Method for generalizing monitoring domain of tool wear state under unknown working condition

Abstract

The invention relates to the field of machining, in particular to an unknown working condition tool wear state monitoring method driven by a bilevel feature enhancement domain generalization network fusing a contrast learning module and dynamic label smoothing. According to the method, based on discriminative feature extraction of the parallel interactive fusion convolutional network, plug-and-play discriminative feature expansion without parameter introduction and multi-source domain data enrichment are adopted. According to the method, a parallel interactive fusion convolutional network structure is designed to be used for constructing the feature extractor, and compared with a classic convolutional neural network, interactive fusion convolution can learn refined local features and a dependency relationship between data in a long range, and the generalization ability of the feature extractor is enhanced; a plug-and-play and parameter-introduction-free data enhancement module is adopted to be embedded into a feature extractor, features obtained through interactive fusion convolution are directly subjected to operation to complete feature enhancement in an implicit mode, and finally a tool wear monitoring model high in generalization ability is established.

Description

Technical Field

[0001] The present invention relates to the field of mechanical processing, and in particular to a method for monitoring tool wear in unknown working conditions driven by a dual-level feature enhancement domain generalization network that integrates a contrastive learning module and dynamic label smoothing. Background Art

[0002] With the intelligent, digital, and networked transformation and upgrading of manufacturing processes, tool wear condition monitoring has become a crucial component of intelligent machining systems. Due to the diverse and complex combinations of machining conditions, such as machining paths, tool-material matching types, machining media, and cutting parameters, tool condition degradation patterns are dynamic, random, and nonlinear, making them difficult to perceive and judge under complex and variable conditions. However, new-generation artificial intelligence technologies, with their powerful nonlinear modeling and adaptive learning capabilities, combined with fast-response and robust anti-interference sensing equipment, have become a key technical approach for tool condition monitoring. To address the problem of tool wear condition monitoring under variable conditions, research institutions both domestically and internationally have proposed a wide variety of transfer learning models based on domain adaptation theory, which have become the primary technical solution. The article "Novel tool wear prediction method based on multimodal information fusion and deep subdomain adaptation," published in Volume 224 of the internationally renowned journal Mechanical Systems and Signal Processing in February 2025, addresses the problem of tool condition monitoring under variable conditions through a transfer learning model based on domain adaptation theory. However, when modeling a transfer learning model based on domain adaptation theory, not only does it require cutting physics data from the old machining conditions (source domain) to participate in the modeling, but it also inevitably requires access to cutting physics data from the new machining conditions (target domain). To mitigate the deviations that occur during domain alignment, sufficient and complete target domain data is often collected, making the data collection process time-consuming and costly. Furthermore, the cutting physics data under some machining conditions may be missing or incomplete, making it difficult to meet the prerequisites for modeling a transfer learning model based on domain adaptation theory. Therefore, transfer learning models based on domain adaptation cannot solve the problem of tool condition monitoring under unknown conditions, severely limiting their promotion and application in the industrial field. Summary of the Invention

