A milling chatter online monitoring method combining energy ratio and amplitude standard deviation

Abstract

The application discloses a milling chatter online monitoring method combining energy ratio and amplitude standard deviation, first, three-direction vibration acceleration sensors are used to obtain three-direction vibration signals of a spindle in a milling process; second, a chatter peak searching method is used to find out corresponding amplitudes of periodic peaks and chatter peaks in a signal spectrum vector, and then the amplitudes are substituted into a calculation formula to obtain an energy ratio; then, amplitude standard deviation is calculated, and the energy ratio and the amplitude standard deviation extracted from each direction of the three-direction vibration signals are fused in a feature layer to obtain a six-dimensional monitoring index set; finally, the index set is input into a trained monitoring model to output a milling processing state monitoring result of a sampling time period. Compared with a traditional chatter monitoring method, the method can simultaneously meet three requirements of early monitoring, insensitivity of the monitoring index to processing parameter changes and real-time monitoring, and can accurately identify amplitudes and frequency sizes of a main chatter frequency.

Description

Technical Field

[0001] This invention belongs to the field of monitoring technology of machining conditions, specifically relating to the field of online monitoring of chatter in high-speed milling, and a method for online monitoring of milling chatter that integrates energy ratio and amplitude standard deviation. Background Technology

[0002] Milling is the most common machining method, boasting advantages such as high efficiency, high precision, and low cost, and is widely used in high-end manufacturing fields such as aerospace. Milling chatter is a strong self-excited vibration generated during the milling process, severely affecting machining efficiency and quality. Therefore, online monitoring technology for milling chatter has been a hot topic in high-end manufacturing for many years. Because chatter signals are complex signals that are nonlinear, non-stationary, and multi-component, it is difficult to directly and accurately identify chatter in real time based on the time-domain waveform or spectrum of the sensor signal. For chatter signals with complex components, signal processing techniques are generally required to separate or identify the chatter components, and then construct appropriate monitoring indicators to reflect the changes in the chatter components in the signal, thereby achieving online monitoring of the milling process status.

[0003] Numerous scholars both domestically and internationally have studied the problem of chatter monitoring in milling based on the energy proportion of chatter components in the signal. The acquired chatter signals typically contain periodic components related to the spindle speed and its harmonics, as well as chatter components related to the low-order natural frequencies of the milling process system. To calculate the energy proportion of the chatter components, it is first necessary to separate the periodic and chatter components in the signal. Current research employs two approaches: one is to directly decompose the signal using various mode decomposition methods and select the component most closely related to chatter as the chatter component; the other approach is to indirectly filter out the periodic components in the signal using various filtering methods, and the remaining components in the signal can be considered the chatter components.

[0004] Caliskan et al. from Canada used a Kalman filter to separate stable periodic signals from the raw vibration acceleration data of the spindle and workpiece. They then used a nonlinear energy operator (NEO) and an energy separation algorithm (ESA) to extract the energy ratio (ER) in the time-domain signal to distinguish between chatter and stability. Rahimi et al. from Canada also used a Kalman filter to separate periodic components from spindle vibration and sound pressure signals. They employed a hybrid model combining a physical model based on energy ratio and a neural network model based on short-time Fourier time-spectrum analysis to detect five states: air cutting, feed, stability, chatter, and retraction. However, all of these methods for extracting the energy ratio are limited by the Kalman filter's inability to adaptively filter. This results in the periodic component filtering effect being affected by changes in machining parameters, requiring manual adjustment of the filtering parameters according to the machining process parameters, and making online monitoring under different operating conditions impossible.

[0005] Based on existing literature searches, the commonly used milling chatter monitoring methods generally have the following problems: 1) they cannot achieve early monitoring; 2) the proposed monitoring indicators are sensitive to changes in machining process parameters; and 3) they are difficult to meet the real-time requirements of online monitoring.

