CN106017879B

CN106017879B - Omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features

Info

Publication number: CN106017879B
Application number: CN201610331391.4A
Authority: CN
Inventors: 孙曙光; 赵黎媛; 于晗; 杜太行; 张强; 丁铭真; 刘建强; 陈云飞
Original assignee: Hebei University of Technology
Current assignee: Hebei University of Technology
Priority date: 2016-05-18
Filing date: 2016-05-18
Publication date: 2018-07-03
Anticipated expiration: 2036-05-18
Also published as: CN106017879A

Abstract

Omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features, include the following steps, the first step, acquire fuselage shaking signal and fuselage voice signal in omnipotent breaker divide-shut brake action process, second step carries out denoising using improved wavelet packet Threshold Filter Algorithms；Third walks, and extracts the intrinsic mode function component of several reflection circuit-breaker switching on-off action status informations to de-noising signal respectively using complementary population mean empirical mode decomposition algorithm；4th step, the quantity Z for the intrinsic mode function component that determines to shake respectively, the 5th step calculate its energy ratio, Sample Entropy and Power Spectral Entropy and are used as three category features；6th step, and using feature samples dimension-reduction treatment of the combination core core pivot element analysis method to three category feature of acoustic signal after reunification, M pivot is obtained, the 7th step establishes the order binary tree multi-categorizer model based on Method Using Relevance Vector Machine.

Description

Omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features

Technical field

Technical scheme of the present invention is related to the mechanical fault diagnosis of breaker, specifically a kind of special based on acoustic signal Levy the method for diagnosing faults of fusion.

Background technology

According to planning of the country to intelligent grid construction, intelligent substation is the important component and key of intelligent grid Link, omnipotent breaker intelligence are the important components of intelligent substation, are the key that electric system low-voltage networks One of equipment is widely used in low-voltage distribution system at total input-wire or at important equipment.Both at home and abroad research shows that breaker is easy Mechanical breakdown occurs, Mechanical Failure of HV Circuit Breaker diagnosis is concentrated mainly on to its research work, but omnipotent breaker operates Reliability be equally related to the safe operation of electric system.Therefore, have for the monitoring of its machine performance and fault diagnosis Important meaning.At present, signal acquisition, feature extraction, fault identification 3 are generally comprised for the fault diagnosis technology of breaker The content of link.

The signal acquired in first link has：Divide-shut brake coil current and voltage, is cut-off moving contact stroke or angle of eccentricity Electric current and arc voltage, contact stress etc..In recent years, it is increasingly becoming using vibration and voice signal detection breaker mechanical state The hot spot studied both at home and abroad.Although breaker run and the composition complicated difficult of the vibration signal in action process and voice signal with It explains, but comprising abundant mechanical property information, detection vibration is with voice signal as non-intrusive measurement mode, it can be achieved that just The machine performance of prompt reliably monitoring breaker, acquisition do not need to be electrically connected with breaker, will not destroy breaker body Structure.Vibration signal is anti-interference strong, high sensitivity, but sensor is added up the easy high frequency distortion of effects by charge；Voice signal Easily by noise jamming, but bandwidth is measured, can preferably avoid distortion effect.Though the information that acoustic signal contains there are redundancy but There is complementarity simultaneously, the two combines and more effectively breaker can be carried out non-destructive testing.Meanwhile for making an uproar in acoustic signal Wavelet packet soft and hard threshold combination noise suppression preprocessing can be used in sound.

In the characteristic extraction procedure of second link, for non-linear, the non-stationary property of vibration signal, frequently with suitable for Wavelet analysis, empirical mode decomposition (empirical mode with Non-stationary Signal Analysis such as transient state, mutation Decomposition, EMD), population mean empirical mode decomposition (ensemble EMD, EEMD), Hilbert-Huang transform (Hilbert-Huang transform, HHT) etc., but there are self-defect, discomposing effects in fault-signal decomposition for small echo Dependent on the selection of wavelet basis and decomposition scale, without adaptivity, in addition, also there are energy leakages for wavelet decomposition.Although Empirical mode decomposition is a kind of adaptive Time-Frequency Localization analysis method, right but there are modal overlap and end effect phenomenon Its population mean empirical mode decomposition being improved can inhibit modal overlap to a certain extent, but the white noise added is not It can be fully neutralized, without completeness, and the integrated empirical mode decomposition supplemented, by adding opposite white of symbol in couples Noise, so as to substantially reduce reconstructed error, realizes the accurate decomposition of each modal components of signal to echo signal.And for signal For the feature extracted, mean value, variance, covariance, the degree of bias, kurtosis, kurtosis etc. are often extracted in time domain；Width is often extracted on frequency domain Frequency phase frequency feature, envelope spectrum signature；In data sequence, energy feature, singularity, comentropy etc. are often extracted.In these methods, Main problem is that carried feature is single, comprising fault message it is limited, and only rely on a certain feature there are accuracy rate it is low, can By property and stability it is poor the problems such as.To further improve the precision and stability of fault identification, need various features integrating profit With, and finally realize the effective integration of fault characteristic information.

The fault recognition method of third link is with the development of artificial intelligence, frequently with neural network, support vector machines etc.. More common neural network has certain antinoise and generalization ability, but training is needed compared with multisample, and there is part Convergence problem.Although support vector machines be suitable for solve small sample, high dimension, it is non-linear the problems such as, algorithm regularization be Number determines that difficult, prediction results do not have statistical significance, kernel function the inherent limitations such as is limited by Mercer conditions.For asking above Topic, Method Using Relevance Vector Machine (relevance vector machine, RVM) can effectively make up drawbacks described above.In addition, associated vector Machine can also be exported with Probability Forms as a result, more have practicability, using Method Using Relevance Vector Machine as pattern classifier, can effectively be carried High fault recognition rate.However, although Method Using Relevance Vector Machine only needs selection kernel functional parameter, its classification results is still quick to nuclear parameter Sense.At present, suitable kernel functional parameter value is mostly empirically chosen for the selection of nuclear parameter parameter, so as to establish based on phase Close vector machine identification model.However, global optimum's kernel functional parameter value is chosen for the Method Using Relevance Vector Machine diagnostic model established Recognition effect can have good promotion.Therefore, the kernel functional parameter value in Method Using Relevance Vector Machine model needs to seek it It is excellent.

To sum up, there are certain defects for existing fault recognition method, high it is urgent to provide fault recognition rate in the prior art, Stability is good, the omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features.

Invention content

The technical problems to be solved by the invention are：A kind of omnipotent breaker based on acoustic signal Fusion Features is provided Mechanical failure diagnostic method is a kind of multichannel multicharacteristic information fusion and parameter adaptive optimizing failure modes identification model phase With reference to method, for the mechanical breakdown occurred during circuit-breaker switching on-off is carried out it is accurate, stablize, reliably diagnosis is known Not and realize effective identification of the faint mechanical breakdown to omnipotent breaker, the joint-detection of vibration signal and voice signal Mode realizes the message complementary sense between different channel source signals, acquires the wavelet packet that the acoustic signal of gained first makes improvements Threshold denoising pre-processes, and then carries out time-frequency using complementary population mean empirical mode decomposition algorithm to the acoustic signal after noise reduction It decomposes, obtains several intrinsic mode functions, extract the energy ratio, Sample Entropy, Power Spectral Entropy of intrinsic mode function, recycle combination Kernel function core principle component analysis is to above-mentioned more characteristic parameters dimensionality reduction and merges composition characteristic vector, solves single features identification open circuit The low accuracy rate and low stability of device divide-shut brake failure, finally using improve quantum telepotation based on Method Using Relevance Vector Machine point Class model parameter establishes the model combined based on improvement quantum particle swarm with Method Using Relevance Vector Machine, and the failure of dimensionality reduction fusion is special Sign is input to identification model as feature vector and carries out fault diagnosis.

The present invention solves the technical problem the technical scheme adopted is that providing a kind of based on acoustic signal Fusion Features Omnipotent breaker mechanical failure diagnostic method, the acoustic signal is passes through acceleration during omnipotent breaker divide-shut brake The collected fuselage shaking signal of sensor and fuselage voice signal is collected by sound pick-up, it is characterized in that including following step Suddenly

The first step acquires fuselage shaking signal and fuselage voice signal in omnipotent breaker divide-shut brake action process, And collected analog signal is separately converted to digital signal, obtain initial vibration s_v' (t) and voice signal s_a' (t), t For divide-shut brake actuation time；

Second step, using improved wavelet packet Threshold Filter Algorithms respectively by collected vibration signal s_v' (t) and sound Signal s_a' (t) as signals and associated noises carry out denoising, obtain noise reduction vibration signal s_v(t) and noise reduction voice signal s_a(t)；

Third walks, by noise reduction vibration signal s_v(t) with voice signal s_a(t) it respectively as signal s (t) to be decomposed, adopts successively Consolidating for several reflection circuit-breaker switching on-off action status informations is extracted respectively with complementary population mean empirical mode decomposition algorithm There is mode function component；

4th step determines noise reduction vibration signal s respectively_v(t) with noise reduction voice signal s_a(t) respectively as signal s to be decomposed (t) when, the quantity Z of the intrinsic mode function component of required extraction characteristic quantity, the energy point according to each intrinsic mode function component Cloth determines Z values when normalized energy value adds up to be more than 90%；When s (t) is respectively s_v(t)、s_a(t) when, Z values are denoted as Z respectively₁ And Z₂,

5th step selects the preceding Z ranks intrinsic mode function component of signal s (t) to be decomposed to be handled, calculates its energy respectively Amount ratio, Sample Entropy and Power Spectral Entropy are simultaneously used as three category features；

6th step according to the first step to the 5th step, is obtained under different divide-shut brake action states, the preceding Z of vibration signal respectively₁ Energy ratio, Sample Entropy and the Power Spectral Entropy of rank intrinsic mode function component and the preceding Z of voice signal₂Rank intrinsic mode function component Energy ratio, Sample Entropy and Power Spectral Entropy, and will uniformly vibrate and form a feature samples with three category features of voice signal, and Using feature samples dimension-reduction treatment of the combination core core pivot element analysis method to three category feature of acoustic signal after reunification, M master is obtained Member, the core core pivot element analysis method that combines is the core pivot element analysis method (KPCA) using compound kernel function, the combination core Function is combined to obtain by local kernel function and global kernel function, and the part kernel function is Polynomial kernel function, the overall situation core Function is RBF kernel functions；

7th step, the Euclidean distance between selected pivot by calculating different divide-shut brake action states is come quantitative assessment Sample pivot average distance between class establishes the order binary tree multi-categorizer model based on Method Using Relevance Vector Machine, by Method Using Relevance Vector Machine Kernel functional parameter υ be denoted as particle to be optimized, using the quanta particle swarm optimization of converging diverging coefficient automatic adjusument to phase The kernel functional parameter υ closed in the disaggregated model of vector machine is optimized.

The omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features, described in second step Improved wavelet packet Threshold Filter Algorithms specifically include following steps：

1) basic function of Daubechies (dbN) wavelet basis as wavelet decomposition is selected, Decomposition order is 5 layers, to noisy Signal carries out WAVELET PACKET DECOMPOSITION, respectively obtains respective wavelet packet coefficient W_j,i, i.e., the wavelet packet system of i-th of frequency range on jth layer Number；

2) using threshold function table to wavelet packet coefficient W_j,iThreshold value quantizing processing is carried out, obtains that treated through threshold function table Wavelet packet coefficientWherein threshold function table is the threshold function table for being combined soft-threshold function with hard threshold function, construction New threshold function table is as follows：

In formula (1)For through threshold function table treated wavelet packet coefficient, wherein α_j,iFor on WAVELET PACKET DECOMPOSITION jth layer I-th of frequency range sub-belt energy normalization coefficient；W_j,iFor the wavelet packet coefficient before processing；Using the inspiration based on heursure Formula threshold value Selection of Function chooses threshold value λ,For adjustment factor,

Wavelet packet energy coefficient α in formula (1)_j,iIt can be by wavelet packet coefficient W_j,iIt acquires, specially：Signals and associated noises are through j layers After WAVELET PACKET DECOMPOSITION, wavelet packet coefficient be W (j, 0), W (j, 1) ..., W (j, 2^j- 1), the ENERGY E of each sub-band wavelet packet_j,_i =| | W (j, i) | |², total wavelet-packet energy isWavelet-packet energy coefficientIt will be through threshold value letter Number treated wavelet packet coefficient rebuilds wavelet packet tree, and inverse transformation reconstructs the signal after denoising, obtains noise reduction vibration signal s_v (t) with noise reduction voice signal s_a(t)。

The omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features, described in third step mutually Population mean empirical mode decomposition algorithm is mended to include the following steps：

1) the opposite white noise signal of symbol is added to signal s (t) to be decomposed in couples, forms two new signal s₁ (t), s₂(t)；

2) to echo signal s₁(t), s₂(t) empirical mode decomposition is carried out, is specifically included

2.1) echo signal s is determined₁(t) all maximum point and minimum point, then by all maximum points with one The smooth curve of item connects to obtain coenvelope line, and all minimum points are connected to obtain with another smooth curve Lower envelope line makes to include all signals between upper and lower envelope line.The average value of upper and lower envelope is denoted as m₁(t), mesh is obtained Mark signal s₁(t) one-component h₁(t), i.e. s₁(t)-m₁(t)=h₁(t) (2)

2.2) by h₁(t) as s₁(t) it brings into formula (2), repeats screening k times of step 2.1), obtain h_1k(t)= h_1(k-1)(t)-m_1k(t) so that h_1k(t) become an intrinsic mode function component.Remember c₁(t)=h_1k(t), then c₁(t) it is signal s₁(t) first component for meeting intrinsic mode function condition, i.e. intrinsic mode function component (IMF).

2.3) by c₁(t) from s₁(t) it is separated in, obtains r₁(t)=s₁(t)-c₁(t) (3)

By r₁(t) as s₁(t) step 2.1) obtains s with 2.2) before repeating₁(t) second meets intrinsic mode function The component c of condition₂(t)；

2.4) circulating repetition step 2.3) n times obtain signal s₁(t) n intrinsic function modal components, i.e.,It is in the end condition of the repetitive cycling of step 2.4), as component c_n(t) or residual r_n(t) It is sufficiently small so that as residual r_n(t) for a monotonic function cannot therefrom extract again meet intrinsic function modal components when；

The intrinsic mode function condition is that the extreme point in the entire time serieses of a. is at most differed with the quantity of zero crossing One, the mean value of upper and lower envelope that b. any moment is obtained by local maximum and local minimum is zero；

Echo signal s₂(t) also according to s₁(t) processing mode, according to step 2.1) to 2.4) to s₂(t) located Reason；

3) above-mentioned steps 1 are recycled)~2)；It repeats to the opposite pairs of white noise signal of signal s (t) to be decomposed addition symbol Then obtained new signal is subjected to empirical mode decomposition；Decomposition result is finally subjected to population mean operation, obtains decomposing knot Fruit such as formula (5), i.e.,In formula, s (t) is signal to be decomposed；c_j(t) (i=1,2 ..., n) be J-th of intrinsic mode function component (IMF), r_n(t) it is residual components.

The omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features, it is characterized in that the 4th The calculation formula of the energy of the i-th rank intrinsic mode function component is in stepFormula (6) Middle n is intrinsic mode function component exponent number of the signal to be decomposed through complementary population mean empirical mode decomposition, and N is each natural mode The data points of state function component；

The energy summation of each rank intrinsic mode function component isThe then intrinsic mode function component of the i-th rank Energy ratio be defined asWork as R₁+R₂+…+R_ZZ values when >=90%, as normalized energy value add up to be more than 90% Z values.

The omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features, it is characterized in that the 5th step The computational methods of the middle extraction intrinsic mode function component sample entropy are as follows：

1) it is c to remember i-th of intrinsic mode function component_i(t)={ c (n) }=c (1), c (2) ..., the N point datas of c (N) Sequence { c (n) } is formed the vector of m dimensions, C by sequence in order_m(1),…,C_m(N-m+1), i.e. C_m(i)={ c (i), c (i+ 1),…,c(i+m-1)},1≤i≤N-m+1(9)

2) definition vector C_m(i) and C_mThe distance between (j) d [C_m(i),C_m(j)] (i ≠ j) is two groups of vector corresponding elements The absolute value of middle maximum difference.I.e.

3) similar tolerance r (r ＞ 0) is given, to each i Data-Statistics d [C_m(i),C_m(j)] C of ＜ r_m(j)(1≤j≤N-m+ 1, j ≠ i) number, then calculate its ratio with total distance N-m, be denoted asI.e.

4) all average value B (m) (r) are calculated, i.e.,

5) vector is increased into dimension to m+1, repeats step 1)~3), thenAverage value such as formula (13) shown in, i.e.,When measured data N is finite value, the estimated value of Sample Entropy for SampEn (m, R, N)=- ln [B^(m+1)(r)/B^(m)(r)] (14), m be taken as 1 or 2, r take 0.1-0.25 times of SD；

The computational methods for extracting intrinsic mode function component power spectrum entropy are：

I-th of IMF component of signal s (t) to be decomposed is denoted as c respectively_i(t), c_i(t) discrete Fourier transform is C_i (w), the power spectrum for further deriving vibration signal and voice signal is：That is S_i(w)={ S_i (1),S_i(2),…,S_i(N) } Power Spectral Entropy H, is thus defined_i(f), i.e.,In formula (15), Subscript f represents frequency domain.q_i(w) it is w-th of power spectrum percentage in entire spectrum.

The omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features, combination described in the 6th step Core core pivot element analysis method is as described below：

If input training characteristics sample X_k(k=1,2 ..., N) it is mapped as φ (X_k), i.e.,：X_k→φ(X_k), after transformation Data meet：Training sample φ (X after then converting_k) total population scatter matrix C be

Characteristic equation λ V=CV (17) are solved, wherein V is feature vector, and V is by φ (X₁)、φ(X₂)…φ(X_N) represent I.e.Wherein α₁、α₂…α_NFor constant；

According to theory of reproducing kernel space, feature vector V mono- is positioned at by φ (X₁)、φ(X₂)…φ(X_N) composition feature space It is interior,

It defines a N × N and meets the kernel matrix K, K of Mercer conditions_ij=φ (X_i)^Tφ(X_j) (19), by formula (16), formula (18), formula (19), which substitute into, obtains N λ α=K α (20) in (17),

The problem of seeking formula (17) feature vector V in this way is converted to the feature vector α of solution formula (20).

One group of nonzero eigenvalue λ is obtained by solving characteristic equation (20)_iAnd corresponding meet normalizing condition λ_i(αi, α i)=1 feature vector αⁱ(i=1,2 ..., N ', N '≤N).

The projection principal component V on feature space is obtained according to formula (18)_j(j=1,2 ..., N), if X is a test sample, Then it is in V_jOn be projected asIt is asked in formula (21) Kernel function K (X are solved in solution preocess_i,X_j)=(φ (X_i)^Tφ(X_j)) definite expression formula method it is as follows：

Centralization as shown in formula (22), i.e. K '=K-K × L-L × K+L are carried out to the kernel matrix K obtained after transformation × K × L (22), wherein L are N ranks matrix and satisfaction

Global kernel function is Polynomial kernel function K₁(X_i,X_j)=[(X_i·X_j)+1]^q, local kernel function is RBF kernel functions K₂(X_i,X_j)=exp (- | | X_i-X_j||²/δ²), local kernel function and global kernel function are combined to obtain kernel function K (X_i,X_j)= (φ(X_i)^Tφ(X_j)) definite expression formula, that is, compound kernel function K (X_i,X_j)=θ K₁(X_i,X_j)+(1-θ)K₂(X_i,X_j) (23), In, θ is compound kernel function ratio, and q is Polynomial kernel function most high-order term number, and δ is RBF width parameters.

