A method for monitoring tool wear in varying operating conditions

Abstract

The present application relates to the field of mechanical processing, in particular to a variable working condition tool wear monitoring method fusing a feature conversion method and a contrastive unsupervised domain adaptation regression strategy. The present application uses the feature conversion method to select features that are not sensitive to process parameter changes, while converting the feature amplitude; a one-dimensional convolutional neural network and a bidirectional long short-term memory neural network are fused to construct a parallel hybrid neural network as a feature extractor; and an unsupervised domain adaptation strategy representing subspace distance and orthogonal basis mismatch penalty is used to align the feature distribution under variable working conditions. The present application uses the unsupervised domain adaptation strategy based on the representation subspace distance and the orthogonal basis mismatch penalty to reduce the distribution difference of the mixed features learned by the feature extractor under different working conditions, and integrates the contrastive learning module to retain the inherent structural information in the data under the new working condition, thereby improving the generalization ability of the model and helping to improve the dynamic perception and autonomous decision-making ability level of the numerical control processing process.

Description

Technical Field

[0001] This invention relates to the field of machining, and more specifically to a method for monitoring tool wear under varying working conditions that integrates a feature transformation method and a comparative unsupervised domain adaptive regression strategy. Background Technology

[0002] Machine tools, as the "mother machines" of manufacturing, are widely used in aviation, aerospace, energy, and other fields, possessing fundamental, universal, and strategic significance. Cutting tools, as the end effector of machine tools, are a key factor in ensuring part machining quality and production efficiency, but they are also components prone to wear and rapid performance degradation. In recent years, the development of advanced sensing equipment, next-generation information technology, and artificial intelligence has led to a revolution in tool wear monitoring technology, becoming a major research direction in tool wear monitoring. Relying on the unique information processing modules and hierarchical feature learning mechanisms of deep learning algorithms, constructing hybrid deep learning models can effectively solve the problem of tool wear monitoring. However, due to complex and varied machining conditions, the probability density distribution of cutting signals varies greatly, and tool wear degradation patterns differ. Even under the same machining conditions, due to geometric and installation errors of the tool, random fluctuations in the material properties of the tool and workpiece, and other factors, the cutting signals will still differ, severely interfering with the performance of hybrid deep learning networks. To solve the problem of tool wear monitoring under multiple working conditions, transfer learning, as a new machine learning paradigm, can effectively reduce the variability in tool wear data distribution under different working conditions and improve the accuracy of tool wear monitoring. However, most existing studies only focus on feature alignment between the source domain (old operating condition) and the target domain (new operating condition), without considering constructing feature quantities that are insensitive to changes in process parameters from the original feature vectors. Moreover, some transfer learning methods rely on the number of sample labels in the target domain, which limits the application of variable operating condition tool wear monitoring methods in practical engineering. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method for monitoring tool wear under varying working conditions by fully mining the effective information in the original feature vector based on the cutting force signal, without relying on the target domain sample label.

[0004] To solve the above technical problems, the present invention adopts the following technical solution:

[0005] ① Construct feature vectors through local feature extraction operations;

[0006] ② Use feature transformation methods to select features that are not sensitive to changes in process parameters, and transform the feature amplitude at the same time;

[0007] A parallel hybrid neural network is constructed by fusing one-dimensional convolutional neural networks and bidirectional long short-term memory neural networks;

[0008] An unsupervised domain adaptation strategy representing subspace distance and orthogonal basis mismatch penalty is adopted to align the feature distribution under varying working conditions, and a contrastive learning module is combined to retain the inherent information in the cutting signal data under the new working conditions;

[0009] Using the transformed feature vectors from step ② The network built in the middle is used for training to improve the model's prediction accuracy.

[0010] Step ① includes:

[0011] 101 The original cutting signal is divided into twenty segments, and time-domain, frequency-domain, and time-frequency-domain features are extracted from each segment of the cutting signal;

[0012] The extracted time-domain features mainly include minimum value, peak-to-peak value, root mean square value, variance, skewness, and kurtosis.

[0013] 103 When extracting frequency domain features, the original cutting signal is first converted into a spectrum using Fourier transform, and then the maximum value, mean, variance, power, skewness and kurtosis of the spectrum are calculated.

[0014] 104. In terms of time-frequency domain feature extraction, a three-level wavelet packet decomposition is performed on the original cutting signal to calculate the energy values ​​of the wavelet packet coefficients of the first two nodes and the last node.

