CN105134619A - Failure diagnosis and health evaluation method based on wavelet power, manifold dimension reduction and dynamic time warping - Google Patents

Abstract

The invention discloses a failure diagnosis and health evaluation method based on wavelet power, manifold dimension reduction and dynamic time warping, and aims to improve the feature separability of bearing failure, impeller failure and the mixed failures of a centrifugal pump and realize diagnosis and health evaluation of various states. The method comprises the following steps: firstly, decomposing collected vibration signals of the centrifugal pump into 8 wavelet components by applying wavelet packet conversion; extracting wavelet energy of each component to be taken as a failure feature to obtain an eight-dimensional failure feature vector; then conducting dimension reduction on the eight-dimensional feature by applying a manifold learning method to obtain a three-dimensional feature vector with better separability, simplicity and stability; finally, based on the feature vector, measuring the distance of test data and training data by applying a dynamic time normalization method so as to determine the current failure state and realize failure diagnosis of a bearing. The distance value can also reflect the health degree of the current state, can realize evaluation of the health state of the centrifugal pump, and has the excellent practical engineering application value.

Description

A kind of fault diagnosis based on wavelet energy, manifold dimension-reducing and dynamic time warping and health evaluating method

Technical field

The present invention relates to the technical field of centrifugal pump fault diagnosis and health evaluating, be specifically related to a kind of fault diagnosis based on wavelet energy, manifold dimension-reducing and dynamic time warping (dynamictimewarping, DTW) and health evaluating method.

Background technique

Centrifugal pump is widely used in the industrial departments such as electric power, petrochemical industry, metallurgy, machinery, is the critical component that influential system runs well.Therefore, Accurate Diagnosis is carried out to centrifugal pump fault, Efficient Evaluation is carried out to centrifugal pump health status, for the stable operation important in inhibiting of industry equipment.Because centrifugal pump can produce vibration in rotary course, and the oscillating signal power produced under different conditions is also different, therefore, is the method for current extensive use based on the fault diagnosis of oscillating signal and health evaluating.In actual applications, the most common failure of centrifugal pump mainly concentrates on bearing or impeller, except single failure state also often has mixed fault state to exist, and collectable oscillating signal often has very strong nonlinear and nonstationary characteristic, make centrifugal pump fault diagnose and health evaluating more difficult.The process of centrifugal pump fault diagnosis and health evaluating mainly comprises fault signature extraction and fault or health status and determines two aspects.The inventive method is intended to extract the characteristic vector having more separability, with the validity of the accuracy and health evaluating that improve centrifugal pump fault diagnosis.

Extracting the matter of utmost importance having more separability fault signature is the bearing vibration signal how processing nonlinear and nonstationary.Single time domain or frequency-domain analysis method are inapplicable in this case.Wavelet transformation is a kind of Time-Frequency Analysis Method, the transient signal of nonlinear and nonstationary is had to the feature of wideband response, when low frequency, high and low to the resolution of time to the resolution of frequency, and when high frequency, low and high to the resolution of time to the resolution of frequency.It is consistent for changing fast feature when changing slow, high frequency when this feature and actual vibration signal low frequency, and therefore wavelet transformation achieves good effect for the process of oscillating signal.Wavelet packet analysis, on the basis of wavelet transformation, can carry out more careful analysis and reconstruct by signal, decomposes simultaneously, be more effectively extracted the time-frequency characteristics of signal than wavelet transformation to the low frequency of signal and HFS.Therefore, select wavelet packet analysis to decompose original vibration signal in the present invention, to obtain effective signal time-frequency characteristics.Because the oscillating signal intensity under the various health status of centrifugal pump is different, the energy of the Wavelet Component that each frequency band obtained after WAVELET PACKET DECOMPOSITION is corresponding also can be different, therefore, can extract the energy value composition fault feature vector of each Wavelet Component.But, the characteristic vector extracted thus higher-dimension often, very effectively can not reflect the difference between each fault state feature, and high dimensional feature directly can increase its complexity greatly as the input vector of consequent malfunction classification or health evaluating algorithm, therefore, dimensionality reduction carries out to high dimensional feature necessary.

