CN-118081482-B - MDRSNet-based cutting tool wear prediction method

CN118081482BCN 118081482 BCN118081482 BCN 118081482BCN-118081482-B

Abstract

The invention relates to a cutting tool abrasion prediction method based on MDRSNet, which comprises the steps of collecting cutting force signals and moment signals of a cutting tool in a milling process, inputting the cutting force signals and the moment signals into a preset MDRSNet tool abrasion prediction model, and outputting predicted abrasion quantity of the cutting tool, wherein a MDRSNet tool abrasion prediction model is obtained based on training of a tool abrasion dataset, the tool abrasion dataset comprises the cutting force signals and the moment signals of X, Y, Z triaxial of different cutting tools in the milling process, and abrasion quantity of a rear cutter surface of the cutting tool after each feeding, and the MDRSNet tool abrasion prediction model is constructed by introducing a multi-branch structure into a depth residual shrinkage network DRSN. The invention enhances the expression capability and learning capability of the model by combining the residual error contraction unit with the multi-branch structure, and realizes the light weight of the model while ensuring that the model has excellent prediction precision.

Inventors

YIN ZENGBIN
CHEN SHAOYANG
YUAN JUNTANG

Assignees

南京理工大学

Dates

Publication Date: 20260508
Application Date: 20240311

Claims (9)

