CN104064186A - Electrical equipment failure tone detection method based on independent component analysis - Google Patents

Abstract

The invention relates to an electrical equipment failure tone detection method based on independent component analysis. The method comprises the following steps: an electrical equipment operation sound signal is acquired by a microphone array; independent sound source signals are separated from the sound signal acquired by the microphone array by adopting the independent component analysis method based on the maximum negentropy, that is the Fast-ICA algorithm; and the Mel Mel-frequency ceptral coefficient MFCC of the independent sound source signals are extracted as sound feature parameters, the sound signal is recognized through a mode matching algorithm, the sound template to be tested is matched with all reference sample templates, and the reference sample template having the minimum matching distance is the electrical equipment working tone recognition result: if the reference sample template having the minimum matching distance is the normal tone, then the working tone of the matched electrical equipment is the normal tone; and if the reference sample template having the minimum matching distance is the failure tone, then the working tone of the matched electrical equipment is the failure tone. According to the invention, the fixed-point iteration algorithm is adopted, so the convergence is more stable and fast.

Description

A kind of electrical equipment malfunction sound detection method based on independent component analysis

Technical field

The present invention relates to a kind of electrical equipment malfunction sound detection method based on independent component analysis, belong to the technical field of maintenance of electric device.

Background technology

Nowadays the most frequently used method of fault diagnosis and state-detection is exactly electric measurement method, but electric parameters can not characterize electrical equipment malfunction information conventionally significantly, and this parameter is relatively difficult to detect, we are difficult to judgment device fault type and information accurately and timely, and electrical equipment can normally be worked before breaking down and not too large variation of electric parameters.Except non-electric quantity, a lot of change informations sends that abnormal voice packet has contained that electrical equipment will break down during as electrical equipment malfunction or out of order signal.Therefore, some parameter indexs of non-electric quantity detection method can be more effectively more convenient we electrical equipment is diagnosed, for example electrical equipment voice signal.

Electrical equipment online monitoring technology is applied already and is played an important role in real time.Electrical equipment malfunction sound diagnostic method in the past generally will detect to realize by touch sensor, under the complex environment at the power transmission and transformation station of high voltage and strong-electromagnetic field, may can produce certain impact to the monitoring result of equipment.In addition, the installation of sensors of contact and maintenance are very inconvenient, once sensor generation problem also may be brought beyond thought consequence.

Since the eighties in 20th century, the fault detect means based on sound signal processing have been widely used in field of industrial production various machinery have been carried out to fault diagnosis, as internal combustion engine and engine etc.; The nineties is fast-developing and extend to other many fields as medical science especially, has obtained huge success.Product based on sound signal processing has been widely used in the every aspect of science and technology, industry, living and education at present, product is also more and more ripe, more and more welcome, as the phonetic entry of the product Google translation of Google company, the phonetic dialing of mobile phone, the technology such as the phonetic search of micro-letter and QQ are applied in the social activity of company of Tengxun, and interactive voice control in various navigation etc.Current power equipment network has had more effective on-line fault monitoring system, propose a kind of method that detects electrical equipment malfunction sound based on Independent component analysis Fast-ICA algorithm herein and monitor in real time the ruuning situation of electrical equipment, it can be used as the effective means of supplementing out economy of one of current electrical equipment monitoring method.

Summary of the invention

For the deficiencies in the prior art, the invention provides a kind of electrical equipment malfunction sound detection method based on independent component analysis.

Technical scheme of the present invention is as follows:

An electrical equipment malfunction sound method based on independent component analysis, comprises that step is as follows:

(1) adopt microphone array, MIC array gathers the voice signal of electric equipment operation;

(2) adopt the Independent component analysis Fast-ICA algorithm based on negentropy maximum to adopt the voice signal of microphone array collection to separate each individual sources signal for step (1);

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

MIC array is by analyzing and process the multi-channel sound signal picking up, the beam pattern main lobe sound source that aims at the mark that array is formed, " zero point " points to interference source to suppress undesired signal, thereby must obtain as far as possible target sound, the problem that needs constantly to regulate microphone directive property while using single or two microphones has not only been eliminated in the formation of wave beam, and the signal to noise ratio (S/N ratio) of output sound signal is significantly improved, obtain high-quality voice signal.

