A tool wear monitoring method for machining centers based on semi-supervised evolutionary learning

Abstract

This invention provides a method for monitoring tool wear in machining centers based on semi-supervised evolutionary learning, belonging to the field of machining condition monitoring technology. A metric learning-based deep learning model is trained using a limited number of labeled tool wear samples, and this model serves as the initial semi-supervised SSL model. Unlabeled tool wear samples are sequentially input into the initial SSL model to obtain pseudo-labels. Based on the confidence level of the output probability and the physical laws governing tool wear, pseudo-labeled samples with higher confidence are selected as an expanded dataset of labeled samples. An improved GAN model is used to adaptively maintain class balance in the dataset. The greatest advantage of this method is its ability to dynamically update the deep learning model with a limited number of labeled samples, improving model generalization performance and achieving accurate prediction of unlabeled tool wear.

Description

Technical Field

[0001] This invention belongs to the field of machining condition monitoring, specifically a method for monitoring tool wear in machining centers based on semi-supervised evolutionary learning. Background Art

[0002] Machining is widely used in manufacturing industries such as aerospace, military, and automotive. Tool wear is unavoidable during material removal. Tool wear affects the surface quality of machined parts and production efficiency. Therefore, it is necessary to study tool wear monitoring technology to effectively monitor tool changeover times and improve economic efficiency.

[0003] In recent years, deep learning methods have been successfully applied to the field of tool wear condition monitoring. Although deep learning methods have the advantage of intelligent monitoring and do not require complex physical modeling and prior knowledge, the following problems still exist: 1) Most deep learning models are offline static models, lacking dynamic learning capabilities and unable to predict sample data with fluctuating data distribution [Mao et al. Online detection of bearing incipient fault with semi-supervised architecture and deep feature representation[J]]; 2) Most deep learning models are supervised learning models that require a large number of labeled samples [Zhou et al. Deep learning fault diagnosis method based on global optimization GAN for unbalanced data[J]]. However, obtaining labels for tool wear samples requires specialized instruments for measurement, which is time-consuming and labor-intensive; 3) The monitoring of tool wear condition does not consider the tool wear pattern, which may lead to a significant error rate. Therefore, it is necessary to propose an effective tool wear monitoring method. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method for monitoring tool wear in machining centers based on semi-supervised evolutionary learning. First, a metric learning-based deep learning model is trained using a limited number of labeled tool wear samples, serving as the initial semi-supervised learning (SSL) model. Unlabeled tool wear samples are then sequentially input into the initial SSL model to obtain pseudo-labels. Based on the confidence level of the output probabilities and the physical laws governing tool wear, pseudo-labeled samples with higher confidence levels are selected as an expanded dataset of labeled samples. An improved Generative Adversarial Network (GAN) model is used to adaptively maintain class balance in the dataset. This paper utilizes an evolutionary learning strategy to continuously update the deep learning model by progressively expanding the labeled sample dataset, thereby improving the generalization performance of the deep learning model.

[0005] The technical solution of this invention:

[0006] A semi-supervised evolutionary learning-based method for tool wear monitoring first trains a metric-based deep learning model using a limited number of labeled tool wear samples, serving as the initial SSL model. Next, unlabeled tool wear samples are sequentially input into the initial SSL model to obtain pseudo-labels. Reliable pseudo-label samples are selected based on the confidence level of the pseudo-labels and the physical laws governing tool wear. The selected pseudo-label samples and labeled tool wear samples are merged to form a new labeled tool wear sample set. An appropriate class imbalance rate is set, and it is determined whether the class imbalance rate of the newly formed labeled tool wear sample set exceeds the set imbalance rate. If so, a data generation method is initiated to generate data for the minority class samples, thereby balancing the class distribution of the sample set. The data generation method used in this invention is a GAN model improved within the SSL framework. The initial SSL model is then updated using the newly formed labeled tool wear sample set. In this invention, a tool wear sample is treated as a batch of unlabeled samples, and the model is updated once per batch. In the evolutionary learning process, by continuously adding pseudo-labeled samples to the labeled tool wear sample set, the diversity of labeled samples can be gradually increased, thereby gradually improving the generalization performance of the SSL model. Finally, the classification accuracy of the semi-supervised evolutionary learning model is tested using test samples from the unlabeled tool wear sample set, verifying the accuracy of tool wear monitoring; the specific steps are as follows:

[0007] The first step is to collect tool wear samples under cutting conditions.