[0006] In summary, there is an urgent need for an online monitoring method for milling chatter that can simultaneously meet the three requirements of early monitoring, insensitivity of the proposed monitoring indicators to changes in process parameters, and real-time monitoring, so as to be used for online monitoring in practical engineering. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides an online monitoring method for milling chatter that integrates energy ratio and amplitude standard deviation. The original vibration acceleration signal is converted from the time domain to the frequency domain using a Fast Fourier Transform (FFT). This chatter peak search method is then used to search for all chatter peaks and periodic peaks in the frequency domain. The energy magnitudes of these chatter peaks and the periodic peaks in the spectrum are calculated. The energy ratio is obtained based on the energy magnitudes of the chatter peaks and periodic peaks. By combining this with the amplitude standard deviation, an auxiliary monitoring indicator, online monitoring of milling chatter can be achieved.

[0008] To achieve the above objectives, the technical solution adopted by this invention is: an online monitoring method for milling chatter based on both energy ratio and amplitude standard deviation, comprising the following steps:

[0009] Step 1: Collect real-time vibration signals of the spindle during the milling process using a triaxial vibration acceleration sensor installed at the end of the spindle box;

[0010] Step 2: Obtain the spectrum of the real-time vibration signal of the main shaft through fast Fourier transform, convert the signal from the time domain to the frequency domain, and use the flutter peak search method to search for all flutter peaks and periodic peaks in the frequency domain. Calculate the energy ratio of the flutter peaks and periodic peaks based on their energy magnitudes to obtain the energy ratio values ​​of the flutter peaks and periodic peaks during the sampling period.

[0011] Step 3: Calculate the standard deviation of the amplitude over the sampling period using the time-domain amplitude of the vibration signal.

[0012] Step 4: Perform feature layer fusion on the energy ratio and amplitude standard deviation values ​​of the three-dimensional vibration signals to obtain a set of 6-dimensional monitoring indicators;

[0013] Step 5: Finally, input the fused 6-dimensional monitoring index set into the trained three-class monitoring model and output the milling processing status for the sampling time period.

[0014] Step 2, the flutter peak search method, includes five parts performed sequentially: median filtering, mean filtering, standard deviation filtering, one-dimensional peak finding, threshold exclusion, and frequency exclusion.

[0015] Before performing median filtering in step 2, the spectrum vector Y is first padded with (n) before and after it. Filter –1) / 2 zero values ​​to ensure that the median filtering operation can be applied before and after the spectral vector Y (n Filter –1) / 2 frequencies.

[0016] In step 2, when searching for chatter peaks, consider frequencies greater than the over-tooth frequency ω. t Furthermore, the amplitude of the chatter peak is greater than the threshold at the corresponding frequency, and if multiple chatter peaks are found between adjacent harmonics of the spindle frequency, only the chatter peak with the largest amplitude is retained.

[0017] In the frequency domain, a one-dimensional peak-finding algorithm is used to obtain all peaks in the signal spectrum, with frequencies greater than the over-tooth frequency ω. t Furthermore, peak values ​​with amplitudes greater than the corresponding frequency threshold are retained, and the nearest principal shaft rotation frequencies ω to the left and right of each retained peak value are calculated. s The interval value of its harmonics, if the interval value is less than the set threshold, the interval value ranges from [0, ω]. s If the peak value is 0, then it is considered to be a periodic peak.

[0018] By using an open CNC system, the spindle speed Ω and the number of tool teeth N are collected in real time. t Perform spindle frequency (ω) s =Ω / 60) and over-tooth frequency (ω) t =N t ω s ) calculation.

[0019] The energy ratio calculation formula in step 1 is defined as follows:

[0020]

[0021] In the formula, N represents the number of flutter peaks found; ω CH,i Represents the i-th flutter frequency; K represents the number of excluded frequencies; ω REJ,k ER represents the k-th exclusion frequency, with the numerator being the energy of the flutter component; the first term in the denominator is the energy of the amplitude at the exclusion frequency that replaces the energy of the periodic component. If no acceptable spectral peak is found in the FFT results, then ER = 0; if no exclusion spectral peak is found, then ER = 1.

[0022] The formula for calculating the standard deviation of amplitude in step 3 is as follows:

[0023]

[0024]

[0025] In the formula, t is the time sequence number, taking values ​​of 1, 2, ..., 2n; S and S m These represent the time-domain amplitude vector and its mean of the original signal, respectively.

[0026] The milling state described in step 5 is either air cutting, stable, or chattering.

