CN105134619B - A kind of fault diagnosis based on wavelet energy, manifold dimension-reducing and dynamic time warping and health evaluating method - Google Patents

Abstract

The invention discloses a fault diagnosis and health assessment method based on wavelet energy, manifold dimensionality reduction and dynamic time warping, so as to improve the feature separability of bearing faults, impeller faults and their mixed faults in centrifugal pumps, and realize various states Diagnosis and health assessment. First, apply the wavelet packet transform to decompose the collected centrifugal pump vibration signal into 8 wavelet components; extract the wavelet energy of each component as the fault feature, and obtain the eight-dimensional fault feature vector; then, apply the manifold learning method to the eight-dimensional The feature is dimensionally reduced to obtain a more separable, more concise and stable three-dimensional feature vector; finally, based on the feature vector, the dynamic time warping method is applied to measure the distance between the test data and the training data, so as to determine the current fault state. Realize bearing fault diagnosis; at the same time, the distance value can also reflect the health of the current state, and realize the evaluation of the health state of the centrifugal pump, which has good practical engineering application value.

Description

A Fault Diagnosis and Analysis Based on Wavelet Energy, Manifold Dimensionality Reduction and Dynamic Time Warping health assessment method

technical field

The invention relates to the technical field of fault diagnosis and health assessment of centrifugal pumps, in particular to a fault diagnosis and health assessment method based on wavelet energy, manifold dimensionality reduction and dynamic time warping (DTW).

Background technique

Centrifugal pumps are widely used in electric power, petrochemical, metallurgy, machinery and other industrial sectors, and are key components that affect the normal operation of the system. Therefore, accurate diagnosis of centrifugal pump faults and effective assessment of the health status of centrifugal pumps are of great significance for the stable operation of industrial equipment. Since the centrifugal pump will vibrate during the rotation process, and the strength of the vibration signal generated in different states is also different. Therefore, the fault diagnosis and health assessment based on the vibration signal are widely used methods at present. In practical applications, the common faults of centrifugal pumps are mainly concentrated in bearings or impellers. In addition to a single fault state, there are often mixed fault states, and the vibration signals that can be collected often have strong nonlinear and non-stationary characteristics, which makes the centrifugal pump failure Diagnosis and health assessment are more difficult. The process of centrifugal pump fault diagnosis and health assessment mainly includes fault feature extraction and fault or health status determination. The method of the invention aims to extract more separable feature vectors, so as to improve the accuracy of fault diagnosis and the effectiveness of health assessment of centrifugal pumps.

The primary problem in extracting more separable fault features is how to deal with nonlinear and non-stationary bearing vibration signals. A single time domain or frequency domain analysis method is not applicable in this case. Wavelet transform is a time-frequency analysis method. It has the characteristics of wide-frequency response to nonlinear and non-stationary transient signals. At low frequencies, it has a high resolution to frequency and low resolution to time. The frequency resolution is low and the time resolution is high. This feature is consistent with the characteristics of the actual vibration signal that changes slowly at low frequencies and changes quickly at high frequencies, so wavelet transform has achieved good results in processing vibration signals. Wavelet packet analysis, on the basis of wavelet transform, can analyze and reconstruct the signal in more detail, decompose the low-frequency and high-frequency parts of the signal at the same time, and extract the time-frequency characteristics of the signal more effectively than wavelet transform. Therefore, in the present invention, wavelet packet analysis is selected to decompose the original vibration signal to obtain effective signal time-frequency characteristics. Because the vibration signal strength of the centrifugal pump is different in various health states, the energy of the wavelet components corresponding to each frequency band obtained after wavelet packet decomposition will also be different. Therefore, the energy value of each wavelet component can be extracted to form a fault feature vector. However, the feature vectors extracted from this are often high-dimensional, which cannot effectively reflect the differences between the characteristics of each fault state, and the high-dimensional features directly used as input vectors for subsequent fault classification or health assessment algorithms will greatly increase their complexity. Therefore, it is necessary to reduce the dimensionality of high-dimensional features.