[0015] Step ② includes:

[0016] 201 Select a set of process parameters as nominal parameters, and use the local feature extraction operation in step ① to convert the cutting signal under the nominal parameter combination into a feature vector;

[0017] 202. Calculate the average of the same feature quantities within all segments for feature transformation;

[0018] 203 When the extracted features are not affected by process parameters, the feature transformation process is expressed as follows:

[0019]

[0020] in, Indicates the transformed features, This represents the features extracted under other combinations of process parameters. , , , and These represent cutting speed, feed per tooth, depth of cut, width of cut, and machining center type, respectively. This represents the features extracted under the nominal combination of process parameters. , , , and These represent the nominal cutting speed, nominal feed per tooth, nominal depth of cut, nominal width of cut, and nominal machining center type, respectively, under the nominal combination of process parameters. These are hyperparameters that need to be optimized.

[0021] The hyperparameter expression for 204 is:

[0022]

[0023] Where n represents the number of combinations of process parameters.

[0024] 205. Set a feature selection threshold to retain insensitive features. The feature selection principle is stated as follows:

[0025]

[0026] 206. Based on the selected feature subset, local feature extraction is performed again to extract features from the original cutting signal. Using the features transformed under unworn tool conditions as a benchmark, features during the tool wear process are transformed:

[0027]

[0028] in, This represents local features extracted during the tool wear process. This indicates the characteristics transformed during the tool wear process.

[0029] Step ③ includes

[0030] 301. A parallel hybrid neural network is constructed by distributing a one-dimensional convolutional neural network and a bidirectional long short-term memory neural network in parallel, which serves as a feature extractor;

[0031] 302 A one-dimensional convolutional neural network performs convolution operations on local regions of the transformed feature vector to generate representative spatial features;

[0032] 303. Learn the long-term dependencies of tool wear degradation characteristics on the time scale through a bidirectional long short-term memory neural network;

[0033] 304 unfolds the spatial features into a one-dimensional feature vector and concatenates it with the temporal features of the last time node of the bidirectional long short-term memory neural network, thereby achieving the mixing of spatial and temporal features without interference.

[0034] Step ④ includes:

[0035] 401. Using the feature extractor in ③, feature learning is performed on the transformed feature vectors in the source and target domains. The resulting mixed features are represented as follows: and ;

[0036] 402 pairs and Perform singular value decomposition to obtain the corresponding orthonormal basis:

[0037]

[0038] in, and These represent the mixed features of the source and target domains, respectively. and Let represent the orthonormal basis after the mixed eigenvalue decomposition of the source and target domains, respectively, and let b represent the number of mini-batch data samples during the training of the feature extractor. and Let these represent the orthogonal matrices in the source and target domains, respectively.

[0039] 403. Using the protagonist to measure the similarity of the orthonormal basis of the source and target domains, the expression is:

[0040]

[0041] in, This represents the main character of the i-th mini-batch sample;

[0042] 404 The summation of the sine values ​​of all main characters yields the subspace distance, expressed as:

[0043]

[0044] in, This represents the subspace distance;

[0045] 405. An orthogonal basis mismatch penalty term is introduced to match orthogonal bases of similarity ranking, ensuring that orthogonal bases of equal importance to the source and target domains are constrained together for matching. The mathematical expression is:

[0046]

[0047] in, This represents the mismatch penalty term for orthogonal bases. and Let represent the weight matrices of the source and target domains, respectively. Denotes the F-norm of a matrix;

[0048] 406 employs information noise contrastive estimation loss to maximize the mutual information between the mixed features of the target domain and the original feature vectors, thereby preserving the inherent information of the target domain when aligning the features of the source and target domains;

[0049] 407 When calculating the information-noise contrastive estimation loss, a fully connected neural network is used to change the dimension of the mixed features in the target domain. The overall information-noise contrastive estimation loss is expressed as:

[0050]

[0051] Here, FC represents a fully connected neural network. This represents a positive sample in the target domain. This represents a negative sample in the target domain.

[0052] Step ⑤ includes:

[0053] After incorporating an unsupervised domain adaptation strategy based on representation subspace distance and orthogonal basis mismatch penalty term, the network model's optimization objective is: (The optimization objective is to compare the information and noise of the contrastive learning loss function.)

