x _i =(x _i1 ,x _i2 ,…,x _iD );

Historical optimal position: p _i = (p _i1 , p _i2 ,...p _iD );

The average of the historical optimal positions of all particles is

The next position of the particle is updated according to the following formula

Among them, g _best is the global optimal position; α is the compression-expansion factor; and u are uniformly distributed values on (0,1); the probability of taking + and - is 0.5.

Further, in step S5, the 1/3 octave band calculates the power transformer audio signal based on the power spectrum method. The power transformer audio signal is windowed and framed according to the method described in step S202. The 1/3 octave band amplitude of each frame _The calculation results _are recorded _as

The T ² statistic measures the change of sample x in the pivot space:

Among them, Λ=(λ ₁ , λ ₂ ,..., λ _p ), is the control limit with a confidence level of α;

The common calculation method for control limits is:

Among them, F _α (p, np) is the F distribution value with p and np degrees of freedom and confidence level α;

The SEP metric measures the change in the projection of the sample vector x in the residual space:

in, Represents the control limit with a confidence level of α;

The calculation formula is:

in, i＝1,2,3，/> is the eigenvalue of the covariance matrix of X, C _α is the threshold of the standard normal distribution at the confidence level α;

The contribution graph based on ^T2 is defined as follows:

The SPE-based contribution graph is defined as follows:

in,

When it is detected that the T ² or SEP statistic exceeds the threshold, the frequency range corresponding to the largest octave of the contribution map is considered to be the fault characteristic frequency. By comparing it with expert experience rules, the possible faults of the power transformer are obtained. type.

A power transformer fault prediction and diagnosis system based on audio characteristics, including a field device layer, a sensing device layer, an edge gateway layer, an AI big data platform, and an external service layer; the sensing device layer collects the power transformer audio in real time The data is uploaded to the edge gateway layer; the edge gateway layer predicts the operating status of the power transformer based on the published prediction model and the received real-time audio data. If the power transformer is in a fault state, the edge gateway uploads the power transformer audio data to the AI The big data platform performs fault diagnosis; the AI big data platform will generate an alarm event after receiving the fault information, and at the same time, perform fault diagnosis on the power transformer based on the real-time data received; the external service layer provides power transformer fault prediction and Diagnostic system display page.

Furthermore, the AI big data platform can perform regular training and release of prediction models, regular assessment of the health status of power transformers, equipment management, data processing, data services, and equipment emergency control.

Furthermore, the equipment sensing layer will also upload real-time collected power transformer temperature data and/or infrared images to the edge gateway layer; fault detection and judgment can also be performed through other measurement data or infrared images based on the power transformer.

Compared with existing technology, the advantages of the present invention are:

1. According to the fact that the perturbation and sensitivity of the chaotic system to parameters are the characteristics that cause essential changes in the phase state of the chaotic system, the present invention adopts the method of calculation and identification to detect the effective audio signal in the operating state of the power transformer for subsequent processing. Failure prediction;

2. The present invention adopts Mel frequency cepstral technology and uses the extracted effective audio signal to calculate its logarithmic energy spectrum on the nonlinear Mel scale as the characteristic quantity of the audio signal to improve the identification accuracy of fault prediction;

3. The present invention uses the PCA method to analyze the principal components of the Mel cepstrum coefficients of the audio signal, reducing the non-main components, further improving the prediction accuracy of the model and reducing the amount of calculation;

4. The present invention uses the quantum particle swarm algorithm to optimize the hyperparameters of the SVM, effectively improving the search efficiency; when a power transformer failure is detected, the power transformer is calculated based on the 1/3 octave frequency range of the audio signal of the power transformer and based on expert experience rules. Types of faults that may occur in the transformer and provide treatment suggestions;

5. The power transformer fault prediction and diagnosis system of the present invention not only supports fault prediction and diagnosis of the operating status of the power transformer based on sound signals, but also supports judgment based on other measurement data or infrared images of the power transformer.

Other advantages, objects, and features of the present invention will be apparent in part from the description below, and in part will be understood by those skilled in the art through study and practice of the present invention.

Description of drawings

Figure 1 is a structural diagram of the power transformer fault prediction and diagnosis system of the present invention;

Figure 2 is a flow chart of the power transformer fault prediction and diagnosis method of the present invention;

Figure 3 is a diagram of the Mel cepstrum calculation steps of the present invention;

Figure 4 is a flow chart of the quantum particle swarm algorithm of the present invention;

Figure 5 is a flow chart of the 1/3 octave algorithm of the present invention.

