A cutting tool wear prediction method based on MDRSNet

Abstract

The application relates to a cutting tool wear prediction method based on an MDRSNet, which comprises the following steps: collecting a cutting force signal and a moment signal of a cutting tool in a milling process; inputting the cutting force signal and the moment signal into a preset MDRSNet tool wear prediction model to output a predicted wear amount of the cutting tool, wherein the MDRSNet tool wear prediction model is obtained by training based on a tool wear dataset, the tool wear dataset comprises cutting force signals and moment signals of X, Y and Z three axes of different cutting tools in the milling process and a wear amount of a tool flank surface of the cutting tool after each tool feed, and the MDRSNet tool wear prediction model is constructed by introducing a multi-branch structure into a deep residual shrinkage network (DRSN). The application enhances the expression ability and learning ability of the model by combining the residual shrinkage unit with the multi-branch structure, so that the model is light-weighted while the excellent prediction precision of the model is ensured.

Description

Technical Field

[0001] This invention relates to the field of machining condition monitoring technology, and in particular to a cutting tool wear prediction method based on MDRSNet. Background Technology

[0002] Machine tools are key equipment in the manufacturing process, and cutting tools are an important component. During parts production, if cutting tool wear reaches a threshold and is not replaced, it will affect the surface finish of the workpiece, and may even lead to product scrap and machine tool damage. Currently, domestic enterprises mainly adopt a preventative maintenance strategy for CNC machine tool cutting tools, replacing tools prematurely before they reach their wear threshold. This results in nearly 62% of tool life not being utilized, leading to tool waste, increased production costs, and reduced production efficiency. Related research indicates that using a tool wear monitoring system to monitor tool wear in real time can reduce downtime by 75% and increase production efficiency by 10%-50%. Therefore, real-time tool condition monitoring and predictive maintenance are of great significance for promoting the development of intelligent cutting technology in my country.

[0003] In recent years, with the continuous development of sensor technology, computer processing technology, and artificial intelligence technology, data such as cutting force, vibration, noise, temperature, tool path texture, and wear images during tool machining can be acquired and stored in real time. This data can fully reflect the entire lifecycle of tool wear, providing necessary data support for tool wear monitoring. Currently, tool wear prediction methods can be divided into classical machine learning models and deep learning models. Compared to machine learning models, deep learning-based tool wear prediction models have superior predictive performance. However, the complex structure of deep learning models, containing a large number of computational parameters, leads to slow training speeds, long inference times, and high demands on computer performance. Furthermore, the large number of parameters requires significant storage space, placing a huge storage burden on actual production. Therefore, when establishing deep learning prediction models for tool wear, how to ensure the accuracy of model predictions while shortening training time, reducing computational costs, and minimizing memory usage under limited computing resources is a current challenge in the development of tool wear prediction technology. Summary of the Invention

[0004] The purpose of this invention is to address the problems of complex structure, large parameters, and large storage space consumption of current deep learning-based tool wear prediction models. It provides a cutting tool wear prediction method based on MDRSNet, which extracts milling force signal features through residual shrinking units, enhances the expressive power of the model through multi-branch structure, and improves the learning ability of the model by integrating the output features of different branches. While ensuring the model has excellent prediction accuracy, it also achieves lightweight model.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A cutting tool wear prediction method based on MDRSNet includes:

[0007] Acquire cutting force and torque signals from the cutting tool during the milling process;

[0008] The cutting force and torque signals are input into a preset MDRSNet tool wear prediction model, which outputs the predicted wear amount of the cutting tool. The MDRSNet tool wear prediction model is trained based on a tool wear dataset, which includes cutting force and torque signals of different cutting tools in the X, Y, and Z axes during milling, as well as the wear amount of the tool flank after each pass. The MDRSNet tool wear prediction model is constructed by introducing a multi-branch structure into the Deep Residual Shrinking Network (DRSN). The DRSN includes several layers, and different layers contain different Residual Shrinking Units (RSBUs). The RSBUs are used to extract the signal features of each layer, and the multi-branch structure is used to fuse the signal features from different layers. The multi-branch structure is divided according to the layers.

[0009] Optionally, before training the tool wear prediction model based on the tool wear dataset, the method further includes preprocessing the tool wear dataset, which includes:

[0010] A variational mode decomposition denoising algorithm based on Bayesian optimization is used to denoise the cutting force and torque signals in the tool wear dataset, and feature extraction is performed on the denoised cutting force and torque signals to obtain signal features.

