Tool wear state recognition method and device cooperating with degeneration mechanism and sample characteristics

Abstract

The application discloses a tool wear state recognition method and device cooperating with degradation mechanism and sample characteristics, which is used for improving the precision of tool wear state recognition. Feature vectors are extracted from tool original signals to generate an original unbalanced training set; sample distance is calculated according to tool wear feature vectors and degradation stage labels; sample local topology hardness is generated according to the sample distance; the original unbalanced training set is screened according to the sample local topology hardness to construct a difficult classification seed set; a degradation direction constraint term is generated according to the tool wear feature vectors at the category center of each degradation stage; a local disturbance term is generated according to the same local neighborhood covariance matrix of the difficult classification seed set; new samples are generated based on the difficult classification seed set and combined with the degradation direction constraint term and the local disturbance term to generate an enhanced training set according to the new samples and the original unbalanced training set; and a tool wear state recognition model is subjected to multi-stage wear state recognition and training through the enhanced training set.

Description

Technical Field

[0001] This application relates to the field of tool wear condition identification, and in particular to a method and apparatus for tool wear condition identification that combines degradation mechanism and sample features. Background Technology

[0002] The processing of parts in the aerospace industry has extremely high requirements, especially the processing of key components for aero engines. The requirements for dimensional accuracy, surface integrity and manufacturing reliability are very strict. In order to ensure the high-quality production of parts, in addition to high-precision production lines, the cutting tools used to process parts are also a major factor in high-quality production.

[0003] During machining, the wear condition of cutting tools directly affects the machining quality and process stability of parts. Therefore, high-precision identification of tool wear conditions has become a major research direction for stable machining of parts. Currently, tool wear condition identification can be improved through multimodal signals and data-driven models. These multimodal signals include vibration signals, image signals, and sound signals for tool wear condition identification.

[0004] The quality of tool wear condition identification using multimodal models and data-driven models typically relies on sufficient and relatively balanced training samples, especially those covering all stages of the tool's operational lifecycle. However, in real-world industrial scenarios, tools spend most of their time in normal or stable wear stages, while the number of training samples for severe wear and the transitional phase before failure is limited, often exhibiting a long-tailed distribution. This imbalanced sample distribution can cause the classifier to favor the majority class, thus weakening its ability to identify the minority class. Therefore, improving the model's ability to identify training samples from critical degradation stages is a crucial issue in early warning systems for tool wear.

[0005] Oversampling methods are widely used to address class imbalance due to their simplicity and decoupling from classifiers. However, they generate new samples by linear interpolation between minority class samples and their nearest neighbors, failing to consider the degradation mechanism of the tool and local sample features. In other words, existing oversampling methods primarily generate samples based on geometric proximity relationships in the feature space, lacking constraints on the physical evolution of tool wear. When adjacent degradation stages overlap, linear interpolation or boundary reinforcement may produce pseudo-samples that deviate from the true degradation path, even increasing the risk of cross-class boundary errors. Particularly in the transition from stable wear to severe wear, the number of samples is limited and overlaps significantly with features from adjacent stages. If the generated samples lack constraints on the degradation direction, it becomes difficult to guarantee the physical rationality of sample expansion, leading to a decline in the training effect of the recognition model and consequently reducing the accuracy of tool wear state recognition. Summary of the Invention

[0006] This application discloses a method and apparatus for identifying tool wear state by combining degradation mechanism and sample features, which is used to improve the accuracy of tool wear state identification.

[0007] In a first aspect, embodiments of this application provide a method for identifying tool wear state based on the synergy of degradation mechanism and sample features, comprising: Feature vectors are extracted from the original tool signal to generate an original unbalanced training set. The original unbalanced training set includes several tool wear feature vectors. The original tool signal is all wear signals collected by the target tool during the complete working cycle. The distance between samples is calculated based on the tool wear feature vector and the degradation stage label of the tool wear feature vector; Local topological hardness of each tool wear feature vector is generated based on the distance between samples; The original imbalanced training set is filtered based on the local topological hardness of the samples to construct a seed set that is difficult to classify. The degradation direction of adjacent stages is calculated based on the tool wear feature vector at the center of each degradation stage category, and degradation direction constraint terms are generated based on the degradation direction of adjacent stages. Generate local perturbation terms based on the local neighborhood covariance matrix of the same type of difficult-to-classify seed set; New samples are generated based on the difficult-to-classify seed set and combined with the degradation direction constraint term and the local perturbation term. An enhanced training set is generated based on the new samples and the original imbalanced training set. The tool wear condition recognition model is trained and recognized in multiple stages by enhancing the training set.

[0008] Optionally, the step of generating local perturbation terms based on the local neighborhood covariance matrix of the difficult-to-classify seed set specifically includes: Calculate the local neighborhood covariance matrix of the same type based on the difficult-to-classify seed set; The characteristic fluctuation range of the local neighborhood is generated based on the covariance matrix of the local neighborhood of the same type, the preset shrinkage coefficient and regularization term; Local disturbance terms are generated based on preset disturbance intensity coefficients and the fluctuation range of similar local neighborhood characteristics.

[0009] Optionally, the step of calculating the inter-sample distance based on the tool wear feature vector and the degradation stage label of the tool wear feature vector specifically includes: Generate the covariance matrix of training samples based on the tool wear feature vector; The inter-sample distance between tool wear feature vectors is calculated based on the training sample covariance matrix, the preset regularization coefficient, the tool wear feature vector, and the degradation stage label of the tool wear feature vector.

[0010] Optionally, the step of generating the local topological hardness of each tool wear feature vector based on the distance between samples specifically includes: Generate a set of nearest neighbor feature vectors for each tool wear feature vector based on the distance between samples; Generate feature vector distance weights between tool wear feature vectors based on the distance between samples; The sample local topological hardness is generated for each tool wear feature vector based on the feature vector distance weight and the set of nearest neighbor feature vectors.

[0011] Optionally, after the step of extracting feature vectors from the original tool signal to generate the original unbalanced training set, and before the step of calculating the inter-sample distance based on the tool wear feature vector and degradation stage label, the tool wear state identification method further includes: The tool wear feature vectors in the original unbalanced training set are subjected to feature standardization to eliminate the influence of different feature dimensions; The wear stage is calibrated for the tool wear feature vector, and an initial degradation stage label is associated with each tool wear feature vector.

[0012] Optionally, after generating new samples based on the difficult-to-classify seed set and combining the degradation direction constraint term and the local perturbation term, and generating an enhanced training set based on the new samples and the original unbalanced training set, and before performing multi-stage wear state recognition and training on the tool wear state recognition model using the enhanced training set, the tool wear state recognition method further includes: A nearest neighbor detection set is generated based on the preset nearest neighbor count value and the tool wear feature vector in the enhanced training set. Calculate the proportion of similar neighbors for each nearest neighbor detection set; The tool wear feature vectors in the enhanced training set are removed based on the proportion of similar nearest neighbors; The isolated forest method was used to assist in screening tool wear feature vectors that deviate from the local manifold structure.

[0013] Optionally, after the step of using the isolated forest method to assist in screening tool wear feature vectors that deviate from the local manifold structure, the tool wear state identification method further includes: Based on the degradation direction of adjacent stages and the difficult-to-classify seed set, degradation consistency parameters are generated to enhance the training set. The consistency between the sample expansion direction and the degradation direction is detected based on the degradation consistency parameter in the augmented training set.

