A cross-domain tool wear monitoring method and system

Abstract

The present invention discloses a cross-domain tool wear monitoring method and system. The method includes: using physically guided wavelet kernel initialization to perform time-frequency decomposition of source domain and target domain vibration signals to extract high-frequency features with physical interpretability; constructing a bat bionic attention mechanism, combining energy guidance, echo alignment and time-frequency attention modules to achieve dynamic enhancement of wear features and cross-domain feature alignment; after splicing the enhanced features of the source domain and target domain, input them into a shared convolution and recurrent network to extract spatiotemporal fusion features; constructing an improved MMD loss function that integrates the similarity of processing parameters, jointly optimizing the regression error, and realizing collaborative training of feature alignment and wear prediction. The present invention can meet the high-precision and strong robustness monitoring tasks of tool wear under different working conditions in industrial sites.

Description

Technical Field

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

[0002] Tool wear is a key factor affecting machining accuracy, production efficiency, and product quality during machining. Especially in the later stages of tool wear, the wear rate increases dramatically, and the signal exhibits significant nonlinear and high-frequency characteristics. Failure to accurately monitor tool wear status directly impacts the timing of tool changes, leading to machining interruptions, reduced workpiece quality, and increased production costs. In industrial scenarios with multiple working conditions and tools, the diversity of tool types, machining materials, and working conditions further exacerbates the complexity of wear monitoring. In particular, in cross-domain scenarios, significant sample distribution differences exist between the source domain (e.g., laboratory environment) and the target domain (e.g., actual production environment). Traditional monitoring methods struggle to effectively adapt to this domain shift, resulting in insufficient model generalization and difficulty meeting the high-precision and robust monitoring requirements of industrial sites. Therefore, there is an urgent need to develop a monitoring method that can accurately extract time-frequency features of tool wear, achieve cross-domain feature alignment, and improve model robustness to address complex working conditions and cross-domain challenges.

[0003] Existing tool wear monitoring technologies primarily rely on signal processing and machine learning methods to extract wear features and predict status. However, traditional methods are often limited to simple time-domain or frequency-domain analysis when processing the high-frequency and nonlinear characteristics of tool wear signals. This makes it difficult to fully capture the complex time-frequency dynamics of the later stages of wear, resulting in insufficient feature expression capabilities. In cross-domain scenarios, Chinese patent publication CN115351601A assumes that the data distribution of the source and target domains is consistent and lacks an effective domain alignment mechanism. When target domain samples are scarce or operating conditions differ significantly, model performance significantly degrades. Furthermore, Chinese patent publications CN117001420A and CN112801139A, respectively, enhance feature expression capabilities by introducing attention mechanisms and feature fusion. However, these methods lack precise control over high-frequency time-frequency information and struggle to effectively adapt to the distribution differences between the source and target domains in cross-domain scenarios, limiting their generalization and robustness in real-world industrial environments.

[0004] Although existing tool wear monitoring technologies have made certain progress, they still face significant challenges. First, traditional feature extraction methods are difficult to accurately capture the high-frequency and nonlinear time-frequency characteristics of tool wear signals. Especially under complex working conditions in the late stage of wear, the extracted features cannot fully reflect the dynamic changes in wear behavior, resulting in insufficient prediction accuracy. Secondly, when faced with distribution differences between the source domain and the target domain, existing cross-domain monitoring methods lack an effective domain alignment mechanism, making it difficult to adapt to scenarios with limited sample numbers or large changes in working conditions, and the model robustness and generalization capabilities are limited. In addition, when processing high-frequency time-frequency information, existing methods often ignore the periodicity and dynamic characteristics of the signal, resulting in insufficient feature expression capabilities and difficulty in meeting the wear monitoring needs under variable working conditions. Summary of the Invention

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

[0006] A cross-domain tool wear monitoring method, comprising:

[0007] (1) Collect the labeled vibration signals in the source domain and the unlabeled vibration signals in the target domain during the tool processing and perform preprocessing;

[0008] (2) The preprocessed signal is input into the dual-domain wavelet convolution layer, the Morlet wavelet kernel is used for time-frequency decomposition and a smoothing control term is introduced to obtain the labeled source domain features. and unlabeled target domain features ;

