A tool wear state recognition method based on weight sampling

Abstract

The application discloses a kind of weight sampling-based machine tool tool wear state identification method, belongs to machine tool tool wear state identification technical field, specifically includes: constructing multi-scale series-parallel convolutional neural network is used to adaptively extract tool vibration signal feature, and then establish multilayer perception machine classifier to realize tool initial wear, normal wear and sharp wear state identification;In model training phase, a kind of training strategy under the support of weight collection is designed, to realize the class balanced subset training under the class imbalance data.The application designs the structure of multi-scale series-parallel on the basic structure of convolutional neural network, improves the ability of model to extract multi-granularity features, and the designed training strategy based on weight sampling can overcome the tendency of model to multiple classes discrimination problem, while improving the generalization ability of model, realizing high-precision strong generalization tool wear state identification, which has important significance for machine tool health management.

Description

Technical Field

[0001] This invention relates to the field of machine tool wear condition recognition technology, and in particular to a machine tool wear condition recognition method based on weighted sampling. Background Technology

[0002] Traditional tool wear detection methods rely on periodic manual inspections or signal analysis based on simple statistical features. These methods are typically inefficient and inaccurate, failing to reflect the specific state of tool wear in real time. Furthermore, in practical applications, tool wear data often suffers from class imbalance, meaning the number of samples in initial wear, normal wear, and rapid wear states is unevenly distributed. This causes the model to tend to favor the majority class during training, thus affecting the accuracy of minority class recognition. In addition, single-scale feature extraction methods cannot fully capture the multi-scale, multi-granular features in tool vibration signals, limiting the model's recognition performance and generalization ability. Summary of the Invention

[0003] This invention proposes a weighted sampling-based method for identifying machine tool wear states. It extracts features from tool vibration signals by constructing a multi-scale serial-parallel convolutional neural network and combines this with a multilayer perceptron classifier to identify initial wear, normal wear, and rapid wear states. This method introduces a multi-scale serial-parallel design into the network structure, enabling the extraction of features at multiple granularities and improving the model's feature capture capability. Simultaneously, a weighted sampling-based training strategy is employed during the model training phase to effectively address class imbalance, enhance the model's ability to identify minority classes, and improve its generalization performance. Through these technical designs, this invention not only improves the accuracy of tool wear state identification but also significantly enhances the model's generalization ability in practical applications, which is of great significance for the health management of machine tool tools.

[0004] This invention is achieved through the following technical solution: This invention relates to a method for identifying machine tool wear states based on weighted sampling. A multi-scale serial-parallel convolutional neural network is constructed to adaptively extract tool vibration signal features, and then a multilayer perceptron classifier is established to identify different tool wear states. During the model training stage, a training strategy supported by weighted sampling is invented to achieve balanced subset training under class-imbalanced data. This invention improves the model's ability to extract multi-granularity features by designing a multi-scale serial-parallel structure on the basic structure of the convolutional neural network. The designed weighted sampling-based training strategy overcomes the model's bias in class discrimination and enhances the model's generalization ability, achieving highly accurate and strong generalization of tool wear state identification, which is of great significance for machine tool health management.

[0005] The multi-scale serial-parallel convolutional neural network comprises: starting from the input, it first enters a convolutional layer with a kernel size at the hundreds level to capture larger-scale features; then it enters a parallel multi-scale part, which contains three parallel convolutional layers with kernel sizes at the hundreds, tens, and units levels, respectively, capable of extracting features at different scales; then it passes through a convolutional layer with a kernel size at the tens level to further process medium-scale features; finally, it enters a convolutional layer with a kernel size at the units level to capture the finest features. Through this multi-scale serial-parallel structural design, the network can adaptively extract multi-granularity features from tool vibration signals, thereby improving the comprehensiveness and accuracy of feature extraction, which helps the subsequent multilayer perceptron classifier to accurately identify the initial wear, normal wear, and rapid wear states of the tool.