[0008] Milling experiments were conducted to collect tool wear samples. Different tool wear sample datasets were collected for each set of cutting parameters. An accelerometer was fixed to the spindle to collect vibration signals from the tool holder. Vibration signals during the machining process were captured as tool wear sample data.

[0009] The second step is to establish the initial SSL model.

[0010] The deep learning model used in this paper is the Deep Extreme Learning Machine (DELM). This model has the advantage of quickly extracting sample features and is suitable for every evolutionary learning process. A softmax classifier is added to the final output layer of DELM to classify tool wear states. To make the features of samples within the same class more compact and the features of samples in different classes more separable, metric learning is introduced into the training of the deep learning model. Assumptions... N is the feature of the i-th sample belonging to category c. c is the number of samples in category c, and μ is the feature center. c The calculation formula is as follows:

[0011]

[0012] In deep learning models, after the feature extraction layer extracts the features of samples, the features of samples of the same class are close to the feature center. The intra-class loss is calculated as follows:

[0013]

[0014]

[0015] This paper sets a threshold between class centers to separate different classes from each other. A larger threshold increases the difficulty of model training, while a smaller threshold leads to overlap of samples from different classes. This paper designs a threshold selection method that considers the distance distribution pattern from each sample to the class center. Weibull fitting is used to fit the distance distribution for each class.

[0016] Weibull model:ρ c (τ c ,κ c ,λ c ) = FitWeibull(d c (4)

[0017] Where, τ c ,κ c ,λ c Representing the Weibull model ρ c The parameter d cThis represents the distance between all samples belonging to class c. To avoid the influence of sample features significantly deviating from the extreme points of the class center, a threshold is determined based on the cumulative distribution function. As a statistical theory, the threshold is usually located in the extreme overlapping region of the score distribution. The inter-class loss is calculated as follows:

[0018] L inter_class =∑ j,k∈Y,j≠k max(0,d j,k -||μ j -μ k ||2) (5)

[0019]

[0020]

[0021] Where, τ j κ j , λ j and τ k , k k , λ k These are the Weibull model ρ j and ρ k The parameter Ω is the boundary probability threshold. j and n k They are d j and d k The number of distances that satisfy the threshold condition. The largest sample distance that satisfies the boundary probability threshold is selected to separate the different class centers.

[0022] This paper selects the DELM model with metric learning as the SSL model. Assume L... class Let represent the cross-entropy loss function used for classification. The final loss function of the DELM model based on metric learning is:

[0023] L total =L class +L intra_class +L inter_class (8)

[0024] The third step is the selection of pseudo-marker samples based on the physical laws of tool wear.

[0025] Tool wear is a slow and continuous physical process that occurs during cutting, and the tool wear value increases monotonically with time. Therefore, when tool wear samples are input into the SSL model in the cutting sequence, the predicted pseudo-labels will also be ordered. Based on this physical knowledge, the confidence level of the model output probability is combined with the physical laws of tool wear to select pseudo-label samples with high reliability.

[0026] Assume the probability output for unlabeled samples is The pseudo-label selection mechanism proposed in this paper consists of the following three stages.

[0027] Phase 1:

[0028] Phase 2:

[0029] Phase 3:

[0030] In Phase 1, pseudo-label samples with a confidence level higher than λ2 (70%) are selected. In Phase 2, pseudo-label samples with a confidence level higher than λ1 (80%) are fully accepted. When the confidence level is between λ2 and λ1, based on the physical laws of tool wear, if the pseudo-label is consistent with the preceding and following pseudo-labels, we consider the pseudo-label sample to be reliable. Phase 3: The selection of pseudo-label samples is checked to make them purer and more reliable.

[0031] The fourth step is to develop a data generation method based on an improved GAN.

[0032] The addition of pseudo-labeled samples to the labeled sample set can lead to class imbalance in the new sample set. To effectively update the SSL model, it is necessary to evaluate the class imbalance rate of the newly formed sample set and generate minority class samples. GANs are unsupervised learning neural network models that can learn the data distribution of real samples and generate new samples with similar distributions. However, traditional GANs suffer from the vanishing gradient problem, which easily generates meaningless data. Ideally, the generated samples should be within the decision boundary of the SSL model while maintaining the diversity of real samples as much as possible. Therefore, this paper improves GANs by using the decision boundary of the SSL model as a constraint.