[0027] Compared with the prior art, the present invention has at least the following beneficial effects: the monitoring index energy ratio of the present invention directly reflects the energy proportion of the flutter component in the original vibration acceleration signal, enabling early monitoring of flutter; the energy ratio extraction method of the present invention does not require the use of a Kalman filter, and the accuracy of the energy ratio calculation is not affected by changes in processing parameters, allowing monitoring under various working conditions; the real-time performance of the present invention can meet the needs of online monitoring, enabling the processing of the signal from the previous sampling period and the provision of monitoring results of the milling processing status before the signal acquisition of the next sampling period is completed; the method can simultaneously meet the three requirements of early monitoring, insensitivity of the proposed monitoring index to changes in processing parameters, and real-time monitoring, and can accurately identify the amplitude and frequency of the main flutter frequency. Attached Figure Description

[0028] Figure 1 This is a system schematic diagram of the method of the present invention; wherein, 1 represents the spindle box, 2 represents the triaxial accelerometer, 3 represents the milling cutter, 4 represents the data acquisition card, and 5 represents the computer.

[0029] Figure 2The numerical experimental test results of the method of the present invention are shown below; where the horizontal axis represents the time domain sample length in points, the left vertical axis represents the CPU time in seconds, and the right vertical axis represents the ratio of average CPU time to sampling time (time ratio). Dimensionless indicators have no units.

[0030] Figure 3 This is a flowchart of the method of the present invention.

[0031] Figure 4 This is a schematic diagram of variable cut width processing according to an embodiment of the present invention.

[0032] Figure 5 This invention provides a time-domain plot of X-axis acceleration data, along with the variation curves of flutter monitoring indicators (ER and σ(a)) extracted from it, and a given sequence of monitoring results; wherein, Figure 5 (a) is the time-domain plot of the X-axis acceleration data. Figure 5 (b) shows the variation curves of flutter monitoring indicators (ER and σ(a)) extracted from the X-axis acceleration data and the given monitoring result sequence; Figure 5 The horizontal axis of both (a) and (b) represents time, in seconds; Figure 5 (a) The vertical axis represents the amplitude of the vibration signal, in m / s. 2 ; Figure 5 (b) The left ordinate represents the amplitude standard deviation extracted from the vibration signal, in m / s. 2 ; Figure 5 (b) The right ordinate represents the energy ratio extracted from the vibration signal. The dimensionless index has no unit.

[0033] Figure 6 shows the spectrum of X-axis, Y-axis, and Z-axis acceleration data within 12.44–12.64 s according to an embodiment of the present invention; wherein, Figure 6(a) is the spectrum of X-axis acceleration data; Figure 6(b) is the spectrum of Y-axis acceleration data; and Figure 6(c) is the spectrum of Z-axis acceleration data; in Figures 6(a), (b), and (c), the horizontal axis represents frequency in Hz; and in Figures 6(a), (b), and (c), the vertical axis represents vibration signal amplitude in m / s. 2 .

[0034] Figure 7 The diagram shows the confusion matrix of the test results obtained using two types of milling chatter monitoring models: Linear Support Vector Machine (LSVM) and Extreme Gradient Boosting (XGBoost). Figure 7 (a) is a confusion matrix diagram of the test results of LSVM; Figure 7 (b) is the confusion matrix of the XGBoost test results. Detailed Implementation

[0035] A system schematic diagram of the method of the present invention is shown below. Figure 1 As shown.

[0036] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0037] A method for online monitoring of milling chatter that integrates energy ratio and amplitude standard deviation includes the following steps:

[0038] Step 1: Collect real-time vibration signals of the spindle during the milling process using a triaxial vibration accelerometer installed at the end of the spindle box;

[0039] Step 2: Obtain the spectrum of the real-time vibration signal of the spindle through Fast Fourier Transform, convert the signal from the time domain to the frequency domain, and use a flutter peak search algorithm to search for all flutter peaks and periodic peaks in the frequency domain. Calculate the energy ratio between the flutter peaks and periodic peaks based on their energy magnitudes to obtain the energy ratio value of the flutter peaks and periodic peaks during the sampling period. Specifically, this includes the following steps:

