A high-precision lightweight servo turret tool wear monitoring method and system

Abstract

The invention discloses a high-precision lightweight servo turret tool wear monitoring method and system, wherein the method includes: calculating the time domain, frequency domain and time-frequency domain features of the acceleration signal, horizontally splicing to obtain a multi-domain feature matrix, and normalizing; calculating the correlation coefficient between each feature and the label after normalization; calculating the reconstruction weight of each feature to obtain a reconstructed feature weight set; automatically adding the reconstructed features in each domain, horizontally splicing to obtain a reconstructed multi-domain feature matrix; establishing a physical model, and constructing a PINN model in combination with the physical model; establishing a physical loss weight adaptive adjustment function; constructing a total loss function by integrating data-driven loss and physical loss; using the obtained reconstructed multi-domain feature matrix as the input of the PINN model, combining the total loss function for training, and the trained PINN model is used to predict tool wear. The invention can meet the needs of industrial sites for high-precision and lightweight tool wear monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical machining process monitoring, and in particular to a high-precision lightweight servo turret tool wear monitoring method and system. Background Art

[0002] Tool wear is a key factor affecting machine tool processing accuracy and production efficiency, especially in the later stage of tool wear, as the wear rate accelerates, the working conditions become more complicated. At this time, if the tool wear status cannot be accurately monitored, it will directly affect the judgment of the timing of tool change, causing machine tool shutdown, and thus reducing production efficiency and product quality. In a servo turret with multiple tools, due to the frequent replacement of tools, higher requirements are placed on the high precision of wear monitoring. At the same time, the diversity of tool types and quantities makes it necessary to train multiple models, but this also increases the computing and storage burden, requiring the model to have lightweight characteristics while ensuring accuracy in order to achieve efficient monitoring and rapid response.

[0003] In the prior art, a Chinese patent document with publication number CN113609789A discloses a tool wear prediction method based on kernel principal component analysis (KPCA) fused with spatiotemporal features, and obtains a tool wear sensitive feature set through nonlinear dimensionality reduction technology to improve the feature distinguishing ability. However, when faced with complex wear patterns, this method has limited feature reconstruction effect and is difficult to fully reflect the actual wear state of the tool. A Chinese patent document with publication number CN113378725A discloses a tool fault diagnosis method based on a multi-scale-channel attention deep neural network, which enhances the prediction ability of the model through a multi-scale deep network structure and improves the prediction accuracy. However, the model has a complex structure and many parameters, resulting in a large consumption of computing resources, making it difficult to apply efficiently in real-time monitoring systems.

[0004] However, the existing feature reconstruction methods are difficult to effectively reflect the behavioral patterns of tool wear, resulting in the inability of the extracted features to fully reflect the dynamic changes of wear, which not only limits the input quality of the prediction model, but also increases the complexity required for model prediction. Secondly, the existing high-precision prediction methods mainly rely on complex deep neural network structures. These models have many parameters and high computational costs, and are difficult to achieve lightweight, which limits their real-time application in actual production environments. In addition, although lightweight models have advantages in practical applications, their simple structures are difficult to fully express the complex working conditions and nonlinear changes in the later stages of wear, resulting in reduced prediction accuracy, especially in the later stages of tool wear. These problems jointly restrict the further development of tool wear monitoring technology. It is urgent to improve the feature reconstruction method, improve the accuracy of the prediction model, and achieve lightweight models to meet the needs of industrial sites for efficient and accurate monitoring. Summary of the invention

[0005] The present invention provides a high-precision lightweight servo turret tool wear monitoring method and system, which can meet the needs of industrial sites for high-precision and lightweight tool wear monitoring and has broad application prospects.

[0006] A high-precision lightweight servo turret tool wear monitoring method comprises the following steps:

[0007] (1) Collect the acceleration signal of the tool and calculate the time domain characteristics, frequency domain characteristics, and time-frequency domain characteristics of the acceleration signal; horizontally splice these three feature matrices to obtain a multi-domain feature matrix and normalize it;

[0008] (2) Calculate the normalized Pearson correlation coefficient between each feature and the label to obtain a set of multi-domain feature correlation coefficients; where the label represents the tool wear amount corresponding to each sample;

[0009] (3) Based on the multi-domain feature correlation coefficient set, the reconstruction weight of each feature is calculated according to the adaptive feature reconstruction function; the reconstructed feature weight set is obtained through iterative adjustment;

