Cross-domain-oriented tool wear monitoring method and system

Abstract

The invention discloses a cross-domain-oriented tool wear monitoring method and a cross-domain-oriented tool wear monitoring system. The method comprises the following steps: carrying out time-frequency decomposition on vibration signals of a source domain and a target domain by utilizing physically guided wavelet kernel initialization, and extracting high-frequency characteristics with physical interpretability; a bat bionic attention mechanism is constructed, and dynamic enhancement and cross-domain feature alignment of wear features are realized in combination with energy guidance, echo alignment and a time-frequency attention module; splicing enhanced features of the source domain and the target domain, inputting the spliced enhanced features into a shared convolution and circulation network, and extracting space-time fusion features; and constructing an improved MMD loss function fused with the processing parameter similarity, and performing joint optimization by combining regression errors to realize cooperative training of feature alignment and wear prediction. By means of the method, the high-precision and high-robustness monitoring task of tool abrasion under different working conditions on an industrial site can be met.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical processing process monitoring, and in particular to a tool wear monitoring method and system for cross-domain scenarios. Background Art

[0002] Tool wear is a core factor affecting machining accuracy, production efficiency, and product quality during the mechanical processing process. Especially in the later stage of tool wear, the wear rate increases sharply, and the signal exhibits significant non-linear and high-frequency characteristics. If the tool wear state cannot be accurately monitored, it will directly affect the accurate judgment of the tool change time, resulting in machining interruption, workpiece quality decline, and production cost increase. In industrial scenarios with multiple working conditions and multiple tools, the diversity of tool types, machining materials, and working conditions further exacerbates the complexity of wear monitoring. Especially in cross-domain scenarios, there are significant differences in sample distributions between the source domain (such as the laboratory environment) and the target domain (such as the actual production environment). Traditional monitoring methods are difficult to effectively adapt to this domain shift, resulting in insufficient model generalization ability and difficulty in meeting the high-precision and high-robustness monitoring requirements of the industrial site. Therefore, there is an urgent need to develop a monitoring method that can accurately extract the time-frequency characteristics of tool wear, achieve cross-domain feature alignment, and improve the robustness of the model to cope with complex working conditions and cross-domain challenges.

[0003] Existing tool wear monitoring technologies mainly rely on signal processing and machine learning methods to extract wear characteristics and predict the state. However, when dealing with the high-frequency and non-linear characteristics of tool wear signals, traditional methods are often limited to simple time-domain or frequency-domain analysis and are difficult to comprehensively capture the complex time-frequency dynamic changes in the later stage of wear, resulting in insufficient feature expression ability. In cross-domain scenarios, the Chinese patent document with the publication number CN115351601A assumes that the data distributions in the source domain and the target domain are the same and lacks an effective domain alignment mechanism. When the samples in the target domain are scarce or the working condition differences are significant, the model performance drops significantly. In addition, the Chinese patent documents with the publication numbers CN117001420A and CN112801139A enhance the feature expression ability by introducing an attention mechanism and feature fusion respectively. However, these methods have insufficient precise control of high-frequency time-frequency information and are difficult to effectively adapt to the distribution differences between the source domain and the target domain in cross-domain scenarios, limiting their generalization ability and robustness in the actual industrial environment.

[0004] Although certain progress has been made in existing tool wear monitoring technologies, significant challenges still remain. First, traditional feature extraction methods struggle to accurately capture the high-frequency and non-linear time-frequency characteristics of tool wear signals. Especially under complex working conditions in the later stage of wear, the extracted features cannot fully reflect the dynamic changes in wear behavior, resulting in insufficient prediction accuracy. Second, existing cross-domain monitoring methods lack an effective domain alignment mechanism when faced with distribution differences between the source domain and the target domain, making it difficult to adapt to scenarios with limited sample quantities or large variations in working conditions, and the model's robustness and generalization ability are limited. In addition, when dealing with high-frequency time-frequency information, existing methods often ignore the periodicity and dynamic characteristics of the signals, leading to insufficient feature expression ability and difficulty in meeting the wear monitoring requirements under changing working conditions. Summary of the Invention

[0005] The present invention provides a cross-domain oriented tool wear monitoring method and system, aiming to solve the problems of insufficient model generalization ability caused by distribution differences between domains in cross-domain scenarios, and insufficient extraction of high-frequency and non-linear time-frequency characteristics of tool wear signals, and is applicable to industrial applications under complex working conditions.