[0006] The multilayer perceptron classifier comprises: first, features extracted by a multi-scale serial-parallel convolutional neural network enter a first fully connected layer, which contains several neurons to achieve preliminary feature fusion and transformation; then, the output enters a second fully connected layer, which further processes and integrates feature information to improve the expressive power of the features; finally, it passes through a softmax classification layer to output the probability distribution of different tool wear states, achieving the classification objective; the loss function adopts the cross-entropy loss function to measure the difference between the predicted class probability distribution and the true class distribution, and is optimized by the backpropagation algorithm to minimize the loss function, thereby improving the classification accuracy of the model; the tool wear states include: initial wear, normal wear, and rapid wear.

[0007] The weighted sampling-supported training strategy involves collecting tool vibration signal data, labeling the data as 1, 2, and 3 according to the tool's initial wear, normal wear, and rapid wear states, respectively, dividing the labeled data into training and testing sets in an 8:2 ratio, and performing weighted sampling on the training set data to obtain a class-balanced training subset. Each training subset corresponds to one round of training. The weighted sampling refers to designing weights for the data of the three tool wear states. , and According to the designed weights, the sample size is drawn from the complete training set using a non-replacement sampling method. Use the data as a training subset and repeat the above operation. Next, corresponding to get A training subset; among which... The range of values ​​is (0, 3×min{ n 1 , n 2 , n 3 }] , Conditions must be met ,in, This refers to the number of data samples in the training set with a label of 1. This refers to the number of data samples with a label of 2 in the training set. This refers to the number of data samples with a label of 3 in the training set.

[0008] The aforementioned balanced subset refers to the subset whose sampling weight values ​​satisfy... Under the given conditions, the training subset obtained by weight sampling will satisfy the expectation that the amount of data in each category is equal. Attached Figure Description

[0009] Figure 1 This is a flowchart of the invention;

[0010] Figure 2 This is a diagram illustrating weighted sampling;

[0011] Figure 3 This is a schematic diagram of a multi-scale serial-parallel convolutional neural network structure;

[0012] Figure 4 This is a comparison of the confusion matrix of wear state identification results under different methods. Detailed Implementation

[0013] This embodiment, which focuses on tool wear condition identification, is implemented based on the technical solution of the present invention. It provides detailed implementation methods and specific operation procedures, but the scope of protection of the present invention is not limited to the following embodiment.

[0014] The first step was to organize the collected data, resulting in 300 initial wear data entries, 1000 normal wear data entries, and 2000 rapid wear data entries.

[0015] The second step is to label the above three types of data with values ​​of 1, 2 and 3, and divide them into training set and test set in an 8:2 ratio. The training set contains 240 initial wear data, 800 normal wear data, and 1600 rapid wear data. The test set contains 60 initial wear data, 200 normal wear data, and 400 rapid wear data.

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

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

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

[0019] 3-3) Design small-scale convolutional base modules: kernel size The pooling step size is 5. A convolutional base module of type 3;

[0020] 3-4) Design a parallel structure module: to realize multi-scale feature representation learning, specifically including: branch 1 consisting of a large-scale convolutional base module, branch 2 consisting of two cascaded medium-scale convolutional base modules, and branch 3 consisting of two cascaded small-scale convolutional base modules; the output of the parallel structure module. It equals 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 multilayer perceptron classifier, setting the number of neurons in the first and second fully connected layers to 100 and 3, respectively.

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

[0024] The sixth step involves inputting the training subset obtained in the fifth step into the model constructed in the third and fourth steps, using cross-entropy as the loss function, and completing the model training through the backpropagation mechanism of the neural network.

[0025] The seventh step involves inputting the test data from the second step into the model trained in the sixth step, identifying the wear state according to the criterion of the highest class probability, and outputting the wear state identification value.

