An online identification method for milling geometric parameters based on interpretable machine learning

Abstract

The application provides a milling geometric parameter online identification method based on an interpretable machine learning, and specifically comprises the following steps: performing tool geometric modeling by using a discrete micro-element method, and performing milling force mechanism modeling based on the tool geometric modeling; performing quantitative characterization on simulation signals by using characteristic values such as time domain, frequency domain and waveform parameters, and obtaining characteristics sensitive to milling geometric parameters and insensitive to milling force coefficients based on characteristic sensitivity analysis; kurtosis, skewness, waveform factor, peak factor, pulse factor and margin factor of milling resultant force; and establishing a quantitative relationship model between the characteristics and the milling geometric parameters by using machine learning algorithms such as a decision tree and K-nearest neighbor which have inherent interpretability, training the model based on a large labeled simulation data set based on a mechanism model, and realizing online monitoring of the milling geometric parameters according to the collected milling force signals. Test results show that the monitoring accuracy of the established model for the milling width and the milling depth is greater than 89%.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of advanced manufacturing technology in mechanical engineering, artificial intelligence, and data processing and analysis, specifically to an online identification method for milling geometric parameters based on interpretable machine learning. Background Technology

[0002] Monitoring the machining process is the prerequisite and foundation for achieving intelligent monitoring of machining status. Milling, due to its excellent machining efficiency, accuracy, and applicability to complex-shaped parts, is widely used in high-end manufacturing industries such as aerospace and medical abrasives. Milling parameters such as tool speed, feed rate, width at break, and depth of cut are key factors determining machining quality, efficiency, and tool life. Inappropriate milling parameters will reduce the surface quality of the machined parts and may even lead to serious losses such as workpiece scrap and machine tool damage. Online monitoring of milling parameters has become a focus of attention in both industry and academia. While tool speed and depth of cut can be obtained in real time through CNC systems, effective monitoring and identification methods for width at break and depth of cut still lack.

[0003] Milling force is one of the most important process parameters in milling. Milling force information reflects milling status information such as tool wear, tool runout, and milling parameters. Monitoring milling width and depth by comparing real-time acquired milling force signals with theoretically predicted milling force signals is currently one of the main research approaches. However, to obtain an accurate milling force mechanism model, the above methods require experimental calibration of milling force coefficients, resulting in poor model generalization ability. Secondly, researchers have conducted related studies using sensors such as ultrasound, acoustic emission, and lasers.

[0004] Furthermore, with the rapid development of artificial intelligence, constructing a relationship model between monitoring signal features and monitoring targets based on machine learning methods is crucial for achieving effective and accurate monitoring and identification of the machining process status. However, in actual milling engineering, there is a problem of insufficient labeled training data samples, and the lack of interpretability of complex machine learning models further increases the difficulty of making targeted improvements to the models based on experience or mechanistic knowledge. This results in current monitoring algorithms having problems such as low efficiency in multi-dimensional feature optimization and dimensionality reduction, and low model accuracy or generalization ability. Summary of the Invention

[0005] Purpose of the invention: To address the shortcomings of existing technologies, this invention provides an online identification method for milling geometric parameters based on interpretable machine learning. It constructs a large simulation dataset using milling force mechanism modeling, and performs feature extraction and feature sensitivity analysis based on this dataset to establish an interpretable machine learning model. This enables accurate online identification of milling geometric parameters and has good potential for widespread application.

[0006] Technical solution: An online identification method for milling geometry parameters based on interpretable machine learning, comprising the following steps:

[0007] S1: Based on the instantaneous mechanical force model of milling, establish the milling force F along the feed direction at any time. x Milling force F perpendicular to the feed direction y and the resultant force F in the plane perpendicular to the tool axis tot The display expression;

[0008] S2: The feature matrix M composed of parameters obtained from the time and frequency domain analysis of the signal. F0 The milling force signal is quantified and characterized. Sensitivity analysis is used to select the coefficient K for the tangential cutting force. tc Radial cutting force coefficient K rc Spindle speed n, feed rate v f These four variables are insensitive to the radial cutting depth a e and axial cutting depth a p Sensitive feature matrix M F2 ;

[0009] S3: Establishing a simulation big data sample set A based on the instantaneous mechanical force model of milling data The label for each data sample includes a e and a p The feature matrix is ​​constructed as M. F2 ;

[0010] S4: Employ interpretable machine learning methods to establish the feature matrix M. F2 With a e and a p The mapping relationship; interpretable machine learning methods include either decision trees or K-nearest neighbor regression;

[0011] S5: Use force sensors to collect multi-directional milling force signals during the machining process, and compare them with the simulation signals of the instantaneous mechanical force model of milling established in step 1 to verify the accuracy of the milling force mechanism model;

[0012] S6: Based on interpretable machine learning methods, a e and a p Intelligent identification, and evaluation of the rationality of milling geometry parameters based on the predicted output results.

