Signal noise reduction method for acoustic monitoring of tool wear state in milling process

Abstract

The invention provides a signal noise reduction method for acoustic monitoring of a tool wear state in a milling process, which relates to the technical field of intelligent manufacturing, and is characterized in that main shaft noise is eliminated through spectral characteristics of IMF components obtained through adaptive noise complete set empirical mode decomposition in combination with spectral characteristics of signals acquired by a multi-scene milling test; first-stage noise reduction is realized; performing fast independent component analysis on the residual IMF components after the first-stage noise reduction to realize blind source separation, and identifying and removing random noise components in combination with a fuzzy entropy threshold criterion to realize second-stage noise reduction; reconstructing a de-noised signal by using the residual signal components after the second-stage noise reduction, and screening out features conforming to a tool wear trend through a Kendall rank correlation coefficient; and inputting the screened features conforming to the tool wear trend into a convolutional neural network to realize high-precision intelligent monitoring of the tool wear state based on the sound signal. According to the invention, stable denoising performance is maintained under different working conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing, and particularly relates to a signal denoising method for acoustic monitoring of tool wear state in a milling process. BACKGROUND

[0002] Milling is a widely used high-efficiency cutting process in aerospace, automotive and shipbuilding industries. Studies have shown that 10%-40% of the total machine downtime can be attributed to abnormal tool conditions. In the milling process, the contact pressure between the tool, workpiece and chip causes tool deformation, which manifests as progressive wear or sudden fracture. Tool wear not only reduces product machining accuracy, but also leads to tool loss, material waste and increased time cost, so tool condition monitoring is of great necessity.

[0003] The technology aims to identify and predict tool conditions through sensor signal processing and pattern recognition methods to reduce losses caused by wear or failure. Tool condition is usually defined based on geometric deformation, where direct monitoring methods require interrupting the machining process, and image measurement, machine vision and wear area estimation are used to detect geometric dimensions through sensors or cameras. However, such methods are mostly limited to laboratory environments and lack applicability in industrial settings. Indirect monitoring methods use the correlation between cutting force, vibration, sound, acoustic emission (AE), temperature and surface roughness and tool wear state for continuous or intermittent estimation, with the advantages of easy system configuration and strong engineering applicability. Signals generated during the machining process, such as cutting force, temperature, noise and vibration, contain information about tool life, cutting quality and workpiece defects. In particular, audible sound signals generated by continuous friction between the tool, workpiece and flowing chip can serve as effective indicators of tool health. Compared to force, vibration, current and acoustic emission sensors, audible sound signals based on microphones have four key advantages: strong correlation with tool wear, non-contact measurement, flexible and convenient acquisition, and no impact on normal equipment operation. Therefore, acoustic diagnosis technology has become a research focus in the field of tool wear monitoring.

[0004] With the development of artificial intelligence (AI) technology, deep learning (DL)-based methods are increasingly used in tool wear monitoring. However, the effectiveness of deep learning models is highly dependent on data quality. The collected acoustic signals are often contaminated by various complex noises, leading to significant degradation in signal quality and thus reducing the accuracy of tool wear state recognition. The acoustic signals generated during the milling process are composed of multiple source signals, including chip deformation, tool-workpiece friction, tool vibration, workpiece vibration, machine noise, and environmental noise. The frequency bands of these multiple source signals overlap, and the audible sound signals generated by tool-workpiece contact are affected by the spindle, worktable, and box wall in the closed space of the machine tool, resulting in reflection, diffraction, and other physical phenomena. This leads to differences in acoustic signals collected at different positions, further increasing signal complexity. Therefore, the monitored acoustic signals often exhibit high noise, non-stationary, and nonlinear characteristics, leading to decreased diagnostic model performance and increased data uncertainty. Therefore, it is crucial to use appropriate denoising techniques to improve the accuracy of milling tool wear monitoring.

[0005] Current research on acoustic signal denoising methods mainly focuses on spectral subtraction, filters, fast Fourier transform (FFT), wavelet analysis, and empirical mode decomposition (EMD) adaptive decomposition techniques. FFT, as a standard method for frequency domain signal analysis, has been widely studied in the field of tool condition monitoring. Although it has been widely used in early applications, FFT has significant theoretical limitations when processing machining signals: it cannot effectively process time-varying signals, analyze non-stationary signals, and avoid frequency spectrum leakage. Wavelet theory was proposed by Malla and Daubechies in the late 1980s, and because it overcomes the shortcomings of the FFT fixed window function, wavelet transform (WT) exhibits superior performance in the time-frequency domain local feature representation of non-stationary signals. However, the effectiveness of wavelet denoising depends on the reasonable selection of key parameters such as wavelet basis function, threshold strategy, and decomposition level. The optimization of these parameters poses a significant technical challenge.

[0006] Independent component analysis (ICA), as a blind source separation (BSS) method, is widely used as a denoising technique. When the prior information of the signal is unknown, ICA can adaptively separate the mixed signal into independent components (ICs) while preserving the characteristic information of the source signal. This method has been widely used in image processing, signal processing, and other fields. However, ICA has inherent limitations: it assumes statistical independence of multiple source signals, cannot identify noise components, and cannot handle single-channel signals. Furthermore, independent component analysis (ICA) has certain limitations in identifying noise-related independent components (ICs), and selecting an appropriate threshold is also a major challenge.

