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

Abstract

The invention discloses a voice detection method of power equipment failure based on a combined similar diagonalizable blind source separation algorithm. The method comprises the specific steps as follows: (1) adopting a microphone array; (2) separating all independent sound source signals from sound signals collected by the microphone array by adopting the combined similar diagonalizable blind source separation algorithm; (3) extracting Mel-MFC (Frequency Cepstral Coefficients) of the independent sound source signals as sound characteristic parameters, and identifying the sound signals through a model matching algorithm, wherein a reference sample template with a minimal matching distance is a result of identifying the operating sound of the power equipment after a sound template to be tested and all reference sample templates are matched. According to the voice detection method provided by the invention, the characteristics of a non-Gaussian source signal can be enhanced, a source signal which is more clear than a Fast ICA (Independent Component Analysis) can be estimated, the similarity coefficients of the separated signal and the source signals are 0.9 or above, the voice frequency audiometry on the separated signals can be carried out, and the signals separated by a JADE algorithm is clear and distinguishable.

Description

Based on the electrical equipment fault sound detection method of joint approximate diagonalization blind source separation algorithm

Technical field

The present invention relates to the electrical equipment fault sound detection method based on joint approximate diagonalization blind source separation algorithm, belong to the technical field of maintenance of electric device.

Background technology

Electrical equipment breaks down not only can cause damage to equipment itself, and also can produce serious destruction to the safe, stable of whole electric system and economical operation, therefore, detects whether electrical equipment breaks down in time and is of great significance.Electrical equipment malfunction detects and experienced by three phases: have a power failure experimental phase, live testing stage and on-line checkingi stage.Traditional periodic sensing approach exists that the test period is long, trial voltage is low, workload is large and the shortcoming such as validity is poor, be difficult to meet the requirement of electric system to reliability, particularly along with power industry is towards large-sized unit, Large Copacity and high-tensionly to develop rapidly, these shortcomings become particularly evident.Therefore, take state-detection as the attention that the on-line monitoring of benchmark has caused about management, scientific research, operation and engineering technical personnel, real time on-line monitoring and the cost that reduces monitoring system are also the inexorable trends of power system development simultaneously.

Electrical equipment fault sound diagnostic method in the past generally will be detected by touch sensor and realize, but the complex environments such as the high voltage in power regulation station and strong-electromagnetic field may produce certain impact to sensor, thus reduces Detection results.In addition, the installation and maintenance of non-contacting sensor is very inconvenient, once sensor generation problem, also may bring beyond thought consequence.Therefore, non-contact detection method is the inexorable trend of future studies.

The whether normal of electrical equipment duty can be reflected by voice signal during its work, and voice signal can be collected by contactless microphone array, so by voice signal analytical applications in fault diagnosis, when normally can run when not affecting electrical equipment, its running status of accurate reflection, can be applicable to the electrical equipments such as transformer, isolating switch and mutual inductor.

Chinese patent literature CN104064186A discloses a kind of electrical equipment malfunction sound detection method based on independent component analysis, comprises step: adopt microphone array to gather the voice signal of electric equipment operation; The Independent component analysis Fast-ICA algorithm based on negentropy is maximum is adopted to be separated each individual sources signal for adopting the voice signal of microphone array collection; Extract the Mel frequency cepstral coefficient MFCC of individual sources signal as sound characteristic parameter, by pattern matching algorithm sound recognition signal, after sound pattern to be tested and all reference sample template matches, the minimum reference sample template of matching distance is the result of electrical equipment work sound identification: reference sample template as minimum in matching distance is normal sound, then the electrical equipment work sound matched is normal sound; Reference sample template as minimum in matching distance is trouble back tone, then the electrical equipment work sound matched is trouble back tone.But this patent exists following defect: the interference that often there is Gaussian noise in the complex environment of power regulation station, this patent well can not process the Gaussian Background noise that may remain in source signal.