[0040] In the frequency domain, a one-dimensional peak-finding algorithm is used to obtain all peak values ​​in the signal spectrum, with frequencies greater than the over-tooth frequency (ω). t Peaks with amplitudes greater than the corresponding frequency threshold are retained, and the threshold a at each frequency is... th The formula for calculating (i) is as follows:

[0041]

[0042]

[0043] median(Y)(i)=median({Y(k)})

[0044] a th (i)=σ(Y)+median(Y)(i)+Y m

[0045] In the formula, i is the frequency index, taking values ​​of 1, 2, ..., n; k is the frequency index in the median filter window, taking values ​​of i - (n Filter -1) / 2,i-(n Filter -1) / 2+1,…,i+(n Filter -1) / 2; Y represents the amplitude vector of the FFT result of the original signal at [n×1] frequencies; Y mσ(Y) and median(Y)(i) represent the mean, standard deviation, and median filtered value of the FFT result of the original signal, respectively, where the window size n of the median filter is... Filter =25.

[0046] Then, calculate the rotational frequency (ω) of each retained peak value with its nearest left and right pivots. s The interval between the peak value and its harmonics is considered a periodic peak if the interval is less than a set threshold, and a flutter peak if the interval is greater than a set threshold. For example, the set threshold is 0.1ω. s .

[0047] The energy ratio of flutter peaks to periodic peaks is defined as follows:

[0048]

[0049] In the formula, N represents the number of flutter peaks found; ω CH,i Represents the i-th flutter frequency; K represents the number of excluded frequencies; ω REJ,k This represents the k-th exclusion frequency. The numerator is the energy of the flutter component; the first term in the denominator is the energy of the amplitude at the exclusion frequency, which replaces the energy of the periodic component. Furthermore, if no acceptable spectral peak is found in the FFT results, then ER = 0; if no exclusionary spectral peak is found, then ER = 1.

[0050] Step 3: Calculate the standard deviation of the amplitude over the sampling period using the time-domain amplitude of the vibration signal; the definition of the standard deviation of amplitude is as follows:

[0051]

[0052]

[0053] In the formula, t is the time sequence number, taking values ​​of 1, 2, ..., 2n; S and S m These represent the time-domain amplitude vector and its mean of the original signal, respectively.

[0054] Step 4: Perform feature layer fusion on the energy ratio and amplitude standard deviation values ​​of the three-dimensional vibration signals to obtain a set of 6-dimensional monitoring indicators;

[0055] Step 5: Finally, input the fused 6-dimensional monitoring index set into the trained three-class monitoring model and output the milling processing status of the sampling time period, wherein the milling processing status is empty cutting, stable or chattering.

[0056] Time complexity is a key parameter for evaluating the real-time performance of monitoring methods. Since the time consumption of the algorithm is mainly concentrated in the monitoring index extraction part (extracting ER and extracting σ(a)), this invention analyzes the theoretical time complexity of the monitoring index extraction part.

[0057] (1) Time complexity of extracting ER

[0058] This part of the program includes Fast Fourier Transform (O(nlogn)) and Flutter Peak Search Algorithm (Median Filtering (O(nlogn))). 2 The time complexity of extracting the ER is as follows: (O(n)) = logn), (O(n)) = 10 ...

[0059] T ER =O(nlogn) +O(n 2 logn)+O(n)+O(2n)+O(logn)+O(n)+O(n×(N+K))+O(2N)

[0060] (2) Time complexity of extracting σ(a)

[0061] This part of the program only calculates the standard deviation of the time-domain signal amplitude (O(4n)), so the total time complexity of extraction is:

[0062] T σ(a) = O(4n)

[0063] Because O(n 2 The time complexity of logn is much larger than that of other parts. Therefore, the time complexity of the monitoring indicator extraction part can be summarized as follows:

[0064] T≈O(n 2 logn)

[0065] Therefore, the time complexity of the entire flutter detection algorithm should be non-linear, O(n^2). 2 The time complexity of the flutter detection algorithm grows in a mixed quadratic and logarithmic manner with the frequency domain sample length n, as the time domain sample length L is twice the frequency domain sample length n (a linear relationship). Therefore, the time complexity of the entire algorithm still maintains a mixed quadratic and logarithmic growth relationship with the frequency domain sample length L, characterized by a slow initial growth followed by a faster growth. Consider the sliding window parameters: window duration (t... w =L / f s =2n / f s ) and window interval duration (t) s ), using f pRepresenting the number of floating-point operations per second of the processor, the processing time of the signal segment in the sliding window can be expressed as:

