A tool breakage state in-situ real-time monitoring method

Abstract

The application discloses a tool breakage state in-situ real-time monitoring method, and belongs to the technical field of tool breakage state monitoring.The method is based on multi-source information fusion features and multi-algorithm combination, aims to realize efficient and accurate tool breakage state real-time monitoring, and can comprehensively utilize multi-dimensional information in a cutting process by integrating multi-dimensional force signals and multi-dimensional vibration signals, so that the monitoring precision and stability are improved.The application adopts a generative adversarial network to enhance unbalanced samples, effectively solves the problem of insufficient minority class samples in a traditional monitoring method, and improves the generalization ability and judgment accuracy of the model.Deep feature extraction is combined with multi-layer threshold decision and a multi-tooth tool breakage intelligent identification system, so that the accuracy and efficiency of tool breakage type identification are further improved.The method is suitable for various mechanical processing scenes, especially in advanced manufacturing fields with high precision and high stability, and can effectively prevent production accidents caused by tool breakage.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of tool breakage state monitoring, and particularly relates to a tool breakage state in-situ real-time monitoring method. BACKGROUND

[0002] Tool breakage state recognition as part of manufacturing process automation technology has been increasingly valued by people, and a reliable tool breakage state monitoring system can guide the machine tool to complete the stop alarm and automatic tool changing operation in time through timely and accurate reflection of tool breakage state information, realize uninterrupted production process, stabilize the machining quality of workpieces and avoid safety hazards caused by tool breakage and broken tools.

[0003] The two main types of tool breakage are tool shank fracture and cutting edge breakage. Tool shank fracture is the most serious failure, which usually occurs under severe cutting conditions, significant impact load or improper operation. Cutting edge breakage is essentially microscopic, making it difficult to check the defects or damage of the tool with the naked eye. If the cutting edge appears jagged or hollow during machining, it means that cutting edge breakage has occurred. Under light breakage conditions, the tool can still be used for rough machining and semi-finishing operations that require low surface quality. However, when the breakage area further expands, the tool will completely lose its cutting ability.

[0004] For both tool shank fracture and cutting edge breakage, different measures should be taken when they occur. Tool shank fracture can cause serious production accidents and should be stopped immediately when it occurs, requiring the highest speed of the relevant algorithm; cutting edge breakage has relatively small impact on machining when it occurs, and the accuracy of the algorithm is the first consideration. Therefore, for the two different situations, multiple algorithms can be used for monitoring at the same time.

[0005] Tool breakage state monitoring methods can mainly be divided into direct monitoring method and indirect monitoring method. Direct monitoring method relies on machine vision, directly measures the change of cutting edge geometry, and provides better inspection consistency and accuracy than manual inspection based on experience, professional knowledge and convention. However, they are easily affected by cutting fluid, lighting conditions and chips, so they need to be measured during downtime, making it difficult to monitor sudden tool breakage during machining. In contrast, data-driven indirect monitoring method is economical and practical in industrial applications, collects physical signals closely related to tool state during machining through different sensors, and then makes decisions with the help of advanced signal processing algorithms and artificial intelligence models, which can achieve real-time monitoring of tool breakage during machining.

[0006] The tool breakage state monitoring method is mainly divided into single signal monitoring method and multi-signal fusion monitoring method according to the signal source. The single signal monitoring method has single signal type and is easy to process, but the time-varying factors in the cutting process, such as tool wear and machine tool vibration, may cause uncertainty in the decision-making process of the monitoring system; the multi-signal fusion monitoring method fully reflects the change of tool state through complementary tool cutting information, so as to improve the accuracy and robustness of monitoring. SUMMARY

[0007] In view of the above-mentioned defects of the existing direct method for measuring the tool breakage state, the tool breakage monitoring method based on single channel signal has low stability, and the single algorithm cannot cover all cases, the present application provides a tool breakage state in-situ real-time monitoring method based on multi-source information fusion features and multi-algorithm combination, aiming at realizing efficient and accurate real-time monitoring of tool breakage state.