[0007] To overcome these problems, researchers have proposed hybrid denoising methods to break through the above limitations. Based on modal decomposition, single-channel signals can be decomposed into multiple intrinsic mode functions (IMFs), which meet the input assumption of ICA; ICA further analyzes the implicit correlation between IMFs and decouples them into linear combinations of mutually independent ICs, achieving noise component separation. Shi et al. proposed a signal processing method based on empirical mode decomposition (EMD) and ICA, which effectively separates the cutting sound signal in face milling and successfully extracts the different source components of normal / worn tools, achieving effective detection of tool wear state. Zhang et al. developed a multi-scale ICA method based on ensemble empirical mode decomposition (EEMD-MSICA), which not only achieved bearing fault detection but also provided a new mechanism for multi-channel signal denoising, effectively solving the problem of multi-scale dynamic signal processing in early fault detection of large rotary bearings.

[0008] However, these methods still have certain limitations. For example, the EMD method has mode mixing, end effect, and stopping criterion constraints, which directly affect its denoising performance. To this end, Wu et al. proposed an improved noise-assisted signal decomposition technique, ensemble empirical mode decomposition (EEMD), which suppresses mode mixing by continuously adding white noise. However, EEMD cannot effectively eliminate residual noise, causing reconstruction error. To further overcome these problems, researchers have proposed adaptive noise complete ensemble empirical mode decomposition (CEEMDAN), which adds a limited amount of adaptive white noise to each stage of signal decomposition and calculates a unique residual signal to obtain IMFs, effectively overcoming the EMD mode mixing problem. Vanraj et al. based on CEEMDAN achieved local defect diagnosis of fixed-axis gear box sound signals, verifying that it can obtain more accurate IMFs and improve the fault severity diagnosis capability compared to EEMD. Li and Wang used CEEMDAN to decompose sound signals and combined the least mean square criterion with an adaptive filter to achieve denoising, but the CEEMDAN decomposition results still contain a small amount of noise and are time-consuming to calculate.

[0009] In addition, entropy, as a complexity measure of signal analysis, includes Shannon entropy, approximate entropy (ApEn), sample entropy (SampEn), multi-scale entropy (MSE), and fuzzy entropy (FuzzyEn), which can provide support for noise component identification. The calculation of entropy does not depend on linear assumptions and is suitable for the analysis of complex systems. Fuzzy entropy (FuzzyEn) has the best performance among various entropies due to its less dependence on data set size and stronger anti-noise ability. Existing research has proposed calculating the fuzzy entropy values of ICs and combining threshold discrimination methods to identify and eliminate noise components. However, this method still faces the challenge of selecting the appropriate threshold, and in complex variable working conditions and random noise environments, traditional denoising methods still struggle to effectively extract effective signal components, resulting in insufficient classification accuracy of tool wear state.

[0010] Therefore, although the existing hybrid denoising method solves the shortcomings of the traditional method to a certain extent, there are still obvious limitations in noise complexity, multi-source interference and time-varying characteristics, and further technical innovation and optimization are needed. SUMMARY

[0011] In view of the deficiencies in the prior art, the present application provides a signal denoising method for acoustic monitoring of tool wear state in a milling process, which realizes the first stage of denoising by improved adaptive noise complete ensemble empirical mode decomposition (ICEEMDAN) combined with the frequency spectrum characteristics of each IMF and the test results to remove the spindle noise; the second stage of denoising is realized by identifying and removing the random noise component through independent component analysis (ICA) combined with fuzzy entropy threshold criterion; finally, the denoised signal is reconstructed using the remaining signal components, and the features are screened through Kendall rank correlation coefficient, and the high-precision intelligent monitoring of tool wear state based on sound signal is realized using convolutional neural network (CNN).

[0012] A signal denoising method for acoustic monitoring of tool wear state in a milling process, comprising the following steps:

[0013] Step one: the frequency spectrum characteristics of the IMF components obtained by adaptive noise complete ensemble empirical mode decomposition (ICEEMDAN) are combined with the frequency spectrum characteristics of the signals collected in multi-scene milling tests to remove the spindle noise, and the first stage of denoising is realized;

[0014] Step two: the remaining IMF components after the first stage of denoising in step one are subjected to blind source separation through fast independent component analysis (Fast-ICA), and the random noise component is identified and removed through fuzzy entropy threshold criterion, and the second stage of denoising is realized;

[0015] Step three: the denoised signal is reconstructed using the remaining signal components after the second stage of denoising, and the features conforming to the tool wear trend are screened out through Kendall rank correlation coefficient;

[0016] Step four: the features conforming to the tool wear trend screened out are input into the convolutional neural network (CNN) to realize high-precision intelligent monitoring of tool wear state based on sound signal.

[0017] In the above scheme, step one comprises:

[0018] S11: Acquire sound signal and tool wear data: for sound signal data, remove unstable signal data in the cutting-in and cutting-out parts, and retain signal data in the stable milling stage; for tool wear data, observe the cutting edge photo through a microscope and calibrate the wear, and divide the wear interval;

[0019] S12: The original acoustic signal is decomposed into several intrinsic mode functions (IMFs) by the improved adaptive noise complete set empirical mode decomposition (ICEEMDAN);

[0020] S13: Design a multi-scenario noise acquisition experiment, and eliminate the main axis noise based on the spectral characteristics of each IMF and the test results to achieve the first stage of noise reduction.

[0021] In the above scheme, step two specifically includes: the remaining IMF is used as the input source of independent component analysis (ICA), and further separated into uncorrelated independent components (ICs). By calculating the fuzzy entropy (FuzzyEn) value of each IC and applying the fuzzy entropy threshold criterion, random noise components are identified and eliminated to achieve the second stage of noise reduction.

[0022] In the above scheme, step three specifically includes: reconstructing the denoised signal using the retained ICs, extracting time-domain, frequency-domain, and time-frequency-domain features from the denoised signal, and selecting features that conform to the trend of the tool wear curve through Kendall's rank correlation coefficient.

