CN115351601B - A tool wear monitoring method based on transfer learning - Google Patents

Abstract

The tool wear monitoring method based on transfer learning solves the problem that the current tool wear monitoring model based on deep learning can only model for a single working condition, and belongs to the technical field of tool wear state prediction of a numerical control machining center. The invention comprises the following steps: s1, collecting monitoring signals of a machining process under an original working condition and a new working condition; s2, acquiring cutter wear value labels under an original working condition and a new working condition, and integrating the cutter wear value labels with a monitoring signal to obtain an original working condition data set and a new working condition data set with labels; s3, constructing a transfer learning model based on edge distribution self-adaption; s4, training by using the data set with the minimized transfer learning model loss function as a target, and completing training to obtain a new working condition cutter abrasion monitoring model; s5, acquiring a monitoring signal of the current moment of the new working condition, and inputting the monitoring signal into a new working condition cutter abrasion monitoring model for prediction.

Description

A tool wear monitoring method based on transfer learning

Technical Field

The present invention belongs to the technical field of tool wear state prediction of CNC machining centers, and more specifically, relates to a tool wear monitoring method based on transfer learning.

Background technique

Tool performance is the basic guarantee for production efficiency and processing quality in the actual production of enterprises, and supports the production technology level and economic benefits of the entire manufacturing field. Accurately predicting tool wear values can maximize tool utilization and reduce production costs while ensuring part processing accuracy and avoiding equipment damage. In actual factory production, tools of the same model are often used for cutting different workpiece materials. Due to the different physical and chemical properties of workpiece materials, the degradation process of tools during cutting will also show different trends. However, the wear label collection process is cumbersome. Therefore, how to quickly adapt the monitoring model trained using historical data and corresponding wear value labels to the processing of new workpiece materials with only a small amount of data samples is an urgent problem that needs to be solved.

In the presence of a large amount of historical data and corresponding wear label values and background, a small amount of monitoring data on the cutting process of new workpiece materials is collected, and the method of rapid adaptation and migration of the model is studied. This can avoid the need to collect a large amount of monitoring data on the cutting process of new workpiece materials in order to re-establish the monitoring model, save time, and improve the efficiency of model development. This is the development need of the current online tool status monitoring technology.

With the development and application of industrial big data, traditional tool wear monitoring methods are gradually being replaced by deep network prediction methods with feature autonomous learning. In the context of massive processing process monitoring data, the advantages of deep learning for big data processing are more prominent. However, current deep learning models can usually only model a single working condition. For example, when the workpiece material changes, the existing model usually fails.

Summary of the invention

In view of the problem that the current tool wear monitoring model based on deep learning can only be modeled for a single working condition, the present invention provides a tool wear monitoring method based on transfer learning.

A tool wear monitoring method based on transfer learning of the present invention comprises:

S1, collecting monitoring signals of the machining process under the original working condition and the new working condition, the monitoring signals including cutting force signals, vibration signals and acoustic emission signals;

S2, obtaining tool wear value labels under the original working condition and the new working condition, integrating them with the monitoring signal, and obtaining the labeled original working condition data set and the new working condition data set;

S3,

A transfer learning model based on edge distribution adaptation is constructed. The transfer learning model includes a feature adapter, two original working condition tool wear monitoring models, regression network 1 and regression network 2.

The original working condition tool wear monitoring model 1 is used to extract the features of the monitoring signal in the original working condition data set in the high-dimensional embedding space, and the extracted features are input into the regression network 1;

The original working condition tool wear monitoring model 2 is used to extract the features of the monitoring signals in the new working condition data set in the high-dimensional embedding space, and the extracted features are input into the regression network 2;

Regression network 1 is used to fit the nonlinear relationship between the input features and the tool wear values in the original working condition data set;

Regression network 2 is used to fit the nonlinear relationship between input features and tool wear values in the new working condition data set;

The feature adapter is used to obtain the distribution difference between the features in the original working condition data set and the features in the new working condition data set;

S4, with the goal of minimizing the loss function of the transfer learning model, the transfer learning model is trained based on the labeled original working condition data set and the new working condition data set, and the original working condition tool wear monitoring model 2 and regression network 2 obtained after training are the new working condition tool wear monitoring model;

S5. Obtain the monitoring signal of the new working condition at the current moment, and input the monitoring signal into the tool wear monitoring model of the new working condition for prediction.