[0033] Traditional GANs consist of two modules: a generator (G) and a discriminator (D). The generator produces samples, while the discriminator distinguishes between real and generated samples. Through a game between the two, the generator eventually produces samples with a distribution similar to real data, while the discriminator cannot distinguish between real and generated samples. A new module (SSL) is added to the GAN. The network parameters of the SSL model are transferred to the new module of the GAN and frozen during the training of the improved GAN model, thus promoting the GAN to generate sample data within the decision boundary. This improved GAN is named SemiGAN. The loss function of D is:

[0034]

[0035] The loss function of G is:

[0036]

[0037]

[0038] Where G(z) is the sample data generated by mapping from the noisy data z, and D(x) represents the output probability obtained by inputting the real data x. D(G(z)) is the output probability of the generated data. It is the expected value of the probability obtained from actual data. It is the expected value of the probability obtained from the generated data. This is the output probability obtained by inputting the generated sample into the SSL module. y is the label of the generated sample. SSL It is the cross-entropy loss function for classification error. G and D are alternately optimized to generate ideal minority class samples.

[0039] Fifth step: Update the model based on evolutionary learning strategy

[0040] Static SSL methods utilize an initial model trained on a small number of labeled samples to predict all unlabeled samples. However, due to the limited number of labeled samples, the error rate of pseudo-labels increases, and high-confidence-threshold pseudo-label selection methods waste a large number of unlabeled samples. Therefore, we employ an evolutionary learning strategy to dynamically train the SSL model by gradually adding new pseudo-labeled samples to the labeled sample set. This allows the SSL model to adapt to data changes and extract more useful information from unlabeled samples. Based on tool wear patterns, this paper treats a set of tool wear samples collected from a single tool as a batch, and updates the SSL model once per batch.

[0041] Step 6: Verify the prediction accuracy of the semi-supervised evolutionary learning method.

[0042] Unlabeled tool wear test samples were input into the SSL model, which had completed semi-supervised evolutionary learning, to classify tool wear states and verify the accuracy of the proposed tool wear monitoring method based on semi-supervised evolutionary learning.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] 1. The tool wear monitoring method based on semi-supervised evolutionary learning proposed in this invention establishes a deep learning model based on metric learning, which is beneficial for promoting intra-class compactness and inter-class separability; the physical laws of tool wear are considered when selecting pseudo-labels, which improves the credibility of pseudo-label samples; under the SSL framework, the structure of traditional GAN ​​is improved to reduce the damage of invalid samples to the SSL model; the evolutionary learning strategy is used to gradually update the SSL model, which is beneficial for continuously improving the generalization performance of the model.

[0045] 2. This invention is used in the field of tool wear monitoring, and can effectively solve the problem of poor generalization ability of deep learning models due to insufficient tool wear sample labels, thereby improving the accuracy of tool wear monitoring. Attached Figure Description

[0046] Figure 1 This is a flowchart of a tool condition monitoring method based on semi-supervised evolutionary learning.

[0047] Figure 2 This is a schematic diagram showing the sensor installation location.

[0048] Figure 3a () represents the feature extraction distribution based on DELM. Figure 3b ) represents the feature extraction distribution of DELM based on metric learning.

[0049] Figure 4 This is a pseudo-label sample selection method based on the physical laws of tool wear.

[0050] Figure 5a (This is a schematic diagram of the SemiGAN architecture.) Figure 5b (This refers to a spectrum diagram of the generated data and the real data.) Figure 5c () represents the characteristic distribution of real data and generated data in the SSL model.

[0051] Figure 6a () represents the classification result of the initial semi-supervised learning model; Figure 6b () represents the classification results of the model after semi-supervised evolutionary learning.

[0052] Figure 2 In the middle: 1. Spindle; 2. Three-dimensional accelerometer; 3. Tool holder; 4. Tool; 5. Die steel workpiece. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the accompanying drawings are provided. Figure 1 Taking the milling of mold steel on a three-axis vertical machining center as an example, the implementation of the present invention will be described in detail.

[0054] In this experiment, a 10mm diameter cutting tool was used to mill the surface of the mold steel workpiece, such as... Figure 2As shown. An accelerometer was fixed on the spindle to collect vibration signals from the tool holder. Wear values ​​on the blade surface were detected using an ultra-deep microscope. Cutting parameters were: depth of cut 0.4 mm, feed rate 480 mm / min, and spindle speed 3000 r / min. Four sets of tool wear sample datasets were collected. Each sample contained 2500 time-domain data points. Based on the characteristics of tool wear, the samples were divided into three categories: early wear (VB < 0.15 mm), intermediate wear (0.3 mm > VB > 0.15 mm), and severe wear (VB > 0.3 mm). One set of tool wear samples was selected as labeled samples, and the remaining samples were used as unlabeled samples. Training and testing datasets were selected from each set of unlabeled tool wear samples at a ratio of 3:2. Detailed information about the datasets is shown in Table 1.