In 2000, Seung and Lee published a paper entitled "The Manifold ways of perception" in "Science", which opened the era of manifold learning. From the perspective of differential geometry, the effective part of the signal is often distributed on the low-dimensional manifold in the high-dimensional space, and obtaining the signal features on the low-dimensional manifold can better reflect the fault information. Manifold learning algorithms achieve dimensionality reduction for high-dimensional data by discovering the inherent low-dimensional structure in high-dimensional data. At present, manifold learning has been deeply and widely applied and researched, and many classic methods have been formed. Its application range involves face recognition, visual information analysis, finger vein recognition, pattern recognition and other fields. In the present invention, the effects of six nonlinear dimensionality reduction methods are compared, including kernel principal component analysis (kernel principal component analysis, KPCA) method, Laplacian Eigenmaps (Laplacian Eigenmaps, LE) method, local linear embedding (local linear embedding, LLE) method, Hessian-based LLE method (HLLE), local tangent space alignment (LTSA) method and linear local tangent space alignment (linear local tangent space alignment, LLTSA) method. The method of the invention reduces the dimension through the manifold learning method, and reduces the high-dimensional wavelet energy feature vector to a more separable, simple and stable feature vector.

For fault diagnosis and health assessment, the key is to accurately measure the similarity between test data and sample data. The dynamic time warping (DTW) method was proposed in 1978, initially to solve the problem of speech recognition. Then, as a pattern matching technology, DTW has been applied in many other fields, such as fingerprint verification, behavior recognition, online signature verification, data mining, computer vision and computer animation, process monitoring and fault diagnosis, etc. Compared with other pattern matching methods, DTW is simple, easy and has better real-time capability. Therefore, the present invention uses the DTW method to measure the similarity between the feature data in the state to be tested and the feature data of each state sample, so as to determine the current fault state or the health index of the current state.

Contents of the invention

The technical problem to be solved by the present invention is: to overcome the deficiencies of the prior art, to provide a fault diagnosis and health assessment method based on wavelet energy, manifold dimension reduction and dynamic time warping, to extract more separable fault feature vectors, And quickly and accurately measure the distance between the test data and the training data, so as to determine the current fault state, evaluate the health index of the current state, and realize the diagnosis and health assessment of typical faults of centrifugal pumps.

The technical solution adopted by the present invention is: a fault diagnosis and health assessment method based on wavelet energy, manifold dimension reduction and dynamic time warping, the steps are as follows:

Step (1), applying the wavelet packet analysis method to decompose the original vibration signal, obtains several wavelet signal components;

Step (2), for each wavelet signal component, extract its wavelet energy value to form a fault feature vector, and reflect the fault information through the vibration energy under various health states;

Step (3), aiming at the extracted high-dimensional wavelet energy feature vector, apply the manifold learning method to reduce the feature dimension, so as to obtain a more separable, more simple and stable fault feature vector;

Step (4), based on the extracted three-dimensional fault feature vector, apply DTW to measure the similarity between the test data and the training data, so as to determine or evaluate the fault or health state corresponding to the current data, so as to realize fault diagnosis and health assessment.

Further, the step (1) specifically includes: using the wavelet packet analysis method to perform three-layer wavelet decomposition on the nonlinear and non-stationary original vibration signal x(t) of the centrifugal pump to obtain 8 wavelet signal components.

Further, the step (2) specifically includes: for each wavelet signal component, extracting the wavelet energy value to form a fault feature vector to reflect the fault information of each fault state. The process is as follows:

Step (A1), assuming that the original vibration signal is x(t), after three layers of wavelet decomposition processing, x(t) is decomposed into 8 wavelet components, which can be expressed as (x ¹ ,x ² ,x ³ ,x ⁴ ,x ⁵ ,x ⁶ ,x ⁷ ,x ⁸ );

Step (A2), for each wavelet component x ⁱ , its corresponding wavelet energy is In the formula, i=1,2,...,8; k=1,2,...,N, x _ik is the amplitude of the discrete point of the reconstructed signal x ⁱ . Thus, the fault feature vector W=[E ¹ , E ² ,...,E ⁸ ] composed of wavelet energy is obtained, and the vector W is a fault feature vector of the original vibration signal.