2000, Seung and Lee delivered the paper that a section is entitled as " TheManifoldwaysofperception " on " Science ", opens the epoch of manifold learning.From Differential Geometry angle, the live part of signal is often distributed on the low dimensional manifold in higher dimensional space, and obtaining signal characteristic on low dimensional manifold can faults information better.Manifold learning arithmetic realizes the dimensionality reduction to high dimensional data by the inherent low dimensional structures in discovery high dimensional data.At present, manifold learning has obtained study and practice deeply and widely, defines a lot of classical method.Its application area relates to the fields such as recognition of face, visual information analysis, finger vena identification, pattern recognition.The effect of 6 kinds of Method of Nonlinear Dimensionality Reductions is compared in the present invention, comprise core principle component analysis (kernelprincipalcomponentanalysis, KPCA) method, laplacian eigenmaps (LaplacianEigenmaps, LE) method, local linear embeds (locallinearembedding, LLE) method, based on the LLE method (HLLE) of Hessian, local tangent space alignment (localtangentspacealignment, LTSA) method and linear local tangent space alignment (linearlocaltangentspacealignment, LLTSA) method.The Gray-level co-occurrence of higher-dimension vector, by manifold learning dimensionality reduction, is reduced to and is had more separability, more brief stable characteristic vector by the inventive method.

For fault diagnosis and health evaluating, key is the similarity of measuring exactly between test data and sample data.Dynamic time warping (dynamictimewarping, DTW) method is set forth in 1978, is the problem in order to solve speech recognition at first.Then, as a kind of mode-matching technique, DTW obtains application at a lot of other field, as fingerprint authentication, Activity recognition, on-line signature checking, data mining, computer vision and computer animation, process monitoring and fault diagnosis etc.Compared with other method for mode matching, DTW is simple, easy, has good real-time capacity.Therefore, the present invention's application DTW method measures the similarity between characteristic under state to be measured and each state sample characteristic, thus determines the health degree index of current fault state or current state.

Summary of the invention

The present invention wants technical solution problem to be: overcome the deficiencies in the prior art, a kind of fault diagnosis based on wavelet energy, manifold dimension-reducing and dynamic time warping and health evaluating method are provided, in order to extract the fault feature vector having more separability, and the distance of measuring quickly and accurately between test data and training data, thus determine current fault state, the health degree index of assessment current state, realizes diagnosis and the health evaluating of centrifugal pump typical fault.

The technical solution used in the present invention is: a kind of fault diagnosis based on wavelet energy, manifold dimension-reducing and dynamic time warping and health evaluating method, and step is as follows:

Step (1), application analysis method of wavelet packet decompose original vibration signal, obtain several small echo component of signals;

Step (2), for each small echo component of signal, extract its wavelet energy value composition fault feature vector, carry out faults information by the vibrational energy under various health status;

Step (3), for extract higher-dimension Gray-level co-occurrence vector, application manifold learning carry out Feature Dimension Reduction, have more separability, more brief stable fault feature vector to obtain;

Step (4), based on the three-dimensional fault feature vector extracted, application DTW measures the similarity between test data and training data, thus determines or assess fault corresponding to current data or health status, thus realizes fault diagnosis and health evaluating.

Further, described step (1) is specially: application analysis method of wavelet packet carries out three layers of wavelet decomposition to the original vibration signal x (t) of centrifugal pump nonlinear and nonstationary, obtains 8 small echo component of signals.

Further, described step (2) is specially: to each small echo component of signal, extracts wavelet energy value composition fault feature vector, to reflect the fault message of each fault state.Process is as follows:

Step (A1), set original vibration signal as x (t), through three layers of wavelet decomposition process, x (t) is broken down into 8 Wavelet Component, can be expressed as (x from low to high ¹, x ², x ³, x ⁴, x ⁵, x ⁶, x ⁷, x ⁸);

Step (A2), to each Wavelet Component x ⁱ, the wavelet energy of its correspondence is in formula, i=1,2 ..., 8; K=1,2 ..., N, x _ikfor reconstruction signal x ⁱthe amplitude of discrete point.Thus, the fault feature vector W=[E of wavelet energy composition is obtained ¹, E ²..., E ⁸], vectorial W is exactly a fault feature vector of original vibration signal.

Further, described step (3) is specially: application manifold learning carries out dimensionality reduction to the higher-dimension Gray-level co-occurrence vector extracted, and has more separability, more brief stable fault feature vector to obtain.