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

[0027] This invention achieves highly accurate identification of initial wear, normal wear, and rapid wear states of machine tool cutting tools by employing a multi-scale serial-parallel convolutional neural network structure and a weighted sampling-based training strategy. The multi-scale serial-parallel convolutional neural network can adaptively extract multi-granular features from tool vibration signals, improving the comprehensiveness and accuracy of feature extraction. The weighted sampling-based training strategy effectively solves the class imbalance problem, ensuring the balance of class samples in each training iteration and enhancing the model's generalization ability. Compared with existing technologies, this invention has significant advantages. Traditional methods often lead to poor minority class identification when dealing with imbalanced data, while this invention balances the class distribution during training through a weighted sampling strategy, improving the identification accuracy of the minority class. Furthermore, the multi-scale serial-parallel structure can capture multi-scale features of vibration signals, improving the overall recognition performance of the model. Therefore, this invention not only significantly improves accuracy but also has stronger generalization ability in practical applications, making it of great significance and with broad application prospects for machine tool health management.

[0028] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for identifying the wear state of machine tool cutting tools based on weighted sampling, characterized in that, A multi-scale serial-parallel convolutional neural network was constructed to adaptively extract tool vibration signal features, and then a multilayer perceptron classifier was established to identify different tool wear states. In the model training stage, a training strategy supported by weight sampling was proposed to achieve class-balanced subset training under class-imbalanced data. By designing a multi-scale serial-parallel structure on the basic structure of the convolutional neural network, the model's ability to extract multi-granular features was improved. The designed weight sampling-based training strategy can overcome the model's bias in class discrimination and improve the model's generalization ability, achieving highly accurate and strong generalization of tool wear state identification, which is of great significance for machine tool health management. The multi-scale serial-parallel convolutional neural network comprises: starting from the input, it first enters a convolutional layer with a kernel size at the hundreds level to capture larger-scale features; then it enters a parallel multi-scale part, which contains three parallel convolutional layers with kernel sizes at the hundreds, tens, and units levels, respectively, to extract features at different scales; then it passes through a convolutional layer with a kernel size at the tens level to further process medium-scale features; finally, it enters a convolutional layer with a kernel size at the units level to capture the finest features. Through this multi-scale serial-parallel structure design, the network can adaptively extract multi-granularity features from tool vibration signals, thereby improving the comprehensiveness and accuracy of feature extraction, which helps the subsequent multilayer perceptron classifier to accurately identify the initial wear, normal wear, and rapid wear states of the tool. The multilayer perceptron classifier comprises: first, features extracted by a multi-scale serial-parallel convolutional neural network enter a first fully connected layer, which contains several neurons to achieve preliminary feature fusion and transformation; then, the output enters a second fully connected layer, which further processes and integrates feature information to improve the expressive power of the features; finally, it passes through a Softmax classification layer to output the probability distribution of different tool wear states, achieving the classification objective; the loss function adopts the cross-entropy loss function to measure the difference between the predicted class probability distribution and the true class distribution, and is optimized by the backpropagation algorithm to minimize the loss function, thereby improving the classification accuracy of the model; the tool wear states include: initial wear, normal wear, and rapid wear; The weighted sampling-supported training strategy involves collecting tool vibration signal data, labeling the data as 1, 2, and 3 according to the tool's initial wear, normal wear, and rapid wear states, respectively, dividing the labeled data into training and testing sets in an 8:2 ratio, and performing weighted sampling on the training set data to obtain a class-balanced training subset. Each training subset corresponds to one round of training. The weighted sampling refers to designing weights for the data of the three tool wear states. , and According to the designed weights, the sample size is drawn from the complete training set using a non-replacement sampling method. Use the data as a training subset and repeat the above operation. Next, corresponding to get A training subset; among which... The range of values ​​is , Conditions must be met ,in, This refers to the number of data samples in the training set with a label of 1. This refers to the number of data samples with a label of 2 in the training set. This refers to the number of data samples with a label of 3 in the training set.

2. The method for identifying machine tool wear state based on weighted sampling according to claim 1, characterized in that, The aforementioned balanced subset refers to a subset whose sampling weight values ​​satisfy... Under the given conditions, the training subset obtained by weight sampling will satisfy the expectation that the amount of data in each category is equal.

Images