[0014] Optionally, after the step of using the isolated forest method to assist in screening tool wear feature vectors that deviate from the local manifold structure, the tool wear state identification method further includes: The distribution fidelity of the augmented training set was analyzed using the Wasserstein distance evaluation method to measure the overall difference between the distribution of the real minority class samples and the distribution of the generated samples.

[0015] Optionally, after the step of using the isolated forest method to assist in screening tool wear feature vectors that deviate from the local manifold structure, the tool wear state identification method further includes: Generate a boundary safety coefficient for the same type based on the relative distance from the enhanced training set to similar samples; Generate the out-of-class boundary risk coefficient based on the relative distance from the enhanced training set to out-of-class samples; Boundary security analysis is performed on the augmented training set based on the boundary security coefficient of similar boundaries and the risk coefficient of dissimilar boundaries.

[0016] Secondly, embodiments of this application provide a tool wear state identification device that combines degradation mechanism and sample characteristics, comprising: The training set generation unit is used to extract feature vectors from the original tool signal and generate an original unbalanced training set. The original unbalanced training set includes several tool wear feature vectors. The original tool signal is all wear signals collected by the target tool during the complete working cycle. The inter-sample distance generation unit is used to calculate the inter-sample distance based on the tool wear feature vector and the degradation stage label of the tool wear feature vector; The sample local topological hardness generation unit is used to generate sample local topological hardness for each tool wear feature vector based on the distance between samples. The difficult-to-classify seed set construction unit is used to filter the original imbalanced training set based on the local topological hardness of the samples in order to construct the difficult-to-classify seed set. The degradation direction constraint term generation unit is used to calculate the degradation direction of adjacent stages based on the tool wear feature vector at the category center of each degradation stage, and generate degradation direction constraint terms based on the degradation direction of adjacent stages. The local perturbation term generation unit is used to generate local perturbation terms based on the local neighborhood covariance matrix of the same type of seed set that is difficult to classify. The enhanced training set generation unit is used to generate new samples based on the hard-to-classify seed set and combined with the degradation direction constraint term and the local perturbation term, and to generate an enhanced training set based on the new samples and the original imbalanced training set. The recognition model training unit is used to perform multi-stage wear state recognition and training on the tool wear state recognition model by enhancing the training set.

[0017] Optionally, the local perturbation term generation unit specifically includes: Calculate the local neighborhood covariance matrix of the same type based on the difficult-to-classify seed set; The characteristic fluctuation range of the local neighborhood is generated based on the covariance matrix of the local neighborhood of the same type, the preset shrinkage coefficient and regularization term; Local disturbance terms are generated based on preset disturbance intensity coefficients and the fluctuation range of similar local neighborhood characteristics.

[0018] Optionally, the inter-sample distance generation unit specifically includes: Generate the covariance matrix of training samples based on the tool wear feature vector; The inter-sample distance between tool wear feature vectors is calculated based on the training sample covariance matrix, the preset regularization coefficient, the tool wear feature vector, and the degradation stage label of the tool wear feature vector.

[0019] Optionally, the sample local topological stiffness generation unit specifically includes: Generate a set of nearest neighbor feature vectors for each tool wear feature vector based on the distance between samples; Generate feature vector distance weights between tool wear feature vectors based on the distance between samples; The sample local topological hardness is generated for each tool wear feature vector based on the feature vector distance weight and the set of nearest neighbor feature vectors.

[0020] Optionally, after the training set generation unit and before the inter-sample distance generation unit, the tool wear state recognition device further includes: The feature standardization processing unit is used to perform feature standardization processing on the tool wear feature vector in the original unbalanced training set to eliminate the influence of different feature dimensions; The wear stage calibration unit is used to calibrate the wear stage of the tool wear feature vector and associate an initial degradation stage label with each tool wear feature vector.

[0021] Optionally, after the enhancement training set generation unit and before the identification model training unit, the tool wear state identification device further includes: The nearest neighbor detection set generation unit is used to generate a nearest neighbor detection set based on the preset nearest neighbor number value of the tool wear feature vector in the enhanced training set; The same-type nearest neighbor ratio calculation unit is used to calculate the same-type nearest neighbor ratio for each nearest neighbor detection set; The elimination processing unit is used to eliminate tool wear feature vectors in the enhanced training set according to the proportion of similar nearest neighbors; The screening unit is used to assist in screening tool wear feature vectors that deviate from the local manifold structure using the isolated forest method.

[0022] Optionally, after the screening unit, the tool wear condition identification device further includes: The degradation consistency parameter generation unit is used to generate degradation consistency parameters for the augmented training set based on the degradation direction of adjacent stages and the difficult-to-classify seed set. The consistency detection unit is used to detect the consistency between the sample expansion direction and the degradation direction of the augmented training set based on the degradation consistency parameter.

[0023] Optionally, after the screening unit, the tool wear condition identification device further includes: The distribution fidelity analysis unit is used to perform distribution fidelity analysis on the augmented training set according to the Wasserstein distance evaluation method to measure the overall difference between the distribution of the real minority class samples and the distribution of the generated samples.

[0024] Optionally, after the screening unit, the tool wear condition identification device further includes: The same-class boundary safety coefficient generation unit is used to generate the same-class boundary safety coefficient based on the relative distance from the enhanced training set to the same-class samples; The outlier boundary risk coefficient generation unit is used to generate outlier boundary risk coefficients based on the relative distance from the augmented training set to outlier samples. The boundary security analysis unit is used to perform boundary security analysis on the augmented training set based on the boundary security coefficient of similar boundaries and the risk coefficient of dissimilar boundaries.

[0025] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: In this application, feature vectors are first extracted from the original tool signal to generate an original unbalanced training set. This original unbalanced training set includes several tool wear feature vectors, and the original tool signal consists of all wear signals collected by the target tool during a complete work cycle. The distance between samples is calculated based on the tool wear feature vectors and their degradation stage labels. A sample local topological hardness is generated for each tool wear feature vector based on the sample distance. The original unbalanced training set is then filtered based on the sample local topological hardness to construct a difficult-to-classify seed set. The degradation direction of adjacent stages is calculated based on the tool wear feature vectors at the category centers of each degradation stage, and a degradation direction constraint term is generated based on this direction. A local perturbation term is generated based on the covariance matrix of the local neighborhood of the difficult-to-classify seed set. New samples are generated based on the difficult-to-classify seed set, combined with the degradation direction constraint term and the local perturbation term. An enhanced training set is then generated based on these new samples and the original unbalanced training set. The enhanced training set is used to perform multi-stage wear state recognition and training on the tool wear state recognition model.

