A wear state monitoring method for relief tooth turning tools based on conservative adaptive sampling

Abstract

The present invention relates to the field of mechanical numerical control technology, and specifically to a method for monitoring the wear state of a shovel-tooth turning tool based on conservative adaptive sampling. The method comprises the following steps: S1, data acquisition, using an acceleration sensor and a deep depth of field microscope to obtain tool wear and vibration data; S2, data denoising, using a hard threshold wavelet denoising algorithm to denoise the vibration data of the tool wear state; S3, feature extraction, using a principal component analysis method PCA to extract features from the denoised vibration data, and obtain unbalanced tool wear state data as sample data; S41, introducing a Gaussian mixture model GMM of an unsupervised clustering framework, the GMM uses Gaussian probability density to identify the clustering center of a minority sample; S42, according to the near-field density belief level, calculating the number of samples that need to be synthesized for each minority sample; S43, generating artificial samples through a sample synthesis strategy.

Description

Technical Field

[0001] The invention relates to the technical field of mechanical numerical control, and in particular to a wear state monitoring method of a shovel-tooth turning tool based on conservative adaptive sampling. Background Art

[0002] As a special tool for machining gear hobs, the wear state of the tooth-removing turning tool during the tooth-removing process has become a key factor affecting the gear hob size, shape, position accuracy, surface quality, machining efficiency and production safety. Accurate identification of tool wear status is crucial.

[0003] At present, manual identification of tool wear status requires heavy reliance on processing experience and has too high production costs. Therefore, it is of great significance to realize data-driven non-manual identification of tool wear status.

[0004] At present, the technical methods in the field of tool wear identification mainly use the physical signals (such as vibration, force, electrical signals, etc.) generated by the tool during operation as the basis of the monitoring method. Its typical technical strategy is to use algorithms such as wavelet denoising and wavelet packet denoising to denoise the data, use feature extraction methods to extract features from the tool wear state data, and finally use machine learning shallow classifiers to identify data related to the tool wear state. Alternatively, a deep learning related model is used to identify the tool wear state by using the deep model's learning ability for features. It is worth noting that data-driven tool wear state identification methods mainly emphasize data collection, processing and decision-making. Therefore, existing methods often ignore the basic mechanism issues related to tool wear itself. Specifically, throughout the service life of a tool, it usually goes through a series of stages, starting with initial wear, followed by normal wear, and finally severe wear. Among them, the normal wear stage lasts the longest, while the initial wear stage lasts the shortest. Therefore, the tool wear state data presents characteristics such as small samples and imbalance, which poses a huge challenge to data-driven tool wear state identification and monitoring, resulting in inaccurate tool wear state identification and monitoring. Summary of the invention

[0005] The technical problem solved by the present invention is to provide a method for monitoring the wear state of a shovel-tooth turning tool based on conservative adaptive sampling, which can formulate an intelligent sample synthesis strategy for tool wear state data of various unbalanced small-category samples, and solve the problem of tool wear state identification under small sample and unbalanced conditions.

[0006] The basic solution provided by the present invention is a method for monitoring the wear state of a shovel-tooth turning tool based on conservative adaptive sampling, comprising the following steps:

[0007] S1, data acquisition, using acceleration sensor and depth of field microscope to obtain tool wear and vibration data;

[0008] S2, data denoising, using hard threshold wavelet denoising algorithm to denoise the vibration data of tool wear state;

[0009] S3, feature extraction, using the principal component analysis method PCA, to extract features from the denoised vibration data to obtain unbalanced tool wear state data as sample data;

[0010] S4, using the conservative adaptive sampling algorithm CASS to balance the unbalanced tool wear state data, wherein S4 includes the following steps:

[0011] S41, introducing a Gaussian mixture model GMM of an unsupervised clustering framework, wherein the GMM uses Gaussian probability density to identify cluster centers of a small number of samples;

[0012] S42. Calculate the number of samples required to synthesize each minority sample according to the near-field density belief level;

[0013] S43, generating artificial samples through sample synthesis strategy;

[0014] S5. State recognition: using the basic classifier to identify and monitor the tool wear state.