[0011] The recursive feature elimination algorithm based on random forest is used to filter the signal features, obtain the optimal feature subset, and complete the preprocessing.

[0012] Optionally, the Bayesian-optimized variational mode decomposition denoising algorithm for denoising the cutting force and torque signals in the tool wear dataset includes:

[0013] S21: Initialize the Bayesian optimization algorithm, setting the maximum number of iterations, the number of modes, and the range of values ​​for the penalty function;

[0014] S22: Initialize a set of modal numbers and penalty function values ​​as test points, and perform VMD decomposition on the cutting force signal and torque signal to obtain the modal functions;

[0015] S23: Calculate the envelope entropy of the modal function, fit a Gaussian regression model, and estimate the posterior probability distribution of the minimum envelope entropy;

[0016] S24: Maximize the acquisition function to determine the next test point;

[0017] S25: Calculate the minimum envelope entropy of the next test point, add it to the observation data, and update the Gaussian regression model;

[0018] S26: Determine whether the maximum number of iterations has been reached. If it has, stop the iteration and output the optimal number of modes and the optimal penalty function; otherwise, jump to S24.

[0019] S27: Perform VMD decomposition based on the optimal number of modes and the optimal penalty function, and calculate the Pearson correlation coefficient of each mode function;

[0020] S28: Select the three mode functions with the highest Pearson coefficients to reconstruct the signal and complete the denoising process.

[0021] Optionally, the envelope entropy is calculated as follows:

[0022]

[0023]

[0024]

[0025] Where a(j) is the envelope signal sequence of signal x(j) (j=1,2,...,n) after Hibert transform; H(·) is the Hibert transform; p j Let be the normalized form of a(j); E is the envelope entropy.

[0026] Optionally, the Pearson correlation coefficient is calculated as follows:

[0027]

[0028] Where X = {x1, x2, ..., x} n} and Y = {y1, y2, ..., y n Let} be two time signal sequences of length n; xi is the i-th time signal in the time signal sequence X, and y is the ith time signal. i Let i be the i-th time signal in the time signal sequence Y; and Let X and Y represent the mean values ​​of the time signal sequences X and Y, respectively.

[0029] Optionally, the signal features include: time-domain features, frequency-domain features, and time-frequency-domain features. The time-domain features include maximum value, minimum value, peak-to-peak value, peak value, mean value, root square amplitude, variance, root mean square value, kurtosis, skewness, waveform factor, peak factor, margin factor, and energy. The frequency-domain features include frequency mean, frequency variance, centroid frequency, root mean square frequency, frequency standard deviation, standard deviation coefficient, kurtosis index, skewness index, kurtosis factor, skewness factor, and frequency 1 / 2 central moment factor. The time-frequency-domain features include the energy features of eight sub-bands after three-level wavelet packet decomposition and the energy features of four mode functions after three-level empirical mode decomposition.

[0030] Optionally, the recursive feature elimination algorithm based on random forest is used to filter the signal features, including:

[0031] S41: Set the maximum depth and number of classifiers for the random forest model;

[0032] S42: Perform replacement sampling on the signal feature data to generate k training sample sets, and update the training sample sets by feature extraction without replacement;

[0033] S43: Based on the updated training sample set, train k decision tree models respectively, and fuse the decision tree models by averaging.

[0034] S44: Calculate the error of the fused decision tree model and the weights of each feature;

[0035] S45: Determine whether the error of the fused decision tree model has decreased. If the error has decreased, remove the feature with the lowest weight and jump to step S42; otherwise, end the iteration and output the optimal feature subset.

[0036] Optionally, the MDRSNet tool wear prediction model includes: a feature extraction module, a feature fusion module, and a prediction module. The feature extraction module is used to extract features from the input signal to obtain multi-dimensional features. The feature fusion module is used to fuse the multi-dimensional features to obtain a feature vector. The prediction module is used to perform regression prediction based on the feature vector to obtain a prediction result.

[0037] Optionally, the feature extraction module includes several residual shrinking units and several dimension splicing units. The residual shrinking units are used to extract multi-dimensional features from the input signal; the dimension splicing units are used to splice the multi-dimensional features.

[0038] The feature fusion module includes several pooling layers, which are used to perform information fusion on the multi-dimensional features using a global average pooling algorithm.