[0009] (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 enhancement features ;

[0010] (4) Introduce the echo difference alignment module to enhance the features of the target domain Enhanced features from the source domain Perform explicit alignment to obtain the alignment features of the target domain ;

[0011] (5) Introducing the sliding window time-frequency attention module based on the alignment features of the target domain Obtain high-frequency focusing features , and focus the high frequency features Alignment features with the target domain Fusion is performed to form dynamic alignment features of the target domain ;

[0012] (6) Enhance the source domain features Dynamic alignment features with the target domain Perform channel dimension splicing to obtain fusion features ;

[0013] (7) Fusion features Input the shared convolutional network to extract the spatial features of the wear area ; Then the spatial features As the input of the gated recurrent network, the hidden state of the last time step of the gated recurrent network is extracted as the spatiotemporal fusion representation, input into the fully connected network for regression prediction, and output the tool wear value ;

[0014] (8) The total loss function is constructed by combining the root mean square error loss and the improved MMD loss to achieve collaborative training of feature alignment and wear prediction; after the training is completed, the vibration signals of different working conditions are input as the target domain to monitor tool wear.

[0015] By incorporating an adaptive gain module (introducing a smoothing control term) into wavelet kernel initialization, this method introduces physical interpretability and effectively prevents gradient vanishing and explosion. Furthermore, a bat-inspired attention mechanism is constructed, combining energy guidance, echo alignment, and time-frequency attention modules to achieve dynamic enhancement of wear characteristics and cross-domain feature alignment. Simultaneously, an improved MMD loss is used to integrate working condition information, effectively suppressing distribution differences between the source and target domains and significantly improving model generalization and prediction accuracy in cross-domain scenarios. This method is adaptable to industrial environments where samples are scarce or working conditions are variable, meeting the needs of modern industrial manufacturing for high-precision, robust tool wear monitoring.

[0016] In step (2), the preprocessed signal is input into the dual-domain wavelet convolution layer, the Morlet wavelet kernel is used for time-frequency decomposition and a smoothing control term is introduced. The formula is as follows:

[0017] ;

[0018] Where, 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.

[0019] In step (3), the process of enhancing the high-frequency energy response of the source domain features and the target domain features is the same; wherein, the process of enhancing the high-frequency energy response of the source domain features is:

[0020] Calculate source domain characteristics The average energy of each channel , quantizes the strength of the high-frequency sub-band;

[0021] The average energy of each channel Normalized to [0,1], get normalized energy ;

[0022] Using normalized energy Dynamically generate high-frequency channel weight coefficients ;

[0023] Compute source domain enhancement features , the formula is as follows:

[0024] ;

[0025] Where, , represents the amplification factor.

[0026] In step (4), the alignment features of the target domain are obtained , the formula is as follows:

[0027] ;

[0028] Where, represents the alignment adaptive weight, and the formula is:

[0029] ;

[0030] Where, is the Sigmoid activation function, is the amplitude adjustment coefficient, It is a one-dimensional convolution operation; represents the aligned modulation input, and the formula is:

[0031] ;

[0032] Where, represents the absolute difference in high-frequency intensity between the source domain and the target domain, Represents periodic timing coding.

[0033] The specific process of step (5) is:

[0034] First, the alignment features of the target domain are obtained through convolution The query, key, value triple ;

[0035] Calculating attention weights , the formula is as follows:

[0036] ;

[0037] Where, is the scaling factor, is a local attention operator, where each time step is only concerned with its predecessor and successor Attention is calculated for the positions within the step range. Indicates the window size;

[0038] The attention weight Applied to a vector of values , obtain locally enhanced high-frequency focusing features :

[0039] ;

[0040] Finally, the high-frequency focus features Alignment features with the target domain Fusion is performed to form dynamic alignment features of the target domain .

[0041] In step (7), the shared convolutional network uses a two-layer convolutional structure, each layer contains convolution, batch normalization, nonlinear activation and pooling operations.

