MULTI-SPECTRUM FUSION DECISION FAULT DIAGNOSIS METHOD FOR HIGH-SPEED ELECTRIC MULTIPLE-UNIT BEARING

Abstract

A multi-spectrum fusion decision fault diagnosis method for a high-speed electric multiple-unit bearing, including the steps of collecting, by an acceleration sensor vibration acceleration signals of a high-speed electric multiple-unit; performing multi-method spectrum analysis on vibration acceleration signal data of a plurality of faults under a constant working condition, and marking the data to construct a spectrum data set; extracting inherent fault features of the spectrum data set to form a fault feature data set; performing weight fusion on spectrum fault features of different methods to construct a spectrum fault feature function, and fusing each set of spectra in the spectrum data set into a data point; using a criterion that a fisher discrimination ratio minimizes an intra-class spacing and maximizes an inter-class spacing to seek an optimal weight combination under the data set, so that fusion data points distinguish the different types of faults to a maximum degree.

Claims

1. A multi-spectrum fusion decision fault diagnosis method for a high-speed electric multiple-unit bearing, comprising the following steps: step S100: collecting, by an acceleration sensor, vibration acceleration signals of the high-speed electric multiple-unit bearing, and performing classification according to fault types of the bearing, the fault types totally comprising four classes of an inner ring fault, an outer ring fault, a rolling body fault and a holder fault; step S200: performing analysis of 4 kinds of spectra of a Fourier spectrum, an envelope spectrum, a power spectrum and a fast spectral kurtosis-square envelope spectrum on the collected vibration acceleration signals respectively, to obtain a spectrum data set composed of different fault types and different spectrum analysis methods; step S300: extracting strength of a fault feature frequency corresponding to a fault position from the spectrum data set according to the fault feature frequency of the monitored bearing as fault features, simplifying the spectra in the spectrum data set based on the fault features, to obtain a fault feature data set composed of different fault types and different spectrum analysis methods, each piece of data in the fault feature data set being data based on the fault types respectively, and performing analysis of the Fourier spectrum, the envelope spectrum, the power spectrum, and the fast spectral kurtosis-square envelope spectrum on each piece of data in the fault feature data set, to obtain fault features extracted from the 4 kinds of spectra; step S400: constructing a fault feature function by weight fusion of 4 fault features from the 4 kinds of spectra, and converting each piece of data in the fault feature data set into a fusion data point representing a fault type and a diagnosis effect of the data; step S500: optimizing weights of the 4 fault features through a criterion that a fisher discrimination ratio maximizes an intra-class spacing and minimizes an inter-class spacing, to obtain an optimal weight combination of 4 classes of spectrum fault features in the fault feature dataset, and meanwhile obtaining a numerical value interval formed by fusion data points after fusion of the different fault types to form an experience range for judging the fault types; and step S600: inputting test data, performing analysis of the Fourier spectrum, the envelope spectrum, the power spectrum, and the fast spectral kurtosis-square envelope spectrum on the test data, then performing weight fusion on the test data according to the optimal weight combination to form a fusion spectrum, extracting the fault features of the 4 kinds of spectra and performing weight fusion to form fusion data points, judging fault types of the fusion data points according to the experience range, and finally implementing fault diagnosis of the high-speed electric multiple-unit bearing.

2. The method according to claim 1, wherein in step S200, the spectrum data set is represented as: ${}_p=\left\{\begin{array}{c}{X}_{1,p}^1\left(f\right),.Math.,X_{1,p}^k\left(f\right),.Math.,X_{1,p}^K\left(f\right)\\ X_{2,p}^1\left(f\right),.Math.,X_{2,p}^k\left(f\right),.Math.,X_{2,p}^K\left(f\right)\\ .Math.\\ X_{m,p}^1\left(f\right),.Math.,X_{m,p}^k\left(f\right),.Math.,X_{m,p}^K\left(f\right)\\ .Math.\\ X_{M,p}^1\left(f\right),.Math.,X_{M,p}^k\left(f\right),.Math.,X_{M,p}^K\left(f\right)\end{array}\right\}$ $=\left\{{}_1,{}_2,.Math.,{}_P\right\}$ where, p represents a p.sup.th fault type, m represents an m.sup.th spectrum analysis method, representing four methods of analysis of the Fourier spectrum, the envelope spectrum, the power spectrum, and the fast spectral kurtosis-square envelope spectrum, k represents a k.sup.th piece of data under each fault type, where m=1, 2, . . . , M, p=1, 2, . . . , P, and k=1, 2, . . . , K, .sub.p represents a spectrum data set formed by data of the p.sup.th fault type, a frequency domain sequence after spectrum analysis on a k.sup.th group of data by adopting the m.sup.th method is denoted as X.sub.m.sup.k(f), f=1, . . . N, f is a sequence number of the vibration acceleration signals, a length of the sequence number is N, and meanwhile, in order to avoid an influence of a difference of spectrum magnitude orders of different methods on results, normalization processing is performed on the spectra.