Preferred according to the present invention, the concrete steps that separate individual sources signal in described step (2) are:

In a, step (1), adopt microphone array, i.e. the voice signal of MIC array collection electric equipment operation is designated as: x=[x ₁, x ₂..., x _n], n=1,2,3 ... n, wherein,

x ₁＝a ₁₁s ₁+a ₁₂S ₂+…a _1ms _m

x ₂＝a ₂₁s ₁+a ₂₂S ₂+…a _2ms _m

.

.

.

x _n(t)＝a _n1s ₁+a _n2s ₂+…a _nms _m(i)

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

First, to the employing microphone array described in step a, i.e. the voice signal of MIC array collection electric equipment operation carries out centralization processing, the voice signal obtaining through type (II) gained:

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

Then, to the voice signal obtaining after centralization processing carry out albefaction processing, obtain voice signal z, through type (III) gained:

$z=q\stackrel{\‾}{x}={\mathrm{\Λ}}^{-1/2}{U}^T\stackrel{\‾}{x}---\left(\mathrm{iii}\right)$

In formula (III), each component z of z _iuncorrelated mutually, and there is unit variance E{zz ^t}=I; Λ=diag (d ₁, d ₂..., d _n) be Correlation Matrix the diagonal matrix of n eigenvalue of maximum composition; U ∈ C ^{m × n}, be the matrix of n characteristic of correspondence vector composition;

B. the negentropy approximate evaluation J (x) based on obtaining fine compromise between kurtosis and two non-Gauss's tolerance of negentropy, shown in (IV):

J(x)∝[E(G(x))-E(G(v))] ²(ⅳ)

In formula (IV), G is non-quadratic function arbitrarily, and v is the gaussian variable of zero-mean unit variance;

C. approximate Newton iterative is suc as formula shown in (V):

w←w-[E{zg(w ^Tz)}+βw]/[E{g'(w ^Tz)}+β] (ⅴ)

E{g'(w is multiplied by above formula (V) both sides simultaneously ^tz) }+β, is reduced to:

w←w-E{zg(w ^Tz)}-E{g'(w ^Tz)}w (ⅵ)

By the w standardization in formula (VI), obtain formula (VII):

w←w/||w|| (ⅶ)

In above-mentioned formula (V), w is the initialization vector with unit norm that can choose at random, and z is the albefaction data that formula (III) obtains, and obtains voice signal, and g is the derivative of non-quadratic function G, as shown in the formula:

g ₁(x)＝tanh(a ₁x)

g ₂(x)＝uexp(-x ²/2)

g ₃(x)＝x ³；

In above-mentioned formula, β is a constant, and formula (VI) is the fundamental formular of FastICA, if do not restrained, returns to step (V).

From the basic model of ICA, as long as obtain separation matrix w, wz just can infinitely approach the voice signal s sending in independent signal source ₁, s ₂..., s _m, separate thereby realize blind source;

Described Independent component analysis ICA algorithm is the Fast-ICA algorithm based on negentropy maximum direction, and the advantage of this algorithm is to adopt fixed point iteration algorithm, makes convergence more stable and rapid; And then the Mel frequency cepstral coefficient MFCC that extracts sound is as sound characteristic parameter.

