Milling tool wear recognition method based on three-dimensional variational mode decomposition feature fusion

Abstract

The application discloses a kind of based on three-dimensional variation modal decomposition feature fusion milling tool wear identification method, comprising: obtaining the multi-channel force signal and multi-channel vibration signal of tool under different tool path number, and corresponding tool wear amount;With two subsection regression algorithm, tool wear curve is divided into initial wear stage, normal wear stage and sharp wear stage;The principal component signal of multi-channel force signal and multi-channel vibration signal is three-dimensional variation modal decomposition, and main IMF component is selected;The time domain, frequency domain and time-frequency domain features of each main IMF component are calculated, according to the feature selection method of ReliefF, the features with large contribution to tool wear state classification are retained, and secondary dimension reduction is carried out by unsupervised learning SDAE;Feature fusion forms multidimensional feature vector, and tool wear state constitutes the training data of model together;OSELM is optimized using the algorithm of dung beetle optimization, and DBO-OSELM model is trained to realize the identification of tool wear state.The application has higher identification accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of tool wear detection, and particularly relates to a milling tool wear recognition method based on three-dimensional variational modal decomposition feature fusion. BACKGROUND

[0002] Tool wear, as the most important tool failure form, is related to the precision of manufacturing and the surface quality of products. The methods for measuring tool wear mainly include direct observation method and indirect observation method. The direct observation method measures tool wear to measure, including monitoring wear width, wear surface size, etc. It has two main defects, one is that it cannot monitor the tool wear state in real time, and the other is that it needs to stop detection, resulting in a decrease in overall machining efficiency. The indirect observation method is based on the monitoring of signals closely related to wear or damage, and a model of monitoring signals and tool state is constructed to indirectly obtain the current wear state. Therefore, it is of important practical significance to construct a tool wear model and recognize tool wear.

[0003] With the rapid development of sensor technology and signal processing, tool state information can be obtained by sensors installed on the machine tool part, and the tool wear condition can be finally predicted by combining machine learning model and analyzing sensor signal features. However, due to the redundancy of sensor signals, the signals are easily disturbed by the surrounding environment, and the correlation between signal features and tool wear is not strong, which makes the prediction performance of the model not meet the requirements. Therefore, there is an urgent need for an efficient signal preprocessing method and signal feature combination method to meet the tool wear prediction accuracy requirements. SUMMARY

[0004] The application aims to provide a milling tool wear recognition method based on three-dimensional variational modal decomposition feature fusion, which realizes high-precision prediction of tool wear state.

[0005] To achieve the above-mentioned purpose, the technical solution adopted by the application is as follows:

[0006] A milling tool wear recognition method based on three-dimensional variational modal decomposition feature fusion, the method comprising the following steps:

[0007] Obtain the multi-channel force signals and multi-channel vibration signals of the tool under different tool passes, and the corresponding tool wear amount; form a tool wear curve with the tool wear amount as the vertical coordinate and the tool pass as the horizontal coordinate, and divide the curve into the initial wear stage, the normal wear stage and the rapid wear stage by using the double-section regression algorithm, to obtain the tool wear state corresponding to the multi-channel force signals and multi-channel vibration signals under different tool passes; the tool wear state includes the initial wear stage, the normal wear stage and the rapid wear stage;

[0008] The multi-channel force signals and the multi-channel vibration signals under different tool passes are preprocessed to obtain principal component signals of the multi-channel force signals and the multi-channel vibration signals;

[0009] The principal component signals of the multi-channel force signals and the multi-channel vibration signals are subjected to three-dimensional variational modal decomposition to obtain a series of sub-modal components IMF;

[0010] Correlation coefficients between the principal component signals of the multi-channel force signals and the multi-channel vibration signals and respective corresponding IMF components are calculated, and an IMF component with a high correlation coefficient is selected as a principal IMF component;

[0011] Time-domain features, frequency-domain features and time-frequency domain features of each principal IMF component are calculated, and features with large contributions to tool wear state classification are reserved according to a ReliefF feature selection method to realize dimension reduction of feature quantities;

[0012] The reduced features are subjected to secondary dimension reduction through unsupervised learning SDAE;

[0013] The features after secondary dimension reduction are fused to form a multi-dimensional feature vector, and the tool wear state is used to constitute training data of a model;

[0014] In combination with the training data, an OSELM is optimized by using a Scarabaeus optimization algorithm to train a DBO-OSELM model; and the multi-dimensional feature vector of a tool to be identified is input into the trained DBO-OSELM model to realize identification of the tool wear state.

[0015] Further, the multi-channel force signals and the multi-channel vibration signals of the tool are force signals and vibration signals of the tool in X, Y and Z directions.

[0016] Further, the wear values of the tool flank faces of the 4 side edges of the milling tool are measured, and the 4 wear values are averaged to obtain a tool wear amount of each tool pass.