[0026] First, a difficult-to-classify seed set is selected from the original unbalanced training set based on local topological hardness. Then, degradation direction constraint terms and local perturbation terms are generated. New samples are generated under the combined effect of degradation direction constraints (degradation direction constraint terms) and similar local feature perturbations (local perturbation terms). For minority samples in the critical degradation stage, local topological hardness, wear degradation direction, and similar local feature fluctuations are jointly introduced into the sample generation process. This reduces the generation of pseudo-samples that deviate from the true degradation path and pseudo-samples that cross class boundaries, ensuring the physical rationality of sample expansion. This can improve the identification ability of the critical degradation transition stage of the tool under sample scarcity conditions and improve the accuracy of tool wear state identification. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the first embodiment of the tool wear state identification method based on the synergy of degradation mechanism and sample features in this application; Figure 2 A schematic diagram of a first embodiment of the method for generating local perturbation terms in this application; Figure 3 This is a schematic diagram of the first embodiment of the method for calculating the distance between samples in this application; Figure 4 A schematic diagram of a first embodiment of the method for generating local topological hardness of a sample according to this application; Figure 5 This is a schematic diagram of the first embodiment of the method for preprocessing the original imbalanced training set of this application; Figure 6 This is a schematic diagram of the first embodiment of the method for enhancing the selection of the training set according to this application; Figure 7 This is a schematic diagram of the first embodiment of the conformity detection method of this application; Figure 8 This is a schematic diagram of the first embodiment of the distribution fidelity analysis method of this application; Figure 9 This is a schematic diagram of one embodiment of the boundary security analysis method of this application; Figure 10 This is a schematic diagram of an embodiment of the tool wear state identification device that combines degradation mechanism and sample characteristics according to this application; Figure 11This is the overall flowchart of the tool wear state identification method based on the synergy of degradation mechanism and sample characteristics proposed in this application; Figure 12 This is a schematic diagram illustrating the seed selection and degradation direction modeling based on local topological hardness in this application. Figure 13 This is a schematic diagram of the structured manifold purification based on boundary consistency detection and IF-assisted screening in this application; Figure 14 Generate a comparison chart of PCA distribution for samples S3–S4 of the PHM2010 dataset in this application; Figure 15 This is a feature space degradation trajectory diagram of real samples from the TTWD dataset in this application. Detailed Implementation

[0029] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0030] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0031] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0032] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0033] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0034] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0035] The method described in this application can be applied to servers, devices, terminals, or other devices with logical processing capabilities; therefore, this application does not limit its application. For ease of description, the following description uses a terminal as the executing entity.

[0036] Please see Figure 1 This application provides an embodiment of a tool wear state identification method that combines degradation mechanism and sample features, comprising: 101. Extract feature vectors from the original tool signal to generate an original unbalanced training set. The original unbalanced training set includes several tool wear feature vectors. The original tool signal is all wear signals collected by the target tool during the complete working cycle. In this embodiment, the target tool refers to a specific machining tool installed on a CNC machine tool in an aerospace parts production line, covering various industrial machining tools such as milling cutters, turning tools, and drills. An independent full lifecycle data set is established from the first use of the target tool, tracking the state changes of each machining operation and collecting corresponding raw tool signals. Raw tool signals refer to signals collected during the tool's operation that characterize its wear state. These include, for example, raw time-series signals synchronously collected by multimodal sensors such as vibration sensors, current Hall sensors, and acoustic emission sensors installed on the machine tool spindle. These signals cover three core types of signals during tool machining: vibration acceleration, spindle motor current, and tool friction acoustic emission, as well as image data of the tool after each machining stage. The complete operating cycle in this embodiment refers to the entire lifecycle of the target tool, from its brand-new, unused state, through the initial break-in stage, the normal stable wear stage, and the rapid and severe wear stage, ultimately reaching a state of severe wear that prevents further machining or is deemed unusable and scrapped. This covers all machining operation periods from the tool's first use to its final scrapping, completely reconstructing the complete degradation process of tool wear.

[0037] Specifically, in this embodiment, a mechanism of multi-sensor synchronous acquisition and full lifecycle time-series binding can be used to acquire the raw tool signals throughout the tool's entire lifecycle. First, after the target tool is initially installed on the machine, the system automatically triggers the full-cycle acquisition process, binding a unique lifecycle ID to the target tool and synchronously activating various sensors. It should be noted that these sensors can be those originally installed in the production machine tool or additionally installed, such as vibration sensors, current Hall sensors, and acoustic emission sensors at the spindle end. Subsequently, the system acquires all raw signals in real time during each machining operation of the tool. When the tool reaches a preset wear threshold, the system or a manual determination triggers a scrap determination process, marking the end of the target tool's complete operating cycle.

[0038] Feature vector extraction refers to the process of extracting tool wear feature vectors that reflect tool wear changes from the raw signals of multimodal tools, comprehensively covering all signal change characteristics during the tool wear process. The tool wear feature vector is a multidimensional vector composed of wear features from multiple different modes, capable of accurately reflecting the real-time wear status of the target tool from different perspectives. The initial training dataset composed of all tool wear feature vectors is the original imbalanced training set.

[0039] 102. Calculate the distance between samples based on the tool wear feature vector and the degradation stage label of the tool wear feature vector; In this embodiment, the terminal first obtains the degradation stage label of the tool wear feature vector. Then, it calculates the inter-sample distance between training samples based on the tool wear feature vector and the degradation stage label of the tool wear feature vector. The specific calculation method will be described in detail in subsequent embodiments.

[0040] 103. Generate local topological hardness for each tool wear feature vector based on the distance between samples; After the terminal generates the distance between samples, in order to determine the distance between the training samples and the class boundary, the terminal needs to generate the sample local topological hardness for each tool wear feature vector based on the distance between samples. The specific calculation method of the sample local topological hardness will be described in detail in subsequent embodiments.

[0041] 104. The original imbalanced training set is screened based on the local topological hardness of the samples in order to construct a seed set that is difficult to classify; In this embodiment, after the terminal calculates the local topological hardness of the sample, in order to avoid redundant expansion of safe samples within the class and misselection of abnormal samples, a dual threshold strategy is used to construct a seed set for difficult classification, as shown in the following formula:

[0042] In the formula For a set of seed samples that are difficult to classify into category c and These represent the lower and upper thresholds for local topological hardness. The final hard-classifying seed sample set is generated by integrating the hard-classifying seed sample sets of all categories. .

[0043] 105. Calculate the degradation direction of adjacent stages based on the tool wear feature vector at the category center of each degradation stage, and generate degradation direction constraint terms based on the degradation direction of adjacent stages; In this embodiment, to introduce a degradation trend of tool wear from light to heavy, the terminal is defined as follows: Let the category center of the c-th degradation stage be... Then the direction of degradation in adjacent stages is defined as:

[0044] In the formula, It is a very small constant. The category center is for the (c+1)th degradation stage. The S3 to S4 direction is used in the PHM2010 dataset, while the C3 to C4 direction is used in the HMoTP and TTWD datasets. Specific data will be described in detail in subsequent examples. When c is the 3rd degradation stage, i.e., the degradation direction of adjacent stages... The degradation direction of adjacent stages is used to constrain the generated samples to extend along the direction of increased wear, reducing the risk of generating meaningless pseudo-samples. The degradation direction constraint term is generated based on the degradation direction of adjacent stages. ,in This is the degradation direction coefficient.

[0045] 106. Generate local perturbation terms based on the local neighborhood covariance matrix of the same type of seed set that is difficult to classify; In obtaining a set of seed samples that are difficult to classify and the direction of degradation Furthermore, this embodiment introduces local sample feature perturbation to ensure that the generated samples simultaneously satisfy the consistency of degradation direction and the rationality of local distribution. Specifically, the terminal generates local perturbation terms based on the covariance matrix of similar local neighborhoods in the difficult-to-classify seed set. The specific generation method will be described in detail in subsequent embodiments.

