CN103674511B - A kind of mechanical wear part Performance Evaluation based on EMD-SVD and MTS and Forecasting Methodology - Google Patents

Abstract

The invention provides a kind of mechanical wear part Performance Evaluation based on EMD-SVD and MTS and Forecasting Methodology, belong to mechanical wear part fault diagnosis technology field。First the signal of the monitored target gathered is carried out noise reduction process, then signal is carried out EMD decomposition, choose effective IMF component and survival function composition initial matrix, initial matrix is carried out singular value decomposition, the eigenvalue obtained is normalized and obtains characteristic vector；Then utilize MTS method to calculate mahalanobis distance, and utilize Taguchi's method that characteristic vector is optimized peace treaty to subtract；Mahalanobis distance is converted into the value of the confidence, by following the tracks of the trend of the value of the confidence, the performance of mechanical wear part is estimated, by the monitored target the value of the confidence relational model with operating mode or mate matrix, fault is predicted。Present invention, avoiding existing method process nonlinear and non local boundary value problem and the problem of mistake easily occurs, be suitably applied industry monitoring in real time, minimizing fault occurrence probability。

Description

A kind of mechanical wear part Performance Evaluation based on EMD-SVD and MTS and Forecasting Methodology

Technical field

The present invention relates to the technology of the Performance Evaluation of a kind of mechanical wear part and prediction, belong to mechanical wear part fault diagnosis technology field。

Background technology

Along with the development of science and technology, modern industry is towards the direction fast development of automatization, and the popularity of industrial automation is more and more higher。Industrial automation does not require nothing more than the manufacturing process such as machine, equipment and can automatically run, and meanwhile, matched control process is also essential。Control process is mainly used to the running status of the equipment of monitoring, it is ensured that it is properly functioning, to realize the control to final product quality。Cutter in the mechanical wear part such as lathe such as numerically controlled lathe, CNC milling machine, and the gear commonly used, bearing etc., as the critical component of various kinds of equipment, its Performance Evaluation and Predicting Technique also just become the indispensable technology of control process。

In industrial processes, shut down detection and production can be caused certain loss, on-line monitoring then can display device state in real time, thus making early warning before fault occurs, in order to maintenance and more changing jobs is arranged, reduces the loss because shutdown causes。On-line monitoring is typically via the signal gathering associated components, by signal is analyzed, noise reduction, characteristic information extraction, thus the source block Progressive symmetric erythrokeratodermia that comes of signal can be assessed and predict。

Mechanical wear part is cutter, bearing and gear etc. such as, and its fault generally shows in vibration signal, and fault-signal generally presents the characteristic of nonlinear and nonstationary, is difficult to obtain effective information by conventional time domain or frequency domain decomposition method。

The empirical mode decomposition (EmpiricalModeDecomposition is called for short EMD) proposed in 1998 by N.E.Huang et al., is suitable for analyzing non-linear, non-stationary signal sequence, has significantly high signal to noise ratio。Sophisticated signal can be decomposed into limited intrinsic mode functions (IntrinsicModeFunction by it, it is called for short IMF), the each IMF component being decomposed out contains the local feature signal of the different time scales of original signal, owing to decomposing the characteristic being based on signal sequence its temporal yardstick, therefore there is adaptivity。EMD can make Non-stationary Data carry out tranquilization process, compared with the method such as short time discrete Fourier transform, wavelet decomposition, this method be intuitively, direct, posterior and adaptive。

Singular value decomposition (SingularValueDecomposition, be called for short SVD) is a kind of method mathematically extracting matrix exgenvalue, it can the characteristic information of preservation matrix preferably, be now widely used for signal processing field。

