Milling chatter online monitoring method and system based on stable mapping and deep learning
By combining stable mapping and deep learning, the problem of insufficient physical meaning in feature extraction and inadequate model generalization ability in milling chatter monitoring is solved, achieving high-precision chatter identification and monitoring with clear physical interpretation and strong adaptability to working conditions.
Patent Information
- Application Number
- CN202610709676.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-22
- Publication Date
- 2026-06-19
AI Technical Summary
Existing milling chatter monitoring methods lack physically meaningful feature extraction, have poor model generalization ability, use simplistic signal processing methods with severe noise interference, have limited model recognition capabilities, and are difficult to adapt to complex nonlinear relationships and time-series dependencies.
A milling dynamics model based on stability mapping was established, modal parameters were obtained by combining modal hammering experiments, a flutter stability lobe diagram was plotted, features were screened by stability margin, and signals were decomposed using EMD-correlation coefficient and VMD. A long short-term memory network model optimized by genetic algorithm was constructed for flutter identification.
It achieves high-precision and robust flutter monitoring, can capture long-term time-series dependent patterns, improves the physical interpretability and generalization ability of the model, has significant noise reduction effect, and high recognition accuracy.
Smart Images
Figure CN122241135A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of machining vibration monitoring technology, and in particular to an online monitoring method and system for milling chatter based on stability mapping and deep learning. Background Technology
[0002] Milling is one of the core machining methods in the field of mechanical manufacturing. During milling, chatter is highly likely to occur due to the interaction of dynamic milling forces between the tool and the workpiece. Chatter is a self-excited vibration; once it occurs, it can severely degrade the surface finish, accelerate tool wear, generate harsh noise, and even damage the machine tool spindle, seriously restricting machining efficiency and product accuracy. Therefore, accurate and online monitoring of milling chatter is of paramount engineering significance for timely adjustment of process parameters, suppression of chatter, and ensuring the stability of the machining process.
[0003] Existing chatter monitoring methods are mainly divided into two categories: those based on mechanistic models and those based on data. Mechanism-based methods, such as stability lobe diagrams, can provide theoretical guidance, but they rely on precise system parameters and are difficult to adapt to time-varying and nonlinear factors such as tool wear and material inhomogeneity in actual machining. Data-driven methods, especially machine learning methods, identify chatter by extracting and classifying features from vibration signals and other state information.
[0004] However, existing methods have the following shortcomings: (1) Feature extraction lacks physical meaning: Most methods rely on experience or pure statistical analysis to extract features. These features are not strongly related to the dynamic mechanism of chatter, resulting in poor model generalization ability and insufficient adaptability to new processing conditions. (2) Single signal processing method: For strong noise and non-stationary milling vibration signals, traditional signal processing methods (such as signal denoising methods that only use Empirical Mode Decomposition (EMD)) have problems such as mode mixing and endpoint effects, making it difficult to effectively extract pure chatter features. (3) Limited recognition model capability: Traditional machine learning models, such as Support Vector Machine (SVM), have limited learning ability for complex nonlinear relationships, while ordinary recurrent neural networks (RNNs) have gradient vanishing / exploding problems, making it difficult to effectively capture the long-term temporal dependence of chatter evolution.
[0005] Therefore, it is necessary to make improvements to address the problems existing in the current technology. Summary of the Invention
[0006] The purpose of this invention is to provide an online monitoring method for milling chatter based on stable mapping and deep learning, aiming to provide an engineering solution with clear physical meaning, high precision and high robustness to solve at least one of the problems mentioned in the background art; and also to provide an online monitoring system for milling chatter based on this online monitoring method.