[0017] Further, the double-section regression algorithm includes:

[0018] 1) establishing a tool wear curve;

[0019] 2) selecting a first section regression analysis region according to the tool pass number;

[0020] 3) dividing sample points in the first section regression analysis region, taking the tool pass number 1-10 and the first multiple of the tool wear amount corresponding to the tool pass number, performing least square regression fitting, and calculating the slope of the fitted straight line;

[0021] 4) Repeat step 3), and select 5-15, 10-20, 15-25 in turn until the end of the first segmented regression analysis region, and find a group of sample points such that the difference between the slope of the fitting line of the previous group of sample points is less than the first threshold value ε1;

[0022] 5) The number of tool passes in the middle of the group of sample points found in step 4) is taken as the initial tool wear division point;

[0023] 6) According to the number of tool passes, select the initial tool wear division point to the last tool pass in the tool wear curve as the second segmented regression analysis region;

[0024] 7) Divide the sample points in the second segmented regression analysis region, take the initial tool wear division point to the next 10 points and the first multiple of the tool wear amount corresponding to the tool pass number, perform least squares regression fitting, and calculate the slope of the fitting line;

[0025] 8) Repeat step 7), and select the last tool pass according to the selection method of step 4) to find a group of sample points such that the difference between the slope of the fitting line of the previous group of sample points is greater than the second threshold value ε2;

[0026] 9) The number of tool passes in the middle of the group of sample points found in step 8) is taken as the tool rapid wear division point;

[0027] 10) By determining the two division points, the tool wear curve is divided into initial wear stage, normal wear stage and rapid wear stage.

[0028] Further, the first two-thirds of the tool wear curve is selected as the first segmented regression analysis region.

[0029] Further, the first multiple is 1000 times, ε1≤0.01, and ε2≥0.4.

[0030] Further, the three-dimensional variational modal decomposition is as follows:

[0031]

[0032] In the formula, X(t) represents the principal component signal of the multi-channel force signal or the multi-channel vibration signal, C i (t) represents the i-th IMF component, and R(t) represents the residual component;

[0033] For a multi-channel signal, each IMF component C i (t) is a vector composed of IMF components of multiple channels, that is:

[0034] C i (t) = [c i,1 (t), c i,2(t), c i,3 (t)] T

[0035] In the formula, c i,1 (t) represents the i th IMF component of the first channel, c i,2 (t) represents the i th IMF component of the second channel, c i,3 (t) represents the i th IMF component of the third channel.

[0036] Further, the time domain features include mean, variance, square root amplitude, absolute average amplitude, maximum value, minimum value, peak-to-peak value, root mean square value, skewness, kurtosis, standard deviation, peak factor, waveform factor, pulse factor, margin factor and kurtosis factor; the frequency domain features include frequency domain amplitude average value, barycenter frequency, mean square frequency, frequency variance, root mean square frequency, frequency amplitude variance, frequency domain amplitude skewness index, frequency domain amplitude kurtosis index, frequency standard deviation, frequency domain frequency skewness, frequency domain frequency kurtosis and square root ratio; the time-frequency domain features include energy entropy, approximate entropy, sample entropy, fuzzy entropy, information entropy and envelope entropy.

[0037] Further, the decomposition modal number K is 6, and the first r IMF components of the correlation coefficient are selected as the main IMF components for each channel; the feature selection method of ReliefF is used for feature selection of the features of the force signal and the vibration signal respectively, wherein the selection threshold of the force signal is 0.160, and the selection threshold of the vibration signal is 0.065.

[0038] Further, the OSELM is optimized by using the OBO algorithm, and the OSELM includes:

[0039] The multi-dimensional feature vector is taken as input, and the tool wear state is taken as a division label.

[0040] The parameters of the OBO algorithm and the OSELM are initialized.

[0041] The positions of the rolling ball OBO, the breeding OBO, the foraging OBO and the stealing OBO are updated.

[0042] The fitness value is calculated, and the global optimal position is updated.

[0043] It is judged whether the maximum iteration number is reached, and if not, the iteration is continued, and the optimal hidden layer input weight and bias are output.

[0044] The optimal hidden layer input weight and bias are taken as the parameters of the OSELM, and the DBO-OSELM model is obtained by training the training data.

[0045] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0046] 1. The double segmented regression algorithm selects the most suitable calibration point to calibrate the tool wear stage, provides better parameters for subsequent feature dimension reduction and classification recognition, and has higher tool wear recognition accuracy and faster efficiency.

[0047] 2. The three-dimensional variational modal decomposition method can directly decompose multi-channel vibration signals and force signals, and further extract signal features, not only limited to single channel and single signal, so that the tool wear recognition efficiency is higher.

[0048] 3. The algorithm filters different channels and different IMF component features, automatically assigns different weights to the extracted features, realizes adaptive fusion of features, removes invalid feature values, and uses unsupervised learning for secondary feature dimension reduction, combines supervised learning and unsupervised learning algorithm, realizes feature optimization, integrates time domain, frequency domain and time-frequency domain features, expands feature dimension, and improves the prediction accuracy of real tool wear.

[0049] 4. The DBO is used to realize tool wear feature classification and recognition of the OSELM model, compared with other classification algorithms, the method has comprehensive advantages in classification accuracy and calculation efficiency, OSELM updates the model with new data, without discarding previous learning results, and through optimization, can quickly and accurately identify three wear states of initial wear, normal wear and rapid wear. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 It is a milling tool wear recognition method flow chart of the application;

[0051] Figure 2 It is a double segmented regression algorithm for tool real wear curve interval division schematic diagram of the application;

[0052] Figure 3 It is a three-dimensional variational modal decomposition signal decomposition diagram of the application; Figure 3 (a) is a vibration signal first channel decomposition component diagram, Figure 3 (b) is a vibration signal second channel decomposition component diagram, Figure 3 (c) is a vibration signal third channel decomposition component diagram;

[0053] Figure 4 It is a relief F feature extraction importance arrangement diagram of the application;

[0054] Figure 5 It is a relief F feature extraction accuracy diagram of the application;

[0055] Figure 6 The DBO optimizes the OSELM flowchart of the present application;

[0056] Figure 7 The DBO optimizes the OSELM accuracy chart of the present application. DETAILED DESCRIPTION