Finally define average inter-class separability parameter d_M, evaluate the validity that M pivot is extractedIn formula s be omnipotent breaker opening and closing state classification number, d_abFor The between class distance of a classes and b classes M pivot of state,WithM pivot is away from this class pivot center in respectively a, b class Maximum distance；

It is described using combination core core pivot element analysis method to three category feature of acoustic signal feature samples dimensionality reduction after reunification at The basic step of reason includes：

1) the feature samples collection X of three category feature of input noise reduction acoustic signal after reunification_k(k=1,2 ..., N)；

2) the kernel matrix K in combination core core pivot element analysis method is sought；

3) the centralization kernel matrix K ' of combination core core pivot element analysis method is sought；

4) eigenvalue λ and feature vector α of centralization kernel matrix K ' is calculated；

5) standardization feature vector：

6) individual features for calculating step 4) gained are worth descending arrangement；

7) it is shaken according to i characteristic value before accumulation contribution rate extraction and its corresponding preceding i feature vector, composing training noise reduction The principal component direction of the feature samples collection of three category feature of acoustical signal after reunification；

8) by training three category feature of noise reduction acoustic signal feature samples collection after reunification project to the 7) step acquire it is main into Divide on direction, extract the pivot of the feature samples collection of trained three category feature of noise reduction acoustic signal after reunification.

9) it three category feature of input test noise reduction acoustic signal feature samples after reunification and carries out corresponding kernel mapping and obtains Then the nuclear matrix is projected on the principal component direction that 7) step acquires, obtains accordingly testing noise reduction acoustic signal three by nuclear matrix The feature samples pivot of category feature after reunification, M contribution rate of accumulative total reaches 90% pivot before only choosing.

The omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features is set in the 7th stepFor the pivot of the feature samples collection of three category feature of noise reduction acoustic signal after reunification, as input vector, t=[t₁, t₂,…,t_N]^TFor object vector, then Method Using Relevance Vector Machine disaggregated model such as formula (25) is shown, i.e.,

In formula (25), w is weight vectors, w=[w₀,w₁,…,w_N]^T；K(x,x_i)=exp (- | | x-x_i||²/υ²) it is core letter Number, υ are kernel functional parameter, noise ε_n~N (0, σ 2)；

Shown in the likelihood function such as formula (26) of entire data set, i.e., In formula, Φ=[φ (x₁),φ(x₂),…,φ(x_N)]^TTo design square Battle array, φ (x_i)=[1, K (x_i,x₁),K(x_i,x₂),…,K(x_i,x_N)]^T；Work as desired valueWhen being 0 or 1, likelihood function such as formula (27) shown in, i.e.,In formula, δ () is sigmoid letters Number；

Using sparse Bayesian method shown in zero-mean gaussian prior distribution such as formula (28), i.e., weight vectors w is assignedIn formula (28), α ties up hyper parameter vector, α=(α for N+1₀,α₁,…,α_N)^T, N () For normal distyribution function；

For new input vector x_*, corresponding desired value t_*Probabilistic forecasting formula be p (t_*| t)=∫ p (t_*|w,α, σ²)p(w,α,σ²|t)dwdαdσ²(29), p (w, α, σ can be obtained by Bayes's derivation being carried out to formula (29)²| t)=p (w | t, α, σ²)p (α,σ²| t) (30) carry out formula (30) approximate processing, and the learning process of Method Using Relevance Vector Machine is maximizes p (α, σ²|t)∝p(t| α,σ²)p(α)p(σ²) process, that is, find α_MP、σ² _MP, meet

Using numerical method approximate solution α_MP、σ² _MP, α, σ can be obtained²Iteration more new formula such as formula γ_i=1- α_iΣ_i,i(34)；

In formula (34), Σ_i,iFor Σ=[σ^-2Φ^TΦ+diag(α₀,α₁,…,α_N)]^-1In i-th diagonal entry, μ_iFor Weight vectors μ=σ^-2ΣΦ^TI-th of element of t；After enough updates, most α_iInfinitely great, correspondence will be approached W_iIt is 0, and others α_iFinite value can then be leveled off to, corresponding x_iCollection be collectively referred to as associated vector, and then can obtain related The disaggregated model of vector machine, sample mean Euclidean distance is as separable measures using between class, by omnipotent breaker fault diagnosis This more classification problem is converted into multiple two classification problems, establishes the order binary tree multi-categorizer failure based on Method Using Relevance Vector Machine Diagnostic model；

For two class sample setsWithWherein x_i∈ A classes,Class, then sample between the class of A classes and B classes Average Euclidean distance is as shown in formula (35)In formula (35)It is 2 Euclidean distance between a different classes of sample.

The process of establishing of the order binary tree multi-categorizer model based on Method Using Relevance Vector Machine is to be acted with divide-shut brake Input vector of the preceding M pivot selected under different mechanical breakdown states and normal condition in journey as Method Using Relevance Vector Machine, with Omnipotent breaker is normal, falseness is closed a floodgate, separating brake is not thorough or certain mutually asynchronous divide-shut brake action state is Method Using Relevance Vector Machine Output, establish omnipotent breaker fault identification model successively, the order of Method Using Relevance Vector Machine is selected according to the size of Euclidean distance It is fixed；

Using converging diverging coefficient automatic adjusument quanta particle swarm optimization (hereinafter referred to as improving QPSO) to correlation to Kernel functional parameter υ in the disaggregated model of amount machine is optimized, and kernel functional parameter υ is denoted as to particle to be optimized；

The fitness function for improving QPSO rolls over cross validation discrimination function for K-, i.e.,Formula Middle x_lrAnd x_lwIt is by the sample size of correct and wrong classification, fitness respectively in the l times of Method Using Relevance Vector Machine verification sample set Value is bigger, represents that solution is more excellent；

The kernel functional parameter υ in two sorter model of order based on Method Using Relevance Vector Machine is optimized using QPSO is improved The specific steps are：

1) one group of particle, i.e. kernel functional parameter υ are initialized；

2) particle fitness value is calculated, obtains individual extreme value and group's extreme value；

3) particle position is updated according to the evolution equation for improving QPSO；

4) judge whether the iterations for reaching setting.It is to exit, group's extreme value at this time is required optimal core Function parameter；It is no, return to step 2)；

5) it exports optimal kernel functional parameter and corrects to two sorter model of order based on Method Using Relevance Vector Machine；

Calculating function of the discrimination that the cross validation of RVM is classified as particle fitness is selected in step 2), is calculated Step is as follows：

2.1) kernel functional parameter is given according to newer particle；

2.2) it uses the core pivot training of known divide-shut brake action state classification and establishes sorter model；

2.3) discrimination of the K- folding cross validations of two sorter model of order is calculated；

2.3) value is returned as the particle fitness for improving QPSO models.

The above-mentioned omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features, it is characterized in that described Method is realized by the method for Labview and Matlab hybrid programmings.

The above-mentioned omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features, the vibration and sound Signal captures the time identifier that can be used as effective vibration signal and voice signal by setting the threshold value of acquisition signal in real time, The electric signal of breaker closing coil can be given as effective vibration signal and the time identifier of voice signal.

It is a kind of to be used to perform the above-mentioned omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features Omnipotent breaker divide-shut brake fault detection system, the system comprises for fix the operation console of omnipotent breaker, accelerate Degree sensor, sound pick-up, solid-state relay group, industrial personal computer, I/O control cards and USB capture cards, the connection mode of each component are： Industrial personal computer is connect by isa bus with I/O control cards, and I/O control cards also connect and operate on it with solid-state relay group, For being fixed on omnipotent breaker, the sound pick-up is mounted on operation console acceleration transducer, acceleration transducer and Sound pick-up is connected by usb bus with USB capture cards, and the USB capture cards are connected with industrial personal computer, the solid-state relay Group include being respectively used to control breaker energy storage, combined floodgate, separating brake and under-voltage energy storage relay, closing relay, separating brake after Electric appliance and under-voltage relay.

The USB capture cards be USB7648A capture cards, the PCL720 control cards of the I/O control cards, the acceleration Sensor model number is LC0159, the model KZ520B of the sound pick-up.

The I/O control cards give electric signal in solid-state relay group energy storage relay, closing relay, separating brake after Electric appliance and under-voltage relay are operated；The analog signal of acceleration transducer and sound pick-up is converted by USB capture cards Digital signal, and pass through usb bus and send industrial personal computer to.

Beneficial effects of the present invention

Compared with prior art, the omnipotent breaker mechanical breakdown provided by the invention based on acoustic signal Fusion Features The substantive distinguishing features of the protrusion of diagnostic method are：First, wavelet packet soft and hard threshold is carried out respectively with voice signal to vibration to be combined Noise suppression preprocessing, and using complementary overall experience mode decomposition algorithm, to treated, acoustic signal decomposes, and extracts natural mode of vibration Function energy than, Sample Entropy, Power Spectral Entropy composition composite character parameter；Then, pass through compound kernel function core principle component analysis To more characteristic parameters dimensionality reduction and by its Fusion Features composition characteristic vector, solution single features identify circuit-breaker switching on-off failure Low accuracy rate and low stability；Finally, using improving quantum telepotation Method Using Relevance Vector Machine disaggregated model kernel functional parameter value, The model combined based on improvement quantum particle swarm with Method Using Relevance Vector Machine is established, and the feature vector for merging fault signature is input to Identification model carries out fault diagnosis.The experimental results showed that this method effectively improves the reliability of diagnostic result, and the side of detection Formula is simple, highly practical.

