Search

CN-122020410-A - Cross-machine tool cutter wear state identification method based on transfer learning

CN122020410ACN 122020410 ACN122020410 ACN 122020410ACN-122020410-A

Abstract

The invention relates to the technical field of tool wear state identification for machining, in particular to a cross-machine tool wear state identification method based on transfer learning. The method comprises the steps of taking the machine tool power consumption growth rate as a cutter health index, adopting a Newton interpolation and random dithering combined time sequence random enhancement method from the data preprocessing angle, training a C4.5 decision tree classification model, carrying out fine adjustment on a pre-classification training model by utilizing a small amount of abrasion label data on a target machine tool to realize model migration, obtaining a decision tree classification model based on migration stacking, realizing the optimal model expression effect, and finally, taking a real-time classification result as input information by utilizing a sliding window and a data density theory, accurately identifying transition points among different abrasion states of a cutter, determining the correction rule of the classification result according to the transition point identification result, effectively reducing the number of error classification samples of an abrasion state transition region, and further improving the classification precision of the cutter abrasion state under the machine tool.

Inventors

  • SHI KAINING
  • Qiang Biyao
  • ZHANG ZIXIN
  • LIU JIACHENG
  • REN JUNXUE

Assignees

  • 西北工业大学

Dates

Publication Date
20260512
Application Date
20260415

Claims (9)