[0054]

[0055] in, This represents the total error loss function. This represents the error function between the model's monitored values ​​and the measured values. , and This represents the weighting coefficients to balance the contributions of different modules;

[0056] 502 The parameters of the feature extractor are trained using wear data throughout the entire tool lifecycle. Based on the error backpropagation mechanism, the parameter update process of the feature extractor is as follows:

[0057]

[0058] in, Indicates the parameters of the feature extractor. This represents the learning rate of the feature extractor. This indicates finding the derivative;

[0059] 503 The generalization ability of the proposed model is verified using wear data over the entire tool life cycle under new working conditions, and quantitatively characterized using absolute mean error and root mean square error:

[0060]

[0061] Where MAE represents the absolute mean error. This represents the root mean square error. and Let represent the measured value and the model predicted value of tool wear for the i-th target domain sample, respectively, and N represent the total number of samples.

[0062] The positive effects of this invention are as follows: First, this invention utilizes local feature extraction to transform the original cutting signal into a feature vector, which not only contains multi-level effective features but also retains the temporal correlation information in the cutting signal sequence. A feature transformation method is proposed to remove features sensitive to changes in process parameters from the original feature vector, while simultaneously transforming the feature amplitude, so that the information contained in the feature vector is as relevant as possible to tool wear degradation. The spatial learning capability of a one-dimensional convolutional neural network and the temporal series learning capability of a bidirectional long short-term memory neural network are integrated to construct a parallel hybrid neural network model for learning the spatial and temporal features in the transformed feature vector. Finally, an unsupervised domain adaptation strategy based on representation subspace distance and orthogonal basis mismatch penalty terms is adopted to reduce the distribution differences of the hybrid features learned by the feature extractor under different working conditions, and a contrastive learning module is incorporated to retain the inherent structural information in the data under new working conditions, improving the model's generalization ability. The proposed variable working condition tool wear monitoring method helps to improve the dynamic perception and autonomous decision-making capabilities of CNC machining processes. Attached Figure Description

[0063] Figure 1 The specific process of feature transformation method;

[0064] Figure 2 It is a feature extractor based on parallel hybrid neural networks;

[0065] Figure 3 This outlines the overall framework for monitoring tool wear under varying operating conditions.

[0066] Figure 4 The tool wear monitoring results are shown at a cutting speed of 50 m / min.

[0067] Figure 5 The tool wear monitoring results are shown at a cutting speed of 60 m / min.

[0068] Figure 6 The tool wear monitoring results are shown at a cutting speed of 70 m / min. Detailed Implementation

[0069] The present invention will now be described in detail with reference to the accompanying drawings and specific cutting examples. The main steps are as follows:

[0070] ① Construct feature vectors through local feature extraction operations

[0071] 101 The cutting signal collected during the experiment is the cutting force. Based on the sampling frequency and sampling time of the cutting signal, the original cutting force signal is divided into multiple segments (twenty segments). Then, the time domain, frequency domain and time-frequency domain features of each segment of the cutting force signal are extracted.

[0072] The extracted time-domain features mainly include minimum value, peak-to-peak value, root mean square value, variance, skewness, and kurtosis.

[0073] 103 When extracting frequency domain features, the original cutting signal is first converted into a spectrum using Fourier transform, and then the maximum value, mean, variance, power, skewness and kurtosis of the spectrum are calculated.

[0074] 104. In terms of time-frequency domain feature extraction, a three-level wavelet packet decomposition is performed on the original cutting signal to calculate the energy values ​​of the wavelet packet coefficients of the first two nodes and the last node.

[0075] ②Use feature transformation methods to select features that are insensitive to changes in process parameters, and simultaneously transform the feature amplitude.

[0076] 201 First, a set of process parameters is randomly selected as nominal parameters. The cutting force signal under the nominal parameter combination is converted into a feature vector using the local feature extraction operation in step ①.

[0077] 202 Considering the influence of factors such as the spindle feed system, spindle speed harmonics, and the natural frequencies of the tool and workpiece, the acquired cutting force signal is very likely to contain noise. The average value of the same characteristic quantity in all segments is used for feature transformation.

[0078] 203 If the extracted features are not affected by process parameters, the features extracted under other combinations of process parameters will be as close as possible to the features under the nominal parameter combinations. Therefore, the feature transformation process can be expressed as:

[0079] (1)

[0080] in, Indicates the transformed features, This represents the features extracted under other combinations of process parameters. , , , and These represent cutting speed, feed per tooth, depth of cut, width of cut, and machining center type, respectively. This represents the features extracted under the nominal combination of process parameters. , , , and These represent the nominal cutting speed, nominal feed per tooth, nominal depth of cut, nominal width of cut, and nominal machining center type, respectively, under the nominal combination of process parameters. These are hyperparameters that need to be optimized.