The mode identification technology that horse field system (Mahalanobis-TaguchiSystem is called for short MTS) is developed the nineties in last century by Japanese famous quality control specialist field profound (GenichiTaguchi) doctor of mouth。Horse field system is based on mahalanobis distance (MahalanobisDistance, it is called for short MD), with mahalanobis distance signal to noise ratio (SignaltoNoiseRatio, it is called for short SNR) for optimizing index, application two-level orthogonal array carries out the selection of validity feature, and carries out data classification and discriminant analysis on this basis。Compared with mode identification technology conventional at present, horse field system has the function optimizing characteristic vector, and compared with the neutral net based on iteration thought, it is simple that it calculates process, is suitable for commercial Application。Additionally, the foundation of mahalanobis distance reference space only needs the data under mechanical wear part normal condition, thus solving at parts initial operating stage, it is more difficult to the problem obtaining life-cycle data。

The value of the confidence (ConfidenceValue is called for short CV) is the parameter weighing component health proposed by JayLee team。In the present invention, by corresponding normalized function, the mahalanobis distance that horse field system draws finally characterizes the health degree of parts by being normalized to the value of the confidence。

In industrial control process, traditional mechanical wear-out part health evaluating exists non-stationary Nonlinear harmonic oscillator is relatively difficult with Forecasting Methodology, the inaccurate problem of health evaluating, and poor real is there is, is not suitable for the problems such as commercial Application in existing employing based on the method such as neutral net, chaos。

Summary of the invention

The invention aims to solve in industrial control process, traditional mechanical wear-out part health evaluating exists non-stationary Nonlinear harmonic oscillator is relatively difficult with Forecasting Methodology, the inaccurate problem of health evaluating, solve existing to there is poor real based on the method such as neutral net, chaos, be not suitable for commercial Application problem, using on the EMD basis effectively signal processed, SVD is used to preserve the feature of signal, and use MTS that feature is differentiated, define a set of mechanical wear part Performance Evaluation based on EMD-SVD and MTS and Forecasting Methodology。

A kind of mechanical wear part Performance Evaluation based on EMD-SVD and MTS provided by the invention and Forecasting Methodology, comprise the steps:

Step one, collection monitored target signal, carry out noise reduction process to the signal gathered；

Step 2, the signal that step one is obtained carry out feature extraction, specifically: first, signal are carried out empirical mode decomposition, obtain n IMF component and survival function；Secondly, effective IMF component and survival function composition initial matrix A are chosen；Then, initial matrix A is carried out singular value decomposition, obtain the eigenvalue of signal；Finally, eigenvalue is normalized, obtains the characteristic vector after signal normalization；

Step 3, utilize the characteristic vector that obtains under monitored target normal condition to build mahalanobis distance reference space, then utilize test data to calculate mahalanobis distance；And utilize Taguchi's method that characteristic vector is optimized peace treaty to subtract；

Step 4, by normalized function, mahalanobis distance being converted into the value of the confidence, the value of the confidence characterizes the performance state of monitored target, by following the tracks of the trend of the value of the confidence, the performance of mechanical wear part is estimated；

Step 5, mechanical wear part failure predication: collect the mahalanobis distance of monitored target life-cycle, the value of the confidence data and work information, it is established that corresponding relation model or coupling matrix；Passing through relational model or the coupling matrix set up, mahalanobis distance and the value of the confidence trend to monitored target are made prediction, thus fault is predicted；

Step 6, Real-time Collection monitored target signal, after the signal gathered is carried out noise reduction process, feature extraction is carried out through step 2, corresponding mahalanobis distance is calculated through step 3, in step 4, according to this mahalanobis distance, the performance of monitored target is estimated, in step 5, fault is predicted。

In described step one, gather each signal more than two under monitored target normal condition and malfunction。

In described step 2, initial matrix A is expressed as: A=(c₁,c₂,...,c_n,r)^T, wherein, c₁,c₂,...,c_nFor n IMF component c₁(t),c₂(t),...,c_nT writing a Chinese character in simplified form of (), r is writing a Chinese character in simplified form of r (t), and superscript T represents transposition。

The mechanical wear part Performance Evaluation of the present invention and Forecasting Methodology, its advantage with have the active effect that

