CN104863840A - Reciprocating compressor intelligent diagnosis method based on EMD-PCA - Google Patents

Abstract

The invention relates to a reciprocating compressor intelligent diagnosis method based on EMD-PCA. Intelligent diagnosis of faults of a reciprocating compressor is realized by comprehensively using an empirical mode decomposition (EMD) method, a principal component analysis (PCA) method and a multi-class classification support vector method; the method is characterized in that common faults of the reciprocating compressor are subjected to feature extraction, the fault alarm and diagnosis are performed according to different sensitive features of different faults, acceleration signals of a reciprocating compressor cylinder are subjected to feature extraction by using empirical mode decomposition, and each group of acceleration signals have 30 features; then dimension reduction is performed by using principal component analysis to obtain a three-dimensional feature space; finally, five working conditions obtained by experiments are diagnosed by using multi-class classification support vector, and the intelligent diagnosis process framework of the reciprocating compressor is summarized and proposed. Through the adoption of the method, fault diagnosis accuracy under the five working conditions is high.

Description

A kind of reciprocating compressor intelligent diagnosing method based on EMD-PCA

Technical field

The present invention relates to a kind of intelligent diagnosing method of reciprocating compressor, integrated use EMD method, PCA method and SVM method realize the intelligent diagnostics of reciprocating compressor fault.

Background technique

Reciprocating compressor is the key equipment of many process industries, its complex structure, and vibration stimulus source is many, once break down very easily bring massive losses to enterprise.Therefore, research reciprocating compressor on-line monitoring and intelligent trouble diagnosis are necessary.Reciprocating compressor status monitoring is generally realized by physical quantitys such as monitoring Vibration Cylinder Body, valve gap temperature, dynamic pressures.Monitoring reciprocating compressor cylinder body acceleration signal and analyze is the important means of reciprocating compressor fault diagnosis, by the discovery reciprocating compressor fault that the monitoring of reciprocating compressor cylinder body acceleration signal can be more early stage, thus to the reciprocating compressor fault of serious accident may be caused to carry out early warning.

It is generally acknowledged that reciprocating compressor cylinder body acceleration signal belongs to non-stationary signal.The prerequisite realizing reciprocating compressor Intelligent Diagnosis Technology carries out effective feature extraction work to its acceleration signal.The people such as NASA Huang propose a kind of self-adapting signal Time-Frequency Analysis Method-empirical mode decomposition (EMD), and this method is mainly used in the analysing and processing of nonlinear and non local boundary value problem.It is thought, any signal is made up of different natural modes, simple oscillator signal.Each linear or nonlinear model will have extreme point and the zero cross point of equal number.The object that EMD decomposes is exactly the problem solving instantaneous frequency, and it is different from short time FFT transform and little conversion, and it directly can carry out instantaneous analysis, has the advantage that short time discrete Fourier transform and wavelet transformation cannot compare.

Because the feature obtained after carrying out feature extraction based on EMD method to reciprocating compressor acceleration signal is more, exceeded 3, therefore this is a high-dimensional feature space characteristic processing problem.The phenomenon of dimension disaster is there is, this phenomenon inevitable causing trouble feature identification inaccurate, so under the prerequisite of feature main information is not lost in guarantee, need to carry out dimension reduction to feature, then carrying out feature identification when high dimensional feature is identified.

The object of dimension reduction avoids dimension disaster problem.Dimension reduction has 2 kinds of methods usually: a kind of method directly selects representative data from initial data, do not calculate, and the method for this dimension reduction is called feature selection; Another method is by a series of calculating and conversion, high-dimensional original feature vector is reduced to the characteristic vector of required low dimension, this organizes new characteristic vector is one group of representative brand-new data, and the method for this reduction dimension is called as " feature extraction ".PCA belongs to second method and namely carries out feature extraction to characteristic vector, PCA is called principal component analysis, a kind of multivariate statistical method from statistical angle, an its large advantage can carry out the process of dimensionality reduction number to data, the main component element of feature space data acquisition system is obtained by PCA method, choose wherein topmost part, remaining dimension is deleted, finally reach the object reducing dimension and simplify feature space, while feature space being carried out to compress process, maintain the main information of original feature space.

Realize the intelligent diagnostics of reciprocating compressor fault, on the basis of feature extraction realizing non-stationary signal, also need the identification realizing fault signature.At field of diagnosis about equipment fault, support vector machine is compared to other the conventional method such as neuron network, and have larger advantage, wherein the most outstanding advantage is, the study of SVM model parameter can be converted into a convex optimization problem.It can draw optimal solution from the angle of the overall situation.

SVM can with multiple methods combining, realize fault diagnosis, intrinsic mode envelope spectrum and single category support vector machines combine by the people such as support vector machine is divided into single category support vector machines and multi-category support vector machines, Yang Y, diagnose rolling bearing fault; The people such as Widodo A propose a kind of intelligent failure diagnosis method based on SVM; The people such as Abbasion S use multi-category support vector machines, realize the intelligent diagnostics of rolling bearing multiclass fault; The people such as Li-mei use multi-category support vector machines to realize blast furnace fault diagnosis.