[0007] To achieve the above objectives, this invention provides an online monitoring method for milling chatter based on stability mapping and deep learning, comprising the following steps: S1, establishing a milling dynamics model based on regenerative effect and solving the milling stability domain analytically; obtaining modal experimental parameters of the machine tool-tool system through modal hammering experiments; drawing a milling chatter stability lobe diagram based on the milling dynamics model and modal experimental parameters to obtain the milling chatter stability boundary; S2, conducting milling experiments and collecting original vibration signals under different experimental milling parameters; determining the stability boundary based on step S1. S2. Calculate the shortest signed distance from the parameter point of each experimental milling parameter to the milling chatter stability boundary, defining it as the stability margin of that experimental milling parameter; S3. Observe and analyze the samples obtained from the milling experiment in step S2, classifying the milling state into stable milling state, chatter incubation state, and chatter outbreak state, and assigning corresponding state label values to each experimental milling parameter; S4. Perform feature engineering processing on the original vibration signals collected in step S2 to obtain the experimental multidimensional chatter feature vector; establish a genetic algorithm optimized... A long short-term memory network model is used to train the experimental multidimensional chatter feature vector as input and the corresponding state label value obtained in step S3 as output, resulting in a trained chatter recognition model. In actual processing, the original vibration signal is acquired in real time, and feature engineering is performed on the real-time original vibration signal to obtain a real-time multidimensional chatter feature vector. The real-time multidimensional chatter feature vector is input into the chatter recognition model in step S4, and the real-time online monitoring result of milling chatter status is output. The feature engineering process involves using empirical mode decomposition to decompose the original vibration signal into multiple components, calculating the correlation coefficient between each component and the original vibration signal, selecting highly correlated components, and then performing signal superposition and reconstruction. The time-domain and frequency-domain features of the reconstructed signal are extracted, and combined with the stability margin calculated in step S2, the correlation between each feature and the stability margin is calculated using the Pearson correlation coefficient. Features with an absolute correlation coefficient greater than a preset threshold are selected as highly sensitive features. Variational mode decomposition is used to decompose the highly sensitive features, extracting feature components of different scales to construct a multidimensional chatter feature vector.
[0008] Furthermore, in step S1, the obtained modal experimental parameters include natural frequency, damping ratio, and modal stiffness.
[0009] Furthermore, in step S2, the experimental process of each milling experiment includes the tool entry process, the uneven milling process, the smooth milling process, and the tool retraction process, and the original vibration signal collected is taken from the smooth milling process.
[0010] Furthermore, in step S2, the formula for calculating the stability margin is:
[0011] ,
[0012] In the formula, d i C represents the stability margin of the i-th experimental milling parameter; C represents the milling chatter stability boundary curve; P i P represents the parameter point of the i-th experimental milling parameter; s Let be any point on curve C; The minimum Euclidean distance; the sign rule is that when the parameter point is within the stable region, d i <0, when the parameter point is within the flutter region. i >0, when the parameter point is on curve C. i =0.
[0013] Furthermore, the obtained stability margin is normalized using the following formula:
[0014] ,
[0015] In the formula, d i norm d represents the normalized stability margin of the milling parameters in the i-th experiment; max It represents the maximum absolute value of the stability margin across all parameter points.
[0016] Furthermore, the method for selecting highly sensitive features is to calculate F for each feature within the i-th experimental milling operation. k With normalized stability margin d i norm The Pearson correlation coefficient r between them k The calculation formula is as follows:
[0017]
[0018] In the formula, r k For the i-th experiment, the k-th feature within the milling process is related to d. i norm The Pearson correlation coefficient between them ranges from [-1, 1]; F k Let |r be the k-th extracted vibration signal feature; E() is the mathematical expectation operator; filter |r k Features with a value greater than 0.8 are considered highly sensitive features.
[0019] Furthermore, the highly sensitive features include: kurtosis, root mean square (RMS), mean square frequency, peak value, average value, minimum value, waveform factor, and centroid frequency of the X-axis time-domain signal; kurtosis, RMS, amplitude, peak factor, average value, and waveform factor of the X-axis frequency-domain signal; kurtosis, RMS, mean square frequency, amplitude, peak value, average value, waveform factor, relative power spectral entropy, centroid frequency, band energy, and frequency variance of the Y-axis time-domain signal; kurtosis, RMS, amplitude, average value, RMS value, waveform factor, and margin of the Z-axis time-domain signal; and kurtosis and minimum value of the Z-axis frequency-domain signal.
[0020] Furthermore, the variational mode decomposition is used to decompose the highly sensitive features and extract feature components at different scales. Specifically, a variational mode model is established, and for the original constrained variational problem, Lagrange multipliers and a quadratic penalty factor are introduced to transform it into an unconstrained optimization problem. The variational problem is solved using the alternating direction multiplier method to obtain K mode functions. Feature components are extracted from the K mode functions to form a multidimensional flutter feature vector.
[0021] Furthermore, in step S4, the genetic algorithm is used to optimize the hyperparameters of the Long Short-Term Memory network model. The optimized hyperparameters include the number of LSTM hidden layer units, the number of LSTM layers, the learning rate, the training batch size, and the number of training rounds.
