A device for monitoring the state of breakage of a milling tool and a method for monitoring the same

Abstract

A kind of milling tool breakage state monitoring device and its monitoring method belong to tool monitoring technical field.Vibration sensor is fixed on workpiece, micro-displacement sensor and multichannel force sensor are fixed on milling cutter spindle, vibration sensor, micro-displacement sensor and multichannel force sensor are connected with data acquisition card signal transmission;Computer program is stored on memory and runs on processor, sensing signal acquisition module is connected with data acquisition card and industrial camera signal transmission, for obtaining milling cutter processing sensing signal sample data and milling cutter relief wear zone image data according to time sequence acquisition.METHOD AS FOLLOWS:collect data and construct sample data sequence;Signal pre-processing;Feature extraction.The present application is suitable for various mechanical processing scenes, in the field of intelligent manufacturing requiring high precision ultra-precision, can effectively reduce the processing error and production accident caused by tool breakage, improve production efficiency and safety.

Description

Technical Field

[0001] This invention relates to a milling tool damage monitoring device and method, belonging to the field of tool monitoring technology. Background Technology

[0002] Intelligent cutting tools can achieve in-situ monitoring and compensation of dynamic data such as machine tool dynamic cutting performance, tool breakage status, and machining accuracy evolution.

[0003] When studying tool breakage, the main method involves removing the tool from the equipment and measuring the wear band on the flank face, then analyzing the tool image to obtain the measurement results. Direct measurement of tool wear offers high accuracy; however, due to limitations in measurement conditions, the direct measurement results may not always meet expectations. For example, foreign objects on the workpiece can affect the measurement results. In actual production, multiple testing methods should be used comprehensively based on the specific conditions.

[0004] During metal cutting, as the width of the wear band on the tool's flank increases, signals such as cutting force, vibration, and micro-displacement also change accordingly. These signals are also correlated. Multi-source sensor information signals can be used to monitor the tool's damage state, which is an indirect measurement method.

[0005] Indirect measurement methods measure multi-source sensor signals generated during tool machining and then establish relevant analysis and judgment models to determine the tool's condition. Therefore, the accuracy of the final tool breakage condition determination largely depends on the type of sensor used and the correlation between the signal characteristics and the tool breakage condition.

[0006] In most cases, the signals acquired from sensors contain a certain amount of noise. The raw signals acquired from sensors are a superposition of the tool signals and environmental machining noise that we need. Such signals have poor accuracy in characterizing the spatiotemporal state and damage of the tool. Therefore, it is necessary to reduce noise and filter the raw signals.

[0007] Time-domain analysis of signals reflects the changes in various variables during milling as machining time increases, providing a direct view of the instantaneous and continuous fluctuations in the signal. However, relying solely on the amplitude of a single channel signal to monitor tool breakage is insufficient to accurately predict the tool's true condition. In practice, dimensional and dimensionless characteristic parameters are often extracted from various dimensions of the signal through time-domain statistical analysis. Dimensional characteristic parameters directly reflect the changes in the flank wear band during the cutting process. Meanwhile, dimensionless parameters, obtained by dividing two dimensional parameters, eliminate the influence of factors such as signal amplitude, thus reflecting the tool breakage state.

[0008] If these signal feature parameters are used directly for tool condition characterization, the mapping mechanism between tool breakage state and signal feature parameters is unclear, and it is unknown which feature is a better method for evaluating tool breakage state. This will lead to errors in the tool breakage state characterization results and increase the computational load of the system.

[0009] Based on the inconvenience of the direct method for measuring the width of the wear band on the tool's flank and the low accuracy of the tool breakage monitoring method using a single-channel signal, this invention proposes a milling tool breakage condition monitoring device and its monitoring method. Summary of the Invention

[0010] To address the problems existing in the background art, the present invention provides a milling tool breakage condition monitoring device and monitoring method thereof.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: a milling cutter damage monitoring device, comprising a vibration sensor, a micro-displacement sensor, a multi-channel force sensor, a data acquisition card, a computer terminal, and an industrial camera; the vibration sensor is fixed on the workpiece, which is placed on a CNC machine tool; the micro-displacement sensor and the multi-channel force sensor are both fixed on the spindle of the milling cutter, and the vibration sensor, micro-displacement sensor, and multi-channel force sensor are all connected to the data acquisition card for signal transmission; the computer terminal includes a memory, a processor, a computer program, and a sensor signal acquisition module, the computer program is stored in the memory and runs on the processor, and the sensor signal acquisition module is connected to the data acquisition card and the industrial camera for signal transmission, used to acquire time-series milling cutter machining sensor signal sample data and milling cutter back face wear zone image data.

