A cutting chatter recognition method based on optimized VMD and CNN-ICBAM-MHSA-BiLSTM

Abstract

This invention provides a cutting chatter identification method based on optimized VMD and CNN-ICBAM-MHSA-BiLSTM, belonging to the field of machine tool cutting monitoring technology. It solves the technical problem in existing cutting chatter identification methods that rely on manual adjustment of VMD parameters and have low recognition rates for minor chatter. The method includes: multi-sensor signal acquisition, AOA-VMD adaptive decomposition and IMFs screening and reconstruction, multi-attention fusion network training, and chatter condition identification. Specifically, it combines the arithmetic optimization algorithm AOA with variational mode decomposition VMD to achieve adaptive parameter optimization; designs an improved convolutional block attention mechanism ICBAM, enhancing spatial feature extraction through multi-scale convolutional kernels; and constructs a CNN-ICBAM-MHSA-BiLSTM fusion network to integrate local features, temporal dependencies, and key time step attention. Experimental verification shows that this invention can accurately identify three conditions: stable machining, minor chatter, and severe chatter, with an overall recognition accuracy ≥98.40%, a minor chatter recognition accuracy improvement of ≥3.9%, and strong anti-noise interference capability, providing reliable technical support for chatter identification in machine tool cutting.

Description

Technical Field

[0001] This invention relates to the field of machine tool cutting monitoring technology, and in particular to a cutting chatter identification method based on optimized VMD and CNN-ICBAM-MHSA-BiLSTM. Background Technology

[0002] During machine tool cutting, chatter, as a self-excited vibration between the tool and the workpiece, can lead to a decrease in workpiece surface quality, increased tool wear, and reduced production efficiency. In severe cases, it can even damage the machine tool spindle. Therefore, chatter detection is a key technology to ensure machining quality.

[0003] Existing flutter recognition methods suffer from the following drawbacks: First, Variational Mode Decomposition (VMD), a commonly used signal preprocessing method, requires manual adjustment of the number of modes and penalty factor. Under different processing conditions, mode aliasing or loss of effective features can easily occur, affecting subsequent recognition accuracy. Second, traditional deep learning models such as CNN and LSTM can only capture single-scale features, making it difficult to adapt to local details, mid-range correlations, and global contextual information in multi-sensor signals, resulting in a low recognition rate for weak feature signals such as slight flutter. Third, existing attention mechanisms mostly use single-scale convolutional kernels, which cannot fully explore the complex correlations between multi-source signals and have insufficient resistance to noise interference, leading to a decrease in recognition robustness when the processing environment changes.

[0004] Although existing studies have attempted to improve model performance by using gray wolf optimization, genetic algorithms to optimize VMD parameters, or by introducing attention mechanisms, problems still exist, such as slow convergence speed of optimization algorithms, incomplete capture of multi-scale features by attention modules, and high false positive rate of slight flutter, making it difficult to meet the high-precision and high-robust flutter recognition requirements in industrial settings.

[0005] Therefore, developing a cutting chatter identification method with adaptive parameter optimization, multi-scale feature fusion, and high recognition accuracy is of great practical significance. Summary of the Invention

[0006] This invention proposes a cutting chatter recognition method based on optimized VMD and CNN-ICBAM-MHSA-BiLSTM to solve the problems of difficult chatter recognition in existing technologies, especially slight chatter, and the problems of VMD parameters relying on manual adjustment and insufficient multi-scale feature capture.

[0007] This invention provides a cutting chatter identification method based on optimized VMD and CNN-ICBAM-MHSA-BiLSTM, characterized in that the identification method includes the following steps:

[0008] Step S1: Acquire vibration acceleration signals and cutting force signals during the cutting process;

[0009] Step S2: The Arithmetic Optimization Algorithm (AOA) is used to optimize the parameters of Variational Mode Decomposition (VMD). The optimized VMD is then used to adaptively decompose and reconstruct the acquired signal to obtain the reconstructed signal and construct a sample dataset.

[0010] Step S3: Construct a fusion network that includes a convolutional neural network (CNN), an improved convolutional block attention module (ICBAM), a bidirectional long short-term memory network (BiLSTM), and a multi-head self-attention mechanism (MHSA). Input the reconstructed signal into the fusion network for iterative optimization training to obtain the flutter recognition model.

[0011] Step S4: Input the signal to be identified into the trained chatter recognition model and output the working condition recognition result of cutting chatter.

[0012] Furthermore, the specific method for acquiring vibration acceleration signals and cutting force signals during the cutting process in step S1 includes:

[0013] Data is collected from three components of the machine tool using multiple sensors: the spindle, the worktable, and the leadscrew.

[0014] The spindle signals are vibration acceleration signals in the x and y directions and radial cutting force signals.

[0015] The signal types of the worktable are vibration acceleration signals in the x and y directions and cutting force signals in the feed direction;

[0016] The signal types of the lead screw shaft are x-direction vibration acceleration signal and axial cutting force signal;

[0017] A total of 8 sets of data were collected, with each set having a sampling duration of 10 seconds and a sampling frequency of 20kHz. Each set of signals had 9000 sampling points and covered three operating conditions: stable processing, slight chatter, and severe chatter.