[0022] This invention also provides an online monitoring system for milling chatter, which implements the aforementioned online monitoring method for milling chatter. The system includes signal acquisition hardware, comprising a triaxial piezoelectric accelerometer, a signal conditioner, and a data acquisition card. The triaxial piezoelectric accelerometer is installed on the machine tool spindle housing or tool holder to acquire raw vibration signals during the milling process. The signal conditioner and data acquisition card amplify, filter, and perform analog-to-digital conversion on the raw vibration signals. A PC is configured with online milling chatter monitoring software, developed based on the PyQt framework, which includes a data acquisition and control unit, a signal processing and feature extraction unit, a chatter identification and display unit, and an offline analysis subunit. The data acquisition and control unit configures data acquisition card parameters, controls the data acquisition process, displays and stores vibration signals in real time, and sets the data storage location. The signal processing and feature extraction unit performs the aforementioned feature engineering processing. The chatter identification and display unit receives real-time multidimensional chatter feature vectors and outputs the online monitoring results of the milling chatter state. The offline analysis subunit imports historical experimental data, re-sets parameters, trains the model, and compares performance.
[0023] The online monitoring method for milling chatter provided by this invention has the following advantages compared with the prior art:
[0024] (1) By using stability margin as the physical criterion for feature selection, it guides the data-driven model to focus on the sensitive features most related to flutter evolution, realizing the deep integration of prior knowledge and data learning, and enabling the model to have clear physical interpretability and stronger generalization ability under working conditions.
[0025] (2) The EMD-correlation coefficient method is used for adaptive noise reduction and reconstruction, which effectively removes noise and irrelevant components; then the VMD method is used to further decompose and obtain highly sensitive features, which overcomes the limitations of the single decomposition method and significantly improves the quality of feature extraction.
[0026] (3) Introduce a long short-term memory network (GA-LSTM) optimized by a genetic algorithm. The long short-term memory network is naturally suitable for processing time-series signals and can capture the long-term evolution of flutter from incubation to outbreak. The genetic algorithm is used to automatically search for the optimal hyperparameters of the long short-term memory network, avoiding the blindness of manual parameter tuning and ensuring the best performance of the model.
[0027] The present invention provides an online monitoring system for milling chatter, which integrates all functions from data acquisition, signal processing, feature extraction to model recognition and result visualization, forming a complete engineering application solution. Attached Figure Description
[0028] Figure 1 This is the overall flowchart of the online monitoring method for milling chatter of the present invention;
[0029] Figure 2 This is a flap diagram of milling chatter stability;
[0030] Figure 3 This is a schematic diagram of the boundary curve and parameter points for milling chatter stability when calculating the stability margin;
[0031] Figure 4 This is a schematic diagram illustrating the division of milling states;
[0032] Figure 5 This is a flowchart of signal denoising and reconstruction based on EMD-correlation coefficient;
[0033] Figure 6 This is a frequency domain analysis plot;
[0034] Figure 7 This is a flowchart illustrating the structure and optimization of the GA-LSTM model.
[0035] Figure 8 The software interface diagram of the online monitoring system for milling chatter of the present invention. Detailed Implementation
[0036] The embodiments of the present invention will be described in detail below.
[0037] This embodiment provides an online monitoring method for milling chatter based on stable mapping and deep learning, the process of which is as follows: Figure 1 As shown, it includes the following steps.
[0038] Step S1: Establish a milling dynamics model based on the regenerative effect and solve the milling stability domain analytically. The establishment of the milling dynamics model and the solution of the milling stability domain are existing technologies and will not be elaborated upon in this embodiment. Modal experimental parameters of the machine tool-tool system are obtained through modal impact experiments. This embodiment uses a VMC850LA CNC machining center as the object, and modal impact experiments are conducted using a modal hammer and an accelerometer to obtain the modal experimental parameters (including natural frequency, damping ratio, and modal stiffness) of the machine tool-tool system in the X and Y directions. The experiment measured the first-order natural frequency in the X direction to be 895.32Hz and in the Y direction to be 913.54Hz. The modal experimental parameters were then recorded. Based on the milling dynamics model and the modal experimental parameters, the following diagram is drawn: Figure 2 The milling chatter stability lobe diagram shown below obtains the milling chatter stability boundary. In the diagram, the horizontal axis is the spindle speed, the vertical axis is the axial depth of cut, the area below the curve is the stable milling zone, and the area above the curve is the chatter burst zone.
