A method and system for monitoring spindle tool wear considering fiber orientation

Abstract

This invention discloses a method and system for monitoring spindle tool wear considering fiber orientation, comprising: acquiring cutting force and acceleration signals during milling, labeling fiber orientation angles and working parameters, and constructing a dataset through strong labels and monotonically interpolated weak labels; calculating a unified scaling factor to normalize cutting force and constructing an overlapping sliding window, inputting the data into a shared-time convolutional network to extract temporal features, and generating a cutting-level representation through channel modulation and mean aggregation using a FiLM module; based on the cutting-level representation, performing parallel multi-task output to achieve collaborative optimization of wear regression, physical reconstruction consistency constraints, and domain alignment; and employing a joint loss of supervised, weakly supervised, physical consistency, monotonically evolved, and domain adversarial methods, performing end-to-end training in stages of preheating, semi-supervised, and cross-domain training. Using this invention, the high-precision and robust online monitoring requirements for tool wear in CFRP machining with different fiber orientations under the same working conditions can be met.

Description

Technical Field

[0001] This invention relates to the field of machining process monitoring technology, and in particular to a method and system for monitoring spindle tool wear that takes into account fiber orientation. Background Technology

[0002] Carbon fiber reinforced polymer (CFRP) composites have been widely used in aerospace, advanced rail transportation, wind power, and high-end equipment due to their advantages such as high specific strength, high specific stiffness, excellent corrosion resistance, and significant weight reduction. Compared with metallic materials, CFRP exhibits significant anisotropy and lamination heterogeneity. Factors such as the fiber / resin interface, interlaminar bonding, and low thermal conductivity lead to more complex force and thermal field coupling during cutting, making it prone to delamination and burrs, and causing severe wear on the tool's flank and cutting edge. Especially when machining plates with different fiber orientations (e.g., 0°, ±45°, 90°), systematic differences exist in the components of cutting force, interfacial friction, abrasive wear mechanisms, local temperature rise, and material removal modes, resulting in a significant orientation-dependent evolution rate and morphology of tool wear. Achieving stable and transferable online monitoring of tool wear under multi-orientation conditions is crucial for ensuring the surface / interlaminar quality of composites, reducing tool costs and downtime for inspection, and improving equipment availability and processing consistency.

[0003] Existing publicly available technologies mostly focus on tool condition monitoring for homogeneous metals. For example, Chinese patent document CN115922442A discloses a real-time identification method for tool dullness and breakage based on spindle vibration. However, this method assumes material isotropy and signal steady state, and does not consider the signal-wear mapping offset caused by fiber orientation differences, making it difficult to apply to cross-domain monitoring scenarios involving multiple orientations in CFRP. Another example is Chinese patent document CN120190677A, which uses transfer learning and domain adaptation to address changes in operating conditions such as speed and feed. While this method has some robustness, it does not explicitly model the modulation effect of fiber orientation, a structural variable, on the relationship between cutting signals and wear. Therefore, effective methods for cross-domain wear monitoring in CFRP multi-orientation machining still lack sufficient depth.

[0004] Existing cross-domain and migration tool wear monitoring methods still have insufficient applicability in CFRP machining scenarios, mainly in the following aspects: First, they generally ignore the essential domain shift caused by fiber orientation differences and fail to explicitly model or parameterize the relationship between fiber orientation, monitoring signals, and wear states. Second, these methods mostly focus on feature alignment at the data level, lacking physical constraints consistent with cutting forces and wear mechanisms, making it difficult to guarantee the stability and generalization ability of monitoring results during cross-orientation migration. Third, targeted method designs have not yet been developed for data structures characterized by "rich window-level temporal information but sparse cutting-level wear labels," resulting in insufficient utilization of supervisory information and a lack of constraints on the monotonic evolution process of wear. Summary of the Invention

[0005] This invention provides a spindle tool wear monitoring method that considers fiber orientation, solving the problems of insufficient generalization caused by inter-domain distribution differences in cross-fiber orientation scenarios, and insufficient supervision utilization caused by rich window-level time series but sparse cutting-level labels. Under the same working conditions, it realizes online monitoring of the wear state of CFRP machining tools with different fiber orientations and has good cross-orientation generalization ability.