[0018] Furthermore, the specific method for optimizing the parameters of the variational mode decomposition (VMD) using the arithmetic AOA optimization algorithm in step S2, adaptively decomposing and reconstructing the acquired signal using the optimized VMD to obtain the reconstructed signal, and constructing the sample dataset includes:

[0019] S2-1: Construct an arithmetic optimization algorithm (AOA) with the small normalized sample entropy as the fitness function to optimize the number of modes in variational mode decomposition (VMD). and penalty factor ;

[0020] Arithmetic Optimization Algorithm (AOA) parameter settings: population size 20, number of iterations 15. The optimization interval is [3,10]. The optimization range is [200, 3500].

[0021] Normalized minimum sample entropy fitness function formula:

[0022]

[0023] In the formula: For embedded dimensions, This is the tolerance threshold;

[0024] S2-2: The optimized version and Substituting the values ​​into Variational Mode Decomposition (VMD), the multi-sensor signals are decomposed into... Individual intrinsic mode functions (IMFs);

[0025] S2-3: IMFs with natural frequencies close to those of machine tool components are selected by wavelet packet energy entropy and wavelet energy entropy, where wavelet energy entropy represents the correlation between the reconstructed signal and the original signal structure;

[0026] Since the flutter frequency is close to the natural frequency, severe flutter may cause a flutter frequency shift, requiring the selection of other frequency bands;

[0027] The wavelet packet energy entropy threshold is set to 1.2, and IMFs with a threshold less than the threshold are selected because a small wavelet packet energy entropy indicates strong signal structure and less noise.

[0028] IMFs with frequency bands similar to the inherent frequencies of the acquisition components were selected: first-order 48Hz and second-order 519Hz for the spindle; first-order 399Hz and second-order 1200Hz for the worktable; and first-order 43Hz and second-order 650Hz for the lead screw.

[0029] The filtered IMFs are reconstructed to obtain a preprocessed signal dataset, which is then divided into a training set and a test set in a 7:3 ratio.

[0030] Furthermore, the specific method for constructing a fusion network comprising a Convolutional Neural Network (CNN), an Improved Convolutional Block Attention Module (ICBAM), a Bidirectional Long Short-Term Memory Network (BiLSTM), and a Multi-Head Self-Attention Mechanism (MHSA) in step S3, and inputting the reconstructed signal into the fusion network for iterative optimization training to obtain the tremor recognition model includes:

[0031] S3-1: Network Architecture Design: Construct a fusion network of CNN-ICBAM-MHSA-BiLSTM, including the following modules:

[0032] Convolutional Neural Network (CNN): 2 convolutional layers, convolutional kernels [2,1], output channels 32, 64, ReLU activation layer, sequence defolding layer, and tiling layer, used to extract local sensitive features;

[0033] Improved Convolutional Block Attention Module (ICBAM): Channel Attention + Multi-Scale Spatial Attention. Channel attention generates channel weights through GAP / GMP dual-path + fully connected layer. Spatial attention is processed in parallel using 3×1, 5×1, and 7×1 convolutional kernels. After concatenation, spatial weights are generated by 1×1 convolution to reduce dimensionality and enhance key features to suppress noise.

[0034] Multi-head self-attention mechanism (MHSA): 4 heads, giving higher weight to key time steps to highlight the weak characteristics of slight flutter;

[0035] Bidirectional Long Short-Term Memory Network (BiLSTM): 6-dimensional hidden layers, output mode last, captures bidirectional temporal dependencies of signals;

[0036] Classification layer: Fully connected layer (output channel 3) + Softmax activation layer, outputting the probabilities of three operating conditions;

[0037] S3-2: Model Training: The model uses the cross-entropy loss function, Adam optimizer, learning rate 0.001, batch size 128, number of iterations 50, number of rounds 300, parameters are updated through backpropagation, and the optimal model on the training set is saved.

[0038] Furthermore, the improved convolutional block attention module (ICBAM) specifically includes:

[0039] First, the multi-convolutional block attention mechanism CBAM is used, which includes a dual attention module of "channel + space";

[0040] Channel attention is achieved by fusing feature weights from different channels using Global Average Pooling (GAP) and Global Max Pooling (GMP).

[0041]

[0042]

[0043] The results of average pooling and max pooling are fed into the shared fully connected layers FC1 and FC2 to generate channel attention maps. :

[0044]

[0045] in, It is the ReLU activation function. and These are the weights of the fully connected layer. It is the Sigmoid activation function;

[0046] Finally, the channel attention map Applied to input feature map The above yields the channel-weighted feature map. :

[0047]

[0048] Spatial attention is:

[0049] Channel-weighted feature map Perform channel summation to obtain the spatial features of each pixel location:

[0050]

[0051]

[0052] By extracting the weights of key regions in the spatial dimension through convolution, the spatial distribution of flutter features is enhanced, and irregular noise regions are suppressed.

[0053] Improvements were made to the original CBAM, in a single 3 Based on the convolution block of 1, add 5 1, 7 1 convolutional block, and 3 Parallel processing of convolutional blocks and feature fusion of 1 are used to construct multi-scale spatial attention, which more accurately adapts to the feature distribution of cutting chatter signals:

[0054]

[0055] The three convolution results are concatenated to obtain the concatenated feature map. :

[0056]

[0057] Use a 1×1 convolution kernel to concatenate the feature maps Perform convolution to generate a spatial attention map. :

[0058]

[0059] Finally, the spatial attention map Applied to the channel-weighted feature map The final ICBAM feature map is obtained from the above.

[0060] .

[0061] Furthermore, the Bidirectional Long Short-Term Memory (BiLSTM) network specifically includes:

[0062] First, the input matrix is ​​transformed through three different linear transformations, namely the weight matrices. Convert to query ,key Sum :

[0063]

[0064] in, , , These are trainable weight matrices used to generate queries, keys, and values, respectively.

