Summary of the invention
The milling machine process tool abrasion prediction technique based on power detection that the object of the present invention is to provide a kind of, makes
Lathe power signal is acquired with power sensor, difficult cutting and the impacted problem of signal strength can be improved;Extraction and cutter
The strong signal characteristic of abrasion loss correlation, and feature is post-processed to improve the sensibility of feature Cutter wear, it establishes
Tool abrasion prediction model based on management loading method, prediction model is accurate, and effectively save cost simultaneously improves effect
Rate.
A kind of milling machine process tool abrasion prediction technique based on power detection provided by the invention, step
Are as follows:
S1, data acquisition: acquisition power signal and measurement tool abrasion;
S2, feature extraction: the effective value tag of power is extracted, the effective value tag of the power and the tool abrasion are calculated
Related coefficient, obtain sensitive features;
S3, feature post-processing: first carrying out isotonic regression to all characteristic values of the sensitive features, then carries out index and put down
It is sliding;
S4, prediction model: by training sample, the SBL model prediction of management loading method is carried out, verifying sample is carried out
This.
Preferably, the power signal of milling machine process is acquired by power sensor;
Measuring tool abrasion is to be measured after each feed using microscope Cutter wear.
Preferably, the effective value tag of the power includes valid value maximum value, virtual value minimum value and virtual value mean value;
Virtual value maximum value, virtual value minimum value and virtual value mean value after extracting feed each time, calculate separately out whole
The mean value of the mean value of the virtual value maximum value of a feed process, the mean value of virtual value minimum value and virtual value mean value, then count respectively
Calculate the related coefficient of virtual value maximum value, virtual value minimum value and virtual value mean value and the tool abrasion.
Preferably, the formula of the related coefficient of the effective value tag of any one power and the tool abrasion is calculated are as follows:
Wherein, i-feed time;xiThe virtual value characteristic value of-i-th;The mean value of the validity feature value of-entire feed process;VaiThe tool abrasion generated when the feed of-i-th;- entire mistake
The tool abrasion average value of journey;The related coefficient of effective value tag and tool abrasion after ρ-i-th feed.
Preferably, compare related coefficient, virtual value minimum value and the tool wear of virtual value maximum value and tool abrasion
The size of the related coefficient of the related coefficient and virtual value mean value and tool abrasion of amount, will be corresponding to maximum correlation coefficient value
The effective value tag of power as the sensitive features.
Preferably, the isotonic regression is the trend for making all characteristic values of sensitive features keep monotonic nondecreasing;The guarantor
The step of sequence returns are as follows:
If a, having f for arbitrary i=1,2 ..., Ni≤fi+1, then
B, i makes f if it existsi> fi+1, enable
Simultaneously willIt is updated to
C, repeating said steps b, until obtainMeet
Then export result
Wherein, fiFor the ith feature value of sensitive features, N is the number of all characteristic values of sensitive features, f*iIt is fi's
Isotonic regression,It is the sequence of N number of characteristic value,It is the isotonic regression sequence that N number of characteristic value keeps monotonic nondecreasing.
Preferably, exponential smoothing is carried out to all characteristic values for the sensitive features for keeping monotonic nondecreasing, keeps its variation smooth;
The exponential smoothing are as follows:
Given time sequence xt,xt-1,...,x2,x1, sequence after single exponential smoothing are as follows:
yi=α xi+(1-α)yi-1(2≤i≤t),Wherein, 0 < α < 1;α is smoothing factor, yi
For in the smooth value at moment i second, yi-1For in the smooth value at moment i-1 second;y1For initial smoothing value, x1It is the spy at first second moment
Value indicative, x2It is the characteristic value at second second moment, x3It is the characteristic value at moment third second.
Preferably, the smoothing factor α is set as 0.3.
