CN-121705721-B - Spindle cutter abrasion monitoring method based on physical guidance and semi-supervised learning

Abstract

The invention discloses a spindle tool wear monitoring method based on physical guidance and semi-supervised learning, which comprises the steps of collecting tool acceleration signals, extracting multi-domain statistical features, carrying out cross-tool robustness screening, constructing accumulated features reflecting historical trends to form a mixed input set, utilizing a parallel symbol regression network introducing monotonicity penalty to mine an explicit physical equation conforming to a degradation mechanism, carrying out prediction and smooth monotonicity processing on unlabeled data by utilizing the equation to generate a physical weak label, constructing a mixed training set, designing a physical guidance cross-attention mechanism, carrying out weighted fusion on the statistical and accumulated features by taking the physical equation as priori, constructing a lightweight neural network processing fusion feature, and designing differentiated loss function joint training based on the mixed training set. The invention can break through the dependence on a large amount of labeling data, ensure the physical consistency of the prediction result and meet the high-precision monitoring requirement of the industrial field.

Inventors

• FU JIANZHONG
• PENG YEZHEN
• YU FENGWEN
• KANG WEIMIN
• Han Nanjie
• RUAN ZHIBIN
• Shen Yaoqi
• YAO XINHUA
• LUAN CONGCONG

Assignees

• 浙江大学
• 浙江先端数控机床技术创新中心有限公司

Dates

Publication Date

20260508

Application Date

20260210

Claims (5)

1. 1. The main shaft cutter abrasion monitoring method based on physical guidance and semi-supervised learning is characterized by comprising the following steps of: (1) Collecting acceleration signals of a spindle cutter, extracting multi-domain statistical features comprising a time domain, a frequency domain and a time domain, and performing cross-cutter adaptive screening to obtain a statistical feature subset; (2) Aiming at the statistical features in the statistical feature subsets, constructing an accumulated feature for describing the continuous degradation trend of the cutter performance along with time, combining the statistical features and the accumulated feature in the physical channel dimension to obtain a mixed input feature set, wherein the construction formula of the accumulated feature is as follows: ; in the formula, For the current moment of cutting, The number of the cutter is given to the cutter, Is the first Time of day The statistical characteristics of the dimensions are maintained, For the corresponding cumulative features, for characterizing the historical evolution trend of the wear process from the data plane; (3) Performing symbolized regression mining on the mixed input feature set by using a parallel symbolized regression network introducing a monotonicity penalty term to obtain an explicit physical equation conforming to a degradation mechanism, wherein the monotonicity penalty term is defined as follows: ; in the formula, Is the first Monotonicity penalty items corresponding to the individual tools, Is the first The total number of cuts of the individual tools, Regression of network presence for parallel symbols A predicted value of time; (4) Carrying out prediction and smooth monotonic processing on the unlabeled cutter data by using an explicit physical equation to generate a physical weak label, and constructing a mixed training set with the strong label data; (5) The method comprises the steps of designing a physical guidance cross attention mechanism, taking the output of an explicit physical abrasion equation as a physical priori query vector, carrying out weighted screening and gating fusion on statistical features and accumulated features to obtain a fusion feature vector under physical guidance, wherein the specific process is as follows: The method comprises the steps of designing a physical guidance cross attention module, mapping statistical features and accumulated features into keys and values by taking the output of an explicit physical wear equation as a physical priori query vector, calculating attention weights of the physical channels to the statistical feature channels and the accumulated feature channels, then constructing an adaptive gating mechanism, calculating gating coefficients by using a Sigmoid activation function, and fusing original statistical features, weighted statistical features, original accumulated features and weighted accumulated features according to the gating coefficients to obtain fused feature vectors under physical guidance; and calculating the attention weights of the physical channels to the statistical characteristic channels and the cumulative characteristic channels, wherein the formula is as follows: ; ; ; in the formula, And The attention weights representing the statistical and cumulative characteristic channels, respectively, For a query vector mapped to a physical channel, And Key vectors mapped to statistical features and cumulative features respectively, As a dimension of the features, And Respectively representing the normalized attention weights of the statistical characteristic channel and the cumulative characteristic channel; (6) Constructing a lightweight neural network comprising a bidirectional gating circulating unit and a residual error full-connection module, and performing joint training based on a hybrid training set design differentiated hybrid loss function; (7) In the application process, acceleration signals of the spindle cutter are collected, the fusion characteristic vector is obtained through processing, and then the fusion characteristic vector is input into a trained lightweight neural network, and a final abrasion predicted value is output.
2. 2. The method for monitoring the wear of the spindle cutter based on physical guidance and semi-supervised learning according to claim 1, wherein in the step (1), the cross-cutter adaptive screening is performed, and the specific process is as follows: First, each statistical feature is calculated Wear and tear tab Spearman rank correlation coefficient absolute value between Mutual information After normalizing the two indexes, calculating the comprehensive score of the single tool : ; In the formula, For training the tool set, c is the tool number, , And (3) with Respectively normalizing the correlation coefficient and the mutual information; Representing dimensions; subsequently, a cross-tool integrated sensitivity score is calculated And according to the average value of the scores And standard deviation of Setting a threshold value for screening: ; ; in the formula, Is an empirical threshold coefficient.
3. 3. The method for monitoring the wear of a spindle tool based on physical guidance and semi-supervised learning according to claim 1, wherein the specific process of the step (3) is as follows: And in the process of mining, a monotonicity penalty term is introduced on the basis of mean square error loss to restrict a search space, and an explicit physical abrasion equation with high fitting precision and physical evolution rule consistency is optimized in a Pareto front.
4. 4. The method for monitoring the wear of a spindle tool based on physical guidance and semi-supervised learning according to claim 1, wherein the formula for calculating the gating factor is: ; in the formula, Activating a function for Sigmoid; Is a weight matrix which can be learned; Representing a splicing operation; Is a gating coefficient; 、 、 Respectively representing the normalized physical channel characteristics, the statistical channel characteristics and the cumulative channel characteristics.
5. 5. The method for monitoring the wear of a spindle tool based on physical guidance and semi-supervised learning according to claim 1, wherein in the step (6), the differentiated mixing loss function is formulated as follows: ; Employing strong supervision loss for strong label sample set : ; For weak label sample set, weak supervision loss is adopted : ; In the formula, In the case of a strong sample set of tags, As a sample set of weak tags, Is the first The predicted value of the individual samples is calculated, Is the first A true wear measurement of the individual samples, The first step generated for step (4) The physical weakness of the tag of each sample, As a result of the weak tag weight coefficient, , Is the difference between the predicted value and the actual value, The threshold for transition from the secondary loss to the linear loss is controlled.