Preferably, in S4, the loss function of the transfer learning model is:

In the formula,

D _S represents the features extracted by the tool wear monitoring model 1 in the original working condition;

D _T represents the features extracted by the tool wear monitoring model 2 in the original working condition;

MMD ² (D _S ,D _T ) represents the distribution difference obtained by the feature adapter;

represents the regression loss of regression network 1;

represents the regression loss of regression network 2;

y _S represents the tool wear value corresponding to D _S in the original working condition data set;

y _T represents the tool wear value corresponding to D _T in the new working condition data set;

L represents the total loss.

Preferably, the method further comprises:

S6. Combine the tool wear prediction value at the current moment predicted by S5 and the tool wear prediction values in the previous time periods into a time series, perform least squares fitting on the composed time series to obtain the prediction value at the current moment, compare the prediction value of the least squares fitting with the tool wear prediction value of S5, if the difference is greater than the set threshold, use the prediction value of the least squares fitting as the predicted wear value at the current moment, otherwise, retain the tool wear prediction value of S5, transfer to S5, and make prediction at the next moment.

Preferably, in S6, the least squares method is fitted to the composed time series, and the objective function and constraint conditions include:

st2θ ₂ x _i +θ ₁ ≥ 0

Among them, _yi represents the wear prediction value at a certain moment in the tool wear prediction value time series;

_xi represents the corresponding cutting program number; _θ0 , _θ1 , _θ2 represent the coefficients of the quadratic polynomial.

Preferably, in S1, the cutting force signal is acquired by arranging a cutting force sensor at the bottom of the workpiece, the vibration signal is acquired by arranging vibration sensors on the workpiece and the machine tool worktable, and the acoustic emission signal is acquired by arranging an acoustic emission sensor on the machine tool worktable.

Preferably, the monitoring signal in S2 is a preprocessed signal, and the preprocessing includes:

S21, after the cutting feed starts, the monitoring signal offset is eliminated, a time period is set, the average value of each sensor signal is calculated, and the average value is subtracted from the signal as the offset value, and the standard deviation of the signal is calculated as the standard deviation base value to characterize the fluctuation degree of the monitoring signal when no cutting is performed;

S22: periodically calculate the standard deviation of the signal during the feeding process. When the standard deviation is greater than 3 times the base value, mark the moment as the cutting start point, and clear the monitoring signal before the start point to complete the empty cutting elimination;

S23, intercept the monitoring signal of the stable cutting section, and perform Z score standardization on each monitoring signal, the expression is:

Where x represents the value of a sample point on a channel signal in the sample;

μ represents the mean of all sample points of the signal data in the channel x;

σ represents the standard deviation of all sample points of the signal data of the channel x;

S24, downsampling the data after Z-score standardization in a sliding window manner, and calculating the first four moment statistics of the data in the sliding window: mean, variance, skewness and kurtosis, that is, converting the monitoring signal into four-dimensional data.

Preferably, in S2, a portable microscope is used to take photos to record the wear trend of the tool during the cutting process, and the tool wear value is marked in combination with the microscope pixel size, resolution and magnification, and the wear value is then integrated with the tool monitoring signal of the corresponding time period to obtain a labeled original working condition data set and a new working condition data set.