[0010] (4) Reconstruct each feature according to the reconstructed feature weight set and the multi-domain feature matrix. The reconstructed features in each domain are automatically summed to obtain the reconstructed features corresponding to each domain, and then horizontally spliced ​​to obtain the reconstructed multi-domain feature matrix F * ;

[0011] (5) According to the nonlinear growth law of tool wear, a physical model is established to calculate the theoretical wear of the tool; a lightweight PINN model is constructed in combination with the physical model to predict tool wear;

[0012] (6) Based on error uncertainty analysis, an adaptive adjustment function for physical loss weight is established; the total loss function is constructed by integrating data-driven loss and physical loss;

[0013] (7) Reconstruct the multi-domain feature matrix F obtained in step (4) * As the input of the PINN model, it is trained in combination with the total loss function;

[0014] (8) For each tool configured in the servo turret, train an independent PINN model according to the above steps (1) to (7);

[0015] (9) During the operation of the servo turret, the PINN model corresponding to the current machining tool is called in real time to monitor the wear status of the tool in real time.

[0016] The present invention improves the performance of features through adaptive reconstruction of multi-domain features, accurately reflects tool wear behavior, and thus reduces the dependence of the prediction model on complex structures. In addition, physical knowledge is introduced, and an adaptive physical loss weight adjustment strategy is designed based on the uncertainty of the prediction error, which effectively improves the prediction accuracy of the lightweight model under complex working conditions, especially in the later stage of tool wear.

[0017] In step (1), the representation of the multi-domain feature matrix is ​​as follows:

[0018] ;

[0019] In the formula, is the multi-domain feature matrix is the time domain feature matrix, is the frequency domain feature matrix, is the time-frequency domain feature matrix, is the number of samples, is the time domain feature number, is the frequency domain feature number, is the feature number in the time-frequency domain.

[0020] In step (3), the formula of the adaptive feature reconstruction function is:

[0021] ;

[0022] In the formula, It is In the iteration The correlation coefficient of the features, Used to adaptively adjust the degree of enhancement or reduction.

[0023] In step (3), the reconstructed feature weight set is obtained through iterative adjustment, specifically:

[0024] During the weight iteration process, the maximum difference in feature weights is calculated :

[0025] or ;

[0026] when or the maximum number of iterations is reached When , stop the iteration and output the weight of the feature ; The threshold indicating the maximum difference.

[0027] In step (5), the physical model constructed is:

[0028] ;

[0029] In the formula, It is cutting Theoretical tool wear at times, is the initial wear of the tool, and A, B, and C are the linear and quadratic coefficients, respectively.

[0030] In step (6), based on the error uncertainty analysis, the physical loss weight adaptive adjustment function is established, and the formula is:

[0031] ;

[0032] ;

[0033] In the formula, It is n Uncertainty in prediction error during the first cut, represents the error variance, is the wear value predicted by the model; is the physical loss weight, is the initial physical loss weight; To adjust the coefficient, control the change range of physical weight; is a constant that prevents the denominator from being zero.

[0034] In step (6), the total loss function is constructed as follows:

[0035] ;

[0036] In the formula, is the weight constant of the data-driven loss, is the adaptive weight of the physical loss, It is The actual tool wear during the first cutting, is the wear value predicted by the PINN model, It is Theoretical wear volume for sub-physical models.

[0037] A high-precision lightweight servo turret tool wear monitoring system includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned high-precision lightweight servo turret tool wear monitoring method.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. The present invention significantly improves the performance of features through a multi-domain feature adaptive reconstruction method, and can accurately reflect tool wear behavior. This method overcomes the shortcomings of existing feature reconstruction methods that are difficult to fully capture tool wear behavior, ensures the high quality of the input features of the prediction model, and reduces the dependence of the prediction model on complex structures, thus achieving a lightweight model.

[0040] 2. The present invention introduces physical knowledge and designs an adaptive physical loss weight adjustment strategy based on prediction error uncertainty. This strategy dynamically adjusts the loss weight according to the uncertainty of the real-time prediction error, effectively improving the prediction accuracy of the lightweight model under complex working conditions, especially in the later stage of tool wear, and solving the problem that the existing high-precision prediction method relies on complex deep neural network structure and has high computational cost.