Among them, there are field device layer 1, sensing device layer 2, edge gateway layer 3, AI big data platform 4, and external service layer 5.

Detailed ways

The utility model will be described in further detail below in conjunction with the accompanying drawings and specific implementation modes.

Example 1:

As shown in Figure 2, a power transformer fault prediction and diagnosis method based on audio features, the specific steps include:

S1. Based on chaotic oscillator detection of effective signals of power transformer audio data under strong noise background;

S2. For the effective audio signal extracted in step S1, use Mel frequency cepstrum technology to calculate the logarithmic energy spectrum on the nonlinear Mel scale as the characteristic quantity of the power transformer audio signal; during the extraction process Perform Mel scale transformation. This nonlinear transformation makes the sound signal have higher anti-noise performance;

S3. Use the PCA principal component analysis method to calculate the principal component components of the audio signal features of the power transformer;

S4. Use the QPSO quantum particle swarm algorithm to optimize the optimal hyperparameters of the SVM vector machine algorithm to train the power transformer fault prediction model;

S5. If the power transformer is in a fault state, use the 1/3 octave band algorithm to extract the fault characteristic frequency band amplitude, and calculate the T2 and SEP statistics of the power transformer audio signal based on the 1/3 octave band amplitude. Find out the frequency range corresponding to the octave that contributes the most to T ² or SEP statistics when T ² or SEP statistics exceeds the threshold, as the power transformer fault audio characteristics, compare it with the expert experience rule base, and predict or derive power Type of fault that occurs in the transformer.

Step S1 of this embodiment uses the power transformer audio signal as a weak periodic or quasi-periodic disturbance signal of the chaotic system. The perturbation and sensitivity of the chaotic system to parameters are used to cause the phase state of the chaotic system to undergo essential changes. Through calculation and identification, Detect the audio signal from the power transformer. According to the fact that when a nonlinear dynamic system is in a chaotic state and a large-scale periodic state, the corresponding system phase is completely different from the system state, the transition of the system from the chaotic phase to the large-scale periodic phase is used to detect weak periodic or quasi-periodic signals. in accordance with. The chaotic system is based on Duffing-Holmes, and the specific method is:

S101. The Duffing-Holmes oscillator equation is

Among them, k is the damping ratio, -x+x ³ is the nonlinear restoring force term, rcos(ωt) is the periodic driving force or excitation signal, r is the periodic driving force amplitude, Poincare map

There is chaos in the sense of Smale;

S102. Calculate the threshold of chaos through the Melnikov method, r/k satisfies the inequality r/k>R ^m

Among them, R ^m is the bifurcation value of the m-th harmonic orbit of parameter r/k. When r/k crosses the bifurcation value R ^m , the topological properties of the vector field of the chaotic system in the phase space will undergo a jump; when r/k When r/k is less than R ^m , the system is in a chaotic state and there is no periodic closed orbit; when r/k is greater than R ^m , the system is in a large periodic state and there is a cluster of periodic closed orbits.

As shown in Figure 3, the specific calculation method of step S2 Mel frequency cepstrum technology includes the following steps:

S201. Use a high-pass filter to pre-emphasize audio clips to solve the problem of high-frequency component signals losing more than low-frequency signals during transmission:

H(z)＝1-bz ^-1

Among them, b is 0.98;

S202. In order to avoid mutations at the two endpoints during discrete FFT transformation, Hamming windows are used to window the audio clips:

w[n]＝0.54-0.46cos(2nπ/L)

Among them, n represents the sampling point number, and L represents the window length;

S203. Perform N-point discrete Fourier transform DFT on each frame of time domain signal, N≥L:

S204. Set several bandpass filters H _m (k) within the sound spectrum range, 0≤m<M, where M is the number of filters;

The filter has triangular filtering characteristics, the center frequency is f(m), and the filter has equal bandwidth in the Mel frequency range; the actual frequency domain is converted to the Mel frequency domain using the following formula

m(f)＝1125ln(1+f/700)

Map the limited actual frequency range [fl, fh] to the Mel frequency range [Fmell, Fmelh] and divide it equally into M+1 parts to obtain M Mel center frequencies, and then map the obtained M Mel center frequencies to the actual frequency domain , get the center frequencies of M triangular filters in the actual frequency domain;

According to the center frequency of the actual frequency domain, determine the transfer function of the Mel filter bank, where m is the filter number:

Calculate the logarithmic energy of the frequency domain signal X(k) output by the m-th Mel filter. The expression is:

Among them, E(m) is the logarithmic energy, H _m (k) is the Mel filter bank, and X(k) is the frequency domain signal;

S205. Obtain the Mel cepstral coefficient through discrete cosine transform DCT, the expression is:

Among them: C(n) is the nth frequency cepstral coefficient, E(m) is the logarithmic energy, and M is the number of Mel filters, that is, the output dimension.