[0006] A cross-domain oriented tool wear monitoring method includes: (1) Collecting labeled vibration signals in the source domain and unlabeled vibration signals in the target domain during the tool processing and performing preprocessing; (2) Inputting the preprocessed signals into a dual-domain wavelet convolutional layer, performing time-frequency decomposition using a Morlet wavelet kernel and introducing a smoothing control term to obtain labeled source domain features and unlabeled target domain features ; (3) Enhancing the high-frequency energy response of the source domain features and target domain features obtained in step (2) to obtain source domain enhanced features and target domain enhanced features ; (4) Introducing an echo difference alignment module to explicitly align the target domain enhanced features with the source domain enhanced features to obtain aligned features of the target domain ; (5) Introducing a sliding window time-frequency attention module to obtain high-frequency focused features based on the aligned features of the target domain , and fusing the high-frequency focused features with the aligned features of the target domain to form dynamic aligned features of the target domain ; (6) Concatenating the source domain enhanced features and the dynamic aligned features of the target domain in the channel dimension to obtain fused features ; (7) Input the fused features into the shared convolutional network to extract the spatial features of the worn area ; Then, use the spatial features as the input of the gated recurrent network, extract the hidden state at the last time step of the gated recurrent network as the spatio-temporal fusion representation, input it into the fully connected network for regression prediction, and output the tool wear value ; (8) Combine the root mean square error loss and the improved MMD loss to construct the total loss function, and realize the collaborative training of feature alignment and wear prediction; After training, input the vibration signals under different working conditions as the target domain for tool wear monitoring.

[0007] In the present invention, by combining an adaptive gain module (introducing a smoothing control term) during the initialization of the wavelet kernel, physical interpretability is introduced and gradient vanishing and explosion are effectively prevented. In addition, a bat-inspired attention mechanism is constructed, which combines energy guidance, echo alignment, and time-frequency attention modules to realize the dynamic enhancement of wear features and cross-domain feature alignment. At the same time, the improved MMD loss fuses the working condition information, effectively suppresses the distribution difference between the source domain and the target domain, and significantly improves the model generalization ability and prediction accuracy in cross-domain scenarios. The present invention can adapt to industrial environments with scarce samples or variable working conditions, and meet the requirements of modern industrial manufacturing for high-precision and robust tool wear monitoring.

[0008] In step (2), input the preprocessed signal into the dual-domain wavelet convolutional layer, perform time-frequency decomposition using the Morlet wavelet kernel and introduce a smoothing control term. The formula is as follows: ; In the formula, is the time variable, is the scale parameter, is the translation parameter, is the center frequency, is a small positive number, is the normalization constant, represents the smoothing control term, is the normalization parameter, is the regularization coefficient.

[0009] In step (3), the processes of enhancing the high-frequency energy response of the source domain features and the target domain features are the same; among them, the process of enhancing the high-frequency energy response of the source domain features is as follows: Calculate the average energy of each channel in the source domain feature , and quantify the intensity of the high-frequency subbands; For each channel, the average energy Normalize to [0, 1] to obtain the normalized energy ; Utilize the normalized energy to dynamically generate the high-frequency channel weight coefficients ; Calculate the source domain enhanced features , and the formula is as follows: ; In the formula, represents the amplification factor.

[0010] In step (4), obtain the aligned features of the target domain , and the formula is as follows: ; In the formula, represents the alignment adaptive weight, and the formula is: ; In the formula, is the Sigmoid activation function, is the amplitude adjustment coefficient, is the one-dimensional convolution operation; represents the alignment modulation input, and the formula is: ; In the formula, represents the absolute difference in high-frequency intensity between the source domain and the target domain, represents the periodic time series encoding.