[0023] In the above scheme, step four specifically includes: constructing a CNN deep learning model, training the model using the selected features, and predicting the tool wear state based on the denoised signal.

[0024] In the above scheme, the single-channel signal is decomposed into multiple intrinsic mode functions (IMFs) using the ICEEMDAN method. The ICEEMDAN method is used to decompose the original signal into x[n], where ω... (i) (i = 1, 2, ..., l) represents the Gaussian white noise added in the i-th group. In each iteration of IMF component extraction, the introduced noise signal is the IMF component obtained by decomposing the original Gaussian white noise signal, which is E. k (ω (i) ), (i = 1, 2, ..., l), operator E k (·) represents the extraction of the k-th IMF component from the Empirical Mode Decomposition (EMD) of the signal; the parameter β = ε / std(x) represents the multiplier coefficient of the noise component, which is defined as the ratio of the signal-to-noise ratio (SNR) between the added noise and the analyzed signal to the standard deviation of the analyzed signal, where std(·) is the standard deviation operator, whose mathematical definition is... In the formula, N is the number of sampling points; the operator M·) represents the local mean of the signal, which is defined as the residual obtained by subtracting the IMF component from the original signal, i.e., satisfying M(·)=x[n]-E k Mathematical relationships (·);

[0025] S201: Calculate "X" through l iterations of Empirical Mode Decomposition (EMD). 1i =x[n]+β0E1(ω (i)The local mean of [n], (i = 1, 2, ..., l) is denoted as M(X). 1i (i = 1, 2, ...);) to obtain the first-order residual Where β0=ε0 / std(x[n]), ε0 represents the signal-to-noise ratio (SNR) between the added noise and the analyzed signal, and std(·) is the standard deviation operator; E1(ω (i) [n]) is the first IMF component extracted by Empirical Mode Decomposition (EMD) of the Gaussian white noise added to the i-th group;

[0026] S202: Calculate the first-order intrinsic mode components:

[0027] S203: Estimate the second-order residual as "X" 2i =r1[n]+β1E2(ω (i) The local mean of [n], (i = 1, 2, ..., l) is denoted as M(X). 2i (i = 1, 2, ..., l), to obtain the second-order residual. Where β1=ε1 / std(r1[n]), ε1 represents the signal-to-noise ratio (SNR) between the added noise and the analyzed signal, and std(·) is the standard deviation operator; E2(ω (i) [n]) represents the second IMF component extracted from the empirical mode decomposition (EMD) of the Gaussian white noise added to the i-th group; solve for the second-order intrinsic mode components:

[0028] S204: For the k-th order residual (k = 3, 4, ... K), calculate "M(r k-1 [n]+β k-1 E k (ω (i) [n])),(i=1,2,...l)”, where the coefficient β k-1 =ε k-1 / std(r k-1 ), ε k-1 The signal-to-noise ratio (SNR) between the added noise and the analyzed signal is represented by std(·), where std(·) is the standard deviation operator.

[0029] S205: Calculate the kth eigenmode components:

[0030] S206: Perform step S204 for subsequent k-value iterations until the final residual r is reached. k The termination conditions are met;

[0031] Based on the structure of ICEEMDAN, the final residuals satisfy:

[0032]

[0033] Where K is the total number of IMFs, therefore, the signal x can be represented as:

[0034]

[0035] In the above scheme, steps two and three specifically include:

[0036] S211: The observed signal vector is x. If A represents the mixing matrix, the source signal vector s obtained according to the mixing model formula (3) is called the independent component (ICs).

[0037] x = A·s (3)

[0038] S212: Under the premise that the source signals are statistically independent and the mixing matrix A is invertible, the ICA algorithm estimates the mixing matrix A and the separation matrix W based on the observed mixed signal x, thereby reconstructing the data matrix y—that is, representing the mathematical expression of the estimated independent components (ICs)—and the relationship can be expressed as follows:

[0039] 1 = W·x(4)

[0040] S213: After calculating the fuzzy entropy (FuEn) value of each independent component (ICs), a threshold discrimination method is introduced to identify noise-related independent components. The discrimination criteria are as follows:

[0041]

[0042] Symbol here Let g represent the fuzzy entropy value of the g-th independent component (IC) arranged in ascending order, where g is the fuzzy entropy value that satisfies The minimum integer of the condition, where N is the total number of ICs. If there is no g value that satisfies the condition within the specified range, g = 1 is taken by default. When the g value satisfies equation (5), the corresponding IC is determined to be a noise component and is filtered by setting it to zero to obtain the denoised source signal vector y′. Then, the mixing matrix is ​​multiplied by the ICs matrix to obtain the denoised observation signal x′, as shown in equation (3). Finally, the signal c′ is reconstructed by superposition operation of the denoised observation signal.

[0043] S214: ICA successfully separated several independent components (ICs). The number of ICs obtained by blind source separation is usually less than or equal to the number of input IMFs.

[0044] S215: The noise-reduced signal exhibits significantly smoothed temporal characteristics while maintaining the overall trend of the original signal.

[0045] In the above scheme, a CNN deep learning model is constructed: a deep learning model based on convolutional neural network (CNN) is used for tool wear classification training. By adaptively extracting the latent patterns of input features, a nonlinear mapping relationship between features and wear state is established to realize tool wear state monitoring.

