CN104614069A - Voice detection method of power equipment failure based on combined similar diagonalizable blind source separation algorithm - Google Patents

Abstract

Based on the joint approximate diagonal blind source separation algorithm for electric equipment fault sound detection method, the specific steps include: (1) using a microphone array; (2) using a joint approximate diagonal blind source separation algorithm for step (1) using a microphone The sound signal collected by the array is separated from each independent sound source signal; (3) the Mel frequency cepstral coefficient MFCC of the independent sound source signal is extracted as the sound characteristic parameter, and the sound signal is identified by the pattern matching algorithm, and the sound template to be tested is compared with all reference samples After the templates are matched, the reference sample template with the smallest matching distance is the result of the recognition of the working sound of the power equipment. The present invention enhances the characteristics of the non-Gaussian source signal, and can estimate a clearer source signal than the fast independent component analysis FastICA algorithm. The similarity coefficient between the separated signal and the source signal is above 0.9, and the audio audiometry is performed on the separated signal, and the JADE algorithm Separated signals are clearly discernible.

Description

Fault Tone Detection Method of Power Equipment Based on Joint Approximate Diagonalization Blind Source Separation Algorithm

technical field

The invention relates to a fault sound detection method of electric equipment based on a joint approximate diagonalization blind source separation algorithm, and belongs to the technical field of electric equipment maintenance.

Background technique

The failure of electrical equipment will not only cause damage to the equipment itself, but also cause serious damage to the safe, stable and economical operation of the entire power system. Therefore, it is of great significance to detect whether electrical equipment fails in time. Electrical equipment fault detection has gone through three stages: blackout experiment stage, live test stage and online detection stage. The traditional periodic detection method has the shortcomings of long test period, low test voltage, heavy workload and poor effectiveness, and it is difficult to meet the reliability requirements of the power system, especially with the rapid development of the power industry towards large units, large capacity and high voltage. development, these shortcomings become more apparent. Therefore, the online monitoring method based on status detection has attracted the attention of relevant management, scientific research, operation and engineering technicians. At the same time, real-time online monitoring and reducing the cost of the monitoring system are also inevitable trends in the development of power systems.

In the past, the fault sound diagnosis method of power equipment was generally realized by contact sensor detection. However, the complex environment such as high voltage and strong electromagnetic field in the transmission and transformation substation may have a certain impact on the sensor, thereby reducing the detection effect. In addition, the installation and maintenance of non-contact sensors are very inconvenient, and once the sensor fails, it may also bring unexpected consequences. Therefore, non-contact detection methods are an inevitable trend in future research.

Whether the working status of electrical equipment is normal or not can be reflected by the sound signal when it is working, and the sound signal can be collected through a non-contact microphone array, so the application of sound signal analysis to electrical equipment fault diagnosis can be done without affecting When electrical equipment is in normal operation, it can accurately reflect its operating status, and can be applied to electrical equipment such as transformers, circuit breakers, and transformers.

Chinese patent document CN104064186A discloses a method for detecting fault sounds of electrical equipment based on independent component analysis, including the steps of: using a microphone array to collect sound signals of electrical equipment running; using the independent component analysis method Fast-ICA algorithm based on the largest negentropy The sound signal collected by the microphone array is separated from each independent sound source signal; the Mel frequency cepstral coefficient MFCC of the independent sound source signal is extracted as the sound characteristic parameter, and the sound signal is identified by the pattern matching algorithm, and the sound template to be tested is matched with all reference sample templates Finally, the reference sample template with the smallest matching distance is the result of electrical equipment working sound recognition: if the reference sample template with the smallest matching distance is a normal sound, then the matching electrical equipment working sound is a normal sound; if the reference sample template with the smallest matching distance If the template is a fault sound, the matching electrical equipment working sound is a fault sound. However, this patent has the following defects: Gaussian noise interference often exists in the complex environment of power transmission and substations, and this patent cannot well deal with the Gaussian background noise that may remain in the source signal.