[0011] The specific process of step (5) is as follows: First, obtain the query, key, and value triples of the aligned features of the target domain ; Calculate the attention weights , and the formula is as follows: ; In the formula, is the scaling factor, is the local attention operator, and at each time step, it only performs attention calculation with the positions within the range of its previous and next step lengths, represents the window size; Apply the attention weights to the value vector to obtain the locally enhanced high-frequency focused features : ; Finally, use the high-frequency focused features Alignment features with the target domain are fused to form the dynamic alignment features of the target domain .

[0012] In step (7), the shared convolutional network uses a two-layer convolutional structure, and each layer contains convolutional, batch normalization, non-linear activation, and pooling operations.

[0013] In step (8), the gated recurrent network updates the state as follows: ; wherein, is the hidden state at the current moment, is the hidden state at the previous moment, is the update gate, is the candidate hidden state at the current moment.

[0014] In step (8), the formula for the total loss function is: ; wherein, is the total loss function, is the root mean square error loss, is the weight coefficient of the improved MMD loss, is the improved MMD loss.

[0015] The formula for the improved MMD loss is: ; wherein, the feature similarity and the weight formula are as follows: ; ; wherein, , represents the source domain augmented feature corresponding to the th source domain sample; represents the dynamic alignment feature corresponding to the th target domain sample; and are the number of samples in the source domain and the target domain respectively, is the high-dimensional feature mapping; represents the process parameter corresponding to the th source domain sample, represents the process parameter corresponding to the th target domain sample, , , respectively represent the encodings of the spindle speed, feed rate, and material type, is the scaling hyperparameter.

[0016] A cross - domain tool wear monitoring system includes a memory and one or more processors. Executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement the above - mentioned cross - domain tool wear monitoring method.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention adopts a physical prior - driven wavelet convolution kernel initialization method, which effectively enhances the time - frequency feature extraction ability of vibration signals, accurately captures the high - frequency dynamic changes caused by tool wear. Compared with traditional methods, it has stronger physical interpretability and cross - domain adaptability, significantly reduces the dependence on complex artificial feature engineering, realizes efficient and lightweight feature modeling, and improves the accuracy and robustness of cross - working - condition monitoring.

[0018] 2. The present invention constructs a bat - bionic attention mechanism, simulates the echolocation process, focuses on the high - frequency signal regions sensitive to wear, and improves the transferability and anti - interference ability of features. By dynamically adjusting the attention weights through energy enhancement, echo alignment, and sliding window time - frequency attention modules, it solves the problem that the existing attention mechanism has insufficient response to high - frequency wear features, reduces the false detection rate in complex scenarios, and meets the requirements of industrial - level precision monitoring.

[0019] 3. The present invention proposes a weighted MMD loss method that fuses process and material parameters to achieve explicit alignment of feature distributions between the source domain and the target domain. By introducing process information such as cutting depth, feed rate, and material type for weight adjustment, it dynamically reflects the influence of working - condition differences on feature distribution offset, improves the prediction stability and accuracy on the unlabeled target domain, and enhances the practicality and adaptability of the method in the intelligent manufacturing scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flowchart of a cross - domain tool wear monitoring method according to an embodiment of the present invention.

[0021] Figure 2 It is the overall framework diagram of the present invention.

[0022] Figure 3 It is a diagram of the initialization result of the wavelet kernel in an embodiment of the present invention.

[0023] Figure 4 It is a diagram of the tool wear recognition result in a cross - domain working condition in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] The following further describes the present invention in detail with reference to the drawings and embodiments. It should be noted that the following embodiments are intended to facilitate the understanding of the present invention, but do not limit it in any way.