[0046] 107. Generate new samples based on the difficult-to-classify seed set and combine them with the degradation direction constraint term and the local perturbation term. Generate an enhanced training set based on the new samples and the original imbalanced training set. When the generated degradation direction constraint term is and local disturbance terms Then, new samples can be generated. The generated sample is defined as follows:

[0047] in, For local disturbance terms, The disturbance intensity coefficient is... This is a new sample. This generation method combines the direction of wear and degradation with fluctuations in similar local features, which can reduce the risk of generating meaningless pseudo-samples using traditional linear interpolation.

[0048] 108. Multi-stage wear state recognition and training of the tool wear state recognition model by enhancing the training set.

[0049] Finally, the terminal can then use the enhanced training set to perform multi-stage wear state recognition and training on the tool wear state recognition model. This paper uses OVO-SVM to complete multi-stage wear state recognition. For K degradation stages, OVO-SVM constructs a total of k(k-1) / 2 binary classifiers, and determines the final category through voting. Please refer to [reference needed]. Figure 11 , Figure 11 This is a flowchart illustrating the overall process of a tool wear state identification method that combines degradation mechanisms with sample features. Please refer to the following: Figure 12 , Figure 12 This is a schematic diagram illustrating the seed selection and degradation direction modeling based on local topological hardness in this application.

[0050] In this embodiment, feature vectors are first extracted from the original tool signal to generate an original unbalanced training set. This original unbalanced training set includes several tool wear feature vectors, and the original tool signal consists of all wear signals collected by the target tool during a complete work cycle. The distance between samples is calculated based on the tool wear feature vectors and their degradation stage labels. A sample local topological hardness is generated for each tool wear feature vector based on the sample distance. The original unbalanced training set is then filtered based on the sample local topological hardness to construct a difficult-to-classify seed set. The degradation direction of adjacent stages is calculated based on the tool wear feature vectors at the category centers of each degradation stage, and a degradation direction constraint term is generated based on the adjacent stage degradation directions. A local perturbation term is generated based on the covariance matrix of the local neighborhood of the difficult-to-classify seed set. New samples are generated based on the difficult-to-classify seed set and combined with the degradation direction constraint term and the local perturbation term. An enhanced training set is then generated based on the new samples and the original unbalanced training set. The enhanced training set is used to perform multi-stage wear state recognition and training on the tool wear state recognition model.

[0051] First, a difficult-to-classify seed set is selected from the original unbalanced training set based on local topological hardness. Then, degradation direction constraint terms and local perturbation terms are generated. New samples are generated under the combined effect of degradation direction constraints (degradation direction constraint terms) and similar local feature perturbations (local perturbation terms). For minority samples in the critical degradation stage, local topological hardness, wear degradation direction, and similar local feature fluctuations are jointly introduced into the sample generation process. This reduces the generation of pseudo-samples that deviate from the true degradation path and pseudo-samples that cross class boundaries, ensuring the physical rationality of sample expansion. This can improve the identification ability of the critical degradation transition stage of the tool under sample scarcity conditions and improve the accuracy of tool wear state identification.

[0052] Please see Figure 2 This application provides an embodiment of a method for generating local perturbation terms, comprising: 201. Calculate the local neighborhood covariance matrix of the same type based on the difficult-to-classify seed set; 202. Generate the characteristic fluctuation range of the local neighborhood of the same type based on the covariance matrix of the local neighborhood of the same type, the preset shrinkage coefficient and regularization term; 203. Generate local disturbance terms based on the preset disturbance intensity coefficient and the fluctuation range of similar local neighborhood characteristics.

[0053] The local perturbation term generated in this embodiment is a small-range random perturbation generated in the local neighborhood of the difficult-to-classify seed sample. This allows the newly generated sample to be reasonably distributed in the local neighborhood, without completely overlapping with the original seed sample, and without deviating from the local distribution range of similar samples.

[0054] For seed-like Search for its in similar samples The formula for the range of local feature fluctuations in the same local neighborhood is as follows: (The formula is missing from the provided text.)

[0055] In the formula For local neighborhood covariance matrices of the same type, The shrinkage coefficient, This is a regularization term. The range of local neighborhood feature fluctuations within this class is used to characterize the feature fluctuations of minority class samples within their local neighborhoods. Based on this, the generated sample is defined as:

[0056] In the formula, For the degradation direction constraint term, For local disturbance terms, This represents the disturbance intensity coefficient.

[0057] Please see Figure 3 This application provides an embodiment of a method for calculating the distance between samples, comprising: 301. Generate the training sample covariance matrix based on the tool wear feature vector; 302. Calculate the inter-sample distance between tool wear feature vectors based on the training sample covariance matrix, the preset regularization coefficient, the tool wear feature vector, and the degradation stage label of the tool wear feature vector.

[0058] In this embodiment, the terminal is set Let be the tool wear feature vector of the i-th training sample (original tool signal). Let be the degradation stage label for the i-th training sample. To reduce the impact of feature dimensions and correlation on neighborhood search, this embodiment uses a regularized Mahalanobis distance metric to fully adapt the inter-sample distance to the wear feature distribution. Unlike the traditional Euclidean distance, this method can more effectively eliminate multicollinearity interference between features.

[0059]

[0060] in, To generate the training sample covariance matrix based on the tool wear feature vector, The preset regularization coefficients are used. It is an identity matrix. Let be the distance between the i-th training sample and the j-th training sample.

[0061] The training sample covariance matrix is ​​a multidimensional covariance matrix calculated based on all calibrated tool wear feature vectors. Each element in the matrix corresponds to the covariance between two different feature dimensions, which can accurately reflect the linear correlation between different wear features and solve the problem of distance calculation deviation caused by multicollinearity between features.

[0062] Please see Figure 4 This application provides an embodiment of a method for generating local topological hardness of a sample, comprising: 401. Generate a set of nearest neighbor feature vectors for each tool wear feature vector based on the distance between samples; Calculate the inter-sample distance between training samples Subsequently, based on the precise inter-sample distance For each feature vector, match the corresponding set of nearest neighbor samples, based on Building samples k-nearest neighbor set Where k is the number of nearest neighbors. The nearest neighbor feature vector set refers to the set of the k nearest feature vectors selected from the entire original unbalanced training set for a given target tool wear feature vector. All samples are distances from the target sample in the feature space. The nearest local neighborhood samples can accurately reflect the characteristic distribution topology of the local area where the target sample is located.

[0063] Specifically, this embodiment uses a hierarchical nearest neighbor retrieval mechanism to perform retrieval on each target sample ( This process generates a set of nearest-neighbor feature vectors that takes into account both similar and dissimilar distribution characteristics, avoiding the loss of neighborhood information caused by a single retrieval logic. Based on the distance between samples and for each target tool wear feature vector, the distance sequence between the tool wear feature vector of the training sample and the tool wear feature vectors of all other samples is extracted and sorted in ascending order of distance.

[0064] The terminal then selects the k nearest neighbors from the sorted distance sequence based on the preset number of nearest neighbors K, and uses them as the nearest neighbor set.

[0065] 402. Generate feature vector distance weights between tool wear feature vectors based on the distance between samples; In this embodiment, the terminal also needs to generate weight coefficients for nearest neighbor samples based on the distance between samples. Nearest neighbor samples closer to the target sample have higher weights, while samples farther away have lower weights, accurately reflecting the contribution of different nearest neighbor samples to the local topology of the target sample. In this embodiment, the feature vector distance weight refers to the weight coefficient assigned to each nearest neighbor sample in the nearest neighbor feature vector set. The magnitude of the weight is inversely proportional to the distance of the nearest neighbor sample to the target sample. Samples closer to the target sample have higher weights and a greater impact on the local topology of the target sample, while samples farther away have lower weights and a smaller impact.