[0015] Further, the S41 comprises the following steps:

[0016] S411, set the number of minority class sample clusters, and calculate the central sample of each cluster ;

[0017] S412, let the minority class sample set be , where the number of minority class samples is , each minority class sample The cluster distribution density is:

[0018]

[0019] in is the nth component of the Gaussian mixture model GMM, is the weight of each category, is the mean of the samples under Gaussian distribution, is the variance of the sample under Gaussian distribution;

[0020] S413, use the expectation maximization EM algorithm to estimate the three parameters of GMM:

[0021] a. Set the number of minority sample clusters n, and , , The initial value of , , Calculate the posterior probability :

[0022]

[0023] b. Update , , :

[0024]

[0025]

[0026]

[0027] judge , , Convergence or not? If not, execute steps a and b until convergence. If yes, then according to the updated , , Divide the samples into different clusters and calculate the central sample of each cluster :

[0028]

[0029] in Indicates the sample format within the cluster to which the sample belongs. Represents the minority class samples in the same cluster.

[0030] Further, the S42 comprises the following steps:

[0031] S421. Use the k-nearest neighbor algorithm to find minority samples The k nearest neighbor samples of , count the number of minority samples in k samples and the number of samples of the majority class ;

[0032] S422, according to the confidence of the near-field density :

[0033]

[0034]

[0035] Where G is the total number of samples that need to be synthesized for the imbalanced dataset, For minority class samples The number of minority class samples that need to be synthesized around.

[0036] Further, the S43 comprises the following steps:

[0037] S431. Using minority samples The judgment factor N, when the minority class sample When it is a trustworthy conservative sample, a new sample is synthesized using the following formula:

[0038]

[0039] When the minority class samples When it is a radical sample, a new sample is synthesized by the following formula:

[0040]

[0041] in for A randomly selected minority class sample from the neighborhood sample, is a random number between 0 and 1. is the sample at the center of the sample cluster.

[0042] The present invention proposes a framework for the wear state monitoring method of the tooth-cutting turning tool based on conservative adaptive sampling (CASS). The framework mainly includes six parts: data acquisition, data denoising, feature extraction, data balancing, basic classifier construction and tool wear state recognition and monitoring. In the data acquisition part, the acceleration sensor and the depth of field microscope are used to obtain quantitative tool wear and vibration data; then the data denoising part is the data denoising part, and the vibration data of the tool wear state is denoised using the hard threshold wavelet denoising algorithm; based on the denoising, the feature extraction part is the principal component analysis (PCA) method to extract features from the denoised data to obtain the unbalanced tool wear state data; after the feature extraction is completed, the data balancing stage is entered, and the conservative adaptive sampling (CASS) algorithm proposed in this paper is used to balance the unbalanced tool wear state data to achieve balancing and enhancement of the original tool wear state data; then the basic classifier, such as the support vector machine (SVM), is used to identify and monitor the tool wear state. Therefore, the proposed method includes a complete tool wear state identification and monitoring technology, and effectively solves the small sample and imbalance problems of tool wear state data through the conservative adaptive sampling (CASS) method, thereby improving the accuracy of tool wear state identification and monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of an embodiment of a method for monitoring wear state of a relief-tooth turning tool based on conservative adaptive sampling according to the present invention;

[0044] Figure 2It is a schematic diagram of denoising results of an embodiment of a method for monitoring wear state of a relief-tooth turning tool based on conservative adaptive sampling according to the present invention;

[0045] Figure 3 It is a flow chart of EEMD algorithm calculation of an embodiment of a wear state monitoring method of a relief tooth turning tool based on conservative adaptive sampling of the present invention;

[0046] Figure 4 This is a graph showing the feature extraction results of an EEMD algorithm of an embodiment of a wear state monitoring method for a skive-tooth turning tool based on conservative adaptive sampling according to the present invention;

[0047] Figure 5 A visualization diagram of a PHM2010 tool data set in an embodiment of a wear state monitoring method for a relief-tooth turning tool based on conservative adaptive sampling of the present invention;