[0025] AsFigure 1 and Figure 2 As shown in Figure 2 , a cross - domain tool wear monitoring method mainly includes three sub - steps, namely sub - step (a): physical - information - guided smooth wavelet kernel initialization; sub - step (b): joint dynamic feature enhancement and working - condition - guided domain alignment; sub - step (c): spatio - temporal feature fusion based on a shared network layer.

[0026] Sub - step (a): physical - information - guided smooth wavelet kernel initialization, including: Collect the labeled vibration signals in the source domain and the unlabeled vibration signals in the target domain, with the length of each signal segment being 4000. Normalize the signals to obtain the pre - processed signals.

[0027] Add Gaussian noise to the pre - processed signals. The noise intensity is dynamically adjusted according to the statistical characteristics of the signals to generate the smoothed signals.

[0028] Use the Morlet wavelet kernel to perform time - frequency decomposition on the smoothed signals. By setting different scale and shift parameters, the wavelet kernel generates a time - frequency feature matrix.

[0029] Normalize the generated time - frequency feature matrix, adjust the eigenvalue range, and generate a normalized time - frequency feature representation.

[0030] Specifically, as shown in (a) of Figure 2 , sub - step (a) specifically includes the following steps: Figure 2 Sub - step (a) specifically includes the following steps: S1: Collect the labeled vibration signals in the source domain and the unlabeled vibration signals in the target domain , with the length of each signal segment being 4000, and normalize the signals to obtain the pre - processed signals : ; where is the -th sampling point of the -dimensional signal.

[0031] S2: Use the Morlet wavelet kernel as the initial convolution kernel to perform time - frequency decomposition on the pre - processed signals . The initial form of the Morlet wavelet kernel is: ; where is the Gaussian envelope input, is the time variable, is the scale parameter, is the translation parameter, is the central frequency, is a small positive number, is the normalization constant.

[0032] S3: To address the problem of the Gaussian envelope being too narrow / too sharp, a smoothing control term is introduced , and its form is: ; In the formula, where is the normalization parameter, is the regularization coefficient, is the scale parameter.

[0033] This factor is used to dynamically adjust the sharpness of the wavelet kernel envelope term, so that more smoothing is performed when the amplitude of the input signal is small, and it remains sensitive when the amplitude is large.

[0034] S4: Incorporate into the original wavelet envelope to obtain the improved kernel function as: ; Based on the characteristic that the high-frequency components of the vibration signal increase during the tool wear process, the above steps adopt a physically informed smoothing wavelet kernel initialization method. By combining the Morlet wavelet kernel with the AGM smoothing control term, the accuracy and robustness of time-frequency feature extraction are improved, and the high-frequency transient changes induced by wear are accurately captured.

[0035] The smoothing wavelet kernel initialization method proposed in this embodiment utilizes the time-frequency local analysis ability of wavelet transform and incorporates physical prior knowledge, enabling the first-layer convolution of the network to have strong physical interpretability, overcoming the limitations of traditional feature extraction methods lacking physical basis and insufficient cross-domain adaptability. The AGM smoothing control term optimizes the Gaussian envelope shape, avoids gradient vanishing and explosion, reduces the dependence on complex feature engineering, and realizes efficient and lightweight feature extraction. The extracted time-frequency feature matrix comprehensively reflects the dynamic characteristics of tool wear, ensuring the accuracy and stability of cross-domain monitoring.

[0036] Sub-step (b): Combine dynamic feature enhancement and working condition-guided domain alignment, including: According to the time-frequency feature matrix generated by sub-step (a), which is the initial feature representation of the source domain and the target domain, and serves as the input to the subsequent feature enhancement and alignment module.

[0037] Enhance the high-frequency energy response of the dual-domain features. By calculating the high-frequency energy distribution of the features, adjust the feature weights, enhance the high-frequency components related to tool wear, and obtain the enhanced features.

[0038] Simulate the echolocation characteristics of bats, calculate the echo difference between the source domain and target domain features, optimize the time-frequency response consistency of the features, and generate the aligned features of the target domain.