Further, described step (4) is specially: on the basis of the three-dimensional fault feature vector extracted, application DTW method calculates the distance between test data and each sample data, and then judges the health status of current data, thus realizes fault diagnosis and the health evaluating of bearing.Process is as follows:

Step (B1), first, to the original vibration signal under various health status, carry out WAVELET PACKET DECOMPOSITION and extract Gray-level co-occurrence vector, dimensionality reduction is being carried out to high dimensional feature vector, as sample characteristics matrix during follow-up health status classification, if the data of total k kind health status, then this sample characteristics matrix V=[W ₁, W ₂..., W _k], wherein W _iit is the characteristic vector of i-th kind of health status;

Step (B2), then, for the oscillating signal of arbitrary state to be determined, by WAVELET PACKET DECOMPOSITION signal, extracts Gray-level co-occurrence vector and dimensionality reduction;

Step (B3), application DTW algorithm measure the similarity of each characteristic vector in the characteristic vector of state to be determined and sample characteristics matrix, metric is less, prove that the state of current state to be determined and this label characteristics vector is more close, thus determine the health status of current data.Health degree corresponding for eigenvalue under normal state is set to 1, then measures the health degree of current state by the degree of similarity between current data and normal sample notebook data.

The present invention's advantage is compared with prior art:

(1) the present invention is directed to centrifugal pump working condition complicated and changeable, the oscillating signal gathered is by force non-linear, single failure and mixed fault are also deposited, existing centrifugal pump fault diagnosis is few with health evaluating method, lack the present situation of overall procedure, propose the diagnosis of a kind of centrifugal pump fault and the effective ways of health evaluating, be extracted and have more separability, more brief stable fault signature, thus improve the effect of fault diagnosis and health evaluating.

(2), the present invention is directed to the feature of centrifugal pump vibration signal nonlinear and nonstationary non-gaussian, original vibration signal is decomposed several Wavelet Component by application analysis method of wavelet packet, obtains the high-low frequency weight of primary signal; And the wavelet energy value extracting each Wavelet Component is as fault signature, effectively reflects fault message by the vibrational energy under various health status.

(3) the present invention is directed to the higher-dimension Gray-level co-occurrence vector of extraction, application manifold learning carries out Feature Dimension Reduction, has more separability, more brief stable fault feature vector to obtain.

(4) the present invention's application DTW measures the similarity between test data and sample data, makes the process of fault state coupling and health evaluating more simple, improves operation efficiency, ensure that the ease for operation of centrifugal pump fault diagnosis and health evaluating.

Accompanying drawing explanation

Fig. 1 is centrifugal pump fault diagnosis and health evaluating method overall flow figure;

Fig. 2 is centrifugal pump data capture vibration transducer scheme of installation;

Fig. 3 is the original vibration signal under centrifugal pump normal state;

Fig. 4 is the wavelet decomposition result of centrifugal pump normal state oscillating signal;

Fig. 5 is the broken line graph (20 stack features vector) of the octuple Gray-level co-occurrence vector under centrifugal pump different conditions;

Fig. 6 is the dimensionality reduction effect contrast figure of various flows shape method;

Fig. 7 is the centrifugal pump fault diagnostic result figure based on DTW;

Fig. 8 is the centrifugal pump health evaluating result figure based on DTW.

Embodiment

The present invention is further illustrated below in conjunction with accompanying drawing and specific embodiment.

One of the present invention is based on the fault diagnosis of wavelet energy, manifold dimension-reducing and dynamic time warping (dynamictimewarping, DTW) and health evaluating method, and concrete steps are as follows:

1, WAVELET PACKET DECOMPOSITION and wavelet energy

(1) wavelet packet analysis

Wavelet packet analysis (WavalctPacketAnalysis, WPA) based on wavelet multi_resolution analysis, more careful analysis and reconstruct can be carried out to signal, the part that multiresolution analysis does not segment is decomposed further, namely the low frequency of signal and HFS is decomposed simultaneously; And according to analyzed signal characteristic, the resolution of signal at different frequency range can be determined adaptively, constitute complete tree.Wavelet packet analysis, owing to having good signal partial analysis ability, makes it effectively can analyze nonlinear properties.