[0048] Figure 6 A schematic diagram of a forklift tool wear state data acquisition experimental platform according to an embodiment of a forklift tool wear state monitoring method based on conservative adaptive sampling of the present invention;

[0049] Figure 7 The distribution of original PHM2010 data (“initial wear / normal wear”) of an embodiment of a wear state monitoring method for a relief-tooth turning tool based on conservative adaptive sampling of the present invention;

[0050] Figure 8 The data distribution of PHM2010 after CASS balancing processing ("initial wear / normal wear") of an embodiment of a wear state monitoring method for a relief tooth turning tool based on conservative adaptive sampling of the present invention;

[0051] Figure 9 The distribution of original PHM2010 data ("severe wear / normal wear") of an embodiment of a wear state monitoring method for a relief-tooth turning tool based on conservative adaptive sampling of the present invention;

[0052] Figure 10 The data distribution of PHM2010 after CASS balancing processing ("severe wear / normal wear") of an embodiment of a wear state monitoring method for a relief-tooth turning tool based on conservative adaptive sampling of the present invention;

[0053] Figure 11 The experimental results of PHM2010 under the Accuracy index of an embodiment of a wear state monitoring method for a relief-tooth turning tool based on conservative adaptive sampling of the present invention;

[0054] Figure 12 The experimental results of an embodiment of the wear state monitoring method of a relief-tooth turning tool based on conservative adaptive sampling in the present invention are obtained on the PHM2010 public milling cutter data set;

[0055] Figure 13 The experimental results of an embodiment of the wear state monitoring method of a relief-tooth turning tool based on conservative adaptive sampling in the present invention are obtained on the PHM2010 public milling cutter data set;

[0056] Figure 14 The experimental results of PHM2010 under AUC index and ROC curve of an embodiment of a wear state monitoring method for a relief tooth turning tool based on conservative adaptive sampling of the present invention. DETAILED DESCRIPTION

[0057] The following is further described in detail through specific implementation methods:

[0058] The embodiment is basically as shown in the attached Figure 1 As shown:

[0059] A wear state monitoring method for a shaving tool based on conservative adaptive sampling comprises the following steps:

[0060] S1. Data acquisition, using acceleration sensor and depth of field microscope to obtain tool wear and vibration data.

[0061] S2. Data denoising: using hard threshold wavelet denoising algorithm to denoise the vibration data of tool wear state.

[0062] S3. Feature extraction: Use the principal component analysis method PCA to extract features from the denoised vibration data and obtain unbalanced tool wear state data as sample data.

[0063] Specifically, data denoising uses Daubechies8 (db8) wavelet hard thresholding technology to denoise the original collected data. Assume that the original collected noisy vibration signal is , which contains high-quality signals and noisy signals ,but The expression is:

[0064]

[0065] Using the db8 wavelet hard threshold denoising algorithm to perform simultaneous wavelet changes on both sides are:

[0066]

[0067] in, Indicates that the signal Wavelet transform is performed within the interval, and then the denoised signal is reconstructed

[0068] The results of denoising the public tool data (PHM2010) using the db8 wavelet hard threshold denoising algorithm are shown in the figure. Figure 2 As shown:

[0069] On the basis of data denoising, Ensemble Empirical Mode Decomposition (EEMD) is used to obtain features from the denoised signal. The calculation process of the EEMD algorithm is as follows: Figure 3 As shown:

[0070] The steps for feature extraction of denoised data using EEMD are:

[0071] (1) Assume that the denoised data is used as the original signal for feature extraction ;

[0072] (2) Set the number of times the original signal needs to be processed N, and then add random white noise of different amplitudes (with a mean of 0) to each of the N original signals to form a new signal such as ;

[0073] (3) Then, the EMD algorithm is used to decompose the series of new signals to obtain a series of IMF components such as ;

[0074] (4) Then, the IMF components of the corresponding modes are averaged to obtain the decomposition results of EEMD.