Description

Spindle cutter abrasion monitoring method based on physical guidance and semi-supervised learning Technical Field The invention belongs to the technical field of intelligent manufacturing and industrial equipment state monitoring, and particularly relates to a spindle cutter abrasion monitoring method based on physical guidance and semi-supervised learning. Background With the rapid development of high-end manufacturing and intelligent manufacturing, the numerical control machine tool is widely applied to the fields of aerospace, automobiles, dies, precise part machining and the like. As a key execution part in the cutting process, the abrasion state of a cutter directly affects the surface quality, the dimensional accuracy and the machining efficiency of a workpiece, and excessive abrasion can even lead to cutter tipping, workpiece scrapping and equipment failure, thereby causing great economic loss and safety risk. Therefore, on the premise of ensuring the production takt, the tool wear state is monitored online, accurately and reliably, and the method has important significance for realizing predictive maintenance of equipment, improving the stability of processing quality and reducing the manufacturing cost. In the prior art, chinese patent document with publication number of CN111832432A discloses a real-time prediction method of cutter abrasion based on wavelet packet decomposition and deep learning. According to the method, the non-stationary sensing signals are converted into a two-dimensional time-frequency matrix through wavelet packet decomposition, the deep features are extracted by using a convolutional neural network, a monitoring model is built, and the cutter abrasion prediction accuracy is improved to a certain extent. However, the method mainly adopts a full-supervision learning paradigm to mine statistical correlation among data, so that a large amount of unlabeled process data in an industrial field is difficult to fully utilize, and meanwhile, the characteristic learning process lacks explicit constraint on a tool abrasion physical evolution rule, so that generalization capability is limited in a small sample scene. The Chinese patent document with publication number of CN110355608A proposes a cutter abrasion loss prediction method based on a self-attention mechanism and a bidirectional long-short time memory network, which is used for capturing the long time sequence dependency relationship of a monitoring signal and realizing regression prediction of cutter abrasion loss. The method has certain advantages in the aspect of time sequence feature modeling, but basically still belongs to a data fitting model based on implicit feature mapping, and does not integrate the physical priori of the tool abrasion process into network constraint, so that the model is easy to show insufficient robustness under complex working conditions, and popularization and application of the model in engineering sites are limited. Overall, the existing methods still have problems in engineering floor-based mismatch with field data morphology. The characteristic of 'a small amount of labeled samples and a large amount of unlabeled process data' is commonly presented in an industrial field, but the technology relies on full supervision training, so that potential value of the unlabeled data in abrasion modeling is difficult to effectively mine, and application capability of the model under a label scarcity condition is limited. Meanwhile, many schemes based on deep learning still mainly adopt statistical fitting, and lack explicit guidance and consistency constraint on the physical evolution law of tool wear, so that the interpretation and prediction reliability of the model are insufficient. Based on this, it is necessary to provide an online monitoring method capable of fully utilizing a large number of unlabeled samples and fusing the cutter wear physical prior so as to improve the stability and practicality of the wear monitoring model in a real industrial scene. Disclosure of Invention The invention provides a spindle tool wear monitoring method based on physical guidance and semi-supervised learning, which can solve the technical problems that the existing tool wear monitoring technology has high dependence on full-period labeling data, cannot effectively utilize mass unlabeled data in an industrial field, and a pure data driving model lacks physical consistency constraint and has insufficient robustness. A spindle cutter wear monitoring method based on physical guidance and semi-supervised learning comprises the following steps: (1) Collecting acceleration signals of a spindle cutter, extracting multi-domain statistical features comprising a time domain, a frequency domain and a time domain, and performing cross-cutter adaptive screening to obtain a statistical feature subset; (2) Aiming at the statistical features in the statistical feature subsets, constructing an accumulated feature for describing the c