CN104330255A - Gear fault diagnosing method based on multi-sensor information fusion - Google Patents

Abstract

The invention provides a gear fault diagnosing method based on multi-sensor information fusion. The method mainly comprises the following steps: acquiring a sensor vibrating signal through a vibrating sensor; synchronously integrating each sensor vibrating signal by the spline interpolation and synchronous reconstruction method; performing wavelet transformation and decomposition to obtain wavelet coefficients within each frequency band; performing related noise reducing processing for the wavelet coefficient in each frequency band to obtain a high-signal-to-noise ratio wavelet coefficient; calculating an energy spectrum entropy of the high-signal-to-noise ratio wavelet coefficient; inputting the energy spectrum entropy of the coefficient of each layer into an SOM neural network to obtain a secondary evidence; performing fusion analysis for the secondary evidence of each sensor according to the fusion rule in the D-S evidence theory. With the adoption of the method, the trust degree of the fault analysis result is further raised, certain noise interference is adapted, the overall uncertainty caused by error is reduced, the accuracy of fault analysis is greatly improved, and great means are brought to the information fusion field.

Description

A kind of gear failure diagnosing method based on multi-sensor information fusion

Technical field

The present invention relates to a kind of gear failure diagnosing method, be specifically related to a kind of gear failure diagnosing method based on multi-sensor information fusion

Background technology

Gear drive is as the most important kind of drive of heave-load device, and its fault can cause heave-load device work efficiency to reduce, even major accident.When the factors such as friction force, linear Stiffness, non-stationary load that are subject to when gear especially affect the non-stationary characteristic shown, characteristic signal becomes fainter, its failure rate height of a specified duration not under, therefore, realize strong noise background lower gear fault detection and diagnosis early, for reduction gear distress rate, improve equipment dependability, promote the aspects such as production, all have important engineering significance.Current Gear Fault Diagnosis is many, and by analyzing the vibration signal of single-sensor collection, analysis result easily produces uncertainty, correctly cannot judge failure cause.

Summary of the invention

The object of the present invention is to provide a kind of can based on the gear failure diagnosing method of multi-sensor information fusion under real noise background, adapt to very noisy interference, effectively can extract fault characteristic information, eliminate the uncertainty produced in identifying, improve diagnostic accuracy.

For reaching above object, the present invention takes following technical scheme to be achieved:

Based on a gear failure diagnosing method for multi-sensor information fusion, it is characterized in that, the method concrete steps are as follows:

Utilize vibration transducer pick-up transducers vibration signal.

Further, synchronous integration sensor vibration signal, step is as follows:

Adopt the reconstruct of sp1ine spline interpolation mode, synchronous same-speed rate asynchronous collecting and non-integral multiple multi tate collection.

Further, adopt the sensor vibration signal after the integration described in wavelet transformation decomposition, obtain the wavelet coefficient in each frequency band.

Further, carry out noise reduction process to the wavelet coefficient of described each frequency band, obtain the wavelet coefficient of high s/n ratio, step is as follows:

Energy normalized is carried out to wavelet coefficient and related coefficient, obtains energy and the correlation of each layer wavelet coefficient energy and related coefficient, the correlativity between composition wavelet coefficient;

According to the correlativity between wavelet coefficient, by retaining the wavelet coefficient that vibration signal decomposes, eliminating the wavelet coefficient decomposed by noise, reaching noise reduction;

The wavelet coefficient of signal at different adjacent yardsticks is repeatedly multiplied, until wavelet coefficient energy meets a threshold value ratio relevant with noise immune level, finally obtains the wavelet coefficient that signal to noise ratio (S/N ratio) is higher.

Further, the Energy Spectrum Entropy of the wavelet coefficient of high s/n ratio is calculated.

Further, be input in SOM neural network by the Energy Spectrum Entropy of high s/n ratio wavelet coefficient, obtain the sub-evidence of sensor, step is as follows:

Initialization is carried out to SOM neural network, for SOM neural network mapping layer initial weight vector selects less random value;

Using each frequency band wavelet coefficient Energy Spectrum Entropy as the input vector of SOM neural network, input to input layer;

Calculate the weight vector of mapping layer and the distance of input vector, obtain the neuron that has minor increment, as output neuron;

Revise the weights of output neuron and adjacent neurons thereof, obtain the sub-evidence of each sensor;

In order to keep each sensor characteristic data unified, the SOM neural network corresponding with each sensor has identical basic setup; SOM is provided with 1 input layer; 1 output layer, without hidden layer.