Contents of the invention

Aiming at the deficiencies of the prior art, the present invention discloses a fault sound detection method of power equipment based on a joint approximate diagonalization blind source separation algorithm;

Technical scheme of the present invention is:

A fault sound detection method for power equipment based on the joint approximate diagonal blind source separation algorithm, the specific steps include:

(1) Using a microphone array, that is, a MIC array to collect sound signals from the operation of electrical equipment;

(2) Adopting a blind source separation algorithm based on joint approximate diagonalization for step (1) adopting the sound signal collected by the microphone array to separate each independent sound source signal;

(3) Extract the Mel frequency cepstral coefficient MFCC of the independent sound source signal as the sound characteristic parameter, identify the sound signal through the pattern matching algorithm, match the sound template to be tested with all the reference sample templates, and match the reference sample template with the smallest distance It is the result of electrical equipment working sound recognition: if the reference sample template with the smallest matching distance is a normal sound, the matching electrical equipment working sound is a normal sound; if the reference sample template with the smallest matching distance is a fault sound, then The working sound of the matching electrical equipment is the fault sound.

Preferably according to the present invention, in step (1), a microphone array, i.e. a MIC array is used to collect sound signals of electrical equipment operation, specifically referring to:

Use a microphone array, that is, a MIC array to collect the sound signal of the electrical equipment running as: x(t)=[x ₁ (t), x ₂ (t),...,x _n (t)], n is a positive integer, where,

x ₁ (t)=a ₁₁ s ₁

x ₂ (t)＝a ₂₁ s ₁ +a ₂₂ s ₂

.

.   (ⅰ)

.

x _n (t)＝a _n1 s ₁ +a _n2 s ₂ +…+a _nm s _m

In formula (i), s ₁ , s ₂ ,…,s _m are sound signals from independent signal sources, a _ij (i=1,2,…,n; j=1,2,…,m) are real Coefficient, n=m.

Preferably according to the present invention, in step (2), adopt the blind source separation algorithm based on joint approximate diagonalization to separate each independent sound source signal from the sound signal collected by the microphone array in step (1), the specific steps include:

a. Centrally process the sound signals collected by the operation of electrical equipment using the microphone array, that is, the MIC array, and obtain the observation vector after de-averaging Calculated by formula (ii):

$\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; no</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>$

Sound signal obtained after centralized processing Perform whitening processing, whitening processing is to remove the mean value of the observation vector Perform linear transformation Q to obtain the processed observation signal z(t), and obtain it through formula (Ⅲ):

$\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; Q</span>}\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; iii</span>}<span\; class="notranslate">\; )</span>$

In formula (iii), the components in z(t) are not correlated with each other and have unit variance, and the whitening process adopts the PCA method of principal component analysis, which can be obtained by formula (iv):

$\mathrm{<span\; class="notranslate">\; Q</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}^{<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; f</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; iv</span>}<span\; class="notranslate">\; )</span>$

In formula (ⅳ), E is the covariance matrix Diagonal matrix composed of the n largest eigenvalues; F is the covariance matrix A matrix composed of n corresponding eigenvectors;

b. Calculate the fourth-order cumulant matrix of the observed signal z(t), and obtain the processed observed signal in step a: z(t)=[z ₁ (t),z ₂ (t),...,z _n (t) ], four observation signals are randomly selected: z _p , z _q , z _x , z _y (1≤p,q,x,y≤n), and the fourth-order cumulant is defined by formula (ⅴ):

cum(z _p ,z _q ,z _x ,z _y )=E[z _p z _q z _x z _y ]-E[z _p z _q ]E[z _x z _y ]-E[z _p z _x ]E [z _q z _y ]-E[z _p z _y ]E[z _q z _x ] (ⅴ)

Obtain all the fourth-order cumulants, get n ² fourth-order cumulants, set n ² fourth-order cumulants as m ₁ ,m ₂ ,…, m=[m ₁ ,m ₂ ,…, ], the pth, q element [C _z (A)] _pq of the fourth-order cumulant matrix is established by formula (ⅵ):