[0039] Next, based on the sliding window time-frequency attention mechanism, local attention weighting is performed on the aligned features in the target domain. By calculating the attention weights of the query, key, and value, the high-frequency segments sensitive to wear are focused, and an attention-enhanced feature representation is generated.

[0040] Finally, the working condition parameters (such as material hardness, cutting speed) are embedded into the feature representation, and the weighted MMD loss between the source domain enhanced features and the aligned target domain features is calculated.

[0041] Specifically, as Figure 2 shown in (b) of S1: The source domain features and target domain features extracted by the wavelet convolutional layer are respectively expressed as: ; ; In the formula, B is the batch size.

[0042] It should be noted that the source domain and target domain features adopt the same processing process. Subsequently, taking the high-frequency enhancement of the source domain features as an example.

[0043] S2: Calculate the average energy E of each channel, quantify the intensity of the high-frequency subband, and simulate the intensity of the high-frequency echo measured by the bat cochlea: ; In the formula, , representing the sequence length.

[0044] To eliminate the amplitude difference between working conditions, the channel energy is normalized to [0,1] to highlight the relative energy distribution. The normalized energy is defined as: ; In the formula, represents the maximum energy of the channel within the batch, .

[0045] Furthermore, weights are dynamically generated to amplify the high-frequency channels. The high-frequency channel weights are generated as: ; ; In the formula, is the channel convolution response, is the weight coefficient, is the bias term, controls the activation, realizes the normalization clipping.

[0046] The source domain enhanced features with direction awareness and energy amplification after enhancement are expressed as follows: ; Wherein, , represents the amplification factor.

[0047] S3: To explicitly align the high-frequency feature differences between the source / target domains, the high-frequency amplitude differences under the L2 norm were calculated. Denote the source / target high-frequency subband energies as and , defined as follows: ; ; Wherein, , represents the L2 norm along the time dimension.

[0048] Then, calculate the absolute difference of the high-frequency intensities between domains to characterize the high-frequency feature shift between domains: ; Introduce periodic coding as the timing perception to enhance the timing stability of the difference information, and the formula is: ; Wherein, is the fundamental frequency.

[0049] Fuse the inter-domain intensity difference with the periodic timing coding to obtain the alignment modulation input : ; Then, generate the alignment adaptive weight through convolution and non-linear activation: ; Wherein, is the Sigmoid activation function, is the amplitude adjustment coefficient, is the one-dimensional convolution operation.

[0050] Finally, perform a regional alignment operation on the target domain high-frequency features to obtain the aligned features of the target domain: .

[0051] S4: To achieve the interaction and modeling of cross-domain high-frequency information, a local window attention mechanism is further introduced. First, obtain the query, key, and value triples through convolution: ; ; ; Within each sliding window (window length is 𝑊), calculate the dot product correlation of Q and K, and perform local normalization. Define the local attention operator , and the overall expression is as follows: ; where is the scaling factor, is the local attention operator, and at each time step, attention calculation is only performed with positions within the range of the previous and next W / 2 steps

[0052] Apply the attention weight to the value vector to obtain the locally enhanced high-frequency focused feature: ; Finally, fuse the high-frequency focused feature with the aligned feature of the target domain through residual fusion to form the final output of the module, that is, the dynamic alignment feature of the target domain : .

[0053] S5: The classical maximum mean discrepancy (MMD) is defined as: ; In the formula, and , and are the number of samples in the source domain and the target domain respectively, is the high-dimensional feature mapping, is the reproducing kernel Hilbert space

[0054] For the source domain sample 𝑖 and the target domain sample 𝑗, define their process parameters as and respectively. The process similarity is calculated as follows: ; where , , , represent the encodings of the spindle speed, feed rate, and material type respectively, is the scaling hyperparameter