[0006] A method for monitoring spindle tool wear considering fiber orientation includes the following steps:

[0007] (1) Acquire multi-channel cutting signals during CFRP milling and record fiber orientation angles. Operating parameters; tool wear value measured after every 5 cuts as a strong label. Unmeasured weak labels are generated through interpolation. ;

[0008] (2) Use data containing both strong and weak labels as the source domain. The unlabeled data of the next fiber orientation is used as the target domain. ;

[0009] (3) Perform sliding window segmentation on the source domain and target domain signals to obtain the source domain window sequence. With target domain window sequence Furthermore, the cutting force of each window is standardized to eliminate amplitude differences; among them, Indicates the first Secondary cutting signal. Indicates the first A sliding window;

[0010] (4) After standardization processing and Input the shared temporal convolutional network (TCN) separately to extract window-level features from the source domain. and window-level features of the target domain ;

[0011] (5) Construct orientation coding FiLM, for and By performing feature-level modulation, the orientation modulation features of the source domain are obtained. Orientation modulation features of the target domain ;

[0012] (6) Aggregate the orientation modulation features within the same cutting operation to obtain the cutting-level embedding of the source domain. Cutting level embedding of the target domain ;

[0013] (7) Based on and Design a wear regression head, a physical reconstruction head, and a domain alignment head, respectively, to output cutting-level wear identification results. The results of the reconstructed cutting force Domain category prediction results ;

[0014] (8) Construct a system that includes supervised loss Losses due to weak supervision Physical consistency loss Monotonic constraint loss Losses in the fight against the domain The total loss function is used for end-to-end training; after training, the target domain is... The signal is processed according to steps (3)-(6), and the wear regression head outputs the tool wear value and its wear curve with the cutting sequence in step (7).

[0015] This invention first employs a shared TCN for timing encoding based on a tooth-synchronous sliding window and explicitly modulates fiber orientation using FiLM to establish an interpretable mapping between orientation, signal, and wear. Furthermore, it introduces multi-instance aggregation and a three-branch structure consisting of a wear regression head, a physical reconstruction head, and a domain alignment head to collaboratively achieve wear prediction, mechanical reconstruction, and source / target domain feature alignment. Simultaneously, it combines strong and weak supervision, physical consistency, monotonic evolution, and domain adversarial loss for phased optimization, improving cross-orientation generalization ability and prediction stability. This invention is applicable to cross-domain monitoring of tool wear under the same working conditions for CFRP with different fiber orientations, meeting the requirements of high precision, low latency, and real-time performance in composite material machining, and effectively supporting wear prediction and machining decisions under different fiber orientation conditions.

[0016] In step (3), the cutting force of each window is standardized to obtain... The formula is as follows:

[0017] ;

[0018] ;

[0019] ;

[0020] ;

[0021] In the formula, Indicates the first Secondary cutting, window Inner The average value of cutting force in each direction; Indicates the first In the cutting signal, the first The index of the starting sampling point of each sliding window; Indicates the first In the second cutting signal, the first Index of the end sampling point of each sliding window; The radial cutting depth is a parameter in the operating conditions. The axial depth of cut is a parameter in the operating conditions. This refers to the feed per tooth in the operating parameters. As an equivalent benchmark; Corresponding to the three directions of cutting.

[0022] In step (4), window-level features of the source domain are extracted. and window-level features of the target domain The formula is as follows:

[0023] ;

[0024] ;

[0025] In the formula, Indicates a shared TCN mapping. Its parameters; For window-level embedding, This is the number of feature channels output by TCN.

[0026] In step (5), the orientation modulation features of the source domain are obtained. Orientation modulation features of the target domain The formula is as follows:

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] In the formula, It is an orientation encoding vector. For the first Fiber orientation angle during the second cut For a small perceptron network, modulation parameters , It is the number of feature channels output by TCN. This is a channel-by-channel multiplication.