[0081] 204 If it is desired that the transformed characteristic under unworn tool conditions is always as close as possible to 1, then the hyperparameter can be determined by the following expression:

[0082] (2)

[0083] Where n represents the number of combinations of process parameters.

[0084] 205. Setting a feature selection threshold to retain insensitive features, and drawing on the evaluation criteria of the Pearson coefficient, combined with the actual distribution of values, the feature selection principle can be expressed as follows:

[0085] (3)

[0086] 206. Based on the selected feature subset, local feature extraction is performed again to extract features from the original cutting signal. Using the features transformed under unworn tool conditions as a benchmark, features during the tool wear process are transformed:

[0087] (4)

[0088] in, This represents local features extracted during the tool wear process. This represents the transformed features during tool wear. A detailed feature transformation process is as follows: Figure 1 As shown.

[0089] ③ A parallel hybrid neural network is constructed by fusing one-dimensional convolutional neural networks and bidirectional long short-term memory neural networks as a feature extractor.

[0090] 301. A parallel hybrid neural network is constructed by distributing one-dimensional convolutional neural networks and bidirectional long short-term memory neural networks in parallel, which serves as a feature extractor. Figure 2 As shown;

[0091] 302 A one-dimensional convolutional neural network performs convolution operations on local regions of the transformed feature vector to generate representative spatial features;

[0092] 303. Learn the long-term dependencies of tool wear degradation characteristics on the time scale through a bidirectional long short-term memory neural network;

[0093] 304 unfolds the spatial features into a one-dimensional feature vector and concatenates it with the temporal features of the last time node of the bidirectional long short-term memory neural network, thereby achieving the mixing of spatial and temporal features without interference.

[0094] ④ An unsupervised domain adaptation strategy representing subspace distance and orthogonal basis mismatch penalty is adopted to align the feature distribution under varying working conditions, and the inherent information in the cutting signal data under the new working conditions is preserved by combining a contrastive learning module, such as... Figure 3 As shown

[0095] 401. Using the feature extractor in ③, feature learning is performed on the transformed feature vectors in the source and target domains. The resulting hybrid features can be represented as follows: and ;

[0096] 402 pairs and Perform singular value decomposition to obtain the corresponding orthonormal basis:

[0097] (5)

[0098] in, and These represent the mixed features of the source and target domains, respectively. and Let represent the orthonormal basis after the mixed eigenvalue decomposition of the source and target domains, respectively, and let b represent the number of mini-batch data samples during the training of the feature extractor. and Let represent the orthogonal matrices in the source and target domains, respectively.

[0099] 403 To measure the similarity between orthonormal bases of the source and target domains, a simple metric is often used, which can be expressed as:

[0100] (6)

[0101] in, This represents the main character of the i-th mini-batch sample.

[0102] 404 Summing the sine values ​​of all main characters yields the subspace distance, as shown below:

[0103] (7)

[0104] in, This represents the subspace distance.

[0105] 405 Simultaneously, an orthogonal basis mismatch penalty term is introduced to match orthogonal bases of similarity ranking, ensuring that orthogonal bases of equal importance to the source and target domains are constrained together for matching. The mathematical expression can be described as:

[0106] (8)

[0107] in, This represents the mismatch penalty term for orthogonal bases. and Let represent the weight matrices of the source and target domains, respectively. This represents the F-norm of the matrix.

[0108] 406 Forcing the source and target domain features to align through unsupervised domain adaptation strategies is highly likely to destroy the inherent properties of the target domain data, thereby causing negative transfer. Therefore, Information Noise Contrastive Estimation (InfoNCE) loss is used to maximize the mutual information between the mixed features of the target domain and the original feature vectors, so as to preserve the inherent information of the target domain when aligning the source and target domain features.

[0109] 407 Because the original feature vectors and the mixed features learned by the feature extractor have inconsistent dimensions, a fully connected neural network is needed to change the dimension of the mixed features in the target domain when calculating the InfoNCE loss. Therefore, the overall InfoNCE loss can be expressed as:

[0110] (9)

[0111] Here, FC represents a fully connected neural network. This represents a positive sample in the target domain. This represents a negative sample in the target domain.

[0112] ⑤ Use the transformed feature vectors from step ② to pair The network built in the middle is used for training to improve the model's prediction accuracy.

[0113] 501 Due to the incorporation of an unsupervised domain adaptation strategy based on representation subspace distance and orthogonal basis mismatch penalty term, as well as the contrastive learning loss function InfoNCE, the optimization objective of the network model can now be expressed as:

[0114] (10)

[0115] in, This represents the total error loss function. This represents the error function between the model's monitored values ​​and the measured values. , and This represents the weighting coefficients to balance the contributions of different modules.