[0066] 403. Generate sample local topological hardness for each tool wear feature vector based on the feature vector distance weight and the nearest neighbor feature vector set.

[0067] The formula for generating sample local topological hardness for each tool wear feature vector based on feature vector distance weights and the set of nearest neighbor feature vectors is as follows:

[0068] In the formula, Let the distance weights be the distance between the i-th training sample and the j-th training sample. This is an indicator function. The larger the value, the closer the training sample is to the class boundary.

[0069] Please see Figure 5 This application provides an embodiment of a method for preprocessing an original imbalanced training set, comprising: 501. Perform feature standardization on the tool wear feature vectors in the original unbalanced training set to eliminate the influence of different feature dimensions; In this embodiment, the tool wear feature vectors in the original unbalanced training set are first subjected to feature standardization. Feature standardization refers to mapping all tool wear feature vectors with different dimensions and numerical ranges to the same numerical range through a linear transformation algorithm. This ensures that the weight contributed by each tool wear feature vector in the subsequent distance calculation is more reasonable, and avoids the problem that features with large numerical ranges completely dominate the distance calculation results.

[0070] In this embodiment, an adaptive Z-score standardization algorithm can be selected to unify the dimensions of all features while preserving the original distribution characteristics of the features and avoiding disruption of the relative distance relationships between different samples. First, the system iterates through all tool wear feature vectors in the original unbalanced training set. For each feature dimension, the global mean and global standard deviation of all samples in that dimension are calculated as the standardization benchmark parameters for that dimension. Then, the terminal performs Z-score standardization on all sample values ​​for each feature dimension. The original sample value is subtracted from the global mean of that dimension and then divided by the global standard deviation of that dimension. After the transformation, the numerical distribution of all samples in that dimension satisfies a standard normal distribution with a mean of 0 and a standard deviation of 1, unifying the numerical ranges of all feature dimensions to a similar interval. Finally, the system verifies all standardized feature vectors to confirm that there is no numerical distribution imbalance caused by dimensional differences, while preserving the relative distance relationships between the original features, and outputs a set of standardized feature vectors.

[0071] 502. The wear stage is calibrated for the tool wear feature vector, and an initial degradation stage label is associated with each tool wear feature vector.

[0072] The terminal performs wear stage labeling on the tool wear feature vectors. This labeling refers to the process of initially assigning a corresponding wear stage label to each tool wear feature vector by combining the time sequence of the tool's entire life cycle and the detection data of the actual wear amount. This conforms to the natural degradation sequence of tool wear and avoids label misalignment in time sequence. However, for difficult-to-identify and definite stages such as transition stages, labels will be generated directly in subsequent steps. The degradation stage label refers to the initially assigned wear stage label for each tool wear feature vector. These can typically be divided into new break-in stage labels, normal stable wear stage labels, rapid and severe wear stage labels, and scrap warning stage labels. Each label corresponds to a specific degradation stage of the tool. It should be noted that, in addition to the above, degradation stage labels can also be divided according to the characteristics of different tools or manually; this is not limited here.

[0073] The following example illustrates this: The terminal can first generate initial degradation stage labels using a two-layer calibration mechanism combining temporal constraints and measured wear calibration. First, based on the time proportion of the tool's entire lifecycle and according to industry-standard tool wear stage division ratios, the terminal initially divides the feature vectors of the entire lifecycle into four consecutive temporal intervals, each corresponding to a different wear stage, ensuring that the temporal order of the labels fully conforms to the natural degradation law of tool wear. Then, the system combines the actual flank wear data obtained manually at fixed intervals throughout the tool's lifecycle to calibrate the initially divided labels. If the measured wear amount in a certain time period has reached the threshold of severe wear, the labels of all feature vectors corresponding to that time period are adjusted to the rapid severe wear stage label, thereby correcting the stage boundary deviation caused by pure temporal division. Finally, the terminal outputs all calibrated tool wear feature vectors and their corresponding degradation stage labels.

[0074] Please see Figure 6 This application provides an embodiment of an enhanced training set selection method, comprising: 601. Generate a nearest neighbor detection set based on the preset nearest neighbor count value and the tool wear feature vector in the enhanced training set; 602. Calculate the proportion of similar neighbors for each nearest neighbor detection set; 603. Remove tool wear feature vectors from the enhanced training set based on the proportion of similar nearest neighbors; 604. Use the isolated forest method to assist in screening tool wear feature vectors that deviate from the local manifold structure.

[0075] In this embodiment, to reduce the interference of obviously out-of-bounds samples on the classification boundary, boundary consistency detection is introduced, and an isolated forest is used as an auxiliary anomaly scoring tool, as illustrated below. Figure 13 As shown. For generating samples Searching for its in the enhanced training set Find the nearest neighbors and calculate the proportion of neighbors of the same type. The formula is as follows:

[0076] when If a sample is found to be in a high-risk boundary region, it is removed. Isolation forest is used only as an auxiliary anomaly scoring tool to identify samples that significantly deviate from the local manifold structure; it is not used as the final classification model.

[0077] After boundary detection, the retained generated samples and the original training samples (original imbalanced training set) together constitute the augmented training set.

[0078] Please see Figure 7 This application provides an embodiment of a consistency detection method, comprising: 701. Generate degradation consistency parameters for the enhanced training set based on the degradation direction of adjacent stages and the difficult-to-classify seed set; 702. Based on the degradation consistency parameter, perform consistency detection between the sample expansion direction and the degradation direction on the augmented training set.

[0079] To evaluate the reasonableness of the generated samples, this embodiment analyzes them from three aspects: degradation consistency, distribution fidelity, and boundary safety. Degradation consistency is used to determine whether the generated samples extend along the direction of tool wear aggravation. The terminal generates degradation consistency parameters for the augmented training set based on the degradation direction of adjacent stages and the difficult-to-classify seed set. It is mainly measured by the cosine similarity between the generation direction and the degradation mechanism direction:

[0080] The larger the value, the more consistent the expansion direction of the generated samples is with the degradation direction. For multi-stage degradation tasks, the real minority class samples and generated samples are further projected onto the S3-S4 or C3-C4 degradation axis to calculate the degradation trajectory deviation. Its consistency score (TCS) is used to characterize the position and dispersion of generated samples on the degradation trajectory.

[0081] Please see Figure 8 This application provides an embodiment of a method for distribution fidelity analysis, comprising: 801. Perform distribution fidelity analysis on the augmented training set using the Wasserstein distance evaluation method to measure the overall difference between the distribution of the real minority class samples and the distribution of the generated samples.

[0082] Distribution fidelity was evaluated using the Wasserstein distance, which measures the overall difference between the true minority class distribution and the generated sample distribution. The smaller the value, the closer the generated sample is to the true minority class distribution.

[0083] Please see Figure 9 This application provides an embodiment of a method for boundary security analysis, comprising: 901. Generate the boundary safety coefficient of the same type based on the relative distance from the enhanced training set to samples of the same type; 902. Generate the out-of-class boundary risk coefficient based on the relative distance from the enhanced training set to out-of-class samples; 903. Perform boundary security analysis on the augmented training set based on the boundary security coefficient of the same type and the boundary risk coefficient of different types.