Preferably, by the management loading method SBL model prediction, the characteristic value sample of sensitive features is inputted
This, predicts corresponding tool abrasion;
The management loading method SBL model prediction step are as follows:
S41: given input Yn×k=[y1,y2,...,yk];
Wherein y1,y2,...,ykThe characteristic value of k sensitive features respectively after post treatment, every a line y of matrix Yi
=[yi1,yi2,…,yik] (i=1 ..., n) correspondingly export as the tool abrasion t being measured microscopicallyi;
S42: according to the characteristic value sample for the sensitive features for giving input in step S41, Gaussian noise variance σ is initialized2
With hyper parameter vector α=(α1,α2,…,αN), set hyper parameter iteration upper limit αmax;Wherein, if assuming tiBe a unknown function and
Measurement error εiCombination, then obtain formula: ti=F (yi,W)+εi;In formula: εiIt is independent error term, and Gaussian distributed
N(0,σ2);tiObedience is desired for F (yi, W), Gaussian noise variance σ2Gaussian Profile, W=(w0,w1,…,wn) be weight to
Amount;
S43: according to formula u=σ1 -2∑ΦTT, Σ=(σ1 -2ΦTΦ+A)-1Calculate separately weight vector W expectation u and
Variances sigma1 2;Wherein t is tool abrasion, Φ=[Φ (y1),Φ(y2),…,Φ(yn)]TIt is basic function matrix, Φ (yi)=
[1,K(yi,y1),K(yi,y2),…,K(yi,yn)]T;W obeys Gaussian prior distribution, andIn formula:
α=(α0,α1,...,αn) it is hyper parameter vector, αiIt is to wiSpecified hyper parameter;A=diag (α0,α1,...,αn);
S44: being updated, and recalculates Gaussian noise variance σ2;
S45: repeating step S43 and step S44, until reaching specified the number of iterations, or until Gaussian noise variance
σ2And hyper parameter vector α=(α0,α1,...,αN) reach specified computational accuracy;
S46: being screened, and is retained and is less than α in hyper parameter vectormaxHyper parameter corresponding to weight and basic function;
S47: corresponding tool wear value t is predicted*。
Preferably, hyper parameter iteration upper limit α is set in the step S42max=1.0e4。
Compared with prior art, beneficial effects of the present invention are that prediction model is accurate, and sensor installation is simple, are effectively saved
About cost, and improve efficiency.
Specific embodiment
The milling machine process tool abrasion prediction technique based on power detection that the present invention provides a kind of, to make this hair
Bright to be clearer and more comprehensible, the present invention will be further described with specific embodiment with reference to the accompanying drawing.
Milling machine process tool abrasion prediction technique based on power detection of the invention, step specifically:
Step 1: data acquire;
The power signal of milling machine process is acquired using power sensor, and using micro- after each completion of processing
Mirror Cutter wear is taken pictures, and tool abrasion is measured or calculate, such as the tool wear value V generated when i-th feedaiWith it is whole
The tool wear average value of a process
The embodiment of the present invention is related to a CNC milling machine, work piece (No. 45 steel), milling cutter, power sensor and shows
Micro mirror.
Cutting parameter is as shown in table 1:
1 cutting parameter of table
Wherein, n is cutter revolving speed, and f is cutter linear velocity, apWide, a for millingeIt is deep for milling.
The present embodiment acquires the power signal of CNC milling machine using power instrument (HIOKI PW3360), and power instrument is mounted on
Machine power point of incoming cables, the frequency of power instrument sampling are 10.24KHz, while being calculated automatically and to export each second power effective
Value.
After each feed processing, tool abrasion is measured by microscope.
It is processed according to each feed, the virtual value characteristic value x of available i-thiIt (include virtual value mean value, virtual value
Maximum value and virtual value minimum value), the validity feature value mean value of entire feed process(the virtual value mean value comprising whole process
Average value, whole process virtual value maximum value average value and whole process virtual value minimum value average value).
It is illustrated in figure 2 the curve graph of milling machine cutting time S (abscissa) Yu power virtual value RMS (ordinate).
Step 2: feature extraction;
(1) feature of cutting power virtual value is extracted, respectively virtual value maximum value, virtual value minimum value and virtual value is equal
Value.
(2) meter is extracted to three kinds of above-mentioned virtual value maximum value, virtual value minimum value and virtual value mean value characteristic values
It calculates, calculating extraction is carried out using formula (1):
Wherein, i-feed time;xiThe virtual value characteristic value of-i-th (includes virtual value mean value, virtual value maximum value
With virtual value minimum value);The mean value of the validity feature value of-entire feed process be (the virtual value mean value comprising whole process
Average value, whole process virtual value maximum value average value and whole process virtual value minimum value average value);Vai- the
The tool abrasion generated when i feed;The tool abrasion average value of-whole process;It is each after ρ-i-th feed to have
The related coefficient of valid value feature and tool abrasion.
(3) according to the calculating of formula (1), virtual value mean value related coefficient, virtual value maximum value related coefficient are obtained respectively
With virtual value minimum value related coefficient, wherein correlation coefficient ρ value is taken to grind closest to 1 effective value tag as cutter of the invention
The sensitive features of damage amount.
Using formula (1), the correlation coefficient ρ between virtual value maximum value tag and tool abrasion can be calculated1, effectively
The correlation coefficient ρ being worth between minimum value tag and tool abrasion2And between virtual value characteristics of mean and tool abrasion
Correlation coefficient ρ3。
The numerical value specifically calculated is as shown in table 2, because of ρ1<ρ2<ρ3, so selecting virtual value characteristics of mean for the present embodiment
Sensitive features.
The related coefficient of table 2 each feature and tool abrasion
Step 3: feature post-processes;
In order to overcome power network fluctuation to the influence of characteristic value, need to post-process all characteristic values of sensitive features,
The process of post-processing mainly includes isotonic regression and exponential smoothing.