[0046] Compared with existing technologies, it has the following advantages:

[0047] This invention utilizes ICEEMDAN adaptive decomposition of single-channel signals to achieve precise separation of spindle noise based on IMF spectral characteristics and experimental results. It reconstructs the independent components (ICs) generated by ICA using fuzzy entropy as a threshold criterion, achieving automated removal of random noise. The reconstructed signal strictly adheres to the maximum correlation-minimum error criterion, closely approximating the original effective signal. Feature analysis is employed to select features that effectively reflect tool wear trends, significantly improving the accuracy of tool wear state classification. Furthermore, this method maintains stable denoising performance under different operating conditions, highlighting its application value in practical industrial environments. The research results demonstrate that the proposed denoising framework has broad application prospects in the field of intelligent monitoring, providing a reliable theoretical foundation and technical support for tool wear state monitoring and fault diagnosis in complex industrial environments. It also offers an innovative solution for the intelligent development of industrial equipment health management systems. Attached Figure Description

[0048] Figure 1 This is the pseudocode for the ICEEMDAN algorithm;

[0049] Figure 2 This is the pseudocode for the ICA-FuEn algorithm;

[0050] Figure 3 Here is a flowchart of the ICEEMDAN algorithm;

[0051] Figure 4 This is a diagram of the ICA blind source separation and denoising process;

[0052] Figure 5 A diagram showing the noise generated by the milling machine;

[0053] Figure 6 This is the overall framework diagram;

[0054] Figure 7 For milling test platform;

[0055] Figure 8 The tool wear curves are shown in (a) T01, (b) T02, (c) T03, and (d) tool wear is divided into four levels, with T02 as an example: light wear: 0-55 μm, average wear: 55-103 μm, heavy wear: 103-123 μm, and severe wear: 123-150 μm.

[0056] Figure 9 Time-domain and frequency-domain plots of sound signals acquired for independent noise tests;

[0057] Figure 10 Comparison of the three-axis vibration and acoustic signal spectra of the spindle for T02-1 (first milling cycle) and T02-40 (last milling cycle);

[0058] Figure 11 The noise spectrum of the spindle idling at 6000 rpm;

[0059] Figure 12 Kendall coefficients for different characteristics of T01, T02, and T03;

[0060] Figure 13 Time-domain and frequency-domain plots of the ICEEMDAN decomposition of T02-1;

[0061] Figure 14 The time-domain and frequency-domain plots of the T02-1 ICA decomposition;

[0062] Figure 15 Comparison and local amplification of the original signals and denoised signals of the initial and final milling cycles of T01, T02, and T03;

[0063] Figure 16 Comparison of tool wear state classification accuracy before and after noise reduction for T01, T02, and T03;

[0064] Figure 17 The confusion matrix for classifying tool wear states before and after denoising T01, T02, and T03.

[0065] The following labels are used in the attached diagram: 1-Contact force noise source; 2-Spindle noise source; 3-Motor noise source; 4-Lead screw noise source; 5-Gear noise source; 6-Microphone; 7-Spindle; 8-Bracket; 9-Worktable; 10-Tool; 11-Workpiece. Detailed Implementation

[0066] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0067] In the description of this invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "axial", "radial", "vertical", "horizontal", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0068] A robust two-stage noise reduction method for monitoring tool wear conditions in milling processes includes the following steps:

[0069] S1: Acquire sound signals and tool wear data: For sound signal data, remove unstable signal data during the entry and exit phases, and retain signal data during the stable milling phase; For tool wear data, observe the tool edge photographs under a microscope, perform wear calibration, and divide the wear range.

[0070] S2: The original acoustic signal is decomposed into several intrinsic mode functions (IMFs) by the improved adaptive noise complete set empirical mode decomposition (ICEEMDAN);

[0071] S3: Design a multi-scenario noise acquisition experiment, and eliminate spindle noise based on the spectral characteristics of each IMF and the test results;

[0072] S4: The remaining IMF serves as the input source for Independent Component Analysis (ICA), which is further separated into uncorrelated independent components (ICs). By calculating the fuzzy entropy (FuzzyEn) value of each IC and applying the fuzzy entropy threshold criterion, noise components are identified and eliminated.

[0073] S5: Reconstruct the denoised signal using the retained ICs, extract time-domain, frequency-domain, and time-frequency-domain features from the denoised signal, and select features that conform to the trend of the tool wear curve using Kendall's rank correlation coefficient.

[0074] S6: Construct a CNN deep learning model and train the model using the selected features so that the model can predict the tool wear state based on the denoised signal.

[0075] Preferably, the specific steps of S1 are as follows:

[0076] S101: Milling Test Setup and Sound Signal Acquisition:

[0077] To obtain data samples of tool wear acoustic signals, a Yuling D10*30*75*4T four-flute end mill was used for milling experiments with a cutting diameter of 10mm, a cutting edge length of 30mm, and a total length of 75mm. Tool wear data and corresponding acoustic signals were collected simultaneously during the cutting phase of the experiment. Three experimental schemes were designed based on the differences in feed rate and cutting width. Specific parameter configurations are shown in Table 1.

[0078] Table 1 Milling Test Parameters

[0079]

[0080] The experiment was conducted on a DX650 CNC machine tool. The experimental setup and measuring point layout are as follows: Figure 5 As shown. The 45 steel specimen was fixed to the machine tool worktable 9 by a connecting plate. The end mill performed five consecutive unidirectional climb milling operations along the outer contour of the workpiece. Sound signals were synchronously acquired during each milling cycle, with the feed direction aligned with the positive x-axis. After every five milling operations, the tool flank wear was measured using a 5301-D400 industrial microscope. A GRAS microphone, with a sensitivity of 50mV / Pa at 250Hz and a frequency response range of 5Hz-10kHz (±1dB) and 3.15Hz-20kHz (±2dB), was mounted on a bracket 1m above the ground, 68cm from the initial tool coordinates (-448.872, -30.193, -193.080). This microphone was connected to a Dewesoft X3 data acquisition system, and the acquired signals were stored and analyzed using Dewesoft X software at a sampling rate of 50kHz.

