Cutter wear loss prediction method based on physical information self-supervised learning under condition of few marks

Abstract

The invention discloses a tool wear loss prediction method based on physical information self-supervised learning under the condition of few marks, and belongs to the technical field of numerical control machining center tool wear state prediction. Comprising the steps of collecting monitoring signals such as cutting force, vibration and acoustic emission in the machining process; carrying out preprocessing operation of interception and Z score standardization on each sample; a mixed residual convolutional neural network tool wear monitoring model is designed, and information fusion and high-dimensional representation extraction are carried out on multi-source sensor signals; pre-training the tool wear monitoring model by using the auxiliary task I, and guiding the model to extract the invariance feature of the signal; according to the method, under the condition of scarce marked data, high-precision prediction and excellent generalization ability of tool wear are realized, the fluctuation amplitude of prediction errors can be effectively reduced, and the stability of a prediction result is kept. According to the method, a small amount of mark data is effectively utilized in model training, the stability and reliability of tool wear monitoring are ensured, and the requirements for monitoring precision and adaptability in actual production are met.

Description

Technical Field

[0001] The invention belongs to the technical field of tool wear state prediction of numerical control machining centers, and specifically relates to a tool wear amount prediction method based on physical information self-supervised learning under few labels. Background Art

[0002] Manufacturing plays a vital role in the modern economy. It is not only a key engine for economic growth, but also an important pillar for innovation, employment and social development. With the close integration of manufacturing and the new generation of information technology, intelligent manufacturing is becoming a common and key trend, triggering a profound change in manufacturing methods and becoming one of the core driving forces for the deep integration of informatization and industrialization. Among them, the state monitoring of key components in the metal cutting process is a key construction direction for promoting the intelligentization of the manufacturing process. As a cutting actuator that directly acts on the workpiece, the tool will be subjected to the combined action of different load factors when removing the workpiece material, which will inevitably cause tool wear and tool degradation, accompanied by the increase of cutting force, the increase of temperature in the processing area, and the increase of vibration. This will have a negative impact on the dimensional accuracy and surface quality of the product, increase the risk of parts scrapping, increase additional production costs, and even cause unexpected downtime of the machine tool, resulting in unnecessary economic losses. Therefore, timely and effective monitoring of tool wear status is crucial to ensure the reliability of product quality, improve production efficiency and reduce safety risks, and provide important support for the development of intelligent manufacturing.

[0003] The invention patent with the authorization number CN112757053B discloses a tool wear monitoring method and system based on the model fusion of power and vibration signals, including: S1: signal acquisition and screening, synchronous sampling of the power signal and vibration signal of the tool spindle, and synchronous sampling of the PLC of the machine tool to obtain the real-time processing information of the machine tool, and the power signal and the vibration signal are screened by using the processing information to obtain the power signal and the vibration signal when the tool to be monitored is processing; S2: signal denoising, respectively setting thresholds for the power signal and the vibration signal of the processing section signal, and using the threshold filtering to remove the high-frequency area signal for noise reduction; S3: feature calculation, calculating the features of the power signal and the vibration signal after denoising to obtain the original feature data set; S4: feature dimensionality reduction, using the random forest algorithm to calculate the importance scores of all features in the original feature data set, and the obtained The scores are arranged in descending order, and the elimination ratio is determined. According to the elimination ratio, some of the features are eliminated to obtain a new feature set; S5: data partitioning, the data of the new feature set is divided into three parts, which are used as a training set, a validation set and a test set respectively; S6: model training, the data of the training set is input into the Catboost model and the extreme random tree model, and the parameters are adjusted to respectively train the Catboost model and the tool wear monitoring model corresponding to the extreme random tree model, and the Stacking integration algorithm is used to perform fusion learning on the Catboost model and the tool wear monitoring model corresponding to the extreme random tree model; S7: model prediction, according to the fused tool wear monitoring model, the tool wear is predicted for the features of the test set; S8: synchronous update, the prediction result of the tool wear is subjected to Kalman filtering denoising to obtain the tool wear prediction value.