${<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; z</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; A</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}_{\mathrm{<span\; class="notranslate">\; pq</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; cum</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; xy</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; vi</span>}<span\; class="notranslate">\; )</span>$

In formula (ⅵ), a _xy is the x, yth element of matrix A, and A is an n×n matrix, the p, qth element of matrix A is 1, and the remaining elements of matrix A are all zero;

Calculate the fourth-order cumulant matrix for each m _{i ∈} m, and obtain n ² fourth-order cumulant matrices, let them be M ₁ , M ₂ ,…, And let M=[M ₁ ,M ₂ ,…, ], the cumulant matrix composed of M _o ∈ M as the weight matrix is decomposed into:

C _z (M _o )＝λM _o (ⅶ)

In formula (ⅶ), λ is the eigenvalue of C _z (M _o );

c. Carry out joint approximate diagonalization processing on the fourth-order cumulant matrix group M, determine the unitary matrix U, and obtain the estimation of the source signal, obtained by formula (vii), C _z (M _o ) is a symmetric matrix, and C _z ( M _o )=λM _o , the orthogonal separation matrix U makes the fourth-order cumulant matrix C _z (M _o ) diagonalized, as shown in formula (viii):

C _z (M _o ) = U ^T C (M _o ) U = Diag [k ₄ (s ₁ ): k ₄ (s ₂ ): ...: k ₄ (s _m )] (viii)

In formula (viii), Diag[k ₄ (s ₁ ):k ₄ (s ₂ ):…:k ₄ (s _m )] is the orthogonal separation matrix U so that the fourth-order cumulant matrix C _z (M _o ) pairs Cornerization calculation function, which belongs to existing functions;

Find the orthogonal separation matrix U, which simultaneously diagonalizes all fourth-order cumulant matrices C _z (M _o ), and the calculation process is shown in formula (ix):

$\mathrm{<span\; class="notranslate">\; min</span>}\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; Mo</span>}<span\; class="notranslate">\; \&Subset;</span>\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; off</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; u</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; ix</span>}<span\; class="notranslate">\; )</span>$

In formula (ix), the off-diagonal component off( ) is defined as The A represents a matrix, a _ij is each element of the matrix A, and minC (U) is the result of joint diagonalization of all fourth-order cumulant matrices C _z (M _o );

Consider the off-diagonal elements of U ^T C(M _o )U. If the off-diagonal elements of U ^T C(M _o )U are close to zero, it indicates that the degree of diagonalization is good.

An orthogonal separation matrix U is required to jointly diagonalize all fourth-order cumulant matrices C _z (M _o ) at the same time. In actual calculations, due to factors such as environmental noise and calculation errors, complete diagonalization cannot be achieved. Only Approximate diagonalization can be performed instead of complete diagonalization, so that each transformed C _z (M _o ) can be diagonalized as much as possible at the same time, so how to measure the degree or effect of diagonalization? A natural criterion is to consider the off-diagonal elements of U ^T C(M _o )U. If these elements are close to zero, it indicates a good degree of diagonalization.

The optimization of the algorithm is completed by using Givens rotation, and the unitary matrix U is obtained;

The source signal y(t) is estimated by formula (x):

y(t) = U ^T · Q · x(t) (x).

In step b, all the fourth-order cumulants are obtained according to the sequential structure, and n ⁴ fourth-order cumulants are obtained. According to the characteristics of formula (v), there are repetitions among the n ⁴ fourth-order cumulants, and finally n ² are obtained Fourth order cumulants.