[0039] The prediction module includes a fully connected layer, which is used to perform regression prediction on the feature vector.

[0040] Optionally, the working process of the MDRSNet tool wear prediction model includes:

[0041] The input signal is subjected to dimensional expansion by the first residual shrinkage unit to obtain the first feature vector;

[0042] The first feature vector is passed through the second residual shrinking unit to adjust its width, thereby obtaining the second feature vector. At the same time, the input signal is passed through the third and fourth residual shrinking units to perform information mining, thereby obtaining the third and fourth feature vectors.

[0043] The second and third feature vectors are concatenated using the first dimension concatenation unit to obtain the fifth feature vector; the third and fourth feature vectors are concatenated using the second dimension concatenation unit to obtain the sixth feature vector.

[0044] The fifth feature vector is compressed in dimension by the fifth residual shrinking unit to obtain the seventh feature vector;

[0045] The seventh feature vector is processed through the sixth residual shrinking unit to adjust its width, thereby obtaining the eighth feature vector. At the same time, the sixth feature vector is processed through the seventh residual shrinking unit to extract features, thereby obtaining the ninth feature vector.

[0046] The eighth and ninth feature vectors are concatenated using the third dimension concatenation unit to obtain the tenth feature vector.

[0047] The tenth feature vector is processed by the eighth residual shrinking unit to obtain the eleventh feature vector.

[0048] The fourth, ninth, and eleventh feature vectors are processed through the first, second, and third pooling layers, respectively, to aggregate information and obtain the twelfth, thirteenth, and fourteenth feature vectors.

[0049] The twelfth, thirteenth, and fourteenth feature vectors are concatenated using the fourth dimension concatenation unit to obtain the fifteenth feature vector.

[0050] The fifteenth feature vector is passed through the fully connected layer for regression prediction to obtain the prediction result.

[0051] The beneficial effects of this invention are as follows:

[0052] This invention employs a Bayesian optimization algorithm to adaptively select the number of modes K and the penalty function α for VMD decomposition, effectively solving the problem of difficult hyperparameter selection in traditional VMD algorithms. Simultaneously, by selecting the three mode functions with the highest Pearson correlation coefficients between the VMD decomposed signal and the original signal for signal reconstruction, reliable signal denoising can be achieved.

[0053] This invention employs a random forest-based RFE feature dimensionality reduction algorithm to adaptively select the optimal feature subset from high-dimensional features, effectively eliminating irrelevant signal features, thereby simplifying the model and improving the training speed.

[0054] This invention combines RSBU with a multi-branch structure, ensuring excellent prediction accuracy while reducing computational load, increasing training speed, and minimizing memory usage, thus achieving a lightweight model. Furthermore, the use of dimensional concatenation for feature fusion aggregates feature information from different levels, effectively improving the model's predictive performance. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart illustrating the method for constructing the MDRSNet tool wear prediction model according to an embodiment of the present invention.

[0057] Figure 2 This is an example of a milling force signal acquired according to an embodiment of the present invention;

[0058] Figure 3 This is a convergence curve of the Bayesian optimization of VMD hyperparameters according to an embodiment of the present invention;

[0059] Figure 4 This is an example of the signal within one cutting cycle after denoising, according to an embodiment of the present invention;

[0060] Figure 5 This is a flowchart of the RFE algorithm for random forest in an embodiment of the present invention;

[0061] Figure 6 This is a schematic diagram of the MDRSNet tool wear prediction model structure according to an embodiment of the present invention;

[0062] Figure 7 This is a convergence curve of the training loss of the MDRSNet tool wear prediction model in an embodiment of the present invention.

[0063] Figure 8 This is a comparison chart of the prediction results and actual tool wear values ​​of the MDRSNet tool wear prediction model in this embodiment of the invention.

[0064] Figure 9 This is a schematic diagram of the structure of the Deep Residual Shrinkage Network (DRSN) according to an embodiment of the present invention.

[0065] Figure 10 This is a schematic diagram of the structure of the multi-scale network MSNet according to an embodiment of the present invention;

[0066] Figure 11 This is a graph showing the prediction results of the Deep Residual Shrinking Network (DRSN) according to an embodiment of the present invention.

[0067] Figure 12 This is a graph showing the prediction results of the multi-scale network MSNet according to an embodiment of the present invention. Detailed Implementation

[0068] 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.