[0055] Next, convert the similarity to a weighting factor between [0,1] through the Sigmoid activation function: ; Then, based on the weights, calculate the pairwise weighted MMD loss for the source domain and target domain samples in the feature space: ; Based on the law that the high-frequency signals in the tool wear process are more sensitive to the wear state, in sub-step (b), a bionic mechanism is used to simulate the direction perception ability of the bat cochlea to high-frequency vibrations, so as to achieve efficient focusing and temporal alignment of key wear characteristics. This method combines an energy-guided sub-band enhancement module and an echo difference alignment module based on periodic position encoding to improve the modeling ability of the temporal structure consistency of source / target domain features. Through the local window attention mechanism, accurately capture the time-frequency change features related to wear in the frequency band range, and effectively alleviate the problems of temporal misalignment and response ambiguity existing in traditional feature alignment. Further combined with the process perception weighted MMD loss function guided by processing parameters, introduce the physical working condition similarity information in the feature distribution alignment to improve the robustness of key feature migration under different working conditions.

[0056] Sub-step (c): Spatiotemporal feature fusion based on a shared network layer, including: First, concatenate the source domain enhanced features and the dynamically aligned features of the target domain in the channel dimension to form a unified fused input feature; Use a shared convolutional neural network to extract spatial local features and enhance the recognition ability of local wear patterns; Input the convolutional output into a gated recurrent neural network to extract the wear evolution trend changing with the working condition in the time dimension; Use a fully connected neural network to complete the regression prediction of the tool wear value and output the wear amount estimation result; Construct a joint regression and weighted MMD loss function, and jointly optimize the network parameters by combining the supervision error and the distribution alignment effect to achieve robust modeling under different domain conditions.

[0057] Specifically, as shown in (c) below, sub-step (c) specifically includes the following steps: Figure 2 S1: Concatenate the source domain high-frequency enhanced features obtained in sub-step (b) and the aligned target domain features in the channel dimension to obtain the fused feature Denoted as: ; S2: The fused feature is input into a shared convolutional module for extracting the spatial context pattern of the wear area. Use a two-layer convolutional structure, each layer containing convolution, batch normalization, non-linear activation, and pooling operations: ; S3: Use the convolved feature as the input of the GRU to extract the wear dynamic features changing with time. The GRU updates the state as follows:​ ; In the formula, is the hidden state at the current moment, is the hidden state at the previous moment, is the update gate, is the candidate hidden state at the current moment.

[0058] S4: Extract the hidden state of the last time step as the spatio-temporal fusion representation, input it into the fully connected network for regression prediction, and output the tool wear value : ; ; In the formula, represents the intermediate feature, is the weight matrix, is the bias term.

[0059] S5: Construct a multi-objective loss function, comprehensively consider the regression prediction accuracy and the alignment of feature distributions, and define the total loss as: ; In the formula, is the weight coefficient of the improved MMD loss.

[0060] Based on the spatio-temporal evolution characteristics during the tool wear process, the above steps adopt a spatio-temporal feature fusion method with a shared network layer. By splicing the source domain enhanced features and the target domain aligned features, local spatial features and time dynamic information are extracted using shared convolution and recurrent neural networks. This method combines the pooling compression and sequence modeling mechanisms to improve the continuity and stability of wear feature extraction. A joint loss is constructed by combining the regression error and the weighted MMD to enhance the prediction accuracy and distribution alignment ability under different working conditions, and achieve a high-precision and robust estimation of the tool wear value.

[0061] To verify the feasibility of the methods proposed in sub-steps (a), (b) and (c), the tool wear milling dataset provided by NASA was selected as the research object. This dataset covers the milling experiment data under 16 different working conditions from Case1 to Case16, and records the multi-channel signals and the tool wear value (VB) during each feed. The data preprocessing stage includes removing the samples with missing values, 0 values and repeated VB. It should be noted that after processing, Case6 only contains one sample, which cannot meet the analysis requirements, so it is excluded. Finally, 15 valid cases are retained as the experimental data.

[0062] Before feature extraction, the acceleration signals of two channels in the dataset are selected as the source of monitoring signals. To ensure data quality and uniformity, 4000 sampling points in the steady-state processing stage are extracted from each sample as the analysis object, and the signals are normalized through S1 in sub-step (a) to eliminate the influence of signal amplitude differences in different channels and working conditions on subsequent modeling.