  1. 1. The method for identifying the wear state of the tool across the machine tool based on the transfer learning is characterized by comprising the following steps of: Firstly, constructing a health index representing the degradation degree of a cutter by acquiring a machine tool power signal of a stable cutting stage of a target domain machine tool and using the power consumption increment of the machine tool power signal, and establishing a cutter abrasion state data set through data enhancement and label division processing for balancing the data quantity of a sample of the machine tool power signal; Training a pre-classification training model of C4.5 decision tree classification by using source domain data of a source domain machine tool, performing model migration adjustment on the pre-classification training model by using wear data with labels in a cutter wear state data set of a target domain machine tool to obtain migration classification models from a plurality of source domain machine tools to the target domain machine tool, and integrating the migration classification models by using a migration stacking technology to obtain a decision tree classification model based on migration stacking; the source domain data refers to a cutter abrasion state data set formed by machine tool power signals of a source domain machine tool in a stable cutting stage; and thirdly, windowing a real-time state classification sequence output by the decision tree classification model by adopting a sliding window technology, analyzing the distribution characteristics of classification results in a window by adopting a data density theory, identifying transition points of the cutter in different wear states, and finally correcting the transition points to realize the identification of the wear state of the cutter.
  2. 2. The method for identifying the wear state of a tool across machine tools based on transfer learning according to claim 1, wherein the specific steps of constructing the health index in the first step are as follows: Machine tool machining power consumption in the tool wear state without considering the additional power consumption P loss The method comprises the following steps: , Wherein P idle is idle power consumption, P feed is feeding power consumption, P removal is material removing power consumption, and P increment (VB) is power consumption increment caused by cutter abrasion; Along with the continuous evolution of the tool wear state, the idle power consumption and the feeding power consumption are completely determined by the characteristics of the machine tool, the material removal power consumption is constant when the machining parameters are fixed, and only the power consumption increment is increased by the subsequent increase of the tool face wear amount and is the only component directly related to the tool wear and irrelevant to the machine tool characteristics, so that the power consumption increment is used as an index for measuring the tool health state; Health index constructed by machine tool power consumption increase rate The calculation mode of (a) is as follows: , Wherein P cutting is the machine tool machining power consumption in the ideal tool state.
  3. 3. The method for identifying tool wear state across machine tools based on transfer learning according to claim 1, wherein the specific step of creating the tool wear state data set in the step one is: Taking the power consumption increment and the health index sequence obtained by calculating the machine tool power signal as processing objects, adopting a Newton interpolation method to carry out data enhancement, and assuming that the function of the health index changing along with the sampling point is Given n+1 interpolation nodes { x 0 ,x 1 ,… , x n }, then The newton interpolation formula of (c) is: , in the interpolation process, new points are inserted between adjacent nodes by refining interpolation intervals, new health index data are generated, and the interpolation order is controlled to be not more than a cubic polynomial; After interpolation is completed, noise is added to new health index data points generated by interpolation by adopting a random value dithering method, a random value is added on the basis of an original value, and a calculation formula is as follows: , In the formula, To add the health index value after the random value, The health index value obtained for the original interpolation, Representing a random value, N is Is used for the value range of the (a), the use of the formula is expressed as: , Wherein, the The value range N of the random jitter is used for avoiding the interference of the random jitter on the original distribution; y 1 ,y 2 is the health index value corresponding to the left and right adjacent points of the current interpolation point, L is the interpolation number between the adjacent points, namely the expansion multiple of the interpolation method to the sequence, at the moment, the enhanced health index sequence is obtained, then the enhanced health index sequence is subjected to label division according to the cutter abrasion stage, and a cutter abrasion state data set containing the enhanced health index and the corresponding abrasion state label is constructed.
  4. 4. The method for identifying the wear state of a tool across machine tools based on transfer learning according to claim 1, wherein in the second step, the specific building process of the pre-classification training model is as follows: Definition of information entropy The expression of (2) is: , Wherein D is a wear state set of a tool in which the current source domain machine tool is located, k is a k-th type source domain data sample, k=1, 2, β, and p k are ratios occupied by the source domain data samples; Assuming that the discrete attribute a has V possible values, dividing the wear state set D by using the a, generating V branch nodes, wherein the V branch nodes comprise all tool wear samples with the value of a v on the discrete attribute a in the wear state set D, marking the tool wear samples as D v , and calculating the information entropy of D v ; The weight |D v |/|D| is given to the V branch nodes, the abrasion state set D is divided by the attribute a, and the obtained information Gain (D, a) is calculated according to the calculation model: , Then, the calculation model of gain_ratio (D, a) is: , wherein IV (a) is the inherent value of the discrete attribute a, and the calculation formula is as follows: , the Gain ratio gain_ratio (D, a) is used for reducing attribute preference of information Gain to a large number of available values, so that optimal dividing attributes are selected; at this time, m pre-classification training models generated by the m source domain data of the C4.5 decision tree algorithm are obtained and recorded as , wherein, I=1 for the i-th source domain dataset 2 m。
  5. 5. The method for identifying the wear state of a tool across machine tools based on transfer learning according to claim 1, wherein the specific steps of the decision tree classification model based on transfer stacking obtained in the step two are as follows: firstly, taking labeled wear data in a target domain machine tool wear state data set as target domain samples, and giving the same weight to each target domain sample Training to obtain a first weak classification model and weighting The calculation formula of (2) is as follows: , wherein N is the number of target domain samples which participate in the adjustment of the pre-classification training model in the cutter abrasion state dataset, and then, the expression of the classification error rate e 1 is calculated according to the classification result, and is as follows: , In the formula, For classifying the wrong target domain sample points, and updating the sample weight from the positive and negative aspects The sample weight of the target domain with wrong classification is increased, and the sample weight is increased The expression of (2) is: , in the set maximum iteration times M, the parameters of the pre-classification training model are continuously subjected to iterative updating, when the classification error rate of the decision tree in the target domain machine tool starts to be reduced, the decision tree model with the highest classification precision generated by iteration is taken as output, and the M source domains can generate corresponding tool wear state decision tree classification models after model adjustment, which are expressed as: ; Furthermore, the specific construction process for integrating the plurality of classification models by using the migration stacking technology is as follows: The output of the adjusted pre-classification training model to the target domain sample is as follows: , wherein i and j are the number of target domain samples and pre-classifying training models respectively, and simultaneously, a new decision tree model is additionally trained by using the target domain samples by utilizing a five-fold cross validation method, and the output of which is expressed as h m+1 ,h m+1 ; Weight is distributed to each pre-classifying training model so as to solve the problem of cutter wear state identification error minimization, and an objective function of weight optimization is as follows: , The constraint conditions for optimization are: , wherein gamma j is the weight coefficient of the pre-classification training model, For the output of each pre-classification training model, y i is the real label of the target domain sample; Thus, a decision tree classification model based on migration stacking is obtained The method comprises the following steps: 。
  6. 6. The method for identifying the wear state of a tool across machine tools based on transfer learning according to claim 1, wherein in the third step, the specific process of windowing the real-time state classification sequence output by the decision tree classification model by adopting a sliding window technology is as follows: First, the recognition results are chronologically combined into a recognition result sequence x= , wherein, Representing the cutter abrasion state identification result of the nth sampling point; Then, the recognition result sequence X is sequentially divided into a plurality of mutually overlapping sub-sequences s=by sliding along the observation sequence time direction using a window ω of a fixed size ; During this time, assuming that x t is the first recognition result, x t+ω is the last recognition result, the ith subsequence The method comprises the following steps: ; At this time, the windowing process is completed.
  7. 7. The method for identifying the wear state of a tool across machine tools based on transfer learning as claimed in claim 1, wherein in the third step, the data density theory is applied to analyze the distribution characteristics of classification results in a window and identify transition points of the tool between different wear states, specifically: Assume that The number of target domain samples in the same abrasion state falling into one window, omega is a window with a fixed size, and the approximate number of recognition results in each state is expressed as a data density function p (x) as follows: , And searching for the intersection points of at least 3 corresponding data density curves, namely determining the position of the obtained transition point.
  8. 8. The method for recognizing the wear state of a tool across machine tools based on transfer learning according to claim 1, wherein in the third step, a correction rule for correcting the transition point is: The method comprises the steps of finding a first transition point, correcting a corresponding transition point to be initial abrasion when a recognition result is normal abrasion, finding the first transition point, correcting the corresponding transition point to be normal abrasion when the recognition result is initial abrasion, finding the first transition point, but not finding a second transition point, correcting the corresponding transition point to be normal abrasion when the recognition result is severe abrasion, and correcting the corresponding transition point to be severe abrasion when the recognition result is normal abrasion.
  9. 9. The method for identifying the tool wear state of a cross-machine tool based on transfer learning according to any one of claims 1 to 8, wherein the source domain machine tool is a VMC-850D machine tool and the target domain machine tool is a LEADWELL MCV-1500i+ machine tool.