[0069] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0070] This embodiment provides a cutting tool wear prediction method based on MDRSNet, including:

[0071] Acquire cutting force and torque signals from the cutting tool during the milling process;

[0072] The cutting force and torque signals are input into the preset MDRSNet tool wear prediction model, which outputs the predicted wear amount of the cutting tool. The MDRSNet tool wear prediction model is trained based on a tool wear dataset, which includes the cutting force and torque signals of different cutting tools in the X, Y, and Z axes during milling, as well as the wear amount of the tool flank after each pass. The MDRSNet tool wear prediction model is constructed by introducing a multi-branch structure into the Deep Residual Shrinking Network (DRSN). The DRSN includes several layers, and different layers contain different Residual Shrinking Units (RSBUs). The RSBUs are used to extract the signal features of each layer, and the multi-branch structure is used to fuse the signal features of different layers. The multi-branch structure is divided according to the layers.

[0073] like Figure 1 As shown, constructing the MDRSNet tool wear prediction model specifically includes the following steps:

[0074] S1: Obtain the tool wear dataset;

[0075] Tool wear datasets were obtained through experiments conducted on CNC milling machines, specifically:

[0076] A force gauge is installed on the machine tool, and a milling cutter is used to perform milling operations on the workpiece. The cutting force and torque signals of the X, Y, and Z axes are obtained during the milling process. After each pass, the wear of the tool flank face is measured using a microscope to obtain a tool wear dataset. The maximum value of the wear on the flank face of the milling cutter is used to represent the magnitude of tool wear.

[0077] In this embodiment, the experiment was conducted on a CNC milling machine (KVC 1050N, CCZMT, China). The material being machined was γ-TiAl alloy, with dimensions of 80mm × 80mm × 50mm. A 3-flute carbide end mill was used, with a cutting diameter of 6mm, a cutting length of 18mm, an overall length of 50mm, and a helix angle of 45°. During machining, a force sensor (9119AA2, Kistler, Switzerland) was used to measure the cutting force and torque signals in three axes. These signals were amplified by a multi-channel laboratory charge amplifier (5080A, Kistler, Switzerland) and acquired using a data acquisition card (5697A, Kistler, Switzerland). The data was displayed in real time using DynoWare software, and the sensor sampling frequency was set to 10kHz. After each pass, a digital measuring microscope (ISM-PM200SA, INISE, China) was used to measure the wear value of the end mill's flank face.

[0078] The experiment used climb milling, and a total of 5 sets of tool wear data were collected. For each set, the radial depth of cut was set to 5 mm, the depth of cut to the back of cut to 0.2 mm, and the feed per tooth to 0.2 mm·z⁻¹. The spindle speed settings for each tool are shown in Table 1. Taking the X-axis milling force signal data collected during the first milling operation of tool I as an example, the original data obtained are as follows: Figure 2 As shown.

[0079] Table 1

[0080]

[0081] S2: Denoise the tool wear dataset;

[0082] The Bayesian-optimized variational mode decomposition (VMD) denoising algorithm is used to denoise signals, specifically including:

[0083] S21: Initialize the Bayesian optimization algorithm, and set the maximum number of iterations, the number of modes K, and the range of values ​​for the penalty function α;

[0084] S22: Initialize a set of modal numbers and penalty function values ​​as test points, and perform VMD decomposition on the signal to obtain K intrinsic modal functions (IMFs);

[0085] S23: Calculate the envelope entropy E of the modal function according to equations (1)-(3), and fit a Gaussian regression model to estimate the posterior probability distribution of the minimum envelope entropy;

[0086]

[0087]

[0088]

[0089] Where a(j) is the envelope signal sequence of signal x(j) (j=1,2,...,n) after Hibert transform; H(·) is the Hibert transform; p j Let be the normalized form of a(j); E is the envelope entropy.

[0090] S24: Maximize the expected improvement sampling function and select the test points for the next iteration;

[0091] S25: Calculate the minimum envelope entropy of the test points, add it to the observation dataset, and update the Gaussian regression model;

[0092] S26: Determine if the maximum number of iterations has been reached. If so, stop iterating and output the optimal number of modes K and the penalty function α; otherwise, jump to step S24.

[0093] S27: Perform VMD decomposition based on the optimal number of modes K and the penalty function α, and calculate the Pearson correlation coefficient of each mode function according to equation (4);

[0094]

[0095] Where X = {x1, x2, ..., x} n} and Y = {y1, y2, ..., y n Let} be two time series of length n; and Let X and Y represent the mean values ​​of signals X and Y, respectively.