Preferably, in S4, the feature adapter calculates the distribution difference using the maximum mean difference method:

In the formula, φ(·) represents a mapping, which is used to map the original variable into the reproducing kernel Hilbert space;

D _Si represents the i-th feature extracted by the tool wear monitoring model 1 of the original working condition, with a total of n features;

D _Tj represents the jth feature extracted by the tool wear monitoring model 2 of the original working condition, with a total of m features;

k(·) represents the kernel function;

It means to calculate the sum of squares of functions in Hilbert space.

Beneficial effects of the present invention: The present invention simultaneously realizes the use of the prediction model obtained by training under the original working conditions, combined with a small amount of monitoring data under new working conditions, to quickly adapt to the tool wear value prediction purpose of the new workpiece material processing process, effectively reducing the data volume required for new model training, improving the efficiency of monitoring model development, and ensuring production continuity. The present invention realizes predictive maintenance of CNC tools, which is of great significance to actual production.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a tool wear monitoring model based on transfer learning provided in an embodiment of the present invention.

FIG. 2 is a schematic diagram of eliminating empty cutting of source domain vibration signals according to the present invention.

Figure 3 is a schematic diagram of the deep convolutional residual network structure of the present invention, Input represents the 4-dimensional time domain statistical feature data of the input model, Conv represents the convolution layer, BN represents the batch normalization layer, ReLU represents the linear rectification function, Max Pool represents the maximum pooling layer, Adaptive AvgPool represents the adaptive average pooling layer, Flatten represents the flattening layer, Linear represents the linear layer, and Output represents the tool wear prediction value output by the model.

FIG. 4 is a diagram showing the principle of edge distribution adaptation adopted by the present invention.

FIG5 is a schematic diagram showing the principle of the transfer learning model of the present invention.

FIG. 6 is a smoothing process for predicting wear values according to the present invention.

FIG7 is a diagram showing the training results of the transfer learning model of the present invention.

Detailed ways

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

The tool wear monitoring method based on transfer learning in this embodiment includes:

Step 1, collecting monitoring signals of the machining process under the original working condition and the new working condition, the monitoring signals including cutting force signals, vibration signals and acoustic emission signals;

Step 2: Obtain tool wear value labels under the original working condition and the new working condition, and integrate them with the monitoring signal to obtain the labeled original working condition data set and the new working condition data set;

Step 3: construct a transfer learning model based on edge distribution adaptation, which includes a feature adapter, two original working condition tool wear monitoring models and two regression networks;

The original working condition tool wear monitoring model 1 is used to extract the features of the monitoring signals in the original working condition data set in the high-dimensional embedding space, and the extracted features are respectively input into the regression network 1;

The original working condition tool wear monitoring model 2 is used to extract the features of the monitoring signals in the new working condition data set in the high-dimensional embedding space, and the extracted features are respectively input into two regression networks 2;

Regression network 1 is used to fit the nonlinear relationship between the input features and the tool wear values in the original working condition data set;

Regression network 2 is used to fit the nonlinear relationship between input features and tool wear values in the new working condition data set;

The feature adapter is used to obtain the distribution difference between the features in the original working condition data set and the features in the new working condition data set;

Step 4, with the goal of minimizing the loss function of the transfer learning model, the transfer learning model is trained based on the labeled original working condition data set and the new working condition data set, and the original working condition tool wear monitoring model 2 and regression network 2 obtained after training are used as the new working condition tool wear monitoring model;

Step 5: Obtain the monitoring signal of the new working condition at the current moment, and input the monitoring signal into the tool wear monitoring model of the new working condition for prediction.

This implementation method uses the concept of transfer learning to maximize the use of the inherent similarities between prediction tasks, find the inherent connections and differences in tool degradation trends under two working conditions, and use algorithms to adapt the feature edge distribution under the two working conditions, thereby obtaining a prediction model for new processing materials. This implementation method collects monitoring signals under the original workpiece material working condition and the new workpiece material working condition, obtains tool wear labels under the original working condition and a small number of new working conditions, and integrates them to obtain labeled original working condition data sets and new working condition data sets; constructs a transfer learning model, and then establishes a monitoring model under new working conditions through transfer learning. While reducing the demand for labeled data in the target domain, it also solves the problem of generalization application of the original tool wear monitoring model when the workpiece material changes.