3. The method according to claim 2, wherein in step S300, a step of extracting fault features of corresponding positions of the spectra from the spectrum data set to construct the fault feature data set is as follows: a fault feature frequency of the bearing at a monitoring position is f.sub.p, p=1, . . . P, and for a spectrum X.sub.m,p.sup.k(f), a fault feature h.sub.m,p.sup.k extracted from the spectrum is: $h_{m,p}^k=\underset{l=1}{\overset{L}{.Math.}}{a}_p^l{f}_p^2,$ where, a.sub.p.sup.l is an amplitude of an l.sup.th harmonic wave of the fault feature frequency f.sub.p, L is the number of frequency multiplications to be detected, h.sub.m,p.sup.k is the fault feature extracted from a spectrum of the k.sup.th group of data of the m.sup.th method in the p.sup.th fault, and the spectrum data set is simplified as: ${}_p=\left\{\begin{array}{c}{h}_{1,p}^1,h_{1,p}^2,.Math.,h_{1,p}^k,.Math.h_{1,p}^K\\ h_{2,p}^1,h_{2,p}^2,.Math.,h_{2,p}^k,.Math.h_{2,p}^K\\ .Math.\\ h_{m,p}^1,h_{m,p}^2,.Math.,h_{m,p}^k,.Math.h_{m,p}^K\\ .Math.\\ h_{M,p}^1,h_{M,p}^2,.Math.,h_{M,p}^k,.Math.h_{M,p}^K\end{array}\right\}$ $=\left\{{}_1,{}_2,.Math.,{}_p,.Math.{}_P\right\},$ where, .sub.p represents a p.sup.th fault feature data set extracted from a p.sup.th fault type spectrum data set .sub.p, and represents a simplified global fault feature dataset.

4. The method according to claim 3, wherein in step S400, the constructed fault feature function is represented as: $S_p^k=\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k$ where, S.sub.p.sup.k is a fault feature function of a k.sup.th piece of data in the p.sup.th fault type, and a meaning of the fault feature function is a fusion data point formed by weight fusion of different spectrum features.

5. The method according to claim 4, wherein in step S500, the criterion of the fisher discrimination ratio is represented as: $C=\underset{p=1}{\overset{P}{.Math.}}\underset{lp}{\overset{P}{.Math.}}\frac{{\left(m_p-m_l\right)}^2}{s_p^2+s_l^2}=\frac{{}^T{S}_b}{{}^T{S}_w},$ in the formula, P represents P kinds of different fault types, is a projection direction, that is a weight matrix of the fusion spectrum, where =[.sub.1, .sub.2, . . . .sub.M].sup.T, m.sub.p and m.sub.l represent an intra-class mean value of a p.sup.th class and an intra-class mean value of an l.sup.th class respectively, s.sub.p and s.sub.l represent an intra-class variance of the p.sup.th class and an intra-class variance of the l.sup.th class respectively, S.sub.b represents a total inter-class dispersion matrix, and S.sub.w represents a total intra-class dispersion matrix.

6. The method according to claim 5, wherein optimizing the weights of the 4 fault features through the criterion that the fisher discrimination ratio maximizes the intra-class spacing and minimizes the inter-class spacing comprises: S501: solving the total intra-class dispersion matrix S.sub.w after projection according to an intra-class mean value and each intra-class data point of the fault feature set : $S_w=\underset{p=1}{\overset{P}{.Math.}}{v}_p$ $v_p=\underset{k=1}{\overset{K}{.Math.}}\left(S_p^k-m_p\right){\left(S_p^k-m_p\right)}^T$ where, in the formula, v.sub.p represents an intra-class dispersion of a p.sup.th class of fault data, and the fault feature function is substituted into the above formula to obtain: $\begin{array}{c}{v}_p=\underset{k=1}{\overset{K}{.Math.}}\left(\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k-\frac{\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k}{K}\right){\left(\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k-\frac{\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k}{K}\right)}^T\\ =\frac{{}^T}{K}\left(\underset{k=1}{\overset{K}{.Math.}}\left({\mathrm{KU}}_{p,k}-{}_{p}1\right){\left({\mathrm{KU}}_{p,k}-{}_{p}1\right)}^T\right)\end{array}$ ${}_p=\left[\begin{array}{cccc}{f}_{1,p,1}& f_{1,p,2}& .Math.& f_{1,p,k}\\ f_{2,p,1}& & & f_{2,p,k}\\ .Math.& & & .Math.\\ f_{M,p,1}& f_{M,p,2}& .Math.& f_{M,p,k}\end{array}\right]$ where, U.sub.p,k=[h.sub.1,p,k, h.sub.2,p,k, . . . , h.sub.M,p,k].sup.T and 1=[1, 1, . . . 1].sup.T are column vectors with elements being 1; S502: solving the total inter-class dispersion matrix S.sub.b after projection according to the intra-class mean value and each intra-class data point of the fault feature set : $S_b=\underset{p=1}{\overset{P}{.Math.}}\underset{lp}{\overset{P}{.Math.}}\left(m_p-m_l\right){\left(m_p-m_l\right)}^T$ $\begin{array}{c}{S}_b=\underset{p=1}{\overset{P}{.Math.}}\underset{lp}{\overset{P}{.Math.}}\left(\frac{\underset{k=1}{\overset{K}{.Math.}}{S}_p^k-\underset{k=1}{\overset{K}{.Math.}}{S}_l^k}{K}\right){\left(\frac{\underset{k=1}{\overset{K}{.Math.}}{S}_p^k-\underset{k=1}{\overset{K}{.Math.}}{S}_l^k}{K}\right)}^T\\ =\underset{p=1}{\overset{P}{.Math.}}\underset{lp}{\overset{P}{.Math.}}\left(\frac{\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k-\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k}{K}\right){\left(\frac{\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k-\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k}{K}\right)}^T\\ =\frac{1}{K^2}\underset{p=1}{\overset{P}{.Math.}}\underset{lp}{\overset{P}{.Math.}}\left(\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k-\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k\right){\left(\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k-\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k\right)}^T\\ =\frac{1}{K^2}{}^T\left(\underset{p=1}{\overset{P}{.Math.}}\underset{lp}{\overset{P}{.Math.}}\left({}_p-{}_l\right){\left({}_p-{}_l\right)}^T\right)\end{array}$ and S503: solving an optimal projection direction .sup.T, seeking a maximum value of a discrimination ratio c, and in order to solve the discrimination ratio c, introducing a Lagrange operator here, solving of the discrimination ratio C being able to be converted into: $\underset{}{\max }\left({}^T{S}_b\right),s.t.{}^T{S}_w=c0$ $L\left(,\right)={}^T{S}_b-\left({}^T{S}_b-c\right)$ and after solving a partial derivative of w, obtaining the following formula: $S_b{}^{*}-S_w{}^{*}=0$ hereby, obtaining an optimal projection direction, that is, an optimal spectrum weight *, and performing normalization processing on the optimal spectrum weight, to obtain the most suitable spectrum weight combination w* on the fault feature data set.