[0084] In this embodiment, boundary security is evaluated by the relative distances between the generated new sample and samples of the same and different classes. and Let be the distances from the generated sample to the nearest sample of the same class and the nearest sample of a different class, respectively. Then we have:

[0085] In the formula, The larger the value, the closer the generated sample is to the same type of region. The smaller the value, the lower the risk of generated samples crossing boundaries. For the HMoTP dataset, further... As a boundary safety ratio indicator, this paper not only evaluates the oversampling effect based on classification performance, but also characterizes the quality of generated samples from three aspects: degradation direction, distribution similarity, and boundary risk.

[0086] The effectiveness of the proposed solution is described in detail below. This embodiment verifies the effectiveness of the proposed solution on three datasets: PHM2010, HMoTP, and TTWD. PHM2010 is used to evaluate the main performance of the method on a publicly available milling cutter wear dataset; HMoTP is used to test the applicability of the method on external milling datasets; and TTWD is used to analyze the generalization ability of the method on turning tool wear data. All experiments use OVO-SVM as the final classifier and compare it with methods such as Imbalanced, ROS, SMOTE, B-SMOTE, ADASYN, and A-SUWO, as well as this application (DMSF-A-SUWO). To reduce the randomness caused by random partitioning and generation, the experiments are repeated 30 times, and the mean and standard deviation are reported. Classification performance is evaluated using accuracy, macro-F1 score, balanced accuracy, and recall. Furthermore, the Wilcoxon signed-rank test was used to analyze the statistical differences between the DMSF-A-SUWO and comparative methods, with a significance level set at 0.05. The reasonableness of the generated samples was evaluated from aspects such as consistency of degradation direction, fidelity of Wasserstein distribution, and boundary safety.

[0087] I. Main Experiment of PHM2010 End Mill The PHM2010 dataset was used to validate the main performance of the proposed method in the end mill wear imbalance diagnosis task. Three labeled tool samples (c1, c4, and c6) were selected and divided into four degradation stages (S1–S4) based on the wear curve trend and key inflection points. S3 and S4 correspond to severe wear stage I and severe wear stage II, respectively, and are the critical minority classes in the imbalance diagnosis task. To simulate the scarcity of samples in the critical degradation stages in real-world industrial scenarios, only 25 samples were retained for S3 and S4 in the training set, resulting in an imbalance ratio of approximately 10.08. Experimental results of different methods on the PHM2010 dataset are shown in Table 1.

[0088] Table 1. Classification results of different methods on the PHM2010 dataset.

[0089] As shown in Table 1, the S3 Recall of the Imbalanced method is only 0.8197, indicating that S3, as a critical stage in the transition from stable wear to severe wear, is highly sensitive to sample scarcity. The results of A-SUWO are close to those of Imbalanced, suggesting that the original A-SUWO has limited enhancement effect on the critical transition stages in this dataset. In contrast, DMSF-A-SUWO achieves the best or tied-best results in Accuracy, Macro-F1, Balanced Accuracy, and S3 Recall, with an S3 Recall of 0.9508, demonstrating that the proposed method can enhance the identification ability of the critical S3 degradation stage while maintaining overall diagnostic performance.

[0090] It is worth noting that the S4 recall of DMSF-A-SUWO is not higher than that of Imbalanced and A-SUWO. This indicates that the severe wear stage in S4 is inherently separable and is not a major hard category in the PHM2010 dataset. Therefore, the main advantage of the proposed method on PHM2010 lies in the critical transition stage in S3, rather than simply improving the recall of all categories. To further evaluate the rationality of the samples generated by different methods, this paper calculates the degradation trajectory bias, distribution fidelity, and boundary safety indices, and the results are shown in Table 2.

[0091] Table 2 Comparison of Reasonableness Indicators for Sample Generation in the PHM2010 Dataset

[0092] ROS (Randomized Outcome Response) refers to the random replication of minority class samples. Its metrics primarily reflect the properties of the original minority class samples and do not represent the ability to generate new samples. As shown in Table 2, SMOTE... The better performance on TCS indicates that its generated samples are closer to the true minority class distribution in terms of the statistical position of the degradation axis. ROS has better Wasserstein and safety metrics, but it is essentially a sample replication and does not form a new sample structure. Compared to SMOTE, B-SMOTE, and ADASYN, DMSF-A-SUWO achieves higher scores. and the lowest This indicates that the generated samples have a low risk of boundary intrusion. Furthermore, its Wasserstein distance is lower than that of SMOTE, B-SMOTE, and ADASYN, demonstrating that this method maintains good distribution fidelity while improving boundary security. For further observation of the spatial distribution of generated samples in the S3–S4 critical region, please refer to... Figure 14 , Figure 14 The visualization results of PCA using different methods are presented. PCA is only used to assist in displaying the projected distribution of generated samples on the first two principal components.

[0093] Depend on Figure 14 It is evident that SMOTE, B-SMOTE, and ADASYN generate relatively dense samples in the S3–S4 boundary region, with some samples close to the boundary overlap area; DMSF-A-SUWO generates relatively conservative sample distributions, without forming excessively dense expansion in the boundary region. This result is consistent with the higher values ​​in Table 2. and lower The consistency indicates that the proposed method can reduce the proportion of high-risk samples at the boundary while supplementing samples in the critical degradation stage.

[0094] II. Validation of HMoTP External Milling Dataset To examine the applicability of the proposed method under different milling data sources, the HMoTP high-speed milling thin-walled part tool wear dataset was selected for external milling dataset validation. This dataset contains continuous cutting samples from three tools, totaling 300 data sets. Based on the wear evolution trend, the samples were divided into four degradation stages, C1–C4, and 42-dimensional statistical features were extracted for model training and testing. Nearest neighbor structure analysis results show that the proportion of dissimilar neighbors C2 and C4 in the neighborhood of sample C3 is relatively high, indicating that stage C3 is located in the transition region from stable wear to severe wear, exhibiting significant boundary overlap. Therefore, the HMoTP dataset can be used to verify the applicability of the proposed method in complex boundary milling scenarios.

[0095] In the construction of the imbalanced training set, C3 and C4 were set as minority classes with 16 and 13 training samples, respectively. The results of repeated experiments using different methods on the HMoTP dataset are shown in Table 3.

[0096] Table 3. Classification results of different methods on the HMoTP dataset.

[0097] Table 3 shows that the Imbalanced method has a C3 Recall of only 0.1739, indicating that the model's ability to identify the critical transition stage C3 is significantly limited under conditions of insufficient minority class samples. SMOTE achieves the highest values ​​in Accuracy, Macro-F1, BalancedAccuracy, and C3 Recall, indicating that the local interpolation method is still highly competitive on this dataset. Although A-SUWO achieves the highest value in C4 Recall, its C3 Recall is only 0.2174, indicating that it tends to enhance the severe wear end stage and has limited improvement on the C3 boundary overlap stage. Compared to A-SUWO, DMSF-A-SUWO increases the C3 Recall to 0.3942 and the Macro-F1 from 0.7312 to 0.7689, indicating that the synergistic constraint of degradation mechanism and sample features can improve the problem of insufficient enhancement of critical transition stages by the original A-SUWO.