[0012] The present invention provides a monitoring method for a milling cutter breakage condition monitoring device, the method comprising the following steps:

[0013] S1: Collect data and construct sample data sequences;

[0014] S2: Signal preprocessing;

[0015] S3: Feature extraction.

[0016] Compared with the prior art, the beneficial effects of the present invention are:

[0017] This invention extracts the time-domain, frequency-domain, and time-frequency-domain features of multi-dimensional, multi-channel signals. It evaluates the correlation between signal features and tool breakage through the maximum information coefficient, obtaining a good feature sequence. Then, it uses principal component analysis to reduce the dimensionality of tool signal features and establish health indicators, obtaining the most important evaluation indicators for tool breakage status. It is applicable to various machining scenarios. In the field of intelligent manufacturing that requires high precision and ultra-precision, it can effectively reduce machining errors and production accidents caused by tool breakage, and improve production efficiency and safety. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the monitoring device of the present invention;

[0019] Figure 2 This is a flowchart of the monitoring method of the present invention;

[0020] Figure 3 This is a time-series diagram showing the relationship between some features and the width of the wear band on the back face;

[0021] Figure 4 It is a matrix of correlation coefficients for the feature columns of the cutting tool.

[0022] Figure 5 This is a schematic diagram of the health index curve of the cutting tool. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] A milling cutter damage monitoring device includes a vibration sensor 2, a micro-displacement sensor 4, a multi-channel force sensor 5, a data acquisition card 8, a computer terminal 9, and an industrial camera 10. The vibration sensor 2 is fixed on a workpiece 1, which is placed on a CNC machine tool 7. The micro-displacement sensor 4 and the multi-channel force sensor 5 are both fixed on the spindle 6 of the milling cutter 3, and are all connected to the data acquisition card 8 for signal transmission. The computer terminal 9 includes a memory, a processor, a computer program, and a sensor signal acquisition module. The computer program is stored in the memory and runs on the processor. The sensor signal acquisition module is connected to the data acquisition card 8 and the industrial camera 10 for signal transmission, and is used to acquire time-series milling cutter machining sensor signal sample data and milling cutter back face wear zone image data.

[0025] Vibration sensor 2 uses a Kistler 8762A5 accelerometer, which can measure vibration signals in the X, Y, and Z axes.

[0026] The machining method involves the milling cutter 3 cutting along multiple transverse paths on the upper surface of the workpiece 1.

[0027] The micro-displacement sensor 4 adopts a domestically produced Sanying Precision Control nano-displacement capacitive sensor, which can measure micro-displacements within 200 micrometers.

[0028] The multi-channel force sensor 5 uses a Swiss Kistler 5347A force measuring tool holder, which can measure force and torque signals in the X, Y, and Z axes.

[0029] The CNC machine tool 7 uses a three-axis milling machine VMC855. The milling cutter 3 used is a single-tooth end mill with an APMT1135 insert made of polycrystalline diamond. The workpiece 1 is made of aluminum-based silicon carbide.

[0030] Data acquisition card 8 uses the 5324 and 69529 series acquisition boards from domestic Jianyi Technology to monitor and record force signals, vibration signals and displacement signals in real time. The signal acquisition frequency is 2500Hz and there are a total of seven channels.

[0031] The industrial camera 10 is used to photograph the shape of the milling cutter in-machine mode.

[0032] The monitoring method of the present invention is executed when the processor runs a computer program.

[0033] The width of the wear band on the flank of the milling cutter was measured using a Visionox MV-HS2000GM industrial camera in conjunction with a dual telecentric microscope. Subsequent processing and analysis of the milling cutter sample data were performed on a computing platform with an NVIDIA RTX4060Ti 16GB GPU.