Finally, carry out fusion diagnosis according to the fusion rule of D-S evidence theory to described sense signals evidence, step is as follows:

Structure D-S evidence theory identification framework Ω, step is as follows:

The cluster position and each Jiao of identification framework first basic trust function distribution method that identify uncertain feature, described competition layer different faults type clustering neuron is produced, structure D-S evidence theory identification framework Ω according to described SOM neural network output neuron.

Distribute the first basic trust function of each Jiao in described identification framework Ω, step is as follows:

The present invention proposes the basic trust function distribution method of Corpus--based Method SOM neural network recognization rate, according to the cluster position of SOM neural network competition layer different faults type clustering neuron, SOM neural network competition layer is divided into multiple different region, each region is corresponding with burnt unit each in identification framework, add up the number that often kind of fault type test sample book output neuron falls into each region, and then obtain the probability that often kind of fault type sample output neuron falls into each region, the distribution of the first basic trust function of each Jiao of identification framework is completed with this;

According to the fusion rule of basic trust function and D-S evidence theory, calculate the confidence level size after the sub-evidence fusion of multisensor, judge the Output rusults after merging.

The invention has the beneficial effects as follows that fault diagnosis result is reliable, strong noise background interference can be adapted to, effective extraction fault characteristic information, eliminate the uncertainty produced in identifying, substantially increase the accuracy of fault diagnosis, be applicable to the fault diagnosis of the heave-load device transmission gear under harsh environments such as coalcutter.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the method for the invention;

Fig. 2 is the time domain beamformer of four kinds of state gears that the embodiment of the present invention gathers;

Fig. 3 is the wavelet decomposition result figure of the broken teeth Gearbox vibration signal of the embodiment of the present invention;

Fig. 4 be the embodiment of the present invention noise reduction before and after comparison diagram;

Fig. 5 be the embodiment of the present invention noise reduction before and after FFT spectrogram comparison diagram;

Fig. 6 is the feature Energy Spectrum Entropy of four kinds of state gears of the embodiment of the present invention;

Fig. 7 is SOM neural network recognization result and the input neuron importance degree of the embodiment of the present invention;

Fig. 8 is the first zoning of each Jiao of SOM neural network competition layer of the sensor 1 of the embodiment of the present invention;

Fig. 9 is the gear distress discrimination comparison diagram after the fusion of the embodiment of the present invention.

Figure 10 is the embodiment of the present invention each sensor basic trust function allocation result.

Embodiment

Hereafter will describe embodiments of the invention in detail by reference to the accompanying drawings.It should be noted that the combination of technical characteristic or the technical characteristic described in following embodiment should not be considered to isolated, they can mutually be combined thus be reached better technique effect.In the accompanying drawing of following embodiment, the identical label that each accompanying drawing occurs represents identical feature or parts, can be applicable in different embodiment.

As shown in Figure 1, a kind of gear failure diagnosing method based on multi-sensor information fusion, concrete steps are as follows:

S100 utilizes vibration transducer pick-up transducers vibration signal.

S200 integrates described sensor vibration signal, and step is as follows:

Adopt sp1ine spline interpolation reconstruct mode, synchronous same-speed rate asynchronous collecting and non-integral multiple multi tate collection, described sp1ine spline interpolation reconstruct mode is as follows:

Sp1ine spline interpolation function s (t) is at sub-range [t _i-1, t _i] on expression formula be:

$s_i\left(t\right)=\frac{M_{i-1}}{6h_i}{(t_i-t)}^3+\frac{M_i}{6h_i}{(t-t_{i-1})}^3+(\frac{y_i-1}{h_i}-\frac{M_i-1}{6}{h}_i)(t_i-t)+(\frac{y_i}{h_i}-\frac{M_i}{6}{h}_i)(t-t_{i-1})t_{i-1}\≤t\≤t_i(i=\mathrm{1,2},...,n)$

H in formula _i=t _i-t _i-1, M _i=s ⁿ(t _i), determine M _imethod as follows:

With second derivative structure, M ₀, M ₁..., M _nit is system of linear equations