7. The method according to claim 4, wherein in step S500, the optimal fusion data obtained after acquiring the optimal weight combination is represented as: $S{t}_p^k=\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k={}^{*}.Math.U_{p,k}$ for the p.sup.th class of faults, a mean value of the fusion data points of the p.sup.th class of faults being ${\stackrel{}{m}}_p=\frac{\underset{k=1}{\overset{K}{.Math.}}{\mathrm{St}}_p^k}{K},$ and an interval range in which the data points fall after the faults are fused being able to be represented by minimum and maximum values of the data points of the p.sup.th class of faults, that is, an interval of the fusion data points of the p.sup.th class of faults being able to be represented as .sub.p=[min(St.sub.p.sup.k),max(St.sub.p.sup.k)].

8. The method according to claim 1, wherein in step S600, the obtained fusion spectrum needs to perform analysis of four kinds of spectra of the Fourier spectrum, the envelope spectrum, the power spectrum, and the fast spectral kurtosis-square envelope spectrum on input data to be diagnosed respectively, frequency domain sequences after amplitude normalization on the four kinds of spectra are denoted as X.sub.1(f), X.sub.2(f), X.sub.3(f) and X.sub.4(f) respectively, where f represents a corresponding frequency of the frequency domain sequences, and the fusion spectrum is denoted as: $X\left(f\right)=\underset{m=1}{\overset{M}{.Math.}}{w}_m^{*}.Math.X_m\left(f\right)$ where, M is 4, w*=[w*.sub.1, w*.sub.2*, . . . , w*.sub.m, . . . , w*.sub.M], and fault diagnosis of a high-speed rail bogie bearing is realized by judging a fault feature frequency amplitude in the fusion spectrum.

9. The method according to claim 7, wherein in step S600, when spectrum analysis of newly input data is performed and a fusion data point S.sub.test falls in an interval .sub.p, it is judged that data is of the p.sup.th class of fault type, if fusion numerical value intervals of different types of faults overlap and a newly input data point just falls in an overlapping interval, a relatively-close point may be selected as a judgment conclusion according to an absolute value of a difference value between the fusion data point and the mean value m.sub.p, and when S.sub.test falls in an overlapping interval p,l[1,P] of a fusion data interval .sub.p of the p.sup.th class of faults and a fusion data interval .sub.l of the l.sup.th class of faults, if |S.sub.testm.sub.l|<|S.sub.testm.sub.p|, a fault is judged to be the l.sup.th class of faults, otherwise, the fault is judged to be the p.sup.th class of faults.

Description

BRIEF DESCRIPTION OF FIGURES

[0037] FIG. 1 is a flow diagram of a multi-spectrum fusion decision fault diagnosis method for a high-speed electric multiple-unit bearing provided by an embodiment of the present disclosure.

[0038] FIG. 2 is time domain signal display of part of experiment data of one embodiment of the present disclosure.

[0039] FIG. 3 is diagnosis effect display of a frequency domain analysis method on the experiment data according to the present disclosure.

[0040] FIG. 4 represents an experiment range of fusion data points for each fault type in an experiment data set according to the present disclosure.

[0041] FIG. 5 is time domain signal display of one data to be diagnosed and time domain signal display after analysis of a spectrum, a power spectrum, an envelope spectrum and a fast spectral kurtosis-square envelope spectrum respectively according to the present disclosure.

[0042] FIG. 6 is time domain signal display of another data to be diagnosed and time domain signal display after analysis of a spectrum, a power spectrum, an envelope spectrum and a fast spectral kurtosis-square envelope spectrum respectively according to the present disclosure.

[0043] FIG. 7 is fusion spectrum result display of one data to be diagnosed according to the present disclosure.

[0044] FIG. 8 is display of a comparative result of the present disclosure and other frequency domain methods on the same experiment data set.

[0045] The present disclosure is further explained in combination with accompanying drawings and embodiments.

DETAILED DESCRIPTION

[0046] Specific embodiments of the present disclosure will be described in more detail with reference to accompanying drawings below. Although the specific embodiments of the resent disclosure are shown in the accompanying drawings, it should be understood that the present disclosure may be realized in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable the present disclosure to be understood a more thoroughly and to enable a range of the present disclosure to be completely communicated to those skilled in the art.