Further, the step (3) specifically includes: applying a manifold learning method to reduce the dimensionality of the extracted high-dimensional wavelet energy feature vectors, so as to obtain more separable, simple and stable fault feature vectors.

Further, the step (4) is specifically: on the basis of the extracted three-dimensional fault feature vector, apply the DTW method to calculate the distance between the test data and each sample data, and then judge the health status of the current data, so as to realize the bearing fault diagnosis and health assessment. The process is as follows:

Step (B1), first, perform wavelet packet decomposition on the original vibration signals in various health states and extract wavelet energy feature vectors, and perform dimensionality reduction on the high-dimensional feature vectors as the sample feature matrix for subsequent health state classification, Suppose there are k kinds of health state data, then the sample feature matrix V=[W ₁ ,W ₂ ,…,W _k ], where W _i is the feature vector of the i-th health state;

Step (B2), then, for any vibration signal to be determined, the signal is decomposed by the wavelet packet, and the wavelet energy feature vector is extracted and dimensionally reduced;

Step (B3), applying the DTW algorithm to measure the similarity between the eigenvector of the state to be determined and each eigenvector in the sample feature matrix, the smaller the measurement value, the closer the current state to be determined is to the state of the label eigenvector, thereby determining The health status of the current data. If the health degree corresponding to the characteristic value in the normal state is set to 1, the health degree of the current state can be measured by the degree of similarity between the current data and the normal sample data.

The advantage of the present invention compared with prior art is:

(1) The present invention is aimed at the complex and changeable working conditions of centrifugal pumps, the collected vibration signals are strongly nonlinear, single faults and mixed faults coexist, and the existing methods for fault diagnosis and health assessment of centrifugal pumps are few and lack a complete set of methods. An effective method for fault diagnosis and health assessment of centrifugal pumps is proposed, and more separable, simple and stable fault features are extracted, thereby improving the effects of fault diagnosis and health assessment.

(2), the present invention is aimed at the non-linear non-stationary non-Gaussian characteristics of the centrifugal pump vibration signal, and uses the wavelet packet analysis method to decompose the original vibration signal into several wavelet components, and obtains the high and low frequency components of the original signal; and extracts each wavelet component The wavelet energy value of is used as the fault feature, and the fault information is effectively reflected by the vibration energy in various health states.

(3) For the extracted high-dimensional wavelet energy feature vector, the present invention applies a manifold learning method to perform feature dimensionality reduction to obtain more separable, simple and stable fault feature vectors.

(4) The present invention uses DTW to measure the similarity between test data and sample data, so that the process of fault state matching and health assessment is simpler, the operation efficiency is improved, and the operability of centrifugal pump fault diagnosis and health assessment is ensured.

Description of drawings

Figure 1 is the overall flow chart of the centrifugal pump fault diagnosis and health assessment method;

Figure 2 is a schematic diagram of the installation of the centrifugal pump data acquisition vibration sensor;

Figure 3 is the original vibration signal of the centrifugal pump in normal state;

Figure 4 is the wavelet decomposition result of the vibration signal of the centrifugal pump in normal state;

Fig. 5 is the line diagram (20 groups of feature vectors) of the eight-dimensional wavelet energy eigenvector under the different states of the centrifugal pump;

Figure 6 is a comparison of dimensionality reduction effects of different manifold methods;

Figure 7 is a diagram of the fault diagnosis results of a centrifugal pump based on DTW;

Figure 8 is a graph of the health assessment results of centrifugal pumps based on DTW.

detailed description

The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

A kind of fault diagnosis and health assessment method based on wavelet energy, manifold dimension reduction and dynamic time warping (dynamic time warping, DTW) of the present invention, concrete steps are as follows:

1. Wavelet packet decomposition and wavelet energy

(1) Wavelet packet analysis

Wavelet Packet Analysis (WPA) is based on wavelet multi-resolution analysis, which can analyze and reconstruct the signal in more detail, and further decompose the part that has not been subdivided in multi-resolution analysis, that is, the low frequency of the signal Decompose simultaneously with the high-frequency part; and according to the characteristics of the analyzed signal, it can adaptively determine the resolution of the signal in different frequency bands, forming a complete tree structure. Wavelet packet analysis can effectively analyze nonlinear signals because of its good signal local analysis ability.