In step S3 of this embodiment, the principal component analysis method is used to extract the principal component of the audio Mel cepstrum feature. X represents the Mel cepstral feature matrix of the audio signal, and R represents the correlation coefficient matrix of X. According to the characteristic equation |R-λE|= 0 calculate its eigenvalue, that is, solve the characteristic equation

r _n λ ^m +r _n-1 λ ^m-1 +…+r ₁ λ+r ₀ =0

Obtain the eigenvalues λ ₁ , λ ₂ ,…, λ _m , and arrange the eigenvalues in order from large to small, that is, λ ₁ ≥λ ₂ ≥…≥λ _m ≥0, according to the set cumulative contribution rate threshold Determine the main component; among them, the contribution rate of a single feature is The cumulative contribution of the first k eigenvalues is/> Take the first, second, ..., and p (p≤m) principal components corresponding to the eigenvalues whose cumulative contribution reaches 85 to 95%.

Further, the method of using the vector machine algorithm to train the prediction model in step S4 is:

The given training sample set is {( _xi ,y _i )|i=1,2,...,l}, where x _i ∈R ⁿ represents the input vector, y _i ∈R represents the output result, and nonlinear mapping is used. Map the input vector x _i to a higher-dimensional feature space R ^k (k>n), and construct an optimal hyperplane in this space/> Minimize the "total deviation" of all sample points from the hyperplane,

Among them, ω represents the weight coefficient vector; b is the bias constant;

Using the ε-insensitive loss function, the deviation of the sample point from the optimal hyperplane can be expressed as

c(x,y,f(x))=max(0,|y-f(x)|-ε)

Among them, ε represents the allowable error;

Add relaxation factor ξ _i , When there is an error in the division, ξ _i ,/> are all greater than 0; when there is no error in the division, ξ _i ,/> Take all 0 and convert it into a problem of minimizing the optimization objective function:

Among them, C represents the penalty factor, which is converted into a convex quadratic optimization problem; the Lagrange coefficient method is used to solve the constrained optimal problem according to the optimization condition KKT.

Among them, α _i , represents the Lagrange coefficient, and SV represents the support vector.

In step S4 of this embodiment, the quantum particle swarm algorithm is used to optimize the optimal hyperparameters of SVM, a heuristic intelligent optimization algorithm that simulates the behavior of biological groups or biological evolution rules, and searches or approximates the global situation of nonlinear complex space in a random or approximately random manner. The optimal solution can effectively improve search efficiency. Take the Quantum Particle Swarm Optimization (QSPO) as an example. It cancels the attribute of particle movement direction and increases the randomness of particle position. Not only does it require fewer parameters to be set than the particle swarm algorithm, but it also reduces the particle initialization time. The impact of random location on search results. Assume the size of the particle population is m and the particle dimension is D. The position of each particle represents the current solution of the particle. The i-th particle has the following attributes:

Current position: x _i =(x _i1 ,x _i2 ,…,x _iD );

Historical optimal position: p _i = (p _i1 , p _i2 ,...p _iD );

The average of the historical optimal positions of all particles is

The next position of the particle is updated according to the following formula

Among them, g _best is the global optimal position; α is the compression-expansion factor; and u are uniformly distributed values on (0,1); the probability of taking + and - is 0.5.

The QSPO algorithm flow in this embodiment is shown in Figure 4, and its specific steps are as follows:

(1) Set parameters such as population size m, particle dimension D, compression-expansion factor α, maximum iteration number iter, particle solution space range, etc.;

(2) Initialize the current position of the particle x _i , the historical optimal position of the particle p _i , the historical fitness function value of the particle fit _pi , and the global optimal position g _best ;

(3) Calculate the current fitness function value fit _i of all particles, and search for the current optimal fitness function value fit _best of all particles and their corresponding optimal positions;

(4) Update the global historical optimal position gb _est and its corresponding historical optimal population fitness function value fit _gbest , update the historical optimal position p _i of all particles and its historical optimal population fitness fit _pi ;

(5) Repeat steps (3-4) until the maximum number of iterations iter is reached, and output the current global optimal position g _best .