[0004] The invention patent with the authorization number CN111832432B discloses a real-time prediction method for tool wear based on wavelet packet decomposition and deep learning, including: (1) synchronously collecting relevant sensor signals during the workpiece processing process, and selecting stable signal segments as signal segments to be analyzed, and expanding the signal samples to be analyzed to increase the sample size; performing wavelet packet decomposition transformation on the signals to be analyzed to obtain multiple two-dimensional matrices of wavelet packet coefficients; (2) each two-dimensional matrix of wavelet packet coefficients is used as the input of a feature extraction CNN model block, and the one-dimensional feature matrix output by each feature extraction CNN model block is spliced ​​into a longer one-dimensional matrix, and then feature fusion is performed and a two-layer fully connected network is established to obtain a convolutional neural network model; (3) the signal data to be analyzed is input into the convolutional neural network model to predict the tool wear in real time.

[0005] Traditional methods used in tool remaining life prediction generally require feature extraction, feature selection, and feature dimensionality reduction before building a model. These methods all rely on certain experience and theories, and when data preprocessing is not appropriate, important feature information will be lost, resulting in a reduction in the accuracy of tool remaining life prediction. Summary of the invention

[0006] The purpose of the present invention is to provide a tool wear prediction method based on self-supervised learning of physical information under few labels, taking cutting force, vibration and acoustic emission signals as monitoring signals, and combining a hybrid residual convolutional neural network with a fully connected regression layer to achieve accurate prediction of tool wear under a two-stage self-supervised learning framework, thereby overcoming the shortcomings of the prior art.

[0007] The technical solution adopted by the present invention is:

[0008] The tool wear prediction method based on self-supervised learning of physical information under few labels includes the following steps:

[0009] S1. Collect monitoring signals such as cutting force, vibration and acoustic emission during the processing;

[0010] S2. The collected signals are used as data samples, some samples are marked, and each sample is preprocessed by interception and Z-score standardization;

[0011] S3. Design a hybrid residual convolutional neural network tool wear monitoring model to perform information fusion and high-dimensional representation extraction on multi-source sensor signals;

[0012] S4. Based on the self-supervised learning framework, a pair of auxiliary tasks are used to pre-train the tool wear monitoring model. By enhancing view generation and twin network contrast loss calculation, the model is guided to extract the invariant features of the signal;

[0013] S5. Based on the self-supervised learning framework, the tool wear monitoring model is pre-trained using auxiliary task 2, and the physical consistency of the representation vector is enhanced by introducing physical constraints;

[0014] S6. Based on the self-supervised learning framework, by alternately optimizing auxiliary task 1 and auxiliary task 2, the model is gradually guided to have both efficient feature extraction capabilities and consistency with physical laws;

[0015] S7. Based on the self-supervised learning framework, a limited number of labeled samples are used to fine-tune the pre-trained tool wear monitoring model and obtain the final wear value evaluation results;

[0016] S8. Evaluate the overall performance of the model in the tool wear condition monitoring task by calculating the model evaluation index.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] When it is difficult to obtain tool wear value marks, the present invention makes full use of a large number of unmarked tool cutting process monitoring signals to improve the prediction of tool wear values ​​under wear mark restricted conditions, thereby making use of the excellent nonlinear modeling capabilities of deep neural networks and reducing the dependence of deep neural networks on wear marks.

[0019] The present invention excludes deep network model parameters that do not conform to physical laws by introducing regularization constraints based on physical information in the self-supervised learning framework, effectively limits the range of feasible solutions in the model parameter space, improves the generalization ability of pre-trained features, and maintains the model generalization ability by freezing some parameters in the fine-tuning process.

[0020] The present invention achieves high-precision prediction of tool wear and excellent generalization ability when labeled data is scarce, and can effectively reduce the fluctuation amplitude of prediction error and maintain the stability of prediction results.

[0021] The present invention effectively utilizes a small amount of labeled data in model training, ensures the stability and reliability of tool wear monitoring, and meets the requirements for monitoring accuracy and adaptability in actual production.

[0022] The present invention is of great significance to the monitoring of tool wear status and can provide highly adaptable and extensible prediction capabilities in practical applications.