Summary of the invention

For the deficiencies in the prior art, the invention discloses the electrical equipment fault sound detection method based on joint approximate diagonalization blind source separation algorithm;

Technical scheme of the present invention is:

Based on the electrical equipment fault sound detection method of joint approximate diagonalization blind source separation algorithm, concrete steps comprise:

(1) microphone array, the i.e. voice signal of MIC array acquisition electric equipment operation is adopted;

(2) employing adopts the voice signal of microphone array collection to be separated each individual sources signal based on joint approximate diagonalization blind source separation algorithm for step (1);

(3) the Mel frequency cepstral coefficient MFCC of individual sources signal is extracted as sound characteristic parameter, by pattern matching algorithm sound recognition signal, after sound pattern to be tested is mated with all reference sample templates, the minimum reference sample template of matching distance is exactly the result of electrical equipment work sound identification: if the minimum reference sample template of matching distance is normal sound, then the electrical equipment work sound matched with it is normal sound; If the minimum reference sample template of matching distance is trouble back tone, then the electrical equipment work sound matched with it is trouble back tone.

Preferred according to the present invention, in step (1), adopt microphone array, the i.e. voice signal of MIC array acquisition electric equipment operation, specifically refer to:

Adopt microphone array, namely the voice signal of MIC array acquisition electric equipment operation is designated as: x (t)=[x ₁(t), x ₂(t) ...., x _n(t)], n is positive integer, wherein,

x ₁(t)＝a ₁₁s ₁

x ₂(t)＝a ₂₁s ₁+a ₂₂s ₂

.

. (ⅰ)

.

x _n(t)＝a _n1s ₁+a _n2s ₂+…+a _nms _m

In formula (I), s ₁, s ₂..., s _mfor the voice signal that independent signal source sends, a _ij(i=1,2 ..., n; J=1,2 ..., m) be real coefficient, n=m.

Preferred according to the present invention, in step (2), adopt and adopt the voice signal of microphone array collection to be separated each individual sources signal based on joint approximate diagonalization blind source separation algorithm for step (1), concrete steps comprise:

A, to employing microphone array, namely the voice signal of MIC array acquisition electric equipment operation carries out centralization process, the measurement vector gone after average obtained through type (II) is tried to achieve:

$\stackrel{\‾}{x}\left(t\right)=x\left(t\right)-\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\left(t\right)---\left(\mathrm{ii}\right)$

To the voice signal obtained after centralization process carry out whitening processing, whitening processing to remove the measurement vector after average carry out linear transformation Q and obtain observation signal z (t) after processing, through type (III) is tried to achieve:

$z\left(t\right)=Q\stackrel{\‾}{x}\left(t\right)---\left(\mathrm{iii}\right)$

In formula (III), in z (t), each component is uncorrelated mutually, and has unit variance, and what whitening processing adopted is principal component analysis PCA method, and through type (IV) is tried to achieve:

$Q=E^{-\frac{1}{2}}{F}^T---\left(\mathrm{iv}\right)$

In formula (IV), E is covariance matrix the diagonal matrix of n eigenvalue of maximum composition; F is covariance matrix the matrix of the corresponding eigenvector composition of n;

The fourth order cumulant matrix of b, calculating observation signal z (t), step a obtains the observation signal after processing: z (t)=[z ₁(t), z ₂(t) ..., z _n(t)], appoint and get wherein four observation signal: z _p, z _q, z _x, z _y(1≤p, q, x, y≤n), through type (V) defines fourth order cumulant:

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

Try to achieve all fourth order cumulants, obtain n ²individual fourth order cumulant, if n ²individual fourth order cumulant is m ₁, m ₂..., m=[m ₁, m ₂..., ], through type (VI) sets up the p of fourth order cumulant matrix, q element [C _z(A)] _pqfor:

${\left[C_z\left(A\right)\right]}_{\mathrm{pq}}=\underset{p,q=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{cum}(z_p,z_q,z_x,z_y)a_{\mathrm{xy}}---\left(\mathrm{vi}\right)$

In formula (VI), a _xyfor the xth of matrix A, y element, and A is the p of n × n battle array, matrix A, a q element is 1, and all the other elements of matrix A are zero;

To each m _i∈ m asks fourth order cumulant matrix, obtains n ²individual fourth order cumulant matrix, if it is M ₁, M ₂..., and make M=[M ₁, M ₂..., ], through type (VII) will with M _o∈ M is that the cumulant matrices that weight matrix is formed is decomposed into:

C _z(M _o)＝λM _o(ⅶ)

In formula (VII), λ is C _z(M _o) eigenwert;