As shown in Figure 5, in step S5, the 1/3 octave frequency band is used to calculate the power transformer audio signal based on the power spectrum. The power transformer audio signal is windowed and framed according to the method described in step S202. The 1/3 octave frequency of each frame is _The calculation results of _the range ^amplitude _are recorded as ),

The T ² statistic measures the change of sample x in the pivot space:

Among them, Λ=(λ ₁ , λ ₂ ,..., λ _p ), is the control limit with a confidence level of α;

The common calculation method for control limits is:

Among them, F _α (p, np) is the F distribution value with p and np degrees of freedom and confidence level α;

The SEP metric measures the change in the projection of the sample vector x in the residual space:

in, Represents the control limit with a confidence level of α;

The calculation formula is:

in, i＝1,2,3，/> is the eigenvalue of the covariance matrix of X, C _α is the threshold of the standard normal distribution at the confidence level α;

The T2-based contribution graph is defined as follows:

The SPE-based contribution graph is defined as follows:

in,

When it is detected that the T2 or SEP statistic exceeds the threshold, the frequency range corresponding to the largest octave of the contribution map is considered to be the fault characteristic frequency. By comparing it with expert experience rules, the possible fault types of the power transformer can be obtained. .

Example 2:

As shown in Figure 1, the present invention also provides a power transformer fault prediction and diagnosis system based on audio characteristics, including a field device layer 1, a sensing device layer 2, an edge gateway layer 3, an AI big data platform 4, and an external service layer. 5; The sensing device layer 2 uploads the real-time collected audio data of the power transformer to the edge gateway layer 3; the edge gateway layer 3 predicts the operating status of the power transformer based on the published prediction model and the received real-time audio data, If the power transformer is in a fault state, the edge gateway layer 3 uploads the power transformer audio data to the AI big data platform 4 for fault diagnosis; the AI big data platform 4 will generate an alarm event after receiving the fault information, and at the same time, it will generate an alarm event based on the received fault information. The obtained real-time data is used to perform fault diagnosis on the power transformer; the external service layer 5 provides a display page of the power transformer fault prediction and diagnosis system.

The AI big data platform 4 in this embodiment can perform regular training and release of prediction models, regular assessment of the health status of power transformers, equipment management, data processing, data services, and equipment emergency control.

In this embodiment, the sensing equipment layer 2 will also upload real-time collected power transformer temperature data and/or infrared images to the edge gateway layer 3; it can also perform fault detection and judgment based on other measurement data or infrared images of the power transformer. .

The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which the present invention belongs can make various modifications or additions to the described specific embodiments or substitute them in similar ways, but this will not deviate from the spirit of the present invention or exceed the definition of the appended claims. range.

Claims (9)

1. The power transformer fault prediction and diagnosis method based on the audio frequency characteristics is characterized by comprising the following specific steps:

s1, detecting effective signals of audio data of a power transformer under a noise background based on a chaotic oscillator;

s2, aiming at the audio effective signal extracted in the step S1, calculating a logarithmic energy spectrum on a nonlinear Mel scale by adopting a Mel frequency cepstrum technology, and taking the logarithmic energy spectrum as a characteristic quantity of the audio signal of the power transformer;

s3, calculating principal components of the characteristic quantity of the audio signal of the power transformer by adopting a principal component analysis method;

s4, adopting a quantum particle swarm optimization vector machine algorithm to perform super-parameter training on a power transformer fault prediction model;

s5, if the power transformer is in a fault state, extracting fault characteristic frequency range amplitude by adopting a 1/3 octave algorithm, and calculating T of an audio signal of the power transformer based on the 1/3 octave amplitude ² And SEP statistics, find T ² Or the SEP statistic exceeds a threshold value for T ² Or the frequency range corresponding to the octave with the largest SEP statistic contribution graph is used as the fault audio characteristic of the power transformer, and is compared with an expert experience rule base to predict/obtain the fault type of the power transformer;

the 1/3 octave power spectrum-based method in the step S5 calculates the power transformer audio signal, the power transformer audio signal is windowed and framed according to the method in the step S202, and the 1/3 octave amplitude calculation result of each frame is recorded as x= (x) ₁ ,x ₂ ,…,x ₂₉ ) And calculates T of the power transformer audio signal based on 1/3 octave amplitude ² And SEP statistics, T ² Statistics measure the change in principal component space of sample x:

wherein Λ= (λ) ₁ ,λ ₂ ,...,λ _p )，Is a control limit with a confidence level of alpha;

the calculation method of the control limit comprises the following steps:

wherein F is _α (p, n-p) is the F distribution value with p and n-p degrees of freedom with confidence α;

the SEP index measures the change in projection of the sample vector x in the residual space:

wherein,a control limit indicating a confidence level α;

the calculation formula of (2) is as follows:

wherein, is the eigenvalue of the covariance matrix of X, C _α A threshold value with the confidence degree alpha is distributed for standard normal;

based on T ² The definition of the contribution graph of (2) is:

SPE-based contribution graphs are defined as:

wherein,

when the T is detected ² Or SEP systemAnd after the measurement exceeds the threshold value, the frequency range corresponding to the octave with the largest contribution graph is used as the fault characteristic frequency, and the fault type possibly occurring in the power transformer is obtained by comparing the fault characteristic frequency with expert experience rules.

2. The method for predicting and diagnosing a power transformer fault based on audio features as claimed in claim 1, wherein the step S1 uses the power transformer audio signal as a weak periodic or quasi-periodic disturbance signal of the chaotic system, the power transformer audio signal is detected through calculation and identification, and the transition of the system from the chaotic phase state to the large-scale periodic phase state is used as the detection basis of the weak periodic or quasi-periodic signal, and the chaotic system is based on Duffing-Holmes, and the specific method comprises the following steps:

s101, the Duffing-Holmes oscillator equation is

Wherein k is the damping ratio, -x+x ³ R cos (ωt) is periodic strategy power or excitation signal, r is periodic strategy power amplitude, and Poincare maps chaos in the meaning of Smale horseshoe;

s102, calculating a threshold value for generating chaos by a Melnikov method, wherein R/k satisfies inequality R/k > R ^m ；

Wherein R is ^m For m harmonic rail bifurcation values of parameter R/k, when R/k crosses bifurcation value R ^m When the chaotic system jumps in the topological property of the vector field in the phase space; when R/k is smaller than R ^m When the system is in a chaotic state, a periodic closed rail does not exist; when R/k is greater than R ^m When the system is in a large period state, a cluster of period closed tracks exist.

3. The power transformer fault prediction and diagnosis method based on audio features according to claim 1, wherein the mel frequency cepstrum technique specific calculation method of step S2 comprises the following steps:

s201, pre-emphasis is carried out on the audio fragment by adopting a high-pass filter, so that the problem that the loss of a high-frequency component signal is larger than that of a low-frequency signal in the transmission process is solved:

H(z)＝1-bz ^-1

wherein b is 0.98;

s202, in order to avoid mutation at two end points during discrete FFT conversion, a Hamming window is adopted to carry out windowing treatment on the audio fragment:

w[n]＝0.54-0.46cos(2nπ/L)

wherein n represents a sampling point number, and L represents a window length;

s203, performing N-point discrete Fourier transform DFT on each frame of time domain signal, wherein N is greater than or equal to L:

s204, arranging a plurality of band-pass filters H in the sound frequency spectrum range _m (k) M is more than or equal to 0 and less than M, wherein M is the number of filters;

the filter has triangular filtering characteristics, the center frequency is f (m), and the filter has equal bandwidth in a Mel frequency range; transforming the actual frequency domain into the mel frequency domain

m(f)＝1125ln(1+f/700)

Mapping the limited actual frequency range [ fl, fh ] to the Mel frequency range [ Fsell, fsell ] and equally dividing the Mel frequency range into M+1 parts to obtain M Mel center frequencies, and mapping the M Mel center frequencies to an actual frequency domain to obtain the center frequencies of M triangular filters in the actual frequency domain;

according to the center frequency of the actual frequency domain, determining the transfer function of the Mel filter bank, wherein m is the filter sequence number:

the logarithmic energy of the frequency domain signal X (k) output by the mth Mel filter is calculated, and the expression is:

wherein E (m) is logarithmic energy, H _m (k) For a mel filter bank, X (k) is the frequency domain signal;

s205, obtaining the mel cepstrum coefficient through discrete cosine transform DCT, wherein the expression is

Wherein: c (n) is the nth frequency cepstrum coefficient, E (M) is logarithmic energy, and M is the number of Mel filters, namely the output dimension.