Preferably according to the present invention, the concrete steps of described step (3) are:

d, performing pre-emphasis, framing and windowing operations on the source signal y(t) separated in step (2);

E, carry out FFT transformation to each frame sound signal after step d process, i.e. fast Fourier transform, obtain its frequency spectrum, take the square of modulus again as discrete power spectrum S (k);

f, calculate the power value obtained after S (k) passes through the band-pass filter bank, obtain V parameters Pv, v=0,1...V-1; then calculate the natural logarithm of Pv, obtain Lv, v= 0,1,...V-1; finally calculate the DCT discrete cosine transform of Lv to obtain Dv, v=0,1,...V-1; remove D ₀ and take D ₁ , D ₂ ,...,D _k as Parameters of MFCC;

G, described pattern matching algorithm is that dynamic time warping DTW algorithm carries out the concrete steps of voice recognition as:

Let the sound signal of step d be divided into p frame vectors, namely {T(1):T(2):...:T(n):...:T(p)}: T(n) is the speech feature of the nth frame Vector, 1≦n≦p, the reference sample has q frame vectors, that is, {R(1):R(2):...:R(m):...:R(q)}: R(m) is the mth frame 1≦m≦q, then the dynamic time warping DTW algorithm uses the time warping function j=w(i) to complete the mapping between the test vector and the reference template vector time axis, and this warping function w satisfies the following formula (xi ):

$\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; min</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; l</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; xi</span>}<span\; class="notranslate">\; )</span>$

In formula (ⅸ), d[T(i), R(w(i))] is the distance measure between the test vector T(i) and the reference template vector R(j); T(i) represents T The voice feature vector of frame i in R (w (i)) represents the voice feature vector of frame j in R; D is the minimum distance between the vector to be tested and the reference sample vector;

After using DTW to match the sound template to be tested with all reference sample templates, the reference sample template with the smallest matching distance is the result of power equipment working sound recognition.

The beneficial effects of the present invention are:

1. The present invention adopts the joint approximate diagonalization blind source separation algorithm, constructs the cost function with the unitary matrix U, and completes the optimization of the algorithm through the Givens rotation. The algorithm makes full use of the fourth-order statistics to automatically suppress the Gaussian background noise and enhance the non-Gaussian The characteristics of the source signal can estimate a clearer source signal than the fast independent component analysis FastICA algorithm. The similarity coefficient between the separated signal and the source signal is above 0.9. The audio audiometry of the separated signal shows that the signal separated by the JADE algorithm is clear and can be identify;

2. The present invention adopts the microphone array system to collect the audio signal in the power equipment working site. This method can significantly enhance the target sound source and remove the background noise, thereby obtaining a relatively pure source signal. By analyzing and processing the multi-channel sound signals picked up by the microphone array, the main lobe of the beam pattern formed by the array is aimed at the target sound source, and the "zero point" points to the interference source to suppress the interference signal, so as to obtain the target sound as much as possible.

detailed description

The present invention is further limited below in conjunction with embodiment, but is not limited thereto.

Example 1

A fault sound detection method for power equipment based on the joint approximate diagonal blind source separation algorithm, the specific steps include:

(1) Using a microphone array, that is, a MIC array to collect sound signals from the operation of electrical equipment;

(2) Adopting a blind source separation algorithm based on joint approximate diagonalization for step (1) adopting the sound signal collected by the microphone array to separate each independent sound source signal;

(3) Extract the Mel frequency cepstral coefficient MFCC of the independent sound source signal as the sound characteristic parameter, identify the sound signal through the pattern matching algorithm, match the sound template to be tested with all the reference sample templates, and match the reference sample template with the smallest distance It is the result of electrical equipment working sound recognition: if the reference sample template with the smallest matching distance is a normal sound, the matching electrical equipment working sound is a normal sound; if the reference sample template with the smallest matching distance is a fault sound, then The working sound of the matching electrical equipment is the fault sound.

Example 2

According to the power equipment failure sound detection method described in embodiment 1, the difference is that in step (1), a microphone array, i.e. a MIC array, is used to collect the sound signal of the operation of the electrical equipment, specifically referring to:

Use a microphone array, that is, a MIC array to collect the sound signal of the electrical equipment running as: x(t)=[x ₁ (t), x ₂ (t),...,x _n (t)], n is a positive integer, where,

x ₁ (t)=a ₁₁ s ₁

x ₂ (t)＝a ₂₁ s ₁ +a ₂₂ s ₂

.