Signal and one all have the companding wavelet function of good local character to carry out convolution in time domain and frequency domain by wavelet transformation, are each composition being positioned at different frequency bands and period signal decomposition.Explanation about theory of wavelet transformation has introduction in a lot of documents and materials, just repeats no more here.The basic thought of wavelet theory is: the different signal of nature various signal intermediate frequency rate height has different time-varying characteristics, the spectrum signature of usual lower frequency composition is relatively slower over time, and the spectrum signature of upper frequency composition then changes rapider.Therefore, by such rule anisotropically time division and frequency axis, just under the prerequisite of obeying uncertainty principle, proper time resolution and frequency resolution can be obtained in different time-frequency region.But in the decomposition of wavelet transformation, only the approximation coefficient that last time decomposes is decomposed at every turn, and the detail coefficients that last time decomposes no longer is decomposed, cause the frequency resolution of small scale to can not get improving; And cause time resolution to can not get improving when large scale.Thus, in order to the signal of more accurate analysis different frequency range, wavelet packet analysis is created.

Wavelet packet analysis passes through orthogonal scaling function with wavelet function ψ (x), signal is decomposed, obtain low frequency and the HFS of signal, and be by wave filter h and g, the decomposition of discrete approximation coefficient has been come in sequence space.Their two scaling relations are:

$\mathrm{\φ}\left(t\right)=\sqrt{2}\underset{k}{\Σ}h\left(k\right)\mathrm{\φ}(2t-k)$ (1)

$\mathrm{\ψ}\left(t\right)=\sqrt{2}\underset{k}{\Σ}g\left(k\right)\mathrm{\φ}(2t-k)$

For further genralrlization two-scale equation, define following recurrence relation:

$w_{2n}\left(t\right)=\sqrt{2}\underset{k\∈Z}{\Σ}h\left(k\right)w_{2n}(2t-k)$ (2)

$w_{2n+1}\left(t\right)=\sqrt{2}\underset{k\∈Z}{\Σ}g\left(k\right)w_n(2t-k)$

Wherein, n=0,1,2 ... the sequence number of representative function.

(2) wavelet energy calculates

Due under different faults state, the intensity of centrifugal pump vibration signal in different frequency bands is different, therefore the power of different frequency bands internal vibration signal can be considered as fault signature.By wavelet packet analysis, original vibration signal is broken down into several high-low frequency weight, can calculate the frequency band energy value of each Wavelet Component as fault eigenvalue.For three layers of WAVELET PACKET DECOMPOSITION, 8 frequency band energy E _3jformula as follows:

$E_{3j}=\∫{\left|S_{3j}\left(t\right)\right|}^{2}dt=\underset{k=1}{\overset{n}{\Σ}}{\left|x_{jk}\right|}^2---\left(3\right)$

Wherein, x _jk(j=0,1 ..., 7; K=1,2 ..., n) represent reconstruction signal S _3jdiscrete point amplitude.

To arbitrary fault state, 8 wavelet energy value composition of vector W=[E can be had ₃₀, E ₃₁..., E ₃₇] as the characteristic vector of this fault state.

2, manifold learning

Manifold learning term is that nineteen ninety-five Bregler and Omohundro proposes first when studying visual speech identification, but manifold learning is really furtherd investigate and developed from 3 sections of papers delivered by Science magazine in 2000.There is scholar that principal component analysis, independent component analysis, Fisher discriminant analysis etc. are also classified as manifold learning, be called linear manifold learning method, and the method proposed afterwards for 2000 is referred to as non-linearity manifold study method.Linear manifold learning method, when Dimensionality Reduction, be difficult to the high dimensional nonlinear data in accurate analysis practical application, and non-linearity manifold learning method easily can address this problem.The essence of manifold learning dimension reduction method is in high-dimensional data space, find low dimensional manifold structure, by minimizing the error between high dimensional data and low-dimensional data, removing different classes of interactional priori, and then realizing Dimensionality Reduction.

The object applying manifold learning dimensionality reduction in the present invention, mainly by nonlinear high dimensional feature vector dimensionality reduction, obtains and has more separability, more brief stable fault feature vector.In order to contrast the practical application effect of each dimension reduction method, the effect of 6 kinds of Method of Nonlinear Dimensionality Reductions is compared in the present invention, comprise core principle component analysis (kernelprincipalcomponentanalysis, KPCA) method, laplacian eigenmaps (LaplacianEigenmaps, LE) method, local linear embeds (locallinearembedding, LLE) method, based on the LLE method (HLLE) of Hessian, local tangent space alignment (localtangentspacealignment, LTSA) method and linear local tangent space alignment (linearlocaltangentspacealignment, LLTSA) method.Because these methods have had very detailed introduction in a lot of documents and materials, here just repeat no more.