[0003] When solving the problem of tool wear status monitoring under variable working conditions, the transfer learning model based on domain adaptation theory needs to access target domain data for modeling. However, the collection process of target domain data is time-consuming and costly, or even impossible to obtain, which makes it difficult to promote and apply it in the industrial field. The present invention proposes a dual-level feature enhancement domain generalization method that integrates dynamic label smoothing and contrastive learning. The method completes modeling without accessing target domain data, thereby solving the problem of tool wear status monitoring under unknown working conditions.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions: ①Discriminative feature extraction based on parallel interactive fusion convolutional networks; The 101 interactive fusion convolutional network consists of three convolutional layers. The first two convolutional layers have convolution kernels of different sizes and are distributed in parallel. The original cutting signal data is input to the first two convolutional layers at the same time. The features obtained by the convolution operation are passed to the Gaussian error linear unit activation function. After performing the random drop operation on the obtained features, they are interactively multiplied to complete the interaction of features of different scales, as shown in the following formula: Where f represents the Gaussian error linear unit activation function, ⊙ represents the dot product operation, x represents the multi-channel cutting signal data, Conv1 represents the first convolutional layer module, Conv2 represents the second convolutional layer module, O1 represents the features obtained under the first branch, and O2 represents the features obtained under the second branch; 102 Add the outputs O1 and O2 obtained from the two branches and use them as the input of the third convolutional layer, as shown in the following formula: F=Conv3(O1+O2) (2) Among them, Conv3 represents the third convolutional layer module, and F represents the added features; 103 Design a parallel interactive fusion convolution structure, one branch contains one interactive fusion convolution block, and the other branch contains two interactive fusion convolution blocks; ② Using a plug-and-play, parameter-free approach to discriminative feature expansion and multi-source domain data enrichment; 201 A hybrid feature enhancement strategy is used to improve the generalization ability of the feature extractor. The mean and standard deviation statistics of the feature map obtained by the feature extractor are calculated as follows: Where H is the total number of channels of the cutting signal, W is the length of the cutting signal in each sample, f is the feature map obtained by the feature extractor, and f h,w Represents the value at the h and w positions in the feature map, μ(f) represents the mean of the feature map, and σ(f) represents the variance of the feature map; 202 The mixed value of the characteristic statistics is obtained by using the weight of random sampling in the Beta distribution in a linear weighted manner, as shown in the following formula: Among them, λ represents the linear weight value, represents a randomly shuffled feature map, represents the variance of the feature map after random shuffling, Represents the mean of the feature map after random shuffling, δ mix represents the variance after linear weighting, η mix represents the linearly weighted mean; 203 According to the example standardization calculation process, the enhanced feature output is: Among them, f aug represents the enhanced feature map, and ⊙ represents the dot product operation. The feature extractor proposed in 204 contains two branches, which obtain feature maps with different depth levels, and perform feature enhancement operations on the feature maps obtained by each branch respectively; ③ Enhancement of key discriminative features by integrating Convformer modules; 301 Use the Convformer module to further process the combined enhanced features f aug , a learnable linear mapping is used to convert the input feature map into a query vector matrix Q, a key vector matrix K and a value vector matrix V, and the attention weight Z of each attention head is obtained by scaling the dot product operation and the Softmax function j ; 302 uses a convolutional neural network to downsample the input feature map and effectively learn the local features within the feature map; 303 uses convolutional mapping to replace the forward feedback mapping in transformer to further reduce the number of model parameters. Through the standardization layer and residual connection, the output f of the Convformer module is obtained. d . ④Domain-invariant feature extraction driven by dynamic label smoothing domain discriminator; 401 The discriminative features obtained by the feature extractor are converted into a probability vector related to the number of source domains through a fully connected layer, and further processed using a normalization layer and a maximum function; The output of the 402 normalization layer is multiplied by the bitwise true domain label, the output of the maximum function is multiplied by the true domain label, and the multiplication results are added together. The predicted probability vector information is used to adjust the hard label to the form of a soft label, as shown in the following formula: Among them, onehot() represents the one-hot encoding function, y t represents the original hard tag, represents the adjustment coefficient, f p represents the predicted probability vector, represents the adjusted soft label, and ~ represents bitwise operation; 403 KL divergence loss is used to evaluate the distribution between the output of the domain discriminator after dynamic label smoothing and the adjusted target domain label, as follows: Among them, N represents the number of samples in a small batch, K represents the number of source domain datasets used for model training, represents the k-th source domain dataset and the adjusted soft label of the t-th sample, Represents the probability vector of the t-th sample of the k-th source domain dataset output by the dynamic label smoothing domain discriminator; ⑤Use supervised contrastive learning loss to constrain class representation information in model modeling; 501 uses supervised contrastive learning loss to aggregate samples of the same category, enhancing the separation of classification boundaries and the generalization ability of the model, as shown in the following formula: in, Represents the mini-batch sample and label y i the same quantity; When y i and y j If they are of the same category, it is 1, otherwise it is 0; f i d , and They represent the discriminative features output by the i-th, j-th, and k-th Convformer modules in a small batch of samples, respectively; exp() represents the exponential function; sim represents the cosine similarity function; and τ represents the temperature coefficient, which is used to control the influence of negative samples on the training process. ⑥ A dual-level feature-enhanced domain generalization network that integrates dynamic label smoothing and contrastive learning through multi-source domain data training; 601 optimizes the parameters of the feature extractor, dynamic label smoothing domain discriminator, and classifier. The classifier consists of multiple layers of fully connected layers, and the cross entropy loss function is used to calculate the classification loss of the classifier. When designing the total loss function, the contribution of the dynamic label smoothing domain discriminator and the supervised contrastive learning loss to the training process is adjusted by the weight coefficient, and the adversarial loss of the joint domain discriminator is used. The total loss function is: L total =α·L sup +(1-α)·L c -β·Ladv (10) Among them, α and β represent weight coefficients, L sup Denotes the supervised contrastive learning loss value, L c represents the classification loss of the classifier, L adv represents the loss value of the dynamic label smoothing domain discriminator during domain adversarial training, L total Indicates the total loss value; 603 optimizes the model parameters through the error back propagation mechanism, as shown below: Among them, θ F ,θ d and θ c Denote the parameters of the feature extractor, domain discriminator, and classifier, respectively, and η denotes the learning rate. After the model training is completed, the performance of the model is verified using data under unknown working conditions (target domain data).