[0008] The technical scheme adopted by the present application is:

[0009] A tool breakage state in-situ real-time monitoring method, comprising the following steps:

[0010] Step 1: building a tool state in-situ real-time monitoring data acquisition system, acquiring cutting processing sensing signal sample data and shooting images of the wear area after each tool movement;

[0011] Step 2: measuring the tool wear area image to obtain the label of tool breakage, and constructing a sample data sequence with the sensing signal sample data;

[0012] Step 3: performing data segmentation and standardization processing on the cutting processing signal, and inputting into an unbalanced sample enhancement system for data enhancement;

[0013] Step 4: performing deep feature extraction on the data set after data enhancement, inputting the training set after feature extraction into a classifier for training, and testing with the test set after feature extraction;

[0014] Step 5: inputting the trained classifier in step 4 into a tool breakage state recognition system;

[0015] Step 6: acquiring real-time signals through the tool state in-situ real-time monitoring data acquisition system, inputting the real-time signals into the tool breakage state recognition system to determine the breakage state of the tool.

[0016] Compared with the prior art, the present application has the following beneficial effects:

[0017] 1. By integrating multi-dimensional force signals and multi-dimensional vibration signals, the present application can comprehensively utilize multi-dimensional information in the cutting process, thereby improving the monitoring accuracy and stability.

[0018] 2、The tool breakage state is judged depending on deep features extracted from real-time collected sensing signals, and the high efficiency and accuracy of the monitoring method are ensured through the data enhancement, feature extraction and classifier training process provided by the application.

[0019] 3、The application generates an adversarial network to enhance unbalanced samples, effectively solves the problem of insufficient minority class samples in traditional monitoring methods, and improves the generalization ability and judgment accuracy of the model.Deep feature extraction combined with multi-layer threshold decision and multi-tooth tool breakage intelligent recognition system further enhances the accuracy and efficiency of tool breakage type identification. This method is suitable for various mechanical processing scenes, especially in advanced manufacturing fields requiring high precision and high stability, which can effectively prevent production accidents caused by tool breakage, ensure production safety and improve production efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is the overall operation logic diagram of the application;

[0021] Figure 2 is the schematic diagram of the tool state in-situ real-time monitoring data acquisition system (taking milling as an example);

[0022] Figure 3 is the flowchart of obtaining data set of the application;

[0023] Figure 4 is the flowchart of training Softmax classifier of the application;

[0024] Figure 5 is the structure diagram of generator network and discriminator network of the application;

[0025] Figure 6 is the flowchart of multi-method fusion to judge real-time signal breakage state of the application. DETAILED DESCRIPTION

[0026] In order to better understand the purpose, structure and function of the application, the application will be further described in detail below in combination with the drawings.

[0027] The running principle of the tool breakage state in-situ real-time monitoring method based on multi-source information fusion features and multi-algorithm combination proposed by the application is as shown in Figure 1

[0028] In this embodiment, a VMC855 three-axis high-speed milling machining center is used to perform the machining task. The tool used is an end mill equipped with an APMT1135 blade made of PCD material. The machining material is an aluminum-based silicon carbide composite material containing 65% SiC by volume, which is processed by transverse cutting along multiple paths using an end mill. During the machining process, intelligent force measuring tool shanks and three-axis vibration sensors from Kistler in Switzerland are used, combined with domestic acquisition board cards, to monitor and record cutting force and vibration data in real time at a sampling rate of 8 kHz. In addition, a VISHNUTECH MV-HS2000GM industrial camera is used in conjunction with a double telecentric microscope to accurately measure and record the flank wear width of the blade, in order to evaluate the degree of tool wear. All monitoring data analysis and processing are completed on a computing platform equipped with NVIDIA RTX4090 24GB GPU, using Python 3.9 and PyTorch environment.

[0029] comprising the steps of:

[0030] Step 1: Build a tool state real-time in-situ monitoring data acquisition system to collect cutting processing sensor signal sample data and take images of the wear area after each cutting; the layout diagram is shown in Figure 2 ;

[0031] Step 2: Measure the tool wear area image to obtain the tool damage label, and construct a sample data sequence with the sensor signal sample data;

[0032] Step 3: Perform data segmentation and standardization processing on the cutting processing signal, and input it into the unbalanced sample enhancement system for data enhancement;

[0033] Step 4: Perform deep feature extraction on the data set after data enhancement, input the training set after feature extraction into the classifier for training, and test with the test set after feature extraction;

[0034] Step 5: The trained classifier in step 4 is input into the tool damage state recognition system;

[0035] Step 6: Collect real-time signals through the tool state real-time in-situ monitoring data acquisition system, and input the real-time signals into the tool damage state recognition system to determine the damage state of the tool.