1. A method of predicting wear of a cutting tool based on MDRSNet, comprising: collecting cutting force signals and moment signals of a cutting tool in the milling process; Inputting the cutting force signal and the moment signal into a preset MDRSNet tool wear prediction model, outputting the predicted wear amount of the cutting tool, wherein the MDRSNet tool wear prediction model is obtained based on tool wear data set training, the tool wear data set comprises X, Y, Z triaxial cutting force signals and moment signals of different cutting tools in the milling process and the wear amount of the rear surface of the cutting tool after each feeding, the MDRSNet tool wear prediction model is constructed by introducing a multi-branch structure into a depth residual shrinkage network DRSN, the depth residual shrinkage network DRSN comprises a plurality of layers, different layers comprise different residual shrinkage units RSBU, the residual shrinkage units RSBU are used for extracting signal characteristics of each layer, the multi-branch structure is used for fusing the signal characteristics of different layers, and the multi-branch structure is used for preprocessing the tool wear data set according to layer division: Denoising the cutting force signals and the moment signals in the cutter wear data set based on a Bayesian optimization variation modal decomposition noise reduction algorithm, and extracting features of the denoised cutting force signals and moment signals to obtain signal features; And screening the signal features based on a recursive feature elimination algorithm of the random forest to obtain an optimal feature subset, and finishing the preprocessing.
2. The MDRSNet-based cutting tool wear prediction method of claim 1, wherein denoising the cutting force signal and the moment signal in the tool wear dataset based on a bayesian optimized variational modal decomposition noise reduction algorithm comprises: S21, initializing a Bayesian optimization algorithm, and setting the maximum iteration times, the mode number and the value range of a penalty function; S22, initializing a group of modal numbers and penalty function values as test points, and performing VMD decomposition on the cutting force signals and the moment signals to obtain modal functions; S23, calculating the envelope entropy of the modal function, fitting a Gaussian regression model, and estimating posterior probability distribution of the minimum envelope entropy; S24, maximizing an acquisition function and determining a next test point; s25, calculating the minimum envelope entropy of the next test point, adding the minimum envelope entropy into observed data, and updating the Gaussian regression model; s26, judging whether the maximum iteration times are reached, if so, stopping iteration, and outputting the optimal mode number and the optimal penalty function, otherwise, jumping to S24; S27, performing VMD decomposition based on the optimal modal number and the optimal penalty function, and calculating Pearson correlation coefficients of the modal functions; and S28, selecting three mode function reconstruction signals with the highest Pearson coefficients, and completing the denoising processing.
3. The MDRSNet-based cutting tool wear prediction method according to claim 2, wherein the envelope entropy calculation method is: (1) (2) (3) Wherein, the Is a signal An envelope signal sequence after Hibert transformation; Is a Hibert transform; Is a signal And E is the envelope entropy.
4. The MDRSNet-based cutting tool wear prediction method according to claim 3, wherein the Pearson correlation coefficient calculation method is as follows: (4) Wherein, the And Is of two lengths X i is the time signal sequence In (3), y i is a time signal sequence The i-th time signal of (a); And Respectively represent time signal sequences And Is a mean value of (c).
5. The MDRSNet-based cutting tool wear prediction method of claim 1, wherein said signal characteristics include time domain characteristics including maximum, minimum, peak-to-peak, peak-to-average, root-mean-square, variance, root-mean-square, kurtosis, skewness, waveform factors, peak factors, margin factors, energy, frequency domain characteristics including frequency mean, frequency variance, center of gravity frequency, root mean-square, frequency standard deviation, standard deviation coefficients, kurtosis index, skew index, kurtosis factor, skew factor, frequency 1/2 center moment factor, and time-to-frequency domain characteristics including eight sub-band energy characteristics after three-layer wavelet packet decomposition and four mode function energy characteristics after three-layer empirical mode decomposition.
6. The MDRSNet-based cutting tool wear prediction method of claim 1, wherein screening the signal features based on a random forest recursive feature elimination algorithm comprises: S41, setting the maximum depth of a random forest model and the number of classifiers; S42, carrying out put-back sampling on the signal characteristic data to generate k training sample sets, and updating the training sample sets through no-put-back characteristic extraction; s43, respectively training k decision tree models based on the updated training sample set, and fusing the decision tree models by an averaging method; s44, calculating the fused decision tree model error and each characteristic weight; s45, judging whether the error of the fused decision tree model is reduced, if so, eliminating the feature with the lowest weight, and jumping to the step S42, otherwise, ending iteration and outputting the optimal feature subset.
7. The MDRSNet-based cutting tool wear prediction method according to claim 1 is characterized in that the MDRSNet tool wear prediction model comprises a feature extraction module, a feature fusion module and a prediction module, wherein the feature extraction module is used for extracting features of an input signal to obtain multi-dimensional features, the feature fusion module is used for carrying out information fusion on the multi-dimensional features to obtain feature vectors, and the prediction module is used for carrying out regression prediction based on the feature vectors to obtain prediction results.
8. The MDRSNet-based cutting tool wear prediction method as set forth in claim 7, wherein the feature extraction module includes a plurality of residual shrinkage units and a plurality of dimension stitching units, the residual shrinkage units being configured to perform multi-dimensional feature extraction on the input signal; The feature fusion module comprises a plurality of pooling layers, wherein the pooling layers are used for carrying out information fusion on the multidimensional features by adopting a global average pooling algorithm; the prediction module comprises a full-connection layer, and the full-connection layer is used for carrying out regression prediction on the feature vector.
9. The MDRSNet-based cutting tool wear prediction method as set forth in claim 8, wherein the working process of the MDRSNet tool wear prediction model includes: performing dimension expansion on the input signal through a first residual error contraction unit to obtain a first feature vector; The width of the first feature vector is adjusted through a second residual error contraction unit to obtain a second feature vector, and meanwhile, the input signal is subjected to information mining through a third residual error contraction unit and a fourth residual error contraction unit to obtain a third feature vector and a fourth feature vector; Performing dimension splicing on the second feature vector and the third feature vector through a first dimension splicing unit to obtain a fifth feature vector; performing dimension splicing on the third feature vector and the fourth feature vector through a second dimension splicing unit to obtain a sixth feature vector; Performing dimension compression on the fifth feature vector through a fifth residual error contraction unit to obtain a seventh feature vector; Performing width adjustment on the seventh feature vector through a sixth residual error contraction unit to obtain an eighth feature vector, and performing feature extraction on the sixth feature vector through the seventh residual error contraction unit to obtain a ninth feature vector; performing dimension splicing on the eighth feature vector and the ninth feature vector through a third dimension splicing unit to obtain a tenth feature vector; Performing information condensation on the tenth characteristic vector through an eighth residual error shrinkage unit to obtain an eleventh characteristic vector; The fourth feature vector, the ninth feature vector and the eleventh feature vector are respectively subjected to information aggregation through a first pooling layer, a second pooling layer and a third pooling layer to obtain a twelfth feature vector, a thirteenth feature vector and a fourteenth feature vector; Performing dimension splicing on the twelfth feature vector, the thirteenth feature vector and the fourteenth feature vector through a fourth dimension splicing unit to obtain a fifteenth feature vector; And carrying out regression prediction on the fifteenth feature vector through the full connection layer to obtain a prediction result.