[0065] The self-attention of each head is calculated as follows:

[0066]

[0067] in, It is the dot product of the query and the key. It is a scaling factor to prevent the gradient from vanishing due to an excessively large dot product.

[0068] Then, the scores are normalized using a softbox operation to obtain the degree of attention each position gives to other positions;

[0069] In MHSA, four heads are used to calculate different attention scores. The outputs of all heads are concatenated together to obtain the final attention output.

[0070]

[0071] in, It is the output linear transformation matrix. The outputs of each head are concatenated to obtain the final multi-head attention result.

[0072] Compared with the prior art, the present invention has the following advantages:

[0073] 1. Adaptive parameter optimization: AOA optimizes VMD parameters, avoiding the subjectivity of manual adjustment, and has a fast convergence speed (7 generations of iterative convergence), ensuring the effectiveness of signal decomposition under different processing conditions.

[0074] 2. Multi-scale feature capture: The ICBAM module fuses local and global features through multi-scale convolutional kernels, and MHSA enhances attention at key time steps, improving the accuracy of slight tremor recognition by ≥3.9% and the recall rate by ≥2.4%.

[0075] 3. High recognition accuracy and robustness: The fusion network integrates local features, temporal dependencies and key attention, with an overall recognition accuracy of ≥98.40%, which is ≥23.8% higher than the traditional CNN-BiLSTM, and the ability to resist noise interference is significantly enhanced.

[0076] Based on the implementation methods provided in the above aspects, this application can be further combined to provide more implementation methods. Attached Figure Description

[0077] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein:

[0078] Figure 1 This is a schematic diagram of the cutting chatter recognition system of the present invention;

[0079] Figure 2 This is the spectrum of the analog signal after AOA-VMD processing;

[0080] Figure 3 To improve the schematic diagram of the CBAM (ICBAM) module structure;

[0081] Figure 4 This is a picture of the cutting test site;

[0082] Figure 5 This is a schematic diagram illustrating the partitioning of a dataset based on workpiece surface quality.

[0083] Figure 6 A schematic diagram of the fitness curve for optimizing VMD for AOA;

[0084] Figure 7 The AOA-VMD decomposition IMF time-frequency plot of the vibration signal in the x-direction of the main axis;

[0085] Figure 8 A schematic diagram of the IMF component spectrum of the vibration signal in the x-direction of the main axis;

[0086] Figure 9 This is a time-domain comparison diagram of the reconstructed signal and the original signal;

[0087] Figure 10 shows a comparison of the confusion matrices of each model;

[0088] Figure 11 Comparison of slight flutter recognition performance of different models;

[0089] Figure 12 A comparison chart showing the effectiveness of different models in identifying severe flutter. Detailed Implementation

[0090] The exemplary embodiments disclosed in this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0091] To make the objectives, technical solutions, and effects of this invention clearer, the following detailed description is provided in conjunction with specific experimental apparatus and procedures:

[0092] I. Experimental Setup

[0093] 1. Machine tool: Vertical machining center (model: Mike CNC);

[0094] 2. Workpiece: Aluminum alloy A16061;

[0095] 3. Cutting tool: 4mm diameter 3-flute flat end mill;

[0096] 4. Sensors: Vibration acceleration sensor (CL-YD-3210), triaxial force sensor, data acquisition card (DH5996);

[0097] 5. Computing equipment: Windows 11 operating system, Intel Core i5-13600KF CPU, NVIDIA GeForce RTX 4070 Super graphics card, 32GB RAM, MATLAB 2024b simulation environment.

[0098] I. Specific Implementation Steps

[0099] 1. Sensor installation and signal acquisition:

[0100] One vibration acceleration sensor is installed in each of the x and y directions on the spindle housing, and a cutting force sensor is installed radially.

[0101] One vibration acceleration sensor is installed on the worktable in the x and y directions, and a cutting force sensor is installed in the feed direction;

[0102] One vibration acceleration sensor is installed in the x-direction of the lead screw axis, and a cutting force sensor is installed in the axial direction.

[0103] Set machining parameters: speed 18000rpm / 21000rpm, feed per tooth 0.1mm and 0.14mm, depth of cut 0.15mm-0.62mm, machining for 10s per set of parameters, collecting 8 sets of data, covering three working conditions.

[0104] 1. AOA-VMD parameter optimization and signal preprocessing:

[0105] Initialize AOA parameters: population size 20, number of iterations 15. , ;

[0106] Using the minimum normalized sample entropy as the fitness function, the optimal parameters are obtained through iterative optimization: the number of decomposition layers of the vibration signal in the principal axis x-direction. =8, penalty factor =2783; vibration signal in the y-direction of the main shaft =6, penalty factor =2574; Spindle cutting force signal =8, penalty factor =3267;

[0107] 2. Model Training:

[0108] Construct a CNN-ICBAM-MHSA-BiLSTM network and input an eight-dimensional reconstructed signal;

[0109] Training parameters: cross-entropy loss function, Adam optimizer, learning rate 0.001, 50 iterations, 300 rounds;

[0110] During training, the model parameters with the highest accuracy are saved after each round of validation set evaluation.

[0111] 3. Flutter identification and result verification:

[0112] Input the test set signal into the trained model and output the working condition recognition result;

[0113] Validation metrics: accuracy 98.40%, precision 98.42%, recall 98.40%, F1 score 98.41%, and slight tremor recognition rate improved by 3.9% compared to VMD-CNN-BiLSTM.