[0036] In this process, as shown in Figure 3 , the data set is obtained by the following steps:

[0037] Step KA1: Build a tool state real-time in-situ monitoring data acquisition system, the layout diagram is shown in Figure 2 ;

[0038] Step KA2: Collecting original data by using the tool state in-situ real-time monitoring data acquisition system, and obtaining the corresponding damage label by observing the image of the wear area;

[0039] Step KA3: Cutting the stable cutting segment of the original signal, and dividing it according to 2000 data points, and there is no overlap between adjacent time windows, obtaining the original sample, and the label of the sample divided from the same segment is the same;

[0040] Step KA4: Standardizing each original sample to make the data distributed in the interval [-1, 1];

[0041] Step KA5: Integrating all standardized samples and their labels into a data set, which contains 550 samples in this example, including 500 normal samples (majority class) and 50 abnormal samples (minority class).

[0042] As shown in the step 1, Figure 2 The tool state in-situ real-time monitoring data acquisition system includes a force signal acquisition system, a vibration signal acquisition system, and an image signal acquisition system;

[0043] The force signal acquisition system collects force signals in the cutting process by installing a three-channel force sensor on the tool holder, including force along the axial feed direction of the workpiece, force along the radial direction of the workpiece, and force along the tangential direction of the workpiece,

[0044] The vibration signal acquisition system collects three-direction vibration signals in the cutting process by installing a three-direction acceleration sensor on the workpiece (milling) or on the tool holder (turning),

[0045] The three-direction vibration includes vibration along the axial feed direction of the workpiece, vibration along the radial direction of the workpiece, and vibration along the tangential direction of the workpiece,

[0046] The image signal acquisition system takes an image of the tool wear area after each pass by an industrial camera or a microscope.

[0047] In the step 2, the label of tool damage includes the tool damage state corresponding to the offline measured wear area image data.

[0048] In the step 3, the data segmentation and standardization processing operation includes:

[0049] The fixed-size time window is used to segment the sensor signal sample data to remove redundant components to obtain stable cutting signals for data segmentation;

[0050] According to the stable cutting signals for data segmentation, the data is segmented according to a certain number of sample points to obtain a plurality of samples;

[0051] According to the sample after the data segmentation, the sample point number is distributed between [-1, 1] by standardizing the sample according to formula (1), and the sequence of the sensing signal sample data is obtained.

[0052] Formula (1) is as follows:

[0053]

[0054] In the formula, x * is the standardized value corresponding to each point in the sample, x is the original value of each point in the sample, max is the maximum value in the sample, and min is the minimum value in the sample.

[0055] The data segmentation according to a certain sample point number comprises the following steps:

[0056] Step 301: Take x data points to segment the stable cutting signal into several segments, and the adjacent time windows do not overlap, and each segment is a sample;

[0057] Step 302: Each sample comprises x data points and a corresponding damage label of the stable cutting signal.

[0058] The training and data enhancement process of the unbalanced sample enhancement system in step 3 comprises:

[0059] Step 311: Establish a data set, each sample in the data set comprises the following parameters: force signal in the cutting process, three-direction vibration signal in the cutting process, and corresponding damage label; divide the data set into a training set and a test set according to a ratio of 8:2, and divide the majority class and the minority class samples based on the same percentage, for example, there are 1000 majority class samples and 100 minority class samples, the training set takes 800 majority class samples and 80 minority class samples, and the training set takes 200 majority class samples and 20 minority class samples;

[0060] Step 312: According to the number of each label in the data set, filter out the minority class samples in the training set to establish a minority class sample set, and train the generator of the generative adversarial network in the unbalanced sample enhancement system by using the minority class sample set;

[0061] The generator and the discriminator objective functions of the generative adversarial network are respectively shown in formula (2) and formula (3) as follows:

[0062]

[0063]

[0064] The generator network of the generative adversarial network comprises an input layer, four hidden layers and an output layer. The input layer accepts a noise vector, the hidden layers are all fully connected layers, each followed by a Leaky ReLU activation function, and the output layer is also a dense layer using tanh as the activation function to generate samples.