[0047] It should be noted that certain vocabularies are used in the specification and claims to refer to specific components. It should be understood by those skilled in the art that technicians may call the same component with different nouns. The present specification and the claims are not based on differences of the nouns as a mode of distinguishing components, but are based on differences in functions of the components as a distinction criterion. If comprising or including mentioned throughout the entire specification and claims is used as an open-ended term, it should be interpreted as including but not limited to. The specification is subsequently described as a better implementation of implementing the present disclosure, but the description is for the purpose of a general principle of the specification and is not intended to limit the scope of the present disclosure. The scope of protection of the present disclosure shall be as defined in the attached claims.

[0048] In an embodiment, as shown in FIG. 1 to FIG. 8, the present disclosure provides a multi-spectrum fusion decision fault diagnosis method for a high-speed electric multiple-unit bearing, including the following steps: [0049] S100: vibration acceleration signals of the high-speed electric multiple-unit bearing are collected through an acceleration sensor; and [0050] S200: analysis of a Fourier spectrum, an envelope spectrum, a power spectrum and a fast spectral kurtosis-square envelope spectrum is performed on the collected vibration acceleration signals respectively, to obtain a spectrum data set composed of different fault types and different spectrum analysis methods.

[0051] In the step, the constructed spectrum data set may be represented as:

[00015]$\begin{array}{c}{}_p=\left\{\begin{array}{c}{X}_{1,p}^1\left(f\right),.Math.,X_{1,p}^k\left(f\right),.Math.,X_{1,p}^K\left(f\right)\\ X_{2,p}^1\left(f\right),.Math.,X_{2,p}^k\left(f\right),.Math.,X_{2,p}^K\left(f\right)\\ .Math.\\ X_{m,p}^1\left(f\right),.Math.,X_{m,p}^k\left(f\right),.Math.,X_{m,p}^K\left(f\right)\\ .Math.\\ X_{M,p}^1\left(f\right),.Math.,X_{M,p}^k\left(f\right),.Math.,X_{M,p}^K\left(f\right)\end{array}\right\}\\ =\left\{{}_1,{}_2,.Math.,{}_P\right\}\end{array}$ [0052] where, p represents a p.sup.th fault type, m represents an m.sup.th spectrum analysis method, k represents a k.sup.th piece of data under each fault type, m=1, 2, . . . , M, p=1, 2, . . . , P, k=1, 2, . . . , K, .sub.p, represents a spectrum data set formed by data of the p.sup.th fault type, a frequency domain sequence after spectrum analysis on a k.sup.th group of data by adopting the m.sup.th method is denoted as X.sub.m.sup.k(f), f=1, . . . N, f is a sequence number of the vibration acceleration signals, and a length of the sequence number is N.

[0053] Before model training, it is necessary to conduct spectrum analysis and processing on classified data, and label and classify the spectra according to its fault types and frequency domain analysis methods, so as to facilitate obtaining an optimal weight by a subsequent trained model.

[0054] Exemplarily, the vibration acceleration signals mentioned in step S001 and step S002 may be time domain signals whose object is the high-speed electric multiple-unit bearing, its amplitude changes with time, and different signals correspond to different pieces of information of fault types and fault degrees.

[0055] Specifically, the fast spectral kurtosis-square envelope spectrum method mentioned in step S200 represents the use of a square envelope spectrum method to perform spectrum analysis on a resonance frequency band preferred by fast spectral kurtosis, so as to improve fault feature frequency clarity in the spectrum.

[0056] S300: strength of a fault feature frequency corresponding to a fault position is extracted from the spectra according to a fault feature frequency of a monitored bearing as fault features, so that the spectra in the spectrum data set are simplified, to obtain a fault feature data set composed of different fault types and different spectrum analysis methods.

[0057] In the step, the step of extracting fault features of corresponding positions of the spectra from the spectrum data set to construct the fault feature data set is as follows: [0058] a fault feature frequency of a bearing at a monitoring position is f.sub.p, p=1, . . . P, and for a spectrum X.sub.m,p.sup.k(f), a fault feature h.sub.m,p.sup.k extracted from the spectrum is:

[00016]$h_{m,p}^k=\underset{l=1}{\overset{L}{.Math.}}{a}_p^l{f}_p^2,$ [0059] where, a.sub.P.sup.l is an amplitude of an l.sup.th harmonic wave of the fault feature frequency f.sub.p, L is the number of frequency multiplications to be detected, and h.sub.m,p.sup.k is the fault feature extracted from a spectrum of the k.sup.th group of data of the m.sup.th method in the p.sup.th fault. Therefore, the spectrum data set constructed in the step S200 may be simplified as:

[00017]$\begin{array}{c}{}_p=\left\{\begin{array}{c}{h}_{1,p}^1,h_{1,p}^2,.Math.,h_{1,p}^k,.Math.h_{1,p}^K\\ h_{2,p}^1,h_{2,p}^2,.Math.,h_{2,p}^k,.Math.h_{2,p}^K\\ .Math.\\ h_{m,p}^1,h_{m,p}^2,.Math.,h_{m,p}^k,.Math.h_{m,p}^K\\ .Math.\\ h_{M,p}^1,h_{M,p}^2,.Math.,h_{M,p}^k,.Math.h_{M,p}^K\end{array}\right\}\\ =\left\{{}_1,{}_2,.Math.,{}_p,.Math.{}_P\right\},\end{array}$ [0060] .sub.p represents a p.sup.th fault feature data set extracted from a p.sup.th fault type spectrum data set .sub.p, represents a simplified global fault feature dataset, which represents a fault diagnosis effect of each spectrum in the spectrum data set, meanwhile, in order to avoid the influence of spectrum magnitude order differences of different methods on the results, the above spectrum is normalized. [0061] S400: At this time, each piece of data in the fault feature data set corresponds to 4 fault features, which are respectively from four kinds of spectra of Fourier spectrum, envelope spectrum, power spectrum and fast spectral kurtosis-square envelope spectrum analysis constructing, a fault feature function is constructed by weight fusion of 4 fault features, and each group of data in the data set is converted into a data point representing a fault type and a diagnosis effect of the data.