[0042] In step (8), the gated recurrent network updates its state as follows:

[0043] ;

[0044] Where, is the hidden state at the current moment, is the hidden state at the previous moment, It is the update gate. is the candidate hidden state at the current moment.

[0045] In step (8), the formula of the total loss function is:

[0046] ;

[0047] Where, is the total loss function, is the root mean square error loss, To improve the weight coefficient of MMD loss, To improve the MMD loss.

[0048] The formula for improving MMD loss is:

[0049] ;

[0050] In the formula, feature similarity And weight formula The formula is:

[0051] ;

[0052] ;

[0053] Where, , indicating the Source domain enhancement features corresponding to source domain samples; Indicates the Dynamic alignment features corresponding to target domain samples; and are the number of samples in the source domain and the target domain, is a high-dimensional feature map; Indicates the The process parameters corresponding to the source domain samples, Indicates the The process parameters corresponding to the target domain samples, ,, 、 The codes for spindle speed, feed rate and material type respectively. is the scaling hyperparameter.

[0054] A cross-domain 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 cross-domain tool wear monitoring method.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] 1. This invention utilizes a wavelet convolution kernel initialization method driven by physical priors to effectively enhance the ability to extract time-frequency features from vibration signals, accurately capturing the high-frequency dynamic changes caused by tool wear. Compared to traditional methods, this method offers greater physical interpretability and cross-domain adaptability, significantly reducing reliance on complex manual feature engineering. It achieves efficient and lightweight feature modeling, improving the accuracy and robustness of cross-condition monitoring.

[0057] 2. This invention employs a bat-inspired attention mechanism, simulating the echolocation process and focusing on wear-sensitive high-frequency signal regions, improving feature transferability and anti-interference capabilities. Through energy enhancement, echo alignment, and a sliding window time-frequency attention module, attention weights are dynamically adjusted. This addresses the issue of existing attention mechanisms being insufficiently responsive to high-frequency wear signatures, reduces false positives in complex scenarios, and meets the demands of industrial-grade precision monitoring.

[0058] 3. This paper proposes a weighted MMD loss method that integrates process and material parameters to achieve explicit alignment of feature distributions between the source and target domains. By incorporating process information such as cutting depth, feed rate, and material type for weight adjustment, it dynamically reflects the impact of operating conditions on feature distribution shifts, improving prediction stability and accuracy in unlabeled target domains, and enhancing the method's practicality and adaptability in intelligent manufacturing scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a flow chart of a cross-domain tool wear monitoring method according to an embodiment of the present invention.

[0060] Figure 2 It is the overall framework diagram of the present invention.

[0061] Figure 3 Graph showing the initialization result of the wavelet kernel in an embodiment of the present invention.

[0062] Figure 4 This is a diagram of tool wear identification results in a cross-domain working condition in an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The present invention will be described in further detail below with reference to the accompanying drawings and examples. It should be noted that the following examples are intended to facilitate understanding of the present invention and do not have any limiting effect on the present invention.

[0064] like Figure 1 and Figure 2 As shown in the figure, 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): spatiotemporal feature fusion based on shared network layer.

[0065] Sub-step (a): Physical information guided smoothing wavelet kernel initialization, including:

[0066] The source domain marked vibration signal and the target domain unmarked vibration signal are collected, and the length of each signal segment is 4000. The signals are normalized to obtain the preprocessed signals.

[0067] Adds Gaussian noise to the preprocessed signal. The noise intensity is dynamically adjusted based on the statistical properties of the signal, resulting in a smoothed signal.

[0068] The Morlet wavelet kernel is used to perform time-frequency decomposition on the smoothed signal. The wavelet kernel generates a time-frequency feature matrix by setting different scale and offset parameters.

[0069] The generated time-frequency feature matrix is standardized, the eigenvalue range is adjusted, and a standardized time-frequency feature representation is generated.

[0070] Specifically, such as Figure 2 As shown in (a), sub-step (a) specifically includes the following steps:

[0071] S1: Collecting labeled vibration signals in the source domain and unlabeled vibration signals in the target domain , the length of each signal is 4000, the signal is normalized to obtain the preprocessed signal :

[0072] ;

[0073] Where, For the dimensional signal sampling points.