The wavelet transform convolves the signal with a stretched wavelet function that has good localization properties in both the time domain and the frequency domain, and decomposes the signal into components located in different frequency bands and time periods. The description of the principle of wavelet transform has been introduced in many literatures, so I won't go into details here. The basic idea of wavelet theory is: in nature, signals with different frequencies have different time-varying characteristics. Usually, the spectral characteristics of lower frequency components change slowly with time, while the spectral characteristics of higher frequency components change more slowly. fast. Therefore, by dividing the time and frequency axes non-uniformly according to such a law, we can obtain more appropriate time resolution and frequency resolution in different time-frequency regions under the premise of obeying the uncertainty principle. However, in the decomposition of wavelet transform, only the approximate coefficients of the last decomposition are decomposed each time, and the detail coefficients of the last decomposition are no longer decomposed, resulting in that the frequency resolution of the small scale cannot be improved; while in the large scale time resolution is not improved. Therefore, in order to analyze signals in different frequency bands more precisely, wavelet packet analysis is produced.

Wavelet Packet Analysis via Orthogonal Scaling Functions and wavelet function ψ(x) to decompose the signal to obtain the low-frequency and high-frequency parts of the signal, and in the sequence space, it is completed by decomposing the discrete approximation coefficients through the filters h and g. Their two-scale relationship is:

(1)

In order to further generalize the two-scale equation, the following recurrence relation is defined:

(2)

Among them, n=0, 1, 2, ... represents the serial number of the function.

(2) Wavelet energy calculation

Since the intensity of vibration signals of centrifugal pumps in different frequency bands is different under different fault states, the strength of vibration signals in different frequency bands can be regarded as fault characteristics. Through wavelet packet analysis, the original vibration signal is decomposed into several high and low frequency components, and the frequency band energy value of each wavelet component can be calculated as the fault characteristic value. Taking the three-layer wavelet packet decomposition as an example, the calculation formula of the eight frequency band energies E _3j is as follows:

Wherein, x _jk (j=0,1,...,7; k=1,2,...,n) represents the discrete point amplitude of the reconstructed signal S _3j .

For any fault state, there can be 8 wavelet energy value composition vector W=[E ₃₀ , E ₃₁ ,…,E ₃₇ ] as the feature vector of the fault state.

2. Manifold learning method

The term manifold learning was first proposed by Bregler and Omohundro when they were studying visual speech recognition in 1995, but manifold learning was really deeply researched and developed from three papers published in Science magazine in 2000. Some scholars also classify principal component analysis, independent component analysis, and Fisher discriminant analysis as manifold learning methods, called linear manifold learning methods, and the methods proposed after 2000 are collectively called nonlinear manifold learning methods. It is difficult for the linear manifold learning method to accurately analyze the high-dimensional nonlinear data in practical applications when the dimensionality is reduced, but the nonlinear manifold learning method can easily solve this problem. The essence of the manifold learning dimensionality reduction method is to find the low-dimensional manifold structure in the high-dimensional data space, by minimizing the error between the high-dimensional data and the low-dimensional data, and removing the prior knowledge of the mutual influence between different categories, and then realizing Dimension reduction.

The purpose of applying manifold learning dimensionality reduction in the present invention is mainly to obtain more separable, simple and stable fault feature vectors by reducing the dimensionality of nonlinear high-dimensional feature vectors. In order to compare the actual application effects of various dimensionality reduction methods, the effects of 6 nonlinear dimensionality reduction methods are compared in the present invention, including kernel principal component analysis (kernel principal component analysis, KPCA) method, Laplacian Eigenmaps (Laplacian Eigenmaps, LE) method, local linear embedding (local linear embedding, LLE) method, Hessian-based LLE method (HLLE), local tangent space alignment (local tangent space alignment, LTSA) method and linear local tangent space alignment (linear local tangent space alignment, LLTSA) method. Because these methods have been introduced in detail in many literatures, they will not be repeated here.