[0062] In the step, the constructed fault feature function may be represented as:

[00018]$S_p^k=\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k$ [0063] where, S.sub.p.sup.k is a fault feature function of a k.sup.th piece of data in the p.sup.th fault type, and a meaning of the fault feature function is a data point formed by weight fusion of different spectrum features.

[0064] The purpose of this step is to perform weighted fusion on the fault features from the different spectra to form data points containing a plurality of spectrum diagnostic features. The meaning of each data point is a fusion of diagnosis effects of different frequency domain analysis methods on the same data, which facilitates subsequent classification processing.

[0065] Step S500: weights of the 4 fault features are optimized through a criterion that a fisher discrimination ratio maximizes an intra-class spacing and minimizes an inter-class spacing, to obtain an optimal weight combination of 4 classes of spectrum fault features in the fault feature dataset, and meanwhile a numerical interval formed by the data points after fusion of the different fault types is obtained to form an experience range for judging the fault types.

[0066] The core idea of the fisher discrimination ratio in this step is to seek the best projection direction, which may maximize each class spacing after projection and minimize the intra-class distance in each class, so as to distinguish each fault type to the greatest extent. Therefore, the criterion of the fisher discrimination ratio may be represented as follows:

[00019]$C=\underset{p=1}{\overset{P}{.Math.}}\underset{lp}{\overset{P}{.Math.}}\frac{{\left(m_p-m_l\right)}^2}{s_p^2+s_l^2}=\frac{{}^T{S}_b}{{}^T{S}_w},$ [0067] in the formula, P represents P kinds of different fault types, is a projection direction, that is a weight matrix of a fusion spectrum, where =[.sub.1, .sub.2, . . . .sub.M].sup.T, m.sub.P and m.sub.l represent an intra-class mean value of a p.sup.th class, and an intra-class mean value of an l.sup.th class respectively, s.sub.p and s.sub.l represent an intra-class variance of the p.sup.th class, and an intra-class variance of the l.sup.th class respectively, S.sub.b represents a total inter-class dispersion matrix, and S.sub.w represents a total intra-class dispersion matrix.

[0068] Therefore, solving of the optimal weight in step S500 may be disassembled into the following several sub-steps:

[0069] S501: the total intra-class dispersion matrix S.sub.w after projection is solved according to an intra-class mean value and each intra-class data point of the fault feature set :

[00020]$\begin{array}{c}{S}_w=\underset{p=1}{\overset{P}{.Math.}}{v}_p\\ v_p=\underset{k=1}{\overset{K}{.Math.}}\left(S_p^k-m_p\right){\left(S_p^k-m_p\right)}^T\end{array}$ [0070] in the formula, v.sub.p represents an intra-class dispersion of a p.sup.th class of fault data, and the fault feature function constructed in step S400 is substituted into the above formula to obtain:

[00021]$v_p=\underset{k=1}{\overset{K}{.Math.}}\left(\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k-\frac{\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k}{K}\right){\left(\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k-\frac{\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k}{K}\right)}^T=\frac{{}^T}{K}\left(\underset{k=1}{\overset{K}{.Math.}}\left(K{U}_{p,k}-{}_{p}1\right){\left(K{U}_{p,k}-{}_{p}1\right)}^T\right)$${}_p=\left[\begin{array}{cccc}{f}_{1,p,1}& f_{1,p,2}& .Math.& f_{1,p,k}\\ f_{2,p,1}& & & f_{2,p,k}\\ .Math.& & & .Math.\\ f_{M,p,1}& f_{M,p,2}& .Math.& f_{M,p,k}\end{array}\right]$ [0071] where, U.sub.p,k=[h.sub.1,p,k, h.sub.2,p,k, . . . , h.sub.M,p,k].sup.T and 1=[1, 1, . . . 1].sup.T are column vectors with elements being 1; [0072] S502: the total inter-class dispersion matrix S.sub.b after projection is solved according to the intra-class mean value and each intra-class data point of the fault feature set :