[0034] The present invention provides a monitoring method for a milling cutter breakage condition monitoring device, the method comprising the following steps:

[0035] S1: Collect data and construct sample data sequences;

[0036] S101: Install a milling cutter damage monitoring device;

[0037] S102: Acquire milling cutter machining sensor signal sample data: acquire raw signal sample data during the milling cutter cutting process. The raw signal sample data includes the sensor data of each sensor and each channel corresponding to each cut of the milling cutter.

[0038] S103: Capture and measure the image of the wear zone area on the back face of the milling cutter after each cut: acquire image data of the wear zone area on the back face of the milling cutter during the cutting process. The image data of the wear zone area on the back face of the milling cutter includes the image data of the milling cutter corresponding to the sensor data at each cut.

[0039] S104: Obtain the label of the width VB of the flank wear band. The label of the width VB of the flank wear band includes the width value of the flank wear band corresponding to the offline measured image data of the flank wear band area.

[0040] S105: Construct a sample data sequence using tag and sensor signal sample data.

[0041] S2: Signal preprocessing;

[0042] S201: Identify the entry point and retraction point in the sensor signal sample data during the milling cutter cutting process, and remove the sensor signals before the milling cutter enters and after the milling cutter retracts during the milling process to obtain sensor signal sample data for noise reduction filtering;

[0043] S20101: The original signal sample data is analyzed using the sliding window method to obtain the variance value of the time window corresponding to each data point;

[0044] S20102: Thresholding is applied to the variance values, and the corresponding time window variance is:

[0045]

[0046] In formula (1):

[0047] i represents time i of the time window of the signal in one of the channels;

[0048] X i This represents the signal value at time i;

[0049] X represents i The sample mean within the given time window;

[0050] X k This represents the signal value at time k;

[0051] k represents time k of the time window of the signal from one of the channels.

[0052] l represents the radius of the time window. The points used to calculate the time window are 2l+1 points centered at time i with a radius of l on both sides. Therefore, the length of each time window is 2l+1.

[0053] If a signal boundary is encountered, the signal is padded with zeros before the calculation is performed;

[0054] Threshold segmentation prediction uses the quartile method:

[0055] The time window variance of each data point is sorted from smallest to largest. The value Q1, which is the 25th percentile of the total time window variance signal, is found. This indicates that 25% of the signal in the entire time window variance signal is less than or equal to Q1. Then, the time window variance is traversed sequentially from front to back and from back to front. The index corresponding to the first value exceeding Q1 is found. The two indices obtained are recorded. The signal between these two indices is the signal of the milling cutter.

[0056] S20103: Removes signals during milling cutter feed and retraction.

[0057] The milling cutter does not make sufficient contact with the workpiece during the feed and retraction phases, so this part of the signal needs to be removed. Based on experience, 2% of the length of the milling cutter machining signal can be removed before and after the cutter to obtain the machining signal when the milling cutter makes sufficient contact.

[0058] S202: Based on the sensor signal sample data used for noise reduction filtering, the sensor signal sample data obtained in S201 is subjected to noise reduction filtering operation using the wavelet soft threshold denoising method to obtain the noise-reduced sensor signal sample data sequence.

[0059] S20201: Perform noise reduction operation;

[0060] S2020101: The signal is decomposed using a single-layer db18 wavelet basis function, and the decomposition mode adopts a symmetric mode;

[0061] S2020102: Determine the soft thresholding function for wavelet basis function denoising. Compare the threshold of the soft thresholding function with the decomposition coefficients of the wavelet basis function obtained after decomposition, and remove noise smaller than the threshold of the soft thresholding function. The soft thresholding function is as follows:

[0062]

[0063] In formula (2):

[0064] W jk Represents the wavelet basis function decomposition coefficients;

[0065] sgn represents the sign function;

[0066] λ represents the threshold of the soft thresholding function;

[0067] The threshold of the soft threshold function is determined using the Minimaxi method.

[0068] S2020103: The wavelet basis function decomposition coefficients that have undergone thresholding by the soft thresholding function are reconstructed using the db18 wavelet basis function to restore the original signal length, thus completing the noise reduction operation.