[0013] By utilizing the milling force mechanism model, a labeled simulation dataset is established, with milling parameters as labels and milling force signals as features. This dataset can be used for qualitative analysis of the mapping relationship between milling geometric parameters and milling force signal features, while also providing training data for modeling the quantitative relationship between the two.

[0014] By employing four dimensionless features—waveform factor, peak factor, impulse factor, and margin factor—and utilizing K-nearest neighbor, decision tree, and random forest interpretable learning algorithms, the model is trained using simulation data, enabling the monitoring of two parameters: milling width and milling depth.

[0015] Using dimensionless features as model input parameters eliminates the need for milling force coefficient calibration, enhancing the model's generalization ability and interpretability. An inherently interpretable machine learning algorithm establishes a mapping model between monitoring signals and milling geometric parameters, making the model's decision-making process traceable. This avoids the uninterpretability of complex machine learning models and facilitates targeted improvements based on experience or mechanistic knowledge.

[0016] In a preferred embodiment, S1 specifically comprises:

[0017] S1.1. Represent the tool mathematically and establish the tool's Cartesian coordinate system X. T Y T Z T O T ;

[0018] Among them, Z T The axis coincides with the tool axis and its positive direction points towards the tool holder, X T -Y T The plane and the plane formed by the blade tip coincide, Y T The axis coincides with the tool tip, and the cutting edges are numbered sequentially as i=N, where N is the number of cutting edges. The numbering order is a counterclockwise rotation around the positive Z-axis, and the cutting edge numbered i=1 coincides with the Y-axis. T The axes intersect;

[0019] The cutting edge is discretized along its axial direction, and the discretized elements are numbered in two dimensions.

[0020] P Edge _ i,j =[i, j]={[1,1],[1,2]…},

[0021] Where i represents the cutting edge, and j represents the cutting edge along the Z-axis. T The discrete element index in the positive direction of the axis, [i, j]=[1,1] represents the j=1 discrete point on the i=1th blade, denoted as P. Edge _ 1,1 .

[0022] S1.2 Establish the workpiece coordinate system X W Y W Z W O W ,

[0023] Among them, X W The positive direction of the axis points to the tool feed direction, YW The positive direction of the axis points outward from the material, Z W The positive direction of the axis points towards the worktable. Let point P be... Edge _ 1,1 At time t0 in the positive Y-axis direction, then at any time t, point P Edge _ i,j The corresponding angle Φ i,j for:

[0024]

[0025] Where, Φ i,j Let P be the point Edge _ i,j By Y W The axis begins to rotate around Z W The direction angle of axis rotation ranges from [0°, 360°], ω is the angular velocity of tool rotation (rad / s), β is the tool helix angle, D is the tool diameter, and dz is the discrete spacing of the cutting edge along its axis.

[0026] Point P Edge _ i,j The corresponding instantaneous cutting thickness is expressed as:

[0027]

[0028] Among them, f t This refers to the feed rate per tooth, expressed in mm / tooth.

[0029] Angle of approach It can be represented as:

[0030] .

[0031] S1.3. Using the instantaneous mechanical force model of infinitesimal cutting force, the milling force corresponding to the infinitesimal element of the cutting edge at any time t is expressed as:

[0032]

[0033] Among them, dF t It is the tangential force of the infinitesimal cutting edge element, and its direction coincides with the direction of the instantaneous cutting velocity; The radial force of the blade element is perpendicular to the direction of the blade. And pointing to the workpiece, K tc and K rc These are the tangential force coefficient and the radial force coefficient, respectively.

[0034] Furthermore, the tangential and radial forces are transformed into forces F along the feed direction in the workpiece coordinate system through coordinate transformation. x and the force F perpendicular to the feed direction y The transformation formula is:

[0035]

[0036] (5)

[0037]

[0038] Considering the helix angle of the milling cutter and multi-edge cutting, establish the time-domain signal expression of the milling force:

[0039]

[0040] (6)

[0041]

[0042]

[0043] In the formula, g[Φ i,j ] represents the unit step function used to indicate whether the current cutting edge micro-element participates in cutting, and equation (6) is the milling force mechanism model.

[0044] In a preferred embodiment, S2 specifically comprises:

[0045] S2.1. Extract features from the time-domain signal of integer multiples of the tool rotation cycle, establish a quantitative feature matrix for milling force, and extract F respectively. x F y F tot Nineteen features in the time and frequency domains of the three force signals are extracted, specifically: maximum value, minimum value, average value, peak-to-peak value, average absolute value, variance, standard deviation, root mean square, kurtosis, skewness, waveform factor, peak factor, impulse factor, margin factor, centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation. Fifty-seven original features are extracted for each set of milling parameters, forming a feature matrix M. F0 .