[00022]$S_b=\underset{p=1}{\overset{P}{.Math.}}\underset{lp}{\overset{P}{.Math.}}\left(m_p-m_l\right){\left(m_p-m_l\right)}^T$$S_b=\underset{p=1}{\overset{P}{.Math.}}\underset{lp}{\overset{P}{.Math.}}\left(\frac{\underset{k=1}{\overset{K}{.Math.}}{S}_p^k-\underset{k=1}{\overset{K}{.Math.}}{S}_l^k}{K}\right){\left(\frac{\underset{k=1}{\overset{K}{.Math.}}{S}_p^k-\underset{k=1}{\overset{K}{.Math.}}{S}_l^k}{K}\right)}^T=\underset{p=1}{\overset{P}{.Math.}}\underset{lp}{\overset{P}{.Math.}}\left(\frac{\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k-\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k}{K}\right){\left(\frac{\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k-\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k}{K}\right)}^T=\frac{1}{K^2}\underset{p=1}{\overset{P}{.Math.}}\underset{lp}{\overset{P}{.Math.}}\left(\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k-\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k\right){\left(\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k-\underset{k=1}{\overset{K}{.Math.}}\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k\right)}^T=\frac{1}{K^2}{}^T\left(\underset{p=1}{\overset{P}{.Math.}}\underset{lp}{\overset{P}{.Math.}}\left({}_p-{}_l\right){\left({}_p-{}_l\right)}^T\right)$ [0073] and S503: an optimal projection direction .sup.T is solved, that is a maximum value of a discrimination ratio c is searched, in order to solve the discrimination ratio c, a Lagrange operator is introduced here, and solving of the discrimination ratio C may be converted into:

[00023]$\begin{array}{c}\underset{}{\max }\left({}^T{S}_b\right),s.t.{}^T{S}_w=c0\\ L\left(,\right)={}^T{S}_b-\left({}^T{S}_b-c\right)\end{array}$ [0074] after solving a partial derivative of w, the following formula is obtained:

[00024]$S_b{}^{*}-S_w{}^{*}=0$ [0075] hereby, an optimal projection direction is obtained, that is, an optimal spectrum weight *, since data after dimensionality reduction on the data points loses its original physical meaning, and it only represents a discrimination classification result, and normalization processing needs to be performed on the optimal spectrum weight, to obtain the most suitable spectrum weight combination w* on the data set.

[0076] In this step, the numerical interval formed by the data points after the different fault types are fused is obtained to form the experience range. The optimal fusion data pint obtained after acquiring the optimal spectrum weight * may be represented as:

[00025]$S{t}_p^k=\underset{m=1}{\overset{M}{.Math.}}{}_m{h}_{m,p}^k={}^{*}.Math.U_{p,k}$ [0077] for the p.sup.th class of faults, a mean value of the fusion data points of the p.sup.th class of faults is

[00026]${\stackrel{}{m}}_p=\frac{\underset{k=1}{\overset{K}{.Math.}}{\mathrm{St}}_p^k}{K},$

and an interval range in which the data points fall after the faults are fused may be represented by minimum and maximum values of the data points of the faults, that is, an interval of the fusion data points of the p.sup.th class of faults may be represented as

[00027]${}_p=\left[\min \left({\mathrm{St}}_p^k\right),\max \left({\mathrm{St}}_p^k\right)\right].$ [0078] S600, data to be diagnosed is input, analysis of the Fourier spectrum, the envelope spectrum, the power spectrum, and the fast spectral kurtosis-square envelope spectrum is performed on the data, then weight fusion is performed on the data according to the optimal weight combination to form a fusion spectrum, the fault features of the 4spectra are extracted, weight fusion is performed to form the data points, fault types of the data points are judged according to the experience range obtained in step S500, and finally fault diagnosis of the high-speed electric multiple-unit bearing is implemented.

[0079] The fusion spectrum obtained in the step needs to perform analysis of four kinds of spectra of the Fourier spectrum, the envelope spectrum, the power spectrum, and the fast spectral kurtosis-square envelope spectrum, frequency domain sequences after amplitude normalization on the four kinds of spectra are denoted as X.sub.1(f), X.sub.2(f), X.sub.3(f) and X.sub.4(f) respectively, where f represents a corresponding frequency of the frequency domain sequences, and the fusion spectrum is denoted as:

[00028]$X\left(f\right)=\underset{m=1}{\overset{M}{.Math.}}{w}_m^{*}.Math.X_m\left(f\right)$ [0080] where, M is 4, w*=[w*.sub.1, w*.sub.2, . . . , w*.sub.m, . . . , w*.sub.M], and fault diagnosis of a high-speed rail bogie bearing is implemented by judging a fault feature frequency amplitude in the fusion spectrum.

[0081] Discrimination of the fault types obtained in the step is judged by the interval in which the fusion data points fall. When spectrum analysis of the newly input data is performed and the a fusion data point S.sub.test falls in an interval .sub.p, it is judged that the data is of the p.sup.th class of fault type, if fusion numerical value intervals of different types of faults overlap, and a newly input data point just falls in the overlapping interval, a relatively-close point may be selected as a judgment conclusion according to an absolute value of a difference value between the fusion data points and the mean value m.sub.p, and when S.sub.test falls in the overlapping interval p,l[1,P] of a fusion data interval .sub.p of the p.sup.th class of faults and a fusion data interval .sub.l of the l.sup.th class of faults, if |S.sub.testm.sub.l|<|S.sub.testm.sub.p|, a fault is judged to be the l.sup.th class of faults, otherwise, the fault is judged to be the p.sup.th class of faults.

[0082] The present embodiment defines the fusion spectrum fault diagnosis method for the high-speed electric multiple-unit bearing, which can not only adaptively assign weights to the spectra of different diagnosis effects and form the fusion spectrum according to working conditions of the data set, but also judge the fault types to which the data belongs according to the numerical values of the fusion data points of the newly input data.