[0069] S20202: Perform filtering operation;

[0070] The sensor signal sample data were sequentially subjected to Hampel filtering and Butterworth low-pass filtering.

[0071] S2020201: After initially identifying outliers in the entire signal, select time windows of corresponding lengths for all outliers and calculate the median and standard deviation of the data in the time window corresponding to each outlier.

[0072] S2020202: If the difference between an outlier and the median of the window exceeds three times the standard deviation of the window data, the outlier is confirmed as an outlier. The outlier is then replaced by the median of the time window corresponding to the outlier to eliminate the outlier.

[0073] S2020203: Determine the passband cutoff frequency and stopband cutoff frequency for the Butterworth low-pass filter, the maximum passband attenuation (typically 3dB) and the minimum stopband attenuation (typically 20dB).

[0074] S2020204: Determine the transfer function coefficients, order, and cutoff frequency used for filtering using the above parameters;

[0075] S2020205: Use the transfer function to filter and obtain the noise-reduced and filtered sensor signal sample data sequence.

[0076] Wavelet analysis always involves approximate components and detail components. Approximate components reflect the low-frequency information of the signal, while detail components reflect the high-frequency information. For noisy signals, the main energy of the noise components is concentrated in the detail components of the wavelet decomposition. The main goal of wavelet denoising is to filter out the noise in the high-frequency components while preserving the true value of the original signal, thus achieving the purpose of noise reduction. After denoising, the signal is filtered to remove high-frequency information, retaining only the low-frequency signals related to tool machining, thereby improving the representation accuracy of the tool machining signal.

[0077] S3: Feature extraction.

[0078] S301: Extract features from the milling cutter machining signal;

[0079] S30101: Calculate the mean, sample variance, peak-to-peak value, minimum value, maximum value, root mean square, sample skewness, sample kurtosis, waveform factor, peak factor, impulse factor, and margin factor of the sensor signal sample data sequence respectively. The calculation process is shown in Table 1.

[0080] S30102: Perform Fast Fourier Transform (FFT) on the sensor signal sample data sequence and calculate the amplitude to obtain the frequency spectrum sequence. Calculate the frequency centroid, root mean square of the spectrum, variance of the spectrum, skewness of the spectrum, kurtosis of the spectrum, and energy of the frequency spectrum sequence. The calculation process is shown in Table 1.

[0081] S30103: Perform a three-level db4 wavelet packet transform on the sensor signal sample data sequence, decompose it to obtain eight decomposition coefficients, reconstruct the decomposed signal to obtain signal components of different frequency bands, extract the energy features of each frequency band signal after decomposition to obtain wavelet energy values ​​of eight different frequency intervals.

[0082] S30104: Arrange the above features in columns to obtain the signal feature matrix;

[0083] S30105: Standardize the signal feature matrix and solve its correlation coefficient matrix. Specifically, use the maximum information coefficient (MIC) to evaluate the signal features and the wear band width of the milling cutter back face through the correlation coefficient, screen the feature matrix columns with high correlation, and remove indicators that reflect duplicate information to obtain a good feature data sequence.

[0084] Compared to the Pearson correlation coefficient, the maximum information coefficient is better able to capture the nonlinear correlation of a series. Compared to the Spearman correlation coefficient, it can capture not only monotonic relationships but also the rate of change of the series (i.e., the rate of change of the derivative).

[0085] S30106: Solve the correlation coefficient matrix for all feature matrix columns (because each column has a correlation coefficient obtained through the MIC method, arrange the correlation coefficients into a matrix, solve the MIC correlation coefficient, and arrange the multiple MIC correlation coefficients into a correlation coefficient matrix), select N features with the largest correlation coefficients as good feature sequences, then obtain the feature matrix of the good feature sequences, solve the eigenvalues ​​and eigenvectors of the good feature matrix, and obtain the standardized feature vector matrix. Each column of the feature vector matrix is ​​the principal component column. Calculate the cumulative score of the principal component column to characterize the width of the tool face wear band, and compare it with the milling cutter failure measurement for use in milling cutter condition monitoring to reflect the degree of damage.