(1) EMD method is adopted, it is to avoid existing method processes the problem that mistake easily occurs in nonlinear and non local boundary value problem；

(2) adopt SVD method, effectively save the mathematical feature of primary signal, and improve the real-time of method；

(3) adopting horse field system, wherein the calculating process of mahalanobis distance is simple, is suitably applied industry monitoring in real time；

(4) adopt Taguchi's method that system is optimized, improve the accuracy of differentiation, also make calculating easier about subtracting of characteristic vector dimension；

(5) achieve the Performance Evaluation to mechanical wear part and prediction, reduce fault odds, it is ensured that industrial safety。

Accompanying drawing explanation

Fig. 1 is the mechanical wear part Performance Evaluation flow chart with Forecasting Methodology of the present invention；

Fig. 2 is primary signal and EMD decomposition result schematic diagram；

Fig. 3 is the trendgram of the value of the confidence；

The structural representation of opening relationships model when Fig. 4 is the failure predication of cutter。

Detailed description of the invention

Below in conjunction with drawings and Examples, the present invention is described in further detail。

The present invention is a kind of performance state for mechanical wear part, adopt EMD-SVD and the MTS method carrying out Performance Evaluation and prediction, for the mechanical wear part in work, the sensors such as vibration, acoustic emission, mechanics are installed, gather signal and study signal characteristic, the matrix that IMF component forms is carried out singular value decomposition, extract its singular value, by the singular value composition characteristic vector after normalization, utilize horse field system that signal is carried out state recognition, utilize the value of the confidence that current state is estimated, and comprehensive historical data opening relationships model carries out trend prediction。

The mechanical wear part Performance Evaluation based on EMD-SVD and MTS of the present invention and Forecasting Methodology, overall flow is as it is shown in figure 1, be specifically described each step below。

Step one, collection monitored target normal state signal and fault status signal, and the signal gathered is carried out noise reduction process。

Feature according to mechanical wear part, installs corresponding sensor, gathers vibration, acoustic emission or mechanical signal, and the signal gathered is carried out preliminary noise reduction process。Generally, the signal under normal condition and malfunction is all gathered many groups。

Step 2, to step one process after signal carry out feature extraction。

Collection in step one signal after noise reduction process are carried out EMD, it is decomposed into IMF component and survival function, choose effective IMF component and survival function composition initial matrix, initial matrix is carried out singular value decomposition and draws eigenvalue, eigenvalue is normalized and composition characteristic vector。

The process that the signal that step one is obtained carries out EMD decomposition is as follows: setting pending original signal sequence as x (t), t is acquisition time, then carry out EMD decomposition by below step:

Step 2.1: using original signal sequence x (t) as pending sequence x'(t), and set residual signal sequence r (t) as x (t), the initial value arranging enumerator p is 1；

Step 2.2: find out pending sequence x'(t) all of maximum point and minimum point, and with cubic spline function to all maximum points and minimum point matching respectively, obtain pending sequence x'(t) envelope up and down；

Step 2.3: calculate pending sequence x'(t) the mean μ (t) of envelope up and down, and by pending sequence x'(t) deduct mean μ (t), obtain new data sequence y (t): y (t)=x'(t removing low frequency)-μ (t)；

Step 2.4: judge whether y (t) meets IMF component condition, if it is not, then using sequences y (t) as pending sequence x'(t), then go to step 2.2 execution；If so, pth IMF component c is then obtained_pT (), then proceedes to perform step 2.5。

First the IMF component c obtained₁The component of highest frequency in (t) representation signal x (t)。

IMF component condition is: in whole time range, and the number of extreme point and zero crossing must be equal, or differs one at most；Point at any time, the meansigma methods of the envelope (coenvelope line) of local maximum and the envelope (lower envelope line) of local minimum is necessary for zero。

Step 2.5: update residual signal sequence r (t), by current residual signal sequence and pth IMF component c_pT () subtracts each other, the residual signal sequence after being updated: r (t)=r (t)-c_p(t)；