[0066]

[0067] The actual duration of the update signal is t. s For real-time systems, the processing time is at most equal to the update sampling time, that is:

[0068] t cpu ≤t s ≤t w

[0069] When t w =t s At this point, the critical floating-point operations per second (FLOPS) can be obtained. c :

[0070]

[0071] If the time-domain sample length is L = 2048, and the sampling frequency is f s =10240Hz, spindle speed ω s =138.33Hz, over-tooth frequency ω t =276.66Hz, then f c =[1024×log2(1024)×10240] / 2FLOPS=0.0524GFLOPS.

[0072] The test platform configuration is as follows: Windows 10 x64, Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz (2C, 4T, 2.7GHz, 2.8GHz, IMC, 2×256kB L2, 3MB L3, 100MHz FSB), 16g DDR3. Using SiSoftware Sandra Professional Home 2018 software, the single-core computing power of the tested platform was obtained through aggregate algorithm performance testing. p It is 10.84 GFLOPS, while the theoretically calculated critical value f c It is 0.0524 GFLOPS, which satisfies f p >f c The real-time requirements.

[0073] For each time-domain sample length, many different types of samples were used to test the CPU time running on a single core (this CPU time is for the entire flutter monitoring algorithm, including spindle vibration data reading, sample construction, monitoring index extraction, and monitoring model prediction), and the average curve was plotted. The numerical experimental test results are as follows: Figure 2 As shown.

[0074] from Figure 2 It can be clearly seen that the average CPU timeline is below the sampling timeline, indicating the existence of a feasible region (e.g., Figure 2 The mid-section line region is formed by the sampling timeline (e.g., Figure 2 (double underline) and CPU time limit (e.g.) Figure 2 The area enclosed by the dashed line (in the middle) constitutes the area, which shows that the present invention can achieve real-time monitoring.

[0075] When the time-domain sample length is too short, there is an unstable region, so the time-domain sample length should not be too short. At the same time, as the time-domain sample length increases, the lower bound of the time window interval increases, and the feasible region shows a trend of first increasing and then decreasing, which is consistent with the previous theoretical analysis. In addition, it is noted that the CPU time margin of this algorithm is still very large, and the configuration of the test platform can be further reduced.

[0076] Example:

[0077] The main process of the online monitoring method for milling chatter based on the energy ratio and amplitude standard deviation of the spindle vibration acceleration signal of this invention is as follows: Figure 3 As shown.

[0078] Online chatter monitoring was performed on a 7050 aerospace aluminum alloy thin-walled plate during high-speed milling. The sampling frequency was 10240Hz. A 3-flute carbide end mill with a diameter of 20mm, a helix angle of 45°, and a body length of 104mm was used, with a tool overhang of 75mm during clamping. The thin-walled plate, 40mm thick, was clamped on the worktable using a vise. The cutting parameters were as follows: spindle speed 6000r / min, tool feed rate maintained at 240mm / min during milling, and tool path... Figure 4 In the medium feed direction, the reverse milling is performed, and the radial cutting width continuously increases from 0 mm to 15 mm. The milling process is dry cutting.

[0079] (1) Signal acquisition

[0080] Vibration information during milling is collected by a triaxial vibration acceleration sensor (with sensitivities of 50.2 mV / g, 48.6 mV / g, and 48.5 mV / g in the X, Y, and Z directions, respectively) located at the end of the spindle box. The collected X-axis vibration acceleration signal is as follows: Figure 5 As shown in (a). From Figure 5As shown in (a), during the 0–0.4s phase, the tool is in an idling state, and the signal amplitude is very small. After 0.4s, the tool transitions from idling to milling, and at 10.95s, the signal amplitude increases sharply. After 14.7s, the tool completely withdraws from the workpiece, and the signal amplitude decreases rapidly.