[0083] FIG. 2 is time domain signal display of part of experiment data of one embodiment of the present disclosure. The experiment data contains 4 classes of faults, each class of fault data contains 75 groups of data, sampling time of each group of data is 1s, and 1 group of data is selected from each group of fault types for display in the figure. FIG. 3 is diagnosis effect display of a frequency domain analysis method on the experiment data according to the present disclosure, including four frequency domain analysis methods, respectively the spectrum, the power spectrum, the envelope spectrum and the fast spectral kurtosis-square envelope spectrum, and its meaning is diagnosis effects of different spectrum analysis methods on the data used in the experiment. FIG. 4 represents an experiment range of fusion data points for each fault type in an experiment data set according to the present disclosure, including four fault types: an inner ring fault, an outer ring fault, a rolling body fault and a holder fault. FIG. 5 is time domain signal display of one data to be diagnosed and time domain signal display after analysis of a spectrum, a power spectrum, an envelope spectrum and a fast spectral kurtosis-square envelope spectrum respectively according to the present disclosure. and fast spectral kurtosis-square envelope spectrum analysis respectively. FIG. 6 is time domain signal display of another data to be diagnosed and time domain signal display after analysis of a spectrum, a power spectrum, an envelope spectrum and a fast spectral kurtosis-square envelope spectrum respectively according to the present disclosure. FIG. 7 is a fusion spectrum result of one data to be diagnosed according to the present disclosure, which is obtained by weighing the spectrum, the power spectrum, the envelope spectrum, and fast spectral kurtosis-square envelope spectrum of the data to be diagnosed according to the optimal weight obtained by the present disclosure. FIG. 8 is a comparison between the present disclosure and other frequency domain analysis methods on the same experiment data set, including a comparison between the stability of the diagnosis effects of the present disclosure and the envelope spectrum and the spectrum to verify the robustness and anti-interference of the present disclosure.

[0084] Specifically, the experiment effect verification and the effect comparison are mainly aimed at the analysis and the comparison of CFIC evaluation indicators, the meaning of which is the clarity of a fault feature frequency in the spectra, which may be used for representing the diagnosis effect of the spectra. The definition of the CFIC is: CFIC=amplitudes of the fault feature frequency and its frequency multiplier in the spectra and/the sum of the amplitudes of all frequencies in the spectra. When the fault feature frequency in the spectra is clearer in the whole spectra, the larger the CFIC value is, the better the diagnosis effect of the spectrum is.

[0085] In the preferred example of the multi-spectrum fusion decision fault diagnosis method for the high-speed electric multiple-unit bearing, the data used is a group of traction motor bearing data, which is a prefabricated fault experiment with four classes of fault types of the inner ring fault, the outer ring fault, the rolling body fault and the holder fault performed under a rotating speed condition of 4100 rpm. The bearing model is NU214. The fault feature frequency of each position at the rotating speed of 4100 rpm is shown as Table 1.

TABLE-US-00001 TABLE 1 Fault position Inner ring Outer ring Rolling body Holder Fault feature 619.57 473.75 251.67 26.1 frequency

[0086] The vibration acceleration signals are collected by the acceleration sensor installed on a bearing sleeve with a sampling frequency of 25600 Hz, and original data is intercepted with is as a data sample, that is, 25600 points in each group of data. 75 groups of data of each fault type are selected for analysis, as shown in Table 2. The data set constructed from this is shown in the figure.

TABLE-US-00002 TABLE 2 Fault type Data volume Data duration Inner ring fault 75 groups 1 s Outer ring fault 75 groups 1 s Rolling body fault 75 groups 1 s Holder fault 75 groups 1 s

[0087] The above data set is simplified according to the step S300, and the obtained fault feature data set is partly shown as Table 3.

TABLE-US-00003 TABLE 3 Fault Power Envelope Fast spectral kurtosis-square type Spectrum spectrum spectrum envelope spectrum Inner 26240.46986 2368.272953 191223.3808 628166.7769 ring 67276.93627 2377.607588 376523.5572 40072.15936 fault 47487.35668 1404.72127 196740.1997 861172.0281 25303.68906 334.593432 355193.3283 91394.85138 80395.36453 4017.964376 209753.6648 562271.2158 21986.73369 351.2331881 234254.4274 641642.9219 68015.6184 2425.165212 343937.7028 21172.5859 46563.86547 1254.839292 210728.5369 532320.7335 22108.1852 257.443797 340663.3199 92881.17668 70866.94504 939.7028633 313017.0286 124346.3573 . . . . . . . . . . . . Outer 33550.84251 1067.528809 250832.3656 500021.3536 ring 304967.2219 102126.8471 304468.964 111880.8534 fault 43838.88916 1777.703363 263486.4018 198770.2839 98132.17038 10293.90837 345376.4988 125129.772 15721.8897 220.2773699 154778.8344 823747.1867 31047.03968 927.9619372 189236.481 558433.8182 364833.2182 146339.4 311658.1027 40278.05495 31417.06189 890.405166 226380.9981 283376.4123 119489.3583 15284.04784 308135.6096 75213.75259 19390.28923 337.3682025 231781.6402 542712.8916 . . . . . . . . . . . . Rolling 32006.03758 65929.32046 85929.32046 71558.32193 body 22964.84243 52790.55577 82790.55577 106755.9463 fault 37342.35272 68767.34954 68767.34954 152667.6245 6823.466055 128615.4622 88615.46224 128980.9664 17381.76733 59822.61048 59822.61048 184981.4427 32478.05926 80528.12418 80528.12418 78398.04682 27033.24127 95407.2903 95407.2903 59783.62418 29682.17889 100070.8356 100070.8356 70414.60027 5431.371807 120560.1302 90560.13017 26338.94067 18827.47196 105781.3286 85781.32856 114795.5821 . . . . . . . . . . . . Holder 852.862754 131.9393834 16565.02676 21157.6592 fau 732.893499 125.4413316 19390.1675 7947.86201 526.0584755 11.65033384 20419.58084 8951.588108 968.2235281 288.2783756 23451.28649 18489.0403 1039.45356 222.6292709 21930.897 14548.56436 687.4118986 188.2502815 25766.74535 13793.63693 777.5488247 180.5736243 25822.28709 8146.23067 328.6232912 12.24730239 23604.76025 10275.36466 802.8466152 153.6612797 24783.45146 8747.926443 736.6613112 129.4124587 25276.98018 19715.10467 . . . . . . . . . . . .