[0032] In step (6), the cutting-level embedding of the source domain is obtained. Cutting level embedding of the target domain The formula is as follows:

[0033] ;

[0034] ;

[0035] In the formula, , It is the number of feature channels output by TCN. It is the first The total number of windows obtained from each cut.

[0036] In step (7), the cutting-level wear identification result is output. The results of the reconstructed cutting force Domain category prediction results The formula is as follows:

[0037] Wear-return head:

[0038] ;

[0039] In the formula, express or , For weights and biases;

[0040] Physical reconstruction head:

[0041] ;

[0042] ;

[0043] In the formula, This is the cutting coefficient vector, corresponding to three components; For the first Fiber orientation angle during the second cut; It is a learnable vector; The axial depth of cut is a parameter in the operating conditions. As an equivalent benchmark;

[0044] Domain alignment header:

[0045] ;

[0046] In the formula, For the domain category probability, It is a gradient inversion layer. For weights and biases.

[0047] In step (8), a monitoring loss is constructed. Losses due to weak supervision Physical consistency loss Monotonic loss Losses in the fight against the domain The total loss function is used for end-to-end training, and the formula is as follows:

[0048] ;

[0049] In the formula, , respectively, represent the weight coefficients of each item.

[0050] Monitoring losses The formula is as follows:

[0051] ;

[0052] In the formula, For strongly labeled cutting sets; The wear value is for the strong label. The wear prediction value output by the wear regression head in step (7);

[0053] Losses due to weak supervision The formula is as follows:

[0054] ;

[0055] In the formula, For weakly labeled cutting set, The wear value for weak tags;

[0056] Physical consistency loss The formula is as follows:

[0057] ;

[0058] In the formula, For the set of window indices participating in the constraints, Total number of windows; Indicates the first Secondary cutting, window Inner Predictive power reconstructed from each direction; This represents the normalized window force statistic.

[0059] Monotonic constraint loss The formula is as follows:

[0060] ;

[0061] In the formula, It is the set of tool or sequence indices used in training. It is a sequence Number of cuts, It is a sequence No. Predicted wear during the second cut It is a monotonic relaxation constant;

[0062] Domain confrontation loss The formula is as follows:

[0063] ;

[0064] In the formula, and This is a set of source and target domain cutting-level sample indexes used for training the domain alignment head. It is the domain class probability vector output by the domain alignment header. It is a truth value label for a field. It is cross-entropy.

[0065] In step (8), a phased strategy is adopted to optimize the training process: first, a warm-up phase of supervision and physical consistency is carried out, then weak supervision and monotonic evolution are added for semi-supervised training, and finally domain adversarial training is enabled for cross-domain alignment optimization.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] 1. This invention proposes a timing coding framework based on tooth-synchronous sliding window and shared TCN, and uses FiLM to modulate fiber orientation, constructing a conditional mapping between orientation, signal, and wear. Under the same working conditions, this method achieves low-latency, lightweight online modeling. Parameter sharing effectively reduces model size and inference overhead, is compatible with multi-channel signals such as force and acceleration, reduces reliance on manual feature engineering, and facilitates rapid deployment and maintenance at the machine tool edge.

[0068] 2. This invention employs multi-instance aggregation oriented towards "numerous windows and sparse labels," coupled with a three-branch structure (wear regression head, physical reconstruction head, and domain alignment head). Under a unified scale, it introduces physical consistency and wear monotonicity constraints to stably output cutting-level wear curves. This method is robust to noise and outliers, significantly suppressing false positives and false negatives caused by spurious correlations and drift, thus meeting the real-time and reliability requirements of continuous monitoring of production lines.

[0069] 3. This invention integrates strong and weak supervision with domain adversarial optimization to achieve explicit feature alignment between different fiber orientations and improve the prediction accuracy of unlabeled target domains. Cross-orientation migration can be completed with only a few strong labels, significantly reducing labeling costs and downtime measurement frequency. It has rapid adaptation capabilities to new tools and orientations, improves cross-scenario generalization and long-term operational stability, and has good engineering usability and scalability. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 This is a flowchart illustrating a spindle tool wear monitoring method considering fiber orientation, as described in an embodiment of the present invention.

