Tool wear state identification method based on weight sampling

Through multi-scale serial and parallel convolutional neural network and multi-layer perceptron classifier, combined with the training strategy of weight sampling, the low efficiency and class imbalance of tool wear detection in traditional methods are solved, achieving high accuracy and strong generalization The tool wear status recognition improves the effectiveness of machine tool health management.

CN119048840AActive Publication Date: 2024-11-29SHANDONG HONGYUNLAI IND TECH CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411237250.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-05
Publication Date
2024-11-29
Estimated Expiration
2044-09-05

AI Technical Summary

Technical Problem

Traditional tool wear detection methods are inefficient and have poor accuracy, making it difficult to reflect wear status in real time. Due to category imbalance, the accuracy of model recognition is poor. Single-scale feature extraction cannot capture multi-scale and multi-grained features, which limits the recognition of the model. Performance and generalization capabilities.

Method used

Multi-scale serial and parallel convolutional neural network is used to extract tool vibration signal characteristics, and combined with multi-layer perceptron classifiers, the category imbalance problem is solved through the training strategy of weight sampling, and the generalization ability of the model is enhanced.

Benefits of technology

It improves the accuracy of tool wear status recognition and the generalization ability of model, significantly improves the ability to identify initial wear, normal wear and rapid wear status, and is of great significance to the health management of machine tool tools.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119048840A_ABST
    Figure CN119048840A_ABST
Patent Text Reader

Abstract

A machine tool cutter wear state identification method based on weight sampling belongs to the technical field of machine tool cutter wear state identification, and specifically comprises the following steps: constructing a multi-scale series-parallel convolutional neural network for adaptively extracting cutter vibration signal features; establishing a multi-layer perceptron classifier to realize the identification of initial wear, normal wear and sharp wear states of the cutter; in the model training stage, a training strategy supported by weight collection is designed, and class balance subset training under class imbalance data is achieved. According to the invention, a multi-scale series-parallel structure is designed on the basis of the basic structure of the convolutional neural network, so that the ability of the model to extract multi-granularity features is improved, the designed training strategy based on weight sampling can overcome the tendency discrimination problem of the model to majority classes, and the generalization ability of the model is improved at the same time. High-accuracy and strong-generalization tool wear state recognition is achieved, and the method has important significance on machine tool tool health management.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of machine tool tool wear state identification, and in particular to a machine tool tool wear state identification method based on weight sampling. Background Art

[0002] Traditional tool wear detection methods rely on regular manual inspections or signal analysis based on simple statistical features. These methods are usually inefficient and inaccurate, and it is difficult to reflect the specific status of tool wear in real time. At the same time, in practical applications, tool wear data often have the problem of class imbalance, that is, the number of samples of initial wear, normal wear and acute wear is unevenly distributed, which makes the model tend to be biased towards the majority class during training, thereby affecting the recognition accuracy of the minority class. In addition, the single-scale feature extraction method cannot fully capture the multi-scale and multi-granularity features in the tool vibration signal, which limits the recognition performance and generalization ability of the model. Summary of the invention

[0003] The present invention proposes a method for identifying the wear state of machine tool tools based on weight sampling. The method extracts the features of tool vibration signals by constructing a multi-scale series-parallel convolutional neural network, and recognizes the initial wear, normal wear and acute wear of the tool by combining a multi-layer perceptron classifier. The method introduces a multi-scale series-parallel design into the network structure, which can extract features of multiple granularities and improve the feature capture capability of the model. At the same time, a training strategy based on weight sampling is adopted in the model training stage to effectively solve the problem of class imbalance, enhance the model's recognition capability for minority classes, and improve the generalization performance of the model. Through the above-mentioned technical design, the present invention not only improves the accuracy of tool wear state identification, but also significantly enhances the generalization capability of the model in practical applications, which is of great significance to the health management of machine tool tools.