[0081] S102: Tool wear results and wear zone division:

[0082] Figure 8 The wear degradation trajectories of the three sets of cutting tools used in the experiment are shown. These trajectories were obtained by measuring the wear width of the four cutting edges after each milling cycle and calculating their average value. The flank wear width (VB) was determined using the "Vogel-Burke" measurement method—averaging 10 measurements taken at different positions along the cutting edge. Tools T01 and T03 underwent 35 cutting cycles each, while T02 completed 40 cycles, with edge wear measurements performed after each cycle. The wear state zones of T01 and T02 were more pronounced, with the break-in period, steady-state zone, and failure zone being more distinct, while T03 exhibited a more uniform wear evolution process. Figure 8 As shown, based on the average flank wear width value, the tool wear state is divided into four categories.

[0083] Preferably, the specific steps of S2 are as follows:

[0084] For the original datasets T01, T02, and T03, containing 35, 40, and 35 milling cycles respectively, 50,000 sampling points were extracted from the stable segment of each cycle, with a duration of 1 second, for analysis. The single-channel signal was decomposed into multiple intrinsic mode functions (IMFs) using the ICEEMDAN method. The ICEEMDAN algorithm parameters were set as follows: the number of implementations and the maximum number of filtering iterations were 50, and the added noise amplitude was 0.02 times the standard deviation of the signal.

[0085] The pseudocode and flowchart of the ICEEMDAN algorithm are as follows: Figure 1 and Figure 3 As shown. The ICEEMDAN method assumes the original signal is x[n], where ω(i) (i = 1, 2, ..., l) represents the Gaussian white noise added in the i-th group. In each iteration of IMF component extraction, the introduced noise signal is the IMF component obtained by decomposing the original Gaussian white noise signal, which is E k (ω (i) ), (i = 1, 2, ..., l), operator E k (·) represents the extraction of the k-th IMF component from the Empirical Mode Decomposition (EMD) of the signal; the parameter β = ε / std(x) represents the multiplier coefficient of the noise component, which is defined as the ratio of the signal-to-noise ratio (SNR) between the added noise and the analyzed signal to the standard deviation of the analyzed signal, where std(·) is the standard deviation operator, whose mathematical definition is... In the formula, N is the number of sampling points; the operator M·) represents the local mean of the signal, which is defined as the residual obtained by subtracting the IMF component from the original signal, i.e., satisfying MM(·)=x[n]-E k Mathematical relationships (·).

[0086] S201: Calculate "X" through l iterations of Empirical Mode Decomposition (EMD). 1i =x[n]+β0E1(ω (i) The local mean of [n], (i = 1, 2, ..., l) is denoted as M(X). 1i (i = 1, 2, ..., l), to obtain the first-order residual. Where β0=ε0 / std(x[n]), s0 characterizes the signal-to-noise ratio (SNR) between the added noise and the analyzed signal, and std(·) is the standard deviation operator; E1(ω (i) [n]) is the first IMF component extracted by Empirical Mode Decomposition (EMD) of the Gaussian white noise added to the i-th group;

[0087] S202: Calculate the first-order intrinsic mode components:

[0088] S203: S201: Calculate "X" through l iterations of Empirical Mode Decomposition (EMD).1i =x[n]+β0E1(ω (i) The local mean of [n], (i = 1, 2, ..., l) is denoted as M(X). li (i = 1, 2, ..., l), to obtain the first-order residual. Where β0=ε0 / std(x[n]), ε0 represents the signal-to-noise ratio (SNR) between the added noise and the analyzed signal, and std(·) is the standard deviation operator; E1(ω (i) [n]) is the first IMF component extracted by Empirical Mode Decomposition (EMD) of the Gaussian white noise added to the i-th group;

[0089] S204: For the k-th order residual (k = 3, 4, ... K), calculate "M(r)". k-1 [n]+β k-1 E k (ω (l) [n]))″. The coefficient β is among them. k-1 From ε k-1 / std(r k-1 ) definition, ε k-1 The signal-to-noise ratio (SNR) between the added noise and the analyzed signal is represented by std(·), which is the standard deviation operator, mathematically defined as follows: In the formula, N is the number of sampling points;

[0090] S205: Calculate the kth eigenmode components:

[0091] S206: Perform step 4 for subsequent k-value iterations until the final residual r is reached. k The termination conditions have been met.

[0092] Based on the structure of ICEEMDAN, the final residuals satisfy:

[0093]

[0094] Where K is the total number of IMFs. Therefore, signal x can be represented as:

[0095]

[0096] Preferably, the specific steps of S3 are as follows:

[0097] S301: To obtain the internal noise of the machine tool, milling-related noise, and environmental interference noise during the milling process, a separate noise acquisition experiment was conducted on a DX650 CNC machine tool. The acoustic signal generated during the milling process is composed of the superposition of multi-source excitation signals, including chip deformation, tool-workpiece friction, tool vibration, workpiece vibration, machine tool noise, and environmental noise. The noise generation mechanism of the milling machine is as follows: Figure 5As shown. A CNC machine tool consists of multiple subsystems, including the machine tool body, CNC system, drive system, transmission system, auxiliary system, programming, and peripheral equipment. Internal noise sources include motor noise from the drive system, operating noise from the hydraulic system, mechanical motion noise from the transmission system, spindle rotation noise, and workpiece machining noise; external environmental noise mainly includes human conversation, ambient background noise, and random transient noise.

[0098] This study designed a multi-scenario milling experiment to detect various noise signals. The specific experimental parameters are detailed in Table 2. The data acquisition system configuration is consistent with that of the previous experiment.

