CN111300146A - Numerical control machine tool cutter abrasion loss online prediction method based on spindle current and vibration signal - Google Patents

Abstract

A method for online prediction of tool wear of CNC machine tools based on spindle current and vibration signals, by using multiple tools of the same type on the spindle of CNC machine tools provided with sensors to repeat test run processing under the same working conditions and measure the original processing wear According to the three wear stages of the initial rapid wear stage of the tool, the normal wear stage and the sharp wear stage, the optimal feature set for training is extracted from the data, and the support vector regression machine prediction model is trained, and finally the trained model is used. Carry out online real-time prediction of tool wear in the actual processing process; the invention can not only fully mine the characteristic parameters related to tool wear in each wear stage in the original signal by repeating the data obtained by multiple tools of the same type under the same conditions, but also The correlation degree with the tool wear amount is further enhanced by post-processing, so that the constructed SVM prediction model can achieve high prediction accuracy and good generalization.

Description

Online prediction of tool wear of CNC machine tools based on spindle current and vibration signals method

technical field

The invention relates to a technology in the field of machining, in particular to a high-precision online prediction method for the wear amount of CNC machine tools based on spindle current and vibration signals.

Background technique

CNC machine tool is an important basic equipment of modern advanced manufacturing technology, and the machine tool downtime caused by tool failure accounts for about 1/5-1/3 of the total machine downtime. For CNC machine tools equipped with a relatively mature tool monitoring system, the downtime can be reduced by 75%, and the processing efficiency can be increased by 10-60%. The existing tool state monitoring technology can generally predict the tool wear state category more accurately, but the prediction accuracy of the wear amount is not high, and the real-time performance and popularization are poor, so it is difficult to be applied to industrial production on a large scale.

SUMMARY OF THE INVENTION

Aiming at the defects of low precision of the prediction model of the existing tool monitoring technology and poor popularization of the model established based on the experimental data of a single tool, the invention proposes an online prediction method of the tool wear amount of a numerically controlled machine tool based on the spindle current and vibration signal. The data obtained by repeating the experiment of the model tool under the same conditions can not only fully mine the characteristic parameters related to the tool wear in each wear stage in the original signal, but also further enhance the correlation degree with the tool wear amount through post-processing methods, so that the constructed support vector The regression machine prediction model can achieve high prediction accuracy and good generalization.

The present invention is achieved through the following technical solutions:

The invention relates to an online prediction method for the wear amount of CNC machine tools based on spindle current and vibration signals. From the original machining wear data, according to the three wear stages of the initial rapid wear stage, the normal wear stage and the sharp wear stage, the optimal feature set for training is extracted from it and the support vector regression machine prediction model is trained, and finally the training is used. After the model is carried out online real-time prediction of tool wear during the actual machining process.

The raw machining wear data includes the spindle vibration time domain signal measured by the acceleration sensor, the spindle single-phase current time domain signal measured by the current sensor, and the corresponding tool wear amount measured by the microscope.

The extraction includes: the wear curve obtained by fitting the tool wear amount and the corresponding number of tool passes, extracting the time domain feature of the time domain signal of the single-phase current of the spindle, extracting the time domain feature of the spindle vibration time domain signal, and the frequency domain. features as well as time-frequency domain features.

The optimal feature set refers to: according to the correlation coefficients between all features and the wear curve, the features corresponding to the correlation coefficients that are greater than the threshold in multiple experiments are obtained as valid features; it is preferable to perform Gaussian weighted moving average and normalization on the valid features. One-off processing is obtained.

The support vector regression machine prediction model is preferably a support vector regression model based on genetic algorithm parameter optimization.

For the training, all the cutting data of some tools are used as the training set, the rest is used as the test set, the mean square error of the model wear amount prediction is selected as the population fitness, and selection, crossover, and mutation operations are performed according to the fitness of the individual. , generate a new population and calculate the current fitness of each individual; select the optimal individual and compare it with the recorded optimal fitness to decide whether to update the optimal parameters and optimal fitness, and obtain the optimal prediction model through iterative training .

technical effect

The technical problem solved by the present invention as a whole is: avoiding the use of cutting force sensors that are expensive and have poor industrial applicability but can most directly reflect tool wear conditions, and choose to use vibration sensors and current sensors that are economical and highly industrially applicable for tool wear. predict. Although the model established based on the data of a single tool in the prior art has high prediction accuracy, due to the complexity of the cutting experiment and the interference of various accidental factors, if the same type of tool is used to repeat the test under the same experimental conditions, the model's performance The prediction performance is poor, that is, a model built with a single tool is often not suitable for another tool, even under the exact same machining conditions.