3. Dynamic time warping method

Dynamic time warping (dynamic time warping, DTW) is a pattern matching method proposed by Sakoe and Chiba for speech recognition, and has been widely used in other fields. Based on dynamic programming technology, DTW calculates the shortest distance between two time series by extending and shortening the time series, and then realizes the similarity measurement. The principle of DTW algorithm is described as follows:

For two sequences C=c ₁ ,c ₂ ,..., _ci ,...,c _m and Q=q ₁ ,q ₂ ,...,q _j ,...,q _n , between them The distance d(C _i , Q _j ) between corresponding elements can be calculated through a distance function, so as to obtain an n×m distance matrix. In the traditional DTW algorithm, the distance function is the square of the Euclidean distance. Then, by minimizing the cumulative distance, a regular path U=(u ₁ ,u ₂ ,...,u _k ,...,u _L ) can be determined, where max(m,n)≤L≤m+n -1. This path needs to satisfy some local constraints, such as:

(a) End point restriction: the start and end points of the path should correspond to the first point and the last point of the distance matrix, ensuring that the order of the sequence does not change, that is, u ₁ =(c ₁ ,q ₁ ), u _L =(c _m ,q _n ).

(b) Continuity restriction: each time, the path can only advance one step, and the matching process must be continuous, and cannot match across points, that is, when u _k = (c _i ,q _j ), u _k+1 = (c _i+1 ,q _j+1 ), then c _i+1 -c _i ≤1, q _j+1 -q _j ≤1.

(c) Monotonicity restriction: the matching process is carried out monotonously along the sequence, that is, when u _k =( _{ci ,q j )} , u _k+1 =( _ci+1 ,q _j+1 ), then c _i ≤ c _i+1 , q _j ≤ q _j+1 .

Finally, the DTW cumulative distance is defined as:

In practical applications, computing all possible paths is too time-consuming and unnecessary, so global constraints on regularized paths are applied in the matching process to reduce the computed paths.

Because in the traditional DTW algorithm, the distance function is the Euclidean distance squared, it treats the feature vectors of all dimensions equally but in fact these features are not equal. In order to solve this problem, before calculating the distance, the sequences C=c ₁ ,c ₂ ,..., _ci ,...,c _m and Q=q ₁ ,q ₂ ,..., q _j ,...,q _n are normalized. The normalization formula is as follows:

Among them, x _i ^* is the point value after standardization, m is the mean value of the sequence elements, and s is the standard deviation of the sequence.

By standardizing the distance function, in the similarity measure, when the distance between the test data and several different categories is relatively small, we hope that the distinguishability between these distances can be increased, so as to obtain better classification results.

4. Centrifugal pump fault diagnosis and health assessment method based on wavelet energy, manifold dimensionality reduction and DTW

The overall flow of the centrifugal pump fault diagnosis and health assessment method proposed by the present invention is shown in Fig. 1 . The specific steps are as follows:

(1) First, apply the wavelet packet analysis method to decompose the original vibration signal, and obtain 8 wavelet sub-signal components through three-layer wavelet decomposition;

(2) Then, for each wavelet component, extract the wavelet energy value to form the fault feature vector, and reflect the fault information through the vibration energy under various health states;

(3) For the extracted high-dimensional wavelet energy feature vector, apply the manifold learning method to reduce the feature dimension, so as to obtain more separable, simple and stable fault feature vector;

(4) Finally, based on the extracted 3D fault feature vector, DTW is used to measure the similarity between the test data and the training data, so as to determine or evaluate the fault or health state corresponding to the current data, thereby realizing fault diagnosis and health assessment.