$\left\{\begin{array}{c}2M_0+a_0{M}_1=d_0\\ b_i{M}_{i-1}+2M_i+a_i{M}_{i+1}=d_i\\ b_n{M}_{i-1}+2M_n=d_n\end{array}\right.$

Solution, in formula

$a_0=0,d_0=2y_0^n;b_n=0,d_n=2y_n^n$

$a_i=\frac{h_{i+1}}{h_i+h_{i+1}},b_i=1-a_i,i=\mathrm{1,2},...,n-1$

$d_i=\frac{6}{h_i+h_{i+1}}(\frac{y_{i+1}-y_i}{h_{i+1}}-\frac{y_i-y_{i-1}}{h_i}),i=\mathrm{1,2},...,n-1$

Obtain a _i, b _i, d _iafter, can M be tried to achieve ₀, M ₁..., M _nvalue; Cubic spline functions s (t) expression can be determined.

S300 adopts the signal after the integration described in wavelet transformation decomposition, obtains the wavelet coefficient in each frequency band.

S400 carries out noise reduction process to the wavelet coefficient of each frequency band, and obtain the wavelet coefficient of high s/n ratio, step is as follows:

S401 utilizes the wavelet coefficient after decomposing, and to the wavelet coefficients multiplication of adjacent yardstick, obtains related coefficient,

Corr ₂(m,n)＝Y(m,n)Y(m+1,n)n＝1,2...,N

In formula, N represents counting of discrete signal, and n represents the time, and m represents yardstick, and Y represents the wavelet transformation of signal;

S402 carries out energy normalized to wavelet coefficient and related coefficient, obtains energy and the correlation of each layer wavelet coefficient energy and related coefficient, the correlativity between composition wavelet coefficient:

To Corr ₂(m, n) carries out energy normalized, its energy normalizing to Y (m, n) is got on, the correlation Corr after normalization ₂for

$\left\{\begin{array}{c}\mathrm{NewCor}{r}_2(m,n)=\mathrm{Cor}{r}_2(m,n)\sqrt{\mathrm{PY}\left(m\right)/\mathrm{PCorr}\left(m\right)}\\ \mathrm{PY}\left(m\right)=\underset{n}{\mathrm{\Σ}}Y{(m,n)}^2\\ \mathrm{PCor}{r}_2\left(m\right)=\underset{n}{\mathrm{\Σ}}\mathrm{Cor}{r}_2{(m,n)}^2\end{array}\right.$

In formula, PY (m), PCorr ₂m () is the energy of m layer wavelet coefficient and related coefficient respectively;

S403 is according to the correlativity between wavelet coefficient, and by retaining the wavelet coefficient that vibration signal decomposes, eliminate the wavelet coefficient decomposed by noise, reach noise reduction, step is as follows:

If | NewCorr2 (m, n) | >|Y (m, n) |, then think that n point place wavelet transformation is caused by vibration signal, Y (m, n) given the relevant position of Yf, and Y (m, n) is set to 0; Otherwise think to be caused by noise, Y (m, n) retains;

The wavelet coefficient of signal at different adjacent yardsticks is repeatedly multiplied by S404, until wavelet coefficient energy meets a threshold value ratio relevant with noise immune level, finally obtains the wavelet coefficient that signal to noise ratio (S/N ratio) is higher.

S500 calculates the Energy Spectrum Entropy of the wavelet coefficient of high s/n ratio, and computing method are as follows:

If e _ijbe the energy of i-th wavelet coefficient jth data point, e _i={ e _i1, e _i2..., e _ijthe synthesis of i-th all data point energy of wavelet coefficient, i-th wavelet coefficient jth probability that data point signal energy exists in i-th wavelet coefficient gross energy is the small echo correlated characteristic Energy Spectrum Entropy then defined is:

$W_{\mathrm{CFSEi}}=-\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{p}_{\mathrm{ij}}\log p_{\mathrm{ij}}.$

Each layer coefficients Energy Spectrum Entropy is input in SOM neural network by S600, and obtain sub-evidence, concrete steps are as follows:

S601 carries out initialization to SOM neural network, is SOM neural network competition layer initial weight vector w _j(0) less random value is selected, j=1,2 ..., l, w _j(0) different, l is neuronic number in network;