Compared with prior art, the omnipotent breaker mechanical breakdown provided by the invention based on acoustic signal Fusion Features The significant progress of diagnostic method is：

(1) the method for the present invention is believed with the vibration comprising abundant mechanical property information generated during circuit-breaker switching on-off Number with voice signal as signal source, while acceleration transducer is easy for installation and in the situation for not destroying breaker body Under, the installation of sound pick-up is even more contactless, it is achieved that omnipotent breaker non-intruding monitor and fault diagnosis.

(2) the method for the present invention for vibration signal and voice signal nonlinear and nonstationary the characteristics of, using improved small Wave packet Denoising Algorithm, preferably improves Signal-to-Noise, reduces root mean square mistake the advantages of taking into account soft, hard threshold method noise reduction Difference.

(3) complementary population mean empirical mode decomposition algorithm is reduced due to adding white noise used by the method for the present invention Caused reconstructed error has obtained better mode decomposition effect.

(4) the method for the present invention is extracted three kinds of IMF component samples entropy, energy ratio, Power Spectral Entropy characteristic synthetics and is utilized respectively, and With compound kernel function core principle component analysis dimensionality reduction, the effective integration of fault characteristic information is realized, overcome and only rely on a certain spy There are the problems such as accuracy rate is low, reliability and stability are poor for sign.

(5) the order binary tree fault diagnosis model based on Method Using Relevance Vector Machine used by the method for the present invention is introduced and is improved The kernel functional parameter of quantum telepotation diagnostic model, non-artificial experience value, the quick global optimizing for realizing kernel function carry The discrimination and reliability of high score class model.

Description of the drawings

The present invention is further described with the present embodiment below in conjunction with the accompanying drawings.

Fig. 1 is omnipotent breaker divide-shut brake failure acoustic signal detecting system hardware architecture diagram of the present invention；

Fig. 2 is vibration signal and its corresponding spectrogram under breaker difference divide-shut brake action state in embodiment 1；

Fig. 3 is voice signal and its corresponding spectrogram under breaker difference divide-shut brake action state in embodiment 1；

Fig. 4 is vibration and the normalization energy of IMF components obtained by audio-signal resolution under A phases in embodiment 1 not same period state Measure block diagram；

Fig. 5 is the main IMF components of vibration signal and its corresponding spectrogram under A phases in embodiment 1 not same period state；

Fig. 6 is the main IMF components of voice signal and its corresponding spectrogram under A phases in embodiment 1 not same period state；

Fig. 7 is the pivot contribution rate of accumulative total tendency chart of the lower three kinds of pca methods of normal condition in embodiment 1；

Fig. 8 is the pivot separability Parameters variation comparison diagram of three kinds of pca methods in embodiment 1；

Fig. 9 is the order Binary tree classifier structure diagram based on Method Using Relevance Vector Machine in embodiment 1.

Figure 10 is that the fitness curve graph that QPSO carries out RVM models parameter optimization is improved in embodiment 1.

Specific embodiment

The present invention provides a kind of omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features, described Acoustic signal by the collected fuselage shaking signal of acceleration transducer and leads in the process for omnipotent breaker divide-shut brake It crosses sound pick-up and collects fuselage voice signal, include the following steps

The first step acquires fuselage shaking signal and fuselage voice signal in omnipotent breaker divide-shut brake action process, And collected analog signal is separately converted to digital signal, obtain initial vibration s_v' (t) and voice signal s_a' (t), t For divide-shut brake actuation time.

Second step, using improved wavelet packet Threshold Filter Algorithms respectively by collected vibration signal s_v' (t) and sound Signal s_a' (t) as signals and associated noises carry out denoising, obtain noise reduction vibration signal s_v(t) and noise reduction voice signal s_a(t)。

The improved wavelet packet Threshold Filter Algorithms specifically include following steps：

In formula (1)For through threshold function table treated wavelet packet coefficient, wherein α_j,iFor on WAVELET PACKET DECOMPOSITION jth layer I-th of frequency range sub-belt energy normalization coefficient；W_j,iFor the wavelet packet coefficient before processing；Using the inspiration based on heursure Formula threshold value Selection of Function chooses threshold value λ,For adjustment factor,Basic thought is the higher frequency band letter of energy It makes an uproar than high, is more likely to hard -threshold processing；The smaller frequency band signal-to-noise ratio of energy is low, is more likely to soft-threshold processing.

Wavelet packet energy coefficient α in formula (1)_j,iIt can be by wavelet packet coefficient W_j,iIt acquires, specially：Signals and associated noises are through j layers After WAVELET PACKET DECOMPOSITION, wavelet packet coefficient be W (j, 0), W (j, 1) ..., W (j, 2^j- 1), the ENERGY E of each sub-band wavelet packet_j,i =| | W (j, i) | |², total wavelet-packet energy isWavelet-packet energy coefficient

Wavelet packet tree will be finally rebuild through threshold function table treated wavelet packet coefficient, and inverse transformation reconstructs the letter after denoising Number, obtain noise reduction vibration signal s_v(t) with noise reduction voice signal s_a(t)。

Third walks, by noise reduction vibration signal s_v(t) with voice signal s_a(t) it respectively as signal s (t) to be decomposed, adopts successively Consolidating for several reflection circuit-breaker switching on-off action status informations is extracted respectively with complementary population mean empirical mode decomposition algorithm There is mode function (intrinsic mode function, IMF) component.

The complementation population mean empirical mode decomposition algorithm includes the following steps：

2.2) by h₁(t) it is brought into formula (2) as s1 (t), repeats screening k times of step 2.1), obtain h_1k(t)= h_1(k-1)(t)-m_1k(t) so that h_1k(t) become an intrinsic mode function component.Remember c₁(t)=h_1k(t), then c₁(t) it is signal s₁(t) first component for meeting intrinsic mode function condition, i.e. intrinsic mode function component (IMF).

By r₁(t) step 2.1) obtains s with 2.2) before being repeated as s1 (t)₁(t) second meets intrinsic mode function The component c of condition₂(t)；

2.4) circulating repetition step 2.3) n times obtain signal s₁(t) n intrinsic function modal components, i.e.,It is in the end condition of the repetitive cycling of step 2.4), as component c_n(t) or residual r_n(t) It is sufficiently small so that as residual r_n(t) for a monotonic function cannot therefrom extract again meet intrinsic function modal components when.

Decompose the condition (i.e. intrinsic mode function condition) that obtained intrinsic mode function component needs meet：A. when entire Between extreme point in sequence at most differ one with the quantity of zero crossing.B. any moment passes through local maximum and local minimum The mean value of envelope for being worth and obtaining is zero.c₁(t) most fine in signal or shortest component should be included.

Echo signal s₂(t) also according to s₁(t) processing mode, according to step 2.1) to 2.4) to s₂(t) it is handled

4th step determines noise reduction vibration signal s respectively_v(t) with noise reduction voice signal s_a(t) respectively as signal s to be decomposed (t) when, the quantity Z of the intrinsic mode function component of required extraction characteristic quantity, the energy point according to each intrinsic mode function component Cloth feature determines Z values when normalized energy value adds up to be more than 90%, when s (t) is respectively s_v(t)、s_a(t) when, Z values are remembered respectively It is Z₁And Z₂,

The calculation formula of the energy of i-th rank intrinsic mode function component is

N is intrinsic mode function component exponent number of the signal to be decomposed through complementary population mean empirical mode decomposition in formula (6), N is that the data of each intrinsic mode function component are counted；

The energy summation of each rank intrinsic mode function component isThe then intrinsic mode function component of the i-th rank Energy ratio be defined asWork as R₁+R₂+…+R_ZZ values when >=90%, as normalized energy value add up to be more than 90% Z values；When s (t) is respectively s_v(t)、s_a(t) when, Z values are denoted as Z respectively₁And Z₂

5th step selects the preceding Z ranks intrinsic mode function component of signal s (t) to be decomposed to be handled, calculates its energy respectively Amount ratio, Sample Entropy and Power Spectral Entropy are simultaneously used as three category features, and the energy ratio is included vibration signal s_v(t) it decomposes Preceding Z₁The energy ratio of rank intrinsic mode function component and by voice signal s_a(t) the preceding Z decomposed₂Rank intrinsic mode function The energy ratio of component

The computational methods for extracting intrinsic mode function component sample entropy are as follows：

4) all average value B are calculated^(m)(r), i.e.,

5) vector is increased into dimension to m+1, repeats step 1)~3), thenAverage value such as formula (13) shown in, i.e.,When measured data N is finite value, the estimated value of Sample Entropy for SampEn (m, R, N)=- ln [B^(m+1)(r)/B^(m)(r)](14).M be taken as 1 or 2, r take 0.1-0.25 times of SD.

The value of Sample Entropy and the value of m, r are related, but Sample Entropy has good consistency, entropy increase and reduction Trend is not influenced by m and r, under normal circumstances when, SD be data sequence standard deviation, calculating obtained Sample Entropy has More rational statistical nature.

Power Spectral Entropy is that the different IMF components spectral patterns of acoustic signal are characterized obtained from the frequency domain character of signal extracts Structure situation.

The computational methods of extraction intrinsic mode function component power spectrum entropy are：

I-th of IMF component of signal s (t) to be decomposed is denoted as c respectively_i(t), c_i(t) discrete Fourier transform is C_i (w), the power spectrum for further deriving vibration signal and voice signal is：That is S_i(w)={ S_i (1),S_i(2),…,S_i(N) } Power Spectral Entropy H, is thus defined_i(f), i.e.,Formula (15) In, subscript f represents frequency domain.q_i(w) it is w-th of power spectrum percentage in entire spectrum.