[0072] Figure 2 This is a framework diagram of a spindle tool wear monitoring method considering fiber orientation, according to an embodiment of the present invention.

[0073] Figure 3 This is a screenshot showing the result without field alignment.

[0074] Figure 4 The diagram shows the effect of using the method of the embodiment of the present invention. Detailed Implementation

[0075] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] It should be noted that, unless otherwise specified, the features in the following embodiments and implementation methods can be combined with each other.

[0077] like Figure 1As shown, a method for monitoring spindle tool wear considering fiber orientation includes the following steps:

[0078] S1: Before machining, the fiber orientation angle is marked, and parameters such as feed per tooth, radial depth of cut, and axial depth of cut are recorded. Triaxial acceleration signals are acquired using an NI device, and triaxial cutting forces are acquired using a Kistler rotary force gauge. After every 5 cuts, the tool wear value is measured using a handheld microscope as a strong label; unmeasured values ​​are interpolated to generate weak labels, thus constructing a "cutting sequence and wear value" dataset.

[0079] S2 merges strongly labeled and weakly labeled data of multiple orientations as the source domain for supervised and semi-supervised training. Unlabeled data of the next orientation is selected as the target domain, used only for cross-domain alignment and inference validation, ensuring a difference in orientation distribution between training and testing.

[0080] S3 calculates a uniform scale factor based on the working condition parameters, normalizes the cutting force of each cutting window to eliminate amplitude differences, constructs overlapping sliding windows with fixed length and step size, and generates a multi-window sequence of source and target domain signals.

[0081] S4 inputs the sliding window sequences of the two domains into the shared TCN, extracts the temporal features, characterizes the high-frequency and transient components in the signal, and outputs the window-level feature sequence.

[0082] S5 encodes the fiber orientation angle of the cut fiber and feeds it into the FiLM small perceptron network to generate channel-level scaling and bias. It modulates the window-level features channel by channel, explicitly models the relationship between orientation and signal and its impact on wear mapping, and improves the interpretability and shareability across orientations.

[0083] S6 performs mean aggregation on the modulated window features of the same cut to obtain the cut-level representation, which alleviates the mismatch of "many windows and sparse labels" and provides stable input for subsequent regression and alignment.

[0084] S7 sets up three branches in parallel on the cutting level representation: the wear regression head outputs wear estimates, the physical reconstruction head reconstructs cutting-related forces and equivalent quantities and applies consistency constraints, and the domain alignment head achieves feature alignment between the source domain and the target domain through gradient inversion and a domain classifier.

[0085] S8 employs a joint objective that includes supervised learning, weakly supervised learning, physical consistency, wear monotonic evolution, and domain adversarial training for end-to-end training, and optimizes the training process according to a phased strategy: first, a warm-up phase of supervised and physical consistency is performed, then weak supervision and monotonic evolution are added for semi-supervised training, and finally domain adversarial training is enabled for cross-domain alignment optimization; after training, the target domain signal is processed through S3 to S6, and finally the predicted curve of wear changing with the cutting sequence is obtained through S7.

[0086] Specifically, it includes the following steps:

[0087] S1: Acquire multi-channel cutting signals during CFRP milling, including triaxial cutting force and triaxial acceleration; record fiber orientation angle. and operating parameters ( Wear values ​​are measured every 5 cutting operations as a strong label. The remaining cutting operations obtained weak labels through nonlinear monotonic interpolation. .

[0088] S2: Use the data containing the label as the source domain. The remaining unlabeled data of the first orientation will be used as the target domain. .

[0089] S3: Perform sliding window segmentation on the source and target domain signals obtained in S2 to obtain a window sequence. and For the first Secondary cutting division A length of Step size is The formula for overlapping windows is as follows:

[0090] ;

[0091] ;

[0092] ;

[0093] In the formula, Indicates the first In the second cutting signal, the first The index of the starting sampling point of each sliding window; This indicates the index of the end sampling point for the corresponding window.