[0023] The model training of the present invention is divided into two stages: the first stage is to pre-train the hybrid residual convolutional neural network through the representation learning of unlabeled data to optimize the signal feature extraction capability; the second stage is to fine-tune the model using limited labeled data and achieve high-precision prediction of tool wear value through mean square error loss. Experimental results show that this method performs well in wear monitoring and remaining life prediction and has significant industrial application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flow chart of the present invention;

[0025] Figure 2 It is a schematic diagram of the network architecture of the tool wear monitoring model of the present invention;

[0026] Figure 3 This is a network structure diagram of auxiliary task 1 of the present invention;

[0027] Figure 4 This is the network structure diagram of auxiliary task 2 of the present invention;

[0028] Figure 5 It is the tool wear monitoring model framework and fine-tuning strategy diagram of the present invention;

[0029] Figure 6 is a schematic diagram of data preprocessing of a milling data set involved in an example of the present invention;

[0030] Figure 7 is a schematic diagram of Pearson correlation analysis of tool wear monitoring involved in an example of the present invention;

[0031] Figure 8 is a tool wear prediction diagram of labeled samples under different degrees of scarcity involved in an example of the present invention;

[0032] Fig. 9 It is a schematic diagram of the change of evaluation indicators of six tool wear monitoring methods (ResNet, LSTM, XGBoost, SVR, KNN and PISSL) as the percentage of training data increases. DETAILED DESCRIPTION

[0033] In order to better understand the purpose, structure and function of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings.

[0034] Deep learning can use massive amounts of unprocessed data to autonomously learn features through deep networks. With the widespread application of tool condition monitoring technology, the tool condition monitoring data obtained during the machining process has increased significantly. In the context of massive tool condition monitoring data, applying deep learning methods to tool remaining life prediction helps overcome the reliance on specific prior knowledge and expertise, and has more obvious advantages than traditional methods.

[0035] like Figure 1 As shown, the present invention provides a tool wear prediction method based on self-supervised learning of physical information under few labels, the method comprising:

[0036] Step S1: collecting monitoring signals such as cutting force, vibration and acoustic emission during the machining process;

[0037] Step S2: taking the collected signals as data samples, marking some samples, and performing preprocessing operations of truncation and Z-score standardization on each sample;

[0038] The step S2 specifically includes:

[0039] In the experiment, each tool performed 315 cutting cycles with a cutting length of 108 mm, and the sensor data collected in each cutting cycle was regarded as a data sample. The experiment generated a total of 1,890 data samples. In order to obtain the true wear state of the samples, a small number of tools were randomly selected to mark the tool wear during their life cycle, and a data set consisting of a small number of labeled samples and a large number of unlabeled samples was constructed. Among them, each sample consists of 7-channel sensor signals at a certain time step, and the channels include three-dimensional cutting force signals, three-dimensional vibration signals, and acoustic emission signals. Specifically, after the tool completes a cutting cycle, the flank wear value of each cutting edge is measured as the state label of the sample. Accordingly, the data set is divided into labeled samples (including C1, C4 and C6 tools) and unlabeled samples (including C2, C3 and C5 tools);

[0040] Each sample in the data set is intercepted along the time step dimension. In the process of intercepting the sample along the time step dimension, 5% of the sample length is discarded at the beginning and end of the sample to exclude the cutting-in and cutting-out stages. Then, considering that the marking work is carried out after the cutting cycle is completed, the 80,000 sample points at the end of the remaining data are retained as the initial sample to maximize the matching of the wear mark.

[0041] The sample data is z-score standardized, and the expression is:

[0042]

[0043] Where x represents the value of a sampling point on a sensor in the sample;

[0044] μ represents the mean of all sampling points of sensor data along the path x;

[0045] σ represents the standard deviation of all sampling points of the sensor data x.