[0099] Table 2 Multi-Scene Noise Acquisition Experimental Setup

[0100]

[0101] S302: The time and frequency domain representations of the acquired signal are as follows: Figure 9 As shown, the analysis indicates that the noise energy is mainly concentrated in the low-frequency range of 0-2500Hz. In the spectrum diagrams of samples SN-5, SN-6, and SN-7, a significant spectral peak was observed in the 8000-10000Hz range, a frequency band not found in other noise signals. Based on spectral feature comparison, this peak can be preliminarily identified as characteristic noise generated by the spindle rotation.

[0102] S303: Figure 10 This paper presents a comparative analysis of the spindle triaxial vibration and acoustic signal spectra for the first milling cycle (T02-1) and the final milling cycle (T02-40) using tool T02. The data shows that the characteristic frequencies of 8803Hz and 9202Hz exhibit stable responses in both the vibration and acoustic spectra, with their amplitudes remaining constant throughout the milling process and showing no significant correlation with tool wear. Furthermore, Figure 11 In the spectrum of the spindle at 6000 rpm idle, dominant frequency components still exist at 8803 Hz and 9202 Hz, a characteristic consistent with observations during the milling process. This chain of evidence confirms that the spectral responses at 8803 Hz and 9202 Hz are generated by spindle rotation and should be filtered out during signal noise reduction.

[0103] Figure 13 The decomposition results of the first milling cycle of T02 are shown, generating 12 IMFs and 1 residual component. Among them, the highest energy frequency of IMF1 coincides with the spindle noise frequency, as marked by the red box, and is therefore discarded.

[0104] Preferably, the specific steps of S4 are as follows:

[0105] S401: The remaining 12 IMFs are processed by Fast-ICA blind source separation. Independent Component Analysis (ICA), as a classic blind source separation technique, can decompose the observed signal into a linear combination of several unknown independent source signals. Let the observed signal vector be x (in this study, the intrinsic mode function IMF obtained by the ICEEMDAN model), the unknown source signal vector obtained after separation is called independent component (ICs). If A represents the unknown mixing matrix, the source signal vector s obtained according to the mixing model formula (3) is called independent component (ICs);

[0106] x = A·s (3)

[0107] S402: Under the premise that the source signals are statistically independent and the mixing matrix A is invertible, the ICA algorithm estimates the mixing matrix A and the separation matrix W based on the observed mixed signal x. This process realizes the reconstruction of the data matrix y—that is, the mathematical expression representing the estimated independent components (ICs)—and its relationship can be expressed as:

[0108] y = W·x (4)

[0109] S403: After calculating the fuzzy entropy (FuEn) value of each independent component (ICs), a threshold discrimination method is introduced to identify noise-related independent components. The discrimination criteria are as follows:

[0110]

[0111] Symbol here Let g represent the fuzzy entropy value of the g-th independent component (IC) arranged in ascending order, where g is the fuzzy entropy value that satisfies The minimum integer of the condition, N is the total number of ICs. If there is no g value that satisfies the condition within the specified range, g = 1 is taken by default. When the g value satisfies equation (5), the corresponding IC is determined to be a noise component and is filtered by setting it to zero to obtain the denoised source signal vector y'; then the mixing matrix is ​​multiplied by the ICs matrix to obtain the denoised observation signal x', as shown in equation (3); finally, the signal c' is reconstructed by superposition operation of the denoised observation signal. The pseudocode and flow of the algorithm are as follows. Figure 2 and Figure 4 As shown.

[0112] S404: ICA successfully separated 12 independent components (ICs), such as Figure 14 As shown. It should be noted that the number of ICs obtained after blind source separation is usually less than or equal to the number of input IMFs. This method uses the fuzzy entropy threshold criterion as the screening standard, and ICs that satisfy equation (4) are judged as noise components and are zeroed out. Figure 14 The part marked in the red box is the noise component that was filtered out. The remaining ICs are reconstructed to form a denoised signal.

[0113] S405:Figure 15 The results of the comparative analysis of the original and denoised signals of the first and last milling cycles for tools T01, T02, and T03 are presented. The denoised signal exhibits significantly smoothed temporal characteristics while maintaining the overall trend of the original signal.

[0114] Preferably, the specific steps of S5 are as follows:

[0115] S501: Feature Extraction

[0116] Feature extraction was performed on the sampled data, covering three main categories: time domain, frequency domain, and time-frequency domain features. The mathematical definitions of each feature are detailed in Table 3. Time domain features include: absolute mean, average maximum, root mean square (RMS) value, root mean square (RMS) amplitude, skewness, kurtosis, arithmetic mean, margin factor, kurtosis factor, waveform factor, impulse factor, skewness factor, peak factor, standard deviation, and amplitude factor. Frequency domain features include: centroid frequency, mean square frequency, RMS frequency, frequency variance, power spectrum mean, power spectrum RMS, power spectrum peak factor, corrected equivalent bandwidth, stabilization ratio, band power skewness, and band power kurtosis. Time-frequency analysis employed db3 wavelet decomposition at 8 levels to calculate the Euclidean norm of each sub-band coefficient to quantify its "signal strength," i.e., energy. For each set of milling parameters, 34 original features were extracted from the acoustic signal under three working conditions, resulting in a 102-dimensional feature set. Each milling cycle contains 50,000 sampling points. A feature value is extracted for every 10,000 points, resulting in 175, 200, and 175 sets of feature values ​​for tools T01, T02, and T03, respectively.