[0065] The discriminator network of the generative adversarial network comprises an input layer, four hidden layers and an output layer. The input layer receives real or generated samples, the hidden layers are all composed of fully connected layers, each followed by a LeakyReLU activation function and a Dropout layer. The output layer is a dense layer using a sigmoid activation function to predict whether the sample is real or generated.

[0066] The generator and the discriminator of the generative adversarial network are both trained using the Adam optimizer.

[0067] Step 313: Use the generator trained in step 312 to generate new samples, and through similarity evaluation with the samples in the minority class sample set in step 312, filter out suitable samples to input into the training set in step 1 as expansion samples;

[0068] The similarity evaluation comprises the following steps:

[0069] Step 320: According to the actual situation, set a threshold value for K-L divergence (KLD), Euclidean distance (ED) and Pearson correlation coefficient (PCC) respectively;

[0070] Step 321: Take the new sample generated by the generator, randomly select a real sample from the minority class sample set, and calculate whether the K-L divergence (KLD) between the generated sample and the real sample is less than the threshold value. If it is less than the threshold value, proceed to step 322;

[0071] Step 322: Calculate whether the Euclidean distance (ED) between the generated sample and the real sample in step 321 is less than the threshold value. If it is less than the threshold value, proceed to step C3;

[0072] Step 323: Calculate whether the Pearson correlation coefficient (PCC) between the generated sample and the real sample in step 321 is greater than the threshold value. If it is greater than the threshold value, the sample will be operated subsequently;

[0073] The K-L divergence (KLD), Euclidean distance (ED) and Pearson correlation coefficient (PCC) in the similarity evaluation are calculated according to the following formula (4), formula (5) and formula (6) respectively:

[0074]

[0075]

[0076]

[0077] In the formula, p(x) is the generated data, and p(y) is the generated data.

[0078] In step 4, the data set after data enhancement is subjected to deep feature extraction, comprising:

[0079] The mean, standard deviation, root mean square, skewness, kurtosis, peak value, waveform factor, peak factor and pulse factor of the sensor signal sample data sequence are calculated;

[0080] The sensor signal sample data sequence is subjected to fast Fourier transform (FFT) and the amplitude is calculated to obtain a frequency spectrum line sequence, and the power spectrum mean, frequency center of gravity and mean square frequency of the frequency spectrum line sequence are calculated;

[0081] The sensor signal sample data sequence is subjected to 3-layer db4 wavelet packet transform, and 8 decomposition coefficients are obtained after decomposition, and the decomposed signal is reconstructed to obtain signal components of different frequency bands, and the energy features of the decomposed frequency band signals are extracted to obtain 8 wavelet energy values in different frequency intervals;

[0082] These features are arranged in columns to obtain a signal feature matrix.

[0083]

[0084]

[0085] As shown in Figure 4 , the classifier model is obtained by the following steps:

[0086] Step 401: The data set is randomly divided into a training set and a test set in a ratio of 8:2 according to each label, obtaining a sample quantity of 440 in the data set, including 400 normal samples and 40 abnormal samples, and a sample quantity of 110 in the training set, including 100 normal samples and 10 abnormal samples;

[0087] Step 402: The minority class sample set in the training set is extracted, and the minority class sample set is generally an abnormal label (damage) in damage monitoring, and the majority class sample set generally corresponds to a normal label (wear);

[0088] Step 403: The minority class sample set is input into the generative adversarial network for training, and the structure of the generative adversarial network generator and discriminator is as shown in Figure 5In the embodiment, the number of nodes of the full connection layer of the generator is 400, 800, 1200, 1600, 2000 respectively, the input noise dimension is 100, the parameters of the LeakyReLU layer are all set to 0.2, the number of nodes of the full connection layer of the discriminator is 1600, 1200, 800, 400, 1 respectively, the parameters of the LeakyReLU layer are all set to 0.2, and the parameters of the Dropout layer are all set to 0.4.