[0081] Figure 6 shows the spectra of the vibration acceleration signal in the X, Y, and Z directions within 12.44–12.64 s in this case. The X-direction acceleration spectrum (Figure 6(a)) shows a dominant chatter peak. The Y-direction acceleration spectrum (Figure 6(b)) also shows a chatter peak, but it is not yet dominant. The Z-direction acceleration spectrum (Figure 6(c)) shows a very small amplitude chatter peak. This is likely because the vibration caused by chatter is mainly distributed within the machining plane (xoy plane), while the Z-direction is along the tool axis, resulting in a very small component of the vibration acceleration in the Z-direction. Based on the X-direction acceleration spectrum and the surface texture of the machined workpiece, it can be concluded that chatter occurred during milling after 10.95 s.

[0082] (2) Extraction of monitoring indicators

[0083] Using 2048 sampling points as a sample, sliding time-window sampling was performed on the X, Y, and Z acceleration data to extract the energy ratio and the amplitude standard deviation of auxiliary monitoring indicators. The variation curves of the flutter monitoring indicators (ER and σ(a)) of the extracted X-axis acceleration data and the given monitoring result sequence are shown below. Figure 5 As shown in (b), the coding meanings of the monitoring results in the figure are as follows: 0 indicates air shear, 1 indicates stability, and 2 indicates flutter (the monitoring results share a vertical axis with ER).

[0084] from Figure 5 As can be seen from the monitoring result sequence in (b), the method of the present invention can detect the generation of flutter in the early stage (transition stage) and give an accurate warning, that is, the method of the present invention can achieve early monitoring.

[0085] Since two monitoring indicators can be extracted from each acceleration data direction, this invention considers first performing feature layer fusion on the two monitoring indicators extracted from the acceleration data direction, and then using the fused set of 6 monitoring indicators as the input of the monitoring model. This makes the operation more convenient in application, rather than using the two monitoring indicators extracted from the acceleration data direction as the input of the monitoring model separately, obtaining the monitoring results of the acceleration data direction, and then performing decision layer fusion.

[0086] (3) Monitoring of flutter state

[0087] Finally, the set of monitoring indicators after feature layer fusion is input into the trained monitoring model. The monitoring model then outputs the milling machining status monitoring results for the sampling period. If the result is empty cutting or stable, no warning is needed. If the result is chatter, a warning can be issued through the software platform. This invention selects two types of machine learning models, linear support vector machine and extreme gradient boosting tree, as milling chatter monitoring models. The effectiveness of the monitoring indicators proposed in this invention is verified through their test results. These two types of monitoring models are not intended to limit this invention.

[0088] Using a large dataset of indicators containing samples of air shuffle, stability, and flutter, the two monitoring models were trained, validated, and tested. The test results show that, for the LSVM and XGBoost constructed using the energy ratio and amplitude standard deviation extracted in this paper, the following conclusions can be drawn: LSVM achieves an average monitoring accuracy of 92.59% for air shuffle, stability, and flutter, with individual monitoring accuracies of 100.00%, 77.78%, and 100.00% for air shuffle, stability, and flutter, respectively (e.g., ...). Figure 7 (a)); XGBoost achieves an average monitoring accuracy of 87.58% for air cut, stabilization, and flutter, with individual monitoring accuracies of 100.00%, 68.94%, and 93.79% for air cut, stabilization, and flutter, respectively (e.g., Figure 7 (b)).

[0089] In summary, this invention provides an online monitoring method for milling chatter that integrates energy ratio and amplitude standard deviation. First, the original vibration acceleration signal is converted from the time domain to the frequency domain using a Fast Fourier Transform (FFT). Second, the chatter peak search algorithm proposed in this invention searches for all chatter peaks (spectral peaks related to the low-order natural frequencies of the milling process system) and periodic peaks (spectral peaks related to the spindle rotation frequency and its harmonics) in the frequency domain. Third, the energy magnitudes of these chatter peaks are calculated. Then, the energy magnitude of the periodic peaks in the spectrum is calculated. Finally, the energy magnitudes of the chatter peaks and periodic peaks are substituted into the energy ratio calculation formula constructed in this invention to obtain the energy ratio value. By combining this with the auxiliary monitoring index amplitude standard deviation, online monitoring of milling chatter can be achieved. The purpose of the search algorithm is to find the chatter peaks among all peaks; the periodic peaks are a secondary result, referring to all other peaks that are not identified as chatter peaks.