[0041] 3. The present invention ensures the stability and accuracy of the lightweight model under complex working conditions and nonlinear changes in the late stage of wear through the synergistic effect of multi-domain feature adaptive reconstruction and adaptive physical loss weight adjustment strategy, significantly reduces the uncertainty in the prediction error, and meets the needs of industrial sites for high-precision, lightweight tool wear monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The present invention is a flowchart of a method for monitoring tool wear of a high-precision lightweight servo turret according to an embodiment of the present invention.

[0043] Figure 2 This is the correlation coefficient matrix diagram calculated for the multi-domain feature matrix and the target vector.

[0044] Figure 3 The first three-dimensional data distribution diagram shows the effects of the traditional feature reconstruction method and the adaptive feature reconstruction method of the present invention.

[0045] Figure 4 The figure is a comparison chart of the correlation, monotonicity and mutual information indicators between the feature set and the target vector reconstructed based on different methods.

[0046] Figure 5 Comparison chart of the regression curves of the prediction results obtained by different methods. DETAILED DESCRIPTION

[0047] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0048] like Figure 1As shown, a high-precision lightweight servo turret tool wear monitoring method mainly includes two sub-steps, namely sub-step (a): multi-domain feature adaptive reconstruction method, sub-step (b): physical loss weight adaptive adjustment method based on knowledge guidance and error uncertainty.

[0049] A multi-domain feature adaptive reconstruction method, comprising:

[0050] Calculate the time domain, frequency domain and time-frequency domain features of the acceleration signal, horizontally splice them to obtain a multi-domain feature matrix, and normalize them to eliminate feature differences;

[0051] Calculate the normalized Pearson correlation coefficient between each feature and the label to obtain a set of multi-domain feature correlation coefficients;

[0052] According to the adaptive feature reconstruction function, the reconstruction weight of each feature is calculated. Through iterative adjustment, the features with excellent performance are enhanced and the features with weak performance are weakened until the reconstruction weight change meets the set convergence condition or reaches the maximum number of iterations, and the reconstructed feature weight set is obtained;

[0053] Each feature is reconstructed according to the reconstruction feature weight set and the multi-domain feature matrix. The reconstructed features in each domain are automatically summed to obtain the reconstructed features corresponding to each domain, and then horizontally spliced ​​to obtain the reconstructed multi-domain feature matrix.

[0054] Specifically, Figure 1 As shown in (a), the following steps are included:

[0055] S1: Calculate the characteristics of the acceleration signal in three domains, including the time domain characteristic matrix , frequency domain feature matrix , time-frequency domain feature matrix , these features are combined by horizontal splicing to obtain a multi-domain feature matrix , the matrix representation is as follows:

[0056] ;

[0057] In the formula, is the number of samples, is the time domain feature number, is the frequency domain feature number, is the feature number in the time-frequency domain.

[0058] S2: In order to eliminate the dimensional differences between different features, the multi-domain feature matrix 𝐹 needs to be normalized. , the normalized formula is:

[0059] ;

[0060] Normalized multi-domain feature matrix It is expressed as:

[0061] ;

[0062] S3: Calculate each feature With the target variable The correlation of the target variable is calculated by Pearson coefficient. is a size of m ×1 vector, representing the tool wear amount corresponding to each sample. The correlation coefficient calculation formula is as follows:

[0063] ;

[0064] In the formula, It is a feature With the target variable l The covariance of . It is a feature The standard deviation of is the target variable l For each domain feature, the corresponding correlation coefficient is calculated, which is expressed as:

[0065] ;

[0066] S4: Calculate the weight of each feature according to the adaptive feature reconstruction function. The adaptive adjustment of the weight iteratively enhances the features with excellent performance and weakens the features with weaker performance until the set convergence conditions are met. The weight calculation formula is as follows:

[0067] ;

[0068] In the formula, It is In the iteration The correlation coefficient of the features, Used to adaptively adjust the degree of enhancement or reduction.

[0069] S5: During the weight iteration process, calculate the maximum difference in feature weights :

[0070] or ;

[0071] when or the maximum number of iterations is reached When , stop the iteration and output the weight of the feature ; The threshold indicating the maximum difference.

[0072] Among them, the weight matrix of time domain, frequency domain and time-frequency domain features is , and .