[0089] Step 404: stop training when the generative adversarial network reaches Nash Equilibrium, and continuously generate new samples by the trained generator, and perform similarity evaluation on the samples in the minority class sample set, screen out new samples, and stop until the screened samples reach a certain number, and the samples are filled into the training set, in this example, the threshold of K-L divergence (KLD) is 0.3, the threshold of Euclidean distance (ED) is 0.25, the threshold of Pearson correlation coefficient (PCC) is 0.85, and the number of samples passing the screening is 360;

[0090] Step 405: perform deep feature extraction on the training set after filling the samples and the test set divided in step KB1, the extracted features include mean, sample variance, root mean square, skewness, kurtosis, peak value, waveform factor, peak factor, pulse factor, power spectrum mean, frequency center of gravity, mean square frequency, and 8-section wavelet packet energy, a total of 20 features, the features and labels extracted in the training set are used to train the Softmax classifier, and the features and labels extracted in the test set are used to test the performance of the Softmax classifier, and finally a Softmax classifier with good performance is obtained.

[0091] In step 6, the tool breakage state recognition system includes a classifier decision system, a multi-layer threshold decision system, and a multi-tooth tool breakage intelligent recognition system,

[0092] The classifier decision system makes decisions on the deep features of the data through the trained classifier to determine the tool breakage state,

[0093] The multi-layer threshold decision system includes dynamic threshold decision and fixed threshold decision, the dynamic threshold decision identifies abnormal data based on Gaussian distribution in mathematical statistics to identify tool breakage, and the fixed threshold decision identifies abnormal data based on fixed values determined by experiments to identify tool breakage,

[0094] The multi-tooth cutter breakage intelligent identification system only works in milling processing, uses multi-dimensional force and vibration sensing data to build a real-time multi-tooth milling cutter state change cloud map, proposes a fuzzy c-means clustering algorithm (FCM), solves the feature matrix of each component, realizes adaptive segmentation of patterns in polar coordinates, and realizes high-speed intelligent identification of tooth breakage and breakage by monitoring the change of characteristic values.

[0095] In the multi-layer threshold decision system, the calculation of the dynamic threshold can be expressed as formula (7) as follows:

[0096]

[0097] In the above formula, N is the number of past monitoring indicators, c is a correction factor, which can be adjusted according to actual conditions, μ is the average of past monitoring data, σ is the standard deviation of past monitoring data, and k is a proportional coefficient.

[0098] The fuzzy c-means clustering algorithm (FCM) is given a data set X = {x1, x2,..., x n}, k is the number of categories, m j (j = 1, 2,..., k) is the center of each cluster, μ j (x i ) is the membership function of the i-th sample corresponding to the j-th class, then the clustering loss function based on the membership function can be written as formula (8) as follows:

[0099]

[0100] Here b refers to the weighting coefficient, also known as the smoothing factor, which controls the sharing degree of the pattern between fuzzy classes.

[0101] Finally, as shown in Figure 6 , in step 6, the multi-method fusion judgment real-time signal breakage state is performed by the following steps:

[0102] Step 601: Use the previously built cutter state in-situ real-time monitoring data acquisition system to perform real-time data acquisition, and then perform steps 602, 603 and 604 simultaneously, which correspond to classifier judgment, multi-layer threshold judgment and multi-tooth cutter breakage intelligent identification judgment, respectively;

[0103] Step 602: when the collected data reaches 2000 points, the data is standardized to the range of [-1, 1], and 20 features including the average value, sample variance, root mean square, skewness, kurtosis, peak value, waveform factor, peak factor, pulse factor, power spectrum mean, frequency center of gravity, mean square frequency, and 8-section wavelet packet energy are extracted, the features are input into the previously trained Softmax classifier for judgment to obtain result A, then the latest 2000 points are taken, and the step is repeated until stopping;

[0104] Step 603: every point collected is compared with a threshold to make a judgment to obtain result B;

[0105] Step 604: when the collected data reaches 2000 points, according to the main shaft speed and the sampling rate of the acquisition card, the collected points are plotted into a polar coordinate graph, and image feature recognition is performed according to an intelligent algorithm to obtain result C;

[0106] Step 605: if result A is damaged, a related interface is called to perform alarm processing to remind an operator to change a tool; if the dynamic threshold decision in result B is damaged, a related interface is called to perform alarm processing to remind the operator to check the situation; if the fixed threshold decision in result B is damaged, a related interface is called to shut down the machine and remind the operator to check the situation; if result C is damaged, a related interface is called to perform alarm processing to remind the operator to check the situation, and the next step is performed according to the specific situation.