Description

MDRSNet-based cutting tool wear prediction method Technical Field The invention relates to the technical field of machining state monitoring, in particular to a cutting tool abrasion prediction method based on MDRSNet. Background Machine tools are key devices in the production and machining processes, and cutters are an important component part of the machine tools. In the production process of parts, once the abrasion of a cutter reaches a threshold value and the cutter is not replaced, the processing quality of the surface of a workpiece is affected, and even the product is scrapped and the machine tool is damaged. At present, domestic enterprises mainly adopt preventive maintenance strategies to maintain the cutters of the numerical control machine tool, namely, the cutters are replaced in advance when the cutters do not reach the abrasion threshold, so that nearly 62% of the service life of the cutters is not utilized, the cutters are wasted, the production cost of the enterprises is increased, and the production efficiency is reduced. Related researches indicate that if the cutter wear monitoring system is adopted to monitor the wear state of the cutter in real time, the downtime of 75% can be reduced, and the production efficiency of 10% -50% can be improved. Therefore, the real-time state monitoring of the cutter and the predictive maintenance are implemented, and the method has important significance for promoting the development of intelligent cutting technology in China. In recent years, with the continuous development of sensor technology, computer processing technology and artificial intelligence technology, cutting force, vibration, noise, temperature, feed texture, abrasion images and other data in the cutter processing process are acquired and stored in real time. These data can fully reflect the full life cycle process of tool wear, providing the necessary data support for achieving tool wear monitoring. The current tool wear prediction modes can be divided into classical machine learning models and deep learning models. The tool wear prediction model based on deep learning has more excellent prediction performance than the machine learning model. However, the complex structure of the deep learning model contains a large number of calculation parameters, so that the training speed of the model is slow, the reasoning time is long, and the performance requirement on a computer is high. In addition, the huge parameter amount needs to occupy a large amount of storage space, and huge storage pressure is brought to actual production. Therefore, when the tool wear deep learning prediction model is established, the accuracy of model prediction can be guaranteed, the training time can be shortened, the calculation cost can be reduced, and the memory occupation can be reduced under the condition of limited calculation resources, so that the tool wear deep learning prediction model is a difficulty in the development of the current tool wear prediction technology. Disclosure of Invention The invention aims to solve the problems that a cutter abrasion prediction model based on deep learning at the present stage is complex in structure, huge in parameters, large in occupied storage space and the like, and provides a cutting cutter abrasion prediction method based on MDRSNet, milling force signal characteristics are extracted through a residual error contraction unit, the expression capacity of the model is enhanced through a multi-branch structure, meanwhile, the learning capacity of the model is improved through fusion of output characteristics of different branches, and the model is lightened while excellent prediction precision is ensured. In order to achieve the above object, the present invention provides the following solutions: a MDRSNet-based cutting tool wear prediction method, comprising: collecting cutting force signals and moment signals of a cutting tool in the milling process; Inputting the cutting force signal and the moment signal into a preset MDRSNet tool wear prediction model, outputting the predicted wear amount of the cutting tool, wherein the MDRSNet tool wear prediction model is obtained based on tool wear data set training, the tool wear data set comprises the cutting force signal and the moment signal of X, Y, Z triaxial of different cutting tools in the milling process and the wear amount of the rear surface of the cutting tool after each feeding, the MDRSNet tool wear prediction model is constructed by introducing a multi-branch structure into a depth residual shrinkage network DRSN, the depth residual shrinkage network DRSN comprises a plurality of layers, different layers comprise different residual shrinkage units RSBU, the residual shrinkage units RSBU are used for extracting signal characteristics of each layer, the multi-branch structure is used for fusing the signal characteristics of different layers, and the multi-branch structure is divided according to the layers. Opt