[0039] Step S2: Conduct milling experiments under different milling parameters and collect the original vibration signals under different experimental milling parameters. The workpiece material in the milling experiments is aluminum alloy, and the cutting tool is a three-flute carbide end mill. Multiple sets of spindle speeds (e.g., 2000-3500 rpm), feed rates (200-500 mm / min), and axial depths of cut (1-3.5 mm) are set. A triaxial accelerometer is attached to the machine tool spindle housing, and vibration signals are collected at a sampling frequency of 6400 Hz. At the same time, the surface morphology of the workpiece after each set of milling experiments is recorded. Based on the milling chatter stability lobe diagram determined in step S1, as shown... Figure 3 The parameter point P marked for each experimental milling parameter is shown. i (n) i ,a pi Then calculate the parameter point P. i The shortest signed distance to the milling chatter stability boundary curve is defined as the stability margin of the milling parameters in this experiment; the formula for calculating the stability margin is...
[0040] ,
[0041] In the formula, d i C represents the stability margin of the i-th experimental milling parameter; C represents the milling chatter stability boundary curve; P i P represents the parameter point of the i-th experimental milling parameter; s Let be any point on curve C; The minimum Euclidean distance to curve C; the sign rule is that when the parameter point is within the stable milling zone, d i <0, when the parameter point is within the flutter burst region. i >0, when the parameter point is on curve C. i =0.
[0042] Next, the obtained stability margin is normalized using the following formula:
[0043] ,
[0044] In the formula, d i norm d represents the normalized stability margin of the milling parameters in the i-th experiment; max It represents the maximum absolute value of the stability margin across all parameter points.
[0045] The stability margin quantifies the distance of parameter points from the theoretical stability boundary, thereby making each signal feature F k Both can be viewed as functions of stability margin, meaning that a mapping relationship is established between characteristic behavior and dynamically stable milling state.
[0046] ,
[0047] In the formula, F k f is the k-th signal feature; k This is a mapping function.
[0048] Preferably, in step S2, each milling experiment includes the tool entering the workpiece, an uneven milling process, a smooth milling process, and a tool retraction process. The collected raw vibration signal is taken from the smooth milling process. Throughout each milling experiment, during the processes of tool not participating in machining, just entering the workpiece, uneven milling, and tool retraction, the vibration signal is significantly affected by changes in the machine tool's motion state and contact transitions, resulting in obvious signal fluctuations, poor regularity, and difficulty in reflecting the true machining state. In contrast, when the experiment enters the smooth milling process, the tool and workpiece are in continuous contact, the system vibration response is relatively stable, and the signal characteristics can more objectively characterize the dynamic characteristics of the machining process. Therefore, in this embodiment, when processing the data, it is necessary to truncate the raw vibration signal, preferably only truncate the effective data from the smooth milling process, and use this portion of the signal as the basis for subsequent signal analysis and feature extraction.
[0049] Step S3: Observe and analyze the samples obtained from the milling experiment in Step S2, classifying the milling state into stable milling state, chatter incubation state, and chatter outbreak state, and assigning corresponding state label values to each experimental milling parameter. This step requires manual labeling by staff based on experience and the surface morphology of the workpiece after the milling experiment. Figure 4 As shown, based on the characteristics of the vibration marks, the milling state is divided into three categories: (1) Stable milling state: the surface is smooth, there are no vibration marks, the spectrum energy is evenly distributed, and the state label value is 0; (2) Flutter incubation state: slight and discontinuous strip vibration marks appear, and abnormal frequency peaks of non-frequency conversion and harmonics appear in the spectrum, and the state label value is 1; (3) Flutter burst state: obvious, large-area messy and irregular vibration marks appear, and the flutter frequency energy in the spectrum is significantly concentrated, and the state label value is 2.
[0050] Step S4: Perform feature engineering processing on the original vibration signal acquired in step S2. Feature engineering processing includes signal denoising and reconstruction based on EMD-correlation coefficient and feature decomposition and vector construction based on VMD.