[0094] The six-channel sequences of the source and target domains are denoted as follows: and According to the channel order Stacked 6-channel timings. Define the sliding window extraction operator as follows: Indicates from multichannel sequence interval Take one The window matrix is ​​used to obtain the source and target domain sample sequences after the sliding window is applied.

[0095] ;

[0096] ;

[0097] Then, based on the working condition parameters recorded in step S1, the force statistics for each window are scaled uniformly to obtain... The formula is as follows:

[0098] ;

[0099] ;

[0100] ;

[0101] ;

[0102] In the formula, Radial cutting depth, The axial cutting depth, The diameter of the cutting tool. For each tooth feed, As an equivalent benchmark, Corresponding to the three directions of cutting.

[0103] S4: The pre-processed... and (Uniform scale processing is already included) Input the shared temporal convolutional network (TCN) to extract window-level features, and obtain and The formula is as follows:

[0104] ;

[0105] ;

[0106] In the formula, Indicates a shared TCN mapping. Its parameters; For window-level embedding, This is the number of feature channels output by TCN.

[0107] S5: Construct orientation encoding FiLM, for and Perform feature-level modulation to obtain orientation modulation features. and :

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] In the formula, It is an orientation encoding vector. For the first Fiber orientation angle during the second cut For a small perceptron network, modulation parameters , This is a channel-by-channel multiplication.

[0113] S6: Orientation modulation feature within the same cutting operation Aggregation is performed to form a cutting-level embedding. and The mean aggregation method is used, and the formula is as follows:

[0114] ;

[0115] ;

[0116] In the formula, , It refers to the number of cuts. It is an overlapping sliding window. It is the first The total number of windows obtained from each cut.

[0117] S7: Based on Design a wear regression head, a physical reconstruction head, and a domain alignment head to output cutting-level wear identification results, respectively. The result of reconstructing the cutting force Domain category prediction results The formulas are as follows:

[0118] Wear-return head:

[0119] ;

[0120] In the formula, express or , For weights and biases.

[0121] Physical reconstruction head:

[0122] ;

[0123] ;

[0124] In the formula, This is the cutting coefficient vector, corresponding to three components. For learnable vectors, This is the force reconstruction quantity at the window level.

[0125] Domain alignment header:

[0126] ;

[0127] In the formula, For the domain category probability, It is a gradient inversion layer. For weights and biases.

[0128] S8: Employs a method that includes supervised loss. Losses due to weak supervision Physical consistency loss Monotonic loss Losses in the fight against the domain The joint target is trained end-to-end. A phased strategy of prediction, semi-supervised learning, and cross-domain alignment is adopted. After training, the target domain is... The signal is processed according to S3-S6, and the tool wear value is output via S7. The wear curves follow the cutting sequence. The formulas for each loss and the total loss are as follows:

[0129] Monitoring losses:

[0130] ;

[0131] In the formula, For strongly labeled cutting sets, The wear value is the actual measured value of the strong tag in S1. This refers to the wear prediction value output by the wear regression head in S7.

[0132] Loss due to weak supervision:

[0133] ;

[0134] In the formula, This is a set of strongly labeled cuts.

[0135] Physical consistency loss:

[0136] ;

[0137] In the formula, For the set of window indices participating in the constraints, This represents the total number of windows.

[0138] Monotonic constraint loss:

[0139] ;

[0140] In the formula, It is the set of tool or sequence indices used in training. It is a sequence Number of cuts, It is a sequence No. Predicted wear during the second cut It is a monotonic relaxation constant.

[0141] Domain adversarial losses:

[0142] ;

[0143] In the formula, and The source domain and target domain cutting-level sample index set used for domain alignment head training. It is the domain class probability vector output by the domain alignment header. It is a truth value label for a field. It is cross-entropy.

[0144] Total loss:

[0145] ;

[0146] In the formula, , respectively, represent the weight coefficients of each item.