[0075] The results of feature extraction using EEMD on the denoised public tool data set (PHM2010) are shown in the figure below. Figure 4 As shown:

[0076] S4, using the conservative adaptive sampling algorithm CASS to balance the unbalanced tool wear state data, wherein S4 includes the following steps:

[0077] S41, introducing a Gaussian mixture model GMM of an unsupervised clustering framework, wherein the GMM uses Gaussian probability density to identify cluster centers of a small number of samples;

[0078] S42. Calculate the number of samples required to synthesize each minority sample according to the near-field density belief level;

[0079] S43, generating artificial samples through sample synthesis strategy;

[0080] S5. State recognition: using the basic classifier to identify and monitor the tool wear state.

[0081] The S41 comprises the following steps:

[0082] S411, set the number of minority class sample clusters, and calculate the central sample of each cluster ;

[0083] S412, let the minority class sample set be , where the number of minority class samples is , each minority class sample The cluster distribution density is:

[0084]

[0085] in is the nth component of the Gaussian mixture model GMM, is the weight of each category, is the mean of the samples under Gaussian distribution, is the variance of the sample under Gaussian distribution;

[0086] S413, use the expectation maximization EM algorithm to estimate the three parameters of GMM:

[0087] a. Set the number of minority sample clusters n, and , , The initial value of , , Calculate the posterior probability :

[0088]

[0089] b. Update , , :

[0090]

[0091]

[0092]

[0093] judge , , Convergence or not? If not, execute steps a and b until convergence. If yes, then according to the updated , , Divide the samples into different clusters and calculate the central sample of each cluster :

[0094]

[0095] in Indicates the sample format within the cluster to which the sample belongs. Represents the minority class samples in the same cluster.

[0096] The S42 comprises the following steps:

[0097] S421. Use the k-nearest neighbor algorithm to find minority samples The k nearest neighbor samples of , count the number of minority samples in k samples and the number of samples of the majority class ;

[0098] S422, according to the confidence of the near-field density :

[0099]

[0100]

[0101] Where G is the total number of samples that need to be synthesized for the imbalanced dataset.

[0102] This embodiment mainly deals with the classification of unbalanced data, which is mainly divided into two types of data (majority class samples and minority class samples). In the entire large data (or original data), the number of minority class samples is expressed as ns, and the number of majority class samples is expressed as nl. In order to balance these two types of data, the sampling method (CASS) of this application is used to expand the minority class data. For the entire unbalanced data, the total number of minority class samples that need to be synthesized is G, which is equal to G=ns-nl, that is, G is equal to the number of majority class samples minus the number of minority class samples. This is a main idea of ​​this application. The main idea of ​​expansion can be understood as: in each minority class sample Synthesize a certain number of minority samples around ,and Together they make G.

[0103] In the process of expanding the data, it is necessary to use the k-nearest neighbor algorithm (KNN) to find minority class samples. The k nearest neighbor samples of represents a minority class sample), find the k nearest neighbor samples, which can be understood as surrounding See what categories the k samples around it are (how many minority class samples are there, how many majority class samples are there, this is around After finding these k samples, count the number of minority samples in the k samples. and the number of samples of the majority class (This is around The samples in a small circle are divided into the number of minority samples and majority samples) and then the near-field density belief level is calculated according to the first formula in S422 , and finally calculate each The number of samples to be synthesized For the formula in S422, there are two cases. In the first case, when When, that is to say, around Among the k samples, when there are both minority class samples and many class samples, calculate the near-field density belief level Is to use around The number of samples with a small number among the k samples is divided by (k-1). It can be understood that this number is a number greater than 0 and less than or equal to 1, which is a ratio. The second case is when , that is to say, around In the case of k samples, only minority class samples or free majority class samples are used to calculate the near-field density belief level , is to use around The number of minority class samples in the k samples is divided by 2k, which is a number equal to 0 or half.

[0104] Then according to the second formula in S422, For minority class samples The number of minority samples that need to be synthesized around, Equal to the total number of samples G that need to be synthesized for the entire imbalanced dataset, multiplied by one Weighted (that is, each A ratio in the collection) This calculates each minority class sample The number of samples to be synthesized , Put together they are G, which is the number of samples G needed to synthesize the entire data.