[0051] The signal denoising and reconstruction process based on EMD-correlation coefficient is as follows: Figure 5 As shown: Empirical Mode Decomposition (EMD) is used to decompose the original vibration signal into multiple components. The correlation coefficient between each component and the original vibration signal is calculated, and highly correlated components are selected for signal superposition and reconstruction. Typically, the first few high-frequency components contain flutter information and noise, while the last low-frequency components and the remainder are trend terms. In this embodiment, IMF1 and IMF2 with correlation coefficients greater than 0.1 are selected for superposition and reconstruction to obtain the denoised signal. Figure 6 The frequency domain analysis plot shown indicates that the reconstructed signal effectively suppresses low-frequency interference and random noise, highlighting the flutter characteristic frequency (approximately 870-900Hz). The time-domain and frequency-domain features of the reconstructed signal are extracted, and combined with the stability margin calculated in step S2, the correlation between each feature and the stability margin is calculated using the Pearson correlation coefficient. Specifically, the correlation F for each feature within the i-th experimental milling step is calculated. k With normalized stability margin d i norm The Pearson correlation coefficient r between them k The calculation formula is as follows:
[0052] ,
[0053] In the formula, r k For the i-th experiment, the k-th feature within the milling process is related to d. i norm The Pearson correlation coefficient between them ranges from [-1, 1]; F krepresents the k-th extracted vibration signal feature (such as kurtosis, root mean square, etc.); E() is the mathematical expectation operator.
[0054] Features with correlation coefficients whose absolute values are greater than a preset threshold (preferably 0.8) are selected as high-sensitivity features. In this embodiment, the coefficients of high-sensitivity features that meet the preset threshold are shown in Table 1 below. Thus, high-sensitivity features include: kurtosis, root mean square (RMS), mean square frequency, peak value, average value, minimum value, waveform factor, and centroid frequency of the X-axis time-domain signal; kurtosis, RMS, amplitude, peak factor, average value, and waveform factor of the X-axis frequency-domain signal; kurtosis, RMS, mean square frequency, amplitude, peak value, average value, waveform factor, relative power spectral entropy, centroid frequency, band energy, and frequency variance of the Y-axis time-domain signal; kurtosis, RMS, amplitude, average value, RMS value, waveform factor, and margin of the Z-axis time-domain signal; and kurtosis and minimum value of the Z-axis frequency-domain signal.
[0055] Table 1 Correlation Coefficients of Highly Sensitive Features
[0056]
[0057] The process of feature decomposition and vector construction based on VMD is as follows: Highly sensitive features are decomposed using variational mode decomposition, feature components at different scales are extracted, and multidimensional flutter feature vectors are constructed. Specifically, a variational mode model is first established, and then the constructed variational problem is solved.
[0058] ,
[0059] In the formula, {u k Let} be the set of K band-limited eigenmode functions obtained from the decomposition, u k Represents the k-th modal function; {ω k} represents the set of center frequencies corresponding to each modal function; To minimize the objective function; K is the number of decomposition levels; Let be the instantaneous value of the k-th mode function at time t; The center frequency of the k-th modal function; It is the first-order partial derivative with respect to time t; Let be the impact function; The imaginary unit; It is a time variable; For exponential modulation terms; These are constraints; π represents the original signal to be decomposed (the filtered high-sensitivity feature signal); π is the constant of pi; * is the convolution operator.
[0060] For the original constrained variational problem, we introduce Lagrange multipliers and a quadratic penalty factor to transform it into an unconstrained optimization problem, as shown in the following equation:
[0061] ,
[0062] In the formula, To augment the Lagrange function; It is a secondary penalty factor; For Lagrange multipliers; Norm operations; This is for inner product operations.
[0063] The above variational problem is solved using the alternating direction multiplier method, with the following steps:
[0064] initialization ; Execute loop ;
[0065] For all Its calculation formula is,
[0066]
[0067] In the formula, This represents the update value of the k-th mode function in the frequency ω domain during the (n+1)-th iteration; This is the Fourier transform of the original signal x(t); Let be the Fourier transform of the i-th mode function; The Fourier transform of the Lagrange multipliers in the nth iteration; For frequency variables; It represents the center frequency of the k-th mode in the n-th iteration.
[0068] renew Its calculation formula is,
[0069]
[0070] In the formula, The center frequency of the k-th mode in the (n+1)-th iteration; for The model; The integral is the center frequency from 0 to infinity; It is the differential element of the center frequency.
[0071] renew Its calculation formula is,
[0072]
[0073] In the formula, This represents the Fourier transform of the Lagrange multipliers in the (n+1)th iteration; To update the step size parameter, it is usually set to 10. -3 Order of magnitude.
[0074] K modal functions are obtained, and feature components are extracted from the K modal functions to form a multidimensional flutter feature vector.