[0005] The positive effects of the present invention are as follows: The present invention designs a parallel interactive fusion convolutional network structure for constructing a feature extractor. Compared with classic convolutional neural networks, interactive fusion convolution can learn refined local features and long-range dependencies between data, thereby enhancing the generalization ability of the feature extractor. A plug-and-play, parameter-free data enhancement module is embedded in the feature extractor, directly operating on the features obtained by interactive fusion convolution to implicitly complete feature enhancement. The enhanced features expand and enrich the source domain data, allowing the feature extractor to obtain a wider range of knowledge, reducing the dependence of the domain generalization model on diverse source domain data, and thus improving the accuracy of tool wear monitoring. By integrating the advantages of domain adversarial training and dynamic label smoothing, a domain label dynamic smoothing adversarial network is proposed to reflect the differences between inter-class and intra-class samples, learning domain-invariant features from the sample level, thereby enhancing the classification ability of the model. Furthermore, a supervised contrastive learning loss is incorporated into the overall loss function to constrain the class representation differences caused by variable processing conditions in the model training process, ultimately establishing a tool wear monitoring model with strong generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 Interactive fusion convolution structure Figure 2 Feature enhancement flow chart Figure 3 Flowchart of the implementation of dynamic label smooth domain discriminator Figure 4 Overall framework diagram of tool wear status monitoring under unknown working conditions Figure 5 Milling tool wear test platform Figure 6 Tool wear status monitoring results under unknown working conditions during milling Figure 7 Turning tool wear test platform Figure 8 Tool wear status monitoring results under unknown working conditions during turning DETAILED DESCRIPTION

[0007] The present invention is described in detail below with reference to the accompanying drawings and specific cutting examples. The main steps are as follows:

[0008] Discriminative feature extraction based on parallel interactive fusion convolutional network The 101 interactive fusion convolution includes three convolutional layers, such as Figure 1 As shown in the figure. The first two convolutional layers have convolution kernels of different sizes and are distributed in parallel. The original cutting signal data is simultaneously input to the first two convolutional layers. The features obtained by the convolution operation are then passed to the Gaussian Error Linear Unit (GELU) activation function. Due to the introduction of the idea of ​​random regularization in GELU, it can alleviate the gradient vanishing problem and accelerate the convergence of the model. After performing the random drop operation on the features obtained from the two branches, they are multiplied mutually to complete the interaction of features of different scales, as shown below: Where f represents the GELU activation function, ⊙ represents the dot product operation, x represents the multi-channel cutting signal data, Conv1 represents the first convolutional layer module, Conv2 represents the second convolutional layer module, O1 represents the features obtained under the first branch, and O2 represents the features obtained under the second branch; 102 adds the outputs O1 and O2 obtained from the two branches and uses them as the input of the third convolutional layer to further enhance the features, as shown below: F=Conv3(O1+O2) (11) Among them, Conv3 represents the third convolutional layer module, and F represents the added features; 103 Features of different scales can be considered as expansions of features in the width direction, and do not capture features at different levels in the depth direction. Therefore, a parallel interactive fusion convolutional structure is designed. One branch contains one interactive fusion convolutional block, and the other branch contains two interactive fusion convolutional blocks. These two branches have convolution kernels of different sizes, thus ensuring feature richness in two dimensions.

[0009] A plug-and-play, parameter-free discriminative feature expansion and multi-source domain data enrichment 201 is limited by factors such as processing time and processing cost. It is difficult to collect enough diverse source domain data, which contains data distribution characteristics under unknown working conditions. Therefore, a hybrid feature enhancement strategy is adopted to improve the generalization ability of the feature extractor, such as Figure 2 As shown; 202 First, the mean and standard deviation statistics of the feature map obtained by the feature extractor are calculated as follows: Among them, H can be regarded as the total number of channels of the cutting signal, W represents the length of the cutting signal in each sample, f represents the feature map obtained by the feature extractor, and f h,w Represents the value at the h and w positions in the feature map, μ(f) represents the mean of the feature map, and σ(f) represents the variance of the feature map; 203 Then, the weights of random sampling in the Beta distribution are used to obtain the mixed values ​​of the feature statistics in a linear weighted manner, as shown below: Among them, λ represents the linear weight value, represents a randomly shuffled feature map, represents the variance of the feature map after random shuffling, Represents the mean of the feature map after random shuffling, δ mix represents the variance after linear weighting, η mix represents the linearly weighted mean; 204Finally, according to the instance normalization calculation process, the enhanced feature output is obtained: Among them, f aug represents the enhanced feature map, and ⊙ represents the dot product operation. The feature extractor proposed in 205 contains two branches, which obtain feature maps with different depth levels. Feature enhancement operations are performed on the feature maps obtained by each branch, thereby completing dual-level feature enhancement.

[0010] Enhancement of key discriminative features by integrating Convformer modules The 301Convformer module combines the Transformer's ability to model global long-range dependencies with the convolutional neural network's ability to learn local features. The Convformer module is used to further process the combined enhanced features f aug ;302 The multi-head attention mechanism is the core component of the Convformer module. It uses a learnable linear mapping to convert the input feature map into a query vector matrix Q, a key vector matrix K and a value vector matrix V, and then obtains the attention weight Z of each attention head through the scaled dot product operation and the Softmax function. j ; 303 When calculating the key matrix K and the value matrix V, in order to alleviate the computational burden of the original multi-head attention, a convolutional neural network is used to downsample the input feature map and effectively learn the local features within the feature map; 304 On the other hand, the forward feedback mapping in the transformer is replaced by the convolutional mapping, which further reduces the number of model parameters. Then, the output f of the Convformer module is obtained through the normalization layer and residual connection. d .

