WO2016004749A1 - Method for recognizing tool abrasion degree of large numerical control milling machine - Google Patents

Abstract

A method for recognizing the tool abrasion degree of a large numerical control milling machine. The method comprises: acquiring vibration time domain signals in the running state of the large numerical control milling machine; obtaining frequency domain distribution of the vibration signals by means of a fast Fourier transform; selecting multiple statistical characteristic parameters in the time domain and the frequency domain to perform diffusion mapping method-based dimensionality reduction of tool abrasion characteristic parameters; determining selection of scale parameters of diffusion mapping by using a leave-one-out cross validation method and an optimization search algorithm; and recognizing the abrasion degree of an unknown tool to be measured by combining with an Nystrom expansion and kernel regression algorithm. The disadvantage that a tool abrasion sample of a large numerical control milling machine is absent can be effectively overcome, the recognition accuracy of the tool abrasion degree of the large numerical control milling machine can be improved, and the maintenance cost and time caused by not-timely tool abrasion recognition can be reduced.

Description

Method for identifying wear degree of large CNC milling machine tool Technical field

The invention belongs to the field of large-scale numerical control milling machines, and particularly relates to a method for identifying the wear degree of a large-scale numerical control milling machine tool.

Background technique

In addition to milling planes, grooves, teeth, threads and spline shafts, CNC milling machines can also process more complex profiles with high production efficiency and are widely used in the mechanical manufacturing industry. In particular, large-scale CNC milling machines (such as portal milling machines) are often used for mass production of large workpieces due to their high machining accuracy and productivity.

As the most vulnerable part of large CNC milling machines, tools are especially important for timely and effective status recognition and monitoring. According to statistics, tool wear is the primary cause of milling machine failure, resulting in downtime of 20%-30% of the total downtime of the milling machine. Moreover, in the milling process, once the tool is damaged and not found in time, it will directly affect the product processing quality, machining accuracy and production efficiency, and will seriously lead to the failure of the milling machine function and shutdown, scrapping the workpiece, and even endangering personnel safety. Therefore, how to effectively identify the wear degree of large CNC milling machine tools and monitor the running state of the tool has become an urgent problem to be solved in the intelligent development of large CNC milling machines.

In recent years, domestic and foreign scholars have done a lot of research work on the recognition of milling tool wear degree, and put forward many effective high-precision, high-reliability diagnostic methods, such as time series analysis, spectrum analysis, wavelet analysis, neural network, support. Vector machine, hybrid intelligence, etc., which provides a certain technical basis for the recognition of the degree of wear of large CNC milling machines. However, for large CNC milling machines, the identification of tool wear degree faces the following problems: (1) The processing objects of large CNC milling machines are generally large, making data collection under different tool wear levels more difficult and cost-effective. High, training sample data is small; (2) Most research methods need to manually determine the value of key parameters (such as the support vector machine's penalty function and scale factor), subjective, and large in the absence of more experience for reference. The misdiagnosis rate of the tool wear degree of CNC milling machine is higher.

Summary of the invention

The invention aims at the above-mentioned deficiencies of the prior art, and provides a method for identifying the wear degree of a large-scale numerical control milling machine tool with higher accuracy.

The invention is achieved by the following technical solutions:

A method for identifying the wear degree of a large-scale numerical control milling machine tool comprises the following steps:

(1) Acquiring the vibration time domain signal of the large CNC milling machine in the wear state of M kinds of tools; intercepting the continuous non-overlapping S group signals with the number of samples t from the vibration time domain signal of each wear state; The Fourier transform converts the waveform of each set of time domain signals into a frequency domain distribution; wherein M, t, and S are positive integers greater than one;

(2) Calculate the 8 time-frequency domain dimensionless statistical feature parameters of the S group signal data in each wear state, and combine them into the sample data set {(C _i , m _i )} (i=1, 2,...,N , N=M×S), where C _i ={c _i1 ,c _i2 ,...,c _i8 } is the characteristic parameter set of the i-th sample, and m _i is the maximum wear amount of the tool flank face corresponding to C _i ;

The eight time-frequency domain dimensionless statistical feature parameters include four time-domain dimensionless statistical parameters (C _i1 , C _i2 , C _i3 , C _i4 ) and four frequency-domain dimensionless statistical parameters (C _i5 , C _i6 , C _i7 , C _i8 ); Let the i-th sample signal data be: x _i ={x _i1, x _i2, ..., x _in }, and the frequency domain signal data after FFT transformation is f _i = {f _i1, f _i2, ..., f _in },

x _ik is the k-th signal sample points x _i, n is an amount of data of the sample points x _i,

x _im =max{x _ik |k=1,...,n},

P _ik is the power spectrum of the frequency f _ik , P _im =max{P _ik |k=1,...,n},