[0086] Table 1

[0087]

[0088] In Table 1:

[0089] N represents the length of the signal;

[0090] f i This represents the i-th frequency component in the spectrum obtained after the Fourier transform;

[0091] P(f i ) represents the amplitude corresponding to this frequency component fi;

[0092] d i,k It is the reconstructed signal value of wavelet frequency band k at time i, where the value of k is 1-8;

[0093] S302: Filter out good feature data sequences for characterizing the breakage state of milling cutters;

[0094] S30201: Perform correlation analysis on milling cutter machining signals;

[0095] S3020101: Divide one of the feature sequences of the signal feature matrix into segment a, and divide the wear band width value of the milling cutter back face into segment b to obtain an a*b resolution network G. The values ​​of a and b are generally determined empirically.

[0096] Then, the signal value x and the width value y are sample points (x, y) in network G. The probability that the sample point (x, y) is located in the i-th column of network G is P(x, y). i The probability that a sample point (x, y) is located in the j-th row of network G is P(y). j The probability that the sample point (x, y) is located in the i-th column and j-th row of network G is P(x, y). i y j );

[0097] Then the mutual information I(x, y) of sample points (x, y) in network G i y j The formulas for calculating the mutual information entropy I(a;b) of the entire sample sequence are as follows:

[0098]

[0099]

[0100] S3020102: Adjust the segmentation points of the feature sequence and the back face wear band width value sequence to obtain the maximum value of mutual information entropy. In order to eliminate the influence of different resolutions on information entropy, calculate the maximum value of mutual information entropy and normalize it.

[0101] The calculation process for the maximum value of the mutual information entropy is as follows:

[0102] Substituting equation (3) into equation (4), we get:

[0103]

[0104] Similarly, we can obtain: I(a; b) ≤ log₂a. The inequality holds as follows: for all 1 ≤ i ≤ a, 1 ≤ j ≤ b, P(yi|xi) = 1 and Therefore, the maximum value of the mutual information entropy I(a; b) is:

[0105]

[0106] Therefore, for each resolution, the normalized mutual information entropy I can be obtained. * (a; b):

[0107] I * (a;b)=I(a;b) / I max (a;b) (7)

[0108] S3020103: Gradually increase the network resolution, i.e., gradually increase the values ​​of a and b from 2, and calculate the maximum mutual information entropy I at different resolutions. * max :

[0109]

[0110] Considering the entire sample sequence has a length of n, the maximum number of segments ab is... max for:

[0111] ab max =n 0.3 (9)

[0112] Output maximum mutual information entropy I * max The maximum information coefficient of the aforementioned feature columns is used to evaluate the correlation of signal features. The maximum information coefficient of each feature column in the feature matrix is ​​calculated, and the top N features with the largest maximum information coefficients are selected as the good feature data sequence.

[0113] S30202: Perform principal component analysis on the milling cutter machining signal;

[0114] S3020201: Subtract the average value of all features of each sample signal point from each feature in the good feature data sequence to achieve zero mean, and obtain a zero mean feature matrix;

[0115] S3020202: Calculate the covariance between each feature of the zero-mean feature matrix to form a covariance matrix;

[0116] S3020203: Solve for the covariance matrix by finding its eigenvalues, corresponding eigenvectors, and principal component contribution rates. The principal component contribution rate is the proportion of the eigenvalues ​​to the sum of the absolute values ​​of all eigenvalues, sorted according to the size of the absolute values ​​of the eigenvalues.

[0117] S3020204: Starting with the largest absolute eigenvalue, select M eigenvalues ​​whose absolute eigenvalues ​​exceed a certain threshold and their corresponding eigenvectors as principal component vectors;

[0118] S3020205: Multiply the new M principal component columns by the corresponding principal component contribution rates and then add them together to obtain the cumulative principal component scores.

[0119]

[0120] In formula (10):

[0121] score i The score representing time i;

[0122] feature_PCA j This represents the value of principal component j;

[0123] rate j This represents the contribution rate of principal component j to the whole.