[0057] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0058] The application discloses a milling tool wear recognition method based on three-dimensional variational modal decomposition feature fusion, first, the vibration signal and force signal of the tool in different wear stages are obtained through multiple sensors, the tool wear curve is fitted, and a double subsection regression algorithm is used for classifying the signals; the principal component signal is obtained by intercepting the signals; then, based on three-dimensional variational modal decomposition, the multiple signals are decomposed into a series of intrinsic modal functions; the principal IMF component is screened out by using the correlation coefficient; the time-frequency domain characteristics of the principal IMF component are obtained by analyzing the principal IMF component; next, the time domain, frequency domain and time-frequency domain characteristic value calculation mode are used to extract the characteristic values of the signals; then, the extracted features are automatically given different weights by using the ReliefF algorithm to realize the first dimension reduction of the features; a new SDAE neural network is trained by using the first dimension reduction features to realize the second dimension reduction to obtain low-dimensional features, the characteristic values used are fused into a feature vector, and the feature vector is input into a DBO-OSELM for training and testing, and the obtained model is used for processing guidance. The application can effectively recognize the tool wear condition through different types of signals, can directly process multiple channel signals and is more efficient, solves the problems of relying on manual experience to select wear labels and characteristic values, and optimizes the OSELM by using the DBO, compared with other traditional algorithms, the recognition accuracy is higher, and a new technical means is provided for tool wear detection.

[0059] The milling tool wear recognition method based on three-dimensional variational modal decomposition feature fusion of the present application comprises the following steps:

[0060] Step 1: first, a monitoring system is laid out, multiple sensors are used to obtain signal data of the tool in different health states and obtain the tool wear amount, and then a double subsection regression algorithm is used to classify the unlabeled samples;

[0061] Step 2: the signal data is preprocessed to obtain a principal component signal data set;

[0062] Step 3: three-dimensional variational modal decomposition is carried out on the multi-channel principal component signal data set, and a series of modal components IMF are obtained;

[0063] Step 4: the correlation coefficients of the IMF components and the original signal are calculated, and r IMF components with higher correlation coefficients are selected as the main IMF components for each channel;

[0064] Step 5: the time domain features, frequency domain features and time-frequency domain features of the main IMF components are calculated, the primary selected features with a weight value greater than a threshold H are reserved according to the ReliefF feature selection method, and the dimensionality reduction of the feature quantities is realized;

[0065] Step 6: secondary dimensionality reduction is performed on the features after the first dimensionality reduction through unsupervised learning SDAE;

[0066] Step 7: a feature value space under different categories is constructed, and a multi-dimensional feature vector is formed through multi-feature value fusion;

[0067] Step 8: the feature vector is input into the DBO-OSELM model, and then the tool wear state recognition is realized. The tool wear states include initial wear, normal wear and rapid wear.

[0068] In some embodiments of the application, the signal data in step 1 includes vibration signals and cutting force signals in X, Y and Z directions, respectively denoted as x n (t), wherein n=1 and 2, t=1, 2, 3…N, and N is the number of sampling points. The tool wear value is measured while monitoring the signal and with the number of passes, the wear values of the four flank faces of the milling cutter are measured, and the average of the four wear values is taken as the milling cutter wear amount of each pass.

[0069] The vibration signal and force signal samples during tool milling are collected, and the wear amounts are sequentially corresponding according to the number of passes. The obtained is an unlabeled sample. In order to better realize tool wear recognition, the present application classifies the unlabeled sample by using a double segmented regression algorithm to obtain a sample with a label, and inputs it into a neural network for training to obtain a neural network model.

[0070] The unlabeled sample is classified by the double segmented regression algorithm according to the number of passes and the tool wear value, and the label is used to represent: initial wear, normal wear and rapid wear.

[0071] The specific steps of classifying the unlabeled sample by using the double segmented regression algorithm are as follows:

[0072] 1) establish a tool wear curve corresponding to the number of passes and the milling cutter wear amount;

[0073] 2) according to the number of passes, select two-thirds of the wear curve as the first segmented regression analysis area;

[0074] 3) Sample point division is performed on the first segmented regression analysis region, the tool number 1-10 and the 1000 times of the actual tool wear value corresponding to the tool number are taken, the sample points are fitted by the least square method, and the slope of the fitted straight line is calculated;

[0075] 4) Step 3) is repeated, 5-15, 10-20, 15-25 are sequentially selected until two-thirds of the total tool number is selected, and a group of sample points is found such that the slope difference of the fitted straight line of the previous group of sample points is less than the threshold value ε1;

[0076] 5) The middle tool number of the group of sample points found in step 4) is taken as the tool initial wear division point;

[0077] 6) According to the tool number, the tool wear curve is selected from the tool initial wear division point to the last tool as the second segmented regression analysis region;

[0078] 7) Sample point division is performed on the second segmented regression analysis region, the tool number from the initial wear division point to the next 10 points and the 1000 times of the actual tool wear value corresponding to the tool number are taken, the sample points are fitted, and the slope of the fitted straight line is calculated;

[0079] 8) Step 7) is repeated according to the selection method of step 4) until the last tool is selected, and a group of sample points is found such that the slope difference of the fitted straight line of the previous group of sample points is greater than the threshold value ε2;

[0080] 9) The middle tool number of the group of sample points found in step 8) is taken as the tool rapid wear division point;

[0081] 10) By determining the two division points, the tool initial wear, normal wear and rapid wear are calibrated, and the labels of the collected vibration signal and force signal samples can be obtained.