[0063] In the experimental design, the "case-by-case testing" strategy is adopted for evaluation. That is, each time a case is selected as the target domain test set, and the remaining 14 cases are used as the source domain training set to verify the generalization ability of the method under cross-working condition. To prevent the leakage of target domain knowledge and ensure the fairness of the experiment, when a case is selected as the test set, other cases with the same working condition settings will be excluded from the training set, so as to more realistically simulate the working condition migration problem in the actual application scenario.

[0064] Next, according to S2 and S3 in sub-step (a), the wavelet kernel is initialized for the normalized signal. The Morlet wavelet function is used in combination with the AGM smoothing control term to generate an initial convolution kernel with adjustable scale and frequency characteristics.

[0065] Figure 3 The initialization result of the wavelet kernel is shown, with a smooth shape and a clear modulation structure. The initialization guided by physical information avoids the problems of gradient disappearance and explosion during the training process. At the same time, the efficient extraction of initial features reduces the feature offset between the source domain and the target domain.

[0066] Then, according to S1-S4 in sub-step (b), a cross-domain dynamic alignment mechanism between the source domain and the target domain is constructed. First, based on S1, the L2 energy difference of the features in the two domains in the high-frequency region is calculated to capture the frequency domain offset. Subsequently, through S2, periodic encoding is introduced to enhance the temporal consistency and generate a dynamic modulation factor.

[0067] According to S3, combined with local convolution and sliding window attention mechanism, the key difference features in the alignment region are extracted. Finally, based on S4, the weighted MMD loss is used to achieve the explicit alignment of high-frequency features, alleviating the feature distribution offset between different working conditions.

[0068] The enhanced source domain features and the aligned features of the target domain are uniformly input into the shared network for modeling. Based on S1-S5 in sub-step (c), the convolutional layer extracts local spatial features, and the GRU unit captures the temporal information of wear evolution. Finally, the tool wear value is output through the fully connected regression module.

[0069] Figure 4The tool wear prediction results when Case 1 is the target domain are shown. It can be seen that the recognition curve output by the model is highly consistent with the true value in both the overall trend and the detail fluctuations, and the fitting effect is good. The predicted root mean square error is 0.0629 mm, which further verifies that the proposed method has high recognition accuracy and stability under cross-condition conditions, and reflects its good modeling ability and generalization ability for the tool wear evolution process in complex domain migration scenarios.

[0070] The above-described embodiments have detailed 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 used to limit the present invention. Any modifications, supplements, and equivalent replacements made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A tool wear monitoring method for cross - domain, characterized in that, Including: (1) Collecting the labeled vibration signals in the source domain and the unlabeled vibration signals in the target domain during the tool machining process and performing preprocessing; (2) Input the preprocessed signal into the dual-domain wavelet convolutional layer, perform time-frequency decomposition using the Morlet wavelet kernel and introduce a smoothing control term to obtain labeled source domain features and unlabeled target domain features ; (3) Perform high-frequency energy response enhancement on the source domain features and target domain features obtained in step (2) to obtain source domain enhanced features and target domain enhanced features ; (4) Introduce an echo difference alignment module to explicitly align the enhanced features of the target domain with the enhanced features of the source domain to obtain the aligned features of the target domain ; (5) Introduce a sliding window time-frequency attention module, and based on the aligned features of the target domain Obtain high-frequency focused features , and use the high-frequency focused features to fuse with the aligned features of the target domain to form the dynamic aligned features of the target domain ; (6) Concatenate the source domain enhanced features with the dynamic alignment features of the target domain in the channel dimension to obtain the fused features ; (7) Input the fused features into the shared convolutional network to extract the spatial features of the worn area ; then use the spatial features as the input of the gated recurrent network, extract the hidden state of the last time step of the gated recurrent network as the spatio-temporal fusion representation, input it into the fully connected network for regression prediction, and output the tool wear value ; (8) Constructing a total loss function by combining the root mean square error loss and the improved MMD loss to achieve the collaborative training of feature alignment and wear prediction; After the training is completed, input the vibration signals under different working conditions as the target domain for tool wear monitoring.