[0124] The health indicators are calculated as follows:

[0125]

[0126] In equation (11):

[0127] life i Represents the percentage of remaining lifetime at time i;

[0128] len represents the total number of tool passes;

[0129] N signal features achieve feature dimensionality reduction. The cumulative principal component score can effectively reflect the damage state of the milling cutter. According to the definition of milling cutter failure and damage in the national standard GB / T16460-2016, the cumulative principal component score obtained can be divided by the cumulative principal component score corresponding to the specified maximum width (VB=300um) as the health index of the milling cutter, which reflects the degree of damage. That is, when the health index is less than 10%, we consider the milling cutter to be damaged.

[0130] S30203: Filter good feature data sequences for characterizing the breakage state of milling cutters;

[0131] S303: Outputs good feature data sequences and milling cutter health indicators for subsequent analysis and training.

[0132] Extracting features from the original signal enables data dimensionality reduction, decreasing the computational burden of damage identification. Correlation analysis and principal component analysis can be performed on the signal features to clarify the mapping mechanism between tool damage state and signal feature parameters. Correlation analysis removes irrelevant features and features with minimal impact on the flank wear band width, ensuring that the selected features have a certain correlation with the tool flank wear band width. Principal component analysis removes duplicate and unimportant features, ensuring that the selected features are important evaluation indicators of tool damage state. Finally, by mapping the signal to the tool flank wear band width, the remaining tool life can be effectively characterized.

[0133] Example 1:

[0134] S1: Conduct milling cutter cutting experiments. During the machining process, sensor signals are collected simultaneously. After each pass, the milling cutter is photographed. The image data of the wear band area on the back face is measured using a back face wear band width measuring device to obtain a label for the width of the wear band on the back face of the milling cutter.

[0135] S2: Preprocess the sensor signal sample data.

[0136] S201: Set the time window length 2l+1=1001, identify the tool entry and exit points in the sensor signal sample data, and slide the time window to obtain sensor signal sample data for noise reduction filtering.

[0137] S202: Perform wavelet soft thresholding denoising with a threshold of 3dB to reconstruct the signal; perform Hampel filtering with a time window length of 0.4s; perform Butterworth filtering with a maximum passband attenuation of 3dB and a minimum stopband attenuation of 20dB, calculate the transfer function, filter it, and obtain the sensor signal sample data sequence.

[0138] S301: Extract features from the preprocessed sensor signal sample data.

[0139] S30101: Calculate the mean, sample variance, peak-to-peak value, minimum value, maximum value, root mean square, sample skewness, sample kurtosis, waveform factor, peak value, impulse factor, and margin factor of the sensor signal sample data sequence.

[0140] S30102: Perform Fast Fourier Transform (FFT) on the sensor signal sample data sequence and calculate the amplitude to obtain the frequency spectrum sequence. Calculate the frequency centroid, root mean square of the spectrum, variance of the spectrum, skewness of the spectrum, kurtosis of the spectrum, and energy of the frequency spectrum sequence.

[0141] S30103: Perform a three-level db4 wavelet packet transform on the sensor signal sample data sequence, decompose it to obtain eight decomposition coefficients, reconstruct the decomposed signal to obtain signal components of different frequency bands, extract the energy features of each frequency band signal after decomposition to obtain wavelet energy values ​​of eight different frequency intervals.

[0142] S30104: Calculate the signals of all channels of each sensor, arrange these features in columns to obtain a signal feature matrix. This feature matrix has a total of 26*7=182 columns.

[0143] S302: Perform correlation analysis and principal component analysis on the milling cutter machining signal to obtain highly correlated features and health indicators, reflecting the degree of damage.

[0144] S30201: The method of this invention is used to calculate the mutual information number and maximum information coefficient (MIC) of each feature column in the wear band width sequence of the back face. Based on experience, N=24 features with the highest correlation are selected from 182 feature columns for principal component analysis.

[0145] S30202: Calculate the principal components of the new 24-column feature matrix according to the method of the present invention. The threshold for the absolute value of the eigenvalues ​​is 3. Obtain the cumulative score of the principal components and the health index of the milling cutter, which reflects the degree of damage.