[0082] The double segmented regression algorithm is to segment the curve, divide the curve into two parts and fit multiple linear models, and the segmentation positions are set to the first tool number to two-thirds of the total tool number and the initial division point to the last tool. The double segmented regression algorithm is used in the tool wear curve calibration process, the first is to divide the initial wear stage and the normal wear and rapid wear stage, and the second is to divide the normal wear stage and the rapid wear stage. The slope of the least square fitting straight line represents the tool wear rate of the interval, and if the slope value of the next linear model and the slope value of the previous linear model change little or greatly, it indicates that the trend of the sample point data is stable or mutated.

[0083] In some embodiments of the present application, step 2 pre-processes the signal data. The measured signal length is first selected to be 100,000, with the middle stationary cutting signal information points being divided into 1,000 groups, and the mean value of each group being calculated to obtain the principal component signal x n (i), i = 1, 2, 3 …… 1000.

[0084] In some embodiments of the present application, step 3 uses three-dimensional variational modal decomposition to directly process the three-dimensional signal, with the decomposition modal number K being 6 and the penalty factor a being 2000, to obtain T groups of modal components, where T = 3 in the present embodiment, corresponding to 3 channels of vibration signals and force signals, and each channel decomposes to obtain 6 modal components, denoted as j = 1, 2, 3 …… 6, T = 1 represents 6 modal components of the first channel, T = 2 represents 6 modal components of the second channel, and T = 3 represents 6 modal components of the third channel.

[0085] In the present embodiment, the vibration signal has a total of 3 channels, each of which decomposes to obtain 6 modal components, so the vibration signal has a total of 18 modal components; the force signal also obtains 18 modal components.

[0086] The principal signal is subjected to three-dimensional variational modal decomposition, and each signal channel decomposes to obtain 6 IMF components. It should be noted that although three-dimensional variational modal decomposition can only decompose the vibration signal and the force signal in sequence, it can directly decompose multiple channels of a certain signal.

[0087] The three-dimensional variational modal decomposition formula is as follows:

[0088]

[0089] where X(t) represents the original multi-channel signal, C i (t) represents the i-th IMF component, and R(t) represents the residual component.

[0090] For a multi-channel signal, each IMF component C i (t) is a vector composed of IMF components of each channel, i.e.:

[0091] C i (t) = [c i,1 (t), c i,2 (t), c i,3 (t)] T

[0092] The vibration signal and the force signal are both three-channel signals, and compared with single-channel signal decomposition, the IMF components in multi-channel signal decomposition become vectors.

[0093] In some embodiments of the present application, step 4 calculates the decomposition signal x n (i) the correlation coefficient of the corresponding IMF component

[0094]

[0095] wherein, and respectively represent the average value of x n (i) and , and the first r IMF components with larger correlation coefficients are retained. In this embodiment, r is 3, but r can also take other values as needed.

[0096] In some embodiments of the present application, the time-domain features of the main IMF component calculated in step 5 include mean, variance, root amplitude, absolute average amplitude, maximum value, minimum value, peak-to-peak value, root mean square value, skewness, kurtosis, standard deviation, peak factor, waveform factor, pulse factor, margin factor and kurtosis factor; the frequency-domain features of the main IMF component include frequency-domain amplitude average value, barycenter frequency, mean square frequency, frequency variance, root mean square frequency, frequency amplitude variance, frequency-domain amplitude skewness index, frequency-domain amplitude kurtosis index, frequency standard deviation, frequency-domain frequency skewness, frequency-domain frequency kurtosis and square root ratio; and the time-frequency domain features of the main IMF component include energy entropy, approximate entropy, sample entropy, fuzzy entropy, information entropy and envelope entropy.

[0097] ReliefF can select different features for different signals, and the features with high correlation are retained and the features with low correlation are removed. First, a normalized feature value set F is obtained, a data set X is randomly selected from F, k nearest neighbor sample sets H j (j = 1, 2, 3 …… k) of the same class as X and k nearest neighbors M j (C) of different classes from X, wherein j = 1, 2, 3 …… k, C ≠ class(X), wherein C is the number of sample classes, class(X) is the sample class, the weight is repeatedly updated m times, the feature weight W is obtained, the larger the feature weight after the feature set is screened by the ReliefF algorithm, the greater the contribution of the feature to the tool wear classification, and vice versa. The significant features with a weight greater than a threshold H are selected for the second feature dimension reduction.

[0098]

[0099] In the formula: diff(A, X, H j ) is the difference between the sample X and H j on feature A, C is a class different from the sample X, p(C) is the probability of the Cth target class, class(X) is the class of the sample X, and M j ​(C) is the jth nearest neighbor of sample X in the Cth class target.

[0100] The features with a selection probability greater than H are selected as the primary features, wherein the selection threshold of the force signal is 0.160, and the selection threshold of the vibration signal is 0.065.

[0101] In some embodiments of the present application, the features after the first dimension reduction are subjected to secondary dimension reduction by SDAE.

[0102] h (l) =f(W (l) h (l-1) +b (l) )

[0103] wherein h (l) represents the output of the lth layer encoder, W (l) and b (l) respectively represent the weight matrix of the lth layer and the bias vector of the lth layer, and f represents an activation function.

[0104] SDAE refers to a stacked denoising autoencoder, which stacks multiple autoencoders together while masking the real information with noise at the input of the encoder, so as to improve the robustness and generalization ability of the autoencoder and help prevent overfitting, forming a deep neural network. In the process of feature dimension reduction of the feature vector, the random gradient descent method is used to adjust the network parameters, so that the error between the reconstructed features and the input features is minimized.