[0073] S6: Apply weights to the feature matrix , and perform weighted reconstruction on it, where , the reconstructed matrix F * It is expressed as:

[0074] ;

[0075] Among them, the reconstruction features of each domain are obtained by weighted summation: , , .

[0076] The above steps are based on the characteristics of tool wear, and adopt a multi-domain feature adaptive reconstruction method. Through the adaptive feature reconstruction function and weight adjustment mechanism, the performance of the features is significantly improved, and the tool wear behavior is accurately reflected. This method overcomes the shortcomings of existing feature reconstruction methods that are difficult to fully capture tool wear behavior, and helps reduce the dependence of the prediction model on complex structures, thereby providing an effective way to achieve lightweight models. The reconstructed features more comprehensively reflect the dynamic changes of tool wear, ensuring the reliability and accuracy of the prediction results.

[0077] The feature adaptive reconstruction method proposed in this embodiment avoids the information loss in the feature reconstruction process of traditional methods such as kernel principal component analysis and multidimensional scaling, and directly performs weight adaptive adjustment through the correlation between the feature and the target variable. The reconstructed feature matrix can more accurately reflect the dynamic changes of tool wear, especially significantly improving the prediction accuracy of the late wear period.

[0078] A method for adaptively adjusting physical loss weights based on knowledge guidance and error uncertainty, comprising:

[0079] According to the nonlinear growth law of tool wear, a physical model is established to calculate the theoretical wear of the tool. A physical information network (PINN) is constructed in combination with the physical model to predict tool wear.

[0080] Data-driven loss and physical loss are calculated according to the PINN model, and an adaptive adjustment mechanism of physical loss weight is established based on error uncertainty analysis to solve the problems of increased error and prediction uncertainty in the later stage of tool wear.

[0081] Specifically, Figure 1 As shown in (b), the following steps are included:

[0082] S1: The Sipos model is selected as the physical model to describe the cumulative change of tool wear with the number of cutting times 𝑛. The formula is:

[0083] ;

[0084] In the formula, It is cutting Theoretical tool wear at times, is the initial wear of the tool, and A, B, and C are the linear and quadratic coefficients, respectively.

[0085] S2: Introduce uncertainty error suppression mechanism to dynamically adjust the loss weight of the model by calculating the uncertainty of the prediction error. The formula is:

[0086] ;

[0087] In the formula, It is n Uncertainty in prediction error during the first cut, represents the error variance, is the wear value predicted by the model.

[0088] Dynamically adjust physical loss weights based on uncertainty analysis results To adapt to the uncertainty changes with the number of cuts. The adjustment formula is:

[0089] ;

[0090] is the physical loss weight, is the initial physical loss weight; To adjust the coefficient, control the change range of physical weight; is a constant that prevents the denominator from being zero.

[0091] S3: Construct the total loss function. The total loss function combines the characteristics of data-driven and physical-driven, and is defined as follows:

[0092] ;

[0093] In the formula, is the weight constant of the data-driven loss, is the adaptive weight of the physical loss, It is The actual tool wear during the first cutting, is the wear value predicted by the PINN model, It is Theoretical wear volume for sub-physical models.

[0094] S4: Through the PINN model, the reconstructed feature matrix F in substep (a) is transformed * As input, combined with the total loss function for training. The model adjusts parameters by minimizing the loss function until the total loss function converges. Finally, the predicted value of tool wear is obtained. .

[0095] Sub-step (b) introduces an adaptive physical loss weight adjustment strategy to effectively deal with the uncertainty caused by complex working conditions and nonlinear changes in the tool wear process. By dynamically adjusting the loss weight, this method can accurately capture the tool wear behavior in the later stage of wear and significantly reduce the prediction error. Combined with the feature adaptive reconstruction method, this strategy ensures that the model has good robustness and stability in the later stage of tool wear, improves the overall accuracy and reliability of the prediction, and provides support for the efficient application of lightweight models.

[0096] For each tool configured in the servo turret, the above sub-steps (a) and (b) can be used to train independent PINN models. Each model is based on the specific characteristics and usage frequency of the tool, and is trained and verified through historical processing data to ensure that the model can accurately reflect the wear of the tool.

[0097] During the operation of the servo turret, the system monitors the status of the tool in real time. When the turret changes the tool, the system automatically identifies the current tool number and calls the wear monitoring model corresponding to the tool to ensure seamless wear monitoring.