(1) isotonic regression is used, all characteristic values of sensitive features is made to keep the trend of monotonic nondecreasing.The institute of sensitive features
If characteristic value has been monotonic nondecreasing trend in order, the former sequence of all characteristic values is kept.Next spy if it exists
Value indicative is greater than a upper characteristic value, then needs to carry out isotonic regression to it;Circulation carries out, until making all features of sensitive features
Value keeps the trend of monotonic nondecreasing, completes isotonic regression.
Steps are as follows for its specific calculating:
If a, having f for arbitrary i=1,2 ..., Ni≤fi+1, then
B, i makes f if it existsi> fi+1, enable
Simultaneously willIt is updated to
C, repeat step b, until obtained sequenceMeetThen export resultWherein, fiFor the ith feature value of sensitive features, N is the number of all characteristic values of sensitive features,
f*iIt is fiIsotonic regression,It is the sequence of N number of characteristic value,It is the order-preserving of the holding monotonic nondecreasing of N number of characteristic value
Recursive sequence.
Wherein the principle of isotonic regression method is as follows:
If mutually independent Variables Sequence x1,x2…,xnMeet relationship x1≤x2≤…≤xn;Define g (xi) it is xiLetter
It counts, then g (xi) isotonic regression g (xi)*Meet relationship: g (x1)*≤g(x2)*≤…≤g(xn)*.Wherein g (xi)*Using as follows
Formula (2) is calculated:
In formula:Wherein, 1≤i≤n;
(2) double smoothing method is used, when signal characteristic is integrally in increase or reduces trend, uses secondary finger
Number smoothing model is smoothed feature.When being handled for the characteristic value of power sensitive feature of the invention, smoothly
Coefficient takes α=0.3.
Wherein the principle of exponential smoothing method is as follows:
Given time sequence xt,xt-1,...,x2,x1, then shown in such as formula of the sequence after single exponential smoothing (3):
In formula: 0 < α < 1 is smoothing factor, yiFor the smooth value of moment i, yi-1For the smooth value of moment i-1, y1It is initial
Smooth value, t are the time;And pass through the average determining y of first three moment value1。
Step 4: prediction model;
Using management loading method, Abrasion prediction model is constructed, steps are as follows:
1, input Y is givenn×k=[y1,y2,...,yk], wherein y1,y2,...,ykK respectively after post treatment is sensitive
The characteristic value of feature.
Every a line y of matrix Yi=[yi1,yi2,…,yik] (i=1 ..., n), correspondingly output is measured microscopically
Tool abrasion ti。
According to the characteristic value sample data of the sensitive features of input given herein above, Gaussian noise variance σ is initialized2With super ginseng
Number vector α=(α1,α2,...,αN), set hyper parameter iteration upper limit αmax。
It is assumed that tiIt is a unknown function and measurement error εiCombination, it may be assumed that
ti=F (yi;W)+εi (4)
In formula: εiIt is independent error term, and Gaussian distributed N (0, σ2), then tiObedience is desired for F (yi, W), Gauss
Noise variance is σ2Gaussian Profile, W=(w0,w1,...,wn) it is weight vector.
2, according to formula u=σ1 -2∑ΦTT and Σ=(σ1 -2ΦTΦ+A)-1Calculate separately expectation u and the side of weight vector W
Poor σ1 2, wherein t is tool abrasion;Φ=[Φ (y1),Φ(y2),…,Φ(yn)]TIt is basic function matrix, Φ (yi)=[1,
K(yi,y1),K(yi,y2),…,K(yi,yn)]T, A=diag (α0,α1,...,αn)。
W obeys Gaussian prior distribution, shown in following formula (5):
In formula: α=(α0,α1,...,αn) it is hyper parameter vector, αiIt is to wiSpecified hyper parameter.
3, it is updated, that is, recalculates Gaussian noise variance σ2;
4, step 2 and step 3 are repeated, until reaching specified the number of iterations or Gaussian noise variance σ2With hyper parameter to
Measure α=(α0,α1,...,αN) reach specified computational accuracy;
5, it is screened, i.e. reservation hyper parameter vector α=(α0,α1,...,αN) in be less than αmaxHyper parameter corresponding to
Weight and basic function;
It needs that reasonable α is arranged in step 1max, otherwise may generate extra associated vector.Theoretically αmaxIt should be
Infinity sets α in the calculating of this examplemax=1.0e4。
Hyper parameter vector α and Gaussian noise variance σ are obtained by model training2Afterwards, for signal characteristic vector y*, prediction
Corresponding tool wear value t*。
It is discussed in detail although the contents of the present invention have passed through above preferred embodiment, but it should be appreciated that above-mentioned
Description is not considered as limitation of the present invention.After those skilled in the art have read above content, for of the invention
A variety of modifications and substitutions all will be apparent.Therefore, protection scope of the present invention should be limited to the appended claims.