C, joint approximate diagonalization process is carried out to fourth order cumulant matrix group M, determine unitary matrix U, obtain the estimation of source signal, obtained by formula (vii), C _z(M _o) be symmetrical matrix, and C _z(M _o)=λ M _o, orthogonal separation matrix U makes fourth order cumulant matrix C _z(M _o) diagonalization, shown in (viii):

C _z(M _o)＝U ^TC(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)] make fourth order cumulant matrix C for orthogonal separation matrix U _z(M _o) diagonalization computing function, belong to existing function;

Ask orthogonal separation matrix U, orthogonal separation matrix U is simultaneously to all fourth order cumulant matrix C _z(M _o) carry out Joint diagonalization, computation process is such as formula shown in (ix):

$\min C\left(U\right)=\underset{\mathrm{Mo}\⋐M}{\mathrm{\Σ}}\mathrm{off}\left[U^{T}C\left(M\right)U\right]---\left(\mathrm{ix}\right)$

In formula (ix), non-diagonal component off () is defined as described A represents a matrix, a _ijbe each element of matrix A, minC (U) is to all fourth order cumulant matrix C _z(M _o) carry out the result of calculation of Joint diagonalization;

Consider U ^tc (M _o) off-diagonal element of U, if U ^tc (M _o) off-diagonal element of U close to zero, then show that diagonalization degree is fine.

Require that an orthogonal separation matrix U is simultaneously to all fourth order cumulant matrix C _z(M _o) carry out Joint diagonalization, in actual computation, due to the factor such as neighbourhood noise and the error of calculation, cannot Complete Diagonalization be realized, joint approximate diagonalization can only be carried out to replace Complete Diagonalization, make each C after conversion _z(M _o) diagonalization as far as possible simultaneously, so how measure diagonalization degree or effect? a very natural criterion considers U exactly ^tc (M _o) off-diagonal element of U, if these elements are close to zero, then show that diagonalization degree is fine.

Employing Givens has rotated the optimization to algorithm, obtains unitary matrix U;

Source signal y (t) through type (x) is estimated to obtain:

y(t)＝U ^T·Q·x(t) (x)。

In step b, structure tries to achieve all fourth order cumulants in order, obtains n ⁴individual fourth order cumulant, according to the characteristic of formula (v), n ⁴there is repetition in individual fourth order cumulant, finally obtain n ²individual fourth order cumulant.

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

D, pre-emphasis, framing and windowing operation are carried out to isolated source signal y (t) in step (2);

E, FFT conversion is carried out, i.e. Fast Fourier Transform (FFT) to the every frame voice signal after steps d process, obtains its frequency spectrum, then delivery square as discrete power spectrum S (k);

F, calculating S (k), by the performance number of gained after band-pass filter group, obtain V parameter 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, get D ₁, D ₂..., D _kas the parameter of MFCC;

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

If the voice signal of steps d has divided p frame vector, namely T (1): T (2): ...: T (n): ...: T (p) }: T (n) is the speech characteristic vector of the n-th frame, 1≤n≤p, reference sample has q frame vector, namely R (1): R(2): ...: R (m): ...: R (q) }: R (m) is the speech characteristic vector of m frame, 1≤m≤q, then dynamic time warping DTW algorithm utilizes Time alignment function j=w (i) to complete the mapping of vector to be tested and reference template vector time axle, and this warping function w meets following formula (xi):

$D=\underset{w\left(i\right)}{\min }\underset{i=1}{\overset{l}{\mathrm{\Σ}}}d[T\left(i\right),R\left(w\left(i\right)\right)]---\left(\mathrm{xi}\right)$

In formula (IV), d [T (i), R (w (i))] is the distance measure between vector T (i) to be tested and reference template vector R (j); T (i) represents the speech characteristic vector of the i-th frame in T; R (w (i)) represents jth frame speech characteristic vector in R; The minor increment of D then between vector to be tested and reference sample vector;

After utilizing DTW to be mated with all reference sample templates by sound pattern to be tested, the minimum reference sample template of matching distance is exactly the result of power equipment work sound identification.