(3) Combining the leave-one-cross-validation to establish the mathematical optimization model of the scale parameter ε, according to the sample data set Ω, the optimization search algorithm is used to optimize the scale parameter ε, and find the value of ε which makes the objective function the smallest value;

(4) According to the dimension reduction principle of the diffusion mapping (DM) method, construct the adjacency matrix of the sample data set W={w _ij } _N×N :w _ij =exp(-||C _i -C _j || ² /ε );

among them

Is the Euclidean distance of C _i and C _j ; then, normalize W by line, let W' = {w _ij '},

(5) Solving the eigenvalues and eigenvectors of W': W'φ _k = λ _k φ _k , λ _k is the kth largest eigenvalue of W', and there is 1 = λ ₀ > λ ₁ > λ ₂ >... , φ _k is a unit eigenvector corresponding to the eigenvalue λ _k ; the eigenvalue is selected according to a preset dimension reduction dimension K: Λ={λ ₁ , λ ₂ , . . . , λ _K }, and the corresponding eigenvectors constitute a mapping matrix Ω={φ ₁ , φ ₂ , . . . , φ _K } _N×K ;

(6) Calculate the mapping coordinates of each sample point under the mapping matrix Ω:

φ(C _i )={φ _k (C _i ),k=1,2,...,K},

φ _k (C _i ) is the kth component of C _i under the mapping Ω, and φ _jk is the jth component of φ _k ;

(7) Collecting the vibration time domain signal of the large-scale CNC milling machine in the running state at regular intervals, forming the signal X to be diagnosed; and converting the time domain waveform into the frequency domain distribution; then, calculating 8 data of the signal to be diagnosed Time-frequency domain statistical feature parameter C(X)={C _X1 , C _X2 ,...,C _X8 };

(8) Perform Nystrom expansion on X and calculate the mapping coordinates of X under the mapping matrix Ω:

φ(X)={φ _k (X),k=1,2,...,K},

(9) Perform nuclear regression analysis on X to obtain the corresponding wear degree value, that is, the maximum wear amount of the flank face.

for:

Preferably, in step (3), the tabu search algorithm is used to optimize the scale parameter ε, which specifically includes the following steps:

(3.1) Determine the kernel function containing the scale parameter ε as:

Where C _i and C _j are sample points;

(3.2) Calculate the estimated wear level of each sample point by using the leave-one cross-validation method:

(3.3) Establish a mathematical optimization model for the scale parameter ε:

(3.4) Using the tabu search to optimize the above mathematical model, the value of the scale parameter ε which minimizes the prediction error is obtained.

The invention has the following beneficial effects:

(1) At present, the research on the wear of CNC milling cutters focuses on the tool wear of small and medium-sized milling machines. Because the wear and tear sample data of small and medium-sized milling machines are easier to collect, there is little research on the wear identification of large CNC milling tools. At the same time, most of the existing fault diagnosis methods are carried out under the premise of large sample data. In the case of small samples, the training effect of these methods is very poor, and the wear identification of the tools is powerless. The present invention can overcome the above drawbacks, and the invention can effectively identify the wear degree of a large numerical control milling machine tool in a small sample situation.

(2) At present, most of the research on the recognition of tool wear degree only considers the classification of three wear states (initial wear, moderate wear and severe wear), and there is no progressive nonlinear evolution law to study the degree of tool wear. The present invention By establishing a regression model of the degree of tool wear, the evolution of the degree of wear of large CNC milling tools can be effectively revealed.

(3) The invention can effectively overcome the shortcomings of the missing samples of the large-scale numerical control milling machine tool, improve the recognition accuracy of the tool wear degree of the large-scale numerical control milling machine, and reduce the maintenance cost and time caused by the tool wear recognition not timely.

detailed description

The invention provides a method for identifying the wear degree of a large-scale numerical control milling machine tool, comprising the following steps:

(1) Acquiring the vibration time domain signal of a large CNC milling machine (generally a gantry milling machine in general) in the wear state of M kinds of tools;

Among them, M is determined according to the maximum wear amount of the flank of the tool. In this embodiment, M=5 is taken, and the corresponding five tool wear states are divided into normal state, slight wear, moderate wear, large wear and sharp wear according to the maximum wear amount of the flank face, as shown in Table 1. Shown.