[0046] S2.2, Using the coefficient of variation as the evaluation index, and the milling force coefficient K... tc and K rc As an uncertain factor, based on feature sensitivity analysis, the influence of variations in the milling force coefficient on the characteristics of the milling force signal is analyzed, and the influence is determined from the feature matrix M. F0 The dimensionless features of the resultant milling force signal are sought in the model, which are insensitive to changes in the milling force coefficient. The feature matrix formed by the obtained features is denoted as M. F1 ={F ea_tot_9 F ea_tot_10 F ea_tot_11 F ea_tot_12 F ea_tot_13 F ea_tot_14},

[0047] Design with K tc and K rc For milling force simulation experiments with variables, in order to analyze the sensitivity of the characteristics to the milling force coefficient, the coefficient of variation of the experimental characteristics is calculated:

[0048] (8)

[0049] Where, σ fea and μ fea Calculate the standard deviation and mean of the data for the same feature in the experiment.

[0050] After filtering the results of the coefficient of variation solutions corresponding to 57 features, it was found that the dimensionless features of the milling force resultant signal, including kurtosis, skewness, waveform factor, peak factor, impulse factor, and margin factor, are not sensitive to changes in the milling force coefficient. The feature matrix formed by the above six features is denoted as M. F1 ={ F ea_tot_9 F ea_tot_10 F ea_tot_11 F ea_tot_12 F ea_tot_13 F ea_tot_14}

[0051] S2.3, with feature M F1 As an evaluation indicator, with a e a p n, v f Since milling parameters are uncertain factors, a single-factor sensitivity analysis method is used to analyze the influence of variations in milling parameters on the characteristics of the milling force signal.

[0052] The design uses parameter a e a p n, v f The milling force simulation experiment is performed with the variable being n = [1000, 10000], where the values ​​of the four parameters are in the range of n = [1000, 10000]; v f =[100,1000];a e =[0.25,2.5]; a p =[0.5,5], where each parameter takes 10 values ​​in a uniform distribution.

[0053] By screening milling force simulation test data, four characteristics—waveform factor, peak factor, impulse factor, margin factor, and a—were found to be related to a. e and a p There exists an approximately monotonic relationship; based on this, the feature matrix M is selected. F2 ={ F ea_tot_11 F ea_tot_12 F ea_tot_13 F ea_tot_14} Used to solve milling geometric parameters a e and a p .

[0054] In a preferred embodiment, S4 specifically comprises:

[0055] An interpretable multi-output regression machine learning method is used to establish a quantitative relationship model between the feature matrix MF2 and the solution targets ae and ap. The interpretable multi-output regression machine learning method includes any one of K-nearest neighbor, decision tree, and random forest.

[0056] Using cutting depth and cutting width as variables, a theoretical sample set is obtained based on the milling force mechanism model. 80% of the theoretical sample set is extracted as the training set, and 20% is used as the test set. The mean absolute percentage error is used as the evaluation standard. The solution is obtained through a machine learning model. The machine learning model includes any one of K-nearest neighbor, decision tree, and random forest.

[0057] In a preferred embodiment, S5 specifically includes:

[0058] A milling experiment was designed to test titanium alloy. First, the infeed and outfeed portions of the acquired milling force signal were removed. Then, 20 segments of the milling force signal during the stable phase were extracted, each containing 1800 data points. Finally, the 20 data segments were aligned and averaged to obtain the time-domain monitoring signal of the milling force. The milling force coefficient was calibrated using the average milling force coefficient calibration method and then substituted into...

[0059] (6)

[0060]

[0061]

[0062] The theoretical samples of milling force signals were obtained and compared with the experimental data of titanium alloy milling to obtain the accuracy of the milling force mechanism model in reflecting the true value.

[0063] In a preferred embodiment, S6 specifically includes:

[0064] The milling force signal obtained from the experiment is subjected to low-pass filtering to extract features including waveform factor, peak factor, impulse factor and margin factor of the resultant milling force. These features are then input into the established K-nearest neighbor regression model to obtain prediction results. The rationality of the milling geometric parameters is evaluated based on the prediction results.

[0065] Beneficial effects: This invention utilizes a milling force mechanism model to establish a labeled simulation dataset with milling parameters as labels and milling force signals as features. This dataset can be used for qualitative analysis of the mapping relationship between milling geometric parameters and milling force signal features, while also providing training data for modeling the quantitative relationship between the two.

[0066] Using four dimensionless features—waveform factor, peak factor, impulse factor, and margin factor—and employing interpretable learning algorithms such as K-nearest neighbor, decision tree, and random forest, the model is trained on simulation data. This enables the monitoring of two parameters: milling width and milling depth. Experimental data shows that the model's solution accuracy is greater than 89%.