6th step according to the first step to the 5th step, is obtained under different divide-shut brake action states, the preceding Z of vibration signal respectively₁ Energy ratio, Sample Entropy and the Power Spectral Entropy of rank intrinsic mode function component and the preceding Z of voice signal₂Rank intrinsic mode function component Energy ratio, Sample Entropy and Power Spectral Entropy, and will uniformly vibrate and form a feature samples with three category features of voice signal, and Using feature samples dimension-reduction treatment of the combination core core pivot element analysis method to three category feature of acoustic signal after reunification, M master is obtained Member, the combination core core pivot element analysis method is the core pivot element analysis method (combination core KPCA) using compound kernel function, described Compound kernel function is combined to obtain by local kernel function and global kernel function, and the part kernel function is Polynomial kernel function, described Global kernel function is RBF kernel functions.

The combination core core pivot element analysis method is as described below：

Characteristic equation λ V=CV (17) are solved, wherein V is feature vector, and V is by φ (X₁)、φ(X₂)…φ(X_N) represent I.e.Wherein α₁、α₂…α_NFor constant.

One group of nonzero eigenvalue λ is obtained by solving characteristic equation (20)_iAnd corresponding meet normalizing condition λ_i(αⁱ, αⁱThe feature vector α of)=1ⁱ(i=1,2 ..., N ', N '≤N).

The projection principal component V on feature space is obtained according to formula (18)_j(j=1,2 ..., N), if X is a test sample, Then it is in V_jOn be projected asIt is asked in formula (21) It solved

Kernel function K (X are solved in journey_i,X_j)=(φ (X_i)^Tφ(X_j)) definite expression formula method it is as follows：

Finally define average inter-class separability parameter d_M, evaluate the validity that M pivot is extractedIn formula s be omnipotent breaker opening and closing state classification number, d_abFor The between class distance of a classes and b classes M pivot of state,WithM pivot is away from this class pivot center in respectively a, b class Maximum distance.

5) standardization feature vector：

7th step, the Euclidean distance between selected pivot by calculating different divide-shut brake action states is come quantitative assessment Sample pivot average distance between class establishes the order two based on Method Using Relevance Vector Machine (relevance vector machine, RVM) Fork tree multi-categorizer model, that is, setFor the pivot of the feature samples collection of three category feature of noise reduction acoustic signal after reunification, by it As input vector, t=[t₁,t₂,…,t_N]^TFor object vector, then Method Using Relevance Vector Machine disaggregated model such as formula (25) is shown, i.e.,

In formula (25), w is weight vectors, w=[w₀,w₁,…,w_N]^T；K(x,x_i)=exp (- x-x_i ²υ²) for kernel function, υ For kernel functional parameter, noise ε_n~N (0, σ²)；

Shown in entire data set likelihood function such as formula (26), i.e., In formula, Φ=[φ (x₁),φ(x₂),…,φ(x_N)]^TTo design square Battle array, φ (x_i)=[1, K (x_i,x₁),K(x_i,x₂),…,K(x_i,x_N)]^T。

Work as desired valueWhen being 0 or 1, shown in likelihood function such as formula (27), i.e.,In formula, δ () is sigmoid functions.

Method Using Relevance Vector Machine order Two Binomial Tree Model establishes process：With mechanical breakdown states different in divide-shut brake action process with And input vector of the preceding M pivot as Method Using Relevance Vector Machine selected under normal condition, with omnipotent breaker is normal, falseness conjunction The output that lock, separating brake are not thorough or certain mutually asynchronous divide-shut brake action state is Method Using Relevance Vector Machine, establishes universal and breaks successively Road device fault identification model, the order of Method Using Relevance Vector Machine are selected according to the size of Euclidean distance.First Method Using Relevance Vector Machine is first Identify normal condition (Euclidean distance is big), then training sample is respectively normal characteristics data and remaining all fault samples；The Two Method Using Relevance Vector Machine identifications are false to close a floodgate (Euclidean distance takes second place), and first vector machine has been distinguished normally, so the Two Method Using Relevance Vector Machines do not have the sample of normal condition, thus training sample be respectively it is false close a floodgate and remove normally with falseness The remaining fault sample of combined floodgate；And so on.

Using converging diverging coefficient automatic adjusument quanta particle swarm optimization (hereinafter referred to as improving QPSO) to correlation to Kernel functional parameter υ in the disaggregated model of amount machine is optimized, i.e.,：Kernel functional parameter υ is denoted as to particle to be optimized.

The evolution side of the quanta particle swarm optimization (quantum particle swarm optimization, QPSO) Cheng WeiIn formula ForAnd G_kBetween a random order It puts；The optimal location of particle i when being kth time iteration, i=1,2 ..., N, N is population scale；G_kDuring for kth time evolution iteration The global optimum position of population；C_kThe mean value of personal best particle during for all particle kth time iteration, i.e.,u、Equally distributed random number is obeyed between (0,1)；β is converging diverging coefficient, β by from 1 it is linear be decreased to 0.5 in the way of Value introduces evolution velocity factor and the particle buildup degree factor, adaptive in real time according to the evolutionary rate of population and concentration class Converging diverging coefficient should be adjusted, is specifically included,

If F (G_k)、F(G_k-1) be respectively group currently and previous generation adaptive optimal control angle value, introducing evolution velocity factorIf F_avgIt is the fitness average of all particle current individual optimal locations, i.e.,Introduce the concentration class factorWrite β as the function about h and s, i.e. β =f (h, s)=β₀-η_hh+η_sS (39), β in formula₀For the initial value of β, usual value is 1；η_hIt is to adjust the evolutionary rate factor to be Number, η_sIt is the coefficient for adjusting the concentration class factor, usually takes η_h=0.5, η_s=0.2；

The fitness function of improvement QPSO in optimization process rolls over cross validation discrimination function for K-, i.e.,X in formula_lrAnd x_lwIt is by correct and mistake respectively in the l times of Method Using Relevance Vector Machine verification sample set The sample size of classification, fitness value is bigger, represents that solution is more excellent.

Using the quanta particle swarm optimization (hereinafter referred to as improving QPSO) using converging diverging coefficient automatic adjusument to base Kernel functional parameter υ in two sorter model of order of Method Using Relevance Vector Machine optimize the specific steps are：

1) initialize one group of particle (particle is kernel functional parameter υ)；

5) it exports optimal kernel functional parameter and corrects to two sorter model of order based on Method Using Relevance Vector Machine.

2.1) kernel functional parameter is given according to newer particle；

2.3) discrimination of the K folding cross validations of two sorter model of order is calculated；

2.3) value is returned as the particle fitness for improving QPSO models.

It is omnipotent including being used to fix for performing the omnipotent breaker divide-shut brake fault detection system of above-mentioned diagnostic method Operation console 1, acceleration transducer 2, sound pick-up 3, solid-state relay group 4, industrial personal computer 5, I/O control cards 6 and the USB of formula breaker Capture card 7, the connection mode of each component are：Industrial personal computer is connect by isa bus with I/O control cards, I/O control cards also with solid-state Relay group is connected and is operated on it, and acceleration transducer for being fixed on omnipotent breaker, pacify by the sound pick-up On operation console, acceleration transducer and sound pick-up are connected by usb bus with USB capture cards, the USB capture cards with Industrial personal computer is connected, and the solid-state relay group includes being respectively used to control breaker energy storage, combined floodgate, separating brake and under-voltage storage It can relay 4.1, closing relay 4.2, separating brake relay 4.3 and under-voltage relay 4.4.

The I/O control cards give electric signal in solid-state relay group energy storage relay, closing relay, separating brake after Electric appliance and under-voltage relay are operated；The analog signal of acceleration transducer and sound pick-up is converted by USB capture cards Digital signal, and pass through usb bus and send industrial personal computer to.The hardware knot of the omnipotent breaker divide-shut brake fault detection system Structure schematic diagram is as shown in Figure 1

Embodiment 1

Using DW15 Series Air Circuit Breakers DW15-1600 as experimental subjects.In the fault set of circuit-breaker switching on-off state It is more than regulation to show actuation time, can open the machinery away from the typical fault with excess of stroke simulation divide-shut brake by adjusting between contact State.Adjust the false "on" position of cantilever simulation of contact system；Add gasket between the buckle in place of separating brake, simulation separating brake is not Thorough state；The connecting rod length simulation of adjustment three-phase contact a certain phase in three-phase caused by mechanism wear or adjustment are improper respectively Contact acts asynchronous, i.e., the single-phase not same period state in A, B, C three-phase with another two-phase.Acoustic signal feature is based on using aforementioned The omnipotent breaker divide-shut brake method for diagnosing faults of fusion and the omnipotent breaker divide-shut brake fault detect for performing this method The omnipotent breaker that system simulates this each typical fault machine performance carries out fault diagnosis.

By omnipotent breaker divide-shut brake failure acoustic signal detecting system, open circuit is acquired with the sample frequency of 20kHz Vibration and voice signal in device action process, the Typical Vibration signal under 6 kinds of divide-shut brake action states as shown in Figures 2 and 3, Respectively normal condition, the false full excess of stroke 5mm of "on" position, that is, contact, separating brake are not thorough state i.e. spacer thickness 2mm, and A phases are not Same period state, that is, A phases are opened away from being opened with another two-phase away from differing 3mm, B, C phase not the same period with A phases similarly.

In the present embodiment, adjustment factorIt is 0.9, chooses " db25 " wavelet basis carried in Matlab softwares and carry out 5 layers points Solve noise reduction process.

Third walks, noise reduction vibration signal s_v(t) with voice signal s_a(t) it respectively as signal s (t) to be decomposed, uses successively Complementary population mean empirical mode decomposition algorithm distinguishes the noise reduction vibration signal under 6 kinds of divide-shut brake action state states and noise reduction Extract several reflection circuit-breaker switching on-off action status informations intrinsic mode function (intrinsic mode function, IMF) component.

The amplitude of wherein white noise takes 0.2 times of original signal standard deviation, and population mean number takes 500, at this time resolution error Less than 0.01, resolution error is within the acceptable range.