[0146] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0147] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A monitoring method for a milling cutter breakage monitoring device, the method being based on a monitoring device comprising a vibration sensor (2), a micro-displacement sensor (4), a multi-channel force sensor (5), a data acquisition card (8), a computer terminal (9), and an industrial camera (10); the vibration sensor (2) is fixed on a workpiece (1), the workpiece (1) being placed on a CNC machine tool (7); the micro-displacement sensor (4) and the multi-channel force sensor (5) are both fixed on the spindle (6) of the milling cutter (3), and the vibration sensor (2), the micro-displacement sensor (4), and the multi-channel force sensor (5) are all connected to the data acquisition card (8) for signal transmission; the computer terminal (9) comprises a memory, a processor, a computer program, and a sensor signal acquisition module, the computer program being stored in the memory and running on the processor, and the sensor signal acquisition module being connected to the data acquisition card (8) and the industrial camera (10) for signal transmission, used to acquire milling cutter machining sensor signal sample data and milling cutter back face wear zone image data in a time sequence; characterized in that: The monitoring method includes the following steps: S1: Collect data and construct sample data sequences; S2: Signal preprocessing; S3: Feature extraction; S3 includes the following steps: S301: Extract features from the milling cutter machining signal; S301 includes the following steps: S30101: Calculate the mean, sample variance, peak-to-peak value, minimum value, maximum value, root mean square, sample skewness, sample kurtosis, waveform factor, peak factor, impulse factor, and margin factor of the sensor signal sample data sequence respectively. S30102: Perform a fast Fourier transform on the sensor signal sample data sequence and calculate the amplitude to obtain a frequency spectrum sequence. Calculate the frequency centroid, root mean square of the spectrum, variance of the spectrum, skewness of the spectrum, kurtosis of the spectrum, and energy of the frequency spectrum sequence. S30103: Perform a three-level db4 wavelet packet transform on the sensor signal sample data sequence to obtain eight decomposition coefficients. Reconstruct the decomposed signal to obtain signal components in different frequency bands, and extract the energy features of each frequency band signal. Wavelet energy values ​​were obtained in eight different frequency ranges; S30104: Arrange the above features in columns to obtain the signal feature matrix; S30105: Standardize the signal feature matrix and solve for its correlation coefficient matrix; S30106: Solve the correlation coefficient matrix for all feature matrix columns, solve for the eigenvalues ​​and eigenvectors, and obtain the standardized feature vector matrix. Each column of the feature vector matrix is ​​the principal component column. Calculate the cumulative score of the principal component column to characterize the width of the tool face wear band, and compare it with the milling cutter failure measurement for use in milling cutter condition monitoring to reflect the degree of damage. S302: Filter out good feature data sequences for characterizing the breakage state of milling cutters; S30201: Perform correlation analysis on milling cutter machining signals; S30201 includes the following steps: S3020101: Divide one of the feature sequences of the signal feature matrix into segment 'a', and divide the wear band width value of the milling cutter flank face into segment 'b', to obtain... The resolving network G; Then, the probability that the signal value x and the width value y are sample points (x, y) in network G, and that the sample point (x, y) is located in the i-th column of network G, is: The probability that the sample point (x, y) is located in the j-th row of network G is The probability that the sample point (x, y) is located in the i-th column and j-th row of network G is... ; Then the mutual information of sample points (x, y) in network G And the mutual information entropy of the entire sample sequence The calculation formula is as follows: （3） （4） S3020102: Adjust the segmentation points of the feature sequence and the back face wear band width value sequence to obtain the maximum value of mutual information entropy, calculate the maximum value of mutual information entropy, and normalize it; The calculation process for the maximum value of the mutual information entropy described in S3020102 is as follows: Substituting equation (3) into equation (4), we get: （5） Similarly, we can conclude that: , The inequality holds if: for all They all and Therefore, mutual information entropy The maximum value is: （6） Therefore, for each resolution, the normalized mutual information entropy can be obtained. : （7） S3020103: Gradually increase the network resolution and calculate the maximum mutual information entropy at different resolutions. : （8） Considering the entire sample sequence has a length of n, what is the maximum number of segments? for: （9） Output the maximum mutual information entropy The maximum information coefficient of the above feature column is used to evaluate the correlation of signal features; S3020104: Calculate the maximum information coefficient of each feature column in the feature matrix, and select the top N features with the largest maximum information coefficients as good feature data sequences; S30202: Perform principal component analysis on the milling cutter machining signal; S30202 includes the following steps: S3020201: Subtract the average value of all features of each sample signal point from each feature in the good feature data sequence to achieve zero mean, and obtain a zero mean feature matrix; S3020202: Calculate the covariance between each feature of the zero-mean feature matrix to form a covariance matrix; S3020203: Solve for the covariance matrix by finding its eigenvalues, corresponding eigenvectors, and principal component contribution rates. The principal component contribution rate is the proportion of the eigenvalues ​​to the sum of the absolute values ​​of all eigenvalues, sorted according to the size of the absolute values ​​of the eigenvalues. S3020204: Starting with the largest absolute eigenvalue, select M eigenvalues ​​whose absolute eigenvalues ​​exceed a certain threshold and their corresponding eigenvectors as principal component vectors; S3020205: Multiply the new M principal component columns by their corresponding principal component contribution rates and then sum them to obtain the cumulative principal component scores. （10） In formula (10): The score representing time i; This represents the value of principal component j; This represents the contribution rate of principal component j to the whole. The health indicators are calculated as follows: （11） In equation (11): Represents the percentage of remaining lifetime at time i; Represents the total number of cuts; S30203: Filter good feature data sequences for characterizing the breakage state of milling cutters; S303: Outputs good feature data sequences and milling cutter health indicators for subsequent analysis and training.