The input vector of S602 using the Energy Spectrum Entropy of high s/n ratio wavelet coefficient as SOM neural network, input vector X=(x ₁, x ₂, x ₃..., x _m) ^tinput to input layer;

S603 calculates the weight vector of mapping layer and the distance of input vector, and obtain the neuron that has minor increment, as output neuron, step is as follows:

A jth neuron of mapping layer and the range formula of input vector are:

$d_j=\left|\right|X-W_j\left|\right|=\sqrt{\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{(x_i\left(t\right)-{\mathrm{\ω}}_{\mathrm{ij}}\left(t\right))}^2}$

In formula, ω _ijfor the weights between the i neuron of input layer and the j neuron of mapping layer.By calculating, obtaining the neuron that has minor increment, being referred to as neuron of winning, being designated as j ^*, namely determine certain unit k, make, for arbitrary j, to have and provide its adjacent neurons set.

S604 revises the weights of output neuron and adjacent neurons thereof, and the sub-evidence step obtaining each sensor is as follows:

By following formula correction output neuron j ^*and the weights of adjacent neurons.

Δω _ij＝ω _ij(t+1)-ω _ij(t)＝η(t)(x _i(t)-ω _ij(t))

In formula, η is one and is greater than 0 constant being less than 1, along with time variations drops to 0 gradually.

Calculate and export o _k

$o_k=f\left(\underset{j}{\min }\right||X-W_j|\left|\right)$

In formula, f (*) is generally function or other nonlinear functions of 0 ~ 1;

If function obtained above reaches requirement, algorithm terminates, otherwise the Energy Spectrum Entropy of other wavelet coefficients is input to input layer recalculates;

In order to keep each sensor characteristic data unified, the SOM neural network corresponding with each sensor has identical basic setup; SOM is provided with 1 input layer; 1 output layer, without hidden layer.

Under certain fault mode, fault characteristic information is through the information fusion of characteristic layer SOM neural network, correct conclusion has been contained in acquired results, but be correlated with after noise reduction due to small echo and still comprise certain residual noise, SOM neural network there will be output neuron between several master sample clustering neuron in identifying, produce and identify uncertainty, the scope that SOM neural network recognization result is floated is larger, also need to reduce its scope by information fusion, obtain more accurate fault diagnosis conclusion.

S700 carries out fusion diagnosis according to the fusion rule of D-S evidence theory to described sense signals evidence, and step is as follows:

S701 constructs D-S evidence theory identification framework Ω, and method is as follows:

D-S evidence theory identification framework is all possible fault mode set, the corresponding a kind of contingent fault mode of each subset in identification framework; According to SOM neural network theory, if output neuron is identical with certain standard failure clustering neuron position, illustrate that it there occurs corresponding fault; If output neuron does not overlap with clustering neuron, then can think the clustering neuron that this neuron may belong to close with it, its fault is between a lot of standard failure;

Export based on local decision by SOM neural network, the cluster position and each Jiao of identification framework first basic trust function distribution method that identify uncertain feature, competition layer different faults type clustering neuron is produced, structure D-S evidence theory identification framework Ω according to SOM neural network output neuron;

S702 distributes the first basic trust function of each Jiao in described identification framework Ω, and step is as follows:

About the definition of basic trust function requires as follows:

$\left\{\begin{array}{c}\underset{A\∈\mathrm{\Ω}}{\mathrm{\Σ}}f\left(A\right)=1\\ f\left(\mathrm{\φ}\right)=0\end{array}\right.$

If A is a burnt unit in identification framework Ω, then f (A) is called the basic trust function of A, it reflects the confidence level size to A itself.Meanwhile, the confidence value sum defining each subset in A is the reliability function of A:

$\mathrm{Bel}\left(A\right)=\underset{A\∈\mathrm{\Ω}}{\mathrm{\Σ}}f\left(B\right)$

About the trusting degree of a proposition A, it is inadequate that alone reliability function describes, because Bel (A) can not reflect the degree suspecting A, introduces one and represents that A suspects the likelihood function of degree:

$\mathrm{pl}\left(A\right)=1-\mathrm{Bel}\left(\stackrel{\‾}{A}\right)$

The present invention proposes the basic trust function distribution method of Corpus--based Method SOM neural network recognization rate, according to the cluster position of SOM neural network competition layer different faults type clustering neuron, SOM neural network competition layer is divided into multiple different region, each region is corresponding with burnt unit each in identification framework, add up the number that often kind of fault type test sample book output neuron falls into each region, and then obtain the probability that often kind of fault type sample output neuron falls into each region, the distribution of the first basic trust function of each Jiao of identification framework is completed with this;