[0116] 502 The parameters of the feature extractor are trained using wear data throughout the entire tool life cycle. Based on the error backpropagation mechanism, the parameter update process of the feature extractor can be expressed as follows:

[0117] (11)

[0118] in, Indicates the parameters of the feature extractor. This represents the learning rate of the feature extractor. This indicates finding the derivative.

[0119] 503 Then, the generalization ability of the proposed model was verified using wear data over the entire tool life cycle under the new working conditions, and the absolute mean error and root mean square error were used for quantitative characterization:

[0120] (12)

[0121] Where MAE represents the absolute mean error. This represents the root mean square error. and Let represent the measured value and the model predicted value of tool wear for the i-th target domain sample, respectively, and N represent the total number of samples.

[0122] Example 1

[0123] This section mainly focuses on milling experiments conducted as described above. Haynes 230, a nickel-based superalloy, was machined under different combinations of cutting parameters. Cutting force signals were collected during the milling process, and the maximum wear width near the tool tip was measured after the experiment. The combinations of cutting parameters are shown in Table 1.

[0124] Table 1: Range of cutting parameters during milling

[0125]

[0126] The material used in this experiment was Haynes 230, a nickel-based superalloy. The cutting tool was XOMX090308TR-M08F40M. The experimental platform was a DMU 80 mono BLOCK milling machine. A Kistler 9057B force gauge was used to collect the cutting force, and a Dino-Capture 2.0 optical microscope was used to measure the maximum wear width near the tool tip.

[0127] Time-domain, frequency-domain, and time-frequency-domain features are extracted from the original cutting force signal. A feature transformation method is then applied to the extracted feature vectors to select features that are insensitive to changes in the combination of cutting parameters and to transform the feature amplitudes, so that the information contained in the feature vectors is as relevant as possible to tool wear and degradation.

[0128] To fully verify the generalization ability of the proposed model, based on the cross-validation criterion, tool wear datasets under different working conditions were used cyclically as the source and target domains. Furthermore, due to the limited sample data of tool wear under varying cutting parameters, to alleviate the overfitting problem, two sets of tool wear data were used as the source domain, and the other set as the target domain, as shown in Table 2.

[0129] Table 2. Different transfer learning tasks under cross-validation criteria

[0130]

[0131] The proposed model was trained using tool wear data from the source domain, and then its generalization ability was validated using tool wear data from the target domain. Tool wear monitoring results under different working conditions are shown below. Figure 4 , Figure 5 , Figure 6 As shown, the calculated average absolute errors are 0.0143, 0.0138, and 0.016, respectively, which can meet the accuracy requirements for tool wear monitoring in actual production.

[0132] The embodiments described above are merely preferred embodiments of the present invention, and not an exhaustive list of all possible implementations of the present invention. Any obvious modifications made by those skilled in the art 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 method for monitoring tool wear under varying working conditions, characterized in that... The steps are as follows: ① Construct feature vectors through local feature extraction operations; 101 Based on the sampling frequency and sampling time of the cutting signal, the original cutting signal is divided into twenty segments, and the time domain, frequency domain and time-frequency domain features of each segment of the cutting signal are extracted. 102 The extracted time-domain features include minimum value, peak-to-peak value, root mean square value, variance, skewness, and kurtosis; 103 When extracting frequency domain features, the original cutting signal is converted into a spectrum using Fourier transform, and the maximum value, mean, variance, power, skewness and kurtosis of the spectrum are calculated. 104 When extracting time-frequency domain features, perform three-level wavelet packet decomposition on the original cutting signal and calculate the energy values ​​of the wavelet packet coefficients of the first two nodes and the last node; ② Use feature transformation methods to select features that are not sensitive to changes in process parameters, and transform the feature amplitude at the same time; 201 Select a set of process parameters as nominal parameters, and use the local feature extraction operation in step ① to convert the cutting signal under the nominal parameter combination into a feature vector; 202. Calculate the average of the same feature quantities within all segments for feature transformation; 203 When the extracted features are not affected by process parameters, the feature transformation process is expressed as follows: ； in, Indicates the transformed features, This represents the features extracted under other combinations of process parameters. , , , and These represent cutting speed, feed per tooth, depth of cut, width of cut, and machining center type, respectively. This represents the features extracted under the nominal combination of process parameters. , , , and These represent the nominal cutting speed, nominal feed per tooth, nominal depth of cut, nominal width of cut, and nominal machining center type, respectively, under the nominal combination of process parameters. These are hyperparameters that need to be optimized; The hyperparameter expression for 204 is: ； Where n represents the number of combinations of process parameters; 205. Set a feature selection threshold to retain insensitive features. The feature selection principle is stated as follows: ； Based on the selected feature subset, local feature extraction is used to extract features from the original cutting signal. Using the features transformed under unworn tool conditions as a benchmark, features during the tool wear process are transformed. ； in, This represents local features extracted during the tool wear process. This indicates the transformed characteristics during the tool wear process; ③ A parallel hybrid neural network is constructed by fusing one-dimensional convolutional neural networks and bidirectional long short-term memory neural networks as a feature extractor; ④ An unsupervised domain adaptation strategy representing subspace distance and orthogonal basis mismatch penalty is adopted to align the feature distribution under different working conditions, and the inherent information in the cutting signal data under the new working conditions is preserved by combining the contrastive learning module; 401. Using the feature extractor in ③, feature learning is performed on the transformed feature vectors in the source and target domains. The resulting mixed features are represented as follows: and ; 402 pairs and Perform singular value decomposition to obtain the corresponding orthonormal basis: ； in, and These represent the mixed features of the source and target domains, respectively. and Let represent the orthonormal basis after the mixed eigenvalue decomposition of the source and target domains, respectively, and let b represent the number of mini-batch data samples during the training of the feature extractor. and Let these represent the orthogonal matrices in the source and target domains, respectively.