[0107] It can be understood that the present application is described through some embodiments, and those skilled in the art know that various changes or equivalent replacements can be made to the features and embodiments without departing from the spirit and scope of the present application. In addition, the features and embodiments can be modified to adapt to specific conditions and materials under the guidance of the present application without departing from the spirit and scope of the present application. Therefore, the present application is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of the present application are within the scope of the present application.

Claims

1. A method for in-situ real-time monitoring of a tool breakage condition, characterized by: The method comprises the following steps: Step 1: build a tool state in-situ real-time monitoring data acquisition system, acquire cutting processing sensing signal sample data and shoot images of tool wear areas after each cutting; Step 2: measure the tool wear area images to obtain tool damage labels, and construct a sample data sequence with the sensing signal sample data; Step 3: perform data segmentation and standardization processing on the cutting processing signals, and input them into an unbalanced sample enhancement system for data enhancement; Step 4: perform deep feature extraction on the data set after data enhancement, input the training set after feature extraction into a classifier for training, and test the test set after feature extraction; Step 5: input the trained classifier in step 4 into a tool damage state recognition system; Step 6: acquire real-time signals through the tool state in-situ real-time monitoring data acquisition system, input the real-time signals into the tool damage state recognition system to determine the tool damage state, In step 1, the tool state in-situ real-time monitoring data acquisition system comprises a force signal acquisition system, a vibration signal acquisition system and an image signal acquisition system; The force signal acquisition system acquires force signals in the cutting process by installing a three-channel force sensor on the tool holder, the force signals comprising force along the axial feed direction of the workpiece, force along the radial direction of the workpiece and force along the tangential direction of the workpiece, The vibration signal acquisition system acquires three-direction vibration signals in the cutting process by installing a three-direction acceleration sensor on the workpiece or the tool holder, The three-direction vibration comprises vibration along the axial feed direction of the workpiece, vibration along the radial direction of the workpiece and vibration along the tangential direction of the workpiece, The image signal acquisition system shoots images of tool wear areas after each cutting through an industrial camera or a microscope, In step 3, the training and data enhancement process of the unbalanced sample enhancement system comprises: Step 311: establish a data set, each sample in the data set comprising the following parameters: force signals in the cutting process, three-direction vibration signals in the cutting process, and corresponding damage labels; divide the data set into a training set and a test set; Step 312: filter out minority class samples in the training set to establish a minority class sample set, and train a generator of a generative adversarial network in the unbalanced sample enhancement system by using the minority class sample set; Step 313: generate new samples by using the trained generator in step 312, filter out suitable samples by performing similarity evaluation on the new samples and samples in the minority class sample set in step 312, and input the suitable samples into the training set in step 311 as expanded samples, The similarity evaluation comprises the following steps: Step 320: set a threshold for K-L divergence, Euclidean distance and Pearson correlation coefficient according to actual conditions; Step 321: take the new samples generated by the generator, randomly select real samples from the minority class sample set, calculate whether the K-L divergence between the generated samples and the real samples is less than the threshold, and perform step 322 if the K-L divergence is less than the threshold; Step 322: calculate whether the Euclidean distance between the generated samples and the real samples in step 321 is less than the threshold, and perform step 323 if the Euclidean distance is less than the threshold; Step 323: calculating whether the Pearson correlation coefficient between the generated sample in step 321 and the real sample is greater than a threshold value, and if so, performing subsequent operations on the sample, In step 4, the data set after data augmentation is subjected to deep feature extraction, comprising: calculating the mean, standard deviation, root mean square, skewness, kurtosis, peak value, waveform factor, peak factor and pulse factor of the sensor signal sample data sequence; performing fast Fourier transform on the sensor signal sample data sequence and calculating the amplitude to obtain a frequency spectrum sequence, and calculating the power spectrum mean, frequency barycenter and mean square frequency of the frequency spectrum sequence; performing 