S703, according to the fusion rule of basic trust function and D-S evidence theory, calculates the confidence level size after the sub-evidence fusion of multisensor, and judge the Output rusults after merging, step is as follows:

After the multiple sub-evidence body obtained multisensor carries out local decision analysis, need carry out comprehensively according to certain rule of combination its analysis result.If f _i2 ^Ωthe upper basic probability assignment function corresponding to i-th evidence body (total v evidence body), corresponding focus element is respectively A ₁₁..., A _1k1..., A _v1..., A _vkv, then

$K_1=\underset{A_{1/1}\&cap;...\&cap;A_{v/v}}{\mathrm{\&Sigma;}}{f}_1\left(A_{1/1}\right)f\left(A_{v/v}\right)<1$

Confidence level after fusion is:

$f\left(A\right)=\left\{\begin{array}{cc}\frac{\underset{A_{1/1}\&cap;...\&cap;A_{v/v}=A}{\mathrm{\&Sigma;}}{f}_1\left(A_{1t1}\right)f_v\left(A_{\mathrm{vtv}}\right)}{1-K_t}& A=\mathrm{\&phi;}\\ 0& \mathrm{else}\end{array}\right.$

According to the confidence level size after fusion, the Output rusults after the fusion of many evidences body can be judged.

The validity of Binding experiment checking said gear fault method, experimental technique is as follows:

This experiment is carried out on mechanical fault integrated simulation experiment bench, and testing table is equipped with that gear is normal, tooth surface abrasion, broken teeth and few tooth four kinds of state gear distress parts, carries out replacing as required can simulate gear most common failure to the gear in gear case.Mechanical fault integrated simulation experiment bench structure comprises motor, rotary system, is gear case, 3 vibration transducers, and sensor adopts acceleration vibration transducer, and the vibration signal of pick-up transducers, sample frequency 20kHz, gathers 30s altogether, data sample length 8000.

The vibration signal noise that laboratory gathers is general less, and in the condition of work of reality, there is very large noise, therefore outside noise is taked to add testing table signal, Gearbox vibration signal under Reality simulation condition of work, outside noise background is added to the testing table collected 3 sensor signals, then carry out adopting synchronously integrating each sensor signal of sp1ine spline interpolation reconstruction signal method, for sensor 1, synchronous integration backgear is normal, tooth surface abrasion, the time domain waveform of broken teeth and few tooth four kinds of state gears as shown in Figure 2, the vibration signal of gear presents the obvious cycle less as can be seen from Figure 2, but the vibration signal of other three kinds of gears is not obviously distinguished.

Wavelet decomposition is carried out to signal, adopts db5 small echo through test of many times, carry out 6 layers of decomposition.Figure 3 shows that the wavelet decomposition result figure of the broken teeth Gearbox vibration signal of sensor 1, a6 represents the approximate part of signal, d1 ~ d6 represents the HFS of signal, as can be seen from the figure a6, d6 and d5 layer wavelet coefficient shows some cycles, other frequency ranges are due to noise contributions, cannot extract the useful information of signal from the wavelet coefficient of d2 ~ d6 frequency range, and the HFS of vibration signal often contains signal fault information.

Small echo correlativity is utilized to carry out noise reduction process to signal, only retain each wavelet coefficient decomposed by useful vibration signal, carry out wavelet reconstruction, before and after noise reduction, time domain waveform contrast as shown in Figure 4, as can be seen from the figure to be correlated with noise reduction through small echo, signal shows certain periodicity, the noise background due to elimination, cause signal amplitude to have obvious reduction, reservation be useful signal.Fig. 5 is the signal Spectrum Analysis before and after noise reduction, as can be seen from the figure, before noise reduction, signal is subject to real noise background interference, all be submerged in low-frequency range and high band, only at some Frequency point, its amplitude com parison is given prominence to, but can not observe the feature of the whole frequency range of vibration signal, carrying out small echo to signal is correlated with after noise reduction, can embody the characteristic frequency of signal low-frequency range, the characteristic frequency of Mid Frequency also can be observed, and better demonstrates the characteristic frequency feature of signal.