By noise reduction vibration with audio-signal resolution after before 13 rank intrinsic mode function components carry out energy spectrometer, with A phases Not for same period state, Fig. 4 is the A phases not normalized energy block diagram of same order IMF components under same period state.It can be with by Fig. 4 The energy for intuitively going out sound and vibration signal is concentrated mainly on preceding 7 rank mode, and the energy of the 8th rank intrinsic mode function component Very little, foundation normalized energy value add up the R more than 90% to amount_zValue is R₁+R₂+…+R_z>=90%, Z₁=Z₂=7, it chooses Preceding 7 rank intrinsic mode function Component Analysis.

Further respectively by noise reduction vibration with audio-signal resolution after before 8 rank intrinsic mode function components respectively into line frequency Spectrum analysis, equally by taking A the phases not same period as an example, the main IMF components of circuit-breaker switching on-off vibration signal and its corresponding frequency spectrum such as Fig. 5 Shown, the main IMF components of circuit-breaker switching on-off vibration signal and its corresponding frequency spectrum are as shown in Figure 6.It is obtained by spectrum analysis, the 8th Rank intrinsic mode function component frequencies are low, to effect of signals very little.Therefore, further verification takes preceding 7 rank intrinsic mode function point Amount analysis can reflect the main divide-shut brake action status information of breaker.

5th step, respectively selection vibration handled with the preceding 7 rank IMF components in voice signal, calculate its energy coefficient, Sample Entropy, Power Spectral Entropy, obtained characteristic parameter are as shown in table 1

Table 1 vibrates and voice signal property parameter value

6th step according to the first step to the 5th step, is obtained under different divide-shut brake action states respectively, vibration and voice signal Energy coefficient, Sample Entropy and the Power Spectral Entropy of preceding 7 rank intrinsic mode function component, and uniformly will vibration and voice signal three classes spy Sign one feature samples of composition, and using feature sample of the combination core core pivot element analysis method to three category feature of acoustic signal after reunification This dimension-reduction treatment, removal are associated with and retain global information and local feature, obtain M pivot.

Using vibration and the measurement method of sound, while extract the characteristic parameter of the three classes different angle of IMF components.Although The randomness of single detection mode and single features parameter and low accuracy rate are avoided, but characteristic parameter is caused to be multiplied, The complexity of feature space and the input dimension of subsequent classifier are excessively high, cause the real-time of fault identification and stability still compared with Difference.The dimension of the characteristic parameter for sound both detection modes of shaking is reduced using core KPCA is combined, redundancy is can remove and realizes The effective integration of feature.Wherein, combination core KPCA methods setting kernel function portfolio ratio, most high-order term number, RBF width parameters (q=3, δ=0.5).First by the breaker vibration under each opening and closing state and 3 types of voice signal totally 42 features Parameter chooses 40 groups of characteristic parameters of normal condition (every group of 42 characteristic parameters), then by every group of spy as primitive character parameter Sign parameter is divided into 3 pieces according to energy coefficient, Sample Entropy, Power Spectral Entropy, i.e., every piece contains 14 characteristic parameters, carries out z- respectively The processing of score standardized datas, treated data fit standardized normal distribution, i.e. mean value are 0, standard deviation 1, so as to make Different magnitude of index can carry out unified dimension-reduction treatment removal association, while ensure that the local influence of each type feature is special Property.In order to compare the effect of put forward pivot, herein also using more common PCA and monokaryon KPCA (RBF kernel functions, width ginseng Number δ=0.5) method progress pivot extraction, the lower three kinds of pivot extracting methods parameters obtained comparison of normal condition is as shown in table 2, and Contribution rate of accumulative total trend is provided by Fig. 7.Simultaneously using 5 kinds of malfunctions, each 20 groups of each failure, 100 groups of characteristic parameters altogether Data extract pivot, and calculate average inter-class separability parameter, as pivot number increases, separability parameter value variation such as Fig. 8 It is shown.

The parameter of 2 three kinds of pca methods of table

It can be seen that with reference to table 2 and Fig. 7, the contribution rate of accumulative total (99.59%) of preceding 7 Non-linear Kernel pivots of combination core KPCA Although the contribution rate of accumulative total (98.80%) of the Non-linear Kernel pivot higher than monokaryon KPCA, the linear pivot less than PCA adds up Contribution rate (100%).Pivot shows greatly characteristic parameter contribution data rate and its its entrained variation information positive correlation, contribution rate It is strong to the interpretability of data, and integration capability is good.Although the contribution rate of accumulative total of 7 pivots has all reached 100% before PCA, It is undesirable for the differentiation effect between different faults class, can be seen that by the separability parameter value variation curve of Fig. 8, PCA's The Non-linear Kernel pivot of first 7 linear pivots and single kernel function can significantly lower than the preceding 7 core pivots for combining core KPCA Divide property, reason is that carried characteristic parameter does not have linear feature, and obtained by PCA is at this time linear pivot, uncomfortable It closes, exaggerates influence of certain characteristic parameters to working condition instead；Although the local learning ability of RBF kernel functions in monokaryon KPCA It is very strong, it is usually used in prominent different classes of local feature, but in the case of this fault sample of breaker is less, it is also desirable to The overall situation considers extraction global feature.Therefore, the smaller Polynomial kernel function of weight and RBF kernel functions are combined into organic whole, The generalization ability of kernel function can be improved, with ensure pivot extraction when can retention data entirety global characteristics and protrusion not Generic local feature, so as to obtain more useful characteristic of division information in the case of seldom data sample study.

Since preceding 7 core pivots of combination core KPCA extractions remain the characteristic information of former more than 95% data and can divide Property it is good, therefore, other M=7, i.e., with 42 original characteristic parameters of preceding 7 core pivots substitution, i.e. intrinsic dimensionality reduces 35 It is a.

To combine preceding 7 core pivots that core KPCA models obtain as feature vector, it is empty to form new feature for 7th step Between, for the training and identification of RVM.According to the Euclidean distance between different classes of core pivot sample, Gaussian kernel letter is selected herein Number establishes the order Binary tree classifier model based on RVM as shown in Figure 9.The process of foundation is：With in divide-shut brake action process Input vector of preceding 7 pivots selected under different mechanical breakdown states and normal condition as Method Using Relevance Vector Machine, with omnipotent Formula breaker is normal, falseness is closed a floodgate, separating brake is not thorough or certain mutually asynchronous divide-shut brake action state is the defeated of Method Using Relevance Vector Machine Go out, establish omnipotent breaker fault identification model successively, the order of Method Using Relevance Vector Machine is selected according to the size of Euclidean distance.The One Method Using Relevance Vector Machine identifies normal condition (Euclidean distance is big) first, then training sample is respectively normal characteristics data and remains Remaining all fault samples；Second Method Using Relevance Vector Machine identification is false to close a floodgate (Euclidean distance takes second place), and first vector machine is It distinguishes normally, so second Method Using Relevance Vector Machine does not have the sample of normal condition, so training sample is respectively false closes The remaining fault sample that lock and removing are normally closed a floodgate with falseness；And so on.

Due to including 5 graders in the order Binary tree classifier model based on RVM, it is therefore desirable to using improvement QPSO The kernel functional parameter υ of 5 separators is optimized.Setting improves the population scale N=20 of QPSO first, and particle dimension is 5, Maximum iteration is 100, and the initial value of converging diverging factor beta then is set as 1.0, adjusts the evolutionary rate factor and gathers The coefficient of the intensity factor is respectively set to 0.5 and 0.2.Finally using K- folding cross validation mode (K=10), will be used for training and The feature samples collection of optimization is randomly divided into 10 groups of data, and the sample number of every group of data is roughly equal, chooses 9 groups of works in turn in order For training set, one group is used as test set, and the average cross verification discrimination of 10 times, sample each in this way are asked for after 10 iteration It is verified once, so as to improve the error that the reliability of algorithm reduces parameter.

In the present embodiment, normal opening and closing state and 5 kinds of common machinery events are simulated on DW-15 omnipotent breakers Barrier, 6 kinds of divide-shut brake action states in total, respectively normally, it is false close a floodgate, separating brake is not thorough, the A phases not same period, B the phases not same period, C Mutually not same period does 25 groups under each state, totally 150 groups of experiments and extracts characteristic.To avoid data set skew problems, choose Wherein 120 groups, each 20 groups of each state, for model training and parameter optimization, remaining 30 groups of data are used for the test of model. Figure 10 is to improve the fitness convergence curve that QPSO carries out RVM models parameter optimization, and wherein average fitness is all particles Average fitness value in each generation, optimal adaptation write music line as the maximum adaptation angle value of all particles in each generation.

Fitness value curve in Figure 10 rises with evolution number, and stops evolving at the 35th time, demonstrates in model The optimization performance of influence and improvement QPSO of the nuclear parameter to classification accuracy.Meanwhile (change to test this paper optimization algorithms Into QPSO-RVM models) validity, respectively with RVM, PSO-RVM, QPSO-RVM compare.The correlation of wherein three kinds optimization methods Initiation parameter is consistent with verification mode, and the inertia weight factor reduces rule using linear in PSO, i.e. initial value is maximized It is 0.9, it is 0.2 that when termination, which is minimized, and Studying factors are 2.0；Converging diverging coefficient reduces method using linear in QPSO Then, i.e., 1.0 are initially taken as, when termination is taken as 0.5.The diagnostic result of Different Optimization fault identification model is as shown in table 3, is arranged in table Go out nuclear parameter and the final gained nuclear parameter of Different Optimization algorithm, corresponding discrimination without optimization algorithm to be respectively used for The discrimination of the 120 groups of samples and remaining 30 groups of test sample collections of training and optimization.Pass through the knowledge of 3 Different Optimization model of contrast table Rate can not obtain, improve QPSO and nuclear parameter is optimized, under less evolution number, can effectively improve fault diagnosis Accuracy.Fault Pattern Recognition model finally is established using kernel functional parameter obtained by optimizing, realizes the diagnosis to circuit breaker failure.