[0114] III. Result Verification

[0115] The recognition performance of this invention under three working conditions is as follows:

[0116] Stable processing: 99.3% recognition accuracy;

[0117] Slight flutter: 97.8% accuracy rate in detection;

[0118] Severe flutter: 98.5% accuracy rate in detection;

[0119] Compared to existing models such as CNN-BiLSTM, VMD-CNN-BiLSTM, and CNN-CBAM-BiLSTM, this invention demonstrates significant advantages in the recognition of slight tremors and exhibits strong resistance to noise interference. Even under interference from machine tool spindle rotation noise and environmental noise, the recognition accuracy remains high. 98.4%.

[0120] All experiments of this invention were conducted on the experimental platform belonging to the School of Information Science and Engineering, Shenyang University of Technology. This platform is based on a domestically produced vertical machining center (model Mike CNC) and equipped with a complete multi-sensor synchronous acquisition and data analysis system. The experiments used A16061 aluminum alloy, widely used in industry, as the workpiece material, and employed a 4mm diameter three-slot end mill for end milling. This implementation will describe in detail the complete technical chain from physical signal acquisition, data preprocessing, algorithm parameter optimization, neural network model construction and training, to final performance verification and result analysis, aiming to provide a reproducible and operable high-precision chatter recognition solution.

[0121] Multi-source sensor information acquisition and construction of high-quality datasets:

[0122] The data basis of this invention comes from the synchronous monitoring of multiple physical quantities during the machine tool cutting process. For example... Figure 4 In the test setup shown, we deployed sensors on three key components of the machine tool: a vibration acceleration sensor was installed on the spindle to capture radial (x, y directions) vibration, while an integrated triaxial force sensor measured the radial cutting force; a vibration acceleration sensor was installed on the worktable to monitor vibration in the feed direction and correspondingly measure the cutting force in the feed direction; and a vibration acceleration sensor was installed at the end of the leadscrew to acquire axial vibration and simultaneously measure the axial cutting force. All sensor signals were synchronously acquired by a high-precision data acquisition system (DH5956 and CL-YD-3210), with the sampling frequency uniformly set to 20 kHz to ensure the synchronization of the time-domain signals and the accuracy requirements of the frequency-domain analysis (capable of resolving frequency components up to 10 kHz, far exceeding the machine tool's main natural frequencies). The experimental design aimed to cover the complete process from stable machining to chatter instability, and a total of eight milling experiments with different process parameters were conducted. The spindle speed was set to two levels: 18,000 rpm (300 Hz) and 21,000 rpm (350 Hz). The feed per tooth was set to 0.10 mm and 0.14 mm, respectively. The axial depth of cut was gradually adjusted between 0.15 mm and 0.62 mm to induce different degrees of chatter. Raw signals were continuously acquired for 10 seconds for each experiment.

[0123] To obtain supervised learning labels with clear physical meaning, we abandoned the unreliable method of relying solely on signal thresholds and instead adopted workpiece surface morphology as the gold standard for classifying working conditions. For example... Figure 5 As shown, in the continuously collected data segments, the areas with smooth workpiece surfaces during the initial processing stage (data points 0-3000) are labeled "Stable Processing" (label 1); the areas where uniform but visible textures begin to appear on the workpiece surface (data points 3001-6000) are labeled "Slight Chatter" (label 2); and the areas where clear, dense chatter marks appear on the workpiece surface (data points 6001-9000) are labeled "Severe Chatter" (label 3). This classification method based on processing results ensures the objectivity and accuracy of the labels. Finally, the eight sets of experimental data were integrated, and 3000 continuous data points were extracted from each working condition of each set to form a multi-sensor fusion dataset containing 72,000 sample points. In accordance with general machine learning standards, the dataset was randomly shuffled in a 7:3 ratio and divided into a training set (50,400 samples) and a test set (21,600 samples), as detailed in Table 1.

[0124] Adaptive signal preprocessing based on Arithmetic Optimization Algorithm (AOA):

[0125] Cutting chatter signals are often overwhelmed by powerful forced vibrations from spindle rotation, gear meshing noise, and random environmental interference. Direct analysis of the raw signal makes feature extraction extremely difficult. Variational Mode Decomposition (VMD), as an adaptive, quasi-orthogonal signal decomposition method, can decompose complex signals into a series of intrinsic mode functions (IMFs) with specific center frequencies and finite bandwidths, making it well-suited for processing such non-stationary signals. However, the decomposition performance of VMD is highly dependent on two key parameters: the number of mode decompositions, K, and the penalty factor α. A K value that is too small leads to under-decomposition and mode aliasing; a K value that is too large leads to over-decomposition and the generation of spurious components. The α value affects the bandwidth of each IMF; a smaller value results in a wider bandwidth. Traditionally, these two parameters have been set empirically, lacking universality.

[0126] To address this, this invention introduces the Arithmetic Optimization Algorithm (AOA) for adaptive global optimization of VMD parameters. AOA is a metaheuristic optimization algorithm inspired by arithmetic operations (addition, subtraction, multiplication, and division). Its core lies in dynamically balancing the exploration and development phases through the Mathematical Acceleration Function (MOA).

[0127]

[0128] in, This represents the current iteration number. The maximum number of iterations, and These are the minimum and maximum values ​​of MOA (usually 0.2 and 0.8). When random numbers... When the condition is met, the algorithm performs an exploration (global search); otherwise, it performs a development (local search).