Figure 6 shows that the small echo correlated characteristic Energy Spectrum Entropy of each yardstick of HFS, as can be seen from the figure: normal gear d1 ~ d6 layer entropy is minimum, when gear breaks down, its each layer entropy has obvious increase, illustrate that fault causes each layer signal more complicated, wherein broken teeth gear, few gear, the variation tendency of abrased gear each layer entropy is identical, d1 ~ d4 layer reduces gradually, d5 layer entropy is the highest, but observe broken teeth gear, few gear, the small echo correlated characteristic Energy Spectrum Entropy of abrased gear finds, removing d4 layer, other layer of small echo correlated characteristic Energy Spectrum Entropy broken teeth gear entropy is maximum, next is abrased gear, what entropy was minimum is few gear, roughly gear-type can be distinguished according to the size of each layer entropy.

Using HFS small echo correlated characteristic Energy Spectrum Entropy as the input of SOM neural network, SOM network input layer neuron number is 6, and competition layer neuron number is 64, is arranged in the hexagon two-dimensional grid of 8 × 8, after training, cluster result and input neuron importance degree are as shown in Figure 8.As can be seen from the figure 4 kinds of gear condition obtain differentiation substantially, and wherein normal type cluster is in neuron 9, and 17; Broken teeth type cluster in neuron 57,58; Tooth type cluster is in neuron 7 less, and 8; Wear type cluster in neuron 46,54,55.

According to SOM neural network theory, if output neuron is identical with certain standard failure sample position in the position of output layer, illustrate that sample to be checked there occurs corresponding fault; If output neuron does not overlap with clustering neuron, then can think and can not determine this neuronic classification, but likely belong to the clustering neuron close with it, its fault, between a lot of standard failure, illustrates that this several standard failure all likely occurs.Choose the signal verification sample of sensor 1, SOM neural network classification effect is tested, the discrimination that its output neuron and the clustering neuron of master sample overlap completely is few tooth type gear to the maximum, minimum is wear type gear, discrimination is only 0.74, and Fault Identification effect is unsatisfactory.Utilize D-S evidence theory to merge multi-sensor information below, carry out diagnosing information fusion fault.

First, structure D-S evidence theory identification framework, by being input to the recognition result analysis of SOM neural network to every type sample, and the basic trust function distribution method of the Corpus--based Method SOM Network Recognition rate proposed below, conformation identification framework is:

Ω＝{{A},{B},{C},{D},{A,B},{B,D},{C,D},{A,C},{A,B,C,D}}

In formula: A-is normal, B-broken teeth, the few tooth of C-, D-weares and teares

Then, the burnt unit of each in identification framework is carried out to the distribution of basic trust function, according to the SOM neural network competition layer cluster result trained, by the output neuron of competition layer according to the burnt element type in identification framework, be divided into different regions, a burnt unit wherein in the corresponding identification framework in each region; Even output neuron is in a certain region, and the fault type that there occurs the corresponding burnt unit in this region is described.The region that the SOM neural network competition layer that sensor 1 is trained divides as shown in Figure 8.Set up test sample book collection, add up each sensor, every type sample after SOM neural network recognization, its output neuron falls into the number in each region of Fig. 8, complete distribute the basic trust function of each burnt unit of identification framework with the probability that falls in each region, basic trust function allocation result as shown in Figure 10.

According to basic trust function and D-S evidence theory composition rule, by D-S evidence theory composition rule proposed above, add up the discrimination after 2 sensor fusion and 3 sensor fusion calculating as shown in Figure 9.Contrast from figure and can find out, 2 sensor informations merge, all types of discrimination reaches more than 79%, three sensor informations merge, all types of discrimination reaches more than 83%, few its discrimination of tooth type gear reaches 89%, and compare single-sensor fault diagnosis, diagnostic accuracy improves.

Although give some embodiments of the present invention, it will be understood by those of skill in the art that without departing from the spirit of the invention herein, can change embodiment herein.Above-described embodiment is exemplary, should using embodiment herein as the restriction of interest field of the present invention.