3 Different Optimization disaggregated model of table is to the diagnostic result of test data

Embodiment 2

Text is tested for omnipotent breaker divide-shut brake failure, repeatedly adjustment breaker simulation different faults state, Each malfunction under different condition does 30 groups of experiments respectively, and the specific amplitude of accommodation is shown in Table 4, is followed successively by and adjusts the full excess of stroke of contact The false "on" position of simulation, spacer thickness simulation separating brake are not thorough state, and A phases are opened and simulate A away from differ size away from being opened with another two-phase Mutually not same period state, B, C phase not the same period with A phases similarly.

The adjustment parameter of 4 mimic-disconnecting switch failure of table

Under three kinds of different conditions of table 4, using identical order binary tree RVM models, 20 groups of feature samples of each state This for training, 10 groups for testing, by context of methods and frequently with single signal detection and single features extracting method into Row comparison, the wherein multiple features under single signal detection mode equally select energy coefficient, Sample Entropy and Power Spectral Entropy, and single The Sample Entropy of one feature selecting better performances, the experimental result of different diagnostic methods are as shown in table 5.

The accuracy comparison of the different diagnostic methods of table 5

It can be obtained by table 5, under the same terms, the diagnostic result of vibration detection is substantially better than sound detection, and identical inspection Under survey mode, multi-feature extraction is better than single features extracting method；Although vibrate the multiple features fusion method that is combined with sound with The harsh accuracy for condition is declined slightly, but the accuracy under different condition more than 90% and fluctuates smaller.Reason It is：Based on the characteristic for sound detection of shaking, the multi-source identification of energy coefficient, Sample Entropy, Power Spectral Entropy different characteristic has been merged Information avoids single-measurement mode from the influence of contingency easily occur with feature extraction, so as to further improve the accuracy rate of identification With stability.

Above-mentioned steps are realized using software Labview and Matlab.

Software Labview and Matlab are known to those skilled in the art used in above-mentioned the present embodiment 's.

Percentage in examples detailed above is numerical percentage.

Vibration signal is measured using acceleration transducer LC0159 in above-described embodiment, mounted among breaker On the pedestal crossbeam of phase contact

Voice signal is measured using KZ502B sound pick-ups in above-described embodiment, contactless to be installed on 20cm on rear side of breaker Place.

Claims

1. the omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features, the acoustic signal is universal During circuit-breaker switching on-off machine is collected by the collected fuselage shaking signal of acceleration transducer and by sound pick-up Body voice signal, it is characterized in that including the following steps：

The first step acquires fuselage shaking signal and fuselage voice signal in omnipotent breaker divide-shut brake action process, and will Collected analog signal is separately converted to digital signal, obtains initial vibration signal s_v' (t) and voice signal s_a' (t), t For divide-shut brake actuation time；

Second step, using improved wavelet packet Threshold Filter Algorithms respectively by collected vibration signal s_v' (t) and voice signal s_a' (t) as signals and associated noises carry out denoising, obtain noise reduction vibration signal s_v(t) and noise reduction voice signal s_a(t)；

Third walks, by noise reduction vibration signal s_v(t) with noise reduction voice signal s_a(t) it respectively as signal s (t) to be decomposed, adopts successively Consolidating for several reflection circuit-breaker switching on-off action status informations is extracted respectively with complementary population mean empirical mode decomposition algorithm There is mode function component；

4th step determines noise reduction vibration signal s respectively_v(t) with noise reduction voice signal s_a(t) respectively as signal s (t) to be decomposed When, the quantity Z of the intrinsic mode function component of required extraction characteristic quantity, according to the Energy distribution of each intrinsic mode function component, Determine Z values when normalized energy value adds up to be more than 90%；When s (t) is respectively s_v(t)、s_a(t) when, Z values are denoted as Z respectively₁With Z₂,

5th step selects the preceding Z ranks intrinsic mode function component of signal s (t) to be decomposed to be handled, calculates its energy respectively Than, Sample Entropy and Power Spectral Entropy and it is used as three category features；

6th step according to the first step to the 5th step, is obtained under different divide-shut brake action states, the preceding Z of vibration signal respectively₁Rank is consolidated There are energy ratio, Sample Entropy and the Power Spectral Entropy of mode function component and the preceding Z of voice signal₂The energy of rank intrinsic mode function component Amount ratio, Sample Entropy and Power Spectral Entropy, and will uniformly vibrate and form a feature samples with three category features of voice signal, and use Feature samples dimension-reduction treatment of the core core pivot element analysis method to three category feature of acoustic signal after reunification is combined, obtains M pivot, institute It is the core pivot element analysis method using compound kernel function to state combination core core pivot element analysis method, and the compound kernel function is by karyomerite Function and global kernel function combine to obtain, and the part kernel function is Polynomial kernel function, and the overall situation kernel function is RBF core letters Number；

7th step, the Euclidean distance between selected pivot by calculating different divide-shut brake action states is come between quantitative assessment class Sample pivot average distance establishes the order binary tree multi-categorizer model based on Method Using Relevance Vector Machine, by the core of Method Using Relevance Vector Machine Function parameter υ is denoted as particle to be optimized, using converging diverging coefficient automatic adjusument quanta particle swarm optimization to correlation to Kernel functional parameter υ in the disaggregated model of amount machine is optimized.

2. the omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features as described in claim 1, It is characterized in that improved wavelet packet Threshold Filter Algorithms described in second step specifically include following steps：

1) basic function of Daubechies (dbN) wavelet basis as wavelet decomposition is selected, Decomposition order is 5 layers, to signals and associated noises WAVELET PACKET DECOMPOSITION is carried out, respectively obtains respective wavelet packet coefficient W_j,i, i.e., the wavelet packet coefficient of i-th of frequency range on jth layer；

2) using threshold function table to wavelet packet coefficient W_j,iThreshold value quantizing processing is carried out, is obtained through threshold function table treated small echo Packet coefficientWherein threshold function table is the threshold function table for being combined soft-threshold function with hard threshold function, the new threshold of construction Value function is as follows：

In formula (1)For through threshold function table treated wavelet packet coefficient, wherein α_j,iFor i-th on WAVELET PACKET DECOMPOSITION jth layer Frequency range sub-belt energy normalization coefficient；W_j,iFor the wavelet packet coefficient before processing；Using the heuristic threshold value based on heursure Selection of Function chooses threshold value λ, and l is adjustment factor, 0.5≤l≤1；

Wavelet packet energy coefficient α in formula (1)_j,iIt can be by wavelet packet coefficient W_j,iIt acquires, specially：Signals and associated noises are through j layers of small echo Packet decompose after, wavelet packet coefficient be W (j, 0), W (j, 1) ..., W (j, 2^j- 1), the ENERGY E of each sub-band wavelet packet_j,i=| | W(j,i)||², total wavelet-packet energy isWavelet-packet energy coefficientAt will be through threshold function table Wavelet packet coefficient after reason rebuilds wavelet packet tree, and the signal after inverse transformation reconstruct denoising, obtains noise reduction vibration signal s_v(t) with Noise reduction voice signal s_a(t)。

3. the omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features as described in claim 1, special Sign is that complementary population mean empirical mode decomposition algorithm described in third step includes the following steps：

1) the opposite white noise signal of symbol is added to signal s (t) to be decomposed in couples, forms two new signal s₁(t), s₂(t)；

2) to echo signal s₁(t)、s₂(t) empirical mode decomposition is carried out, is specifically included：

2.1) echo signal s is determined₁(t) all maximum point and minimum point are then smooth with one by all maximum points Curve connect to obtain coenvelope line, all minimum points are connected to obtain lower envelope with another smooth curve Line makes to include all signals between upper and lower envelope line；The average value of upper and lower envelope is denoted as m₁(t), echo signal s is obtained₁ (t) one-component h₁(t), i.e. s₁(t)-m₁(t)=h₁(t) (2)

2.2) by h₁(t) as s₁(t) it brings into formula (2), repeats screening k times of step 2.1), obtain h_1k(t)=h_1(k-1) (t)-m_1k(t) so that h_1k(t) become an intrinsic mode function component；Remember c₁(t)=h_1k(t), then c₁(t) it is signal s₁(t) First component for meeting intrinsic mode function condition, i.e. intrinsic mode function component (IMF)；

By r₁(t) as s₁(t) step 2.1) obtains s with 2.2) before repeating₁(t) second meets intrinsic mode function condition Component c₂(t)；

The intrinsic mode function condition is that the extreme point in the entire time serieses of a. at most differs one with the quantity of zero crossing, B. the mean value of upper and lower envelope that any moment is obtained by local maximum and local minimum is zero；

Echo signal s₂(t) also according to s₁(t) processing mode, according to step 2.1) to 2.4) to s₂(t) it is handled；

3) above-mentioned steps 1 are recycled)~2)；It repeats to the opposite pairs of white noise signal of signal s (t) to be decomposed addition symbol then Obtained new signal is subjected to empirical mode decomposition；Decomposition result is finally subjected to population mean operation, obtains decomposition result such as Formula (5), i.e.,In formula, s (t) is signal to be decomposed；c_j(t) (i=1,2 ..., n) it is jth A intrinsic mode function component (IMF), r_n(t) it is residual components.

4. the omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features as described in claim 1, The calculation formula for being characterized in the energy of the i-th rank intrinsic mode function component in the 4th step is N is intrinsic mode function component exponent number of the signal to be decomposed through complementary population mean empirical mode decomposition in formula (6), and N is each The data points of intrinsic mode function component, c_j(t) (i=1,2 ..., n) it is j-th of intrinsic mode function component (IMF)；

The energy summation of each rank intrinsic mode function component isThe then energy of the intrinsic mode function component of the i-th rank Amount ratio is defined asWork as R₁+R₂+…+R_ZZ values when >=90%, as normalized energy value add up more than 90% Z values.