[0098] It should be noted that DMSF-A-SUWO does not outperform SMOTE across the board on the HMoTP dataset, especially in terms of overall average metrics and C4 Recall. Therefore, the HMoTP experimental results mainly demonstrate that the proposed method has certain applicability on external milling datasets, and its advantages are mainly reflected in the enhancement of the C3 critical transition stage compared to A-SUWO, rather than being the best across all metrics.

[0099] To further analyze the rationality of the generated samples on the HMoTP dataset, the degradation direction was calculated to be consistent. Wasserstein distance and boundary safety ratio The results are shown in Table 4.

[0100] Table 4 shows that SMOTE and A-SUWO Both are close to 0, indicating that the generated samples from both lack explicit constraints on the direction of degradation. DMSF-A-SUWO's The value reached 0.3826, indicating that its generated samples better conform to the degenerate evolution direction of C3→C4. Meanwhile, DMSF-A-SUWO achieved the lowest Wasserstein distance and the highest... This indicates that it performs well in terms of both distribution fidelity and boundary security.

[0101] Table 4 Comparison of Reasonableness Indicators for Samples Generated from the HMoTP Dataset

[0102] III. TTWD Turning Tool Generalization Verification To verify the applicability of the proposed method under turning conditions, a generalization experiment was conducted using the TTWD tool lifecycle wear dataset constructed by the team. This dataset was collected using a CNC lathe, a three-axis accelerometer, a vibration signal acquisition system, and a tool wear measurement device, matching a total of 180 valid vibration samples and corresponding flank wear amounts. Based on the wear curve trends and key inflection points, the samples were divided into four degradation stages, C1–C4, with C3 and C4 corresponding to the critical stages after the tool enters accelerated degradation. To simulate the scarcity of samples in the critical degradation stages during actual turning, C3 and C4 were set as minority classes in the repeated hierarchical partitioning, with an average of 7 imbalanced training samples. The experimental results of repeated hierarchical partitioning of different methods on the TTWD dataset are shown in Table 5.

[0103] Table 5 shows the classification results of different methods on the TTWD dataset.

[0104] The results in the table are the mean ± standard deviation of 30 repeated stratified partitioning experiments. B-SMOTE represents Borderline-SMOTE. Bold indicates the best result for the corresponding metric. Table 5 shows that the C3 Recall of the Imbalanced method is only 0.0444, indicating that the C3 severe wear transition stage is easily submerged by the majority class under conditions of scarce samples. DMSF-A-SUWO achieves a C3 Recall of 0.4289, the highest among all methods, indicating that it can enhance the identification ability of the critical transition stage of tool wear. However, DMSF-A-SUWO did not achieve the highest values ​​in Accuracy, Macro-F1, and Balanced Accuracy. This indicates that the TTWD dataset has more complex stage overlaps and operating condition fluctuations, and the advantage of the proposed method is mainly reflected in the enhanced identification of the C3 stage, rather than the overall optimal performance across all metrics. To explain the impact of stage overlap on the classification results in the TTWD dataset, Figure 15 The degradation trajectory of real samples in the RMS–Kurtosis feature space is given. Figure 15 It is evident that there is local overlap between C2, C3, and C4, indicating that the tool wear process exhibits continuous transition and blurred boundary characteristics. (This is further supported by Table 5 and...) Figure 15 It can be seen that although DMSF-A-SUWO did not achieve the best overall performance, it was able to improve the recall rate in the critical transition phase of C3, indicating that it has certain applicability in turning conditions.

[0105] Please see Figure 10 This application provides an embodiment of a tool wear state identification device that combines degradation mechanism and sample characteristics, comprising: The training set generation unit 1001 is used to extract feature vectors from the original tool signal and generate an original unbalanced training set. The original unbalanced training set includes several tool wear feature vectors. The original tool signal is all wear signals collected by the target tool during the complete working cycle. The sample distance generation unit 1002 is used to calculate the sample distance based on the tool wear feature vector and the degradation stage label of the tool wear feature vector; The sample local topological hardness generation unit 1003 is used to generate sample local topological hardness for each tool wear feature vector based on the distance between samples. The difficult-to-classify seed set construction unit 1004 is used to filter the original imbalanced training set according to the local topological hardness of the samples in order to construct the difficult-to-classify seed set. The degradation direction constraint term generation unit 1005 is used to calculate the degradation direction of adjacent stages based on the tool wear feature vector at the category center of each degradation stage, and generate degradation direction constraint terms based on the degradation direction of adjacent stages. The local perturbation term generation unit 1006 is used to generate local perturbation terms based on the local neighborhood covariance matrix of the same type of seed set that is difficult to classify. The enhanced training set generation unit 1007 is used to generate new samples based on the hard-to-classify seed set and combined with the degradation direction constraint term and the local perturbation term, and to generate an enhanced training set based on the new samples and the original imbalanced training set. The recognition model training unit 1008 is used to perform multi-stage wear state recognition and training on the tool wear state recognition model by enhancing the training set.

[0106] Optionally, the local perturbation term generation unit 1006 specifically includes: Calculate the local neighborhood covariance matrix of the same type based on the difficult-to-classify seed set; The characteristic fluctuation range of the local neighborhood is generated based on the covariance matrix of the local neighborhood of the same type, the preset shrinkage coefficient and regularization term; Local disturbance terms are generated based on preset disturbance intensity coefficients and the fluctuation range of similar local neighborhood characteristics.

[0107] Optionally, the inter-sample distance generation unit 1002 specifically includes: Generate the covariance matrix of training samples based on the tool wear feature vector; The inter-sample distance between tool wear feature vectors is calculated based on the training sample covariance matrix, the preset regularization coefficient, the tool wear feature vector, and the degradation stage label of the tool wear feature vector.

[0108] Optionally, the sample local topological hardness generation unit 1003 specifically includes: Generate a set of nearest neighbor feature vectors for each tool wear feature vector based on the distance between samples; Generate feature vector distance weights between tool wear feature vectors based on the distance between samples; The sample local topological hardness is generated for each tool wear feature vector based on the feature vector distance weight and the set of nearest neighbor feature vectors.

[0109] Optionally, after the training set generation unit 1001 and before the inter-sample distance generation unit 1002, the tool wear state recognition device further includes: The feature standardization processing unit is used to perform feature standardization processing on the tool wear feature vector in the original unbalanced training set to eliminate the influence of different feature dimensions; The wear stage calibration unit is used to calibrate the wear stage of the tool wear feature vector, and to assign an initial degradation stage label to each tool wear feature vector.

[0110] Optionally, after the enhancement training set generation unit and before the identification model training unit, the tool wear state identification device further includes: The nearest neighbor detection set generation unit is used to generate a nearest neighbor detection set based on the preset nearest neighbor number value of the tool wear feature vector in the enhanced training set; The same-type nearest neighbor ratio calculation unit is used to calculate the same-type nearest neighbor ratio for each nearest neighbor detection set; The elimination processing unit is used to eliminate tool wear feature vectors in the enhanced training set according to the proportion of similar nearest neighbors; The screening unit is used to assist in screening tool wear feature vectors that deviate from the local manifold structure using the isolated forest method.

[0111] Optionally, after the screening unit, the tool wear condition identification device further includes: The degradation consistency parameter generation unit is used to generate degradation consistency parameters for the augmented training set based on the degradation direction of adjacent stages and the difficult-to-classify seed set. The consistency detection unit is used to detect the consistency between the sample expansion direction and the degradation direction of the augmented training set based on the degradation consistency parameter.