[0105] The S43 comprises the following steps:

[0106] S431. Using minority samples The judgment factor N, when the minority class sample When it is a trustworthy conservative sample, a new sample is synthesized using the following formula:

[0107]

[0108] When the minority class samples When it is a radical sample, a new sample is synthesized by the following formula:

[0109]

[0110] in for A randomly selected minority class sample from the neighborhood samples.

[0111] Using minority class samples The judgment factor N, which can be understood as the near-field density belief level calculated previously ,when When it is a trustworthy conservative sample (a minority sample far away from the classification decision boundary, that is, when N is equal to 0 or half), a new sample is synthesized by the first formula in S431. When it is an aggressive sample (a minority sample with a risk of sample aliasing, that is, when N is greater than 0 and less than or equal to 1), a new sample is synthesized using the second formula in S431. This formula can be understood as selecting The sample at the center of the sample cluster is represented by , then And the samples in this sample cluster The difference is performed on the connection line to generate a new minority class sample. The above is the synthesis strategy of the new sample.

[0112] The present invention proposes a framework for the wear state monitoring method of the tooth-cutting turning tool based on conservative adaptive sampling (CASS). The framework mainly includes six parts: data acquisition, data denoising, feature extraction, data balancing, basic classifier construction and tool wear state recognition and monitoring. In the data acquisition part, the acceleration sensor and the depth of field microscope are used to obtain quantitative tool wear and vibration data; then the data denoising part is the data denoising part, and the vibration data of the tool wear state is denoised using the hard threshold wavelet denoising algorithm; based on the denoising, the feature extraction part is the principal component analysis (PCA) method to extract features from the denoised data to obtain the unbalanced tool wear state data; after the feature extraction is completed, the data balancing stage is entered, and the conservative adaptive sampling (CASS) algorithm proposed in this paper is used to balance the unbalanced tool wear state data to achieve balancing and enhancement of the original tool wear state data; then the basic classifier, such as the support vector machine (SVM), is used to identify and monitor the tool wear state. Therefore, the proposed method includes a complete tool wear state identification and monitoring technology, and effectively solves the small sample and imbalance problems of tool wear state data through the conservative adaptive sampling (CASS) method, thereby improving the accuracy of tool wear state identification and monitoring.

[0113] In order to verify whether CASS can effectively improve the classification and recognition performance of the basic classifier and solve the tool wear state monitoring problem in a small sample imbalance scenario, this example uses a public milling cutter dataset (PHM2010) and a self-made shovel-tooth turning tool dataset.

[0114] The PHM2010 dataset is a public dataset for the milling cutter wear competition released by the New York Association for Prognostics and Health Management in 2010. The experimental conditions are shown in Table 1:

[0115] Table 1

[0116] Hardware conditions Model and parameters Cutting conditions and parameters CNC milling machine High speed CNC machine tool (Roders Tech RFM760) Spindle speed:10400 (r / min) Dynamometer Three direction dynamometer(Kistler 9265B) Feed speed:1555(mm / min) Charge amplifier Multichannel charge amplifier (Kistler 5019A) Axial cutting depth:0.2(mm) Milling materials square (Inconel 718) Radial cutting width:0.125(mm) Tool Ball end carbide milling cutter (3 teeth) Single cycle to feed rate:0.001(mm) data acquisition card NI DAQ Frequency of application:50(kHz) Wear gauge LEICA MZ12electron microscope Cutting condition: dry cutting

[0117] PHM2010 contains experimental data of 6 milling cutters (C1 to C6), as well as 7 features such as milling force, vibration signal, and root mean square value of acoustic emission in the X, Y, and Z directions during milling. This paper uses the C4 dataset with labeled data to construct three small sample unbalanced tool wear state datasets at the 30th, 140th, and 280th cutting times for experiments. The constructed PHM2010 milling cutter experimental training dataset is shown in Table 2.