[0098] Once the system identifies and calls the monitoring model of the corresponding tool, it begins to monitor the wear status of the new tool in real time. This monitoring process will continue until the tool is replaced again. After each tool change, the system automatically repeats this step, thereby realizing continuous wear monitoring of multiple tools in the servo turret.

[0099] In order to verify the feasibility of the methods in sub-steps (a) and (b), a milling data set provided by the Prognostics and Health Management (PHM) Society of New York, USA in 2010 was selected, including the milling data of three tools, C1, C4 and C6.

[0100] Considering that each cutting process is extremely worn or even damaged in the later stage, causing the signal to fluctuate violently, the present invention selects the data of the first 300 cuttings of each data set. In order to meet the cost requirements of industrial applications, the x-direction acceleration is selected as the monitoring signal.

[0101] 14 time domain features and 13 frequency domain features were extracted from the acceleration signal. And 8 wavelet coefficient energies, namely time-frequency domain features, were obtained through wavelet transform. Based on this, the time domain, frequency domain and time-frequency domain feature sets in formula (1) were obtained, where m =300, a =14, b =13, d =8.

[0102] In addition, each cutting pass in the data set corresponds to a wear label, and the average wear value of the three cutting edges is selected as the target vector for method verification, that is, the label in Eq. (4) l .

[0103] Therefore, according to step S3 in sub-step (a), the correlation coefficients between the multi-domain features and the target vector are calculated to obtain a correlation coefficient matrix, such as Figure 2 shown.

[0104] Next, an iterative calculation is performed based on the adaptive reconstruction function in S4 to obtain the reconstruction weight factor of each feature. In this embodiment, γ =28 satisfies the maximum weight difference is less than ( ) requirements, in addition, the maximum number of iterations in this embodiment T max = 200, initial γ =1, iterative step length The reconstruction weight factors in each domain are horizontally spliced ​​to obtain the weight matrices of three different domains ( , and ).

[0105] Based on S6, the weight matrix is ​​applied to the multi-domain feature set to obtain the weighted feature components, and then the feature components in each domain are summed to obtain the reconstructed feature matrix F * At this point, the feature differences are eliminated and the feature representation performance is improved through the adaptive feature reconstruction function and weight adjustment mechanism.

[0106] In order to further verify the superiority of the adaptive feature reconstruction method, the multi-domain feature matrix is ​​reconstructed by using traditional kernel principal component analysis (KPCA), distributed domain embedding (T-SNE), multidimensional scaling (MDS) and the adaptive feature reconstruction method in Example 1. Figure 3 Shows the distribution of the data after reduction to three dimensions.

[0107] pass Figure 3It can be seen that the characteristics of the three data sets C1, C4, and C6 reconstructed by the proposed method all show an obvious monotonically increasing trend in the three-dimensional space, which is consistent with the physical law of tool wear accumulation over time. On the contrary, other methods show nonlinear or chaotic distribution. This method provides a guarantee for establishing a lightweight model.

[0108] In addition, taking the C1 dataset as an example, the correlation, monotonicity and mutual information between the feature matrix and the target vector obtained by different reconstruction methods are calculated. The results are as follows: Figure 4 As shown in the figure, the superiority of the adaptive reconstruction method is verified through quantitative analysis.

[0109] Next, the reconstructed features are used as input to verify the adaptive adjustment method of physical loss weight based on knowledge guidance and error uncertainty proposed in sub-step (b), and further verify the effectiveness of this method in tool wear prediction.

[0110] Based on the reconstructed features, a lightweight PINN model is established, which includes 3 fully connected layers and uses a tanh activation function. According to steps S1-S3 in substep (b), physical loss is introduced, and the physical loss weight is adaptively adjusted based on the uncertainty of the prediction error.

[0111] In order to improve the generalization ability of the model, this embodiment combines C4 and C6 as a training set and C1 as a test set.

[0112] The tool wear prediction curves obtained by considering whether physical loss is considered and whether the physical loss weight adaptive adjustment method is applied are compared. Figure 5 As shown. Among them, consider λ =0, λ =0.01 and the root mean square error (RMSE) of the recognition results corresponding to the proposed adaptive weight adjustment method are 16.32, 14.94, and 5.78 respectively. It can be found that the overall RMSE of the prediction curve corresponding to this embodiment is the lowest, and the accuracy is greatly improved in the later stage of wear, which proves the superiority of the physical loss weight adaptive adjustment method based on knowledge guidance and error uncertainty.