[0147] This embodiment proposes an online tool wear prediction framework for CFRP milling, which simultaneously considers the influence of fiber orientation, consistency of physical mechanisms, and cross-domain generalization ability. Figure 2 As shown, this framework extracts temporal features within a unified network structure and explicitly embeds the influence of different orientations on signal and wear evolution into the model through a fiber orientation conditional modulation mechanism. Simultaneously, it designs multi-task outputs such as wear regression, physical quantity reconstruction, and domain discrimination to jointly characterize tool state and process features. Without requiring complex signal preprocessing, this framework can adapt to various fiber orientations and working conditions, possessing engineering advantages such as easy deployment, scalability, and strong real-time performance, and can stably support online tool monitoring and life management.

[0148] The training strategy in this embodiment employs a phased optimization process: First, in the warm-up phase, the model is trained under supervision using existing labeled data, enabling it to acquire basic wear recognition and physical quantity reconstruction capabilities. Then, in the semi-supervised phase, weak labeling and wear monotonicity constraints are introduced to fully exploit the temporal information of unlabeled samples. Finally, in the cross-domain phase, an adversarial learning mechanism is used to enhance the feature consistency between different orientations and operating conditions, thereby improving the model's adaptability to new orientations and operating conditions. This training strategy allows for the acquisition of more robust wear prediction curves with limited labeling costs, enabling online monitoring and early warning, and providing implementable and scalable intelligent perception and decision support for high-quality CFRP processing.

[0149] To verify the effectiveness of the present invention, a set of visualization examples based on exemplary feature distributions are provided, such as... Figure 3 and Figure 4 As shown. Figure 3This paper illustrates the schematic distribution of features of different fiber orientations (0°, 45°, 90°, 135°) in the two-dimensional t-SNE space under the assumption of no domain alignment. As can be seen from the figure, the features of each orientation exhibit a mutually separated cluster structure, which intuitively reflects the significant distribution differences that may exist under multi-orientation conditions and the resulting orientation domain shift phenomenon. Figure 4 The diagram presents schematic alignment results after incorporating the FiLM orientation modulation and domain adversarial alignment mechanism proposed in this invention. It can be observed that the orientation features exhibit a higher degree of mixing and overlap in the embedding space, and the previously obvious orientation differences are effectively weakened, forming a more consistent shared feature representation.

[0150] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions 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 method for monitoring spindle tool wear considering fiber orientation, characterized in that, Includes the following steps: (1) Acquire multi-channel cutting signals during CFRP milling and record fiber orientation angles. Operating parameters; tool wear value measured after every 5 cuts as a strong label. Unmeasured weak labels are generated through interpolation. ; (2) Use data containing both strong and weak labels as the source domain. The unlabeled data of the next fiber orientation is used as the target domain. ; (3) Perform sliding window segmentation on the source domain and target domain signals to obtain the source domain window sequence. With target domain window sequence Furthermore, the cutting force of each window is standardized to eliminate amplitude differences; among them, Indicates the first Secondary cutting signal. Indicates the first A sliding window; (4) After standardization processing and Input the shared temporal convolutional network (TCN) separately to extract window-level features from the source domain. and window-level features of the target domain ; (5) Construct orientation coding FiLM, for and By performing feature-level modulation, the orientation modulation features of the source domain are obtained. Orientation modulation features of the target domain ; (6) Aggregate the orientation modulation features within the same cutting operation to obtain the cutting-level embedding of the source domain. Cutting level embedding of the target domain ; (7) Based on and Design a wear regression head, a physical reconstruction head, and a domain alignment head, respectively, to output cutting-level wear identification results. The results of the reconstructed cutting force Domain category prediction results ; (8) Construct a system that includes supervised loss Losses due to weak supervision Physical consistency loss Monotonic loss Losses in the fight against the domain The total loss function is used for end-to-end training; after training, the target domain is... The signal is processed according to steps (3)-(6), and the wear regression head outputs the tool wear value and its wear curve with the cutting sequence in step (7).