Beneficial effect of the present invention is:

1, the present invention adopts joint approximate diagonalization blind source separation algorithm, cost function is built with unitary matrix U, the optimization to algorithm has been rotated by Givens, this algorithm takes full advantage of Fourth amount and automatically suppresses Gaussian Background noise, strengthen the feature of non-gaussian source signal, can estimate than Fast Independent Component Analysis FastICA algorithm source signal more clearly, the signal be separated and the similarity coefficient of source signal are all more than 0.9, carry out audio frequency audiometry to separation signal, the signal that JADE algorithm is separated is clear and legible;

2, the present invention adopts microphone array system to gather the sound signal in power equipment working site, and the method significantly can play the effect strengthening target sound source, remove ground unrest, thus obtains purer source signal.Microphone array is by carrying out treatment and analysis to the multi-channel sound signal of pickup, and the beam pattern main lobe that array is formed aims at the mark sound source, and " zero point " points to interference source to suppress undesired signal, thus obtains target sound as much as possible.

Embodiment

Below in conjunction with embodiment, the present invention is further qualified, but is not limited thereto.

Embodiment 1

Based on the electrical equipment fault sound detection method of joint approximate diagonalization blind source separation algorithm, concrete steps comprise:

(1) microphone array, the i.e. voice signal of MIC array acquisition electric equipment operation is adopted;

(2) employing adopts the voice signal of microphone array collection to be separated each individual sources signal based on joint approximate diagonalization blind source separation algorithm for step (1);

(3) the Mel frequency cepstral coefficient MFCC of individual sources signal is extracted as sound characteristic parameter, by pattern matching algorithm sound recognition signal, after sound pattern to be tested is mated with all reference sample templates, the minimum reference sample template of matching distance is exactly the result of electrical equipment work sound identification: if the minimum reference sample template of matching distance is normal sound, then the electrical equipment work sound matched with it is normal sound; If the minimum reference sample template of matching distance is trouble back tone, then the electrical equipment work sound matched with it is trouble back tone.

Embodiment 2

Electrical equipment fault sound detection method according to embodiment 1, its difference is, in step (1), adopts microphone array, the i.e. voice signal of MIC array acquisition electric equipment operation, specifically refers to:

Adopt microphone array, namely the voice signal of MIC array acquisition electric equipment operation is designated as: x (t)=[x ₁(t), x ₂(t) ...., x _n(t)], n is positive integer, wherein,

x ₁(t)＝a ₁₁s ₁

x ₂(t)＝a ₂₁s ₁+a ₂₂s ₂

.

. (ⅰ)

.

x _n(t)＝a _n1s ₁+a _n2s ₂+…+a _nms _m

In formula (I), s ₁, s ₂..., s _mfor the voice signal that independent signal source sends, a _ij(i=1,2 ..., n; J=1,2 ..., m) be real coefficient, n=m.

Embodiment 3

Electrical equipment fault sound detection method according to embodiment 1, its difference is, in step (2), adopt and adopt the voice signal of microphone array collection to be separated each individual sources signal based on joint approximate diagonalization blind source separation algorithm for step (1), concrete steps comprise:

A, to employing microphone array, namely the voice signal of MIC array acquisition electric equipment operation carries out centralization process, the measurement vector gone after average obtained through type (II) is tried to achieve:

$\stackrel{\‾}{x}\left(t\right)=x\left(t\right)-\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\left(t\right)---\left(\mathrm{ii}\right)$

To the voice signal obtained after centralization process carry out whitening processing, whitening processing to remove the measurement vector after average carry out linear transformation Q and obtain observation signal z (t) after processing, through type (III) is tried to achieve:

$z\left(t\right)=Q\stackrel{\‾}{x}\left(t\right)---\left(\mathrm{iii}\right)$

In formula (III), in z (t), each component is uncorrelated mutually, and has unit variance, and what whitening processing adopted is principal component analysis PCA method, and through type (IV) is tried to achieve:

$Q=E^{-\frac{1}{2}}{F}^T---\left(\mathrm{iv}\right)$

In formula (IV), E is covariance matrix the diagonal matrix of n eigenvalue of maximum composition; F is covariance matrix the matrix of the corresponding eigenvector composition of n;

The fourth order cumulant matrix of b, calculating observation signal z (t), step a obtains the observation signal after processing: z (t)=[z ₁(t), z ₂(t) ..., z _n(t)], appoint and get wherein four observation signal: z _p, z _q, z _x, z _y(1≤p, q, x, y≤n), through type (V) defines fourth order cumulant:

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

Try to achieve all fourth order cumulants, obtain n ²individual fourth order cumulant, if n ²individual fourth order cumulant is m ₁, m ₂..., m=[m ₁, m ₂..., ], through type (VI) sets up the p of fourth order cumulant matrix, q element [C _z(A)] pq is:

${\left[C_z\left(A\right)\right]}_{\mathrm{pq}}=\underset{p,q=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{cum}(z_p,z_q,z_x,z_y)a_{\mathrm{xy}}---\left(\mathrm{vi}\right)$

In formula (VI), a _xyfor the xth of matrix A, y element, and A is the p of n × n battle array, matrix A, a q element is 1, and all the other elements of matrix A are zero;

To each m _i∈ m asks fourth order cumulant matrix, obtains n ²individual fourth order cumulant matrix, if it is M ₁, M ₂..., and make M=[M ₁, M ₂..., ], through type (VII) will with M _o∈ M is that the cumulant matrices that weight matrix is formed is decomposed into:

C _z(M _o)＝λM _o(ⅶ)

In formula (VII), λ is C _z(M _o) eigenwert;

C, joint approximate diagonalization process is carried out to fourth order cumulant matrix group M, determine unitary matrix U, obtain the estimation of source signal, obtained by formula (vii), C _z(M _o) be symmetrical matrix, and C _z(M _o)=λ M _o, orthogonal separation matrix U makes fourth order cumulant matrix C _z(M _o) diagonalization, shown in (viii):

C _z(M _o)＝U ^Tc(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)] make fourth order cumulant matrix C for orthogonal separation matrix U _z(M _o) diagonalization computing function, belong to existing function;

Ask orthogonal separation matrix U, orthogonal separation matrix U is simultaneously to all fourth order cumulant matrix C _z(M _o) carry out Joint diagonalization, computation process is such as formula shown in (ix):

$\min C\left(U\right)=\underset{\mathrm{Mo}\⋐M}{\mathrm{\Σ}}\mathrm{off}\left[U^{T}C\left(M\right)U\right]---\left(\mathrm{ix}\right)$

In formula (ix), non-diagonal component off () is defined as described A represents a matrix, a _ijbe each element of matrix A, minc (u) is to all fourth order cumulant matrix C _z(M _o) carry out the result of calculation of Joint diagonalization;

Consider U ^tc (M _o) off-diagonal element of U, if U ^tc (M _o) off-diagonal element of U close to zero, then show that diagonalization degree is fine.

Require that an orthogonal separation matrix U is simultaneously to all fourth order cumulant matrix C _z(M _o) carry out Joint diagonalization, in actual computation, due to the factor such as neighbourhood noise and the error of calculation, cannot Complete Diagonalization be realized, joint approximate diagonalization can only be carried out to replace Complete Diagonalization, make each C after conversion _z(M _o) diagonalization as far as possible simultaneously, so how measure diagonalization degree or effect? a very natural criterion considers U exactly ^tc (M _o) off-diagonal element of U, if these elements are close to zero, then show that diagonalization degree is fine.

Employing Givens has rotated the optimization to algorithm, obtains unitary matrix U;

Source signal y (t) through type (x) is estimated to obtain:

y(t)＝U ^T·Q·x(t) (x)。

In step b, structure tries to achieve all fourth order cumulants in order, obtains n ⁴individual fourth order cumulant, according to the characteristic of formula (v), n ⁴there is repetition in individual fourth order cumulant, finally obtain n ²individual fourth order cumulant.

Embodiment 4

Electrical equipment fault sound detection method according to embodiment 1, its difference is, described step (3) concrete steps are:

D, pre-emphasis, framing and windowing operation are carried out to isolated source signal y (t) in step (2);

E, FFT conversion is carried out, i.e. Fast Fourier Transform (FFT) to the every frame voice signal after steps d process, obtains its frequency spectrum, then delivery square as discrete power spectrum S (k);

F, calculating S (k), by the performance number of gained after band-pass filter group, obtain V parameter 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, get D ₁, D ₂..., D _kas the parameter of MFCC;