.   (ⅰ)

.

x _n (t)＝a _n1 s ₁ +a _n2 s ₂ +…+a _nm s _m

In formula (i), s ₁ , s ₂ ,…,s _m are sound signals from independent signal sources, a _ij (i=1,2,…,n; j=1,2,…,m) are real Coefficient, n=m.

Example 3

According to the power equipment failure sound detection method described in Embodiment 1, the difference is that in step (2), the blind source separation algorithm based on joint approximate diagonalization is used to separate each independent sound from the sound signal collected by the microphone array in step (1). source signal, the specific steps include:

a. Centrally process the sound signals collected by the operation of electrical equipment using the microphone array, that is, the MIC array, and obtain the observation vector after de-averaging Calculated by formula (ii):

$\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; no</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>$

Sound signal obtained after centralized processing Perform whitening processing, whitening processing is to remove the mean value of the observation vector Perform linear transformation Q to obtain the processed observation signal z(t), and obtain it through formula (Ⅲ):

$\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; Q</span>}\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; iii</span>}<span\; class="notranslate">\; )</span>$

In formula (iii), the components in z(t) are not correlated with each other and have unit variance, and the whitening process adopts the PCA method of principal component analysis, which can be obtained by formula (iv):

$\mathrm{<span\; class="notranslate">\; Q</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}^{<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; f</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; iv</span>}<span\; class="notranslate">\; )</span>$

In formula (ⅳ), E is the covariance matrix Diagonal matrix composed of the n largest eigenvalues; F is the covariance matrix A matrix composed of n corresponding eigenvectors;

b. Calculate the fourth-order cumulant matrix of the observed signal z(t), and obtain the processed observed signal in step a: z(t)=[z ₁ (t),z ₂ (t),...,z _n (t) ], four observation signals are randomly selected: z _p , z _q , z _x , z _y (1≤p,q,x,y≤n), and the fourth-order cumulant is defined by formula (ⅴ):

cum(z _p ,z _q ,z _x ,z _y )=E[z _p z _q z _x z _y ]-E[z _p z _q ]E[z _x z _y ]-E[z _p z _x ]E [z _q z _y ]-E[z _p z _y ]E[z _q z _x ] (ⅴ)

Obtain all the fourth-order cumulants, get n ² fourth-order cumulants, set n ² fourth-order cumulants as m ₁ ,m ₂ ,…, m=[m ₁ ,m ₂ ,…, ], the p, q element [C _z (A)]pq of the fourth-order cumulant matrix established by formula (ⅵ) is:

${<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; z</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; A</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}_{\mathrm{<span\; class="notranslate">\; pq</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; cum</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; xy</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; vi</span>}<span\; class="notranslate">\; )</span>$

In formula (ⅵ), a _xy is the x, yth element of matrix A, and A is an n×n matrix, the p, qth element of matrix A is 1, and the remaining elements of matrix A are all zero;

Calculate the fourth-order cumulant matrix for each m _{i ∈} m, and obtain n ² fourth-order cumulant matrices, let them be M ₁ , M ₂ ,…, And let M=[M ₁ ,M ₂ ,…, ], the cumulant matrix composed of M _o ∈ M as the weight matrix is decomposed into:

C _z (M _o )＝λM _o (ⅶ)

In formula (ⅶ), λ is the eigenvalue of C _z (M _o );

c. Carry out joint approximate diagonalization processing on the fourth-order cumulant matrix group M, determine the unitary matrix U, and obtain the estimation of the source signal, obtained by formula (vii), C _z (M _o ) is a symmetric matrix, and C _z ( M _o )=λM _o , the orthogonal separation matrix U makes the fourth-order cumulant matrix C _z (M _o ) diagonalized, as shown in formula (viii):

C _z (M _o ) = U ^T c (M _o ) U = Diag [k ₄ (s ₁ ): k ₄ (s ₂ ): ...: k ₄ (s _m )] (viii)

In formula (viii), Diag[k ₄ (s ₁ ):k ₄ (s ₂ ):…:k ₄ (s _m )] is the orthogonal separation matrix U so that the fourth-order cumulant matrix C _z (M _o ) pairs Cornerization calculation function, which belongs to existing functions;