2. The spindle tool wear monitoring method considering fiber orientation according to claim 1, characterized in that, In step (3), the cutting force of each window is standardized to obtain the normalized window force statistics. The formula is as follows: ； ； ； ； In the formula, Indicates the first Secondary cutting, window Inner The average value of cutting force in each direction; Indicates the first In the cutting signal, the first The index of the starting sampling point of each sliding window; Indicates the first In the second cutting signal, the first Index of the end sampling point of each sliding window; The radial cutting depth is a parameter in the operating conditions. The axial depth of cut is a parameter in the operating conditions. This refers to the feed per tooth in the operating parameters. As an equivalent benchmark; Corresponding to the three directions of cutting.

3. The spindle tool wear monitoring method considering fiber orientation according to claim 1, characterized in that, In step (4), window-level features of the source domain are extracted. and window-level features of the target domain The formula is as follows: ； ； In the formula, Indicates a shared TCN mapping. Its parameters; For window-level embedding, This is the number of feature channels output by TCN.

4. The method for monitoring wear of carbon fiber machining tools considering fiber orientation according to claim 1, characterized in that, In step (5), the orientation modulation features of the source domain are obtained. Orientation modulation features of the target domain The formula is as follows: ； ； ； ； In the formula, It is an orientation encoding vector. For the first Fiber orientation angle during the second cut For a small perceptron network, modulation parameters , It is the number of feature channels output by TCN. This is a channel-by-channel multiplication.

5. The spindle tool wear monitoring method considering fiber orientation according to claim 1, characterized in that, In step (6), the cutting-level embedding of the source domain is obtained. Cutting level embedding of the target domain The formula is as follows: ； ； In the formula, , It is the number of feature channels output by TCN. It is the first The total number of windows obtained from each cut.

6. The spindle tool wear monitoring method considering fiber orientation according to claim 1, characterized in that, In step (7), the cutting-level wear identification result is output. The results of the reconstructed cutting force Domain category prediction results The formula is as follows: Wear return head: ； In the formula, express or , For weights and biases; Physical reconstruction head: ； ； In the formula, This is the cutting coefficient vector, corresponding to three components; For the first Fiber orientation angle during the second cut; It is a learnable vector; The axial depth of cut is a parameter in the operating conditions. As an equivalent benchmark; Domain alignment header: ； In the formula, For the domain category probability, It is a gradient inversion layer. For weights and biases.

7. The method for monitoring wear of carbon fiber machining tools considering fiber orientation according to claim 1, characterized in that, In step (8), a monitoring loss is constructed. Losses due to weak supervision Physical consistency loss Monotonic constraint loss Losses in the fight against the domain The total loss function is used for end-to-end training, and the formula is as follows: ； In the formula, , respectively, represent the weight coefficients of each item.

8. The spindle tool wear monitoring method considering fiber orientation according to claim 7, characterized in that, Monitoring losses The formula is as follows: ； In the formula, For strongly labeled cutting sets; The wear value is for the strong label. The wear prediction value output by the wear regression head in step (7); Losses due to weak supervision The formula is as follows: ； In the formula, For weakly labeled cutting set, The wear value for weak tags; Physical consistency loss The formula is as follows: ； In the formula, For the set of window indices participating in the constraints, Total number of windows; Indicates the first Secondary cutting, window Inner Predictive power reconstructed from each direction; This represents the normalized window force statistic; Monotonic constraint loss The formula is as follows: ； In the formula, It is the set of tool or sequence indices used in training. It is a sequence Number of cuts, It is a sequence No. Predicted wear during the second cut It is a monotonic relaxation constant; Domain confrontation loss The formula is as follows: ； In the formula, and This is a set of source and target domain cutting-level sample indexes used for training the domain alignment head. It is the domain class probability vector output by the domain alignment header. It is a truth value label for a field. It is cross-entropy.

9. The spindle tool wear monitoring method considering fiber orientation according to claim 1, characterized in that, In step (8), a phased strategy is adopted to optimize the training process: first, a warm-up phase of supervision and physical consistency is carried out, then weak supervision and monotonic constraints are added for semi-supervised training, and finally domain adversarial training is enabled for cross-domain alignment optimization.

Images