[0075] A long short-term memory (LSTM) network model optimized by a genetic algorithm is established. The experimental multidimensional tremor feature vector is used as input, and the corresponding state label value obtained in step S3 is used as output. The model is then trained to obtain a trained tremor recognition model. In this embodiment, the hyperparameters of the LSTM model (Long Short-Term Memory Network Model) are encoded using a genetic algorithm; each set of parameter configurations corresponds to an individual LSTM model; the corresponding LSTM model is trained on the training set, and the recognition accuracy is calculated on the validation set; the recognition accuracy on the validation set is used as the fitness function; the parameter combination is continuously optimized through selection, crossover, and mutation operations of the GA (Genetic Algorithm); finally, the LSTM model with the optimal parameter configuration is output for tremor recognition. This method transforms the model parameter search process into a global optimization problem, effectively avoiding the randomness and local optima problems of manual parameter tuning. The model construction and training steps are as follows: Figure 7 As shown.
[0076] The optimization method is as follows: The number of hidden LSTM units, the number of LSTM layers, the learning rate, the batch size, and the number of training epochs are selected as the optimization objects for GA, and these parameters are represented using real-number encoding. The fitness function is selected based on the model's recognition accuracy on the validation set; the higher the accuracy, the greater the individual fitness, thus guiding the genetic algorithm to search for parameter regions with better performance. Encoding: The number of hidden LSTM units (range [32, 256]), the learning rate (range [0.0001, 0.01]), and the batch size ([32, 128]) are encoded as real-number vectors. Fitness function: The fitness value is the model's tremor recognition accuracy on the validation set. Evolutionary operations: Tournament selection, simulated binary crossover, and polynomial mutation are used to generate a new generation of the population. After 100 generations of evolution, the optimal hyperparameters found by GA are: 128 hidden LSTM units, a learning rate of 0.001, and a batch size of 64. The LSTM model with these optimal hyperparameters is trained using the training set (accounting for 80% of the total data). After training, the model was evaluated on the test set (accounting for 20% of the total data). The results are shown in Table 2 below. The VMD-GA-LSTM model proposed in this embodiment achieved an overall recognition accuracy of 98.94% for the three milling states (stable milling, chatter incubation, and chatter burst), which is significantly higher than the unoptimized LSTM model (87.23%) and SVM model (87.23%). Even the GA-LSTM model (95.74%) has a high accuracy.
[0077] Table 2 Evaluation index results for different models
[0078]
[0079] Step S5: In actual machining, the original vibration signal is acquired in real time, and feature engineering processing is performed on the real-time original vibration signal (this feature engineering processing is consistent with the feature engineering processing method in step S4) to obtain the real-time multidimensional chatter feature vector; the real-time multidimensional chatter feature vector is input into the chatter recognition model in step S4, and the real-time online monitoring result of milling chatter status is output.
[0080] Preferably, to improve the robustness of monitoring, a sliding window statistical decision is performed on the output of the flutter identification model: a time window length T is set, the number of times a, b, c of each state of the model output within the time window is counted, and their proportions are calculated. , , Where a, b, and c represent the number of times the stable milling state, chatter incubation state, and chatter burst state occur within the time length, respectively; p1, p2, and p3 represent the proportion values of the stable milling state, chatter incubation state, and chatter burst state within the time length, respectively; if the proportion of a certain state is greater than a preset threshold (such as 0.7), then the state within that time window is determined to be this state.
[0081] In summary, the online monitoring method for milling chatter of the present invention has the following advantages: (1) It uses stability margin as the physical criterion for feature selection, guides the data-driven model to focus on the sensitive features most related to chatter evolution, realizes the deep integration of prior knowledge and data learning, and enables the model to have clear physical interpretability and stronger generalization ability of working conditions; (2) It adopts the EMD-correlation coefficient method for adaptive noise reduction and reconstruction, effectively removing noise and irrelevant components; and then uses the VMD method to further decompose and obtain highly sensitive features, overcoming the limitations of the single decomposition method and significantly improving the quality of feature extraction; (3) It introduces a long short-term memory network (GA-LSTM) optimized by a genetic algorithm. The long short-term memory network is naturally suitable for processing time-series signals and can capture the long-term evolution law of chatter from incubation to outbreak; the genetic algorithm is used to automatically search for the optimal hyperparameters of the long short-term memory network, avoiding the blindness of manual parameter tuning and ensuring the best performance of the model.
[0082] This invention also provides an online monitoring system for milling chatter, such as... Figure 7 As shown, the method for implementing the above-mentioned online monitoring method for milling chatter includes:
[0083] The triaxial piezoelectric accelerometer is mounted on the machine tool spindle housing via a magnetic mount and is used to sense vibrations in the X, Y, and Z directions.