Application examples are as follows:

1. Centrifugal pump data source

In order to verify the validity of the method proposed in the present invention, the method verification results based on laboratory centrifugal pump failure data will be shown below. The centrifugal pump is a typical single-stage self-priming centrifugal pump, which mainly provides delivery power for refueling. In the signal acquisition, the acceleration sensor is installed directly above the bearing seat of the motor housing, and the sensor is fixed on a special base by bolts, and the base is bonded to the motor housing, as shown in Figure 2. In order to obtain data under different fault conditions, fault injection was performed on the centrifugal pump. The injected faults were: bearing inner ring fault, bearing outer ring fault, bearing rolling body fault, impeller fault, bearing inner ring and impeller mixed fault, bearing Outer ring and impeller mixed failure. The sampling frequency of the vibration signal is 10.24KHz, the collection time of each group is 2s, a group of data is collected every 5s, and 20 groups of data are collected for each fault state. The collected vibration signal in normal state is shown in Figure 3.

2. Fault feature extraction based on wavelet energy and manifold dimensionality reduction

In the present invention, the combination of wavelet energy and manifold learning method is used to extract the fault characteristic information in the vibration signal of the bearing.

First, apply the wavelet packet analysis method to process the original vibration signal, and decompose the signal into 8 wavelet components through three-layer wavelet decomposition; in the present invention, in order to obtain better fault diagnosis performance, the original signal is divided into several parts, each part contains 5000 points for signal decomposition. The result of wavelet decomposition in normal state is shown in Fig.4.

Then, the wavelet energy value of each wavelet component is calculated to form the fault feature vector. The broken line graph of the eight-dimensional wavelet energy feature vector in 20 different states is shown in Figure 5. It can be seen from the figure that the fault feature vector composed of wavelet energy is high-dimensional, and the fault information expressed is too complex, which is not conducive to the calculation of subsequent diagnosis and evaluation algorithms. Therefore, manifold dimensionality reduction is performed on the high-dimensional feature vector. In order to compare the advantages and disadvantages of various manifold dimensionality reduction methods and obtain the best dimensionality reduction method, the present invention compares the dimensionality reduction effects of six methods including KernelPCA, Laplacian Eigenmaps, LLE, HLLE, LTSA, and LLTSA, and the results are shown in Figure 6 Show. It can be seen from the figure that among the six methods, the dimensionality reduction effect of the LLTSA method is the best, and can clearly distinguish the characteristics of various fault states, while the dimensionality reduction results of the other five methods are all There are varying degrees of modal confusion. The fault feature vector based on wavelet energy-LLTSA has obvious separability, and the clustering effect of the same state feature is very good, which ensures the accuracy of subsequent fault diagnosis and health assessment. Table 1 lists the eigenvalues based on wavelet energy-LLTSA in various health states as examples.

Table 1 Example table of fault eigenvalues based on wavelet energy-LLTSA

3. DTW-based fault state determination and health state assessment

(1) Determination of fault status based on DTW

Based on the fault feature vector extracted by wavelet energy-LLTSA, DTW is used to measure the distance between the test data and the sample data of each label in the sample data set. The health status of the current test data is consistent with the label of the sample data corresponding to the minimum distance value, so The fault state of the current data can be determined to realize fault diagnosis.

First, design a sample dataset. In fault diagnosis, the centrifugal pump has 7 health states, including normal state and 6 fault states. Therefore, the sample data set contains feature vectors of 7 state labels, and each label contains 5 sets of data. The details are shown in Table 2 shown.

Then, in order to verify the diagnostic performance of the algorithm, a total of 7 sets of test data are prepared, corresponding to 7 health states, and each set of test data is a set of fault feature vectors in a certain state.

Finally, DTW is used to measure the similarity between test data and sample data. The results are shown in Figure 7. It can be seen from the figure that the distance between each test data and the sample data of the same state label is almost 0, while the distance between each test data and the sample data of different state labels is almost 0. The distance value between the sample data of the tag is relatively large, and the fault state to which each test data belongs can be clearly determined.

Table 2 Details of the sample dataset

(2) Health status assessment based on DTW

Based on the fault feature vector extracted by wavelet energy-LLTSA, the distance between the test data and the normal state sample data is measured by DTW. The larger the distance value, the higher the deviation between the current state and the normal state, that is, the more serious the fault in this state. A health indicator that can evaluate the current state. Let the distance between the current test data and the normal state sample data be d _i , define R=1/(d _i +1) as the health index, and let the health value of the normal state be 1, then any fault can be calculated State health indicators to achieve health assessment.