3-layer db4 wavelet packet transform on the sensor signal sample data sequence, obtaining 8 decomposition coefficients after decomposition, reconstructing the decomposed signal to obtain signal components of different frequency bands, and extracting energy features of the decomposed frequency band signals to obtain 8 wavelet energy values in different frequency intervals; arranging these features in columns to obtain a signal feature matrix, In step 4, the classifier model is obtained by the following steps: Step 401: randomly divide the data set into a training set and a test set in a ratio of 8:2 according to each label, obtaining 440 samples in the data set, including 400 normal samples and 40 abnormal samples, and 110 samples in the training set, including 100 normal samples and 10 abnormal samples; Step 402: extract the minority class sample set in the training set, which is generally an abnormal label in damage monitoring, and the majority class sample set generally corresponds to a normal label; Step 403: input the minority class sample set into the generative adversarial network for training; Step 404: stop training when the generative adversarial network reaches Nash equilibrium, and continuously generate new samples with the generator, and perform similarity evaluation with the samples in the minority class sample set, select new samples, and stop until the selected samples reach a certain number, and then fill the samples into the training set; Step 405: perform deep feature extraction on the training set after filling the samples and the test set divided in step 401, use the features extracted in the training set and the labels for training the Softmax classifier, and use the features extracted in the test set and the labels for testing the performance of the Softmax classifier, and finally obtain the Softmax classifier, In step 6, the tool damage state recognition system comprises a classifier decision system, a multi-layer threshold decision system and a multi-tooth tool breakage intelligent recognition system, The classifier decision system makes decisions on the deep features of the data through the trained classifier to determine the tool damage state, The multi-layer threshold decision system includes dynamic threshold decision and fixed threshold decision, the dynamic threshold decision identifies abnormal data based on Gaussian distribution in mathematical statistics to identify tool damage, and the fixed threshold decision identifies abnormal data based on fixed values determined by experiments to identify tool damage, The multi-tooth tool breakage intelligent recognition system realizes high-speed intelligent recognition of tooth damage and breakage by monitoring the change of characteristic values, In step 6, the real-time signal damage state is determined by the following steps: Step 601: real-time data acquisition is performed using the tool state in-situ real-time monitoring data acquisition system, and then steps 602, 603 and 604 are synchronously performed, corresponding to classifier judgment, multi-layer threshold judgment and multi-tooth tool breakage intelligent identification judgment respectively; Step 602: when the collected data reaches 2000 points, the data is standardized and features are extracted, the features are input into the previously trained Softmax classifier for judgment, and result A is obtained, then the latest 2000 points are taken, and the step is repeated until it is stopped; Step 603: each point is compared with the threshold to make a judgment, and result B is obtained; Step 604: when the collected data reaches 2000 points, according to the spindle speed and the sampling rate of the acquisition card, the collected points are plotted into a polar coordinate graph, and the image features are recognized according to the intelligent algorithm to obtain result C; Step 605: if result A is damaged, call the relevant interface to alarm and remind the operator to change the tool; if the dynamic threshold decision in result B is damaged, call the relevant interface to alarm and remind the operator to check the situation, if the fixed threshold decision in result B is damaged, call the relevant interface to stop the machine and remind the operator to check the situation; if result C is damaged, call the relevant interface to alarm and remind the operator to check the situation, and the next step is performed according to the specific situation.

2. A method of in-situ real-time monitoring of a tool breakage condition according to claim 1, characterized in that: In step 2, the tool damage label includes the tool damage state corresponding to the offline measured wear area image data.

3. A method of in-situ real-time monitoring of a tool breakage condition according to claim 1, characterized in that: In step 3, the data segmentation and standardization processing operation includes: The time window of a fixed size is used to segment the sensor signal sample data to remove redundant components to obtain a stable cutting signal for data segmentation; According to the stable cutting signal for data segmentation, the data is segmented according to a certain number of sample points to obtain a plurality of samples; According to the samples after data segmentation, the samples are standardized to obtain the sequence of sensor signal sample data.

Images