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

A. isolated individual sources signal in step (2) is carried out to pre-emphasis, point frame and windowing operation;

B. every frame voice signal after treatment in step (3) a is carried out to FFT conversion, i.e. Fast Fourier Transform (FFT), obtains its frequency spectrum, then delivery square as discrete power spectrum S (k);

C. calculate S (k) by the performance number of gained after band-pass filter group, obtain M parameter P _m, m=0,1 ..., M-1; Then calculate P _mnatural logarithm, obtain L _m, m=0,1 ..., M-1; Finally calculate L _mdCT discrete cosine transform, obtain D _m, m=0,1 ..., M-1, removes D ₀, get D ₁, D ₂..., D _kas the parameter of MFCC;

D. described pattern matching algorithm is that the concrete steps that dynamic time warping DTW algorithm carries out voice recognition are:

If the voice signal of step (3) a has divided the i.e. { T (1) of N frame vector, T (2), T (n), T (N) }, T (n) is the speech characteristic vector of n frame, reference sample has the i.e. { R (1) of M frame vector, R (2),, R (m) ... R (M) }, the speech characteristic vector that R (m) is m frame, 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 shaft, and this warping function w meets following formula (IV):

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

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

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

In order to use non-Gauss in ICA estimates, must be to the index of a quantification of non-Gauss's definition of stochastic variable, what the present invention adopted is the negentropy approximate evaluation based on obtaining fine compromise between kurtosis and two non-Gauss's tolerance of negentropy.

Beneficial effect of the present invention is:

1, the present invention adopts the Fast-ICA algorithm based on negentropy maximum direction, and the advantage of this algorithm is to adopt fixed point iteration algorithm, makes convergence more stable and rapid;

2, in the present invention, array signal adopts idea of generalized side lobe, obtains the voice signal that quality is higher by the mode of " electronic aiming " from sound source position, can suppress other neighbourhood noises simultaneously, has good spatial selectivity.

Embodiment

Below in conjunction with embodiment, the present invention is described in detail, but is not limited to this.

Embodiment

An electrical equipment malfunction sound detection method based on independent component analysis, comprises that step is as follows:

(1) adopt microphone array, MIC array gathers the voice signal of electric equipment operation;

(2) adopt the Independent component analysis Fast-ICA algorithm based on negentropy maximum to use the voice signal of microphone array collection to separate each individual sources signal for step (1);

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

MIC array is by analyzing and process the multi-channel sound signal picking up, the beam pattern main lobe sound source that aims at the mark that array is formed, " zero point " points to interference source to suppress undesired signal, thereby must obtain as far as possible target sound, the problem that needs constantly to regulate microphone directive property while using single or two microphones has not only been eliminated in the formation of wave beam, and the signal to noise ratio (S/N ratio) of output sound signal is significantly improved, obtain high-quality voice signal.

The concrete steps that separate individual sources signal in described step (2) are:

In a, step (1), adopt microphone array, i.e. the voice signal of MIC array collection electric equipment operation is designated as:

X=[x ₁, x ₂..., x _n], n=1,2,3 ... n, wherein,

x ₁＝a ₁₁S ₁+a ₁₂s ₂+…a _1ms _m

x ₂＝a ₂₁s ₁+a ₂₂s ₂+…a _2ms _m

.

.

.

x _n(t)＝a _n1s ₁+a _n2S ₂+…a _nms _m(i)

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

First, to the employing microphone array described in step a, i.e. the voice signal of MIC array collection electric equipment operation carries out centralization processing, the voice signal obtaining through type (II) gained:

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

Then, to the voice signal obtaining after centralization processing carry out albefaction processing, obtain voice signal z, through type (III) gained:

$z=q\stackrel{\‾}{x}={\mathrm{\Λ}}^{-1/2}{U}^T\stackrel{\‾}{x}---\left(\mathrm{iii}\right)$

In formula (III), each component z of z _iuncorrelated mutually, and there is unit variance E{zz ^t}=I; Λ=diag (d ₁, d ₂..., d _n) be Correlation Matrix the diagonal matrix of n eigenvalue of maximum composition; U ∈ C ^{m × n}, be the matrix of n characteristic of correspondence vector composition;

B. the negentropy approximate evaluation J (x) based on obtaining fine compromise between kurtosis and two non-Gauss's tolerance of negentropy, shown in (IV):