[0129] We use the minimum normalized sample entropy as the fitness function of AOA to find the VMD parameter combination that best represents the inherent regularity of the signal and removes noise. Normalized sample entropy is an effective indicator for measuring the complexity and regularity of time series. For near-periodic signals such as flutter, its value reaches its minimum when noise is effectively removed and the dominant oscillator component is highlighted.

[0130] The fitness function is defined as follows:

[0131]

[0132] In implementation, the AOA population size was set to 20, and the maximum number of iterations was set to 15. Considering practical engineering implications, the parameter search range was set as follows: , Taking the vibration acceleration signal in the x-direction of the principal shaft as an example, parameter optimization is performed, and the convergence curve of its fitness with the number of iterations is shown below. Figure 6 As shown in the table, the algorithm rapidly decreases in the early stages of iteration and converges to a stable optimal value around the 7th generation. This indicates that AOA possesses excellent global exploration capabilities and efficient local optimization. Table 1 shows the optimal parameter combinations obtained for different sensor signals. For example, for the vibration signal in the x-direction of the main shaft, the optimal parameters are... .

[0133] Table 1 Optimal Parameter Combinations for Spindle

[0134]

[0135] By substituting the optimized parameters into VMD, the original signal can be adaptively decomposed. Figure 7 This demonstrates the eight IMF components obtained after VMD decomposition of the spindle x-direction vibration signal (corresponding to...). The time-domain waveforms of the components are shown. To filter out the components most relevant to chatter and eliminate redundant noise from these components, this invention proposes a dual-criteria screening method based on wavelet packet energy entropy and natural frequency proximity. First, the wavelet packet energy entropy of each IMF component is calculated; the lower the entropy value, the more ordered the component and the less noise it contains. Table 3 lists the wavelet packet energy entropy values ​​of different IMFs of the spindle sensor signals. Second, the first three natural frequencies of each component of the machine tool were measured through hammer impact tests, as shown in Table 2. The chatter frequency is usually close to a certain natural frequency of the system.

[0136] Table 2. First 3 natural frequencies of each component

[0137]

[0138] Table 3 Wavelet packet energy entropy of principal axis IMFS

[0139]

[0140] Figure 8 The spectra of the aforementioned IMF components are shown, and the center frequencies of each IMF are compared with the natural frequencies in Table 2. For example, the center frequency of IMF1 is approximately 48Hz, which is very close to the first-order natural frequency of the principal axis (48.7Hz), and its wavelet packet energy entropy (0.326) is the lowest among all IMFs, thus it is identified as a critical flutter-sensitive component. Based on these two criteria, we selected IMF1, IMF2, IMF5, and IMF8 for signal reconstruction. The reconstructed signal is as follows: Figure 9 As shown, compared with the original signal, the periodic modulation phenomenon related to flutter in its waveform is clearer, and the background noise is significantly suppressed, laying a high-quality data foundation for subsequent feature extraction.

[0141] Construction and training of the CNN-ICBAM-MHSA-BiLSTM fusion recognition model:

[0142] The preprocessed reconstructed signal is a time-series data containing multiple sensor channels. To fully exploit its deep spatial (inter-sensor) and temporal (sequence sequence) features, this invention designs a hierarchical deep learning fusion network, the overall architecture of which is as follows: Figure 8 As shown.

[0143] 1. CNN Spatial Feature Extraction Module: This module is responsible for extracting local sensitive features from multi-sensor reconstructed signals. The input data dimensions are (batch size, number of sensor channels, sequence length). First, it passes through a convolutional kernel with a kernel size of... The convolutional layers are designed to fuse features across different sensor channels at the same time step, extracting local correlations across sensors and outputting 32 feature maps. A similar convolutional layer is then applied, outputting 64 feature maps. Each convolutional layer is followed by a ReLU activation function to introduce non-linearity. Afterwards, a sequence defolding layer converts the spatial feature maps output by the convolutions into a sequence format, which is then flattened by a tiling layer to form two-dimensional sequence data suitable for temporal network processing.

[0144] 2. Improved Convolutional Block Attention Module (ICBAM):

[0145] First, the multi-convolutional block attention mechanism CBAM is used, which includes a dual attention module of "channel + space";

[0146] Channel attention is achieved by fusing feature weights from different channels using Global Average Pooling (GAP) and Global Max Pooling (GMP).

[0147]

[0148]

[0149] The results of average pooling and max pooling are fed into the shared fully connected layers FC1 and FC2 to generate channel attention maps. :

[0150]

[0151] in, It is the ReLU activation function. and These are the weights of the fully connected layer. It is the Sigmoid activation function;

[0152] Finally, the channel attention map Applied to input feature map The above yields the channel-weighted feature map. :

[0153]

[0154] Spatial attention is:

[0155] Channel-weighted feature map Perform channel summation to obtain the spatial features of each pixel location:

[0156]

[0157]

[0158] By extracting the weights of key regions in the spatial dimension through convolution, the spatial distribution of flutter features is enhanced, and irregular noise regions are suppressed.

[0159] Improvements were made to the original CBAM, in a single 3 Based on the convolution block of 1, add 5 1, 7 1 convolutional block, and 3 Parallel processing of convolutional blocks and feature fusion of 1 are used to construct multi-scale spatial attention, which more accurately adapts to the feature distribution of cutting chatter signals:

[0160]

[0161] The three convolution results are concatenated to obtain the concatenated feature map. :

[0162]

[0163] Use a 1×1 convolution kernel to concatenate the feature maps Perform convolution to generate a spatial attention map. :

[0164]

[0165] Finally, the spatial attention map Applied to the channel-weighted feature map The final ICBAM feature map is obtained from the above.