[0011] Domain-invariant feature extraction driven by dynamic label smoothing domain discriminator 401 The data feature distribution of each source domain is different. In order to effectively transfer the knowledge from multiple source domain data to data under unknown working conditions, the core idea of ​​domain generalization is to learn domain-independent discriminative features, thereby ensuring the performance of the established domain generalization model on unknown working condition data; 402 Combining the advantages of domain adversarial training mode and label smoothing, the present invention integrates dynamic label smoothing to propose a new domain discriminator for learning domain invariant features, such as Figure 3 As shown; 403 First, the discriminative features obtained by the feature extractor are converted into a probability vector related to the number of source domains through a fully connected layer, and then further processed using a normalization layer and a maximum function; The output of the 404 normalization layer is multiplied by the bitwise true domain label, the output of the maximum function is multiplied by the true domain label, and then the two parts are added. To optimize the dynamic label smoothing domain adversarial, the predicted probability vector information is used to adjust the hard label to the form of a soft label, as shown below: Among them, onehot() represents the one-hot encoding function, y t represents the original hard tag, represents the adjustment coefficient, f p represents the predicted probability vector, represents the adjusted soft label, and ~ represents a bitwise operation. 405 Since the dynamic label smoothing domain discriminator contains three fully connected layers, it needs to be trained. Considering that the original hard labels are adjusted to soft labels of probability distribution, the KL divergence loss is used to evaluate the distribution between the output of the domain discriminator improved by dynamic label smoothing and the adjusted target domain labels, as shown below: Where N represents the number of samples in a mini-batch, K represents the number of source domain datasets used for model training, represents the soft label after adjustment of the k-th source domain dataset and the t-th sample, and represents the probability vector of the k-th source domain dataset and the t-th sample output by the dynamic label smoothing domain discriminator.

[0012] Use supervised contrastive learning loss to constrain class representation information in model modeling 501There are differences in the category representation of tool wear labels under different machining conditions. If the inconsistency of category representation is not considered, the generalization ability of the model will be limited; 502 Supervised Contrastive Learning Loss can explicitly separate samples of different categories and aggregate samples of the same category in the representation space, deeply considering the differences in category representations of tool wear states, thereby improving the generalization ability of the model; The 503 supervised contrastive learning loss is calculated based on positive and negative sample pairs. The principle is to maximize the similarity of samples within a positive pair and the difference of samples within a negative pair. Within a small batch of samples, discriminative features of the same category can be regarded as positive sample pairs, while discriminative features of other categories are negative sample pairs. 504 achieves the goal of aggregating samples of the same category and separating samples of different categories as much as possible through supervised contrastive learning loss, thereby enhancing the separation of classification boundaries and the generalization ability of the model, as shown below: in, Represents the mini-batch sample and label y i the same quantity; When y i and y j If they are of the same category, it is 1, otherwise it is 0; f i d , and They represent the discriminative features output by the i-th, j-th, and k-th Convformer modules in a small batch of samples, respectively; exp() represents the exponential function; sim represents the cosine similarity function; and τ represents the temperature coefficient, which is used to control the influence of negative samples on the training process.

[0013] The parameters that need to be optimized for the tool wear monitoring domain generalization method under unknown working conditions proposed in the present invention include the parameters of the feature extractor, the dynamic label smoothing domain discriminator and the classifier; The 602 classifier is generally composed of multiple fully connected layers, which is used to reduce the high-dimensional discriminative features to the same dimension as the number of tool wear categories, and then obtain the probability output of the tool wear state corresponding to the discriminative features. In order to optimize the feature extractor and classifier parameters, the cross entropy loss function is generally used to calculate the classification loss of the classifier; 603 Since the supervised contrast loss compensates for the disadvantages of the cross entropy loss function, when designing the overall loss function, the contribution of the dynamic label smoothing domain discriminator and the supervised contrast learning loss to the training process can be adjusted through the weight coefficient. Combined with the adversarial loss of the domain discriminator, the overall loss function can be expressed as: L total=α·L sup +(1-α)·L c -β·L adv (20) Among them, α and β represent weight coefficients, L sup Denotes the supervised contrastive learning loss value, L c represents the classification loss of the classifier, L adv represents the loss value of the dynamic label smoothing domain discriminator during domain adversarial training, L total Indicates the total loss value; 604 The model parameters are then optimized through the error back propagation mechanism as follows: Among them, θ F ,θ d and θ c Represent the parameters of feature extractor, domain discriminator and classifier respectively, and η represents the learning rate. After the model training is completed, the performance of the model is verified using data under unknown working conditions (target domain data). The overall framework is as follows Figure 4 shown. Example 1