Table 1 Table of the maximum wear amount of the flank face and the wear state of the props

From the vibration time domain signal of each wear state, the number of consecutive samples is t (t is a multiple of the sampling frequency, t=4096 in this embodiment), and the non-overlapping S (depending on the amount of data, may be taken 5 to 10) group signals; and transforming the waveform of each group of time domain signals into a frequency domain distribution by using a fast Fourier transform; wherein M, t, and S are positive integers greater than one;

(2) Calculate the 8 time-frequency domain dimensionless statistical feature parameters of the S group signal data in each wear state, and combine them into the sample data set {(C _i , m _i )} (i=1, 2,...,N , N=M×S), where C _i ={c _i1 ,c _i2 ,...,c _i8 } is the characteristic parameter set of the i-th sample, and m _i is the maximum wear amount of the tool flank face corresponding to C _i ;

The eight time-frequency domain dimensionless statistical feature parameters include four time-domain dimensionless statistical parameters (C _i1 , C _i2 , C _i3 , C _i4 ) and four frequency-domain dimensionless statistical parameters (C _i5 , C _i6 , C _i7 , C _i8 ). Let the ith sample signal data be: x _i ={x _i1, x _i2, ..., x _in }, and the frequency domain signal data after FFT transformation is f _i ={f _i1, f _i2, ... , f _in }, then:

Waveform indicators:

Peak:

Skewness:

Kurtosis:

Stability rate:

Wave height rate:

Power spectrum standard deviation:

Frequency ratio:

Wherein, x _ik is the k-th signal sample points x _i, n is an amount of data of the sample points x _i,

x _im =max{x _ik |k=1,...,n},

P _ik is the power spectrum of the frequency f _ik , P _im =max{P _ik |k=1,...,n},

(3) Combining the leave-one-crossing verification to establish the mathematical optimization model of the scale parameter ε, according to the sample data set Ω, using the optimized search algorithm (such as tabu search, steepest descent method, golden section method, quadratic interpolation method, etc.) on the scale parameter ε Perform optimization to find the value of ε that minimizes the value of the objective function.

The so-called leave-one cross-validation method divides the sample set with N data into two parts: the training set and the verification set, each time selecting one sample data in the sample set as the verification set, and the remaining N-1 data as the training set pair verification. Set to make predictions. Repeatedly performed N times, each time selecting different sample data as a verification set, and summing each prediction error as an indicator of performance.

The tabu search algorithm is used to optimize the scale parameter ε, which includes the following steps:

(3.1) Determine the kernel function containing the scale parameter ε as:

Where C _i and C _j are sample points;

(3.2) Calculate the estimated wear level of each sample point by using the leave-one cross-validation method:

(3.3) Establish a mathematical optimization model for the scale parameter ε:

S.t.ε∈(0,1)

(3.4) Using the tabu search to optimize the above mathematical model, the value of the scale parameter ε which minimizes the prediction error is obtained.

(4) According to the dimension reduction principle of the diffusion mapping (DM) method, construct the adjacency matrix of the sample data set W={w _ij } _N×N :w _ij =exp(-||C _i -C _j || ² /ε );

among them

The Euclidean distance between C _i and C _j . Then, normalize W by line, let W'={w _ij '},

(5) Solving the eigenvalues and eigenvectors of W': W'φ _k = λ _k φ _k , λ _k is the kth largest eigenvalue of W', and there is 1 = λ ₀ > λ ₁ > λ ₂ >... , φ _k is a unit eigenvector corresponding to the eigenvalue λ _k .

If it is determined in advance that the dimensionality reduction dimension is K (generally K=2 or 3), the first K maximum eigenvalues are selected. Considering that the maximum eigenvalue λ ₀ of W' is a trivial eigenvalue (=1), it should be discarded. The eigenvalues: Λ = {λ ₁ , λ ₂ , ..., λ _K }, the corresponding eigenvectors constitute the mapping matrix Ω = {φ ₁ , φ ₂ , ..., φ _K } _{N × K} .

(6) Calculate the mapping coordinates of each sample point under the mapping matrix Ω:

φ(C _i )={φ _k (C _i ),k=1,2,...,K},

φ _k (C _i ) is the kth component of C _i under the map Ω, and φ _jk is the jth component of φ _k .

(7) Start to identify the wear state of the milling tool to be tested, and collect the vibration time domain signal (the number of samples is t) of the large CNC milling machine in the running state (or called the state to be tested) at regular intervals. Diagnose signal X; and convert the time domain waveform into a frequency domain distribution. Then, calculating eight time-frequency domain statistical feature parameters C(X)={C _X1 , C _X2 , . . . , C _X8 } of the signal data to be diagnosed;

(8) Perform Nystrom expansion on X and calculate the mapping coordinates of X under the mapping matrix Ω:

φ(X)={φ _k (X),k=1,2,...,K},

(9) Perform nuclear regression analysis on X to obtain the corresponding wear degree value, that is, the estimated maximum wear amount of the flank face

for:

The invention may be varied in many ways and will be apparent to those skilled in the art, and such changes are not considered to be within the scope of the invention. All such modifications as would be obvious to one skilled in the art are intended to be included within the scope of the appended claims.