[0105] In some embodiments of the present application, the multi-dimensional feature vector comprises a dimension-reduced time-domain feature vector, a dimension-reduced frequency-domain feature vector, and a dimension-reduced time-frequency domain feature vector.

[0106] In some embodiments of the present application, the dung beetle algorithm (DBO) optimizes the OSELM to improve the classification accuracy, and the specific steps are as follows:

[0107] (81) The preferred feature vector obtained in step 7 is used as the training data of OSELM, and the data set is divided into a training set and a test set in a ratio of 6:4, and the tool wear condition is used as the division label;

[0108] (82) Initialize the parameters of the dung beetle optimization algorithm and OSELM;

[0109] (83) Update the positions of the rolling dung beetle, the breeding dung beetle, the foraging dung beetle, and the stealing dung beetle;

[0110] (84) Calculate the fitness value and update the global optimal position;

[0111] (85) Determine whether the maximum number of iterations is reached, if not, continue iteration, and output the optimal hidden layer input weight and bias;

[0112] (86) The optimal hidden layer input weight and bias are taken as parameters of the OSELM, and the DBO-OSELM model is obtained by training with the training set;

[0113] (87) The DBO-OSELM is tested with the test set.

[0114] The tool wear prediction model is constructed by using the signal feature processing method and the neural network, and the tool wear state is predicted with high precision. Since the amount of data to be processed by the learning model is large, the operation speed is slow, the traditional neural network has low and unstable classification accuracy, the OSELM has excellent "learning ability", and the DBO is used to optimize the OSELM, so that better tool wear recognition effect is achieved.

[0115] The application will be described in detail below in combination with the drawings and embodiments.

[0116] As shown in Figure 1 , the milling tool wear detection method based on three-dimensional variational modal decomposition feature fusion of the embodiment comprises the following steps:

[0117] Step 1: Paste force sensors and acceleration sensors on a workpiece of a mechanical milling machine, acquire sensor signals of tool cutting in the milling process, and obtain a real tool wear amount; the acquisition channels of the force sensors and the acceleration sensors are both 3, that is, the force signals and vibration signals x n (t) wherein n = 1 and 2, t = 1, 2, 3,..., N, and N is the number of sampling points.

[0118] The real tool wear amount is divided according to whether a significant turning point appears in the real tool wear amount with the increase of the number of tool passes in the tool milling operation. Specifically, the real tool wear amount is taken as the vertical coordinate, and the number of tool passes is taken as the horizontal coordinate to form a real tool wear amount curve; before the tool wear curve becomes stable and changes slowly for the first time, it is recorded as the initial wear stage; between the first time when the tool wear curve becomes stable and changes slowly and the first time when a significant turning point appears, it is recorded as the normal wear stage; after the first time when a significant turning point appears, it is recorded as the rapid wear stage. As shown in Figure 2 , a double-section regression algorithm is used for regional division, and the steps of the double-section regression algorithm are as follows:

[0119] 1) Establish a tool wear curve corresponding to the number of tool passes and the milling tool wear amount;

[0120] 2) According to the number of tool passes, select two-thirds of the wear curve as the first section regression analysis area, and select the first tool to the 213th tool as the first section regression analysis area when the total number of tool passes is 320;

[0121] 3) sample point division is performed on the first segmented regression analysis region, the number of passes 1-10 and the corresponding 1000 times of the actual tool wear value under the number of passes are taken, the sample points are fitted by least square method, and the slope of the fitted straight line is calculated:

[0122]

[0123] n represents the number of sample points, x i and y i respectively represent the horizontal and vertical coordinates of the sample points, in order to minimize the sum of squares of errors, the partial derivatives of a and b are taken and set to 0:

[0124]

[0125]

[0126] Solving a and b, the formula of the least square fitting straight line is obtained:

[0127]

[0128]

[0129] The first group of sample point data is fitted with a straight line y=ax+b, where a represents the slope of the fitted straight line;

[0130] 4) repeat step 3), select 5-15, 10-20, 15-25 in turn until two-thirds of the total number of passes are selected, find a group of sample points whose slope difference with the fitting straight line of the previous group of sample points is less than threshold value ε1, ε1≤0.01;

[0131] 5) the middle pass number of the sample found in step 4) is taken as the tool initial wear division point, the fitting straight line slope of this group of sample points is 0.01 within the specified range, then the first wear interval division point is determined to be 85;

[0132] 6) according to the number of passes, the tool wear curve tool initial wear division point to the last pass is selected as the second segmented regression analysis region, i.e. 85 to 320 passes as the second analysis region;

[0133] 7) sample point division is performed on the second segmented regression analysis region, the initial wear division point to the next 10 points and the corresponding 1000 times of the actual tool wear value under the number of passes are taken, the sample points are fitted, and the slope of the fitted straight line is calculated;

[0134] 8) Repeat step 7) and select points in the same way as in step 4) until the last cut is selected. Find a set of sample points such that the difference between the slope of the fitted line of the previous set of sample points and the slope of the line is greater than the threshold ε2, where ε2 ≥ 0.4.

[0135] 9) Take the middle number of tool passes in the set of samples found in step 8) as the dividing point of sharp tool wear. The sample points are 275-285 tool passes. The difference between the slope of the fitted line and the slope of the fitted line of the previous set of sample points is 0.4. If it is within the specified range, then the second wear interval dividing point is determined to be 280.

[0136] 10) The initial wear, normal wear, and rapid wear of the tool are calibrated by determining two dividing points.