[0014] This section mainly conducts milling tests based on the above description. The tool wear experiment is carried out on the MC-510V CNC vertical machining center. The diameter of the milling cutter is 70mm, and the types of materials processed are cast iron and stainless steel. In terms of cutting parameters, the cutting speed is fixed at 200m / min, the milling depth is set to 0.75mm and 1.5mm, and the feed rate is set to 0.25mm / r and 0.5mm / r. By combining different process parameters, 8 different processing conditions can be obtained. Each condition is repeated twice, and the total number of tests is 16 times. During the processing, vibration sensors, acoustic emission sensors and current sensors are used to collect cutting signals. An acceleration sensor and an acoustic emission sensor are installed on the workbench and the spindle respectively. A current sensor is installed on the spindle motor to collect AC and DC signals. The sampling frequency is 250Hz, as shown in the figure. Figure 5 As shown in Table 1, the 16 tool wear data sets have different cutting times, and the tool wear values ​​in some groups are missing. In view of the innovation involved in the present invention and the requirements for tool wear, four tool wear data sets with significantly different cutting conditions, Case 1, Case 2, Case 11, and Case 13, are selected, as shown in Table 1: Table 1 Process parameters of NASA tool wear dataset

[0015] Because unstable cutting signals are generated when the tool cuts in and out, cutting signals during the stable cutting phase are selected as the model input. For the NASA dataset, each cutting signal contains 9,000 sampling points, with 5,000 sampling points retained during the intermediate phase. A sliding window operation with a window length of 250 is then used to divide each cutting signal data set into multiple segments, with no overlap between the segments, thereby expanding the amount of tool wear data.

[0016] The NASA dataset contains four different working conditions. To fully verify the superiority of the proposed domain generalization model, according to the cross-validation criterion, the tool wear data under each working condition is used as a test set once, and the tool wear data under the remaining working conditions are used for modeling. It is particularly emphasized that no data in the target domain is accessed during the modeling process. The detailed information is shown in Table 2. Table 2 Milling tool wear monitoring tasks based on cross-validation criteria Task Training set (source domain data) Test set (target domain data) A1 Case 2 / Case 11 / Case 13 Case 1 A2 Case 1 / Case 11 / Case 13 Case 2 A3 Case 1 / Case 2 / Case 13 Case 11 A4 Case 1 / Case 2 / Case 11 Case 13

[0017] Tool wear monitoring results such as Figure 6 As shown in the figure, the tool wear monitoring accuracy rates for tasks A1, A2, A3, and A4 are 81.47%, 86.43%, 89.13%, and 88%, respectively. The accuracy rates of all tool wear monitoring tasks exceed 81%, which can meet the requirements of industrial applications. This has important guiding significance for machining center operators in the manufacturing field to replace cutting tools in a timely manner. Example 2

[0018] The NASA dataset is milling tool wear data, and a turning tool wear dataset was constructed in the laboratory. For turning experiments, cutting speed is the factor that most significantly affects tool wear, so only the cutting speed was changed to construct a variable-condition turning tool wear dataset. The experiment was completed on a Cincinnati HAWK TC-150 CNC precision turning center. The tool used was a carbide insert, which was fixed to the tool arbor by screws. The tool arbor was then fixed to a Kistler 9121 dynamometer, and the turning forces in three directions were collected at a sampling rate of 1000 Hz, such as Figure 7 The material being cut was difficult-to-machine 38CrSi alloy structural steel, resulting in severe friction in the cutting zone and severe tool wear. Based on the material properties and the cutting parameters recommended by the insert manufacturer, four different cutting speeds of 100 m / min, 120 m / min, 140 m / min, and 160 m / min were used, with a feed rate of 0.2 mm / r and a depth of cut of 0.4 mm, as shown in Table 3. Table 3 Turning tool wear test parameters test Cutting speed (m / min) Feed rate (mm / r) Cutting depth (mm) T1 100 0.2 0.4 T2 120 0.2 0.4 T3 140 0.2 0.4 T4 160 0.2 0.4