Claims (2)

1. A method for identifying the wear degree of a large-scale numerical control milling machine tool, comprising the following steps:

   (1) Acquiring the vibration time domain signal of the large CNC milling machine in the wear state of M kinds of tools; intercepting the continuous non-overlapping S group signals with the number of samples t from the vibration time domain signal of each wear state; The Fourier transform converts the waveform of each set of time domain signals into a frequency domain distribution; wherein M, t, and S are positive integers greater than one;

   (2) Calculate the 8 time-frequency domain dimensionless statistical feature parameters of the S group signal data in each wear state, and combine them into the sample data set {(C _i , m _i )} (i=1, 2,...,N , N=M×S), where C _i ={c _i1 ,c _i2 ,...,c _i8 } is the characteristic parameter set of the i-th sample, and m _i is the maximum wear amount of the tool flank face corresponding to C _i ;

   The eight time-frequency domain dimensionless statistical feature parameters include four time-domain dimensionless statistical parameters (C _i1 , C _i2 , C _i3 , C _i4 ) and four frequency-domain dimensionless statistical parameters (C _i5 , C _i6 , C _i7 , C _i8 ); Let the i-th sample signal data be: x _i ={x _i1, x _i2, ..., x _in }, and the frequency domain signal data after FFT transformation is f _i = {f _i1, f _i2, ..., f _in },

   

   

   

   

   

   

   

   

   x _ik is the k-th signal sample points x _i, n is an amount of data of the sample points x _i,

   

   

   x _im =max{x _ik |k=1,...,n},

   

   P _ik is the power spectrum of the frequency f _ik , P _im =max{P _ik |k=1,...,n},

   

   

   

   (3) Combining the leave-one-cross-validation to establish the mathematical optimization model of the scale parameter ε, according to the sample data set Ω, the optimization search algorithm is used to optimize the scale parameter ε, and find the value of ε which makes the objective function the smallest value;

   (4) According to the dimension reduction principle of the diffusion mapping (DM) method, construct the adjacency matrix of the sample data set W={w _ij } _N×N :w _ij =exp(-||C _i -C _j || ² /ε );

   among them

   

   Is the Euclidean distance of C _i and C _j ; then, normalize W by line, let W' = {w' _ij },

   

   (5) Solving the eigenvalues and eigenvectors of W': W'φ _k = λ _k φ _k , λ _k is the kth largest eigenvalue of W', and there is 1 = λ ₀ > λ ₁ > λ ₂ >... , φ _k is a unit eigenvector corresponding to the eigenvalue λ _k ; the eigenvalue is selected according to a preset dimension reduction dimension K: Λ={λ ₁ , λ ₂ , . . . , λ _K }, and the corresponding eigenvectors constitute a mapping matrix Ω={φ ₁ , φ ₂ , . . . , φ _K } _N×K ;

   (6) Calculate the mapping coordinates of each sample point under the mapping matrix Ω:

   φ(C _i )={φ _k (C _i ),k=1,2,...,K},

   

   φ _k (C _i ) is the kth component of C _i under the mapping Ω, and φ _jk is the jth component of φ _k ;

   (7) Collecting the vibration time domain signal of the large-scale CNC milling machine in the running state at regular intervals, forming the signal X to be diagnosed; and converting the time domain waveform into the frequency domain distribution; then, calculating 8 data of the signal to be diagnosed Time-frequency domain statistical feature parameter C(X)={C _X1 , C _X2 ,...,C _X8 };

   (8) Perform Nystrom expansion on X and calculate the mapping coordinates of X under the mapping matrix Ω:

   φ(X)={φ _k (X),k=1,2,...,K},

   

   (9) Perform nuclear regression analysis on X to obtain the corresponding wear degree value, that is, the maximum wear amount of the flank face.

   

   for:

   
2. The method for identifying the wear degree of a large-scale numerical control milling machine according to claim 1, wherein the step (3) uses a tabu search algorithm to optimize the scale parameter ε, and specifically includes the following steps:

   (3.1) Determine the kernel function containing the scale parameter ε as:

   

   Where C _i and C _j are sample points;

   (3.2) Calculate the estimated wear level of each sample point by using the leave-one cross-validation method:

   

   (3.3) Establish a mathematical optimization model for the scale parameter ε:

   

   (3.4) Using the tabu search to optimize the above mathematical model, the value of the scale parameter ε which minimizes the prediction error is obtained.