Compared with the prior art, the present invention proposes to use the data of repeated experiments of the same type of tool, and extract the relevant features repeated in multiple sets of experiments according to the three wear stages as the optimal feature set of the stage. Compared with the method of selecting features from single tool data in the whole wear stage, the present invention can greatly improve the generalizability of the model on the basis of maintaining high prediction accuracy. The resulting unintended technical effects include:

By extracting a large number of characteristic parameters in the time domain, frequency domain, and time-frequency domain of the signal in the three wear stages, the potential features in the signal that are highly correlated with the wear amount in different stages can be fully explored, and the correlation coefficient can be calculated to screen out the high correlation coefficient with the wear amount. The relevant features are used as the optimal feature set, and then post-processing of Gaussian weighted moving average can be performed, which can further improve the correlation degree between the features and the wear amount, and improve the accuracy of the final prediction model.

Description of drawings

Fig. 1 is the schematic flow chart of the present invention;

Fig. 2 is the flow chart of the online prediction method of tool wear amount according to the embodiment;

3 is a schematic diagram of the wear curve of the tool wear amount and the number of tool passes fitted by the embodiment;

Fig. 4 is the flow chart of the construction of the tool wear amount prediction model according to the embodiment;

5 is a schematic diagram of the variation of the fitness of the genetic algorithm parameter optimization population in an embodiment;

FIG. 6 is a schematic diagram of the overall error of the segmentation prediction model of the tool wear amount according to the embodiment.

Detailed ways

As shown in FIG. 1 and FIG. 2 , the present embodiment relates to a specific process of a tool wear amount prediction method, including the following steps:

Step 1. Set the cutting parameters as follows: the spindle speed is 2000RPM, the feed rate is 300mm/min, the cutting depth is 1mm, and the cutting width is 4.5mm. The steel workpiece is processed by face milling, and the length of each pass is 120mm. The vibration signal of the spindle is collected by the Kistler three-axis acceleration sensor, and the single-phase current signal of the spindle is collected by the current sensor.

The sampling frequency of the vibration signal is 20 kHz, and the sampling frequency of the current signal is 5 kHz.

Preferably, the tool is stopped and unloaded once after processing for a period of time, the wear of the tool flank is measured with a microscope and the corresponding number of passes is recorded. Stop collecting.

Step 2. As shown in Figure 3, in order to obtain one of the wear curves according to the measured wear amount and the corresponding number of passes, it is divided into three wear stages according to the change trend of the wear amount: initial rapid wear stage, normal wear stage and acute wear stage.

Step 3: Extract the time domain feature of the current signal, the time domain feature, the frequency domain feature and the time-frequency domain feature of the vibration signal, respectively.

The time domain features include: mean, variance, standard deviation, root mean square, maximum value, peak-to-peak value, rectified mean value, shape factor, crest factor, kurtosis factor, impulse factor, mean square amplitude, margin factor , the skewness index.

The frequency domain feature is preferably obtained by discrete Fourier transform to obtain a spectrogram before extraction.

频率方差

均方频率

其中：f表示频率，w(f)为该频率对应幅值，N为采样频率。The frequency domain features include: the barycentric frequency representing the center of gravity of the spectral position, the frequency variance representing the discrete degree of the spectrum distribution, and the mean square frequency representing the position change of the main frequency band of the spectrum, wherein: when the wear amount of the tool increases, The spectral structure also changes, so that the center of gravity frequency changes, and the center of gravity frequency changes.

Frequency variance

mean square frequency

Among them: f represents the frequency, w(f) is the amplitude corresponding to the frequency, and N is the sampling frequency.

The time-frequency domain features are extracted by wavelet packet decomposition and empirical mode decomposition, specifically: first, each component of the original vibration signal is decomposed by 4 layers of wavelet packets using the db4 wavelet basis, and then the 16 sub-bands are extracted respectively. The energy value, the percentage of energy in each frequency band, and the time domain characteristics of the reconstructed signal in each frequency band.

其中：E_i为第i个子频带的能量值，x_ij为小波包分解后第i个子频带中j个分解系数；各频带能量所占百分比

the stated energy value

Among them: E _i is the energy value of the ith sub-band, x _ij is the j decomposition coefficients in the ith sub-band after wavelet packet decomposition; the percentage of energy in each frequency band

In the empirical mode decomposition in this embodiment, the first eight signal components are obtained, and the total energy of each component is used as a characteristic parameter.