[0074] S2: Use Morlet wavelet kernel as the initial convolution kernel to preprocess the signal Perform time-frequency decomposition. The initial form of the Morlet wavelet kernel is:

[0075] ;

[0076] Where, is the Gaussian envelope input, 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.

[0077] S3: To solve the problem of Gaussian envelope being too narrow / too sharp, a smoothing control term is introduced , which has the form:

[0078] ;

[0079] In the formula, is the normalization parameter, is the regularization coefficient, is the scale parameter.

[0080] This factor is used to dynamically adjust the sharpness of the wavelet kernel envelope term, so that more smoothing is performed when the input signal amplitude is small, and sensitivity is maintained when the amplitude is large.

[0081] S4: Integrating into the original wavelet envelope, the improved kernel function is:

[0082] ;

[0083] Based on the characteristic that the high-frequency components of the vibration signal are enhanced during tool wear, the above steps adopt a physical information-guided 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.

[0084] The smoothing wavelet kernel initialization method proposed in this embodiment leverages the local time-frequency analysis capabilities of the wavelet transform and incorporates physical prior knowledge to impart strong physical interpretability to the first convolution layer of the network. This overcomes the limitations of traditional feature extraction methods, which lack physical basis and insufficient cross-domain adaptability. The AGM smoothing control term optimizes the Gaussian envelope shape, avoiding gradient vanishing and explosion, reducing reliance on complex feature engineering, and enabling 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.

[0085] Sub-step (b): Joint dynamic feature enhancement and condition guidance domain alignment, including:

[0086] The time-frequency feature matrix generated according to sub-step (a) is the initial feature representation of the source domain and the target domain, which serves as the input of the subsequent feature enhancement and alignment module.

[0087] The high-frequency energy response of the dual-domain feature is enhanced. By calculating the high-frequency energy distribution of the feature, adjusting the feature weight, and enhancing the high-frequency component related to tool wear, the enhanced feature is obtained.

[0088] The bat echolocation characteristics are simulated, the echo difference between the source domain and target domain features is calculated, the time-frequency response consistency of the features is optimized, and the aligned features of the target domain are generated.

[0089] Next, a sliding window time-frequency attention mechanism is used to perform local attention weighting on the aligned features of the target domain. By calculating the attention weights of the query, key, and value, it focuses on wear-sensitive high-frequency segments and generates attention-enhanced feature representations.

[0090] Finally, the working condition parameters (such as material hardness and cutting speed) are embedded into the feature representation, and the weighted MMD loss of the source domain enhanced features and the aligned target domain features is calculated.

[0091] Specifically, such as Figure 2 As shown in (b), sub-step (b) specifically includes the following steps:

[0092] S1: The source domain features and target domain features extracted by the wavelet convolution layer are expressed as:

[0093] ;

[0094] ;

[0095] Where, B is the batch size.

[0096] It is worth noting that the source domain and target domain features use the same processing process. The high-frequency enhancement of source domain features is taken as an example below.

[0097] S2: Calculate the average energy E of each channel, quantify the intensity of the high-frequency sub-band, and simulate the intensity of the high-frequency echo measured by the bat cochlea:

[0098] ;

[0099] Where, , which indicates the sequence length.

[0100] To eliminate the amplitude differences between working conditions, the channel energy is normalized to [0,1] to highlight the relative energy distribution. Defined as:

[0101] ;

[0102] Where, represents the maximum energy of the channel within the batch, .

[0103] Furthermore, weights are dynamically generated to amplify high-frequency channels. The high-frequency channel weights are generated as follows:

[0104] ;

[0105] ;

[0106] Where, is the channel convolution response, is the weight coefficient, is the bias term, Control activation, Implement normalized cropping.

[0107] After enhancement, the source domain has the characteristics of direction perception and energy amplification The expression is as follows:

[0108] ;

[0109] Where, , represents the amplification factor.

[0110] S3: To explicitly align the high-frequency feature differences between the source and target domains, the high-frequency amplitude difference under the L2 norm is calculated. The high-frequency subband energy of the source and target is and , defined as follows:

[0111] ;

[0112] ;

[0113] Where, , represents the L2 norm along the time dimension.