[0019] The cutting force data obtained from turning operations were processed in the same manner as the NASA dataset. Similarly, according to the cross-validation criterion, the tool wear data under each working condition was used as a test set once, and the tool wear data under the remaining working conditions were used for modeling, as shown in Table 4. Table 4 Turning tool wear monitoring tasks based on cross-validation criteria Task Training set (source domain data) Test set (target domain data) B1 T2 / T3 / T4 T1(100m / min) B2 T1 / T3 / T4 T2 (120m / min) B3 T1 / T2 / T4 T3 (140m / min) B4 T1 / T2 / T3 T4 (160m / min)

[0020] The confusion matrix is ​​used to characterize the tool wear state monitoring accuracy. The turning tool wear monitoring results are as follows: Figure 8 According to the monitoring results, the monitoring accuracy of all tasks exceeded 88%, and the tool wear state recognition accuracy of the four monitoring tasks was 91.19%, 88.60%, 91.84% and 93.94% respectively, achieving satisfactory results. The monitoring accuracy shows that it can be used in actual industrial applications.

[0021] The above-described embodiments are only preferred embodiments of the present invention and are not exhaustive of all feasible implementations of the present invention. For those skilled in the art, any obvious modifications made thereto without departing from the principles and spirit of the present invention should be considered to be included within the scope of protection of the claims of the present invention.