5. the omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features as described in claim 1, It is characterized in that the computational methods that the intrinsic mode function component sample entropy is extracted in the 5th step are as follows：

1) it is c to remember i-th of intrinsic mode function component_i(t)={ c (n) }=c (1), c (2) ..., the N point data sequences of c (N), Sequence { c (n) } is formed to the vector of m dimensions, C in order_m(1),…,C_m(N-m+1), i.e. C_m(i)=c (i), c (i+1) ..., c (i+m-1)},1≤i≤N-m+1 (9)

2) definition vector C_m(i) and C_mThe distance between (j) d [C_m(i),C_m(j)] (i ≠ j) in two groups of vector corresponding elements most The absolute value of big difference；I.e.

3) similar tolerance r (r ＞ 0) is given, to each i Data-Statistics d [C_m(i),C_m(j)] C of ＜ r_m(j)(1≤j≤N-m+1,j ≠ i) number, then calculate its ratio with total distance N-m, be denoted asI.e.

4) all average value B are calculated^(m)(r), i.e.,

I-th of IMF component of signal s (t) to be decomposed is denoted as c respectively_i(t), c_i(t) discrete Fourier transform is C_i(w), into One step derives that the power spectrum of vibration signal and voice signal is：That is S_i(w)={ S_i(1),S_i (2),…,S_i(N) } Power Spectral Entropy H, is thus defined_i(f), i.e.,In formula (15), subscript f Represent frequency domain；q_i(w) it is w-th of power spectrum percentage in entire spectrum.

6. the omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features as described in claim 1, It is as described below that core core pivot element analysis method is combined described in the 6th step of feature：

If input training characteristics sample X_k(k=1,2 ..., N) it is mapped as φ (X_k), i.e.,：X_k→φ(X_k), the data after transformation Meet：Training sample φ (X after then converting_k) total population scatter matrix C be

According to theory of reproducing kernel space, feature vector V mono- is positioned at by φ (X₁)、φ(X₂)…φ(X_N) composition feature space in,

The problem of seeking formula (17) feature vector V in this way is converted to the feature vector α of solution formula (20)；

One group of nonzero eigenvalue λ is obtained by solving characteristic equation (20)_iAnd corresponding meet normalizing condition λ_i(αⁱ,αⁱ)= 1 feature vector αⁱ(i=1,2 ..., N ', N '≤N)；

The projection principal component V on feature space is obtained according to formula (18)_j(j=1,2 ..., N), if X be a test sample, then its V_jOn be projected asIn formula (21) solution procedure Middle solution kernel function K (X_i,X_j)=(φ (X_i)^Tφ(X_j)) definite expression formula method it is as follows：

Centralization as shown in formula (22), i.e. K '=K-K × L-L × K+L × K are carried out to the kernel matrix K obtained after transformation × L (22), wherein L are N ranks matrix and satisfaction

Global kernel function is Polynomial kernel function K₁(X_i,X_j)=[(X_i·X_j)+1]^q, local kernel function is RBF kernel functions K₂(X_i, X_j)=exp (- | | X_i-X_j||²/δ²), local kernel function and global kernel function are combined to obtain kernel function K (X_i,X_j)=(φ (X_i )^Tφ(X_j)) definite expression formula, that is, compound kernel function K (X_i,X_j)=θ K₁(X_i,X_j)+(1-θ)K₂(X_i,X_j) (23), wherein, θ is compound kernel function ratio, and q is Polynomial kernel function most high-order term number, and δ is RBF width parameters；

Described use combines feature samples dimension-reduction treatment of the core core pivot element analysis method to three category feature of acoustic signal after reunification Basic step includes：

5) standardization feature vector：

7) according to accumulation contribution rate extraction before i characteristic value and its corresponding preceding i feature vector, composing training noise reduction shake sound believe The principal component direction of number feature samples collection of three category features after reunification；

8) training three category feature of noise reduction acoustic signal feature samples collection after reunification is projected into the principal component side that 7) step acquires Upwards, the pivot of the feature samples collection of trained three category feature of noise reduction acoustic signal after reunification is extracted；

9) it three category feature of input test noise reduction acoustic signal feature samples after reunification and carries out corresponding kernel mapping and obtains nuclear moment Then battle array projects to the nuclear matrix on the principal component direction that 7) step acquires, obtain accordingly testing noise reduction acoustic signal three classes special The feature samples pivot of sign after reunification, M contribution rate of accumulative total reaches 90% pivot before only choosing.

7. the omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features as described in claim 1, special Sign is set in the 7th stepFor the pivot of the feature samples collection of three category feature of noise reduction acoustic signal after reunification, as defeated Incoming vector, t=[t₁,t₂,…,t_N]^TFor object vector, then Method Using Relevance Vector Machine disaggregated model such as formula (25) is shown, i.e.,

In formula (25), w is weight vectors, w=[w₀,w₁,…,w_N]^T；K(x,x_i)=exp (- | | x-x_i||²/υ²) for kernel function, υ For kernel functional parameter, noise ε_n~N (0, σ²)；

For new input vector x_*, corresponding desired value t_*Probabilistic forecasting formula be p (t_*| t)=∫ p (t_*|w,α,σ²)p (w,α,σ²|t)dwdαdσ²(29), p (w, α, σ can be obtained by Bayes's derivation being carried out to formula (29)²| t)=p (w | t, α, σ²)p (α,σ²| t) (30) carry out formula (30) approximate processing, and the learning process of Method Using Relevance Vector Machine is maximizes p (α, σ²|t)∝p(t |α,σ²)p(α)p(σ²) process, that is, find α_MP、σ² _MP, meet

In formula (34), Σ_i,iFor Σ=[σ^-2Φ^TΦ+diag(α₀,α₁,…,α_N)]^-1In i-th diagonal entry, μ_iFor weight Vectorial μ=σ^-2ΣΦ^TI-th of element of t；After enough updates, most α_iInfinitely great, corresponding w will be approached_i It is 0, and others α_iFinite value can then be leveled off to, corresponding x_iCollection be collectively referred to as associated vector, and then associated vector can be obtained The disaggregated model of machine, using between class sample mean Euclidean distance as separable measures, by omnipotent breaker fault diagnosis this More classification problems are converted into multiple two classification problems, establish the order binary tree multi-categorizer fault diagnosis based on Method Using Relevance Vector Machine Model；

For two class sample setsWithWherein x_i∈ A classes,Class, then sample mean between the class of A classes and B classes Euclidean distance is as shown in formula (35)In formula (35)For 2 differences Euclidean distance between classification sample；

The process of establishing of the order binary tree multi-categorizer model based on Method Using Relevance Vector Machine is, in divide-shut brake action process Input vector of the preceding M pivot selected under different mechanical breakdown states and normal condition as Method Using Relevance Vector Machine, with omnipotent Formula breaker is normal, falseness is closed a floodgate, separating brake is not thorough or certain mutually asynchronous divide-shut brake action state is the defeated of Method Using Relevance Vector Machine Go out, establish omnipotent breaker fault identification model successively, the order of Method Using Relevance Vector Machine is selected according to the size of Euclidean distance；

Using the quanta particle swarm optimization (hereinafter referred to as improving QPSO) of converging diverging coefficient automatic adjusument to Method Using Relevance Vector Machine Disaggregated model in kernel functional parameter υ optimize, kernel functional parameter υ is denoted as to particle to be optimized；

The fitness function for improving QPSO rolls over cross validation discrimination function for K-, i.e.,X in formula_lr And x_lwIt is got in the l times verification sample set for being respectively Method Using Relevance Vector Machine by the sample size of correct and wrong classification, fitness value Greatly, represent that solution is more excellent；

The tool optimized using QPSO is improved to the kernel functional parameter υ in two sorter model of order based on Method Using Relevance Vector Machine Body step is：

4) judge whether the iterations for reaching setting；It is to exit, group's extreme value at this time is required optimal kernel function Parameter；It is no, return to step 2)；

Calculating function of the discrimination that the cross validation of RVM is classified as particle fitness is selected in step 2), calculates step It is as follows：

2.1) kernel functional parameter is given according to newer particle；

2.3) value is returned as the particle fitness for improving QPSO models.

8. for performing the ten thousand of any of the above-described omnipotent breaker mechanical failure diagnostic method based on acoustic signal Fusion Features Can formula circuit-breaker switching on-off fault detection system, it is characterized in that including for fixing the operation console of omnipotent breaker, acceleration Sensor, sound pick-up, solid-state relay group, industrial personal computer, I/O control cards and USB capture cards, the connection mode of each component are：Work Control machine is connect by isa bus with I/O control cards, and I/O control cards are also connect with solid-state relay group, and acceleration transducer is used for It is fixed on omnipotent breaker, the sound pick-up is mounted on operation console, and acceleration transducer and sound pick-up pass through usb bus It is connected with USB capture cards, the USB capture cards are connected with industrial personal computer, and the solid-state relay group includes being respectively used to control Breaker energy storage processed, combined floodgate, separating brake and under-voltage energy storage relay, closing relay, separating brake relay and under-voltage relay.

9. omnipotent breaker divide-shut brake fault detection system as claimed in claim 8, it is characterized in that the USB capture cards are USB7648A capture cards, the PCL720 control cards of the I/O control cards, the acceleration transducer model LC0159 are described The model KZ520B of sound pick-up；The I/O control cards give electric signal to energy storage relay, the combined floodgate in solid-state relay group Relay, separating brake relay and under-voltage relay are operated；By USB capture cards by acceleration transducer and the mould of sound pick-up Intend signal and be converted into digital signal, and pass through usb bus and send industrial personal computer to.