In order to verify the effectiveness of the health assessment method proposed by the present invention, calculate the distance between the eigenvectors under the 7 kinds of fault states and the eigenvectors under the normal state, and calculate the health degree of these fault states by the health degree index calculation formula, the result As shown in Figure 8. It can be seen from the figure that different faults correspond to different health values, and the health of the current state can be clearly evaluated based on the DTW method.

In summary, the fault diagnosis and health assessment method based on wavelet energy, manifold dimensionality reduction and dynamic time warping proposed by the present invention has achieved good results in the diagnosis and assessment of centrifugal pump hybrid faults.

Parts not described in detail in the present invention belong to the known techniques of those skilled in the art.

Claims (4)

1. a kind of fault diagnosis based on wavelet energy, manifold dimension-reducing and dynamic time warping and health evaluating method, its feature It is：It is as follows that the method realizes step：

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 spy having more separability Levy vector；

Step (4), the three-dimensional fault feature vector based on extraction, application dynamic time warping (DTW) tolerance test data and instruction Practice the similarity between data, so that it is determined that or assessing the corresponding fault of current data or health status, thus realize fault examining Break and health evaluating；

Described step (4) based on extract three-dimensional fault feature vector, application dynamic time warping (DTW) tolerance test data and Similarity between training data, so that it is determined that the corresponding health status of current data, realizes failure modes and health evaluating Detailed process is as follows：

Step (B1), first, to the original vibration signal under various health status, carries out WAVELET PACKET DECOMPOSITION and extracts wavelet energy Characteristic vector, then dimensionality reduction is carried out to high dimensional feature vector, sample characteristics matrix when classifying as follow-up health status, if total The data of k kind health status, then this sample characteristics matrix V=[W₁,W₂,…,W_k], wherein W_iFeature for i-th kind of health status Vector；

Step (B2) and then, for the vibration signal of arbitrary state to be determined, by WAVELET PACKET DECOMPOSITION signal, extract little wave energy Measure feature vector dimensionality reduction；

Each characteristic vector in step (B3), the characteristic vector of application DTW algorithm tolerance state to be determined and sample characteristics matrix Similarity, metric less it was demonstrated that state currently to be determined and this label characteristics vector state closer to, so that it is determined that The health status of current data, corresponding for the eigenvalue under normal condition health degree is set to 1, then can by current data with just Often the degree of similarity between sample data measures the health degree of current state.

2. a kind of fault diagnosis based on wavelet energy, manifold dimension-reducing and dynamic time warping according to claim 1 with Health evaluating method it is characterised in that：Described step (1) applies Nonlinear harmonic oscillator method WAVELET PACKET DECOMPOSITION, to from Heart pump vibration signal is decomposed, and 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 with Health evaluating method it is characterised in that：Described step (2) calculates the wavelet energy value of each Wavelet Component as fault signature Process is as follows：

Step (A1), set original vibration signal as x (t), process through three layers of WAVELET PACKET DECOMPOSITION, in ground floor, x (t) is broken down into One high-frequency signal and a low frequency signal, then, each sub-signal carries out low-and high-frequency decomposition again, finally gives 8 little wavelength-divisions Amount, 8 Wavelet Component from low to high are expressed as (x¹,x²,x³,x⁴,x⁵,x⁶,x⁷,x⁸)；

Step (A2), the energy of each Wavelet Component of calculating, to each Wavelet Component xⁱ, its corresponding wavelet energy isIn formula, i=1,2 ..., 8；K=1,2 ..., N, x_ikFor reconstruction signal xⁱThe amplitude of discrete point, Thus, obtain the fault feature vector W=[E of wavelet energy composition¹,E²,…,E⁸].

4. a kind of fault diagnosis based on wavelet energy, manifold dimension-reducing and dynamic time warping according to claim 1 with Health evaluating method it is characterised in that：In described step (3), the Gray-level co-occurrence vector due to extracting is octuple, in order to Data characteristicses effectively in analysis higher dimensional space, application manifold learning carries out dimensionality reduction to octuple characteristic vector, and acquisition has more The three-dimensional stability fault feature vector of separability.