[0046] Step S3: Design a hybrid residual convolutional neural network tool wear monitoring model to perform information fusion and high-dimensional characterization extraction on multi-source sensor signals (cutting force, vibration and acoustic emission are fused and feature extracted through the designed network model);

[0047] The hybrid residual convolutional neural network tool wear monitoring model includes a hybrid residual convolutional neural network and a fully connected regression layer, and the tool wear monitoring value is output through the fully connected regression layer;

[0048] The hybrid residual convolutional neural network tool wear monitoring model combines shallow one-dimensional convolution, deep two-dimensional convolution and residual connection, aiming to convert the raw data collected by multi-source sensors into efficient representation signals. The model input is a multi-channel one-dimensional time series signal, including cutting force, vibration and acoustic emission signals. First, the input data passes through multiple one-dimensional convolution modules, which are composed of convolution layers, batch normalization layers and linear rectification activation functions to independently extract the shallow features of each sensor signal and increase the number of channels. Subsequently, these shallow features are fused through a two-dimensional convolution module. The two-dimensional convolution module is composed of multiple residual basic blocks, each of which contains multiple convolution layers and their residual connections to extract deep features of multi-source signals. After passing through the deep residual network, the output features are reduced in dimension through an adaptive average pooling layer and converted to one-dimensional vectors through a flattening layer. Subsequently, these one-dimensional vectors are input into a fully connected regression layer. The fully connected layer consists of a two-layer perceptron (MLP). The first layer maps the 512-dimensional input to 128 dimensions, and the second layer maps the 128-dimensional representation to a single tool wear value. The activation function uses LeakyReLU and Softplus. The tool wear value is predicted through the above process. Taking single-channel signal input as an example, the overall architecture parameter settings of the hybrid residual convolutional neural network are shown in Table 1. The network architecture diagram of the hybrid residual convolutional neural network is shown in Figure 2 shown.

[0049] Table 1 Overall architecture parameter settings of hybrid residual convolutional neural network

[0050]

[0051]

[0052] Step S4: Based on the self-supervised learning framework, the tool wear monitoring model is pre-trained using auxiliary task one. By means of enhanced view generation and twin network contrast loss calculation, the model is guided to extract the invariance features of the signal. The network structure diagram of auxiliary task one is shown in Figure 3 As shown;

[0053] The step S4 specifically includes:

[0054] Auxiliary task 1 is a data-driven pre-auxiliary task based on a twin network. The twin network consists of two weight-sharing hybrid residual convolutional neural networks to construct two branches of the twin network.

[0055] Auxiliary task 1 generates two different views x1 and x2 by randomly combining different data enhancement methods including random cropping, inversion, amplitude inversion, noise addition and filtering operations, and inputs them into the two branches of the twin network to obtain the potential tool wear representations z1 and z2 under different views; then, the predictor is introduced in the upper branch of the twin network to generate the prediction vector p1, and the cosine similarity is compared with the representation z2 of the lower branch to guide the hybrid residual convolutional neural network to learn the invariant representation. The cosine similarity is used to measure the similarity between them, which is defined as follows:

[0056]

[0057] In the formula, ‖·‖2 represents Norm.

[0058] Using the stop gradient operation in the lower branch, the augmented views are swapped to calculate the loss function in a symmetric way:

[0059]

[0060] Where sg(·) represents the gradient stop operation on the input variable.

[0061] Step S5: Based on the self-supervised learning framework, the tool wear monitoring model is pre-trained using auxiliary task 2. By introducing physical constraints, the physical consistency of the representation vector is enhanced. The network structure diagram of auxiliary task 2 is shown in the figure. Figure 4 As shown;

[0062] Auxiliary task 2 is a regularized constraint auxiliary task based on physical properties.

[0063] Auxiliary task 2 combines the idea of ​​triplet loss and proposes two distance measurement methods that integrate physical knowledge, which together constitute the regularized constraint loss function. The triplet loss is calculated by degrading the representation of three different samples; in this sample triplet, the reference input is usually defined as the anchor sample x A , corresponding to the cutting process monitoring data at time t0, represented by symbol A; similar inputs close to the reference input are defined as positive samples x P , corresponding to time t1, represented by the symbol P; the input that is significantly different from the reference input is defined as the negative sample x N , corresponding to time t2, represented by the symbol N. For a given triple sample {x A ,x P ,x N}, the embedded representation obtained through the backbone network is represented by {z A ,z P ,z N} indicates that, The first of the two distance measurement methods constrains the Euclidean distance of the representation in the embedded space and requires that the anchor sample x A With positive sample x P The Euclidean distance in the embedding space is less than the negative sample x N . From this we can get the restriction condition:

[0064]