And update the value of p: p=p+1。

Step 2.6: judge whether residual signal sequence r (t) is monotonic function, if it is not, using residual signal sequence r (t) as pending sequence x'(t), then go to step 2.2 execution。If so, then decomposing and terminate, obtain n IMF component and the survival function of original signal sequence x (t), original signal sequence can be expressed as:

$x\left(t\right)=\underset{i=1}{\overset{n}{\Σ}}{c}_i\left(t\right)+r\left(t\right)$

The IMF component that EMD decomposes out contains the composition of signal different frequency range from high to low, and the frequency that each frequency band comprises comprises signal change itself, has self adaptation multiresolution analysis characteristic。

Generally, front several IMF components contain the most information of primary signal, take effective IMF component and survival function composition initial matrix A according to engineering experience。Taking whole n IMF component and survival function composition initial matrix A in the embodiment of the present invention, initial matrix A is expressed as: A=(c₁,c₂,...,c_n,r)^T, wherein, c₁,c₂,...,c_nFor n IMF component c₁(t),c₂(t),...,c_nT writing a Chinese character in simplified form of (), r is writing a Chinese character in simplified form of r (t), and superscript T represents transposition。

Initial matrix A is carried out singular value decomposition (SVD), obtains eigenvalue。

Q=UAV^T

Wherein, Q is the singular value matrix of matrix A, Q=[diag (E₁,E₂,…,E_n), O], E_iFor the eigenvalue of primary signal, O is null matrix, and U and V is orthogonal matrix。To eigenvalue E_iIt is normalized, obtains:

$Q_k=(E_1/\underset{i=1}{\overset{n}{\Σ}}{E}_i,E_2/\underset{i=1}{\overset{n}{\Σ}}{E}_i,...,E_n/\underset{i=1}{\overset{n}{\Σ}}{E}_i)$

Q_kFor the characteristic vector of primary signal, represent the state of tested mechanical wear part。

Obtaining the characteristic vector Q of signal_kAfter, it is possible to the virtual value of signal, peak value, nargin etc. are added characteristic vector as reference value, so that characteristic vector comprises the information of overall signal, obtains final characteristic vector。

Step 3, pattern recognition。

For the characteristic vector drawn in step 2, the characteristic vector drawn by normal state signal is for building mahalanobis distance reference space, and the mahalanobis distance of monitored target is drawn by above reference space。Use Taguchi's method that characteristic vector is about subtracted, build new mahalanobis distance reference space by the characteristic vector after about subtracting。

Step 3.1: build reference space。

Reference space is to judge the benchmark whether event occurs, and utilizes the characteristic vector structure of the normal condition that step 2 obtains。For constructing reference space, obtained the characteristic vector of normal state signal by step 2, then therefrom choose some groups and constitute reference space data matrix。

Step 3.2: utilize test data to calculate mahalanobis distance value, using judgment standard normal as equipment for the mahalanobis distance value obtained, malfunction。

If each N group of the signal obtained from step one equipment normal condition and malfunction, the characteristic vector of N group signal under each state is obtained through step 2, step 3.1 utilize the characteristic vector of k group normal state signal to construct reference space, then in step 3.2, (N-k) will be remained and organize the characteristic vector of normal state signal, the characteristic vector of N group fault status signal, as test data, calculates mahalanobis distance value。

If one group of signal q characteristic of correspondence vector is x_q, the dimension of characteristic vector is m, the mahalanobis distance MD of signal q_qFor:

${\mathrm{MD}}_q=\frac{1}{m}{x}_q^T\·R^{-1}\·x_q$

Wherein, R is characteristic vector x_qCorrelation matrix,For characteristic vector x_qTransposed matrix。

Then, the meansigma methods of mahalanobis distance under normal condition and malfunction, minima and maximum are obtained。