2. The cross-domain oriented tool wear monitoring method according to claim 1, characterized in that In step (2), the preprocessed signal is input into the dual-domain wavelet convolutional layer, and the Morlet wavelet kernel is used for time-frequency decomposition and a smoothing control term is introduced. The formula is as follows: ； wherein, is the time variable, is the scale parameter, is the translation parameter, is the center frequency, is a small positive number, is the normalization constant, represents the smoothing control term, is the normalization parameter, is the regularization coefficient.

3. The tool wear monitoring method for cross-domain according to claim 1, characterized in that In step (3), the processes of enhancing the high-frequency energy response of the source domain features and the target domain features are the same; among them, the process of enhancing the high-frequency energy response of the source domain features is: Calculate the source domain features The average energy of each channel , and quantify the intensity of the high-frequency subbands; The average energy of each channel is normalized to [0, 1] to obtain the normalized energy ; Using normalized energy Dynamically generate high-frequency channel weight coefficients ; Calculate the source domain enhanced features , the formula is as follows: ； wherein, denotes the magnification factor.

4. The tool wear monitoring method for cross-domain according to claim 1, characterized in that In step (4), the aligned features of the target domain are obtained , and the formula is as follows: ； In the formula, represents the alignment adaptive weight, and the formula is: ； In the formula, is the Sigmoid activation function, is the amplitude adjustment coefficient, is the one-dimensional convolution operation; represents the alignment modulation input, and the formula is: ； In the formula, represents the absolute difference in high-frequency intensity between the source domain and the target domain, represents periodic time-series coding.

5. The tool wear monitoring method for cross-domain according to claim 1, wherein The specific process of step (5) is: First, obtain the aligned features of the target domain through convolution query, key, and value triples ; Calculate the attention weights , the formula is as follows: ； In the formula, is the scaling factor, is the local attention operator, which only performs attention calculation with the positions within the range of its previous and next step lengths at each time step, represents the window size; Apply attention weights to the value vector to obtain locally enhanced high-frequency focused features : ； Finally, fuse the high-frequency focus features with the alignment features of the target domain to form the dynamic alignment features of the target domain .

6. The tool wear monitoring method for cross-domain according to claim 1, characterized in that In step (7), the shared convolutional network uses a two-layer convolutional structure, and each layer includes convolution, batch normalization, non-linear activation, and pooling operations.

7. The tool wear monitoring method for cross-domain according to claim 1, characterized in that In step (8), the state of the gated recurrent network is updated as follows: ； wherein, is the hidden state at the current moment, is the hidden state at the previous moment, is the update gate, is the candidate hidden state at the current moment.

8. The tool wear monitoring method for cross-domain according to claim 1, characterized in that In step (8), the formula of the total loss function is: ； In the formula, is the total loss function, is the root mean square error loss, is the weight coefficient of the improved MMD loss, is the improved MMD loss.

9. The tool wear monitoring method for cross-domain according to claim 8, characterized in that The formula of the improved MMD loss is: ； In the formula, the feature similarity and the weight formula are as follows: ； ； In the formula, represents the source domain enhanced feature corresponding to the -th source domain sample; represents the dynamic alignment feature corresponding to the -th target domain sample; and are the numbers of samples in the source domain and the target domain respectively, is the high-dimensional feature mapping; represents the process parameter corresponding to the -th source domain sample, represents the process parameter corresponding to the target domain sample of ; , , , represent the encodings of the spindle speed, feed rate, and material type respectively, is the scaling hyperparameter.

10. A tool wear monitoring system for cross-domain, characterized in that, Including a memory and one or more processors, where executable code is stored in the memory, and when the one or more processors execute the executable code, it is used to implement the cross-domain tool wear monitoring method according to any one of claims 1-9.

Images