[0084] The signal conditioner provides constant current source excitation and signal amplification and filtering.
[0085] The high-precision data acquisition card with a USB interface (NI USB-6001) is responsible for converting the analog voltage signal from the triaxial piezoelectric accelerometer into a digital signal.
[0086] A PC with a Windows operating system and online milling chatter monitoring software installed. The online milling chatter monitoring software is developed based on the PyQt5 framework, and its interface is as follows: Figure 8 As shown, it mainly comprises three units: a data acquisition and control unit, a signal processing and feature extraction unit, and a chatter recognition and display unit. The data acquisition and control unit is used to configure data acquisition card parameters, control the data acquisition process, display and store vibration signals in real time, and set the data storage location. Users can configure the sampling rate, sampling channels, and data storage path in this unit, and control the start and stop of acquisition. The right side of the interface dynamically displays the vibration signal waveforms of the X, Y, and Z axes in real time. The signal processing and feature extraction unit is used to implement the feature engineering processing. This unit runs in the background and integrates algorithms such as EMD, VMD, and correlation analysis. It automatically processes the acquired data and generates feature vectors that meet the model input requirements. The software also supports exporting the raw data and intermediate processing results for offline analysis. The chatter recognition and display unit receives real-time multi-dimensional chatter feature vectors and outputs the online monitoring results of milling chatter status. It encapsulates a pre-trained chatter recognition model. This unit receives feature vectors and performs real-time inference. It may also include an offline analysis sub-unit: this unit is a separate tab that allows engineers to import historical experimental data, reset parameters, train models, and compare performance, facilitating continuous optimization and evaluation of the flutter recognition model.
[0087] Based on the above system settings, during the milling process, once the system is started, it automatically completes the entire process from signal acquisition to status display. Experimental verification shows that the system has high accuracy and fast response speed in online identification of milling chatter, providing operators with timely warnings, effectively guiding the adjustment of process parameters, and preventing chatter from occurring.
[0088] In summary, the online monitoring method and system for milling chatter based on stable mapping and deep learning provided by this invention has clear physical meaning, high precision and high robustness, and is a better engineering solution.
[0089] Where there is no conflict, the above embodiments and features can be combined with each other.
[0090] Finally, it should be noted that the above embodiments are only used to illustrate the preferred technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the present invention.
Claims
1. A method for online monitoring of milling chatter based on stable mapping and deep learning, characterized in that, Includes the following steps: S1. Establish a milling dynamics model based on the regeneration effect and solve the milling stability domain analytically; obtain the modal experimental parameters of the machine tool-tool system through modal hammering experiments; based on the milling dynamics model and modal experimental parameters, draw the milling chatter stability lobe diagram to obtain the milling chatter stability boundary. S2. Conduct milling experiments and collect raw vibration signals under different experimental milling parameters; based on the milling chatter stability lobe diagram determined in step S1, calculate the shortest signed distance from the parameter point of each experimental milling parameter to the milling chatter stability boundary, which is defined as the stability margin of the experimental milling parameter. S3. Observe and analyze the samples obtained from the milling experiment in step S2, divide the milling state into stable milling state, chatter incubation state and chatter burst state, and assign corresponding state label values to each experimental milling parameter. S4. Perform feature engineering processing on the raw vibration signals acquired in step S2 to obtain the experimental multidimensional flutter feature vector; establish The long short-term memory network model optimized by the genetic algorithm takes the experimental multidimensional flutter feature vector as input and the corresponding state label value obtained in step S3 as output to train the model, thus obtaining a trained flutter recognition model. S5. In actual processing, the original vibration signal is acquired in real time, and the real-time original vibration signal is processed by feature engineering to obtain the real-time multidimensional chatter feature vector; the real-time multidimensional chatter feature vector is input into the chatter recognition model in step S4, and the real-time online monitoring result of milling chatter status is output. The feature engineering process involves decomposing the original vibration signal using empirical mode decomposition to obtain multiple components, calculating the correlation coefficient between each component and the original vibration signal, selecting highly correlated components, and then performing signal superposition and reconstruction. The time-domain and frequency-domain features of the reconstructed signal are extracted, and combined with the stability margin calculated in step S2, the correlation between each feature and the stability margin is calculated using the Pearson correlation coefficient. Features with an absolute value of correlation coefficient greater than a preset threshold are selected as highly sensitive features. Variational mode decomposition is used to decompose highly sensitive features, extract feature components at different scales, and construct a multidimensional flutter feature vector.