Claims

1. A method for online monitoring of milling chatter based on a combination of energy ratio and amplitude standard deviation, characterized in that, Includes the following steps: Step 1: Collect real-time vibration signals of the spindle during the milling process using a triaxial vibration acceleration sensor installed at the end of the spindle box; Step 2: Obtain the spectrum of the real-time vibration signal of the main shaft through fast Fourier transform, convert the signal from the time domain to the frequency domain, and use the flutter peak search method to search for all flutter peaks and periodic peaks in the frequency domain. Calculate the energy ratio of the flutter peaks and periodic peaks based on their energy magnitudes to obtain the energy ratio values ​​of the flutter peaks and periodic peaks during the sampling period. Step 3: Calculate the standard deviation of the amplitude over the sampling period using the time-domain amplitude of the vibration signal. Step 4: Perform feature layer fusion on the energy ratio and amplitude standard deviation values ​​of the three-dimensional vibration signals to obtain a set of 6-dimensional monitoring indicators; Step 5: Finally, input the fused 6-dimensional monitoring index set into the trained three-class monitoring model and output the milling processing status of the sampling time period; the chatter peak search method in Step 2 includes five parts in sequence: median filtering, mean filtering, standard deviation filtering, one-dimensional peak finding, threshold elimination and frequency elimination. In the frequency domain, a one-dimensional peak-finding algorithm is used to obtain all peak values ​​in the signal spectrum, with frequencies greater than the over-tooth frequency being selected. ω t Furthermore, peak values ​​with amplitudes greater than the corresponding frequency threshold are retained, and the nearest spindle frequencies to the left and right of each retained peak value are calculated. ω s The interval value of its harmonics, if the interval value is less than the set threshold, the interval value ranges from [0, ... ω s If the peak value is a periodic peak, then it is considered to be a periodic peak. The milling process is described as either air cutting, stable, or chattering.

2. The online monitoring method for milling chatter based on comprehensive energy ratio and amplitude standard deviation according to claim 1, characterized in that, Before performing median filtering in step 2, the spectrum vector Y is first padded before and after ( n Filter –1) / 2 zero values ​​are used to ensure that the median filtering operation can be applied before and after the spectral vector Y. n Filter –1) / 2 frequencies, where n Filter This is the window size for median filtering.

3. The online monitoring method for milling chatter based on comprehensive energy ratio and amplitude standard deviation according to claim 1, characterized in that, In step 2, when searching for chatter peaks, consider frequencies higher than the over-tooth frequency. ω t Furthermore, the amplitude of the chatter peak is greater than the threshold at the corresponding frequency, and if multiple chatter peaks are found between adjacent harmonics of the spindle frequency, only the chatter peak with the largest amplitude is retained.

4. The online monitoring method for milling chatter based on comprehensive energy ratio and amplitude standard deviation according to claim 1, characterized in that, By using an open CNC system, the spindle speed Ω and the number of tool teeth are collected in real time. N t Perform spindle frequency adjustment ( ω s =Ω / 60) and over-tooth frequency ( ω t = N t ω s ) calculation.

5. The online monitoring method for milling chatter based on comprehensive energy ratio and amplitude standard deviation according to claim 1, characterized in that, The energy ratio calculation formula in step 1 is defined as follows: In the formula, N This indicates the number of flutter peaks found. ω CH,i Indicates the first i One flutter frequency; K Indicates the number of frequencies excluded; ω REJ,k Indicates the first k The numerator is the energy of the flutter component; the first term in the denominator is the energy of the amplitude at the excluded frequency, replacing the energy of the periodic component. If no acceptable spectral peak is found in the Fast Fourier Transform results, then... ER =0, no spectral peaks that can be excluded were found. ER =1.

6. The online monitoring method for milling chatter based on comprehensive energy ratio and amplitude standard deviation according to claim 1, characterized in that, The formula for calculating the standard deviation of amplitude in step 3 is as follows: In the formula, t It is a time sequence number, with values ​​1, 2, ..., 2 n S and S m These represent the time-domain amplitude vector and its mean of the original signal, respectively.

Images