Characteristic vector is optimized, about subtracts by step 3.3, use Taguchi's method。Quantity according to measuring featured items selects suitable two-level orthogonal array, optimizes reference space, again calculates mahalanobis distance so that the reference space after optimizing is carried out validation verification。

Quantity according to characteristic vector selects suitable two-level orthogonal array。Orthogonal table L_a(2^b) in, a is line number, represents test number (TN), and b is columns, represents factor number, and each element in the characteristic vector that step 2 finally gives is as a factor。2 represent that number of levels, definition level 1 are " using this factor ", and level 2 is " not using this factor "。For certain factor, if t₁And t₂It is the meansigma methods of this factor signal to noise ratio under horizontal l and level 2 respectively。If factor of influence Δ t=t₁-t₂, if Δ t < 0, it was shown that this factor index has negative effect, then give up this factor。If Δ t ≈ 0, it is possible to give up this factor, if Δ t > 0, retain this factor。Finally it is reconstructed into the reference space after optimization by the factor retained。Reference space after optimizing is carried out validity check, recalculates mahalanobis distance。

Step 4, performance to monitored target are estimated。

The mahalanobis distance of the monitored target for obtaining in step 3, is normalized to the value of the confidence, by the tracking of the value of the confidence completes the Performance Evaluation to mechanical wear part。

Step 5, monitored target is carried out failure predication。

Relation according to the work information of existing sample, cutting parameter etc. with the value of the confidence curve, it is established that corresponding relational model or coupling matrix。Passing through relational model or the coupling matrix set up, mahalanobis distance and the value of the confidence trend to monitored target are predicted, thus the fault of monitored target is predicted。

Step 6, Real-time Collection monitored target signal, after the signal gathered is carried out noise reduction process, feature extraction process is carried out through step 2, corresponding mahalanobis distance is calculated through step 3, then according to this mahalanobis distance, the performance of monitored target is estimated by step 4, in step 5, fault is predicted。

Embodiment:

The collection of this example is verified from the signal of the cutter milling machine。In the employing present invention, cutter is carried out Performance Evaluation and prediction by the method based on EMD-SVD and MTS, draws the value of the confidence curve of cutter。The method of the present invention specifically comprises the following steps that

Step one, collection signal。

Feature according to cutter, installs vibrating sensor, and gathers primary signal, in Fig. 2 shown in x。Abscissa in Fig. 2 represents that collection is counted, and vertical coordinate represents amplitude。

Step 2, feature extraction。

Signal in step one is carried out EMD decomposition, obtains IMF component and survival function, as shown in Figure 2。4 IMF component IMF1～IMF4 and survival function r is formed initial matrix, and makes SVD, draw the eigenvalue of signal, by characteristic value normalization, obtain characteristic vector (E₁,E₂,E₃,E₄,E₅), and the virtual value Y of primary signal, peak K, nargin P are added characteristic vector as reference value, so that characteristic vector comprises the information of overall signal, obtain final characteristic vector as shown in table 1。

Table 1: characteristic vector example

NO.	E1	E2	E3	E4	E5	Y	K	P
									1	0.4744	0.8642	0.1675	0.0049	0.0109	0.9012	0.7393	0.8068
2	0.4332	0.8964	0.0881	0.0329	0.0040	0.9491	0.7095	0.7624
									3	0.4828	0.8629	0.1491	0.0102	0.0041	0.9568	0.7655	0.8175
4	0.5072	0.8504	0.1377	0.0233	0.0069	0.9341	0.7209	0.7720
									5	0.4323	0.8862	0.1665	0.0064	0.0067	0.9151	0.7492	0.8014
6	0.4511	0.8832	0.1270	0.0161	0.0080	0.8899	0.7602	0.8130
									7	0.5661	0.7986	0.2037	0.0135	0.0027	0.9064	0.7711	0.8280
8	0.4645	0.8684	0.1733	0.0093	0.0038	0.9386	0.7630	0.8181