[0065] where δ is the margin of the mandatory condition between sample pairs, express The square of the norm. For the i-th group of triplet samples, the loss function of the Euclidean distance constraint can be obtained as:

[0066]

[0067] The second distance measurement method of the two distance measurement methods constrains the development trend of the representation and requires that the vector With vector The direction is as consistent as possible. Using cosine similarity S c As a measure of the vector angle representing the trend requirement, for the i-th group of triple samples, the vector angle constraint loss function can be obtained as:

[0068]

[0069] This results in a loss function constraining the Euclidean distance The loss function of the angle between the constraint representation vector The two are combined by weighted summation to form the loss function of task 2 to represent the metric constraints for:

[0070]

[0071] Among them, λ1 and λ2 are weighting coefficients.

[0072] Step S6: Based on the self-supervised learning framework, by alternately optimizing auxiliary task 1 and auxiliary task 2, the model is gradually guided to have both efficient feature extraction capabilities and consistency with physical laws;

[0073] The step S6 specifically includes:

[0074] In the pre-training phase, all collected samples are used to complete the self-supervised auxiliary task. The twin encoder part uses the cosine decay schedule dynamic strategy to update the network parameters. The learning rate gradually decreases to zero with the number of training rounds, taking into account the convergence speed in the early stage of the training process and the optimization stability in the later stage. The training of the predictor is carried out in a fixed learning rate manner. In the loss function that represents the metric constraint, the weights and are set to 0.001 and 1.

[0075] The specific settings of the pre-training parameters are shown in Table 2.

[0076] Table 2 Pre-training parameter settings

[0077]

[0078]

[0079] Step S7: Based on the self-supervised learning framework, a limited number of labeled samples are used to fine-tune the pre-trained tool wear monitoring model and obtain the final wear value evaluation result; Figure 5 shown.

[0080] The step S7 specifically includes:

[0081] Freeze the shallow convolutional network parameters of the pre-trained model and only adjust the deep residual module;

[0082] The optimization process of the deep residual module and the fully connected regression layer is to first obtain the model error by calculating the loss function, then use back propagation to calculate the gradient of the loss function with respect to the model parameters, and optimize the parameters of the deep residual module and the fully connected regression layer by the gradient descent method, and finally complete the model training with the goal of minimizing the mean square error (MSE). The optimization goal can be expressed as:

[0083]

[0084] Where n is the labeled dataset D L The number of samples in .

[0085] Finally, the model obtains data representation through a fine-tuned hybrid residual convolutional neural network, and after average pooling and flattening layers, it outputs the predicted wear value through an optimized fully connected regression layer;

[0086] In the target task of tool wear value monitoring, the 5-fold cross-validation method was used for repeated experiments. The labeled samples were randomly divided into 5 equal parts. In each experiment, one of them was selected as the test set and the other 4 were selected as the training set in turn. In order to verify the accuracy of the proposed method in evaluating tool wear value under limited labeling conditions, only part of the data in the training set was taken as input samples to complete the training of the tool wear value monitoring model. In order to more comprehensively compare the different degrees of label scarcity conditions, three different degrees of limited labeling scenarios were set up in the experiment, which were 10%, 20%, 30%, 50% and 100% of the total number of training set samples. The tool wear monitoring model uses a multi-layer perceptron network as the fully connected layer after the backbone network. The non-negativity of the output is guaranteed by the Softplus activation function, and the final wear value evaluation result is obtained. The optimizer parameters are the same as the pre-training. During the training process, the shallow convolution parameters are frozen, the deep residual block parameters are fine-tuned, and the weights and biases of the neurons in the fully connected layer are updated and optimized. The specific parameter settings are shown in Table 3.

[0087] Table 3 Training parameter settings for target tasks

[0088]

[0089] Step S8: Evaluate the overall performance of the model in the tool wear state monitoring task by calculating the model evaluation index.

[0090] The step S8 specifically includes:

[0091] The model evaluation indicators are root mean square error, mean absolute error and mean absolute percentage error, and the calculation formula is:

[0092]

[0093]

[0094] Among them, RMSE is the root mean square error, MAE is the mean absolute error, MAPE is the mean absolute percentage error, and y i , They represent the actual remaining life and predicted remaining life after the i-th cutting stroke respectively; N represents the total number of predictions.