3, dynamic time warping method

Dynamic time warping (dynamictimewarping, DTW) be by Sakoe and Chiba be speech recognition propose method for mode matching, then have also been obtained extensive application at other field.Based on dynamic programming techniques, DTW, by time series is carried out extending and shortening, calculates the beeline between two time serieses, and then realizes similarity measurement.The principles illustrated of DTW algorithm is as follows:

For two sequence C=c ₁, c ₂..., c _i..., c _mand Q=q ₁, q ₂..., q _j..., q _n, the distance d (C between them between corresponding element _i, Q _j) can be calculated by a distance function, thus obtain the distance matrix of a n × m.In traditional DTW algorithm, distance function is Euclidean distance square.Then, by making Cumulative Distance minimum, a regular path U=(u can be determined ₁, u ₂..., u _k..., u _l), wherein max (m, n)≤L≤m+n-1.Some local restrictive conditions of this paths demand fulfillment, such as:

A () end points limits: the terminal in this path should correspond to first point and last point of distance matrix, ensures that the sequencing of sequence does not change, that is, u ₁=(c ₁, q ₁), u _l=(c _m, q _n).

B () continuity limits: each time, path can only take a step forward, the continuous print that the process of coupling is necessary, across Point matching, that is, can not work as u _k=(c _i, q _j), u _k+1=(c _i+1, q _j+1), then there is c _i+1-c _i≤ 1, q _j+1-q _j≤ 1.

C () monotonicity limits: matching process carries out along sequence dullness, that is, work as u _k=(c _i, q _j), u _k+1=(c _i+1, q _j+1), then there is c _i≤ c _i+1, q _j≤ q _j+1.

Finally, DTW Cumulative Distance is defined as:

$DTW(C_i,Q_j)=d(C_i,Q_j)+min\left\{\begin{array}{c}DTW(C_i,Q_{j-1})\\ DTW(C_{i-1},Q_j)\\ DTW(C_{i-1},Q_{j-1})\end{array}\right.---\left(4\right)$

In actual applications, calculate all possible path and expend time in very much, nor necessary, the overall situation restriction of therefore applying regular path in the matching process reduces the path of calculating.

Because in traditional DTW algorithm, distance function is Euclidean distance square, it treats the characteristic vector of all dimensions coequally but in fact these features are unequal.In order to address this problem, calculating distance before can first establishing criteria formula to sequence C=c ₁, c ₂..., c _i..., c _mand Q=q ₁, q ₂..., q _j..., q _ncarry out standardization.Standardization formula is as follows:

${x_i}^{*}=\frac{x_i-m}{s}---\left(5\right)$

Wherein, x _i ^*be the point value after standardization, m is the average of sequential element, and s is the standard deviation of sequence.

By the standardization of function of adjusting the distance, in similarity measurement, when test data is simultaneously all smaller with several different classes of distance, we wish to strengthen the may differentiate between these distances, thus obtain better classifying quality.

4, based on centrifugal pump fault diagnosis and the health evaluating method of wavelet energy, manifold dimension-reducing and DTW

The centrifugal pump fault diagnosis that the present invention proposes and health evaluating method overall flow are as shown in Figure 1.Concrete step is as follows:

(1) first, application analysis method of wavelet packet decomposes original vibration signal, obtains 8 small echo sub-signal components by three layers of wavelet decomposition;

(2) then, to each Wavelet Component, extract wavelet energy value composition fault feature vector, carry out faults information by the vibrational energy under various health status;

(3) for the higher-dimension Gray-level co-occurrence extracted vector, apply manifold learning and carry out Feature Dimension Reduction, have more separability, more brief stable fault feature vector to obtain;

(4) last, based on the three-dimensional fault feature vector extracted, application DTW measures the similarity between test data and training data, thus determines or assess fault corresponding to current data or health status, thus realizes fault diagnosis and health evaluating.