[0137] Ultimately, the tool wear state was determined to be divided into three stages: initial wear stage, normal wear stage, and rapid wear stage, based on two dividing points: 85 and 280.

[0138] Step 2: Preprocess the signal data. First, select intermediate stable cutting signal information points up to 100,000 from the measured signal length, divide them into 1000 groups of 100 each, and calculate the mean of each group. Finally, obtain the principal component signal x. n (i), i=1, 2, 3...1000.

[0139] Step 3: Use three-dimensional variational mode decomposition to decompose the three-channel force signal and the three-channel vibration signal into sub-mode components. The three-dimensional variational mode decomposition steps are as follows:

[0140]

[0141] Where X(t) represents the original multi-channel signal, C i R(t) represents the i-th IMF component, and R(t) represents the residual component.

[0142] For multi-channel signals, each IMF component C i (t) are vectors composed of the IMF components of each channel, i.e.:

[0143] C i (t)=[c i,1 (t),c i,2 (t),c i,3 (t)] T

[0144] Both vibration and force signals are three-channel signals. Compared to single-channel signal decomposition, the IMF components in multi-channel signal decomposition become vectors. Spectral segmentation of the multi-channel signal spectrum is then performed as follows: Figure 3 As shown, where, Figure 3(a) the component diagram of the first channel of the vibration signal, Figure 3 (b) the component diagram of the second channel of the vibration signal, Figure 3 (c) the component diagram of the third channel of the vibration signal.

[0145] The decomposition modal number K of the three-dimensional variational modal decomposition is 6, the penalty factor a is 2000, and 6 sub-modal components are calculated.

[0146] Step 4: Since the correlation coefficient of the sub-modal component reflects the size of the information it contains, the sub-modal component with a smaller correlation coefficient contains less information, and is considered as a noise component. The decomposition signal x n (i) the correlation coefficient of the corresponding IMF component

[0147]

[0148] wherein, and respectively represent the average value of x n (i) and The correlation coefficients of the 6 IMF components in each channel are calculated respectively, as shown in Table 1 and Table 2, and the first 3 components in each channel are selected as the main IMF component according to the correlation coefficient ratio from high to low.

[0149] Table 1 Correlation coefficient of vibration signal

[0150]

[0151]

[0152] Table 2 Correlation coefficient of force signal

[0153]

[0154] The vibration signal and the force signal each select 9 IMF components according to the correlation coefficient.

[0155] Step 5: Calculate the time domain features of the main IMF component, including mean, variance, square root amplitude, absolute average amplitude, maximum value, minimum value, peak value, root mean square value, skewness, kurtosis, standard deviation, peak factor, waveform factor, pulse factor, margin factor and kurtosis factor; further, calculate the frequency domain features of the main IMF component, including frequency domain amplitude average, center of gravity frequency, mean square frequency, frequency variance, root mean square frequency, frequency amplitude variance, frequency domain amplitude skewness index, frequency domain amplitude kurtosis index, frequency standard deviation, frequency domain frequency skewness, frequency domain frequency kurtosis and square root ratio; further, calculate the time-frequency domain features of the main IMF component, including energy entropy, approximate entropy, sample entropy, fuzzy entropy, information entropy and envelope entropy, and the calculation formula of each feature is shown in Table 3 and Table 4.​

[0156] Table 3 Time domain feature table

[0157]

[0158]

[0159] Table 4 Frequency domain feature table

[0160]

[0161]

[0162] ReliefF feature dimensionality reduction includes: obtaining a normalized feature value set F, randomly selecting a data set X from F, finding k nearest neighbor sample set H j (j = 1, 2, 3 … k) and k nearest neighbors M j (C) of X of different classes, where j = 1, 2, 3 … k, C ≠ class(X), where C is the number of sample categories, class(X) is the sample category, and the weight is updated m times, the feature weight W is obtained, the greater the feature weight after the feature set is screened by the ReliefF algorithm, the greater the contribution of the feature to the tool wear classification, and vice versa, the contribution is small.

[0163]

[0164] In the formula: diff(A, X, H j ) is the difference between sample X and H j on feature A, C is a class different from sample X, p(C) is the probability of the Cth target, class(X) is the class of sample X, and M j (C) is the jth nearest neighbor sample of sample X in the Cth target.

[0165] Select the features with a weight greater than H as the initial selected features, where the selection threshold of the force signal is 0.160 and the selection threshold of the vibration signal is 0.065.

[0166] In step 1, the three wear stages of the tool divided by the double segmented regression algorithm are used as the sample categories of the feature data set, whether the time domain features, frequency domain features and time-frequency domain features contribute to the tool wear classification is checked, and they are arranged according to the size of the contribution, such as Figure 4As shown, the vibration signal and the force signal finally selected weight greater than H1=0.065 and H2=0.160 dimensionality reduction retained features include: vibration signal one channel IMF3 energy entropy, vibration signal one channel IMF3 fuzzy entropy, vibration signal one channel IMF3 mean square frequency and a series of characteristics, vibration signal and force signal 50 feature parameters.

[0167] As shown in Figure 5 , the SVM neural network is established, the first 4 features with the largest weight after ReliefF feature dimensionality reduction are selected as the feature data set, the feature data set is input to the input layer of the SVM neural network, the tool wear categories corresponding to the feature data set are input to the output layer of the SVM neural network, the SVM neural network is trained, and a feature dimensionality reduction neural network model is constructed. The percentage of 70 of all feature data sets is used as the training data of the model, and the percentage of 30 is used as the test data, which can verify the accuracy of the ReliefF feature dimensionality reduction.