[0055] Table 1 Machining cutting parameters

[0056]

[0057] Figure 3 shows the feature distribution of labeled samples extracted by DELM and metric learning-based DELM. It can be seen that metric learning can make the features of samples of the same class more compact and the features of samples of different classes more distinct. Figure 3 also shows the feature distribution of unlabeled samples. It can be seen that the three types of tool wear in the labeled samples are clearly separated in the feature space, while the deep features of the unlabeled data are scattered across the three classes. This indicates a difference in data distribution between unlabeled and labeled samples, and that a deep learning model trained with limited data cannot accurately predict a large number of unlabeled samples. However, it is still evident that metric learning helps the DELM model separate unlabeled samples of different classes and makes samples of the same class closer together, which will improve the classification ability of deep learning models for unlabeled samples. Figure 4 The pseudo-label selection mechanism with confidence levels between 70% and 80% is shown. It can be seen that combining the pseudo-label selection method with the physical laws of tool wear can improve the utilization rate of unlabeled samples and filter out erroneous samples. Figure 5aFigures b) and c) show the structure of the improved GAN. The imbalance rate in this paper is 1.5. After obtaining the newly formed sample set, SemiGAN is used for data generation when the imbalance rate of the data classes reaches 1.5. Both G and D in SemiGAN are three-layer backpropagation neural networks with 128 and 256 hidden layer nodes, respectively. The model iterations are 200. Figures b) and c) show that the generated samples conform to the data distribution characteristics of the real samples. By visualizing the feature distribution of real samples and generated samples in the feature space of the SSL model, it can be seen that the generated samples are distributed within the decision boundary of the model and are distributed near the real samples, expanding the distribution range of real samples and improving the diversity of real samples. This proves the effectiveness of the SemiGAN-based data generation method. The test results are shown in the confusion matrix in Figure 6. Figure 6a The results show that the initial SSL model was used directly to predict unmarked tool wear test samples. Figure 6b The results of the model's predictions of unlabeled tool wear test samples after semi-supervised evolutionary learning are shown. It can be seen that the semi-supervised evolutionary learning method can accurately monitor tool wear conditions. This method effectively solves the problems of insufficient labeled samples and poor model generalization ability in the field of tool wear monitoring.