[0088] The model is solved according to experiment data in Table 2 and Table 3, the optimal weight combination between the four frequency-domain analysis methods in the data set is obtained, results are shown in Table 4, and the results are consistent with an evaluation of the diagnosis effect of the frequency-domain analysis method in FIG. 3, which may prove the effectiveness of the diagnosis method.

TABLE-US-00004 TABLE 4 Fast spectral Power Envelope kurtosis-square Method Spectrum spectrum spectrum envelope spectrum Weight 0.3192 0.4145 0 0.2663

[0089] A result after the weight obtained by solving the above experiment data fuses the experiment data is shown in FIG. 4, therefore, the experience range of the fusion data points of each fault type is summarized and shown in Table 5, wherein the experience ranges of the fusion data points of the rolling body fault and the holder fault does not overlap with that of other fault types, it may be directly judged according to the fusion data of experiment data, and the experience ranges of the inner ring fault and the outer ring fault overlap in the range of 3.110.sup.53.610.sup.5, and therefore, the fault type of the newly input experiment data may be judged by the center distance between its fusion data points and the two classes of experience ranges, that is, it is judged according to an absolute value of a difference between the fusion data points of the newly input experiment data and 3.3510.sup.5 (inner ring fault average value and 2.9910.sup.5 (outer ring fault average value).

TABLE-US-00005 TABLE 5 Inner ring Outer ring Rolling body Holder Fault type fault fault fault fault Fusion data 3.8~3.1 2.7~3.6 1.1~0.9 0.3~0.02 point range (10.sup.5)

[0090] In one embodiment, 2 groups of fault data are selected as the data to be diagnosed to verify the effectiveness of the method, and specific data is as shown in Table 6.

TABLE-US-00006 TABLE 6 Serial number Fault type Data Test 1 Inner ring 0.64732 fault 0.93221 0.43775 0.086309 0.012307 0.074725 0.53524 0.18647 0.41768 0.23819 . . . Test 2 Rolling body 0.052208 fault 0.017828 0.49503 0.042393 0.24171 0.55142 0.034381 0.43031 0.12419 0.47757 . . .

[0091] FIG. 5 and FIG. 6 respectively show spectrum results after spectrum analysis of the spectrum, the power spectrum, the envelope spectrum, and the fast spectral kurtosis-square envelope spectrum on the above data to be diagnosed, it may be seen from the figures that the effect displayed by each method is consistent with an evaluation result in Table 3, with the best power spectrum effect and the highest weight, oppositely, the worst envelope spectrum effect and the lowest weight, and it is proved that that the method is effective. In FIG. 7, weight fusion is performed on the above spectra by the weights preferred in Table 4 to form the fusion spectrum. FIG. 7(a) and FIG. 7(b) may clearly observe the inner ring fault feature frequency and the rolling body fault feature frequency. Results after fusion on these according to the optimal weights found are shown in Table 7, it is judged according to the interval range in which the fusion results fall, and it may be accurately diagnosed as the inner ring fault and the rolling body fault.

TABLE-US-00007 TABLE 7 Data to be diagnosed Test 1 Test 2 Fault type Inner ring fault Rolling body fault Fusion data point numerical 3.62 10.sup.4 0.95 10.sup.4 value Fault discrimination result Inner ring fault Rolling body fault

[0092] In order to verify the diagnosis superiority of the fusion spectrum, two indexes of an average diagnosis effect and diagnosis stability are used to compare the present disclosure with two methods of the envelope spectrum and the spectrum. The results are shown in FIG. 8 and Table 8. It may be seen from FIG. 8 that the fusion spectrum of the present disclosure can not only obtain the better diagnosis results, but also maintain the stability of diagnosis in different data. Accordingly, the superiority of the present disclosure may be seen from Table 8.

[0093] The average diagnosis effect is defined as: average diagnosis effect=(CFIC values of all the frequency spectra and/spectrum number); and diagnosis stability is defined as: diagnosis stability=(variance of the CFIC values of all the frequency spectra).

TABLE-US-00008 TABLE 8 Average diagnosis effect Diagnosis stability Fusion spectrum 0.0057 4.3 10.sup.8 Envelope spectrum 0.0029 3.2 10.sup.7 Spectrum 0.0043 7.7 10.sup.6

[0094] Although the embodiments of the present disclosure are described above in combination with the accompanying drawings, the technical solutions of the present disclosure are not limited to two working conditions of different speeds and different loads, but also include other classes of working conditions. The above specific implementation solutions are only schematic and directive, rather than limitation. Under the inspiration of this specification and without deviating from the scope of protection of the claims of the present disclosure, those ordinarily skilled in the art may also make a variety of forms, which are included in the protection of the present disclosure.