J(x)∝[E(G(x))-E(G(v))] ²(ⅳ)

In formula (IV), G is non-quadratic function arbitrarily, and v is the gaussian variable of zero-mean unit variance;

C. approximate Newton iterative is suc as formula shown in (V):

w←w-[E{zg(w ^Tz)}+βw]/[E{g'(w ^Tz)}+β] (ⅴ)

E{g'(w is multiplied by above formula (V) both sides simultaneously ^tz) }+β, is reduced to:

w←w-E{zg(w ^Tz)}-E{g'(w ^Tz)}w (ⅵ)

By the w standardization in formula (VI), obtain formula (VII):

w←w/||w|| (ⅶ)

In above-mentioned formula (V), w is the initialization vector with unit norm that can choose at random, and z is the albefaction data that formula (III) obtains, and obtains voice signal, and g is the derivative of non-quadratic function G, as shown in the formula:

g ₁(x)＝tanh(a ₁x)

g ₂(x)＝uexp(-x ²/2)

g ₃(x)＝x ³；

In above-mentioned formula, β is a constant, and formula (VI) is the fundamental formular of FastICA, if do not restrained, returns to step (V).

From the basic model of ICA, as long as obtain separation matrix w, wz just can infinitely approach the voice signal s sending in independent signal source ₁, s ₂..., s _m, separate thereby realize blind source;

Described Independent component analysis ICA algorithm is the Fast-ICA algorithm based on negentropy maximum direction, and the advantage of this algorithm is to adopt fixed point iteration algorithm, makes convergence more stable and rapid; And then the Mel frequency cepstral coefficient MFCC that extracts sound is as sound characteristic parameter.

Described step (3) concrete steps are:

A. isolated individual sources signal in step (2) is carried out to pre-emphasis, point frame and windowing operation;

B. every frame voice signal after treatment in step (3) a is carried out to FFT conversion, i.e. Fast Fourier Transform (FFT), obtains its frequency spectrum, then delivery square as discrete power spectrum S (k);

C. calculate S (k) by the performance number of gained after band-pass filter group, obtain M parameter P _m, m=0,1 ..., M-1; Then calculate P _mnatural logarithm, obtain L _m, m=0,1 ..., M-1; Finally calculate L _mdCT discrete cosine transform, obtain D _m, m=0,1 ..., M-1, removes D ₀, get D ₁, D ₂..., D _kas the parameter of MFCC;

D. described pattern matching algorithm is that the concrete steps that dynamic time warping DTW algorithm carries out voice recognition are:

If the voice signal of step (3) a has divided the i.e. { T (1) of N frame vector, T (2), T (n), T (N) }, T (n) is the speech characteristic vector of n frame, reference sample has the i.e. { R (1) of M frame vector, R (2),, R (m) ... R (M) }, the speech characteristic vector that R (m) is m frame, 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 shaft, and this warping function w meets following formula (IV):

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

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

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

In order to use non-Gauss in ICA estimates, must be to the index of a quantification of non-Gauss's definition of stochastic variable, what the present invention adopted is the negentropy approximate evaluation based on obtaining fine compromise between kurtosis and two non-Gauss's tolerance of negentropy.

Claims (3)

1. the electrical equipment malfunction sound detection method based on independent component analysis, is characterized in that, comprises that step is as follows:

(1) adopt microphone array, MIC array gathers the voice signal of electric equipment operation;

(2) adopt the Independent component analysis Fast-ICA algorithm based on negentropy maximum to adopt the voice signal of microphone array collection to separate each individual sources signal for step (1);

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

2. a kind of method that detects electrical equipment malfunction sound based on Independent component analysis Fast-ICA algorithm according to claim 1, is characterized in that, the concrete steps that separate individual sources signal in described step (2) are:

In a, step (1), adopt microphone array, i.e. the voice signal of MIC array collection electric equipment operation is designated as: x=[x ₁, x ₂..., x _n], n=1,2,3 ... n, wherein,

x ₁＝a ₁₁S ₁+a ₁₂s ₂+…a _1ms _m

x ₂＝a ₂₁s ₁+a ₂₂S ₂+…a _2mS _m

.