[0067] Using dimensionless features as model input parameters eliminates the need for milling force coefficient calibration, enhancing the model's generalization ability and interpretability. An inherently interpretable machine learning algorithm establishes a mapping model between monitoring signals and milling geometric parameters, making the model's decision-making process traceable. This avoids the uninterpretability of complex machine learning models and facilitates targeted improvements based on experience or mechanistic knowledge. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0069] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0070] Figure 2 For the mathematical and geometric model of the cutting tool;

[0071] Figure 3 A geometric diagram of the "tool-workpiece" relationship in milling machining;

[0072] Figure 4 The graph shows the sensitivity analysis of the Fx feature to the milling force coefficient.

[0073] Figure 5 The graph shows the sensitivity analysis of the Fy feature to the milling force coefficient.

[0074] Figure 6 A graph showing the sensitivity analysis of the Ftot feature to the milling force coefficient;

[0075] Figure 7 For the feature varies with a e and a p The diagram showing the changing patterns;

[0076] Figure 8 Diagram of the test system;

[0077] Figure 9 Comparative analysis of the milling force mechanism model and experimental results for group TTT1;

[0078] Figure 10 Comparative analysis of the milling force mechanism model and experimental results for the TTT2 group;

[0079] Figure 11 Comparative analysis of the milling force mechanism model and experimental results for the TTT3 group. Detailed Implementation

[0080] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0081] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0082] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0083] like Figure 1 As shown, an online method for identifying milling geometry parameters based on interpretable machine learning includes the following steps:

[0084] S1: Based on the instantaneous mechanical force model of milling, establish the milling force F along the feed direction at any time. x Milling force F perpendicular to the feed direction y and the resultant force F in the plane perpendicular to the tool axistot The display expression;

[0085] S2: The feature matrix M composed of parameters obtained from the time and frequency domain analysis of the signal. F0 The milling force signal is quantified and characterized. Sensitivity analysis is used to select the coefficient K for the tangential cutting force. tc Radial cutting force coefficient K rc Spindle speed n, feed rate v f These four variables are insensitive to the radial cutting depth a e and axial cutting depth a p Sensitive feature matrix M F2 ;

[0086] S3: Establishing a simulation big data sample set A based on the instantaneous mechanical force model of milling data The label for each data sample includes a e and a p The feature matrix is ​​constructed as M. F2 ;

[0087] S4: Use decision trees, K-nearest neighbor regression, and interpretable machine learning methods to build the feature matrix M. F2 With a e and a p The mapping relationship;

[0088] S5: Use force sensors to collect multi-directional milling force signals during the machining process, and compare them with the simulation signals of the instantaneous mechanical force model of milling established in step 1 to verify the accuracy of the milling force mechanism model;

[0089] S6: Based on interpretable machine learning methods, a e and a p Intelligent identification, and evaluation of the rationality of milling geometry parameters based on the predicted output results.

[0090] By utilizing the milling force mechanism model, a labeled simulation dataset is established, with milling parameters as labels and milling force signals as features. This dataset can be used for qualitative analysis of the mapping relationship between milling geometric parameters and milling force signal features, while also providing training data for modeling the quantitative relationship between the two.

[0091] By employing four dimensionless features—waveform factor, peak factor, impulse factor, and margin factor—and utilizing K-nearest neighbor, decision tree, and random forest interpretable learning algorithms, the model is trained using simulation data, enabling the monitoring of two parameters: milling width and milling depth.

[0092] Using dimensionless features as model input parameters eliminates the need for milling force coefficient calibration, enhancing the model's generalization ability and interpretability. An inherently interpretable machine learning algorithm establishes a mapping model between monitoring signals and milling geometric parameters, making the model's decision-making process traceable. This avoids the uninterpretability of complex machine learning models and facilitates targeted improvements based on experience or mechanistic knowledge.

[0093] In a preferred embodiment, S1 specifically comprises:

[0094] S1.1, such as Figure 2 The tool is mathematically represented, and a Cartesian coordinate system X is established for the tool. T Y T Z T O T ;

[0095] Among them, Z T The axis coincides with the tool axis and its positive direction points towards the tool holder, X T -Y T The plane and the plane formed by the blade tip coincide, Y T The axis coincides with the tool tip, and the cutting edges are numbered sequentially as i=N, where N is the number of cutting edges. The numbering order is a counterclockwise rotation around the positive Z-axis, and the cutting edge numbered i=1 coincides with the Y-axis. T The axes intersect;

[0096] The cutting edge is discretized along its axial direction, and the discretized elements are numbered in two dimensions.

[0097] P Edge _ i,j =[i, j]={[1,1],[1,2]…},

[0098] Where i represents the cutting edge, and j represents the cutting edge along the Z-axis. T The discrete element index in the positive direction of the axis, [i, j]=[1,1] represents the j=1 discrete point on the i=1th blade, denoted as P. Edge _ 1,1 .