Application example is as follows:

1, centrifugal pump Data Source

In order to verify the validity of put forward the methods of the present invention, the method validation result based on laboratory centrifugal failure of pump data will be shown below.This centrifugal pump is typical single-stage and self-priming centrifugal pump, is mainly oiling and provides transmitting power.In signals collecting, acceleration transducer is arranged on directly over motor housing bearing support, and sensor is secured by bolts on special base, and base is adhered to motor housing, as shown in Figure 2.In order to obtain the data under different faults state, carried out direct fault location to centrifugal pump, the fault of injection is respectively: bearing inner ring fault, outer race fault, bearing roller fault, impeller failure, bearing inner ring and impeller mixed fault, outer race and impeller mixed fault.The sample frequency 10.24KHz of oscillating signal, often organizing acquisition time is 2s, and gather one group of data every 5s, often kind of fault state all gathers 20 groups of data.Oscillating signal under the normal state gathered as shown in Figure 3.

2, the fault signature based on wavelet energy and manifold dimension-reducing extracts

In the present invention, application wavelet energy and manifold learning combine, and extract the fault characteristic information in bearing vibration signal.

First, signal decomposition is 8 Wavelet Component by three layers of wavelet decomposition by application analysis method of wavelet packet process original vibration signal; In the present invention, in order to obtain good performance of fault diagnosis, primary signal is divided into some parts, every part comprises 5000 points for signal decomposition.Wavelet decomposition result under normal state as shown in Figure 4.

Then, calculate the wavelet energy value composition fault feature vector of each Wavelet Component, the broken line graph of the octuple Gray-level co-occurrence vector under 20 groups of different conditions as shown in Figure 5.As can be seen from the figure, the fault feature vector of wavelet energy composition is higher-dimension, and the fault message of expression is too complicated, is unfavorable for the calculating of follow-up diagnosis, assessment algorithm.Thus, manifold dimension-reducing is carried out to this high dimensional feature vector.In order to contrast the quality of each manifold dimension-reducing method, obtain best dimension reduction method, the dimensionality reduction effect of the present invention to KernelPCA, LaplacianEigenmaps, LLE, HLLE, LTSA, LLTSA six kinds of methods contrasts, and result as shown in Figure 6.As can be seen from the figure, in these six kinds of methods, the dimensionality reduction effect of LLTSA method is best, very clearly by the characteristic area under various different faults state separately, and can all there is mode aliasing in various degree in the dimensionality reduction result of other 5 kinds of methods.Fault feature vector based on wavelet energy-LLTSA has obvious separability, and the Clustering Effect of same status flag is fine, ensure that the accuracy of consequent malfunction diagnosis and health evaluating.In table 1, citing lists the eigenvalue based on wavelet energy-LLTSA under each health status.

Table 1 is based on the fault eigenvalue example table of wavelet energy-LLTSA

3, the fault state based on DTW is determined and health state evaluation

(1) fault state based on DTW is determined

Based on the fault feature vector that wavelet energy-LLTSA extracts, application DTW tolerance test data and sample data concentrate the distance between each exemplar data, the label of the sample data that the health status of current test data is corresponding with lowest distance value is consistent, thus the fault state of current data can be determined, realize fault diagnosis.

First, design sample File.In fault diagnosis, this centrifugal pump has 7 kinds of health status, and comprise normal state and 6 kinds of fault states, therefore, sample data concentrates the characteristic vector comprising 7 kinds of state tags, and often kind of label comprises 5 groups of data, and details are as shown in table 2.

Then, in order to the diagnosis performance of verification algorithm, prepared 7 groups of test datas altogether, corresponded respectively to 7 kinds of health status, often group test data is exactly one group of fault feature vector under certain state.

Finally, application DTW measures the similarity between test data and sample data, result as shown in Figure 7, as can be seen from the figure, each test data and its sample data spacing distance values with state tag are almost 0, and larger with the distance value between the sample data of different conditions label, very clearly can determine the fault state belonging to each test data.

The details of table 2 sample data collection

(2) based on the health state evaluation of DTW

Based on the fault feature vector that wavelet energy-LLTSA extracts, application DTW measures the distance between test data and normal state sample data, distance value larger explanation current state and normal state departure degree higher, namely the fault of this state is more serious, thus can assess the health degree index of current state.If the distance value between current test data and normal state sample data is d _i, definition R=1/ (d _i+ 1) be health degree index, and make the healthy angle value of normal state be 1, then can calculate the health degree index of arbitrary fault state, realize health evaluating.

In order to verify the validity of the health evaluating method that the present invention proposes, calculate the distance between characteristic vector under 7 kinds of fault states and the characteristic vector under normal state, and pass through the health degree of these fault states of health degree index calculate formulae discovery, result is as shown in Figure 8.As can be seen from the figure, the healthy angle value difference that different faults is corresponding, clearly can assess the health degree of current state based on DTW method.