[0096] S28: Select the three mode functions with the highest Pearson coefficients to reconstruct the signal.

[0097] In this embodiment, the maximum number of iterations for the Bayesian optimization algorithm is set to 100 rounds, the range of the mode function K is set to [1, 10], the range of the penalty function α is set to [1000, 3000], and the objective function is set to the minimum envelope entropy of the mode function. The convergence curve of the Bayesian optimization algorithm is shown below. Figure 3 As shown, the optimal mode function K is obtained. * and penalty function α * The combination is [7,2957]. The optimal hyperparameter combination is selected for VMD decomposition of the signal, yielding 7 mode functions. The Pearson correlation coefficients between each mode function and the original signal are shown in Table 2. The three mode functions with the highest Pearson coefficients (IMF1, IMF2, and IMF4) are selected to reconstruct the signal. The reconstructed signal data for one cutting cycle is shown in Table 2. Figure 4 As shown.

[0098] Table 2

[0099]

[0100] S3: Extract features from the denoised tool wear dataset;

[0101] The time-domain, frequency-domain, and time-frequency-domain features of the denoised signal data are extracted, including maximum value, minimum value, peak-to-peak value, peak value, mean, root square amplitude, variance, root mean square value, kurtosis, skewness, waveform factor, peak factor, margin factor, energy, frequency mean, frequency variance, centroid frequency, frequency root mean square, frequency standard deviation, standard deviation coefficient, kurtosis index, skewness index, kurtosis factor, skewness factor, frequency 1 / 2 central moment factor, energy features of eight sub-bands (aaa, daa, ada, dda, aad, dad, add, and ddd) after three-level wavelet packet decomposition of the “db3” wavelet function, and energy features of four mode functions (IMF1, IMF2, IMF3, and IMF4) after three-level empirical mode decomposition.

[0102] In this embodiment, feature extraction yields 14 time-domain features, 11 frequency-domain features, 8 wavelet packet energy features, and 4 IMF energy features across six dimensions, for a total of 222 feature values.

[0103] S4: Filter the extracted features;

[0104] The recursive feature elimination (RFE) algorithm based on random forests is used to select the optimal feature subset. The specific process is as follows: Figure 5 As shown, it includes the following steps:

[0105] S41: Set the maximum depth and number of classifiers k for the random forest model;

[0106] S42: Perform sampling with replacement on the original data to generate k training sample sets, and update the training sample sets by extracting features without replacement;

[0107] S43: Based on k sample sets, train k decision tree models respectively, and then fuse the decision tree models using the averaging method;

[0108] S44: Calculate the model error and the weights of each feature;

[0109] S45: Determine if the model error has decreased. If the model error has decreased, remove the feature with the lowest weight and jump to step S42; otherwise, end the iteration and output the optimal feature subset.

[0110] In this embodiment, the maximum depth of the random forest is set to 20, and the number of classifiers is set to 50. After iteratively building the random forest model and removing the features with the lowest weights, the optimal feature subset is shown in Table 3.

[0111] Table 3

[0112]

[0113]

[0114] S5: Divide the dataset into training and testing sets;

[0115] The tool wear dataset is divided into a tool wear training set and a tool wear test set, which are used for training and testing the MDRSNet tool wear prediction model, respectively. Specifically, a set of tool lifecycle wear data from the tool wear dataset is selected for model training, and the remaining tool wear data is used for model testing.

[0116] In this embodiment, the wear dataset of tool II is used as the training set, and the wear datasets of the other four tools are used as the test set.

[0117] S6: Establish an MDRSNet tool wear prediction model;

[0118] like Figure 6As shown, the MDRSNet tool wear prediction model is constructed by introducing a multi-branch structure into the deep residual shrinkage networks (DRSN). The DRSN consists of several layers, and different layers contain different residual shrinkage building units (RSBUs). The RSBUs are used to extract the signal features of each layer, and the multi-branch structure is used to fuse the signal features of different layers. The multi-branch structure is divided according to the layers.

[0119] The MDRSNet tool wear prediction model includes a feature extraction module, a feature fusion module, and a prediction module. The feature extraction module is used to extract features from the input signal and obtain multi-dimensional features. The feature fusion module is used to fuse information from the multi-dimensional features and obtain feature vectors. The prediction module is used to perform regression prediction based on the feature vectors and obtain prediction results.