Table 1 show the final characteristic vector of 8 primary signals。

Step 3, pattern recognition。

For the characteristic vector drawn in step 2, the characteristic vector drawn by normal state signal is for building mahalanobis distance reference space, and the mahalanobis distance of monitored target is drawn by above reference space。Choosing normal condition and malfunction two states uses two-level orthogonal array to carry out orthogonal test, result is as shown in table 2。

Table 2 orthogonal experiments

Characteristic vector is about subtracted, and as shown in table 2, factor 4 is because with negative factor of influence, so disallowable, the characteristic vector after about subtracting builds new mahalanobis distance reference space, and recalculates mahalanobis distance。

Step 4, Performance Evaluation。

Mahalanobis distance is finally normalized to the value of the confidence by selected normalized function, by the tracking of the value of the confidence completes the Performance Evaluation to mechanical wear part。The value of the confidence is typically between 0 to 1, and 1 represents normal condition, and the more low then health status of the value of the confidence is more poor。

Normalized function CV is expressed as:

$CV=\exp (-\frac{\sqrt{MD}}{c_0})$

In above formula, MD is the mahalanobis distance value of monitored cutter signal；C₀Being scale parameter, the signal under normal condition and initial CV value calculate, what represent in different types of cutter due to same mahalanobis distance value is different health degree, so c₀Value to be calculated by the cutter signal of current type:

$c_0=-\frac{\sqrt{mean\left({\mathrm{MD}}_{normal}\right)}}{ln\left({\mathrm{CV}}_{pre}\right)}$

In above formula, MD_normalIt is the mahalanobis distance value of normal state signal, CV_preIt is expert or according to the initial CV value that engineering experience is chosen, is generally set between 0.9 to 0.99。The signal that mahalanobis distance is big as can be seen from the above equation obtains relatively low CV value, the signal that contrary mahalanobis distance is little obtains higher CV value, and the size of mahalanobis distance is to be determined by the discrimination of signal with the signal under normal condition, therefore the cutter that wear intensity is relatively low obtains higher CV value, otherwise CV value is relatively low。Monitor the health evaluating to cutter by CV is worth。In the present embodiment, the trendgram of cutter CV value is as it is shown on figure 3, abscissa represents selected sample。

Step 5, failure predication。

The cutting parameter of cutting-tool's used life and current processing, such as feed speed, cutting depth, rotating speed etc., has direct relation。In cutter life Pre-Evaluation, first have to the corresponding relation model setting up between cutting parameter and CV value by historical evaluation result, as shown in Figure 4。

CV_(t)=(α V+ β D+ λ ω) t+ θ

Wherein, CV_(t)For the value of the confidence of t cutter, V is feed speed, and D is cutting depth, and ω is rotating speed, and α, β, λ and θ are factor of influence parameter, are drawn by the relation of historical data Yu CV value,。Historical data in the present invention requires the data having different cutting parameters, so just can find the impact on CV value of each cutting parameter。After establishing corresponding relation model, for current cutting parameter, it is possible to dope cutter before work and enter critical days after a certain time, thus changing cutter before this period arrives, or carry out degradation use。

Step 6, Real-time Collection cutter signal, after the signal gathered is carried out noise reduction process, feature extraction is carried out through step 2, corresponding mahalanobis distance is calculated through step 3, in step 4, mahalanobis distance is converted to the value of the confidence, according to the value of the confidence, the wear intensity that cutter is current is estimated, in step 5, utilize the value of the confidence obtained and the relational model set up to predict the time that cutter breaks down。

Claims (7)

1. the mechanical wear part Performance Evaluation based on empirical mode decomposition (EMD)-singular value decomposition (SVD) and horse field system (MTS) and Forecasting Methodology, it is characterised in that comprise the steps:

Step one, collection monitored target signal, carry out noise reduction process to the signal gathered；