[0099] S1.2 Establish the workpiece coordinate system X W Y W Z W O W ,

[0100] Among them, X W The positive direction of the axis points to the tool feed direction, Y W The positive direction of the axis points outward from the material, Z W The positive direction of the axis points towards the worktable, such as... Figure 3 As shown. Let point P. Edge _ 1,1At time t0 in the positive Y-axis direction, then at any time t, point P Edge _ i,j The corresponding angle Φ i,j for:

[0101]

[0102] Where, Φ i,j Let P be the point Edge _ i,j By Y W The axis begins to rotate around Z W The direction angle of axis rotation ranges from [0°, 360°], ω is the angular velocity of tool rotation (rad / s), β is the tool helix angle, D is the tool diameter, and dz is the discrete spacing of the cutting edge along its axis.

[0103] Point P Edge _ i,j The corresponding instantaneous cutting thickness is expressed as:

[0104]

[0105] Among them, f t This refers to the feed rate per tooth, expressed in mm / tooth.

[0106] Angle of approach It can be represented as:

[0107] .

[0108] S1.3. Using the instantaneous mechanical force model of infinitesimal cutting force, the milling force corresponding to the infinitesimal element of the cutting edge at any time t is expressed as:

[0109]

[0110] Among them, dF t It is the tangential force of the infinitesimal cutting edge element, and its direction coincides with the direction of the instantaneous cutting velocity; The radial force of the blade element is perpendicular to the direction of the blade. And pointing to the workpiece, K tc and K rc These are the tangential force coefficient and the radial force coefficient, respectively.

[0111] Furthermore, the tangential and radial forces are transformed into forces F along the feed direction in the workpiece coordinate system through coordinate transformation. x and the force F perpendicular to the feed direction y The transformation formula is:

[0112]

[0113] (5)

[0114]

[0115] Considering the helix angle of the milling cutter and multi-edge cutting, establish the time-domain signal expression of the milling force:

[0116]

[0117] (6)

[0118]

[0119]

[0120] In the formula, g[Φ i,j ] represents the unit step function used to indicate whether the current cutting edge micro-element participates in cutting, and equation (6) is the milling force mechanism model.

[0121] In a preferred embodiment, S2 specifically comprises:

[0122] S2.1. Extract features from the time-domain signal of integer multiples of the tool rotation cycle, establish a quantitative feature matrix for milling force, and extract F respectively. x F y F tot Nineteen features in the time and frequency domains of the three force signals are extracted, specifically: maximum value, minimum value, average value, peak-to-peak value, average absolute value, variance, standard deviation, root mean square, kurtosis, skewness, waveform factor, peak factor, impulse factor, margin factor, centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation. Fifty-seven original features are extracted for each set of milling parameters, forming a feature matrix M. F0 .

[0123] S2.2, Using the coefficient of variation as the evaluation index, and the milling force coefficient K... tc and K rc As an uncertain factor, based on feature sensitivity analysis, the influence of variations in the milling force coefficient on the characteristics of the milling force signal is analyzed. From the feature matrix M... F0 The dimensionless characteristics of the resultant milling force signal are sought, which are insensitive to variations in the milling force coefficient. The feature matrix formed by the obtained features is denoted as M. F1 ={F ea_tot_9 F ea_tot_10 F ea_tot_11 F ea_tot_12 F ea_tot_13 F ea_tot_14}

[0124] Design with K tc and K rcThe milling force simulation test was performed with variable values, and the test parameters are shown in Table 1.

[0125] Table 1 Simulation test parameters of milling force

[0126]

[0127] To analyze the sensitivity of the features to the milling force coefficient, the coefficients of variation for 9 sets of experimental features were calculated:

[0128] (7)

[0129] Where, σ fea and μ fea The standard deviation and mean of the data for the same feature are calculated in 9 groups of experiments.

[0130] like Figures 4 to 6 The figure shows the coefficient of variation results for 57 features across 9 experimental groups. Figure 6 It can be seen that F tot The coefficients of variation for features 9-14 are 0, indicating that these features do not change with variations in the milling force coefficient. Therefore, the six dimensionless features of the resultant milling force signal—kurtosis, skewness, waveform factor, peak factor, impulse factor, and margin factor—are insensitive to changes in the milling force coefficient. For ease of subsequent calculations, the feature matrix formed by these six features is denoted as M. F1 ={F ea_tot_9 F ea_tot_10 F ea_tot_11 F ea_tot_12 F ea_tot_13 F ea_tot_14}

[0131] S2.3, with feature M F1 As an evaluation indicator, with a e a p n, v f The four milling parameters are uncertain factors. A single-factor sensitivity analysis method is used to analyze the influence of the variation of milling parameters on the characteristics of the milling force signal.

[0132] The design is based on a e a p n, v f A milling force simulation experiment was conducted with four parameters as variables, the values ​​of which range from n=[1000,10000]; v f =[100,1000];a e =[0.25,2.5]; a p =[0.5,5], each parameter takes 10 values ​​according to a uniform distribution, and the experimental parameters are shown in Table 2.