[0114] Then, the absolute difference of high-frequency intensity between domains is calculated , to characterize the high-frequency feature offset between domains:

[0115] ;

[0116] Introducing periodic coding As time series perception, the time series stability of the difference information is enhanced. The formula is:

[0117] ;

[0118] Where, is the fundamental frequency.

[0119] Fusion of inter-domain intensity differences with periodic temporal coding yields aligned modulation inputs :

[0120] ;

[0121] Then, After convolution and nonlinear activation, alignment adaptive weights are generated :

[0122] ;

[0123] Where, is the Sigmoid activation function, is the amplitude adjustment coefficient, It is a one-dimensional convolution operation.

[0124] Finally, the high-frequency features of the target domain are aligned to obtain the aligned features of the target domain. :

[0125] .

[0126] S4: To achieve cross-domain high-frequency information interaction and modeling, we further introduce a local window attention mechanism. First, we obtain the query, key, and value triples through convolution:

[0127] ;

[0128] ;

[0129] ;

[0130] In 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 , the overall expression is as follows:

[0131] ;

[0132] in, is the scaling factor, It is a local attention operator, and each time step only pays attention to the positions within W / 2 steps before and after it.

[0133] The attention weight Applied to a vector of values , obtain the locally enhanced high-frequency focusing features:

[0134] ;

[0135] Finally, the high-frequency focus features Alignment features with the target domain Perform residual fusion to form the final output of the module, which is the dynamic alignment feature of the target domain :

[0136] .

[0137] S5: The classic maximum mean difference (MMD) is defined as:

[0138] ;

[0139] Where, and , and are the number of samples in the source domain and the target domain, is a high-dimensional feature map, is the reproducing kernel Hilbert space.

[0140] For source domain samples 𝑖 and target domain samples 𝑗, their process parameters are defined as follows: and Process similarity The calculation is as follows:

[0141] ;

[0142] in, , 、 、 The codes for spindle speed, feed rate and material type respectively. is the scaling hyperparameter.

[0143] Next, the similarity Converted to a weighting factor between [0, 1] through the Sigmoid activation function:

[0144] ;

[0145] Then, based on the weights, the pairwise weighted MMD loss is calculated for the source and target domain samples in the feature space:

[0146] ;

[0147] Based on the increasing sensitivity of high-frequency signals to wear status during tool wear, sub-step (b) simulates the directional perception of high-frequency vibrations in the bat cochlea through a biomimetic mechanism, achieving efficient focusing and temporal alignment of key wear features. This method integrates an energy-guided subband enhancement module with a periodic position-encoded echo difference alignment module to enhance the ability to model the temporal structural consistency of source and target domain features. A local window attention mechanism accurately captures wear-related time-frequency variations within the frequency band, effectively alleviating the temporal misalignment and response ambiguity issues inherent in traditional feature alignment. Furthermore, combined with a process-aware weighted MMD loss function guided by machining parameters, physical working condition similarity information is incorporated into feature distribution alignment, improving the robustness of key feature migration across working conditions.

[0148] Sub-step (c): Spatiotemporal feature fusion based on the shared network layer, including:

[0149] First, the enhanced features of the source domain and the dynamic alignment features of the target domain are concatenated in the channel dimension to form a unified fusion input feature.

[0150] Use a shared convolutional neural network to extract spatial local features and enhance the ability to recognize local wear patterns;

[0151] The convolution output is fed into a gated recurrent neural network to extract the wear evolution trend along the time dimension as the working conditions change.

[0152] A fully connected neural network is used to complete the regression prediction of tool wear value and output the wear estimation result;

[0153] A joint regression and weighted MMD loss function is constructed, and the supervision error and distribution alignment effect are combined to jointly optimize the network parameters to achieve robust modeling under different domain conditions.