[0168] Step 6: The features after the first dimensionality reduction in step 5 are subjected to secondary dimensionality reduction by SDAE.

[0169] h (l) =f(W (l) h (l-1) +b (l) )

[0170] Where h (l) represents the output of the l-layer encoder, W (l) and b (l) represent the weight matrix of the l-layer and the bias vector of the l-layer respectively, and f represents the activation function, which is ReLU here.

[0171] The network is composed of multiple autoencoder layers, each of which is composed of an encoder and a decoder. SDAE uses denoising technology in the training process, which means adding noise to the input data to strengthen the model to learn the true structure of the data. For the 50-dimensional feature vector of the vibration signal and the force signal, input the first layer of the SDAE network as the input of the autoencoder. Then, by training the autoencoder, a compact representation of the 50-dimensional feature vector is learned, i.e. encoding. The first layer of the autoencoder output is used again to train the autoencoder to obtain a new encoding. This process is repeated multiple times, finally mapping the high-dimensional features to a low-dimensional encoding space, and obtaining a 5-dimensional feature vector of the vibration signal and the force signal.

[0172] Step 7: The 5-dimensional feature vectors of the vibration signal and the force signal are combined into a multi-dimensional feature vector, which is labeled as 1, 2 and 3 according to the double-section regression algorithm in step 1. The multi-dimensional feature vector is labeled as the DBO-OSELM model data.

[0173] Step 8: as shown in Figure 6 The dung beetle optimization algorithm (DBO) optimizes the OSELM to improve the classification accuracy, and the specific steps are as follows:

[0174] (81) The preferred feature vector obtained in step 7 is used as the training data of the OSELM, and the data set is divided into a training set and a test set in a ratio of 6:4, and the tool wear condition is used as the division label;

[0175] (82) Initialize the parameters of the dung beetle optimization algorithm and the OSELM;

[0176] (83) Update the positions of the rolling dung beetle, the breeding dung beetle, the foraging dung beetle and the stealing dung beetle;

[0177] (84) Calculate the fitness value and update the global optimal position;

[0178] (85) Determine whether the maximum number of iterations is reached, if not, continue iteration, and output the optimal hidden layer input weight and bias;

[0179] (86) The optimal hidden layer input weight and bias are used as the parameters of the OSELM, and the training set is used to train the DBO-OSELM model;

[0180] (87) Test the DBO-OSELM with the test set.

[0181] Since there is no prior knowledge about the given data, the OSELM adopts a Sigmoid function:

[0182]

[0183] Finally, the optimal parameters of the hidden layer input weight and bias are determined, and the accuracy is verified, and different batches of data are verified under the same experimental parameters. The test signal is input into the DBO-OSELM to obtain the accuracy as shown in Figure 7 .

[0184] The feature vector normalization processing adopts the maximum and minimum normalization method, and the numerical values of the time domain feature, the frequency domain feature and the time-frequency domain feature vector are transformed to the interval 0~1, so as to improve the fitting speed of the OSELM neural network.

[0185] The prediction accuracy comparison table of the OSELM neural network model without optimization, the traditional neural network model and the neural network model of the present application is shown in Table 5, and the accuracy of the model evaluation index is improved compared with the commonly used model.

[0186] Table 5 Accuracy table

[0187]

[0188] While the embodiments of the application have been illustrated and described in detail, it will be readily apparent to those skilled in the art that various modifications and changes can be made therein without departing from the spirit and scope of the application as disclosed in the claims. It is therefore desired that what is claimed be understood to be within the true spirit and scope of the application.

[0189] The various embodiments in the specification are described in progressive order, each embodiment highlighting a difference from the previous embodiment, and the same or similar parts between the various embodiments can be mutually referred to. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method.

[0190] The principles and implementation manners of the present application are described by using specific examples in the specification, and the above description of the embodiments is only used to help understand the method of the present application and its core idea; meanwhile, for the general skilled in the art, the specific implementation manners and application ranges can be changed according to the idea of the present application. In conclusion, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for recognizing the wear of a milling tool based on three-dimensional variational mode decomposition feature fusion, characterized in that, The method includes the following steps: The system acquires multi-channel force and vibration signals of the tool at different pass counts, along with the corresponding tool wear. A tool wear curve is formed with tool wear as the ordinate and pass count as the abscissa. A bi-segmented regression algorithm is used to divide the curve into initial wear, normal wear, and rapid wear stages, thus obtaining the tool wear state corresponding to the multi-channel force and vibration signals at different pass counts. The tool wear state includes the initial wear stage, normal wear stage, and rapid wear stage. Preprocessing is performed on the multi-channel force signal and multi-channel vibration signal under different number of passes to obtain the principal component signals of the multi-channel force signal and multi-channel vibration signal; Three-dimensional variational mode decomposition was performed on the principal component signals of multi-channel force signals and multi-channel vibration signals to obtain a series of sub-modal components (IMFs). Calculate the correlation coefficients between the principal component signals of the multi-channel force signal and the multi-channel vibration signal and their respective IMF components, and select the IMF component with the high correlation coefficient as the principal IMF component; Calculate the time-domain, frequency-domain, and time-frequency-domain features of each principal IMF component. Based on the ReliefF feature selection method, retain the features that contribute significantly to the classification of tool wear state, thereby achieving dimensionality reduction of the feature quantity. The dimensionality-reduced features are then subjected to a second dimensionality reduction using unsupervised learning SDAE. The features after the second dimensionality reduction are fused to form a multidimensional feature vector, which together with the tool wear state constitutes the training data of the model; By combining training data, the OSELM model is optimized using the dung beetle optimization algorithm, and the DBO-OSELM model is trained. The multidimensional feature vector of the tool to be identified is input into the trained DBO-OSELM model to realize the identification of tool wear state.