[0117] Table 3. Time-domain, frequency-domain, and time-frequency-domain characteristics

[0118]

[0119]

[0120] S502: Feature Filtering

[0121] The Kendall rank correlation coefficient is used to assess feature monotonicity for feature selection. This method measures the relative order consistency between 34 feature sequences (arranged from the first to the last milling cycle) and their corresponding integer index sequences. The Kendall coefficient is defined as the ratio of the difference between the number of identical and disidentified pairs to the total number of possible pairs, and its mathematical definition is:

[0122]

[0123] Here, C represents the number of identically ordered pairs, D represents the number of disidentically ordered pairs, and n is the total number of data points. When the absolute value of the Kendall coefficient approaches 1, it indicates a stronger monotonicity of the data. For example... Figure 12 As shown, this study selected feature items with an absolute Kendall coefficient > 0.6 in the feature sequences of tools T01, T02, and T03 (corresponding to the numbers in Table 3). Among them, the features selected for T01 include absolute mean, average maximum value, root mean square value, root mean square amplitude, kurtosis, standard deviation, power spectrum mean, power spectrum root mean square, and ret1; T02 adds ret3 and ret8 on this basis; T03 also includes ret2, ret4, and ret6.

[0124] Preferably, the specific steps of S6 are as follows:

[0125] S601: Building a CNN deep learning model:

[0126] A deep learning model based on convolutional neural networks (CNN) was used for tool wear classification training. The model adaptively extracts the latent patterns of input features and establishes a nonlinear mapping relationship between features and wear state to realize tool wear state monitoring. The model hyperparameters are detailed in Table 4. The dataset is randomly divided into training set and test set in a 7:3 ratio. The sample size of each wear category T01, T02 and T03 is shown in Table 5. The training process includes three core steps: (1) Implementing a five-fold cross-validation strategy: The training set is randomly divided into five subsets. One subset is used as the validation set in each round of training, and the other four subsets are used as the training set; (2) 300 iterations per training session: After each iteration, the accuracy and loss value of the validation set and the training set are calculated synchronously to monitor the training process. After 300 iterations, the model classification accuracy is evaluated on the test set; (3) The final model accuracy is the arithmetic mean of the accuracy of the test set obtained from five rounds of training.

[0127] Table 4 CNN Model Parameter Settings

[0128]

[0129] Table 5. Sample size of each wear interval (T01, T02, T03)

[0130]

[0131] S602: Training Results

[0132] Table 6 presents a comparative analysis of the classification accuracy of tools T01, T02, and T03 before and after noise reduction. The accuracy improved by 2.98%, 8.41%, and 6.10%, respectively. After noise reduction, the average classification accuracy of the convolutional neural network (CNN) model for tool wear status can reach 95.15%. Figure 16 The results of this accuracy comparison are presented intuitively. Figure 17This shows in detail the classification confusion matrices of tools T01, T02, and T03 before and after CNN training.

[0133] Table 6 Classification accuracy before and after noise reduction for T01, T02, and T03

[0134]

[0135] The effectiveness and robustness of the proposed method were verified through three sets of milling experiments (T01, T02, and T03) using different cutting parameters. Based on the average wear width of the flank face of the four cutting edges, the tool wear state was divided into four stages: slight wear, average wear, heavy wear, and severe wear. A convolutional neural network (CNN) model was used for classification. The results show that, using the denoised signal, the CNN model achieved an average classification accuracy of 95.15% for tool wear state in the three sets of milling experiments (T01, T02, and T03), representing improvements of 2.98%, 8.41%, and 6.1% respectively compared to the original signal. The proposed denoising method can adapt to complex noise environments under various working conditions, significantly improving the monitoring accuracy of tool wear state.

[0136] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0137] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A signal noise reduction method for acoustic monitoring of tool wear conditions in milling processes, characterized in that, Includes the following steps: Step 1: Eliminate spindle noise by combining the spectral characteristics of the IMF components obtained by adaptive noise complete set empirical mode decomposition (ICEEMDAN) with the spectral characteristics of the signals collected from multi-scenario milling tests, thus achieving the first stage of noise reduction; Step 2: After the first stage of noise reduction in Step 1, the remaining IMF components are separated into blind sources using Fast Independent Component Analysis (Fast-ICA). Combined with the fuzzy entropy threshold criterion, random noise components are identified and removed to achieve the second stage of noise reduction. Step 3: Reconstruct the denoised signal using the remaining signal components after the second stage of noise reduction, and select features that match the tool wear trend using Kendall's rank correlation coefficient; Step 4: Input the selected features that match the tool wear trend into a convolutional neural network (CNN) to achieve high-precision intelligent monitoring of tool wear status based on sound signals.

2. The signal noise reduction method for acoustic monitoring of tool wear conditions in milling processes according to claim 1, characterized in that, Step one includes: S11: Acquire sound signals and tool wear data: For sound signal data, remove unstable signal data during the entry and exit phases, and retain signal data during the stable milling phase; For tool wear data, observe the tool edge photographs under a microscope, perform wear calibration, and divide the wear range. S12: The original acoustic signal is decomposed into several intrinsic mode functions (IMFs) by the improved adaptive noise complete set empirical mode decomposition (ICEEMDAN); S13: Design a multi-scenario noise acquisition experiment, and eliminate the main axis noise based on the spectral characteristics of each IMF and the test results to achieve the first stage of noise reduction.

3. The signal noise reduction method for acoustic monitoring of tool wear conditions in milling processes according to claim 1, characterized in that, Step two specifically includes: using the remaining IMF as the input source for Independent Component Analysis (ICA), further separating it into uncorrelated independent components (ICs), and identifying and removing random noise components by calculating the fuzzy entropy (FuzzyEn) value of each IC and applying the fuzzy entropy threshold criterion, thus achieving the second stage of noise reduction.