[0113] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A high-precision lightweight servo turret tool wear monitoring method, characterized in that: The following steps are involved: (1) Collect the acceleration signal of the tool and calculate the time domain characteristics, frequency domain characteristics and time-frequency domain characteristics of the acceleration signal; The three feature matrices are horizontally concatenated to obtain a multi-domain feature matrix, which is then normalized. (2) Calculate the normalized Pearson correlation coefficient between each feature and the label to obtain a set of multi-domain feature correlation coefficients; where the label represents the tool wear amount corresponding to each sample; (3) Based on the multi-domain feature correlation coefficient set, the reconstruction weight of each feature is calculated according to the adaptive feature reconstruction function; the reconstructed feature weight set is obtained through iterative adjustment; (4) Reconstruct each feature according to the reconstructed feature weight set and the multi-domain feature matrix. The reconstructed features in each domain are automatically summed to obtain the reconstructed features corresponding to each domain, and then horizontally spliced ​​to obtain the reconstructed multi-domain feature matrix F * ; (5) According to the nonlinear growth law of tool wear, a physical model is established to calculate the theoretical wear of the tool; a lightweight PINN model is constructed in combination with the physical model to predict tool wear; (6) Based on error uncertainty analysis, an adaptive adjustment function for physical loss weight is established; the total loss function is constructed by integrating data-driven loss and physical loss; (7) Reconstruct the multi-domain feature matrix F obtained in step (4) * As the input of the PINN model, it is trained in combination with the total loss function; (8) For each tool configured in the servo turret, train an independent PINN model according to the above steps (1) to (7); (9) During the operation of the servo turret, the PINN model corresponding to the current machining tool is called in real time to monitor the wear status of the tool in real time.

2. The high-precision lightweight servo turret tool wear monitoring method according to claim 1 is characterized in that: In step (1), the representation of the multi-domain feature matrix is ​​as follows: ； In the formula, is the multi-domain feature matrix, is the time domain feature matrix, is the frequency domain feature matrix, is the time-frequency domain feature matrix, is the number of samples, is the time domain feature number, is the frequency domain feature number, is the feature number in the time-frequency domain.

3. The high-precision lightweight servo turret tool wear monitoring method according to claim 1 is characterized in that: In step (3), the formula of the adaptive feature reconstruction function is: ； In the formula, It is k In the iteration j The correlation coefficient of the features, Used to adaptively adjust the degree of enhancement or reduction.

4. The high-precision lightweight servo turret tool wear monitoring method according to claim 3 is characterized in that: In step (3), the reconstructed feature weight set is obtained through iterative adjustment, specifically: During the weight iteration process, the maximum difference in feature weights is calculated : or ; when or the maximum number of iterations is reached When , stop the iteration and output the weight of the feature ; The threshold indicating the maximum difference.

5. The high-precision lightweight servo turret tool wear monitoring method according to claim 1 is characterized in that: In step (5), the physical model constructed is: ； In the formula, It is cutting Theoretical tool wear at times, is the initial wear of the tool, and A, B, and C are the linear and quadratic coefficients, respectively.

6. The high-precision lightweight servo turret tool wear monitoring method according to claim 5 is characterized in that: In step (6), based on the error uncertainty analysis, the physical loss weight adaptive adjustment function is established, and the formula is: ； ； In the formula, It is n Uncertainty in prediction error during the first cut, represents the error variance, is the wear value predicted by the model; is the physical loss weight, is the initial physical loss weight; α is the adjustment coefficient, which controls the change range of the physical weight; is a constant that prevents the denominator from being zero.

7. The high-precision lightweight servo turret tool wear monitoring method according to claim 6 is characterized in that: In step (6), the total loss function is constructed as follows: ； In the formula, is the weight constant of the data-driven loss, is the adaptive weight of the physical loss, It is The actual tool wear during the first cutting, is the wear value predicted by the PINN model, It is Theoretical wear volume for sub-physical models.

8. A high-precision lightweight servo turret tool wear monitoring system, characterized in that: It comprises a memory and one or more processors, wherein the memory stores executable codes, and when the one or more processors execute the executable codes, they are used to implement the high-precision lightweight servo turret tool wear monitoring method described in any one of claims 1-7.

Images