[0120] The feature extraction module includes several residual shrinking units and several dimension splicing units. The residual shrinking units are used to extract multi-dimensional features from the input signal; the dimension splicing units are used to splice multi-dimensional features.

[0121] The feature fusion module includes several pooling layers, which are used to fuse information from multi-dimensional features using a global average pooling algorithm.

[0122] The prediction module includes a fully connected layer, which is used to perform regression prediction on the feature vector.

[0123] In this embodiment, the Deep Residual Shrinkage Network (DRSNet) has 5 feature extraction layers. Different levels of feature extraction are achieved through multiple stacked RSBUs, and a 3-branch structure is used to fuse information from different levels. Let the number of channels in the original data be C, the width be W, and the length be 1. The workflow of the MDRSNet tool wear prediction model is as follows:

[0124] S61: In the first layer, the signal feature x of size C×W×1 is expanded to a dimension of 3 after RSBU calculation, resulting in a feature x of size 3C×W×1. 1,1 .

[0125] S62: In layer II, set the stride of the first convolutional layer of RSBU to 2, and set x... 1,1 The width is adjusted to W / 2 to obtain a feature x with dimensions of 3C×W / 2×1. 1,2 Simultaneously, using two distinct RSBUs to mine information from the original signal x, two features x, each with a size of 3C×W / 2×1, were obtained. 2,1 and x 3,1 .

[0126] S63: x 1,2 x 2,1 and x 2,1 x 3,1 By concatenating the dimensions separately, a fused feature x with a size of 6C×W / 2×1 is obtained. 1,3 and x 2,2 .

[0127] S64: In layer III, feature x is fused. 1,3 After RSBU, dimensionality compression is achieved, resulting in a feature vector x of size 3C×W / 2×1. 1,4 ;

[0128] S65: In layer IV, set the stride of the first convolutional layer of RSBU to 2, and transfer the feature vector x 1,4 The width is adjusted to W / 4 to obtain a feature x with dimensions of 3C×W / 4×1. 1,5 At the same time, the fused feature x 2,2 After RSBU feature extraction, a feature vector x of size 3C×W / 4×1 is obtained. 2,3 .

[0129] S66: Transfer the feature vector x 1,5 and x 2,3 Dimensional concatenation is performed to obtain a fused feature x with dimensions of 6C×W / 4×1. 1,6 .

[0130] S67: In layer V, feature x is fused. 1,6 The information is further refined using RSBU to obtain a feature vector x of size 3C×W / 8×1. 1,7 .

[0131] S68: Use global average pooling algorithm to aggregate feature vector x 3,1 x 2,3 and x 1,7 From the information, we obtain feature vectors x, each with a size of 3C×1×1. 3,2 x 2,4 and x 1,8 .

[0132] S69: Fusing x through dimensional splicing 3,2 x 2,4 and x 1,8 The information is used to obtain the feature vector x4, which is then input into the fully connected layer for regression prediction.

[0133] Branch 1: Signal characteristics x→x 1,1 →x 1,2 →x 1,3 →x 1,4 →x 1,5 →x1,6 →x 1,7 →x 1,8 ;

[0134] Branch 2: Signal characteristics x→x 2,1 →x 2,2 →x 2,3 →x 2,4 ;

[0135] Branch 3: Signal characteristics x→x 3,1 →x 3,2 .

[0136] In this embodiment, the parameters of the MDRSNet tool wear prediction model are shown in Table 4, and the loss change curve during the training process is shown in Table 4. Figure 7 As shown.

[0137] Table 4

[0138]

[0139]

[0140] As can be seen, the fitting loss of the MDRSNet tool wear prediction model decreases rapidly and tends to stabilize, indicating that the model has excellent fitting performance.

[0141] S7: Validate the usability of the MDRSNet tool wear prediction model.

[0142] The MDRSNet tool wear prediction model trained with S6 has only 6844 parameters. Its prediction results on the test set are as follows: Figure 8 As shown, the prediction accuracy of the MDRSNet tool wear prediction model was evaluated using the mean absolute error (MAE). In the four sets of milling cutter wear datasets, the MAEs of the MDRSNet tool wear prediction model were 7.9 μm, 13.0 μm, 10.5 μm, and 13.9 μm, respectively, and the mean absolute error of the MDRSNet tool wear prediction model was 9.4 μm. The results show that the MDRSNet tool wear prediction model in this embodiment has excellent performance and can accurately predict tool wear conditions by training a small number of parameters.