Claims (7)

1. based on a gear failure diagnosing method for multi-sensor information fusion, it is characterized in that, comprise the steps:

Utilize vibration transducer pick-up transducers vibration signal;

The described sensor vibration signal of synchronous integration;

Adopt the sensor vibration signal after the integration described in wavelet transformation decomposition, obtain the wavelet coefficient in each frequency band;

Noise reduction process is carried out to the wavelet coefficient of described each frequency band, obtains the wavelet coefficient of high s/n ratio;

The Energy Spectrum Entropy of the wavelet coefficient of the high s/n ratio described in calculating;

The Energy Spectrum Entropy of described high s/n ratio wavelet coefficient is input in SOM neural network, obtains sense signals evidence;

According to the fusion rule of D-S evidence theory, fusion diagnosis is carried out to described sense signals evidence.

2. a kind of gear failure diagnosing method based on multi-sensor information fusion as claimed in claim 1, it is characterized in that, the step of the described sensor vibration signal of described synchronous integration, adopts the reconstruct of sp1ine spline interpolation mode, synchronous same-speed rate asynchronous collecting and non-integral multiple multi tate collection.

3. a kind of gear failure diagnosing method based on multi-sensor information fusion as claimed in claim 1, it is characterized in that, described carries out noise reduction process to the wavelet coefficient of described each frequency band, and the step obtaining the wavelet coefficient of high s/n ratio is as follows:

Utilize the wavelet coefficient after described decomposition, the wavelet coefficients multiplication with adjacent yardstick, obtains related coefficient;

Energy normalized is carried out to the wavelet coefficient after described decomposition and described related coefficient, obtains energy and the correlation of each layer wavelet coefficient energy and related coefficient, the correlativity between composition wavelet coefficient;

According to the correlativity between described wavelet coefficient, by retaining the wavelet coefficient that vibration signal decomposes, eliminating the wavelet coefficient decomposed by noise, reaching noise reduction;

Wavelet coefficient after decomposing described in other is repeatedly multiplied with the wavelet coefficient of adjacent yardstick, until wavelet coefficient energy meets a threshold value ratio relevant with noise immune level, finally obtains the wavelet coefficient that signal to noise ratio (S/N ratio) is higher.

4. a kind of gear failure diagnosing method based on multi-sensor information fusion as claimed in claim 1, it is characterized in that, described is input to the Energy Spectrum Entropy of described high s/n ratio wavelet coefficient in SOM neural network, and the step obtaining sense signals evidence is as follows:

Initialization is carried out to SOM neural network, for described SOM neural network mapping layer initial weight vector selects less random value;

Input vector using the Energy Spectrum Entropy of described high s/n ratio wavelet coefficient as described SOM neural network, is input to input layer;

Calculate the weight vector of described mapping layer and the distance of described input vector, obtain the neuron that has minor increment, as output neuron;

Revise the weights of described output neuron and adjacent neurons thereof, obtain the sub-evidence of each sensor.

5. a kind of gear failure diagnosing method based on multi-sensor information fusion as claimed in claim 1, is characterized in that, described step of carrying out fusion diagnosis to described sense signals evidence according to the fusion rule of D-S evidence theory is as follows:

Structure D-S evidence theory identification framework Ω;

Distribute the first basic trust function of each Jiao in described identification framework Ω;

According to the fusion rule of described basic trust function and described D-S evidence theory, calculate the confidence level size after the sub-evidence fusion of multisensor, judge the Output rusults after merging.

6. a kind of gear failure diagnosing method based on multi-sensor information fusion as claimed in claim 5, it is characterized in that, the step of described structure D-S evidence theory identification framework Ω is as follows:

The cluster position and each Jiao of identification framework first basic trust function distribution method that identify uncertain feature, described competition layer different faults type clustering neuron is produced, structure D-S evidence theory identification framework Ω according to described SOM neural network output neuron.

7. a kind of gear failure diagnosing method based on multi-sensor information fusion as claimed in claim 5, it is characterized in that, in described distribution described identification framework Ω, the step of each Jiao first basic trust function is as follows:

According to the cluster position of described SOM neural network competition layer different faults type clustering neuron, described SOM neural network competition layer is divided into multiple different region, each region is corresponding with burnt unit each in identification framework, add up the number that often kind of fault type test sample book output neuron falls into each region, and then obtain the probability that often kind of fault type sample output neuron falls into each region, the distribution of the first basic trust function of each Jiao of identification framework is completed with this.