[0004] The present invention is achieved through the following technical solutions: the present invention relates to a method for identifying the wear state of machine tool tools based on weight sampling, constructs a multi-scale series-parallel convolutional neural network for adaptively extracting tool vibration signal features, and then establishes a multi-layer perceptron classifier to realize the identification of different wear states of the tool; in the model training stage, a training strategy supported by weight sampling is invented to realize class-balanced subset training under class-imbalanced data; the present invention improves the ability of the model to extract multi-granularity features by designing a multi-scale series-parallel structure on the basic structure of the convolutional neural network, and the designed training strategy based on weight sampling can overcome the problem of the model's tendency to distinguish the majority of classes, while improving the generalization ability of the model, and realizing highly accurate and strongly generalized tool wear state identification, which is of great significance to the health management of machine tool tools.

[0005] The multi-scale serial-parallel convolutional neural network includes: starting from the input, first entering a convolution layer with a convolution kernel size of hundreds, which is used to capture larger-scale features; then entering a parallel multi-scale part, which includes three parallel convolution layers, which are convolution layers with convolution kernel sizes of hundreds, tens and units, respectively, and can extract features of different scales; then passing through a convolution layer with a convolution kernel size of tens, further processing medium-scale features; finally entering a convolution layer with a convolution kernel size of units, which is used to capture the most subtle features; through this multi-scale serial-parallel structural design, the network can adaptively extract multi-granularity features in the tool vibration signal, thereby improving the comprehensiveness and accuracy of feature extraction, and helping the subsequent multi-layer perceptron classifier to accurately identify the initial wear, normal wear and acute wear states of the tool.

[0006] The multi-layer perceptron classifier includes: first, the features extracted by the multi-scale series-parallel convolutional neural network enter the first fully connected layer, which contains a number of neurons to achieve preliminary feature fusion and conversion; then, the output enters the second fully connected layer, which further processes and integrates the feature information to improve the expressiveness of the features; finally, after passing through a Softmax classification layer, the probability distribution of different tool wear states is output to achieve the classification goal; the loss function adopts the cross entropy loss function to measure the difference between the predicted category probability distribution and the true category distribution, and is optimized through the back propagation algorithm to minimize the loss function, thereby improving the classification accuracy of the model; the tool wear states include: initial wear, normal wear and acute wear.

[0007] The training strategy supported by weighted sampling refers to collecting tool vibration signal data, marking the data as 1, 2 and 3 according to the data of initial wear, normal wear and rapid wear of the tool, dividing the marked data into training set and test set in a ratio of 8:2, and performing weighted sampling on the training set data to obtain class-balanced training subsets, each training subset corresponding to one round of training; the weighted sampling refers to designing weights for the data of the three tool wear states respectively. , and , the sample size is drawn from the complete training set in a non-replacement sampling manner according to the designed weights. The data is used as a training subset and the above operation is repeated times, corresponding to training subsets; among them, The value range is (0, 3×min{ n 1 , n 2 , n 3 }] , Conditions to be met ,in, It refers to the number of data samples with label 1 in the training set. It refers to the number of data samples with label 2 in the training set. It refers to the number of data samples with label 3 in the training set.

[0008] The class-balanced subset is a subset whose sampling weights satisfy Under the condition of , the training subset obtained by weight sampling will satisfy the expected data volume of each category of data is equal. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is a flow chart of the present invention;

[0010] Figure 2 It is a schematic diagram of weight sampling;

[0011] Figure 3 It is a schematic diagram of the structure of a multi-scale series-parallel convolutional neural network;

[0012] Figure 4 This is a comparison chart of the confusion matrix of wear state recognition results under different methods. DETAILED DESCRIPTION

[0013] With respect to tool wear status identification, this embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

[0014] The first step is to organize the collected data and obtain 300 pieces of initial wear data, 1000 pieces of normal wear data, and 2000 pieces of rapid wear data.