[0118] Table 2

[0119] Training dataset Data1(Initial / Normal) Data2(Normal / Sharp) Data3(Initial +Normal+Sharp) Imbalanceratio （50 / 500）1 / 10 （50 / 500）1 / 10 (50 / 500 / 50)Classifier training dataset Data aftersampling （500 / 500）1 / 1 （500 / 500）1 / 1 (500 / 500 / 500)Classifier trainingdata set

[0120] The visualization of the C4 dataset in PHM2010 is as follows Figure 5 shown.

[0121] Research on tool wear status recognition and monitoring is mainly focused on milling cutters, with less research on turning tools. In order to evaluate the performance of the CASS algorithm for various tool types, this embodiment collects the actual full life cycle processing data of the shovel-tooth turning tool. The collection conditions of the shovel-tooth turning tool data are shown in Table 3:

[0122] Table 3

[0123] Hardware conditions Model and parameters Cutting conditions and parameters CNC shovel toothlathe CNC shovel tooth lathe(HTM-Ⅱ) Spindle speed: n=(1000V) / (πD)=63.66 (r / min) Tool Shovel tooth turning tool (W6Mo5Cr4V2 highspeed steel) Turning speed: 30 (m / min) Turning material Gear hob (W18Cr4V high speed steel) Diameter 150 (mm) Accelerometer 3-axis accelerometer (INV9832-50) X-axis feed: 0.05 (mm / r) data acquisition card Vibration signal acquisition instrument (INV3062T0) Frequency of application: 10240 (Hz) Wear gauge Super Depth of Field Microscope (ZW-C3600) Cutting condition: dry cutting

[0124] Collection platform such as Figure 6 shown.

[0125] This embodiment selects 1000 initial wear stage data, 4000 normal wear stage data and 1000 severe wear stage data of the relief tooth turning tool to construct the wear state data set of the relief tooth turning tool under the condition of small sample imbalance.

[0126] Traditional classification methods usually rely on the accuracy index to evaluate the effectiveness of the algorithm. However, this index can only reflect the classification performance of the classifier as a whole. For unbalanced data, the classification performance of the minority class is often more important. In order to effectively evaluate the effectiveness of the CASS algorithm, this paper considers F-measure, G-mean, ROC curve, AUC and Accuracy as evaluation indicators. The calculation method of each indicator is as follows:

[0127]

[0128]

[0129]

[0130]

[0131]

[0132] Among them, TP is the number of correctly classified positive (minority class) samples, TN is the number of correctly classified negative (majority class) samples, TN is the number of correctly classified negative samples, and FN is the number of incorrectly classified negative samples. Among them, F-measure, G-mean, AUC The larger the value, the higher the score, indicating a better classification effect.

[0133] Tool wear status identification and monitoring is a complex problem involving multiple classifications. To this end, we adopted the One-vs-One strategy and used a sampling algorithm to balance the samples of "initial wear / normal wear" and "severe wear / normal wear" to ensure that the sample size of each wear state is reasonable. This embodiment uses the grid search method to determine the k value of each sampling algorithm. This embodiment visualizes the sample distribution after the CASS algorithm balances the PHM2010 milling cutter data set. CASS can synthesize new samples based on the original distribution of the PHM2010 milling tool wear data, thereby effectively increasing high-quality new samples and improving the imbalance of tool wear status data.

[0134] The data distribution of the original PHM2010 tool "initial wear / normal wear" is as follows Figure 7 As shown in Figure 2, the data visualization results after data amplification and balancing of “initial wear / normal wear” using CASS are shown in Figure 2. Figure 8 The data distribution of the original PHM2010 tool "severe wear / normal wear" is shown in Figure 9 As shown in Figure 2, the data visualization results after data amplification and balancing of “severe wear / normal wear” using CASS are shown in Figure 2. Figure 10 shown.

[0135] Figure 11 The accuracy results of wear state recognition and monitoring of PHM2010 milling cutter data using different classifiers and sampling algorithms under small sample imbalance are shown. It can be seen that when the sampling algorithm is not used, the accuracy of most classifiers is not high, especially the LR classifier, which shows that the small sample imbalance problem of tool wear state data has brought challenges to the classifier, but at the same time, after being processed by the sampling method, the accuracy of all classifiers has improved, which shows that the sampling method can solve the small sample imbalance of tool wear state data to a certain extent. Further analysis shows that the CASS proposed in this paper performs best among multiple sampling methods, which shows the superiority of the CASS proposed in this paper.