2. The method of claim 1, wherein the method of identifying the wear of the milling tool based on the feature fusion of the three-dimensional variational mode decomposition is characterized by, The multi-channel force signal and multi-channel vibration signal of the cutting tool are the force signal and vibration signal of the cutting tool in the X, Y and Z directions, respectively. 3.The milling tool wear recognition method based on three-dimensional variational mode decomposition feature fusion according to claim 1, characterized in that, Measure the wear value of the back face of the four side edges of the milling cutter, and calculate the average of the four wear values ​​as the tool wear amount for each pass.

4. The method of claim 1, wherein the method of identifying the wear of the milling tool based on the feature fusion of the three-dimensional variational modal decomposition is characterized by, The bi-segmented regression algorithm is divided into the following categories: 1) Establish tool wear curves; 2) Select the initial segmented regression analysis region based on the number of cutter passes; 3) Divide the sample points in the first piecewise regression analysis area, take the number of passes 1-10 and the first multiple of the tool wear amount under the corresponding number of passes, perform least squares regression fitting, and calculate the slope of the fitted line; 4) Repeat step 3) sequentially select 5-15, 10-20, 15-25 until the end of the first segmented regression analysis region, find a set of sample points such that the difference between the slope of the fitted line of the previous set of sample points and the slope of the fitted line is less than the first threshold ε1. 5) Take the middle number of tool passes in the set of sample points found in step 4) as the tool initial wear division point. 6) Based on the number of passes, select the area from the initial tool wear division point in the tool wear curve to the last pass as the region for the second piecewise regression analysis; 7) sample points are divided in the second segmented regression analysis area, the first multiple of tool wear amount corresponding to the tool initial wear division point and the next 10 points and the tool wear amount at the corresponding tool feed number are taken, least square method regression fitting is carried out, and the slope of the fitting straight line is calculated; 8) repeat step 7), select according to the selection mode of step 4) until the last tool feed is selected, find a group of sample points such that the difference between the slope of the fitting straight line of the previous group of sample points is greater than the second threshold value ε2; 9) the middle tool feed number in the group of sample points found in step 8) is taken as the tool rapid wear division point; 10) the tool wear curve is divided into initial wear stage, normal wear stage and rapid wear stage by the two division points determined.

5. The method of claim 4, wherein the method of identifying the wear of the milling tool based on the feature fusion of the three-dimensional variational modal decomposition is characterized by, The first two-thirds of the tool wear curve is selected as the first segmented regression analysis area.

6. The method of claim 4, wherein the method of identifying the wear of the milling tool based on the feature fusion of the three-dimensional variational modal decomposition is characterized by, The first multiple is 1000 times, ε1≤0.01, and ε2≥0.

4.

7. The method of claim 1, wherein the method of identifying the wear of the milling tool based on the feature fusion of the three-dimensional variational modal decomposition is characterized by, Three-dimensional variational modal decomposition is as follows: In the formula, X(t) represents a principal component signal of a multi-channel force signal or a multi-channel vibration signal, C i (t) represents the i-th IMF component, and R(t) represents a residual component. For multi-channel signals, each IMF component C i (t) is a vector consisting of IMF components of the multiple channels, i.e.: C i (t) = [c i,1 (t),c i,2 (t),c i,3 (t)] T where c i,1 (t) denotes the i-th IMF component of the first channel, c i,2 (t) denotes the i-th IMF component of the second channel, c i,3 (t) denotes the i-th IMF component of the third channel.

8. The method of claim 1, wherein, Time domain features include mean, variance, root amplitude, absolute average amplitude, maximum value, minimum value, peak-to-peak value, root mean square value, skewness, kurtosis, standard deviation, peak factor, waveform factor, pulse factor, margin factor and kurtosis factor; Frequency domain features include frequency domain amplitude average, barycenter frequency, mean square frequency, frequency variance, root mean square frequency, frequency amplitude variance, frequency domain amplitude skewness index, frequency domain amplitude kurtosis index, frequency standard deviation, frequency domain frequency skewness, frequency domain frequency kurtosis and square root ratio; Time-frequency domain features include energy entropy, approximate entropy, sample entropy, fuzzy entropy, information entropy and envelope entropy. 9.The milling tool wear recognition method based on three-dimensional variational mode decomposition feature fusion according to claim 1, characterized in that, The number of decomposition modes K is 6, the first r IMFs are selected as the main IMFs in each channel, and r is 3; the features of the force signal and the vibration signal are selected by the ReliefF feature selection method, wherein the selection threshold of the force signal is 0.160, and the selection threshold of the vibration signal is 0.

065.

10. The method of claim 1, wherein, The OSELM is optimized by using the beetle optimization algorithm, which includes: Taking a multi-dimensional feature vector as input and taking a tool wear state as a division label; Initializing each parameter of the beetle optimization algorithm and OSELM; Updating the positions of the rolling ball beetle, breeding beetle, foraging beetle and stealing beetle; Calculating the fitness value and updating the global optimal position; Determine whether the maximum number of iterations is reached, if not, continue iteration, and output the optimal hidden layer input weight and bias; The optimal hidden layer input weight and bias are taken as the parameters of the OSELM, and the DBO-OSELM model is obtained by training the training data.

Images