Step 4. Calculate the correlation coefficient between each feature obtained in step 3 and the corresponding wear amount of the tool in the parameter setting in step 1 for the three wear stages, so as to screen out the respective optimal feature set for each wear stage.

The optimal feature set refers to the set of all features whose correlation coefficient is greater than the threshold value and which appear repeatedly in the above three groups of experiments.

其中：X_i、Y_i分别表示特征参数和磨损量的第i个值，

分别表示该特征向量和磨损量的均值，r为皮尔逊相关系数，取值范围在[-1,1]之间，越接近区间的极值代表其相关性越强，其中1代表完全正相关，-1代表完全负相关，0代表不相关。The correlation coefficient, the Pearson correlation coefficient

Among them: X _i , Y _i represent the i-th value of the characteristic parameter and the wear amount, respectively,

Represent the mean value of the eigenvector and the wear amount, respectively, r is the Pearson correlation coefficient, the value range is between [-1, 1], the closer to the extreme value of the interval, the stronger the correlation, where 1 represents a complete positive correlation , -1 means completely negative correlation, 0 means no correlation.

The thresholds for the three wear stages are preferably 0.9, 0.6 and 0.8.

Step 5: Perform Gaussian weighted moving average and normalization processing on each feature in the optimal feature set in turn to obtain a regular feature set.

In this embodiment, the moving average data sliding window sizes of the three wear stages are selected as 25, 60, and 65, respectively.

其中：x为输入值，y为归一化后的值，x_min、x_max分别表示该序列最小和最大值。The normalization processing refers to the maximum and minimum normalization method, which converts the original data into the [0,1] interval through linear transformation, and the formula is:

Where: x is the input value, y is the normalized value, and x _min and x _max represent the minimum and maximum values of the sequence, respectively.

Step 6. Take the regular feature set as the input, the tool wear amount as the output, and the data obtained in step 1 as the test set and the training set data. Build and train the support vector regression machine model based on genetic algorithm parameter optimization in three wear stages respectively, and The trained model is used for online real-time prediction of tool wear during actual machining.

The test set and training set refer to: in the 890 samples of the three knives collected in step 1, all the samples of the two knives are selected as the training set data, and the samples of the remaining one knife are selected as the test set data.

The parameters that need to be optimized in the support vector regression machine model are the penalty coefficient and the fault tolerance interval, and the specific steps include:

① Given the training samples {(x _i , y _i ), i=1,...,n|x _i ∈R ^N ,y∈R}, I hope to fit a function f such that f( _xi ) and yi _i The difference is as small as possible, and the maximum deviation is set to ε in the support vector regression machine model, and the error is calculated only when |f(x _i )-y _i |>ε. This is equivalent to constructing an interval band with a width of 2ε taking f(x) as the center. When the sample falls into it, it is considered that the prediction is not biased, f(x)=<w,x>+b, where w, b are the weight and bias terms, respectively.

②By introducing the penalty parameter C and the slack variable ξ, the support vector regression problem can be transformed into:

This constrained optimization problem is transformed into an unconstrained problem by introducing Lagrangian multipliers, the Lagrangian function:

③ According to the necessary conditions for the optimal solution of the constraint problem, that is, the KKT condition:

Bringing the above formula into the dual problem can be solved:

get

本实施例中寻优区间设置为C∈[0,35],ε∈[0.005,1]。therefore,

In this embodiment, the optimization interval is set as C∈[0,35],ε∈[0.005,1].

As shown in FIG. 4 , which is a flowchart for constructing a tool wear amount prediction model, the specific steps of the construction and training process in this embodiment include:

Step 6.1: Select the mean square error of the model wear amount prediction as the population fitness, set the number of individuals in the population to 20, the maximum evolutionary generation to 50 and the optimization interval of the two parameters;

Step 6.2: Initialize the population, calculate the fitness value of each individual, record the initial parameter as the optimal parameter, and the initial fitness value as the optimal fitness;

Step 6.3: According to the fitness of the individual, perform selection, crossover, and mutation operations to generate a new population and calculate the current fitness of each individual; select the optimal individual and compare it with the recorded optimal fitness to decide whether to update the optimal fitness. optimal parameters and optimal fitness;

Step 6.4: When the maximum number of iterations is reached or the termination condition is met, output the current optimal parameters, otherwise go back to Step 6.3;

Step 6.5: Use the obtained optimal parameters as the final parameters of the support vector regression machine model to construct an optimal prediction model.