[0136] It should be noted that although Accuracy can express the performance of the model as a whole, it cannot explain the accuracy of the classification algorithm in each category. Especially for the identification and monitoring of tool wear status, accurate identification and monitoring of minority samples (samples in the severe wear stage) is particularly important. Therefore, in addition to the Accuracy indicator, F-measure and G-mean indicators are needed to test the classification accuracy of the algorithm on minority samples. Figure 12 and Figure 13 These are the results of experiments on the PHM2010 public milling cutter dataset. It can be seen that CASS performs best under the F-measure and G-mean indicators and has a smaller algorithm error. This shows that compared with other sampling methods, CASS can better solve the problem of tool wear status identification and monitoring under small sample balance conditions.

[0137] Figure 14 Figure 2 is the ROC curve and AUC. It can be seen that the ROC curve of CASS is mostly located in the upper left corner of the picture, indicating that compared with other sampling algorithms, CASS has better stability.

[0138] The above are only embodiments of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme is not described in detail here. Ordinary technicians in the relevant field know all the common technical knowledge in the technical field to which the invention belongs before the application date or priority date, can obtain all the existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A wear state monitoring method for a shaving tool based on conservative adaptive sampling, characterized in that: The following steps are involved: S1, data acquisition, using acceleration sensor and depth of field microscope to obtain tool wear and vibration data; S2, data denoising, using hard threshold wavelet denoising algorithm to denoise the vibration data of tool wear state; S3, feature extraction, using the principal component analysis method PCA, to extract features from the denoised vibration data to obtain unbalanced tool wear state data as sample data; S4, using the conservative adaptive sampling algorithm CASS to balance the unbalanced tool wear state data, wherein S4 includes the following steps: S41, introducing a Gaussian mixture model GMM of an unsupervised clustering framework, wherein the GMM uses Gaussian probability density to identify cluster centers of a small number of samples; S42. Calculate the number of samples that need to be synthesized for each minority sample according to the near-field density belief level; The S42 comprises the following steps: S421. Use the k-nearest neighbor algorithm to find minority samples The k nearest neighbor samples of , count the number of minority samples in k samples and the number of samples of the majority class ; S422, according to the confidence of the near-field density : Where G is the total number of samples that need to be synthesized for the imbalanced dataset, For minority class samples The number of surrounding minority samples that need to be synthesized; S43, generating artificial samples through sample synthesis strategy; The S43 comprises the following steps: S431. Using minority samples The judgment factor N, when the minority class sample When it is a trustworthy conservative sample, a new sample is synthesized using the following formula: ； When the minority class samples When it is a radical sample, a new sample is synthesized by the following formula: in for A randomly selected minority class sample from the neighborhood sample, is a random number between 0 and 1. is the sample at the center of the sample cluster; S5. State recognition: using the basic classifier to identify and monitor the tool wear state.

2. The wear state monitoring method of a relief tooth turning tool based on conservative adaptive sampling according to claim 1 is characterized in that: The S41 comprises the following steps: S411, set the number of minority class sample clusters, and calculate the central sample of each cluster ; S412, let the minority class sample set be , where the number of minority class samples is , each minority class sample The cluster distribution density is: in is the nth component of the Gaussian mixture model GMM, is the weight of each category, is the mean of the samples under Gaussian distribution, is the variance of the sample under Gaussian distribution; S413, use the expectation maximization EM algorithm to estimate the three parameters of GMM: a. Set the number of minority sample clusters n, and , , The initial value of , , Calculate the posterior probability : b. Update , , : judge , , Convergence or not? If not, execute steps a and b until convergence. If yes, then according to the updated , , Divide the samples into different clusters and calculate the central sample of each cluster : in Indicates the sample format within the cluster to which the sample belongs. Represents the minority class samples in the same cluster.

Images