Description

Cross-machine tool cutter wear state identification method based on transfer learning Technical Field The invention relates to the technical field of tool wear state identification for machining. Background The evolution of the tool, which is an important component in machining, dynamically changes the conditions under which the tool interacts with the workpiece material, directly affecting the cutting force and the generation and distribution of cutting heat. In the field of precision manufacturing of aeroengine parts, when the cutter is worn out beyond a critical threshold, damage such as scratch, micro-crack and the like can occur on the surface of a blade, so that the size of a critical part is beyond a tolerance allowable range, and more seriously, the rapidly deteriorated wear can cause sudden fracture of the cutter, so that expensive blade scrapping is caused, and high-value machine tool equipment can be damaged. Therefore, the research on the abrasion state of the cutter has important research value for guaranteeing the processing efficiency of aviation parts and even the reliability of final products. Meanwhile, the cutter abrasion state identification can remarkably improve the efficiency of the whole machining process and reduce the manufacturing cost. At present, the main research field of the tool wear state identification method aiming at a single machine tool is innovation of signal processing, feature extraction and machine learning models. The method comprises the steps of processing signals such as wavelet transformation, empirical mode decomposition, short-time Fourier transformation and the like, carrying out noise reduction and time-frequency domain feature extraction on various signals such as cutting force, vibration, acoustic emission, power and the like acquired in the cutting process, so as to achieve the aim of accurately reflecting the state change of a cutter, and capturing the relation between the abrasion state of the cutter and the monitored signal features by various machine learning models such as a neural network, a decision tree model, a gating circulation unit, a long-term memory model, a short-term memory model, a support vector machine model and the like. However, in an actual manufacturing scenario, there is a case where a plurality of machine tools are operated in parallel when the parts are mass-processed. Due to the fact that the mechanical structure and the service time of each machine tool are different, and the influence of complex working conditions such as milling technology and the like is added, cutter abrasion label data generated by different machine tools can show more remarkable distribution differences. In order to solve this problem, a tool wear state recognition method across machine tools has received a lot of attention. At present, the multi-working condition tool wear state identification research can be roughly divided into two directions, namely a migration learning method and a model stacking integration method. Aiming at the problem of data distribution difference under multiple working conditions, the inter-domain distribution deviation is reduced through migration learning methods such as an anti-learning method, a Shan Yuanyu generalization network, an unsupervised deep migration learning method and a semi-supervised multi-source meta-domain generalization method, and model generalization is improved, wherein model fine tuning is used as an important branch of migration learning, and aims to quickly adapt to new tasks by combining general features of a pre-classified training model with a small amount of target working condition data, for example, parameter self-adaptive fine tuning is realized by means of a deep forest algorithm, a self-supervised contrast learning framework, a small sample fine tuning method of domain generalization and meta-domain generalization method and the like are constructed, abrasion prediction under a small amount of data is realized, and model stacking is realized by integrating a plurality of fine-tuned pre-classified training models and further improving the reliability and generalization of the model by using methods such as a support vector machine, a decision tree, a naive Bayesian stack generalization integration model and the like. In general, the existing multi-working-condition tool wear state identification method greatly expands the application range of a prediction model, but still has the defects that the existing migration learning research mostly focuses on working condition changes in a limited range, the research on tool wear state classification under a cross-machine tool is not focused at present, the existing multi-working-condition method is difficult to directly adapt, and in multi-source-domain prediction, how to optimize a model stacking mode to achieve the best effect is still to be explored. The problem is urgent to provide a tool wear state identification method which is flexible in ada