[0143] To verify the effectiveness and superiority of the tool wear prediction method and MDRSNet tool wear prediction model proposed in this embodiment, deep residual shrinkage networks (DRSN) and multiscale networks (MSNet) were selected for comparative analysis. The comparison model structures are as follows: Figure 9 and Figure 10As shown in Table 4, the training parameters are as follows. DRSN consists of four stacked RSBUs, with the number of channels in each RSBU set to 3, 3, 6, and 12, and the stacking number set to 4, 4, 4, and 4. MSNet uses residual units for feature extraction, with the number of channels in each layer set to 16, 16, 64, and 64, and the stacking number of residual units set to 4, 4, 4, and 4. The comparison model uses the wear dataset of tool II for training, and the wear datasets of the other four tools for testing. The prediction results of the comparison model are shown in Table 4. Figure 11 , Figure 12 As shown in Table 5.

[0144] Table 5

[0145]

[0146] As can be seen, compared to MSNet, MDRSNet, which incorporates the RSBU structure, exhibits superior prediction accuracy, reducing prediction error by 21.3%. Comparing the performance of DRSN and MDRSNet, it is evident that introducing a multi-branch structure into DRSN improves the model's prediction accuracy by 8.36% while reducing the number of parameters by 74.3%, thus minimizing model storage space and achieving network lightweighting. In conclusion, the MDRSNet tool wear prediction model proposed in this invention can achieve accurate tool wear value prediction with fewer trained hyperparameters, demonstrating both good prediction accuracy and lightweight design.

[0147] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A cutting tool wear prediction method based on MDRSNet, characterized in that, include: Acquire cutting force and torque signals from the cutting tool during the milling process; The cutting force and torque signals are input into a pre-defined MDRSNet tool wear prediction model, which outputs the predicted wear amount of the cutting tool. The MDRSNet tool wear prediction model is trained on a tool wear dataset, which includes cutting force and torque signals along the X, Y, and Z axes for different cutting tools during milling, as well as the wear amount on the tool flank after each pass. The MDRSNet tool wear prediction model is constructed by introducing a multi-branch structure into the Deep Residual Shrinking Network (DRSN). The DRSN comprises several layers, each containing different Residual Shrinking Units (RSBUs). The RSBUs are used to extract signal features from each layer. The multi-branch structure is used to fuse the signal features from different layers. Based on the layer division, the multi-branch structure preprocesses the tool wear dataset. A variational mode decomposition denoising algorithm based on Bayesian optimization is used to denoise the cutting force and torque signals in the tool wear dataset, and feature extraction is performed on the denoised cutting force and torque signals to obtain signal features. The recursive feature elimination algorithm based on random forest is used to filter the signal features, obtain the optimal feature subset, and complete the preprocessing.

2. The cutting tool wear prediction method based on MDRSNet according to claim 1, characterized in that, The Bayesian-optimized variational mode decomposition denoising algorithm denoises the cutting force and torque signals in the tool wear dataset, including: S21: Initialize the Bayesian optimization algorithm, setting the maximum number of iterations, the number of modes, and the range of values ​​for the penalty function; S22: Initialize a set of modal numbers and penalty function values ​​as test points, and perform VMD decomposition on the cutting force signal and torque signal to obtain the modal functions; S23: Calculate the envelope entropy of the modal function, fit a Gaussian regression model, and estimate the posterior probability distribution of the minimum envelope entropy; S24: Maximize the acquisition function to determine the next test point; S25: Calculate the minimum envelope entropy of the next test point, add it to the observation data, and update the Gaussian regression model; S26: Determine whether the maximum number of iterations has been reached. If it has, stop the iteration and output the optimal number of modes and the optimal penalty function; otherwise, jump to S24. S27: Perform VMD decomposition based on the optimal number of modes and the optimal penalty function, and calculate the Pearson correlation coefficient of each mode function; S28: Select the three mode functions with the highest Pearson coefficients to reconstruct the signal and complete the denoising process.

3. The cutting tool wear prediction method based on MDRSNet according to claim 2, characterized in that, The method for calculating the envelope entropy is as follows: (1) (2) (3) in, For signal The envelope signal sequence after Hibert transform; This is the Hibert transform; For signal The normalized form; E Let be the envelope entropy.