[0133] Table 2 Milling process parameters

[0134]

[0135] The variables in groups TT1 and TT2 are rotational speed n and feed rate v, respectively. f The experimental results are shown in Tables 3 and 4. As can be seen from the tables, feature M... F1 Not because of n and v f Changes with the change, i.e., feature M F1 For variables n and v f Not sensitive.

[0136] Table 3 Results of the TT1 group of tests

[0137]

[0138] Table 4 Results of the TT2 group test

[0139]

[0140] The results of the TT3 group test are as follows: Figure 7 As shown, feature M is illustrated. F1 With a e and a p The changing pattern. To enhance visualization, feature M in the figure... F1 It has undergone normalization. As shown in the figure, feature M... F1 With a e and a p Changes occur with the change of a, among which the four characteristics of waveform factor, peak factor, impulse factor, margin factor and a are related to the change of a. e and a p An approximately monotonic relationship exists. Based on this, the feature matrix M is selected. F2 ={ F ea_tot_11 F ea_tot_12 F ea_tot_13 F ea_tot_14} Used to solve milling geometric parameters a e and a p .

[0141] S3: Establish a simulation big data sample set based on the mechanism model, where each data sample is labeled as a. e and a p Characteristic M F2 ;

[0142] Specifically, S4 is:

[0143] Milling geometric parameters a e and a p With F ea_tot_11 F ea_tot_12 Fea_tot_13 F ea_tot_14 The four features exhibit a non-linear relationship. Interpretable multi-output regression machine learning algorithms such as K-nearest neighbors, decision trees, and random forests are employed to establish the feature matrix M. F2 With the goal a e and a p A quantitative relationship model between the parameters was established. Taking a 4-flute solid carbide end mill with a diameter of 5mm and a helix angle of 40° as an example, the cutting width was set to [0.1D, 0.5D], and the cutting depth to [0.2D, 0.5D]. 100 samples were evenly selected for each parameter within its range, resulting in 10,000 data samples obtained through the milling force mechanism model. 9,000 samples were selected as the training set, and 1,000 as the test set. The mean absolute percentage error was used as the evaluation criterion. Each algorithm was repeated 5 times, and the solution errors are shown in Table 5. The table shows that the K-nearest neighbor, decision tree, and random forest models achieved a solution accuracy greater than 90%.

[0144] Table 5. Error of regression algorithm solution (%)

[0145]

[0146] In a preferred embodiment, S5 specifically includes:

[0147] Multi-directional milling force signals during the machining process are collected using force sensors and compared with simulation signals from the milling mechanism model established in S1 to verify the accuracy of the milling force mechanism model; the specific steps are as follows:

[0148] S5.1: Design three groups of titanium alloy milling tests. The test parameters are shown in Table 6, and the test procedures are as follows. Figure 8 As shown. The test cutter was a 4-flute flat-end carbide end mill from the Fraisa brand, with a diameter of 5mm and a helix angle of 40°; the machine tool was a MIKRON HSM600U high-speed machining center; the force gauge was a Kistler 9119AA2; climb milling and dry milling were used; and the sampling frequency was 15kHz.

[0149] Table 6 Milling Experiment Design

[0150] S5.2: First, the entry and exit portions of the acquired milling force signal are deleted. Then, 20 segments of the milling force signal in the stable phase are extracted, each containing 1800 data points (integer multiples of the tool rotation cycle). Finally, the 20 data segments are aligned and averaged to obtain the time-domain monitoring signal of the milling force. The milling force coefficient is calibrated using the average milling force coefficient calibration method and substituted into...

[0151]

[0152] (6)

[0153]

[0154]

[0155] Theoretical samples of milling force signals were obtained, and the theoretical samples of milling force signals were compared with experimental data, for example... Figures 9 to 11 As shown in the figure. The experimental results show that the established milling force mechanism model can reflect its true value well.

[0156] In a preferred embodiment, S6 specifically includes:

[0157] Based on interpretable machine learning algorithms, a e and a p Intelligent identification and evaluation of the rationality of milling geometric parameters based on the predicted output results enable monitoring and maintenance of the milling system. Low-pass filtering was applied to the milling force signals obtained from three sets of experiments to extract four features: waveform factor, peak factor, impulse factor, and margin factor of the resultant milling force. These features were then input into the K-nearest neighbor regression model established in Section 4, yielding prediction results as shown in Table 7. The table shows that, using the dimensionless features waveform factor, peak factor, impulse factor, and margin factor as input, the interpretable machine learning model trained based on simulation big data can be effectively applied to milling geometric parameters a in practical engineering. e and a p The online recognition accuracy is greater than 89%.