2. The method according to claim 1, characterized in that: S2 includes the following steps: S201: Identify the entry point and retraction point in the sensor signal sample data during the milling cutter cutting process, and remove the sensor signals before the milling cutter enters and after the milling cutter retracts during the milling process to obtain sensor signal sample data for noise reduction filtering; S202: After performing noise reduction filtering on the sensor signal sample data obtained in S201, the noise-reduced sensor signal sample data sequence is obtained.

3. The method according to claim 2, characterized in that: S201 includes the following steps: S20101: Analyze the original signal sample data to obtain the variance value of the time window corresponding to each data point; S20102: Thresholding is applied to the variance values, and the corresponding time window variance is: （1） In formula (1): This represents the time window of the signal in one of the channels. time; express The signal value corresponding to the given time; express The sample mean within the given time window; express The signal value corresponding to the given time; This represents the time window of the signal in one of the channels. time, Indicates the radius of the time window; If a signal boundary is encountered, the signal is padded with zeros before the calculation is performed; S20103: Removes signals during milling cutter feed and retraction.

4. The method according to claim 3, characterized in that: S202 includes the following steps: S20201: Perform noise reduction operation; S2020101: Decompose the signal using wavelet basis functions; S2020102: Determine the soft thresholding function for wavelet basis function denoising. Compare the threshold of the soft thresholding function with the decomposition coefficients of the wavelet basis function obtained after decomposition, and remove noise smaller than the threshold of the soft thresholding function. The soft thresholding function is as follows: （2） In formula (2): Represents the wavelet basis function decomposition coefficients; Represents a symbolic function; The threshold value represents the threshold of the soft thresholding function; S2020103: The wavelet basis function decomposition coefficients that have undergone thresholding by the soft thresholding function are reconstructed using wavelet basis functions to restore the original signal length, thus completing the noise reduction operation. S20202: Perform filtering operation; S2020201: After initially identifying outliers in the entire signal, select time windows of corresponding lengths for all outliers and calculate the median and standard deviation of the data in the time window corresponding to each outlier. S2020202: If the difference between an outlier and the median of the window exceeds three times the standard deviation of the window data, the outlier is confirmed as an outlier. The outlier is then replaced by the median of the time window corresponding to the outlier to eliminate the outlier. S2020203: Determine the passband cutoff frequency and stopband cutoff frequency, maximum passband attenuation value and minimum stopband attenuation value for the low-pass filter. S2020204: Determine the transfer function coefficients, order, and cutoff frequency used for filtering; S2020205: Use the transfer function to filter and obtain the noise-reduced and filtered sensor signal sample data sequence.

Images