Claims

1. A method for monitoring tool wear in machining centers based on semi-supervised evolutionary learning, characterized in that, First, a metric learning-based deep learning model is trained using a limited number of labeled tool wear samples, serving as the initial SSL model. Second, unlabeled tool wear samples are sequentially input into the initial SSL model to obtain pseudo-labels. Reliable pseudo-label samples are selected based on the confidence level of the pseudo-labels and the physical laws governing tool wear. The selected pseudo-label samples and labeled tool wear samples are merged to form a new labeled tool wear sample set. A dataset class imbalance rate is set, and it is determined whether the class imbalance rate of the new labeled tool wear sample set exceeds the set imbalance rate. If it does, a data generation method is initiated to generate minority class samples to balance the new labeled tool wear sample set. The method involves classifying the tool wear sample set using a newly formed labeled tool wear sample set. The data generation method is an improved GAN model within the SSL framework. The initial SSL model is then updated using this newly formed labeled tool wear sample set. In this method, a tool wear sample is treated as a batch of unlabeled samples, and the model is updated once per batch. During evolutionary learning, pseudo-labeled samples are continuously added to the labeled tool wear sample set to gradually increase the diversity of labeled samples, thereby gradually improving the model's generalization performance. Finally, the classification accuracy of the semi-supervised evolutionary learning model is tested using test samples from the unlabeled tool wear sample set, verifying the accuracy of tool wear monitoring. The specific steps are as follows: The first step is to collect tool wear samples under cutting conditions. An accelerometer is fixed to the spindle to collect vibration signals from the tool holder. Vibration signals during the machining process are captured as tool wear sample data. Different tool wear sample datasets are collected for each set of cutting parameters. The second step is to establish the initial SSL model. The deep learning model is a Deep Extreme Learning Machine (DELM). A softmax classifier is added to the final output layer of the DELM to classify tool wear states. To make the features of samples within the same class more compact and the features of samples in different classes more separable, metric learning is introduced into the training of the DELM. Assuming... It is the feature of the i-th sample belonging to category c, N c The number of samples in category c, and the feature centers μ. c The calculation formula is as follows: In deep learning models, after the feature extraction layer extracts the features of samples, the features of samples of the same category are close to the feature center. The intra-class loss is calculated as follows: A threshold is set between category centers to separate different categories from each other; a threshold selection method is designed that considers the distance distribution pattern from each sample to the category center; Weibull fitting is used to fit the distance distribution for each category: Weibull model:p c (t c ,k c ,l c )=FitWeibull(d c ) (4) Where, τ c ,κ c ,λ c Representing the Weibull model ρ c The parameter d c This represents the distance to all samples belonging to class c; to avoid the influence of sample features significantly deviating from the class center extreme point, a threshold is determined based on the cumulative distribution function; the inter-class loss is calculated as follows: L inter_class =∑ j,k∈Y,j≠k max(0,d j,k -||m j -m k ||2) (5) Where, τ j k j , λ j and τ k k k , λ k These are the Weibull model ρ j and ρ k The parameters, Ω is the boundary probability threshold, n j and n k They are d j and d k The number of distances that satisfy the threshold condition; the maximum sample distance that satisfies the boundary probability threshold is selected to separate different class centers; Choose the DELM model with metric learning as the SSL model; assume L class Let represent the cross-entropy loss function used for classification. The final loss function of the DELM model based on metric learning is: L total L class +L intra_class +L inter_class (8) The third step is the selection of pseudo-marker samples based on the physical laws of tool wear. Tool wear is a slow and continuous physical process that occurs during cutting, and the tool wear value increases monotonically with time. Therefore, when tool wear samples are input into the SSL model in the cutting sequence, the predicted pseudo-labels will also be ordered. Therefore, by combining the confidence level of the SSL model output probability with the physical law of tool wear, pseudo-label samples with high reliability can be selected. Assume the probability output for unlabeled samples is The pseudo-label selection mechanism proposed in this method consists of the following three stages: Phase 1: Phase 2: Phase 3: Phase 1: Select pseudo-label samples with a confidence level higher than λ2 = 70%; Phase 2: Pseudo-label samples with a confidence level higher than λ1 = 80% are fully accepted; When the confidence level is between λ2 and λ1, according to the physical laws of tool wear, if the pseudo-label is consistent with the previous and subsequent pseudo-labels, the pseudo-label sample is considered credible; Phase 3: Check the selection of pseudo-label samples to make them purer and more reliable. The fourth step is to develop a data generation method based on an improved GAN. Adding pseudo-labeled samples to the labeled sample set will cause class imbalance in the new sample set. In order to effectively update the SSL model, it is necessary to evaluate the class imbalance rate of the newly formed sample set and generate minority class samples. The decision boundary of the SSL model is used as a constraint to improve the GAN. By adding an SSL module to the GAN, the network parameters of the SSL model are transferred to the SSL module of the GAN and frozen during the training of the improved GAN model, which helps the GAN generate sample data within the decision boundary; the improved GAN is named SemiGAN; the loss function of D is: The loss function of G is: Where G(z) is the sample data generated by mapping from the noisy data z, D(x) represents the output probability obtained by inputting the real data x, and D(G(z)) is the output probability of the generated data. It is the expected value of the probability obtained from actual data. It is the expected value of the probability obtained from the generated data. This is the output probability obtained by inputting the generated sample into the SSL module, where y is the label of the generated sample, and L is the output probability. SSL It is the cross-entropy loss function of the classification error, and G and D are alternately optimized to generate ideal minority class samples; Fifth step: Update the model based on evolutionary learning strategy An evolutionary learning strategy is adopted to dynamically train the SSL model by gradually adding new pseudo-labeled samples to the labeled sample set, so that the SSL model can adapt to data changes and extract more useful information from unlabeled samples. This method is based on the tool wear pattern, and the tool wear sample set collected from a tool is treated as a batch, and the SSL model is updated once for each batch. Step 6: Verify the prediction accuracy of the semi-supervised evolutionary learning method. Unlabeled tool wear test samples were input into the SSL model, which had completed semi-supervised evolutionary learning, to classify tool wear states and verify the accuracy of the proposed tool wear monitoring method based on semi-supervised evolutionary learning.

Images