4. The method for predicting and diagnosing faults of a power transformer based on audio features as claimed in claim 1, wherein the step S3 extracts principal components of the audio mel-frequency cepstrum features by principal component analysis method, uses X to represent mel-frequency cepstrum feature matrix of audio signal, R to represent correlation coefficient matrix of X, calculates eigenvalues according to the eigenvalue |r- λe|=0, i.e. solves the eigenvalue

r _n λ ^m +r _n-1 λ ^m-1 +…+r ₁ λ+r ₀ ＝0

Obtaining a characteristic value lambda ₁ ,λ ₂ ,…,λ _m And arranging the characteristic values in order from large to small, i.e. lambda ₁ ≥λ ₂ ≥…≥λ _m The method comprises the steps that (1) the principal component is determined according to a set cumulative contribution rate threshold value; wherein the contribution rate of the single feature isThe first k eigenvalues add up to contribute +.>Taking the first, second and … corresponding to the characteristic value with the accumulated contribution reaching 85-95 percent,The p (p.ltoreq.m) th principal component.

5. The method for predicting and diagnosing a fault of a power transformer based on audio features as claimed in claim 1, wherein the method for training the prediction model by using a support vector machine algorithm in step S4 is as follows:

given a training sample set of { (x) _i ,y _i ) I=1, 2, …, l }, where x _i ∈R ⁿ Representing the input vector, y _i E R represents the output result, and nonlinear mapping is adoptedInput vector x _i Mapping to higher dimensional feature space R ^k (k>n) constructing an optimal hyperplane in space>The "total deviation" of all sample points from the hyperplane is minimized,

wherein ω represents a weight coefficient vector; b is a bias constant;

using the epsilon-insensitive loss function, the deviation of the sample point from the optimal hyperplane can be expressed as

c(x,y,f(x))＝max(0,|y-f(x)|-ε)

Wherein ε represents the allowable error;

adding relaxation factor xi _i 、When there is an error in the division, ζ _i 、/>Are all greater than 0; when the division is error-free, xi _i 、/>All take 0, convert to solve the optimization objective function minimization problem:

wherein, C represents penalty factor, and is converted into convex quadratic optimization problem; adopting Lagrange coefficient method and solving constraint optimization problem according to optimization condition KKT:

wherein alpha is _i 、Representing Lagrange coefficients, SV representing the support vector.

6. The method for predicting and diagnosing faults of a power transformer based on audio features as claimed in claim 1, wherein the super-parameters optimized by adopting the quantum particle swarm algorithm in the step S4 are as follows: let the particle population scale be m, the particle dimension be D, the position of each particle represents the current solution of the particle, the current position and the historical optimal position of the ith particle are:

current position: x is x _i ＝(x _i1 ,x _i2 ,…,x _iD )；

Historical optimal position: p is p _i ＝(p _i1 ,p _i2 ,…p _iD )；

The average value of the historical optimal positions of the particles is as follows:

the next position update of the particle is:

wherein g _best Is a global optimal position; alpha is a compression-expansion factor;and u is the value of the uniform distribution on (0, 1); the probability of taking the sum is 0.5.

7. An audio feature-based power transformer fault prediction and diagnosis system using the audio feature-based power transformer fault prediction and diagnosis method according to any one of claims 1 to 6, characterized by comprising a field device layer (1), a sensing device layer (2), an edge gateway layer (3), an AI big data platform (4), an external service layer (5); the sensing equipment layer (2) uploads the real-time collected power transformer audio data to the edge gateway layer (3); the edge gateway layer (3) predicts the running state of the power transformer according to the published prediction model and the received real-time audio data, and if the power transformer is in a fault state, the edge gateway layer (3) uploads the audio data of the power transformer to the AI big data platform (4) for fault diagnosis; the AI big data platform (4) generates an alarm event after receiving the fault information, and performs fault diagnosis on the power transformer according to the received real-time data; the external service layer (5) provides a display page of the power transformer fault prediction and diagnosis system.

8. The audio feature-based power transformer fault prediction and diagnosis system according to claim 7, wherein the AI big data platform (4) can perform timing training and release prediction model, evaluate power transformer health status at timing, manage equipment, process data, service data, and control equipment urgently.

9. The power transformer fault prediction and diagnosis system based on audio features according to claim 7, characterized in that the device sensing layer (2) also uploads power transformer temperature data collected in real time and/or infrared images to the edge gateway layer (3).