In a preferred embodiment, in step 1, the cutting force signal is acquired by arranging a cutting force sensor at the bottom of the workpiece, the vibration signal is acquired by arranging vibration sensors on the workpiece and the machine tool worktable, and the acoustic emission signal is acquired by arranging an acoustic emission sensor on the machine tool worktable.

In a preferred embodiment, in step 2, the monitoring signal is a preprocessed signal, and a portable microscope is used to take pictures to record the wear trend of the tool during the cutting process, and the tool wear value is marked in combination with the microscope pixel size, resolution and magnification. The wear value is then integrated with the tool monitoring signal of the corresponding time period to obtain a labeled original working condition data set and a new working condition data set.

The preprocessing steps include:

Step 21, after the cutting feed starts, the monitoring signal offset is eliminated. The average value of the sensor signal in the 120ms period after the feed motion starts is calculated, and the average value is subtracted from the signal as the offset value. At the same time, the standard deviation of the signal in this process is calculated as the standard deviation base value, which characterizes the fluctuation degree of the monitoring signal when no cutting is performed;

Step 22: Calculate the standard deviation every 400ms during the feeding process. When the standard deviation is greater than 3 times the base value, mark the moment as the cutting start point, and clear the monitoring signal before the start point to complete the empty cutting elimination. Figure 2 shows the process of empty cutting elimination.

Step 23: intercept the monitoring signal of the stable cutting segment, retain the sampling points with a monitoring signal length of 100,000, and perform Z-score standardization on each monitoring signal. The expression is:

Where x represents the value of a sample point on a channel signal in the sample;

μ represents the mean of all sample points of the signal data in the channel x;

σ represents the standard deviation of all sample points of the signal data of the channel x;

Step 24, downsample the data after Z score standardization in a sliding window manner, and calculate the first four moment statistics of the data in the sliding window: mean, variance, skewness and kurtosis, that is, convert the monitoring signal into four-dimensional data, so as to reduce the sample length and the amount of model calculation. The calculation formula of the statistics includes:

in, represents the data mean; x _ms represents the data variance; α represents the data skewness; β represents the data kurtosis;

_Xi represents the i-th sample point of a certain sensor signal; n represents the sampling length of the sliding window.

In a preferred embodiment, the original working condition tool wear monitoring model in step 3 uses a deep convolution residual network to extract features of the input signal in a high-dimensional embedding space. The network is mainly composed of a stack of residual basic blocks. The overall architecture parameter settings of the residual convolution neural network are shown in Table 1, and the network structure is shown in Figure 3;

Table 1 Residual neural network structure settings

The feature adapter uses the maximum mean difference as a measure of the difference in the distribution of source domain features and target domain features. The principle of edge distribution adaptation used is shown in Figure 4;

The feature adapter uses the maximum mean difference method to calculate the distribution difference:

In the formula, φ(·) represents a mapping, which is used to map the original variable into the reproducing kernel Hilbert space;

D _Si represents the i-th feature extracted by the tool wear monitoring model 1 of the original working condition, with a total of n features;

D _Tj represents the jth feature extracted by the original working condition tool wear monitoring model 2, with a total of m features; k(·) represents the kernel function; It means to calculate the sum of squares of functions in Hilbert space.

Among them, the kernel function selects the Gaussian radial basis function to map the finite-dimensional data to the high-dimensional space. The calculation formula of the Gaussian radial basis function includes:

In the formula, x′ represents the center point of the kernel function;

x represents any point in space;

σ represents the bandwidth, which is used to control the radial range of the function.

The regression network consists of a fully connected neural network, which is used to fit the nonlinear relationship between high-dimensional embedded features and tool wear values.

The principle of the transfer learning model is shown in Figure 5.