Claims

1. A generalized method for monitoring tool wear status in unknown working conditions, characterized by Here are the steps: ①Discriminative feature extraction based on parallel interactive fusion convolutional networks; The 101 interactive fusion convolutional network consists of three convolutional layers. The first two convolutional layers have convolution kernels of different sizes and are distributed in parallel. The original cutting signal data is input to the first two convolutional layers at the same time. The features obtained by the convolution operation are passed to the Gaussian error linear unit activation function. After performing the random drop operation on the obtained features, they are interactively multiplied to complete the interaction of features of different scales, as shown in the following formula: Where f represents the Gaussian error linear unit activation function, ⊙ represents the dot product operation, x represents the multi-channel cutting signal data, Conv1 represents the first convolutional layer module, Conv2 represents the second convolutional layer module, O1 represents the features obtained under the first branch, and O2 represents the features obtained under the second branch; 102 Add the outputs O1 and O2 obtained from the two branches and use them as the input of the third convolutional layer, as shown in the following formula: F=Conv3(O1+O2) (2) Among them, Conv3 represents the third convolutional layer module, and F represents the added features; 103 Design a parallel interactive fusion convolution structure, one branch contains one interactive fusion convolution block, and the other branch contains two interactive fusion convolution blocks; ② Using a plug-and-play, parameter-free approach to discriminative feature expansion and multi-source domain data enrichment; 201 A hybrid feature enhancement strategy is used to improve the generalization ability of the feature extractor. The mean and standard deviation statistics of the feature map obtained by the feature extractor are calculated as follows: Where H is the total number of channels of the cutting signal, W is the length of the cutting signal in each sample, f is the feature map obtained by the feature extractor, and f h,w Represents the value at the h and w positions in the feature map, μ(f) represents the mean of the feature map, and σ(f) represents the variance of the feature map; 202 The mixed value of the characteristic statistics is obtained by using the weight of random sampling in the Beta distribution in a linear weighted manner, as shown in the following formula: Among them, λ represents the linear weight value, represents a randomly shuffled feature map, represents the variance of the feature map after random shuffling, Represents the mean of the feature map after random shuffling, δ mix represents the variance after linear weighting, η mix represents the linearly weighted mean; 203 According to the example standardization calculation process, the enhanced feature output is: Among them, f aug represents the enhanced feature map, ⊙ represents the dot product operation; The feature extractor proposed in 204 contains two branches, which obtain feature maps with different depth levels, and perform feature enhancement operations on the feature maps obtained by each branch respectively; ③ Enhancement of key discriminative features by integrating Convformer modules; ④Domain-invariant feature extraction driven by dynamic label smoothing domain discriminator; 401 The discriminative features obtained by the feature extractor are converted into a probability vector related to the number of source domains through a fully connected layer, and further processed using a normalization layer and a maximum function; The output of the 402 normalization layer is multiplied by the bitwise true domain label, the output of the maximum function is multiplied by the true domain label, and the multiplication results are added together. The predicted probability vector information is used to adjust the hard label to the form of a soft label, as shown in the following formula: Among them, onehot() represents the one-hot encoding function, y t represents the original hard label, represents the adjustment coefficient, f p represents the predicted probability vector, represents the adjusted soft label, and ~ represents bitwise operation; 403 KL divergence loss is used to evaluate the distribution between the output of the domain discriminator after dynamic label smoothing and the adjusted target domain label, as follows: Among them, N represents the number of samples in a small batch, K represents the number of source domain datasets used for model training, represents the k-th source domain dataset and the adjusted soft label of the t-th sample, Represents the probability vector of the t-th sample of the k-th source domain dataset output by the dynamic label smoothing domain discriminator; ⑤Use supervised contrastive learning loss to constrain class representation information in model modeling; ⑥ A dual-level feature-enhanced domain generalization network that integrates dynamic label smoothing and contrastive learning through multi-source domain data training; 601 optimizes the parameters of the feature extractor, dynamic label smoothing domain discriminator, and classifier. The classifier consists of multiple layers of fully connected layers, and the cross entropy loss function is used to calculate the classification loss of the classifier. When designing the total loss function, the contribution of the dynamic label smoothing domain discriminator and the supervised contrastive learning loss to the training process is adjusted by the weight coefficient, and the adversarial loss of the joint domain discriminator is used. The total loss function is: L total =α·L sup +(1-a)·L c -β·L adv (10) Among them, α and β represent weight coefficients, L sup Denotes the supervised contrastive learning loss value, L c represents the classification loss of the classifier, L adv represents the loss value of the dynamic label smoothing domain discriminator during domain adversarial training, L total Indicates the total loss value; 603 optimizes the model parameters through the error back propagation mechanism, as shown below: Among them, θ F ,θ d and θ c Denote the parameters of the feature extractor, domain discriminator, and classifier, respectively, and η denotes the learning rate. After the model training is completed, the performance of the model is verified using data under unknown working conditions (target domain data).

2. The domain generalization method for monitoring tool wear status under unknown working conditions according to claim 1 is characterized in that: The specific scheme in step ③ is as follows: 301 Use the Convformer module to further process the combined enhanced features f aug , a learnable linear mapping is used to convert the input feature map into a query vector matrix Q, a key vector matrix K and a value vector matrix V, and the attention weight Z of each attention head is obtained by scaling the dot product operation and the Softmax function j ; 302 uses convolutional neural network to downsample the input feature map to effectively learn the local features in the feature map; 303 uses convolution mapping to replace the forward feedback mapping in transformer to further reduce the number of model parameters, and obtains the output f of Convformer module through normalization layer and residual connection. d .

3. The domain generalization method for monitoring tool wear status under unknown working conditions according to claim 1 is characterized in that: The specific solution in step ⑤ is as follows: using supervised contrastive learning loss to aggregate samples of the same category, thereby enhancing the separation of classification boundaries and the generalization ability of the model, as shown in the following formula: in, Represents the mini-batch sample and label y i the same quantity; When y i and y j If they are of the same category, it is 1, otherwise it is 0; f i d , and They represent the discriminative features output by the i-th, j-th, and k-th Convformer modules in a small batch of samples, respectively; exp() represents the exponential function; sim represents the cosine similarity function; and τ represents the temperature coefficient, which is used to control the influence of negative samples on the training process.

Images