4. The signal noise reduction method for acoustic monitoring of tool wear conditions in milling processes according to claim 1, characterized in that, Step 3 specifically includes: reconstructing the denoised signal using the retained ICs, extracting time-domain, frequency-domain, and time-frequency-domain features from the denoised signal, and selecting features that conform to the trend of the tool wear curve using Kendall's rank correlation coefficient.

5. The signal noise reduction method for acoustic monitoring of tool wear conditions in milling processes according to claim 1, characterized in that, Step four specifically includes: constructing a CNN deep learning model, training the model using the selected features, and predicting the tool wear state based on the denoised signal.

6. The signal noise reduction method for acoustic monitoring of tool wear conditions in milling processes according to claim 2, characterized in that, The ICEEMDAN method decomposes a single-channel signal into multiple intrinsic mode functions (IMFs). The ICEEMDAN method is used to decompose the original signal into multiple intrinsic mode functions (IMFs). Let the original signal be x[n], where ω... (i) (i = 1, 2, ..., l) represents the Gaussian white noise added in the i-th group. In each iteration of IMF component extraction, the introduced noise signal is the IMF component obtained by decomposing the original Gaussian white noise signal, which is E. k (ω (i) ), (i = 1, 2, ..., l), operator E k (·) represents the extraction of the k-th IMF component from the Empirical Mode Decomposition (EMD) of the signal; the parameter β = ε / std(x) represents the multiplier coefficient of the noise component, which is defined as the ratio of the signal-to-noise ratio (SNR) between the added noise and the analyzed signal to the standard deviation of the analyzed signal, where std(·) is the standard deviation operator, whose mathematical definition is... In the formula, N is the number of sampling points; the operator M(·) characterizes the local mean of the signal, which is defined as the residual obtained by subtracting the IMF component from the original signal, i.e., satisfying M(·)=x[n]-E k Mathematical relationships (·); S201: Calculate "X" through l iterations of Empirical Mode Decomposition (EMD). 1i =x[n]+β0E1(ω (i) The local mean of [n], (i = 1, 2, ..., l) is denoted as M(X). 1i (i = 1, 2, ..., l), to obtain the first-order residual. Where β0=ε0 / std(x[n]), ε0 represents the signal-to-noise ratio (SNR) between the added noise and the analyzed signal, and std(·) is the standard deviation operator; E1(ω (i) [n]) is the first IMF component extracted by Empirical Mode Decomposition (EMD) of the Gaussian white noise added to the i-th group; S202: Calculate the first-order intrinsic mode components: S203: Estimate the second-order residual as "X" 2i =r1[n]+β1E2(ω (i) The local mean of [n], (i = 1, 2, ..., l) is denoted as M(X). 2i (i = 1, 2, ..., l), to obtain the second-order residual. Where β1=ε1 / std(r1[n]), ε1 represents the signal-to-noise ratio (SNR) between the added noise and the analyzed signal, and std(·) is the standard deviation operator; E2(ω (i) [n]) represents the second IMF component extracted from the empirical mode decomposition (EMD) of the Gaussian white noise added to the i-th group; solve for the second-order intrinsic mode components: S204: For the k-th order residual (k = 3, 4, ... K), calculate "M(r k-1 [n]+β k-1 E k (ω (i) [n])),(i=1,2,…l)”, where the coefficient β k-1 =ε k-1 / std(r k-1 ), ε k-1 The signal-to-noise ratio (SNR) between the added noise and the analyzed signal is represented by std(·), where std(·) is the standard deviation operator. S205: Calculate the kth eigenmode components: S206: Perform step S204 for subsequent k-value iterations until the final residual r is reached. k The termination conditions are met; Based on the structure of ICEEMDAN, the final residuals satisfy: Where K is the total number of IMFs, therefore, the signal x can be represented as:

7. The signal noise reduction method for acoustic monitoring of tool wear conditions in milling processes according to claims 3 and 4, characterized in that, Steps two and three specifically include: S211: The observed signal vector is x. If A represents the mixing matrix, the source signal vector s obtained according to the mixing model formula (3) is called the independent component (ICs). x = A·s (3) S212: Under the premise that the source signals are statistically independent and the mixing matrix A is invertible, the ICA algorithm estimates the mixing matrix A and the separation matrix W based on the observed mixed signal x, thereby reconstructing the data matrix y—that is, representing the mathematical expression of the estimated independent components (ICs)—and the relationship can be expressed as follows: y = W·x (4) S213: After calculating the fuzzy entropy (FuEn) value of each independent component (ICs), a threshold discrimination method is introduced to identify noise-related independent components. The discrimination criteria are as follows: Symbol here This represents the fuzzy entropy value of the g-th independent component (IC) arranged in ascending order, where g is the fuzzy entropy value that satisfies " "The minimum integer of the condition, where N is the total number of ICs. If there is no g value that satisfies the condition within the specified range, g = 1 is taken by default. When the g value satisfies equation (5), the corresponding IC is determined to be a noise component and is filtered by setting it to zero to obtain the denoised source signal vector y'. Then, the mixing matrix is ​​multiplied by the ICs matrix to obtain the denoised observation signal x', as shown in equation (3). Finally, the signal c' is reconstructed by superposition operation of the denoised observation signal. S214: ICA successfully separated several independent components (ICs). The number of ICs obtained by blind source separation is usually less than or equal to the number of input IMFs. S215: The noise-reduced signal exhibits significantly smoothed temporal characteristics while maintaining the overall trend of the original signal.

8. The signal noise reduction method for acoustic monitoring of tool wear conditions in milling processes according to claim 5, characterized in that, Constructing a CNN deep learning model: A deep learning model based on convolutional neural networks (CNN) is used for tool wear classification training. By adaptively extracting the latent patterns of input features, a nonlinear mapping relationship between features and wear state is established to realize tool wear state monitoring.

Images