Summary of the invention

Integrated use EMD method of the present invention, PCA method and SVM method, can identify reciprocating compressor typical fault accurately.The method carries out feature extraction to reciprocating compressor most common failure, and carries out malfunction alarm and diagnosis according to the different sensitive features of different faults, it is characterized in that comprising the following steps:

1) based on the feature extraction of EMD: the five kinds of operating modes choosing reciprocating compressor carry out fault signature extraction, first EMD decomposition is carried out to the non-stationary signal collected, choose the first six layer of IMF Wave data that often kind of fault decomposites, calculate every one deck IMF Wave data frequency spectrum, calculate 0-1000Hz, 1000-2000Hz, 2000-3000Hz in every layer of IMF waveform frequency spectrum composition respectively, the energy of 3000-4000Hz, 4000-5000Hz.An Acceleration pulse of often kind of operating mode obtains 6x5 feature, totally 30 eigenvalues, and 30 eigenvalues often organizing waveform generation constitute the characteristic vector of 30 dimensions;

2) realization of dimension reduction: 30 characteristic vectors extracted after decomposing adopting EMD, PCA method is used to carry out dimensionality reduction to feature, the characteristic vector space that 30 tie up is reduced to 3 dimensional feature vector spaces, these 3 characteristic vectors have certain representativeness, need the key feature retaining initial data;

The step that utilization PCA carries out principal component analysis is as follows:

(a) extract from non-stationary signal 30 characteristic parameter Xi (i=1,2,3 ..., 30), characteristic parameter Xi dimension be N (j=1,2 ..., N);

B () sets up sample matrix M:

$M_{\mathrm{ij}}=\frac{X_{\mathrm{ij}}-\frac{\mathrm{\Σ}{X}_{\mathrm{ij}}}{N}}{\sqrt{\mathrm{\Σ}{(X_{\mathrm{ij}}-\frac{\mathrm{\Σ}{X}_{\mathrm{ij}}}{N})}^2}}$

C () asks the covariance matrix of sample matrix M:

M’＝M×M ^T

D () asks the eigenvalue of covariance matrix M ' and corresponding characteristic vector;

E the eigenvalue of covariance matrix M ' is carried out descending according to order of magnitude by (), adjust the sequence of its characteristic of correspondence vector simultaneously, obtains matrix P;

F eigenmatrix that () calculates:

K＝P×M

G the first three rows of () eigenmatrix K is can 3 characteristic parameters of representing fault sign after more characteristic parameters process;

3) based on the failure modes identification of SVM:

Support vector machine is used once to learn out whole parameters of decision function, then the classification of new feature is judged with this function, Output rusults particular value is obtained after feature to be discriminated is input to the support vector machine trained, can failure judgement classification according to result particular value.

This method, for reciprocating compressor acceleration signal, the basis of EMD, PCA and support vector machine technology have studied a kind of new reciprocating compressor intelligent diagnosing method; Obtain intrinsic mode function (IMF) after using EMD method to decompose non-stationary signal, calculate the different frequency scope of each rank IMF energy and as signal characteristic; Use the major character composition in PCA method extraction feature, reach the object of reduction dimension; SVM method carries out discriminator to the higher-dimension fault signature after dimensionality reduction.

Accompanying drawing explanation

Fig. 1 is that Acceleration pulse EMD decomposes the first six layer of IMF waveform obtained;

Fig. 2 Acceleration pulse EMD decomposes the spectrogram obtaining the first six layer of IMF waveform;

The feature extraction flow process of the acceleration signal that Fig. 3 decomposes based on EMD;

30 eigenvalues that Fig. 4 mono-group of acceleration signal extracts;

Fig. 5 reciprocating compressor fault diagnosis flow scheme framework;

Fig. 6 reduces the 3 dimensional feature vector training samples obtained based on EMD and PCA dimension;

Fig. 7 reduces based on EMD-HHT feature extraction and PCA dimension the 3 dimensional feature vector test sample books obtained;

Embodiment

Be described in detail below in conjunction with the simulation use of accompanying drawing to the inventive method.

Empirical mode decomposition (EMD decomposition) is carried out to the non-stationary signal collected, choose the first six layer of IMF Wave data that often kind of fault decomposites, calculate 0-2000Hz in every one deck IMF Wave data frequency spectrum and every layer of IMF waveform frequency spectrum composition, 2000-4000Hz, 4000-6000Hz, the energy of 6000-8000Hz, 8000-10000Hz.An Acceleration pulse of often kind of operating mode obtains 6x5 feature, totally 30 eigenvalues.30 eigenvalues often organizing waveform generation constitute the characteristic vector of 30 dimensions.After carrying out EMD decomposition to scuffing of cylinder bore fault Acceleration pulse, the first six layer of IMF oscillogram obtained is shown in Fig. 1, and the spectrogram of the first six layer of IMF waveform is shown in Fig. 2, and the feature extraction flow chart based on EMD is shown in Fig. 3, and 30 eigenvalues of one group of acceleration signal extraction as shown in Figure 4.