4. The cutting tool wear prediction method based on MDRSNet according to claim 3, characterized in that, The Pearson correlation coefficient is calculated as follows: (4) in, and For two lengths Time signal sequence; x i Time signal sequence The i-th time signal in y i Time signal sequence The i-th time signal in; and Representing time signal sequences respectively and The mean.

5. The cutting tool wear prediction method based on MDRSNet according to claim 1, characterized in that, The signal features include: time-domain features, frequency-domain features, and time-frequency-domain features. The time-domain features include maximum value, minimum value, peak-to-peak value, peak value, mean value, root square amplitude, variance, root mean square value, kurtosis, skewness, waveform factor, peak factor, margin factor, and energy. The frequency-domain features include frequency mean, frequency variance, centroid frequency, root mean square frequency, frequency standard deviation, standard deviation coefficient, kurtosis index, skewness index, kurtosis factor, skewness factor, and frequency 1 / 2 central moment factor. The time-frequency-domain features include the energy features of eight sub-bands after three-level wavelet packet decomposition and the energy features of four mode functions after three-level empirical mode decomposition.

6. The cutting tool wear prediction method based on MDRSNet according to claim 1, characterized in that, The recursive feature elimination algorithm based on random forest is used to filter the signal features, including: S41: Set the maximum depth and number of classifiers for the random forest model; S42: Perform replacement sampling on the signal feature data to generate k training sample sets, and update the training sample sets by feature extraction without replacement; S43: Based on the updated training sample set, train k decision tree models respectively, and fuse the decision tree models by averaging. S44: Calculate the error of the fused decision tree model and the weights of each feature; S45: Determine whether the error of the fused decision tree model has decreased. If the error has decreased, remove the feature with the lowest weight and jump to step S42; otherwise, end the iteration and output the optimal feature subset.

7. The cutting tool wear prediction method based on MDRSNet according to claim 1, characterized in that, The MDRSNet tool wear prediction model includes a feature extraction module, a feature fusion module, and a prediction module. The feature extraction module is used to extract features from the input signal to obtain multi-dimensional features. The feature fusion module is used to fuse the multi-dimensional features to obtain a feature vector. The prediction module is used to perform regression prediction based on the feature vector to obtain the prediction result.

8. The cutting tool wear prediction method based on MDRSNet according to claim 7, characterized in that, The feature extraction module includes several residual shrinking units and several dimension splicing units. The residual shrinking units are used to extract multi-dimensional features from the input signal; the dimension splicing units are used to splice the multi-dimensional features. The feature fusion module includes several pooling layers, which are used to perform information fusion on the multi-dimensional features using a global average pooling algorithm. The prediction module includes a fully connected layer, which is used to perform regression prediction on the feature vector.

9. The cutting tool wear prediction method based on MDRSNet according to claim 8, characterized in that, The working process of the MDRSNet tool wear prediction model includes: The input signal is subjected to dimensional expansion by the first residual shrinkage unit to obtain the first feature vector; The first feature vector is passed through the second residual shrinking unit to adjust its width, thereby obtaining the second feature vector. At the same time, the input signal is passed through the third and fourth residual shrinking units to perform information mining, thereby obtaining the third and fourth feature vectors. The second and third feature vectors are concatenated using the first dimension concatenation unit to obtain the fifth feature vector; the third and fourth feature vectors are concatenated using the second dimension concatenation unit to obtain the sixth feature vector. The fifth feature vector is compressed in dimension by the fifth residual shrinking unit to obtain the seventh feature vector; The seventh feature vector is processed through the sixth residual shrinking unit to adjust its width, thereby obtaining the eighth feature vector. At the same time, the sixth feature vector is processed through the seventh residual shrinking unit to extract features, thereby obtaining the ninth feature vector. The eighth and ninth feature vectors are concatenated using the third dimension concatenation unit to obtain the tenth feature vector. The tenth feature vector is processed by the eighth residual shrinking unit to obtain the eleventh feature vector. The fourth, ninth, and eleventh feature vectors are processed through the first, second, and third pooling layers, respectively, to aggregate information and obtain the twelfth, thirteenth, and fourteenth feature vectors. The twelfth, thirteenth, and fourteenth feature vectors are concatenated using the fourth dimension concatenation unit to obtain the fifteenth feature vector. The fifteenth feature vector is passed through the fully connected layer for regression prediction to obtain the prediction result.

Images