Find the orthogonal separation matrix U, which simultaneously diagonalizes all fourth-order cumulant matrices C _z (M _o ), and the calculation process is shown in formula (ix):

$\mathrm{<span\; class="notranslate">\; min</span>}\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; Mo</span>}<span\; class="notranslate">\; \&Subset;</span>\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; off</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; u</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; ix</span>}<span\; class="notranslate">\; )</span>$

In formula (ix), the off-diagonal component off( ) is defined as The A represents a matrix, a _ij is each element of the matrix A, and minc(u) is the result of joint diagonalization of all fourth-order cumulant matrices C _z (M _o );

Consider the off-diagonal elements of U ^T C(M _o )U. If the off-diagonal elements of U ^T C(M _o )U are close to zero, it indicates that the degree of diagonalization is good.

An orthogonal separation matrix U is required to jointly diagonalize all fourth-order cumulant matrices C _z (M _o ) at the same time. In actual calculations, due to factors such as environmental noise and calculation errors, complete diagonalization cannot be achieved. Only Approximate diagonalization can be performed instead of complete diagonalization, so that each transformed C _z (M _o ) can be diagonalized as much as possible at the same time, so how to measure the degree or effect of diagonalization? A natural criterion is to consider the off-diagonal elements of U ^T C(M _o )U. If these elements are close to zero, it indicates a good degree of diagonalization.

The optimization of the algorithm is completed by using Givens rotation, and the unitary matrix U is obtained;

The source signal y(t) is estimated by formula (x):

y(t) = U ^T · Q · x(t) (x).

In step b, all the fourth-order cumulants are obtained according to the sequential structure, and n ⁴ fourth-order cumulants are obtained. According to the characteristics of formula (v), there are repetitions among the n ⁴ fourth-order cumulants, and finally n ² are obtained Fourth order cumulants.

Example 4

According to the described power equipment fault sound detection method of embodiment 1, its difference is that described step (3) specific steps are:

d, performing pre-emphasis, framing and windowing operations on the source signal y(t) separated in step (2);

E, carry out FFT transformation to every frame sound signal after step d process, i.e. fast Fourier transform, obtain its frequency spectrum, take the square of modulus again as discrete power spectrum S (k);

f, calculate the power value obtained after S (k) passes through the band-pass filter bank, obtain V parameters Pv, v=0,1...V-1; then calculate the natural logarithm of Pv, obtain Lv, v= 0,1,...V-1; finally calculate the DCT discrete cosine transform of Lv to obtain Dv, v=0,1,...V-1; remove D ₀ and take D ₁ , D ₂ ,...,D _k as Parameters of MFCC;

G, described pattern matching algorithm is that dynamic time warping DTW algorithm carries out the concrete steps of voice recognition as:

Let the sound signal of step d be divided into p frame vectors, namely {T(1):T(2):...:T(n):...:T(p)}: T(n) is the speech feature of the nth frame Vector, 1≦n≦p, the reference sample has q frame vectors, that is, {R(1):R(2):...:R(m):...:R(q)}: R(m) is the mth frame 1≦m≦q, then the dynamic time warping DTW algorithm uses the time warping function j=w(i) to complete the mapping between the test vector and the reference template vector time axis, and this warping function w satisfies the following formula (xi ):

$\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; min</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; l</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; xi</span>}<span\; class="notranslate">\; )</span>$

In formula (ⅸ), d[T(i), R(w(i))] is the distance measure between the test vector T(i) and the reference template vector R(j); T(i) represents T The voice feature vector of frame i in R (w (i)) represents the voice feature vector of frame j in R; D is the minimum distance between the vector to be tested and the reference sample vector;

After using DTW to match the sound template to be tested with all reference sample templates, the reference sample template with the smallest matching distance is the result of power equipment working sound recognition.