Diagnostic process frame diagram of the present invention as shown in Figure 5, with reciprocating compressor Laboratory Furniture for research object, is carried out fault signature and is extracted and automated Classification.

The reciprocating compressor that this method experiment uses is the 2D type reciprocating flow that Nanjing Compressors Factory produces, and its technical parameter is as follows:

Actual exhaust air amount: 12m ³/ min

Rated speed: 500r/min

Boundary dimension: 4500 × 1200 × 1500mm

Use pressure: 0.3MPa

Volume of gas storage tank: 1.5m ³

Stroke: 180mm

Net weight: 3500kg

This method experiment is by carrying out fault simulation to reciprocating compressor and carrying out to laboratory data the diagnostic accuracy that diagnostic analysis verifies this intelligent diagnosing method.

Experiment acquires 5 groups of laboratory datas altogether, and one group is the laboratory data gathered under nominal situation, and other four groups is carry out collection in case of a fault to obtain laboratory data, fault type is respectively piston rod tightening nut and loosens, and connecting-rod bolts rupture, and hit cylinder, scuffing of cylinder bore, rod fracture fault.Often kind of working condition acquiring 500 groups of acceleration wave graphic data, 300 data are used for training, and 200 groups of data are used for test.

First dimension reduction is carried out based on EMD and PCA method, accompanying drawing 6 is reduce based on EMD and PCA dimension the 3 dimensional feature vector training samples obtained, accompanying drawing 7 is reduce based on EMD and PCA dimension the 3 dimensional feature vector test sample books obtained, and often kind of fault is separated substantially clearly as we can see from the figure.

Next is based on SVM method to fault signature identification, and 5 kinds of operating modes that fault simulation experiment obtains are diagnosed, and often organize operating mode 200 groups of laboratory datas, 5 kinds of operating modes are respectively: nominal situation, piston rod tightening nut loosen, and connecting-rod bolts rupture, and hit cylinder, scuffing of cylinder bore, the oscillating signal of rod fracture fault.The fault diagnosis accuracy rate of five kinds of operating modes is respectively:

Scuffing of cylinder bore accuracy rate is 98.5%;

Hitting cylinder accuracy rate is 100.0%;

Rod fracture accuracy rate is 95.5%;

It is 100.0% that piston rod tightening nut loosens accuracy rate;

Connecting-rod bolts fracture accuracy rate is 95.0%.

Claims (1)

1., based on a reciprocating compressor intelligent diagnosing method of EMD-PCA, it is characterized in that comprising the following steps:

1) based on the feature extraction of EMD: the five kinds of operating modes choosing reciprocating compressor carry out fault signature extraction, first EMD decomposition is carried out to the non-stationary signal collected, choose the first six layer of IMF Wave data that often kind of fault decomposites, calculate every one deck IMF Wave data frequency spectrum, calculate 0-1000Hz, 1000-2000Hz, 2000-3000Hz in every layer of IMF waveform frequency spectrum composition respectively, the energy of 3000-4000Hz, 4000-5000Hz; An Acceleration pulse of often kind of operating mode obtains 6x5 feature, totally 30 eigenvalues, and 30 eigenvalues often organizing waveform generation constitute the characteristic vector of 30 dimensions;

2) realization of dimension reduction: 30 characteristic vectors extracted after decomposing adopting EMD, uses PCA method to carry out dimensionality reduction to feature, the characteristic vector space that 30 tie up is reduced to 3 dimensional feature vector spaces;

The step that utilization PCA carries out principal component analysis is as follows:

A 30 characteristic parameter Xi that () extracts from non-stationary signal, i=1,2,3 ..., 30, characteristic parameter Xi dimensions are N, j=1,2 ..., N;

B () sets up sample matrix M:

$M_{\mathrm{ij}}=\frac{X_{\mathrm{ij}}-\frac{\mathrm{\Σ}{X}_{\mathrm{ij}}}{N}}{\sqrt{\mathrm{\Σ}{(X_{\mathrm{ij}}-\frac{\mathrm{\Σ}{X}_{\mathrm{ij}}}{N})}^2}}$

C () asks the covariance matrix of sample matrix M:

M’＝M×M ^T

D () asks the eigenvalue of covariance matrix M ' and corresponding characteristic vector;

E the eigenvalue of covariance matrix M ' is carried out descending according to order of magnitude by (), adjust the sequence of its characteristic of correspondence vector simultaneously, obtains matrix P;

F eigenmatrix that () calculates:

K＝P×M

G the first three rows of () eigenmatrix K is can 3 characteristic parameters of representing fault sign after more characteristic parameters process;

4) based on the failure modes identification of SVM:

Support vector machine is used once to learn out whole parameters of decision function, then the classification of new feature is judged with this function, Output rusults particular value is obtained after feature to be discriminated is input to the support vector machine trained, can failure judgement classification according to result particular value.