[0112] Optionally, after the screening unit, the tool wear condition identification device further includes: The distribution fidelity analysis unit is used to perform distribution fidelity analysis on the augmented training set according to the Wasserstein distance evaluation method to measure the overall difference between the distribution of the real minority class samples and the distribution of the generated samples.

[0113] Optionally, after the screening unit, the tool wear condition identification device further includes: The same-class boundary safety coefficient generation unit is used to generate the same-class boundary safety coefficient based on the relative distance from the enhanced training set to the same-class samples; The outlier boundary risk coefficient generation unit is used to generate outlier boundary risk coefficients based on the relative distance from the augmented training set to outlier samples. The boundary security analysis unit is used to perform boundary security analysis on the augmented training set based on the boundary security coefficient of similar boundaries and the risk coefficient of dissimilar boundaries.

[0114] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0115] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0116] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0117] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0118] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for identifying tool wear state by combining degradation mechanism and sample features, characterized in that, include: Feature vectors are extracted from the original tool signal to generate an original unbalanced training set. The original unbalanced training set includes several tool wear feature vectors. The original tool signal is all wear signals collected by the target tool during a complete working cycle. The distance between samples is calculated based on the tool wear feature vector and the degradation stage label of the tool wear feature vector; Based on the distance between samples, a sample local topological hardness is generated for each of the tool wear feature vectors; The original imbalanced training set is filtered based on the local topological hardness of the samples to construct a seed set that is difficult to classify. The degradation direction of adjacent stages is calculated at the center of each degradation stage category based on the tool wear feature vector, and a degradation direction constraint term is generated based on the degradation direction of adjacent stages. A local perturbation term is generated based on the local neighborhood covariance matrix of the same type of the difficult-to-classify seed set; Based on the difficult-to-classify seed set, new samples are generated by combining the degradation direction constraint term and the local perturbation term. An enhanced training set is then generated based on the new samples and the original imbalanced training set. The enhanced training set is used to perform multi-stage wear condition identification and training on the tool wear condition identification model.

2. The tool wear condition identification method according to claim 1, characterized in that, The step of generating local perturbation terms based on the covariance matrix of the local neighborhood of the same type in the difficult-to-classify seed set specifically includes: Calculate the local neighborhood covariance matrix of the same type based on the difficult-to-classify seed set; The fluctuation range of local neighborhood features is generated based on the covariance matrix of the same type of local neighborhood, the preset shrinkage coefficient, and the regularization term. A local disturbance term is generated based on a preset disturbance intensity coefficient and the fluctuation range of the same type of local neighborhood characteristics.

3. The tool wear condition identification method according to claim 1, characterized in that, The step of calculating the distance between samples based on the tool wear feature vector and the degradation stage label of the tool wear feature vector specifically includes: Generate a training sample covariance matrix based on the tool wear feature vector; The inter-sample distance between the tool wear feature vectors is calculated based on the training sample covariance matrix, the preset regularization coefficient, the tool wear feature vector, and the degradation stage label of the tool wear feature vector.

4. The tool wear condition identification method according to claim 3, characterized in that, The step of generating local topological hardness for each tool wear feature vector based on the inter-sample distance specifically includes: A nearest neighbor feature vector set is generated for each tool wear feature vector based on the distance between the samples; Generate feature vector distance weights between the tool wear feature vectors based on the distances between the samples; Based on the feature vector distance weights and the nearest neighbor feature vector set, a sample local topological hardness is generated for each of the tool wear feature vectors.

5. The tool wear condition identification method according to claim 1, characterized in that, After the step of extracting feature vectors from the original tool signal to generate an original unbalanced training set, and before the step of calculating the inter-sample distance based on the tool wear feature vector and degradation stage label, the tool wear state identification method further includes: The tool wear feature vectors in the original unbalanced training set are subjected to feature standardization to eliminate the influence of different feature dimensions; The tool wear feature vector is labeled with wear stage, and an initial degradation stage label is associated with each tool wear feature vector.

6. The tool wear condition identification method according to any one of claims 1 to 5, characterized in that, After the steps of generating new samples based on the difficult-to-classify seed set and combining the degradation direction constraint term and the local perturbation term, and generating an enhanced training set based on the new samples and the original unbalanced training set, and before the steps of performing multi-stage wear state recognition and training on the tool wear state recognition model using the enhanced training set, the tool wear state recognition method further includes: A nearest neighbor detection set is generated based on the preset nearest neighbor count value and the tool wear feature vector in the enhanced training set; Calculate the proportion of similar neighbors for each nearest neighbor detection set; The tool wear feature vectors in the enhanced training set are removed based on the ratio of similar nearest neighbors. The isolated forest method was used to assist in screening tool wear feature vectors that deviate from the local manifold structure.

7. The tool wear condition identification method according to claim 6, characterized in that, Following the step of using the isolated forest method to assist in screening tool wear feature vectors that deviate from the local manifold structure, the tool wear state identification method further includes: Based on the degradation direction of adjacent stages and the difficult-to-classify seed set, degradation consistency parameters are generated for the enhanced training set. The consistency between the sample expansion direction and the degradation direction is detected in the augmented training set based on the degradation consistency parameter.

8. The tool wear condition identification method according to claim 6, characterized in that, Following the step of using the isolated forest method to assist in screening tool wear feature vectors that deviate from the local manifold structure, the tool wear state identification method further includes: The distribution fidelity analysis of the augmented training set was performed using the Wasserstein distance evaluation method to measure the overall difference between the distribution of the real minority class samples and the distribution of the generated samples.

9. The tool wear condition identification method according to claim 6, characterized in that, Following the step of using the isolated forest method to assist in screening tool wear feature vectors that deviate from the local manifold structure, the tool wear state identification method further includes: Generate a boundary safety coefficient for the same type based on the relative distance from the enhanced training set to samples of the same type; The out-of-class boundary risk coefficient is generated based on the relative distance from the enhanced training set to the out-of-class samples; Boundary security analysis is performed on the enhanced training set based on the same-type boundary security coefficient and the different-type boundary risk coefficient.

10. A tool wear state identification device that combines degradation mechanism and sample characteristics, characterized in that, include: The training set generation unit is used to extract feature vectors from the original tool signal and generate an original unbalanced training set. The original unbalanced training set includes several tool wear feature vectors, and the original tool signal is all wear signals collected by the target tool during a complete working cycle. The sample distance generation unit is used to calculate the sample distance based on the tool wear feature vector and the degradation stage label of the tool wear feature vector; The sample local topological hardness generation unit is used to generate sample local topological hardness for each of the tool wear feature vectors based on the distance between the samples. A difficult-to-classify seed set construction unit is used to filter the original imbalanced training set according to the local topological hardness of the samples in order to construct a difficult-to-classify seed set. The degradation direction constraint term generation unit is used to calculate the degradation direction of adjacent stages at the center of each degradation stage category based on the tool wear feature vector, and generate degradation direction constraint terms based on the degradation direction of adjacent stages. The local perturbation term generation unit is used to generate local perturbation terms based on the local neighborhood covariance matrix of the same type of the difficult-to-classify seed set; An enhanced training set generation unit is used to generate new samples based on the hard-to-classify seed set and in combination with the degradation direction constraint term and the local perturbation term, and to generate an enhanced training set based on the new samples and the original imbalanced training set. The identification model training unit is used to perform multi-stage wear state identification and training on the tool wear state identification model using the enhanced training set.

Images