In a preferred embodiment, the transfer learning model loss function in step 4 is:

Where D _S represents the features extracted by the tool wear monitoring model 1 in the original working condition;

D _T represents the features extracted by the tool wear monitoring model 2 in the original working condition;

MMD ² (D _S ,D _T ) represents the distribution difference obtained by the feature adapter;

represents the regression loss of regression network 1;

represents the regression loss of regression network 2;

y _S represents the tool wear value corresponding to D _S in the original working condition data set;

y _T represents the tool wear value corresponding to D _T in the new working condition data set;

L represents the total loss.

In a preferred embodiment, the method of this embodiment further includes:

As shown in Figure 6, step 6, the tool wear prediction value at the current moment predicted in step 5 and the tool wear prediction values in the previous time periods are combined into a time series, and the composed time series is fitted by the least squares method to obtain the prediction value at the current moment, and the prediction value of the least squares fitting is compared with the tool wear prediction value of step 5. If the difference is greater than the set threshold, the prediction value of the least squares fitting is used as the predicted wear value at the current moment, otherwise, the tool wear prediction value of step 5 is retained, and the process goes to step 5 to make the prediction for the next moment.

In step 6, the least squares method is used to fit the composed time series, and the objective function and constraints include:

st2θ ₂ x _i +θ ₁ ≥ 0

Among them, _yi represents the wear prediction value at a certain moment in the tool wear prediction value time series;

_xi represents the corresponding cutting program number; _θ0 , _θ1 , _θ2 represent the coefficients of the quadratic polynomial.

The tool wear monitoring model obtained by transfer learning is used to monitor the tool wear status of the cutting process of new workpiece materials. As shown in Figure 7, the prediction result is close to the actual value of tool wear, which shows that the tool wear status monitoring model provided by the present invention has the ability to migrate working conditions and can accurately predict the wear status of the tool under new working conditions, which is of great significance for actual production.

Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in other described embodiments.

Claims (7)

1. A tool wear monitoring method based on transfer learning, the method comprising:

s1, collecting monitoring signals of machining processes under an original working condition and a new working condition, wherein the monitoring signals comprise cutting force signals, vibration signals and sound emission signals;

S2, acquiring cutter wear value labels under an original working condition and a new working condition, and integrating the cutter wear value labels with a monitoring signal to obtain an original working condition data set and a new working condition data set with labels;

S3, constructing an edge distribution self-adaption-based transfer learning model, wherein the transfer learning model comprises a characteristic adapter, two original working condition cutter wear monitoring models, a regression network 1 and a regression network 2;

the original working condition cutter abrasion monitoring model 1 is used for extracting features of a high-dimensional embedding space of monitoring signals in an original working condition data set, and the extracted features are input into the regression network 1;

the original working condition cutter abrasion monitoring model 2 is used for extracting features of a high-dimensional embedding space of monitoring signals in a new working condition data set, and the extracted features are input into the regression network 2;

the regression network 1 is used for fitting a nonlinear relation between the input characteristics and the cutter abrasion values in the original working condition data set;

the regression network 2 is used for fitting a nonlinear relation between the input characteristics and the cutter abrasion values in the new working condition data set;

the feature adapter is used for acquiring the distribution difference of the features in the original working condition data set and the features in the new working condition data set;

The feature adapter calculates distribution differences by adopting a maximum mean difference method:

Wherein, phi (·) represents a mapping for mapping the original variable into a regenerated kernel hilbert space;

d _Si represents the ith feature extracted by the original working condition tool wear monitoring model 1, and n features are added;

d _Tj represents the j-th feature extracted by the original working condition tool wear monitoring model 2, and m features are taken as the total; k (·) represents a kernel function; Representing the calculation of the sum of squares of the functions under the hilbert space;

S4, training the migration learning model based on the original working condition data set with the label and the new working condition data set by taking the minimized migration learning model loss function as a target, and training to obtain an original working condition tool wear monitoring model 2 and a regression network 2 as new working condition tool wear monitoring models;

the transfer learning model loss function is:

Wherein D _S represents the characteristics extracted by the original working condition tool wear monitoring model 1;

d _T represents the features extracted by the original working condition tool wear monitoring model 2;

MMD ²(D_S,D_T) represents the distribution differences obtained by the feature adapter;

Representing regression loss of regression network 1;

representing regression loss of regression network 2;

y _S represents the tool wear value corresponding to D _S for the original operating mode dataset;

y _T represents the tool wear value corresponding to D _T for the new operating mode dataset;

L represents the total loss;

s5, acquiring a monitoring signal of the current moment of the new working condition, and inputting the monitoring signal into a new working condition cutter abrasion monitoring model for prediction.

2. The method of tool wear monitoring based on transfer learning of claim 1, further comprising:

S6, forming a time sequence array by the cutter abrasion predicted value at the current moment predicted by the S5 and cutter abrasion predicted values in a plurality of previous time periods, carrying out least square fitting on the formed time sequence array to obtain the predicted value at the current moment, comparing the least square fitted predicted value with the cutter abrasion predicted value of the S5, if the difference value is larger than a set threshold value, using the least square fitted predicted value as the predicted abrasion value at the current moment, otherwise, reserving the cutter abrasion predicted value of the S5, and carrying out the next moment prediction.

3. The tool wear monitoring method based on transfer learning according to claim 2, wherein in S6, the least square fitting is performed on the time series sequence of the components, and the objective function and the constraint condition include;

Wherein y _i represents the predicted wear value at a certain time in the time sequence of predicted wear values of the tool;

x _i represents the corresponding cutting stroke number; θ ₀、θ₁、θ₂ represents the quadratic polynomial coefficient.

4. The tool wear monitoring method based on the transfer learning according to claim 1, wherein in S1, the cutting force signal is acquired by arranging a cutting force sensor at the bottom of the workpiece, the vibration signal is acquired by arranging vibration sensors at the workpiece and the machine table, and the acoustic emission signal is acquired by arranging acoustic emission sensors at the machine table.

5. The tool wear monitoring method based on transfer learning according to claim 4, wherein the monitoring signal in S2 is a pre-processed signal, and the pre-processing includes:

S21, performing offset elimination on the monitoring signal after cutting feeding starts, setting a time period, calculating an average value of each sensor signal, subtracting the average value from the signal as an offset value, and simultaneously calculating the standard deviation of the signal as a standard deviation basic value to represent the fluctuation degree of the monitoring signal when cutting is not performed;

S22, periodically calculating the standard deviation of the signal in the feeding process, and marking the moment as a cutting starting point and clearing a monitoring signal before the starting point when the standard deviation is more than 3 times of a basic value, so as to finish the elimination of the empty cutting segment;

s23, intercepting monitoring signals of the stable cutting section, and carrying out Z fraction standardization on each path of monitoring signals, wherein the expression is as follows:

wherein x represents the value of a certain sample point on a certain channel signal in the sample;

Mu represents the average value of all sample points of the channel signal data of x;

sigma represents the standard deviation of all sample points of the channel signal data of x;

S24, carrying out downsampling processing on the data after Z score standardization in a sliding window mode, and calculating the first fourth order moment statistic of the data in the sliding window: mean, variance, skewness, and kurtosis, i.e.: the monitoring signal is converted into four-dimensional data.

6. The tool wear monitoring method based on transfer learning according to claim 5, wherein in S2, a portable microscope photographing mode is adopted to record a wear change trend of a tool in a cutting process, a tool wear value is marked by combining a microscope pixel size, a resolution and a magnification, and then the wear value and a tool monitoring signal of a corresponding time period are integrated to obtain an original working condition data set and a new working condition data set with labels.

7. A computer-readable storage device storing a computer program, characterized in that the computer program, when executed, implements the tool wear monitoring method based on transfer learning according to any one of claims 1 to 6.