.

.

x _n(t)＝a _n1s ₁+a _n2S ₂+…a _nmS _m(i)

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

First, to the employing microphone array described in step a, i.e. the voice signal of MIC array collection electric equipment operation carries out centralization processing, the voice signal obtaining through type (II) gained:

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

Then, to the voice signal obtaining after centralization processing carry out albefaction processing, obtain voice signal z, through type (III) gained:

$z=q\stackrel{\‾}{x}={\mathrm{\Λ}}^{-1/2}{U}^T\stackrel{\‾}{x}---\left(\mathrm{iii}\right)$

In formula (III), each component z of z _iuncorrelated mutually, and there is unit variance E{zz ^t}=I; Λ=diag (d ₁, d ₂..., d _n) be Correlation Matrix the diagonal matrix of n eigenvalue of maximum composition; U ∈ C ^{m × n}, be the matrix of n characteristic of correspondence vector composition;

B. the negentropy approximate evaluation J (x) based on obtaining fine compromise between kurtosis and two non-Gauss's tolerance of negentropy, shown in (IV):

J(x)∝[E(G(x))-E(G(v))] ²(ⅳ)

In formula (IV), G is non-quadratic function arbitrarily, and v is the gaussian variable of zero-mean unit variance;

C. approximate Newton iterative is suc as formula shown in (V):

w←w-[E{zg(w ^Tz)}+βw]/[E{g'(w ^Tz)}+β] (ⅴ)

E{g'(w is multiplied by above formula (V) both sides simultaneously ^tz) }+β, is reduced to:

w←w-E{zg(w ^Tz)}-E{g'(w ^Tz)}w (ⅵ)

By the w standardization in formula (VI), obtain formula (VII):

w←w/||w|| (ⅶ)

In above-mentioned formula (V), w is the initialization vector with unit norm that can choose at random, and z is the albefaction data that formula (III) obtains, and obtains voice signal, and g is the derivative of non-quadratic function G, as shown in the formula:

g ₁(x)＝tanh(a ₁x)

g ₂(x)＝uexp(-x ²/2)

g ₃(x)＝x ³；

In above-mentioned formula, β is a constant, and formula (VI) is the fundamental formular of FastICA, if do not restrained, returns to step (V).

3. a kind of method that detects electrical equipment malfunction sound based on Independent component analysis Fast-ICA algorithm according to claim 2, is characterized in that, described step (3) concrete steps are:

A. isolated individual sources signal in step (2) is carried out to pre-emphasis, point frame and windowing operation;

B. every frame voice signal after treatment in step (3) a is carried out to FFT conversion, i.e. Fast Fourier Transform (FFT), obtains its frequency spectrum, then delivery square as discrete power spectrum S (k);

C. calculate S (k) by the performance number of gained after band-pass filter group, obtain M parameter P _m, m=0,1 ..., M-1; Then calculate P _mnatural logarithm, obtain L _m, m=0,1 ..., M-1; Finally calculate L _mdCT discrete cosine transform, obtain D _m, m=0,1 ..., M-1, removes D ₀, get D ₁, D ₂..., D _kas the parameter of MFCC;

D. described pattern matching algorithm is that the concrete steps that dynamic time warping DTW algorithm carries out voice recognition are:

If the voice signal of step (3) a has divided the i.e. { T (1) of N frame vector, T (2), T (n), T (N) }, T (n) is the speech characteristic vector of n frame, reference sample has the i.e. { R (1) of M frame vector, R (2),, R (m) ... R (M) }, the speech characteristic vector that R (m) is m frame, 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 shaft, and this warping function w meets following formula (IV):

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

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

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