[0095] The smaller the values ​​of RMSE, MAE, and MAPE, the smaller the prediction error of the model, that is, the higher the accuracy of the model.

[0096] like Figure 6 to Figure 9As shown, the present invention first adopts a high-speed CNC machine tool platform based on the experiment of the PHM2010 data set, and uses multiple sensors to collect multi-source signals, including three-dimensional cutting force, three-dimensional vibration and acoustic emission signals, with a sampling rate of 50kHz. The experimental data covers 315 cutting cycles, in which severe tool wear and breakage occurred in the later stage. The model training is divided into two stages: the first stage pre-trains the hybrid residual convolutional neural network through representation learning of unlabeled data (including data-driven pre-tasks and regularization constraint tasks) to optimize the signal feature extraction capability; the second stage uses limited labeled data to fine-tune the model, and achieves high-precision prediction of tool wear values ​​through mean square error loss. The experimental results show that this method performs superiorly in wear monitoring and remaining life prediction, and has significant industrial application value.

[0097] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A tool wear prediction method based on self-supervised learning of physical information under few labels, characterized by: The following steps are involved: S1. Collect monitoring signals such as cutting force, vibration and acoustic emission during the processing; S2. The collected signals are used as data samples, some samples are marked, and each sample is preprocessed by interception and Z-score standardization; S3. Design a hybrid residual convolutional neural network tool wear monitoring model to perform information fusion and high-dimensional representation extraction on multi-source sensor signals; S4. Based on the self-supervised learning framework, a pair of auxiliary tasks are used to pre-train the tool wear monitoring model. By enhancing view generation and twin network contrast loss calculation, the model is guided to extract the invariant features of the signal; S5. Based on the self-supervised learning framework, the tool wear monitoring model is pre-trained using auxiliary task 2, and the physical consistency of the representation vector is enhanced by introducing physical constraints; S6. Based on the self-supervised learning framework, by alternately optimizing auxiliary task 1 and auxiliary task 2, the model is gradually guided to have both efficient feature extraction capabilities and consistency with physical laws; S7. Based on the self-supervised learning framework, a limited number of labeled samples are used to fine-tune the pre-trained tool wear monitoring model and obtain the final wear value evaluation results; S8. Evaluate the overall performance of the model in the tool wear condition monitoring task by calculating the model evaluation index.

2. The tool wear prediction method based on self-supervised learning of physical information with few labels according to claim 1 is characterized by: The S2 comprises the following steps: S21. Randomly select a small number of tools to mark tool wear during their life cycle, and build a data set consisting of a small number of marked samples and a large number of unmarked samples; S22. Each sample in the data set is truncated along the time step dimension. In the sample truncation process, the beginning and end data points of the sample are first discarded; then, some sample points at the end of the remaining data are retained as initial samples; S23. The sample data was z-score standardized, and the expression is: Where x represents the value of a sampling point on a sensor in the sample; μ represents the mean of all sampling points of sensor data along the path x; σ represents the standard deviation of all sampling points of the sensor data x.

3. The tool wear prediction method based on self-supervised learning of physical information with few labels according to claim 1 is characterized by: The hybrid residual convolutional neural network tool wear monitoring model in S3 includes a hybrid residual convolutional neural network and a fully connected regression layer, and the tool wear monitoring value is output through the fully connected regression layer; The hybrid residual convolutional neural network is composed of one-dimensional convolution, two-dimensional convolution and residual connection, and is responsible for converting raw data into effective data representation.

4. The tool wear prediction method based on self-supervised learning of physical information with few labels according to claim 3 is characterized by: The specific steps of performing information fusion and high-dimensional representation extraction on multi-source sensor signals in S3 are: First, the input data passes through multiple one-dimensional convolution modules to independently extract the shallow features of each sensor signal and increase the number of channels; then, these shallow features are fused through a two-dimensional convolution module to extract the deep features of multi-source signals; after passing through a deep residual network, the output features are reduced in dimension through an adaptive average pooling layer and converted into a one-dimensional vector through a flattening layer; These one-dimensional vectors are then fed into a fully connected regression layer.