[0154] Specifically, such as Figure 2 As shown in (c), sub-step (c) specifically includes the following steps:

[0155] S1: Concatenate the source domain high-frequency enhancement features obtained in substep (b) with the aligned target domain feature channel dimensions to obtain the fused features express:

[0156] ;

[0157] S2: Fusion features The input shared convolution module is used to extract the spatial context pattern of the wear area. A two-layer convolutional structure is used, each layer contains convolution, batch normalization and nonlinear activation and pooling operations:

[0158] ;

[0159] S3: Use the convolutional features as the input to the GRU to extract the wear dynamic features that change over time. The GRU update status is as follows:

[0160] ;

[0161] Where, is the hidden state at the current moment, is the hidden state at the previous moment, It is the update gate. is the candidate hidden state at the current moment.

[0162] S4: Extract the hidden state of the last time step As a spatiotemporal fusion representation, the fully connected network is input for regression prediction and the tool wear value is output. :

[0163] ;

[0164] ;

[0165] Where, represents the intermediate features, is the weight matrix, is the bias term.

[0166] S5: Construct a multi-objective loss function, align the comprehensive regression prediction accuracy with the feature distribution, and define the total loss as:

[0167] ;

[0168] Where, is the weight coefficient for improving the MMD loss.

[0169] Based on the spatiotemporal evolution of tool wear, the above steps employ a spatiotemporal feature fusion method using a shared network layer. By concatenating enhanced features from the source domain with aligned features from the target domain, a shared convolutional and recurrent neural network is used to extract local spatial features and temporal dynamic information. This method integrates pooling compression with sequence modeling to improve the continuity and stability of wear feature extraction. A joint loss is constructed by combining regression error with weighted multivariate multivariate loss, enhancing prediction accuracy and distribution alignment across operating conditions, achieving high-precision and robust estimation of tool wear values.

[0170] To verify the feasibility of the methods proposed in substeps (a), (b), and (c), a tool wear milling dataset provided by NASA was selected as the research object. This dataset covers milling experimental data under 16 different working conditions, Case 1 to Case 16, recording multi-channel signals and tool wear values (VB) during each tool pass. The data preprocessing stage includes removing missing values, zero values, and samples with repeated VBs. Notably, after processing, Case 6 contains only one sample, which does not meet the analysis requirements and is therefore excluded. Ultimately, 15 valid cases were retained as experimental data.

[0171] Before feature extraction, acceleration signals from two channels in the dataset were selected as the monitoring signal source. To ensure data quality and consistency, 4,000 sampling points during the steady-state processing phase were extracted from each sample for analysis. The signals were normalized using S1 in substep (a) to eliminate the impact of signal amplitude differences across channels and operating conditions on subsequent modeling.

[0172] The experimental design employed a case-by-case testing strategy, selecting one case at a time as the target domain test set and the remaining 14 cases as the source domain training set to verify the method's generalization across operating conditions. To prevent target domain knowledge leakage and ensure experimental fairness, when a case was selected as the test set, other cases with the same operating condition were removed from the training set, thereby more realistically simulating the operating condition transfer problem encountered in real-world application scenarios.

[0173] Next, according to S2 and S3 in sub-step (a), the wavelet kernel is initialized on the normalized signal. The Morlet wavelet function is combined with the AGM smoothing control term to generate an initial convolution kernel with adjustable scale and frequency.

[0174] Figure 3 The initialization results of the wavelet kernel are shown, showing smooth morphology and clear modulation structure. Physical information-guided initialization avoids the vanishing and exploding gradient problems during training. Furthermore, efficient initial feature extraction reduces feature shift between the source and target domains.

[0175] Then, based on steps S1-S4 in substep (b), a cross-domain dynamic alignment mechanism is constructed between the source and target domains. First, based on S1, the L2 energy difference between the features in the high-frequency region of the two domains is calculated to capture the frequency domain offset. Subsequently, through S2, periodic coding is introduced to enhance temporal consistency and generate a dynamic modulation factor.

[0176] Based on S3, local convolution and sliding window attention mechanism are combined to extract key difference features of the aligned regions. Finally, based on S4, weighted MMD loss is used to achieve explicit alignment of high-frequency features to alleviate the feature distribution shift between different working conditions.

[0177] The enhanced source domain features and the aligned target domain features are fed into a shared network for modeling. Based on steps S1-S5 in substep (c), the convolutional layers extract local spatial features, the GRU units capture the temporal information of wear evolution, and the tool wear value is finally output through a fully connected regression module.