The online real-time prediction refers to: obtaining the feature set as the input of the support vector regression machine model according to the feature collection, extraction, and screening post-processing methods in steps 1, 3, 4, and 5, and obtaining its output in real time is the current model. Predicted value of tool wear in state.

As shown in Figure 4, in the process of genetic algorithm parameter optimization in this embodiment, the best fitness and average fitness of the population in the normal wear stage change with algebra. As shown in Figure 5, the tool wear amount in this embodiment Error map of the piecewise optimal prediction model over the entire test set.

As shown in the following table, the prediction accuracy on the test set in this example:

Evaluation indicators value mean squared error 175.2592μm² mean relative error 14.56% mean absolute error 9.3552μm maximum absolute error 57.4960μm

It can be seen that the model constructed by this method still has high prediction accuracy, with an average relative error of 14.56% and an average absolute error of 9.3552 μm on the entire test set. Although the maximum absolute error is as high as 57.4960μm, except for the larger error at the beginning and near the blunt stage, the absolute error in the middle part is below 12μm.

Compared with most methods that divide the training set and the test set with the experimental data of the full life of a single tool, and use the prediction accuracy on the test set as the evaluation index, the present invention utilizes two kinds of processing signals that are easy to collect, namely the spindle vibration and the current. The invention focuses more on the generalizability of the model, so the experimental data of multiple knives are used as the training set, and the established model is predicted on another knife to verify the accuracy of the model. The prediction results in the examples prove the invention. The generalizability of the modeling method; compared with most methods, the whole process of the tool from new to blunt is used as a whole for feature extraction. The present invention performs feature extraction according to the three wear stages respectively, and under the condition of ensuring high prediction accuracy , which can better eliminate the interference of processing environment and accidental factors, and has good adaptability and generalization for tool wear prediction under the experimental conditions of the same type of tool and the same working condition.

The above-mentioned specific implementation can be partially adjusted by those skilled in the art in different ways without departing from the principle and purpose of the present invention. The protection scope of the present invention is based on the claims and is not limited by the above-mentioned specific implementation. Each implementation within the scope is bound by the present invention.

Claims (7)

1. A numerical control machine tool wear amount on-line prediction method based on spindle current and vibration signals is characterized in that trial run processing is repeatedly carried out on a numerical control machine tool spindle provided with a sensor by utilizing a plurality of tools of the same type under the same working condition, original processing wear data is measured, according to three wear stages of an initial rapid wear stage, a normal wear stage and a rapid wear stage of the tools, an optimal feature set for training is respectively extracted from the wear stages and a support vector regression model is trained, and finally the trained model is adopted to carry out on-line real-time prediction on the tool wear amount in the actual processing process;

the original processing wear data comprise a main shaft vibration time domain signal measured by an acceleration sensor, a main shaft single-phase current time domain signal measured by a current sensor and a corresponding cutter wear amount measured by a microscope;

the extraction comprises the following steps: and according to a wear curve obtained by fitting the tool wear amount and the corresponding feed times, extracting the time domain characteristics of the single-phase current time domain signal of the spindle, and extracting the time domain characteristics, the frequency domain characteristics and the time-frequency domain characteristics of the vibration time domain signal of the spindle.

2. The method for on-line prediction of abrasion loss of numerical control machine tool cutter according to claim 1, characterized in that the plurality of cutters of the same model, in particular three 3-edge flat-end milling cutters with a diameter of 8mm, are used for face milling 45 steel workpieces, the length of each feed is 120mm, a Kistler triaxial acceleration sensor is used for collecting spindle vibration signals, and a current sensor is used for collecting spindle single-phase current signals.

3. The method for on-line prediction of abrasion loss of numerical control machine tool cutter according to claim 1 or 2, characterized in that the optimal feature set is: the correlation coefficient is greater than the threshold value and the set of all the characteristics which repeatedly appear in a plurality of groups of experiments, namely the characteristic corresponding to the correlation coefficient greater than the threshold value is obtained as an effective characteristic according to the correlation coefficients of all the characteristics and the wear curve, and the effective characteristic is obtained through Gaussian weighted moving average and normalization processing;

the correlation coefficient, i.e. Pearson's correlation coefficient

Wherein: x_i、Y_iThe ith values representing the characteristic parameter and the wear amount respectively,

respectively representing the mean value of the characteristic vector and the abrasion loss, wherein r is a Pearson correlation coefficient and has a value range of [ -1,1]The more extreme values in the interval represent stronger correlation, wherein 1 represents complete positive correlation, -1 represents complete negative correlation, and 0 represents no correlation.