[0015] In the second step, the above three types of data are labeled with values ​​1, 2, and 3, and divided into training and test sets in a ratio of 8:2; among them, the training set contains 240 initial wear data, 800 normal wear data, and 1600 acute wear data; the test set contains 60 initial wear data, 200 normal wear data, and 400 acute wear data.

[0016] The third step is to construct a multi-scale serial-parallel convolutional neural network. Based on the convolution base module, a serial-parallel convolution module is designed. The convolution base module consists of a convolution kernel with a size of And a one-dimensional convolution layer with a convolution step of 1, a batch normalization layer, and a pooling step of The maximum pooling layers are arranged in sequence.

[0017] 3-1) Design of large-scale convolutional base module: convolution kernel size The pooling step is 128. The convolutional base module is 5;

[0018] 3-2) Design of mesoscale convolution base module: convolution kernel size The pooling step is 32. The convolutional base module is 3;

[0019] 3-3) Design of small-scale convolutional base module: convolution kernel size 5, pooling step length The convolutional base module is 3;

[0020] 3-4) Design parallel structure module: realize the expression learning of multi-scale features, specifically including: branch 1 composed of a large-scale convolution base module, branch 2 composed of two serially connected medium-scale convolution base modules, and branch 3 composed of two serially connected small-scale convolution base modules; the output of the parallel structure module It is equal to the concatenation of the outputs of each branch in the last dimension, expressed as ;

[0021] 3-5) Design a multi-scale serial-parallel convolution module: from input to output, it includes a large-scale convolution base module, a parallel structure module, a medium-scale convolution base module and a small-scale convolution base module.

[0022] The fourth step is to build a multi-layer perceptron classifier and set the number of neurons in the first and second fully connected layers to 100 and 3 respectively.

[0023] Step 5: Set the sample size of the training subset and the number of training rounds ;in, Set to 528, Set to 200.

[0024] In the sixth step, the training subset obtained in the fifth step is input into the model constructed in the third and fourth steps, and the cross entropy is used as the loss function to complete the model training through the back propagation mechanism of the neural network.

[0025] In the seventh step, the test data in the second step is input into the model trained in the sixth step, the wear state is identified according to the judgment criterion of maximum category probability, and the wear state identification value is output.

[0026] like Figure 4As shown, a comparison between this embodiment and two methods is provided; the compared methods include a convolutional autoencoder and a three-dimensional modal variational decomposition method: the convolutional autoencoder method first uses principal component analysis to reduce data, and then constructs a convolutional autoencoder model to identify the tool wear state; the three-dimensional modal variational decomposition method performs variational modal decomposition on the vibration and cutting force signals, selects and retains the main modal components, constructs a DBO-OSELM model and uses the dung beetle optimization algorithm to optimize the model. Figure 4 The first figure is the classification result of the convolutional autoencoder method, the second figure is the classification result of the three-dimensional modal variational decomposition method, and the third figure is the classification result of the method proposed in the present invention. The comparison shows that the classification result of this example is better than the other two methods.

[0027] The present invention realizes high-accuracy recognition of initial wear, normal wear and acute wear of machine tool tools by adopting a multi-scale serial-parallel convolutional neural network structure and a training strategy based on weight sampling. The multi-scale serial-parallel convolutional neural network can adaptively extract multi-granularity features of tool vibration signals, improving the comprehensiveness and accuracy of feature extraction; the training strategy based on weight sampling effectively solves the problem of class imbalance, ensures the balance of class samples in each training iteration, and enhances the generalization ability of the model. Compared with the prior art, the present invention has significant advantages. Traditional methods often lead to poor recognition of minority classes when processing class imbalance data, while the present invention balances the class distribution during the training process through a weight sampling strategy, improving the recognition accuracy of minority classes; in addition, the multi-scale serial-parallel structure can capture the multi-scale features of vibration signals, improving the overall recognition performance of the model. Therefore, the present invention not only greatly improves accuracy, but also has stronger generalization ability in practical applications, which is of great significance and broad application prospects for machine tool tool health management.