5. The tool wear prediction method based on self-supervised learning of physical information with few labels according to claim 1 is characterized by: The auxiliary task 1 in S4 is a data-driven pre-auxiliary task based on a twin network, and the twin network is composed of two weight-sharing hybrid residual convolutional neural networks to construct two branches of the twin network; The specific steps of S4 are: The auxiliary task 1 generates two different views x1 and x2 by randomly combining different data enhancement methods, and inputs them into the two branches of the twin network to obtain the potential tool wear representations z1 and z2 under different views; then, a predictor is introduced in the upper branch of the twin network to generate a prediction vector p1, and a cosine similarity comparison is performed with the representation z2 of the lower branch to guide the hybrid residual convolutional neural network to learn an invariant representation, and the cosine similarity is used to measure the similarity between them, which is defined as follows: Where ‖·‖2 represents the l2 norm; Using the stop gradient operation in the lower branch, the augmented views are swapped to calculate the loss function in a symmetric way: Where sg(·) represents the gradient stop operation on the input variable.

6. The tool wear prediction method based on self-supervised learning of physical information with few labels according to claim 1 is characterized by: The auxiliary task 2 in S5 is a regularized constraint auxiliary task based on physical properties. The auxiliary task 2 proposes two distance measurement methods integrating physical knowledge by combining the idea of ​​triple loss, which together constitute a regularized constraint loss function. The triplet loss is calculated by degrading representations of three different samples; in this sample triplet, the reference input is usually defined as the anchor sample x A , corresponding to the cutting process monitoring data at time t0, represented by symbol A; similar inputs close to the reference input are defined as positive samples x P , corresponding to the moment t1, represented by the symbol P; The input that is significantly different from the reference input is defined as a negative sample x. N , corresponding to time t2, represented by the symbol N; for a given triple sample {x A ,x P ,x N }, the embedded representation obtained through the backbone network is represented by {z A ,z P ,z N } indicates that, The first distance measurement method of the two distance measurement methods constrains the Euclidean distance of the representation in the embedded space, requiring that the anchor sample x A With positive sample x P The Euclidean distance in the embedding space is less than the negative sample x N , thus obtaining the restriction condition: where δ is the margin of the mandatory condition between sample pairs, represents the square of the l2 norm; for the i-th group of triplet samples, the loss function of the Euclidean distance constraint is: The second distance measure of the two distance measures constrains the development trend of the representation and requires that the vector With vector The direction is as consistent as possible; cosine similarity S is used c As a measure of the vector angle representing the trend requirement, for the i-th group of triple samples, the vector angle constraint loss function is obtained as: This results in a loss function constraining the Euclidean distance The loss function of the angle between the constraint representation vector The two are combined by weighted summation to form the loss function of task 2 to represent the metric constraints for: Among them, λ1 and λ2 are weighting coefficients.

7. The tool wear prediction method based on self-supervised learning of physical information with few labels according to claim 1 is characterized by: The S7 fine-tuning using limited-scale labeled samples includes the following steps: S71. Freeze the shallow convolutional network parameters of the pre-trained model and only adjust the deep residual module; S72. The optimization process of the deep residual module and the fully connected regression layer is to first obtain the model error by calculating the loss function, then use back propagation to calculate the gradient of the loss function with respect to the model parameters, and optimize the parameters of the deep residual module and the fully connected regression layer by the gradient descent method, and finally complete the model training with the goal of minimizing the mean square error. The optimization target can be expressed as: Where n is the labeled dataset D L The number of samples in y i , They represent the actual remaining life and predicted remaining life after the i-th cutting stroke respectively; S73. The model obtains data representation through a fine-tuned hybrid residual convolutional neural network, and after average pooling and flattening layers, it outputs the predicted wear value through an optimized fully connected regression layer.

8. The tool wear prediction method based on self-supervised learning of physical information with few labels according to claim 1 is characterized by: The model evaluation indicators in S8 are root mean square error, mean absolute error and mean absolute percentage error, and their calculation formulas are: Among them, RMSE is the root mean square error, MAE is the mean absolute error, MAPE is the mean absolute percentage error, and y i , They represent the actual remaining life and predicted remaining life after the i-th cutting stroke respectively; N represents the total number of predictions.

Images