[0158] Table 7 Comparison of Predicted Values ​​and Actual Values

[0159]

[0160] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0161] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for online identification of milling geometry parameters based on interpretable machine learning, characterized in that: It includes the following steps: S1: Based on the instantaneous mechanical force model of milling, establish the milling force F along the feed direction at any time. x Milling force F perpendicular to the feed direction y and the resultant force F in the plane perpendicular to the tool axis tot The display expression; S2: The feature matrix M composed of parameters obtained from the time and frequency domain analysis of the signal. F0 The milling force signal is quantified and characterized. Sensitivity analysis is used to select the coefficient K for the tangential cutting force. tc Radial cutting force coefficient K rc Spindle speed n, feed rate v f These four variables are insensitive to the radial cutting depth a e and axial cutting depth a p Sensitive feature matrix M F2 ; S3: Establishing a simulation big data sample set A based on the instantaneous mechanical force model of milling data The label for each data sample includes a e and a p The feature matrix is ​​constructed as M. F2 ; S4: Employ interpretable machine learning methods to establish the feature matrix M. F2 With a e and a p The mapping relationship; interpretable machine learning methods include either decision trees or K-nearest neighbor regression; S5: Use force sensors to collect multi-directional milling force signals during the machining process, and compare them with the simulation signals of the instantaneous mechanical force model of milling established in step 1 to verify the accuracy of the milling force mechanism model; S6: Based on interpretable machine learning methods, a e and a p Intelligent identification, and evaluation of the rationality of milling geometry parameters based on the predicted output results.

2. The online identification method for milling geometry parameters based on interpretable machine learning according to claim 1, characterized in that... Specifically, S1 is: S1.

1. Represent the tool mathematically and establish the tool's Cartesian coordinate system X. T Y T Z T O T ; Among them, Z T The axis coincides with the tool axis and its positive direction points towards the tool holder, X T -Y T The plane and the plane formed by the blade tip coincide, Y T The axis coincides with the tool tip, and the cutting edges are numbered sequentially as i=N, where N is the number of cutting edges. The numbering order is a counterclockwise rotation around the positive Z-axis, and the cutting edge numbered i=1 coincides with the Y-axis. T The axes intersect; The cutting edge is discretized along its axial direction, and the discretized elements are numbered in two dimensions. P Edge _ i,j =[i, j]={ [1,1],[1,2]…}, Where i represents the cutting edge, and j represents the cutting edge along the Z-axis. T The discrete element index in the positive direction of the axis, [i, j]=[1,1] represents the j=1 discrete point on the i=1th blade, denoted as P. Edge _ 1,1 .

3. The online identification method for milling geometry parameters based on interpretable machine learning according to claim 2, characterized in that... Specifically, S1 is: S1.2 Establish the workpiece coordinate system X W Y W Z W O W , Among them, X W The positive direction of the axis points to the tool feed direction, Y W The positive direction of the axis points outward from the material, Z W The positive direction of the axis points towards the worktable. Let point P be... Edge _ 1,1 At time t0 in the positive Y-axis direction, then at any time t, point P Edge _ i,j The corresponding angle Φ i,j for: ； Where, Φ i,j Let P be the point Edge _ i,j By Y W The axis begins to rotate around Z W The direction angle of axis rotation ranges from [0°, 360°], ω is the angular velocity of tool rotation (rad / s), β is the tool helix angle, D is the tool diameter, and dz is the discrete spacing of the cutting edge along its axis. Point P Edge _ i,j The corresponding instantaneous cutting thickness is expressed as: ； Among them, f t This refers to the feed rate per tooth, expressed in mm / tooth. Angle of approach It can be represented as: 。 4. The online identification method for milling geometry parameters based on interpretable machine learning according to claim 3, characterized in that... Specifically, S1 is: S1.

3. Using the instantaneous mechanical force model of infinitesimal cutting force, the milling force corresponding to the infinitesimal element of the cutting edge at any time t is expressed as: ； Among them, dF t It is the tangential force of the infinitesimal cutting edge element, and its direction coincides with the direction of the instantaneous cutting velocity; The radial force of the blade element is perpendicular to the direction of the blade. And pointing to the workpiece, K tc and K rc These are the tangential force coefficient and the radial force coefficient, respectively. Furthermore, the tangential and radial forces are transformed into forces F along the feed direction in the workpiece coordinate system through coordinate transformation. x and the force F perpendicular to the feed direction y And the combined force of the two F tot The transformation formula is: ； ； (5) ； Considering the helix angle of the milling cutter and multi-edge cutting, establish the time-domain signal expression of the milling force: ； ； （6） ； ； In the formula, For the tool along Z T The number of infinitesimal elements of the cutting edge after discretization of the axis, g[Φ i,j ] represents the unit step function used to indicate whether the current cutting edge micro-element participates in cutting, and equation (6) is the milling force mechanism model.

5. The online identification method for milling geometric parameters based on interpretable machine learning according to claim 1, characterized in that... Specifically, S2 is: S2.