403. Using the protagonist to measure the similarity of the orthonormal basis of the source and target domains, the expression is: ； in, This represents the main character of the i-th mini-batch sample; 404 The summation of the sine values ​​of all main characters yields the subspace distance, expressed as: ； in, This represents the subspace distance; 405. An orthogonal basis mismatch penalty term is introduced to match orthogonal bases of similarity ranking, ensuring that orthogonal bases of equal importance to the source and target domains are constrained together for matching. The mathematical expression is: ； in, This represents the mismatch penalty term for orthogonal bases. and Let represent the weight matrices of the source and target domains, respectively. Denotes the F-norm of a matrix; 406 employs information noise contrastive estimation loss to maximize the mutual information between the mixed features of the target domain and the original feature vectors, thereby preserving the inherent information of the target domain when aligning the features of the source and target domains; 407 When calculating the information-noise contrastive estimation loss, a fully connected neural network is used to change the dimension of the mixed features in the target domain. The overall information-noise contrastive estimation loss is expressed as: ； Here, FC represents a fully connected neural network. This represents a positive sample in the target domain. This represents a negative sample in the target domain; ⑤ Use the transformed feature vectors from step ② to train the network built in step ③, thereby improving the model's prediction accuracy; After incorporating an unsupervised domain adaptation strategy based on representation subspace distance and orthogonal basis mismatch penalty term, the network model's optimization objective is: (The optimization objective is to compare the information and noise of the contrastive learning loss function.) ； in, This represents the total error loss function. This represents the error function between the model's monitored values ​​and the measured values. , and This represents the weighting coefficients to balance the contributions of different modules; 502 The parameters of the feature extractor are trained using wear data throughout the entire tool lifecycle. Based on the error backpropagation mechanism, the parameter update process of the feature extractor is as follows: ； in, Indicates the parameters of the feature extractor. This represents the learning rate of the feature extractor. This indicates finding the derivative; 503 The generalization ability of the proposed model is verified using wear data over the entire tool life cycle under new working conditions, and quantitatively characterized using absolute mean error and root mean square error: ； Where MAE represents the absolute mean error. This represents the root mean square error. and Let represent the measured value and the model predicted value of tool wear for the i-th target domain sample, respectively, and N represent the total number of samples.

2. The method for monitoring tool wear under varying working conditions according to claim 1, characterized in that, Step ③ includes:

301. A parallel hybrid neural network is constructed by distributing a one-dimensional convolutional neural network and a bidirectional long short-term memory neural network in parallel, which serves as a feature extractor; 302 A one-dimensional convolutional neural network performs convolution operations on local regions of the transformed feature vector to generate representative spatial features; 303. Learn the long-term dependencies of tool wear degradation characteristics on the time scale through a bidirectional long short-term memory neural network; 304 expands the spatial features into a one-dimensional feature vector and concatenates it with the temporal features of the last time node of the bidirectional long short-term memory neural network, thereby achieving the mixing of spatial and temporal features without interference.

Images