[0166] .

[0167] Specifically, standard convolutional neural networks treat all channels and spatial locations "equally," which may fail to focus on key frequency bands and key sensors when processing flutter signals. To address this, we integrate an improved CBAM attention mechanism after the CNN. Traditional CBAM performs channel attention and spatial attention sequentially. We have made multi-scale improvements to its spatial attention part, with the structure as follows: Figure 2 As shown. Specifically, we used three different convolutional kernel sizes—3×1, 5×1, and 7×1—in parallel within the spatial attention branch to process the feature sequence. These three parallel paths focus on local details, mid-range dependencies, and global contextual information, respectively. Their outputs are fused through concatenation and convolution to generate the final spatial attention weight map. This multi-scale design enables the model to adaptively capture key regions of the flutter signal in different spatial ranges (corresponding to different sensors and frequencies), significantly improving the discriminative power of the features. The improved attention process can be summarized as follows: ,in This is a channel attention map. This is a multi-scale spatial attention map. This indicates element-wise multiplication.

[0168] 3. Bidirectional Long Short-Term Memory (BiLSTM) Temporal Modeling Module:

[0169] First, the input matrix is ​​transformed through three different linear transformations, namely the weight matrices. Convert to query ,key Sum :

[0170]

[0171] in, , , These are trainable weight matrices used to generate queries, keys, and values, respectively.

[0172] The self-attention of each head is calculated as follows:

[0173]

[0174] in, It is the dot product of the query and the key. It is a scaling factor to prevent the gradient from vanishing due to an excessively large dot product.

[0175] Then, the scores are normalized using a softbox operation to obtain the degree of attention each position gives to other positions;

[0176] In MHSA, four heads are used to calculate different attention scores. The outputs of all heads are concatenated together to obtain the final attention output.

[0177]

[0178] in, It is the output linear transformation matrix. The outputs of each head are concatenated to obtain the final multi-head attention result.

[0179] Specifically, flutter is a dynamic process with strong time dependence, where the vibration at the current moment is influenced by past cutting marks (regeneration effect). BiLSTM can learn the long-term dependencies of sequences simultaneously from both forward and backward directions. Its core unit, the LSTM, is structured through an input gate... Forgotten Gate Output gate and cell state To control the flow of information:

[0180]

[0181] in, For the sigmoid function, This is element-wise multiplication. BiLSTM consists of two independent LSTM layers, one for processing the forward sequence and the other for processing the reverse sequence (structure as follows). Figure 4 Finally, the hidden states in the two directions are concatenated to form a feature representation containing complete contextual information: .

[0182] 4. Multi-Head Self-Attention Mechanism (MHSA): Although BiLSTM can capture long-range dependencies, it is essentially a recursive model, and its "attention" to all time steps in the sequence is implicit and relatively uniform. However, the incubation and outbreak of flutter are often closely related to abrupt changes at certain specific moments. Therefore, we introduce a multi-head self-attention mechanism after BiLSTM. MHSA allows the model to pay attention to information at different positions (time steps) in the sequence in parallel, and explicitly assigns higher importance to key time steps through attention weights. Its calculation process is as follows:

[0183]

[0184] in, The query, key, and value are obtained from the input sequence through linear transformations. Multi-head attention repeats this process h times, each time using a different linear transformation, and finally combines the results. This allows the model to collaboratively focus on temporal information from different subspaces (e.g., evolutionary patterns of different frequency components), and it is particularly adept at capturing weak but crucial temporal patterns that mark the onset or exacerbation of flutter.

[0185] The four modules described above are connected in series to form an end-to-end flutter detection model. The output layer of the model is a Softmax classifier that maps the learned high-level features to three categories: "stable," "mild flutter," and "severe flutter."

[0186] Model training, results analysis, and performance validation:

[0187] Model training was performed on a workstation equipped with an Intel Core i5-13600KF CPU, an NVIDIA GeForce RTX 4070 Super GPU, and 32GB of RAM, using Python 3.8 and the PyTorch framework. The cross-entropy loss function was used, with the Adam optimizer employed. The initial learning rate was set to 0.001, the batch size to 128, and a total of 300 training epochs were completed. During training, performance was evaluated every 50 epochs on an independent validation set (split from the training set), and the weights of the model with the highest validation accuracy were saved for final test set evaluation.

[0188] To quantitatively evaluate model performance, we use four metrics: precision, recall, F1 score, and accuracy.

[0189]

[0190]

[0191]

[0192]

[0193] in, These represent true positive, true negative, false positive, and false negative, respectively.

[0194] We first conducted systematic ablation experiments, and the results are shown in Table 4. The baseline model CNN-BiLSTM directly processed the raw signal, achieving an accuracy of only 74.57%. After introducing AOA-VMD preprocessing (Improved Model 1), the accuracy significantly improved to 96.66%, irrefutably demonstrating the fundamental contribution of optimized signal preprocessing to fremitus detection. Building on this, we gradually added attention mechanisms: adding the standard CBAM module (Improved Model 2) improved the accuracy to 97.81%; further adding the MHSA module (Improved Model 3) achieved 97.89%; finally, using our proposed combination of ICBAM and MHSA (Improved Model 4), the accuracy reached a maximum of 98.40%, with precision, recall, and F1 score all simultaneously improving to over 98.4%. This clearly demonstrates the effectiveness of each improved module and their synergistic enhancement effect.