1. Extract features from the time-domain signal of integer multiples of the tool rotation cycle, establish a quantitative feature matrix for milling force, and extract F respectively. x F y F tot Nineteen features in the time and frequency domains of the three force signals are extracted, specifically: maximum value, minimum value, average value, peak-to-peak value, average absolute value, variance, standard deviation, root mean square, kurtosis, skewness, waveform factor, peak factor, impulse factor, margin factor, centroid frequency, mean square frequency, root mean square frequency, frequency variance, and frequency standard deviation. Fifty-seven original features are extracted for each set of milling parameters, forming a feature matrix M. F0 .

6. The online identification method for milling geometry parameters based on interpretable machine learning according to claim 5, characterized in that... Specifically, S2 is: S2.2, Using the coefficient of variation as the evaluation index, and the milling force coefficient K... tc and K rc As an uncertain factor, based on feature sensitivity analysis, the influence of variations in the milling force coefficient on the characteristics of the milling force signal is analyzed, and the influence is determined from the feature matrix M. F0 The dimensionless features of the resultant milling force signal are sought in the model, which are insensitive to changes in the milling force coefficient. The feature matrix formed by the obtained features is denoted as M. F1 ={F ea_tot_9 F ea_tot_10 F ea_tot_11 F ea_tot_12 F ea_tot_13 F ea_tot_14 }, Design with K tc and K rc For milling force simulation experiments with variables, in order to analyze the sensitivity of the characteristics to the milling force coefficient, the coefficient of variation of the experimental characteristics is calculated: ； （8） Where, σ fea and μ fea Calculate the standard deviation and mean of the data for the same feature in the experiment. After filtering the results of the coefficient of variation solutions corresponding to 57 features, it was found that the dimensionless features of the milling force resultant signal, including kurtosis, skewness, waveform factor, peak factor, impulse factor, and margin factor, are not sensitive to changes in the milling force coefficient. The feature matrix formed by the above six features is denoted as M. F1 ={ F ea_tot_9 F ea_tot_10 F ea_tot_11 F ea_tot_12 F ea_tot_13 F ea_tot_14 } 7. The online identification method for milling geometry parameters based on interpretable machine learning according to claim 5, characterized in that... Specifically, S2 is: S2.3, with feature M F1 As an evaluation indicator, with a e a p n, v f Since milling parameters are uncertain factors, a single-factor sensitivity analysis method is used to analyze the influence of variations in milling parameters on the characteristics of the milling force signal. The design uses parameter a e a p n, v f The milling force simulation experiment is performed with the variable being n = [1000, 10000], where the values ​​of the four parameters are in the range of n = [1000, 10000]; v f =[100,1000];a e =[0.25,2.5]; a p =[0.5,5], where each parameter takes 10 values ​​in a uniform distribution. By screening milling force simulation test data, four characteristics—waveform factor, peak factor, impulse factor, margin factor, and a—were found to be related to a. e and a p There exists an approximately monotonic relationship; based on this, the feature matrix M is selected. F2 ={ F ea_tot_11 F ea_tot_12 F ea_tot_13 F ea_tot_14 } Used to solve milling geometric parameters a e and a p .

8. The online identification method for milling geometry parameters based on interpretable machine learning according to claim 1, characterized in that... Specifically, S4 is: A feature matrix M is established using an interpretable multi-output regression machine learning method. F2 With the goal a e and a p A quantitative relationship model between them, with interpretable multi-output regression machine learning methods including any one of K-nearest neighbors, decision trees, and random forests; Using cutting depth and cutting width as variables, a theoretical sample set is obtained based on the milling force mechanism model. 80% of the theoretical sample set is extracted as the training set, and 20% is used as the test set. The mean absolute percentage error is used as the evaluation standard. The solution is obtained through a machine learning model. The machine learning model includes any one of K-nearest neighbor, decision tree, and random forest.

9. The online identification method for milling geometric parameters based on interpretable machine learning according to claim 1, characterized in that... Specifically, S5 is: A milling experiment was designed to test titanium alloy. First, the infeed and outfeed portions of the acquired milling force signal were removed. Then, 20 segments of the milling force signal during the stable phase were extracted, each containing 1800 data points. Finally, the 20 data segments were aligned and averaged to obtain the time-domain monitoring signal of the milling force. The milling force coefficient was calibrated using the average milling force coefficient calibration method and then substituted into... ； ； （6） ； ； The theoretical samples of milling force signals were obtained and compared with the experimental data of titanium alloy milling to obtain the accuracy of the milling force mechanism model in reflecting the true value.

10. The online identification method for milling geometry parameters based on interpretable machine learning according to claim 1, characterized in that... Specifically, S6 is: The milling force signal obtained from the experiment is subjected to low-pass filtering to extract features including waveform factor, peak factor, impulse factor and margin factor of the resultant milling force. These features are then input into the established K-nearest neighbor regression model to obtain prediction results. The rationality of the milling geometric parameters is evaluated based on the prediction results.

Images