Step 2, the signal that step one is obtained carry out feature extraction, specifically: first, signal are carried out empirical mode decomposition, obtain n intrinsic mode functions (IMF) component and survival function；Secondly, effective IMF component and survival function composition initial matrix A are chosen；Then, initial matrix A is carried out singular value decomposition, obtain the eigenvalue of signal；Finally, eigenvalue is normalized, obtains the characteristic vector after signal normalization；

Step 3, utilize the characteristic vector that obtains under monitored target normal condition to build mahalanobis distance reference space, then utilize test data to calculate mahalanobis distance；And utilize Taguchi's method that characteristic vector is optimized peace treaty to subtract；

Step 4, mahalanobis distance is converted into the value of the confidence by normalized function, by following the tracks of the trend of the value of the confidence, the performance of mechanical wear part is estimated；

Step 5, the collection mahalanobis distance of monitored target life-cycle, the value of the confidence data and work information, set up corresponding relational model or coupling matrix；Passing through relational model or the coupling matrix set up, mahalanobis distance and the value of the confidence trend to monitored target are made prediction, it is achieved the prediction to fault；

Step 6, Real-time Collection monitored target signal, after the signal gathered is carried out noise reduction process, feature extraction is carried out through step 2, corresponding mahalanobis distance is calculated through step 3, in step 4, according to this mahalanobis distance, the performance of monitored target is estimated, in step 5, fault is predicted。

2. mechanical wear part Performance Evaluation according to claim 1 and Forecasting Methodology, it is characterised in that in described step one, gathers each signal more than two under monitored target normal condition and malfunction。

3. mechanical wear part Performance Evaluation according to claim 1 and Forecasting Methodology, it is characterised in that in described step 2, initial matrix A is expressed as: A=(c₁,c₂,...,c_n,r)^T, wherein, c₁,c₂,...,c_nFor n IMF component c₁(t),c₂(t),...,c_nT writing a Chinese character in simplified form of (), r is writing a Chinese character in simplified form of r (t), and superscript T represents transposition。

4. mechanical wear part Performance Evaluation according to claim 1 or 3 and Forecasting Methodology, it is characterised in that in described step 2, add the virtual value of signal, peak value and nargin in the characteristic vector after signal normalization, obtain final characteristic vector。

5. mechanical wear part Performance Evaluation according to claim 1 or 3 and Forecasting Methodology, it is characterised in that in described step 3, utilize Taguchi's method that characteristic vector is optimized peace treaty and subtract, specifically: select two-level orthogonal array L_a(2^b), a is line number, represents test number (TN), and b is columns, represents factor number, and each element in characteristic vector is as a factor；2 represent that number of levels, definition level 1 are " using this factor ", and level 2 is " not using this factor "；For certain factor, if t₁And t₂It is the meansigma methods of this factor signal to noise ratio under horizontal l and level 2 respectively；If factor of influence Δ t=t₁-t₂, if Δ t < 0, it was shown that this factor index has negative effect, gives up this factor；If Δ t > 0, retain this factor；Finally it is reconstructed into the reference space after optimization by the factor retained, the reference space after optimizing is carried out validity check, recalculates mahalanobis distance。

6. mechanical wear part Performance Evaluation according to claim 1 and Forecasting Methodology, it is characterised in that in described step 4, when monitored target is cutter, normalized function CV is expressed as:MD is the mahalanobis distance value of monitored cutter signal；C₀It is scale parameter,MD_normalIt is the mahalanobis distance value of normal state signal, CV_preIt is initial CV value, is set between 0.9 to 0.99。

7. mechanical wear part Performance Evaluation according to claim 1 or 6 and Forecasting Methodology, it is characterised in that in described step 5, when monitored target is cutter, the relational model of foundation is:

CV_(t)=(α V+ β D+ λ ω) t+ θ

Wherein, CV_(t)For the value of the confidence of t cutter, V is feed speed, and D is cutting depth, and ω is rotating speed；α, β, λ and θ are factor of influence parameter, are drawn by the relation of historical data Yu the value of the confidence。