2. The online monitoring method for milling chatter according to claim 1, characterized in that: In step S1, the obtained modal experimental parameters include natural frequency, damping ratio, and modal stiffness.
3. The online monitoring method for milling chatter according to claim 1, characterized in that: In step S2, the experimental process of each milling experiment includes the tool entry process, the uneven milling process, the smooth milling process, and the tool retraction process. The original vibration signal collected is taken from the smooth milling process.
4. The online monitoring method for milling chatter according to claim 1, characterized in that: In step S2, the formula for calculating the stability margin is: ， In the formula, d i C represents the stability margin of the i-th experimental milling parameter; C represents the milling chatter stability boundary curve; P i P represents the parameter point of the i-th experimental milling parameter; s Let be any point on curve C; The minimum Euclidean distance; the sign rule is that when the parameter point is within the stable region, d i <0, when the parameter point is within the flutter region. i >0, when the parameter point is on curve C. i =0.
5. The online monitoring method for milling chatter according to claim 4, characterized in that: The obtained stability margin is normalized using the following formula: ， In the formula, d i norm d represents the normalized stability margin of the milling parameters in the i-th experiment; max It represents the maximum absolute value of the stability margin across all parameter points.
6. The online monitoring method for milling chatter according to claim 5, characterized in that: The screening method for highly sensitive features is as follows: Calculate F for each feature within the i-th experimental milling operation. k With normalized stability margin d i norm The Pearson correlation coefficient r between them k The calculation formula is as follows: ， In the formula, r k For the i-th experiment, the k-th feature within the milling process is related to d. i norm The Pearson correlation coefficient between them ranges from [-1, 1]; F k Let be the k-th extracted vibration signal feature; E() is the mathematical expectation operator; Filter | r k Features with a value greater than 0.8 are considered highly sensitive features.
7. The online monitoring method for milling chatter according to claim 6, characterized in that: High sensitivity features include, The X-axis time-domain signal contains kurtosis, root mean square (RMS), mean square frequency, peak value, average value, minimum value, waveform factor, and centroid frequency; the X-axis frequency-domain signal contains kurtosis, RMS, amplitude, peak factor, average value, and waveform factor; the Y-axis time-domain signal contains kurtosis, RMS, mean square frequency, amplitude, peak value, average value, waveform factor, relative power spectral entropy, centroid frequency, band energy, and frequency variance; the Z-axis time-domain signal contains kurtosis, RMS, amplitude, average value, RMS value, waveform factor, and margin; and the Z-axis frequency-domain signal contains kurtosis and minimum value.
8. The online monitoring method for milling chatter according to claim 1, characterized in that: The method employs variational mode decomposition to decompose highly sensitive features and extract feature components at different scales. Specifically, the method is as follows: A variational modal model is established. For the original constrained variational problem, Lagrange multipliers and a quadratic penalty factor are introduced to transform it into an unconstrained optimization problem. The variational problem is solved using the alternating direction multiplier method to obtain K modal functions. Feature components are extracted from the K modal functions to form a multidimensional flutter feature vector.
9. The online monitoring method for milling chatter according to claim 1, characterized in that: In step S4, the genetic algorithm is used to optimize the hyperparameters of the Long Short-Term Memory network model. The optimized hyperparameters include the number of LSTM hidden layer units, the number of LSTM layers, the learning rate, the training batch size, and the number of training rounds.
10. An online monitoring system for milling chatter, characterized in that: The method for implementing the online monitoring method for milling chatter as described in any one of claims 1 to 9 includes, The signal acquisition hardware includes a triaxial piezoelectric accelerometer, a signal conditioner, and a data acquisition card. The triaxial piezoelectric accelerometer is installed on the machine tool spindle housing or tool holder to acquire the raw vibration signal during the milling process. The signal conditioner and data acquisition card are used to amplify, filter, and convert the raw vibration signal to digital. A PC equipped with online milling chatter monitoring software, developed based on the PyQt framework, includes a data acquisition and control unit, a signal processing and feature extraction unit, a chatter identification and display unit, and an offline analysis subunit. The data acquisition and control unit is used to configure data acquisition card parameters, control the data acquisition process, display and store vibration signals in real time, and set the data storage location. The signal processing and feature extraction unit is used to implement the feature engineering processing. The chatter identification and display unit is used to receive real-time multi-dimensional chatter feature vectors and output the online monitoring results of milling chatter status. The offline analysis subunit is used to import historical experimental data, reset parameters, train models, and compare performance.