In sum, the fault diagnosis based on wavelet energy, manifold dimension-reducing and dynamic time warping that the present invention proposes and health evaluating method, achieve good effect in the diagnosis and assessment of centrifugal pump mixed fault.

Non-elaborated part of the present invention belongs to the known technology of those skilled in the art.

Claims (5)

1., based on fault diagnosis and the health evaluating method of wavelet energy, manifold dimension-reducing and dynamic time warping, it is characterized in that: the method implementation step is as follows:

Step (1), application Wavelet Packet Transform Method decompose original vibration signal, obtain 8 Wavelet Component;

Step (2), calculate the wavelet energy of each Wavelet Component as fault signature, obtain octuple fault feature vector;

Step (3), application manifold learning carry out dimensionality reduction to fault feature vector, obtain the three-dimensional fault feature vector having more separability;

Step (4), based on extract three-dimensional fault feature vector, the similarity between test data and training data is measured in application dynamic time warping (DTW), thus determine or assess fault corresponding to current data or health status, thus realize fault diagnosis and health evaluating.

2. a kind of fault diagnosis based on wavelet energy, manifold dimension-reducing and dynamic time warping according to claim 1 and health evaluating method, it is characterized in that: described step (1) application Nonlinear harmonic oscillator method---WAVELET PACKET DECOMPOSITION, centrifugal pump vibration signal is decomposed, obtains 8 Wavelet Component.

3. a kind of fault diagnosis based on wavelet energy, manifold dimension-reducing and dynamic time warping according to claim 1 and health evaluating method, is characterized in that: the wavelet energy value that described step (2) calculates each Wavelet Component is as follows as the process of fault signature:

Step (A1), set original vibration signal as x (t), through three layers of WAVELET PACKET DECOMPOSITION process, in first layer, x (t) is broken down into a high-frequency signal and a low frequency signal, then, each sub-signal carries out low-and high-frequency decomposition again, finally obtains 8 Wavelet Component, and the Wavelet Component of 8 is from low to high expressed as (x ¹, x ², x ³, x ⁴, x ⁵, x ⁶, x ⁷, x ⁸);

Step (A2), calculate the energy of each Wavelet Component, to each Wavelet Component x ⁱ, the wavelet energy of its correspondence is in formula, i=1,2 ..., 8; K=1,2 ..., N, x _ikfor reconstruction signal x ⁱthe amplitude of discrete point, thus, obtains the fault feature vector W=[E of wavelet energy composition ¹, E ²..., E ⁸].

4. a kind of fault diagnosis based on wavelet energy, manifold dimension-reducing and dynamic time warping according to claim 1 and health evaluating method, it is characterized in that: in described step (3), because the Gray-level co-occurrence vector extracted is octuple, in order to effectively analyze the data characteristics in higher dimensional space, application manifold learning carries out dimensionality reduction to octuple characteristic vector, obtains the three-dimensional stability fault feature vector having more separability.

5. a kind of fault diagnosis based on wavelet energy, manifold dimension-reducing and dynamic time warping according to claim 1 and health evaluating method, it is characterized in that: described step (4) is based on the three-dimensional fault feature vector extracted, the similarity between test data and training data is measured in application dynamic time warping (DTW), thus determine the health status that current data is corresponding, the detailed process realizing failure modes and health evaluating is as follows:

Step (B1), first, to the original vibration signal under various health status, carry out WAVELET PACKET DECOMPOSITION and extract Gray-level co-occurrence vector, dimensionality reduction is being carried out to high dimensional feature vector, as sample characteristics matrix during follow-up health status classification, if the data of total k kind health status, then this sample characteristics matrix V=[W ₁, W ₂..., W _k], wherein W _iit is the characteristic vector of i-th kind of health status;

Step (B2), then, for the oscillating signal of arbitrary state to be determined, by WAVELET PACKET DECOMPOSITION signal, extracts Gray-level co-occurrence vector and dimensionality reduction;

Step (B3), application DTW algorithm measure the similarity of each characteristic vector in the characteristic vector of state to be determined and sample characteristics matrix, metric is less, prove that the state of current state to be determined and this label characteristics vector is more close, thus determine the health status of current data, health degree corresponding for eigenvalue under normal state is set to 1, then measures the health degree of current state by the degree of similarity between current data and normal sample notebook data.