[0195] Table 4 Comparison of ablation test results

[0196]

[0197] For a broader comparison, we compared our proposed complete method (AOA-VMD + CNN-ICBAM-MHSA-BiLSTM) with several common models on the same test set, and the results are shown in Table 6. Whether it's BiLSTM, CNN, or CNN-BiLSTM, the accuracy hovers between 71% and 75% when directly processing the raw signal. Even with VMD preprocessing (such as VMD-BiLSTM and VMD-CNN), the accuracy only improves to 96.5%–96.7%, still lower than our proposed model. This clearly demonstrates that the fusion network architecture proposed in this paper has a significant advantage in feature learning capabilities.

[0198] Figure 10 visualizes the recognition details of the three typical models using confusion matrices. Figure 10(a) shows that the CNN-BiLSTM model severely confuses "mild tremor" and "severe tremor". Figure 10(b) shows that after VMD preprocessing, the VMD-CNN-BiLSTM model significantly improves its ability to recognize the tremor category, but some confusion still exists. Figure 10(c) shows that the improved model proposed in this paper not only maintains a near-perfect recognition rate for stable processing categories, but more importantly, it greatly distinguishes between "mild tremor" and "severe tremor", with highly concentrated values ​​on the diagonal and very few misclassifications.

[0199] at last, Figure 11 The paper specifically compares the performance of various models on the two key categories of "mild tremor" and "severe tremor." The bar chart clearly shows that, compared to the VMD-CNN-BiLSTM model, the proposed model improves precision by 3.9%, recall by 2.4%, and F1 score by 3.2% in "mild tremor" identification; and significantly improves the F1 score by 12.2% in "severe tremor" identification. This strongly demonstrates the superior ability of the proposed method, especially the MHSA mechanism, to capture early, weak features of tremor and distinguish different degrees of tremor.

[0200] In summary, this specific implementation method, through rigorous and detailed steps, from hardware construction, data engineering, algorithm optimization to model design, fully reproduces and verifies the proposed cutting chatter identification method. Experimental results consistently demonstrate that this method, through the technical route of "optimized preprocessing + deep feature fusion," achieves high-precision and robust identification of three machine tool cutting conditions with nearly 98.4% accuracy, providing an effective and advanced solution to the problem of chatter monitoring in industrial production.

[0201] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A cutting chatter recognition method based on optimized VMD and CNN-ICBAM-MHSA-BiLSTM, characterized in that, The identification method includes the following steps: Step S1: Acquire vibration acceleration signals and cutting force signals during the cutting process; Step S2: The Arithmetic Optimization Algorithm (AOA) is used to optimize the parameters of Variational Mode Decomposition (VMD). The optimized VMD is then used to adaptively decompose and reconstruct the acquired signal to obtain the reconstructed signal and construct a sample dataset. Step S3: Construct a fusion network that includes a convolutional neural network (CNN), an improved convolutional block attention module (ICBAM), a bidirectional long short-term memory network (BiLSTM), and a multi-head self-attention mechanism (MHSA). Input the reconstructed signal into the fusion network for iterative optimization training to obtain the flutter recognition model. Step S4: Input the signal to be identified into the trained chatter recognition model and output the working condition recognition result of cutting chatter.

2. The cutting chatter recognition method based on optimized VMD and CNN-ICBAM-MHSA-BiLSTM according to claim 1, characterized in that, The specific methods for acquiring vibration acceleration signals and cutting force signals during the cutting process as described in step S1 include: Data is collected from three components of the machine tool using multiple sensors: the spindle, the worktable, and the leadscrew. The spindle signals are vibration acceleration signals in the x and y directions and radial cutting force signals. The signal types of the worktable are vibration acceleration signals in the x and y directions and cutting force signals in the feed direction; The signal types of the lead screw shaft are x-direction vibration acceleration signal and axial cutting force signal; A total of 8 sets of data were collected, with each set having a sampling duration of 10 seconds and a sampling frequency of 20kHz. Each set of signals had 9000 sampling points and covered three operating conditions: stable processing, slight chatter, and severe chatter.

3. The cutting chatter recognition method based on optimized VMD and CNN-ICBAM-MHSA-BiLSTM according to claim 1, characterized in that, The specific method described in step S2, which involves using the arithmetic AOA optimization algorithm to optimize the parameters of variational mode decomposition (VMD), and then using the optimized VMD to adaptively decompose and reconstruct the acquired signal to obtain the reconstructed signal and construct the sample dataset, includes: S2-1: Construct an arithmetic optimization algorithm (AOA) with the small normalized sample entropy as the fitness function to optimize the number of modes in variational mode decomposition (VMD). and penalty factor ; Arithmetic Optimization Algorithm (AOA) parameter settings: population size 20, number of iterations 15. The optimization interval is [3,10]. The optimization range is [200, 3500]. Normalized minimum sample entropy fitness function formula: In the formula: For embedded dimensions, This is the tolerance threshold; Indicates the embedding dimension The condition is satisfied that the distance is less than the threshold. The number of template vector pairs, Indicates the embedding dimension The condition is satisfied that the distance is less than the threshold. The number of template vector pairs; S2-2: The optimized version and Substituting the values ​​into Variational Mode Decomposition (VMD), the multi-sensor signals are decomposed into... Individual intrinsic mode functions (IMFs); S2-3: IMFs with natural frequencies close to those of machine tool components are selected by wavelet packet energy entropy and wavelet energy entropy, where wavelet energy entropy represents the correlation between the reconstructed signal and the original signal structure; Since the flutter frequency is close to the natural frequency, severe flutter may cause a flutter frequency shift, requiring the selection of other frequency bands; The wavelet packet energy entropy threshold is set to 1.2, and IMFs with a threshold less than the threshold are selected because a small wavelet packet energy entropy indicates strong signal structure and less noise. IMFs with frequency bands similar to the inherent frequencies of the acquisition components were selected: first-order 48Hz and second-order 519Hz for the spindle; first-order 399Hz and second-order 1200Hz for the worktable; and first-order 43Hz and second-order 650Hz for the lead screw. The filtered IMFs are reconstructed to obtain a preprocessed signal dataset, which is then divided into a training set and a test set in a 7:3 ratio.