[0178] Figure 4 The tool wear prediction results for Case 1 are presented. The model outputs highly consistent identification curves with the true values in both overall trends and detailed fluctuations, demonstrating a good fit. The predicted root mean square error is 0.0629 mm, further validating the proposed method's high recognition accuracy and stability across operating conditions. This demonstrates its excellent modeling and generalization capabilities for tool wear evolution in complex domain migration scenarios.

[0179] 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 scope of protection of the present invention.

Claims

1. A cross-domain tool wear monitoring method, characterized in that: include: (1) Collect the labeled vibration signals in the source domain and the unlabeled vibration signals in the target domain during the tool processing and perform preprocessing; (2) The preprocessed signal is input into the dual-domain wavelet convolution layer, the Morlet wavelet kernel is used for time-frequency decomposition and a smoothing control term is introduced to obtain the 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 enhancement features ; (4) Introduce the echo difference alignment module to enhance the features of the target domain Enhanced features from the source domain Perform explicit alignment to obtain the alignment features of the target domain ; (5) Introducing the sliding window time-frequency attention module based on the alignment features of the target domain Obtain high-frequency focusing features , and focus the high-frequency features Alignment features with the target domain Fusion is performed to form dynamic alignment features of the target domain ; (6) Enhance the source domain features Dynamic alignment features with the target domain Perform channel dimension splicing to obtain fusion features ; (7) Fusion features Input the shared convolutional network to extract the spatial features of the wear area ; Then the spatial features As the input of the gated recurrent network, the hidden state of the last time step of the gated recurrent network is extracted as the spatiotemporal fusion representation, input into the fully connected network for regression prediction, and output the tool wear value ; (8) Combining 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 the training is completed, the vibration signals of different working conditions are input as the target domain to monitor the tool wear.

2. The cross-domain tool wear monitoring method according to claim 1, characterized in that: In step (2), the preprocessed signal is input into the dual-domain wavelet convolution layer, the Morlet wavelet kernel is used for time-frequency decomposition and a smoothing control term is introduced. The formula is as follows: ； Where, 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 cross-domain tool wear monitoring method according to claim 1, characterized in that: In step (3), the process of enhancing the high-frequency energy response of the source domain features and the target domain features is the same; wherein, the process of enhancing the high-frequency energy response of the source domain features is: Calculate source domain characteristics The average energy of each channel , quantizes the strength of the high-frequency sub-band; The average energy of each channel Normalized to [0,1], get normalized energy ; Using normalized energy Dynamically generate high-frequency channel weight coefficients ; Compute source domain enhancement features , the formula is as follows: ； Where, , represents the amplification factor.

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

5. The cross-domain tool wear monitoring method according to claim 1, characterized in that: The specific process of step (5) is: First, the alignment features of the target domain are obtained through convolution The query, key, value triple ; Calculating attention weights , the formula is as follows: ； Where, is the scaling factor, is a local attention operator, where each time step is only concerned with its predecessor and successor Attention is calculated for the positions within the step range. Indicates the window size; The attention weight Applied to a vector of values , obtain locally enhanced high-frequency focusing features : ； Finally, the high-frequency focus features Alignment features with the target domain Fusion is performed to form dynamic alignment features of the target domain .

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

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

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

9. The cross-domain tool wear monitoring method according to claim 8, characterized in that: The formula for improving MMD loss is: ； In the formula, feature similarity And weight formula The formula is: ； ； Where, , indicating the Source domain enhancement features corresponding to source domain samples; Indicates the Dynamic alignment features corresponding to target domain samples; and are the number of samples in the source domain and the target domain, is a high-dimensional feature map; Indicates the The process parameters corresponding to the source domain samples, Indicates the The process parameters corresponding to the target domain samples, , 、 、 The codes for spindle speed, feed rate and material type respectively. is the scaling hyperparameter.

10. A cross-domain tool wear monitoring system, characterized in that: It comprises a memory and one or more processors, wherein the memory stores executable code, 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 to 9.

Images