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

If the voice signal of steps d has divided p frame vector, namely T (1): T (2): ...: T (n): ...: T (p) }: T (n) is the speech characteristic vector of the n-th frame, 1≤n≤p, reference sample has q frame vector, namely R (1): R (2): ...: R (m): ...: R (q) }: R (m) is the speech characteristic vector of m frame, 1≤m≤q, then dynamic time warping DTW algorithm utilizes Time alignment function j=w (i) to complete the mapping of vector to be tested and reference template vector time axle, and this warping function w meets following formula (xi):

$D=\underset{w\left(i\right)}{\min }\underset{i=1}{\overset{l}{\mathrm{\Σ}}}d[T\left(i\right),R\left(w\left(i\right)\right)]---\left(\mathrm{xi}\right)$

In formula (IV), d [T (i), R (w (i))] is the distance measure between vector T (i) to be tested and reference template vector R (j); T (i) represents the speech characteristic vector of the i-th frame in T; R (w (i)) represents jth frame speech characteristic vector in R; The minor increment of D then between vector to be tested and reference sample vector;

After utilizing DTW to be mated with all reference sample templates by sound pattern to be tested, the minimum reference sample template of matching distance is exactly the result of power equipment work sound identification.

Claims (4)

1., based on the electrical equipment fault sound detection method of joint approximate diagonalization blind source separation algorithm, it is characterized in that, concrete steps comprise:

(1) microphone array is adopted, i.e. the voice signal of MIC array acquisition power equipment operation;

(2) employing adopts the voice signal of microphone array collection to be separated each individual sources signal based on joint approximate diagonalization blind source separation algorithm for step (1);

(3) the Mel frequency cepstral coefficient MFCC of individual sources signal is extracted as sound characteristic parameter, by pattern matching algorithm sound recognition signal, after sound pattern to be tested is mated with all reference sample templates, the minimum reference sample template of matching distance is exactly the result of power equipment work sound identification: if the minimum reference sample template of matching distance is normal sound, then the power equipment work sound matched with it is normal sound; If the minimum reference sample template of matching distance is trouble back tone, then the power equipment work sound matched with it is trouble back tone.

2. electrical equipment fault sound detection method according to claim 1, is characterized in that, in step (1), adopts microphone array, i.e. the voice signal that runs of MIC array acquisition power equipment, specifically refers to:

Adopt microphone array, the voice signal that namely MIC array acquisition power equipment runs is designated as: x (t)=[x ₁(t), x ₂(t) ...., x _n(t)], n is positive integer, wherein,

x ₁(t)＝a ₁₁s ₁

x ₂(t)＝a ₂₁s ₁+a ₂₂s ₂

·

·

·

x _n(t)＝a _n1s ₁+a _n2s ₂+…+a _nms _m

In formula (i), s ₁, s ₂..., s _mfor the voice signal that independent signal source sends, a _ij(i=1,2 ..., n; J=1,2 ..., m) be real coefficient, n=m.

3. electrical equipment fault sound detection method according to claim 1, it is characterized in that, in step (2), adopt and adopt the voice signal of microphone array collection to be separated each individual sources signal based on joint approximate diagonalization blind source separation algorithm for step (1), concrete steps comprise:

A, to employing microphone array, the voice signal that namely MIC array acquisition power equipment runs carries out centralization process, the measurement vector gone after average obtained through type (ii) is tried to achieve:

$\stackrel{\‾}{x}\left(t\right)=x\left(t\right)-\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\left(t\right)---\left(\mathrm{ii}\right)$

To the voice signal obtained after centralization process carry out whitening processing, whitening processing to remove the measurement vector after average carry out linear transformation Q and obtain observation signal z (t) after processing, through type (iii) is tried to achieve:

$z\left(t\right)=Q\stackrel{\‾}{x}\left(t\right)---\left(\mathrm{iii}\right)$

In formula (iii), in z (t), each component is uncorrelated mutually, and has unit variance, and what whitening processing adopted is principal component analysis PCA method, and through type (iv) is tried to achieve:

$Q=E^{-\frac{1}{2}}{F}^T---\left(\mathrm{iv}\right)$

In formula (iv), E is covariance matrix the diagonal matrix of n eigenvalue of maximum composition; F is covariance matrix the matrix of the corresponding eigenvector composition of n;

The fourth order cumulant matrix of b, calculating observation signal z (t), step a obtains the observation signal after processing: z (t)=[z ₁(t), z ₂(t) ..., z _n(t)], appoint and get wherein four observation signal: z _p, z _q, z _x, z _y(1≤p, q, x, y≤n), through type (v) defines fourth order cumulant:

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

Try to achieve all fourth order cumulants, obtain n ²individual fourth order cumulant, if n ²individual fourth order cumulant is through type (vi) sets up the p of fourth order cumulant matrix, q element [C _z(A)] _pqfor:

${\left[C_z\left(A\right)\right]}_{\mathrm{pq}}=\underset{p,q=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{cum}(z_p,z_q,z_x,z_y)a_{\mathrm{xy}}---\left(\mathrm{vi}\right)$

In formula (vi), a _xyfor the xth of matrix A, y element, and A is the p of n × n battle array, matrix A, a q element is 1, and all the other elements of matrix A are zero;

To each m _i∈ m asks fourth order cumulant matrix, obtains n ²individual fourth order cumulant matrix, if it is and make through type (vii) will with M _o∈ M is that the cumulant matrices that weight matrix is formed is decomposed into:

C _z(M _o)＝λM _o(vii)

In formula (vii), λ is C _z(M _o) eigenwert;

C, joint approximate diagonalization process is carried out to fourth order cumulant matrix group M, determine unitary matrix U, obtain the estimation of source signal, obtained by formula (vii), C _z(M _o) be symmetrical matrix, and C _z(M _o)=λ M _o, orthogonal separation matrix U makes fourth order cumulant matrix C _z(M _o) diagonalization, shown in (viii):

C _z(M _o)＝U ^TC(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)] make fourth order cumulant matrix C for orthogonal separation matrix U _z(M _o) diagonalization computing function, belong to existing function;

Ask orthogonal separation matrix U, orthogonal separation matrix U is simultaneously to all fourth order cumulant matrix C _z(M _o) carry out Joint diagonalization, computation process is such as formula shown in (ix):

$\min C\left(U\right)=\underset{\mathrm{Mo}\⋐M}{\mathrm{\Σ}}\mathrm{off}\left[U^{T}C\left(M\right)U\right]---\left(\mathrm{ix}\right)$

In formula (ix), non-diagonal component off () is defined as described A represents a matrix, a _ijbe each element of matrix A, minC (U) is to all fourth order cumulant matrix C _z(M _o) carry out the result of calculation of Joint diagonalization;

Employing Givens has rotated the optimization to algorithm, obtains unitary matrix U;

Source signal y (t) through type (x) is estimated to obtain:

y(t)＝U ^T·Q·x(t) (x)。

4. electrical equipment fault sound detection method according to claim 1, it is characterized in that, described step (3) concrete steps are:

D, pre-emphasis, framing and windowing operation are carried out to isolated source signal y (t) in step (2);

E, FFT conversion is carried out, i.e. Fast Fourier Transform (FFT) to the every frame voice signal after steps d process, obtains its frequency spectrum, then delivery square as discrete power spectrum S (k);

F, calculating S (k), by the performance number of gained after band-pass filter group, obtain V parameter 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, get D ₁, D ₂..., D _kas the parameter of MFCC;

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

If the voice signal of steps d has divided p frame vector, i.e. { T (1), T (2), T (n), T (p) }, T (n) is the speech characteristic vector of the n-th frame, 1≤n≤p, reference sample has q frame vector, i.e. { R (1), R (2), R (m), R (q) }, R (m) is the speech characteristic vector of m frame, 1≤m≤q, then dynamic time warping DTW algorithm utilizes Time alignment function j=w (i) to complete the mapping of vector to be tested and reference template vector time axle, and this warping function w meets following formula (xi):

$D=\underset{w\left(i\right)}{\min }\underset{i=1}{\overset{l}{\mathrm{\Σ}}}d[T\left(i\right),R\left(w\left(i\right)\right)]---\left(\mathrm{xi}\right)$

In formula (ix), d [T (i), R (w (i))] is the distance measure between vector T (i) to be tested and reference template vector R (j); T (i) represents the speech characteristic vector of the i-th frame in T; R (w (i)) represents jth frame speech characteristic vector in R; The minor increment of D then between vector to be tested and reference sample vector;

After utilizing DTW to be mated with all reference sample templates by sound pattern to be tested, the minimum reference sample template of matching distance is exactly the result of power equipment work sound identification.