4. The method for on-line prediction of the abrasion loss of a numerical control machine tool cutter according to claim 1, wherein the time domain features comprise: the method comprises the following steps: mean, variance, standard deviation, root mean square, maximum, peak-to-peak, rectified mean, form factor, peak factor, kurtosis factor, impulse factor, mean square amplitude, margin factor, skewness index;

the frequency domain features include: characterizing spectral localityA center of gravity frequency of the heart, a frequency variance characterizing a degree of dispersion of the spectral distribution, and a mean square frequency characterizing a change in position of the spectral main band, wherein: when the abrasion loss of the cutter is increased, the frequency spectrum structure is changed, so that the center of gravity frequency is changed

Frequency variance

Mean square frequency

Wherein: f represents frequency, w (f) is amplitude corresponding to the frequency, and N is sampling frequency;

the time-frequency domain features are extracted by wavelet packet decomposition and empirical mode decomposition, and specifically the method comprises the following steps: firstly, 4-layer wavelet packet decomposition is carried out on each component of the original vibration signal by using a db4 wavelet basis, and then the energy value, the percentage of energy of each frequency band and the time domain characteristics of a reconstructed signal of each frequency band are respectively extracted from 16 frequency bands.

5. The method for on-line prediction of the abrasion loss of the numerical control machine tool cutter according to claim 1, characterized in that the training uses part of the characteristic data as a training set, the rest part of the characteristic data as a test set, the mean square error of model abrasion loss prediction is selected as population fitness, and according to the individual fitness, selection, crossing and variation operations are performed to generate a new population and calculate the current fitness of each individual; and selecting the optimal individual, comparing the optimal individual with the recorded optimal fitness, determining whether to update the optimal parameter and the optimal fitness, and performing iterative training to obtain an optimal prediction model.

6. The method for on-line prediction of abrasion loss of numerical control machine tool cutter according to claim 1 or 5, wherein the support vector regression prediction model is a support vector regression model based on genetic algorithm parameter optimization, and the construction and training process comprises the following specific steps:

step 1: selecting the mean square error of model abrasion loss prediction as population fitness, setting the number of individuals in a population to be 20, the maximum evolution algebra to be 50 and the optimization interval of two parameters;

step 2: initializing a population, calculating the fitness value of each individual, and recording an initial parameter as an optimal parameter and an initial fitness value as an optimal fitness;

and step 3: according to the fitness of the individual, carrying out selection, crossing and mutation operations to generate a new population and calculate the current fitness of each individual; selecting the optimal individual and comparing the optimal fitness with the recorded optimal fitness to determine whether to update the optimal parameters and the optimal fitness;

and 4, step 4: outputting the current optimal parameters when the maximum iteration times are reached or the termination condition is met, otherwise returning to the step 3;

and 5: and constructing an optimal prediction model by taking the obtained optimal parameters as final parameters of the support vector regression model.

7. The method for on-line predicting the abrasion loss of the numerical control machine tool cutter according to claim 6, wherein the parameters to be optimized of the support vector regression model based on genetic algorithm parameter optimization are penalty coefficients and fault-tolerant spacing bands, and the specific optimization steps comprise:

① given training sample { (x)_i,y_i),i＝1,…,n|x_i∈R^NY ∈ R } fitting function f, let f (x)_i) And y_iHas a maximum deviation of ε, i.e. when | f (x)_i)-y_i|>E, the error is calculated, which is equivalent to constructing a spacing band with a width of 2 e, centered on f (x), and when the sample falls into it, the prediction is considered as not biased, f (x) is equal to<w,x>+ b, where w, b are weight and bias terms, respectively;

② by introducing a penalty parameter C and a slack variable ξ, the support vector regression problem can be transformed into:

converting the constrained optimization problem into an unconstrained problem by introducing a lagrange multiplier, lagrange function:

③ the conditions necessary for the optimal solution to the constraint problem, namely the KKT condition:

the dual problem can be solved by substituting the formula:

to obtain

Thus, it is possible to provide

Images