4. The cutting chatter recognition method based on optimized VMD and CNN-ICBAM-MHSA-BiLSTM according to claim 1, characterized in that, The specific method for constructing a fusion network comprising a Convolutional Neural Network (CNN), an Improved Convolutional Block Attention Module (ICBAM), a Bidirectional Long Short-Term Memory Network (BiLSTM), and a Multi-Head Self-Attention Mechanism (MHSA) in step S3, and inputting the reconstructed signal into the fusion network for iterative optimization training to obtain the tremor recognition model includes: S3-1: Network Architecture Design: Construct a fusion network of CNN-ICBAM-MHSA-BiLSTM, including the following modules: Convolutional Neural Network (CNN): 2 convolutional layers, convolutional kernels [2,1], output channels 32, 64, ReLU activation layer, sequence defolding layer, and tiling layer, used to extract local sensitive features; Improved Convolutional Block Attention Module (ICBAM): Channel Attention + Multi-Scale Spatial Attention. Channel attention generates channel weights through GAP / GMP dual-path + fully connected layer. Spatial attention is processed in parallel using 3×1, 5×1, and 7×1 convolutional kernels. After concatenation, spatial weights are generated by 1×1 convolution to reduce dimensionality and enhance key features to suppress noise. Multi-head self-attention mechanism (MHSA): 4 heads, giving higher weight to key time steps to highlight the weak characteristics of slight flutter; Bidirectional Long Short-Term Memory Network (BiLSTM): 6-dimensional hidden layers, output mode last, captures bidirectional temporal dependencies of signals; Classification layer: Fully connected layer (output channel 3) + Softmax activation layer, outputting the probabilities of three operating conditions; S3-2: Model Training: The model uses the cross-entropy loss function, Adam optimizer, learning rate 0.001, batch size 128, number of iterations 50, number of rounds 300, parameters are updated through backpropagation, and the optimal model on the training set is saved.

5. The cutting chatter recognition method based on optimized VMD and CNN-ICBAM-MHSA-BiLSTM according to claim 4, characterized in that, The improved convolutional block attention module (ICBAM) specifically includes: First, the multi-convolutional block attention mechanism CBAM is used, which includes a dual attention module of "channel + space"; Channel attention is achieved by fusing feature weights from different channels using Global Average Pooling (GAP) and Global Max Pooling (GMP). in, Indicates the input feature map, and These represent the height and width of the feature map, respectively. Indicates the number of channels; The feature map represents the first time. The first channel, the first The eigenvalue at the location; This indicates the spatial dimension of the input feature map. The channel vector obtained after global average pooling; This indicates the spatial dimension of the input feature map. The channel vector obtained after performing global max pooling on it; This represents an accumulation operation along the spatial feature dimension. This represents the maximum value operation in the spatial dimension; The results of average pooling and max pooling are fed into the shared fully connected layers FC1 and FC2 to generate channel attention maps. : in, It is the ReLU activation function. and These are the weights of the fully connected layer. It is the Sigmoid activation function; Finally, the channel attention map Applied to input feature map The above yields the channel-weighted feature map. : in, Indicates the input feature map, Represents the channel attention feature map. This indicates element-wise multiplication, which involves weighting each channel. Spatial attention is: Channel-weighted feature map Perform channel summation to obtain the spatial features of each pixel location: By extracting the weights of key regions in the spatial dimension through convolution, the spatial distribution of flutter features is enhanced, and irregular noise regions are suppressed. Improvements were made to the original CBAM, in a single 3 Based on the convolution block of 1, add 5 1, 7 1 convolutional block, and 3 Parallel processing of convolutional blocks and feature fusion of 1 are used to construct multi-scale spatial attention, which more accurately adapts to the feature distribution of cutting chatter signals: The three convolution results are concatenated to obtain the concatenated feature map. : Use a 1×1 convolution kernel to concatenate the feature maps Perform convolution to generate a spatial attention map. : Finally, the spatial attention map Applied to the channel-weighted feature map The final ICBAM feature map is obtained from the above. 。 6. The cutting chatter recognition method based on optimized VMD and CNN-ICBAM-MHSA-BiLSTM according to claim 4, characterized in that, The bidirectional long short-term memory network BiLSTM specifically includes: First, the input matrix is ​​transformed through three different linear transformations, namely the weight matrices. Convert to query ,key Sum : in, , , These are trainable weight matrices used to generate queries, keys, and values, respectively. The self-attention of each head is calculated as follows: in, It is the dot product of the query and the key. It is a scaling factor to prevent the gradient from vanishing due to an excessively large dot product. Then, the scores are normalized using a softbox operation to obtain the degree of attention each position gives to other positions; In MHSA, four heads are used to calculate different attention scores. The outputs of all heads are concatenated together to obtain the final attention output. in, It is the output linear transformation matrix. The outputs of each head are concatenated to obtain the final multi-head attention result.

Images