[0028] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. A method for identifying machine tool wear status based on weight sampling, It is characterized in that A multi-scale series-parallel convolutional neural network is constructed for adaptively extracting tool vibration signal features, and then a multi-layer perceptron classifier is established to recognize different tool wear states. In the model training stage, a training strategy supported by weight sampling is invented to realize class-balanced subset training under class-imbalanced data. The present invention improves the model's ability to extract multi-granularity features by designing a multi-scale series-parallel structure on the basic structure of a convolutional neural network. The designed training strategy based on weight sampling can overcome the problem of the model's tendency to discriminate between most classes, while improving the generalization ability of the model, thereby realizing highly accurate and strongly generalized tool wear state recognition, which is of great significance to machine tool tool health management.

2. According to claim 1, a method for identifying the wear state of machine tool tools based on weight sampling, Its characteristics are: The multi-scale serial-parallel convolutional neural network includes: starting from the input, first entering a convolution layer with a convolution kernel size of hundreds, which is used to capture larger-scale features; then entering a parallel multi-scale part, which includes three parallel convolution layers, which are convolution layers with convolution kernel sizes of hundreds, tens and units, respectively, and can extract features of different scales; then passing through a convolution layer with a convolution kernel size of tens, further processing medium-scale features; finally entering a convolution layer with a convolution kernel size of units, which is used to capture the most subtle features; through this multi-scale serial-parallel structural design, the network can adaptively extract multi-granularity features in the tool vibration signal, thereby improving the comprehensiveness and accuracy of feature extraction, and helping the subsequent multi-layer perceptron classifier to accurately identify the initial wear, normal wear and acute wear states of the tool.

3. The method for identifying the wear state of machine tool tools based on weight sampling according to claim 1, Its characteristics are: The multi-layer perceptron classifier includes: first, the features extracted by the multi-scale series-parallel convolutional neural network enter the first fully connected layer, which contains a number of neurons to achieve preliminary feature fusion and conversion; then, the output enters the second fully connected layer, which further processes and integrates the feature information to improve the expressiveness of the features; finally, after passing through a Softmax classification layer, the probability distribution of different tool wear states is output to achieve the classification goal; the loss function adopts the cross entropy loss function to measure the difference between the predicted category probability distribution and the true category distribution, and is optimized through the back propagation algorithm to minimize the loss function, thereby improving the classification accuracy of the model; the tool wear states include: initial wear, normal wear and acute wear.

4. The method for identifying machine tool wear status based on weight sampling according to claim 1, Its characteristics are: The training strategy supported by weighted sampling refers to collecting tool vibration signal data, marking the data as 1, 2 and 3 according to the data of initial wear, normal wear and rapid wear of the tool, dividing the marked data into training set and test set in a ratio of 8:2, and performing weighted sampling on the training set data to obtain class-balanced training subsets, each training subset corresponding to one round of training; the weighted sampling refers to designing weights for the data of the three tool wear states respectively. , and , the sample size is drawn from the complete training set in a non-replacement sampling manner according to the designed weights. The data is used as a training subset and the above operation is repeated times, corresponding to training subsets; among them, The value range is , Conditions to be met ,in, It refers to the number of data samples with label 1 in the training set. It refers to the number of data samples with label 2 in the training set. It refers to the number of data samples with label 3 in the training set.

5. The method for identifying machine tool wear status based on weight sampling according to claim 1, Its characteristics are: The class-balanced subset is a subset whose sampling weights satisfy Under the condition of , the training subset obtained by weight sampling will satisfy the expected data volume of each category of data is equal.

Citation Information

Patent Citations

  • Visual scene recognition method based on structured information feature decoupling and knowledge migration

    CN114049541A

  • Mechanical fault diagnosis method based on multi-scale component analysis

    CN115391955A

  • Physical and numerical feature combined end-to-end underwater acoustic target classification and identification method

    CN117591941A

  • Method for classifying hyperspectral images on basis of adaptive multi-scale feature extraction model

    US20230252761A1