Claims (4)

1. based on the power equipment fault sound detection method of joint approximate diagonalization blind source separation algorithm, it is characterized in that, concrete steps comprise: (1) Using a microphone array, that is, a MIC array to collect sound signals of power equipment operation; (2) Adopting a blind source separation algorithm based on joint approximate diagonalization for step (1) adopting the sound signal collected by the microphone array to separate each independent sound source signal; (3) Extract the Mel frequency cepstrum coefficient MFCC of the independent sound source signal as the sound characteristic parameter, identify the sound signal through the pattern matching algorithm, match the sound template to be tested with all the reference sample templates, and match the reference sample template with the smallest distance It is the result of the recognition of the working sound of the power equipment: if the reference sample template with the smallest matching distance is a normal sound, then the matching working sound of the power equipment is a normal sound; if the reference sample template with the smallest matching distance is a fault sound, then The matching working sound of electric equipment is fault sound. 2. according to the described power equipment failure sound detection method of claim 1, it is characterized in that, in step (1), adopt microphone array, i.e. the sound signal that MIC array gathers power equipment operation, specifically refers to: Using a microphone array, that is, a MIC array to collect sound signals of power equipment operation is recorded as: x(t)=[x ₁ (t), x ₂ (t), ......, x _n (t)], n is a positive integer, where, x ₁ (t)=a ₁₁ s ₁ x ₂ (t)＝a ₂₁ s ₁ +a ₂₂ s ₂ · · · x _n (t)＝a _n1 s ₁ +a _n2 s ₂ +…+a _nm s _m In formula (i), s ₁ , s ₂ ,..., s _m are sound signals from independent signal sources, and a _ij (i=1, 2,..., n; j=1, 2,..., m) is a real Coefficient, n=m. 3. according to the described power equipment failure tone detection method of claim 1, it is characterized in that, in step (2), adopt the sound signal separation that adopts microphone array to collect based on joint approximate diagonalization blind source separation algorithm for step (1) Independent sound source signal, the specific steps include: a. Centrally process the sound signal collected by the microphone array, that is, the MIC array to collect the operation of the power equipment, and obtain the observation vector after de-averaging Obtained by formula (ii): $\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; no</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>$ Sound signal obtained after centralized processing Perform whitening processing, whitening processing is to remove the mean value of the observation vector Perform linear transformation Q to obtain the processed observation signal z(t), and obtain it through formula (iii): $\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; Q</span>}\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; iii</span>}<span\; class="notranslate">\; )</span>$ In formula (iii), the components in z(t) are not correlated with each other and have unit variance. The whitening process adopts the PCA method of principal component analysis, and is obtained by formula (iv): $\mathrm{<span\; class="notranslate">\; Q</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}^{<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; f</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; iv</span>}<span\; class="notranslate">\; )</span>$ In formula (iv), E is the covariance matrix Diagonal matrix composed of the n largest eigenvalues; F is the covariance matrix A matrix composed of n corresponding eigenvectors; b. Calculate the fourth-order cumulant matrix of the observed signal z(t), step a obtains the processed observed signal: z(t)=[z ₁ (t), z ₂ (t), ..., z _n (t) ], any four observation signals: z _p , z _q , z _x , z _y (1≤p, q, x, y≤n), define the fourth-order cumulant by formula (v): cum(z _p ，z _q ，z _x ，z _y )＝E[z _p z _q z _x z _y ]-E[z _p z _q ]E[z _x z _y ]-E[z _p z _x ]E [z _q z _y ]-E[z _p z _y ]E[z _q z _x ] (v) Get all the fourth-order cumulants, get n ² fourth-order cumulants, set n ² fourth-order cumulants as The pth, q element [C _z (A)] _pq of the fourth-order cumulant matrix is established by formula (vi): ${<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; z</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; A</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}_{\mathrm{<span\; class="notranslate">\; pq</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; cum</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; z</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; xy</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; vi</span>}<span\; class="notranslate">\; )</span>$ In formula (vi), _axy is the x, y elements of matrix A, and A is an n×n matrix, the p and q elements of matrix A are 1, and the remaining elements of matrix A are zero; Find the fourth-order cumulant matrix for each m _{i ∈} m, and get n ² fourth-order cumulant matrices, let them be and order By formula (vii), the cumulant matrix composed of M _o ∈ M as the weight matrix is decomposed into: C _z (M _o ) = λM _o (vii) In formula (vii), λ is the eigenvalue of C _z (M _o ); c. Carry out joint approximate diagonalization processing on the fourth-order cumulant matrix group M, determine the unitary matrix U, and obtain the estimation of the source signal, obtained by formula (vii), C _z (M _o ) is a symmetric matrix, and C _z ( M _o )=λM _o , the orthogonal separation matrix U makes the fourth-order cumulant matrix C _z (M _o ) diagonalized, as shown in formula (viii): C _z (M _o ) = U ^T C (M _o ) U = Diag [k ₄ (s ₁ ), k ₄ (s ₂ ), ..., k ₄ (s _m )] (viii) In formula (viii), Diag[k ₄ (s ₁ ), k ₄ (s ₂ ),…, k ₄ (s _m )] is the orthogonal separation matrix U so that the fourth-order cumulant matrix C _z (M _o ) pairs Cornerization calculation function, which belongs to existing functions; Find the orthogonal separation matrix U, which simultaneously diagonalizes all fourth-order cumulant matrices C _z (M _o ), and the calculation process is shown in formula (ix): $\mathrm{<span\; class="notranslate">\; min</span>}\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; Mo</span>}<span\; class="notranslate">\; \&Subset;</span>\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; off</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; u</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; ix</span>}<span\; class="notranslate">\; )</span>$ In formula (ix), the off-diagonal component off( ) is defined as The A represents a matrix, a _ij is each element of the matrix A, and minC (U) is the result of joint diagonalization of all fourth-order cumulant matrices C _z (M _o ); The optimization of the algorithm is completed by using Givens rotation, and the unitary matrix U is obtained; The source signal y(t) is estimated by formula (x): y(t) = U ^T · Q · x(t) (x). 4. according to claim 1 described power equipment failure sound detection method, it is characterized in that, described step (3) concrete steps are: d, performing pre-emphasis, framing and windowing operations on the source signal y(t) separated in step (2); E, carry out FFT transformation to every frame sound signal after step d process, i.e. fast Fourier transform, obtain its frequency spectrum, take the square of modulus again as discrete power spectrum S (k); f, calculate S (k) by the power value gained after the band-pass filter bank, obtain V parameters Pv, v=0,1, ... V-1; Then calculate the natural logarithm of Pv, obtain Lv, v= 0, 1, ... V-1; finally calculate the DCT discrete cosine transform of Lv, obtain Dv, v=0, 1, ... V-1; remove D ₀ , take D ₁ , D ₂ , ..., D _k as Parameters of MFCC; G, described pattern matching algorithm is that dynamic time warping DTW algorithm carries out the concrete steps of voice recognition as: Let the sound signal of step d be divided into p frame vectors, namely {T(1), T(2),..., T(n),..., T(p)}, T(n) is the speech feature of the nth frame Vector, 1≤n≤p, the reference sample has q frame vectors, namely {R(1), R(2), ..., R(m), ..., R(q)}, R(m) is the mth frame , 1≤m≤q, then the dynamic time warping DTW algorithm uses the time warping function j=w(i) to complete the mapping between the vector to be tested and the time axis of the reference template vector, and this warping function w satisfies the following formula (xi ): $\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; min</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; l</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; w</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; xi</span>}<span\; class="notranslate">\; )</span>$ In formula (ix), d[T(i), R(w(i))] is the distance measure between the test vector T(i) and the reference template vector R(j); T(i) represents T The speech feature vector of frame i in R; R(w(i)) represents the speech feature vector of frame j in R; D is the minimum distance between the vector to be tested and the reference sample vector; After using DTW to match the sound template to be tested with all reference sample templates, the reference sample template with the smallest matching distance is the result of power equipment working sound recognition.