Method of milling chatter monitoring with feature fusion and squeeze excitation attention mechanism

Abstract

The application discloses a kind of milling chatter monitoring methods of feature fusion and extrusion excitation attention mechanism, belong to machining state monitoring and intelligent manufacturing technical field, the present application is realized to the fine representation of milling vibration signal multiscale time-frequency feature by continuous wavelet transform, it is suitable for non-stationary vibration signal analysis;Using multiscale feature fusion mechanism, local mutation characteristics and overall time-frequency structure characteristics in the evolution process of chatter can be captured simultaneously;The three extrusion excitation attention mechanisms proposed significantly enhance the attention ability of the model to chatter sensitive features, improve the accuracy and robustness of chatter recognition;The method does not need artificial setting feature parameters and threshold, strong adaptability, suitable for on-line monitoring under complex milling conditions, with good engineering application prospect.

Description

Technical Field

[0001] This invention relates to the field of machining condition monitoring and intelligent manufacturing technology, and in particular to a milling chatter monitoring method based on feature fusion and extrusion-excitation attention mechanism. Background Technology

[0002] In the aerospace, automotive, and high-end equipment manufacturing industries, thin-walled parts are widely used due to their light weight and high material removal rate. However, because thin-walled parts have low overall stiffness and high structural flexibility, they are prone to chatter during high-speed milling. Chatter not only significantly reduces workpiece surface quality and dimensional accuracy but also shortens tool life and can even cause equipment damage, becoming a major problem restricting the efficient and high-quality machining of thin-walled parts.

[0003] Existing methods for monitoring milling chatter mainly include those based on frequency domain analysis, empirical feature parameter extraction, and threshold discrimination. These methods typically rely on manually selected features and empirical thresholds, resulting in problems such as high feature redundancy, strong subjectivity, and poor adaptability to complex working conditions. With the development of deep learning technology, chatter monitoring methods based on convolutional neural networks have gradually attracted attention. However, existing methods generally suffer from heavy reliance on single-scale features and insufficient response to local chatter abrupt changes, making it difficult to simultaneously consider both local time-frequency features and overall structural features during chatter evolution.

[0004] Therefore, this invention proposes a chatter monitoring method that can fully mine multi-scale time-frequency information of milling vibration signals and adaptively enhance key chatter features to improve the accuracy and robustness of chatter identification. Summary of the Invention

[0005] The purpose of this invention is to address the technical deficiencies in the existing technology by providing a milling chatter monitoring method based on feature fusion and squeeze-excitation attention mechanism.

[0006] The technical solution adopted to achieve the purpose of this invention is: A method for monitoring milling chatter based on feature fusion and squeeze-excitation attention mechanism includes the following steps: Step 1: During the milling process, the milling vibration signal is collected in real time. The milling vibration signal is a one-dimensional non-stationary vibration signal. Step 2: Perform multi-scale time-frequency analysis on the preprocessed milling vibration signal to obtain a two-dimensional time-frequency image; Step 3: Input the two-dimensional time-frequency image into a convolutional neural network model containing multiple parallel convolutional layers, extract multi-scale features from the two-dimensional time-frequency image, fuse the multi-scale features to obtain multi-scale fused features; Step 4: Input the multi-scale fusion features into the triple squeeze-excited attention mechanism. The triple squeeze-excited attention mechanism outputs a weight vector containing three dimensions: global, local, and adjacent. The weight vectors are fused to obtain a fusion weight vector that comprehensively reflects the importance of each channel. After normalization, weighted processing is performed to identify flutter.

[0007] In the above technical solution, in step 1, a Kistler 8636C vibration acceleration sensor is used to collect milling vibration signals.

[0008] In the above technical solution, step 2, the preprocessing includes outlier removal, detrending, bandpass wavelet denoising, normalization, and segment alignment.

[0009] In the above technical solution, in step 2, the Continuous Wavelet Transform (CWT) is used to perform multi-scale time-frequency analysis on the preprocessed milling vibration signal, converting the one-dimensional non-stationary vibration signal into a two-dimensional time-frequency image, and extracting the time-domain signals of the three typical stages of milling stability, chatter incubation, and chatter outbreak, which are used to characterize the distribution characteristics of vibration energy in the time and frequency domains during the milling process.

[0010] In the above technical solution, wavelet transform is expressed as: dt= dt; Where x(t) is the acquired milling vibration signal, a is the scale factor, which determines the size and location of the time-frequency window, b is the translation factor, which determines the position of the time-frequency window in the time domain, and t is the time variable. (t) is the wavelet basis function, also called the mother wavelet, and its expression is: (t)= ,b 0,a .

[0011] In the above technical solution, in step 3, four parallel convolutional layers (conv1, conv2, conv3, and conv4) are used to extract features at different scales from the input two-dimensional time-frequency image. After each convolutional layer, the ReLU activation function is used as the activation value of each layer's features. The four extracted features at different scales are then fed into the Triple SE Block attention network to increase the attention weights of the features.

[0012] In the above technical solution, step 4, the triple squeeze-incentivized attention mechanism includes three parallel feature weighting branches: a channel global weight modeling branch based on a fully connected layer, a channel local correlation modeling branch based on 1×1 convolution, and an adjacent channel association feature modeling branch based on one-dimensional convolution.

[0013] In the above technical solution, the channel global weight modeling branch based on the fully connected layer performs global average pooling on the input multi-scale fusion features, and inputs the pooled one-dimensional features into two-level fully connected layers. Through parameter learning of the fully connected layers, the output is a global weight vector with the same number of input channels.

[0014] In the above technical solution, the channel local correlation modeling branch based on 1×1 convolution sets a 1×1 convolution kernel that matches the number of input channels, locally weights the features of each channel, and outputs a local weight vector.

[0015] In the above technical solution, the adjacent channel association feature modeling branch based on one-dimensional convolution adopts one-dimensional convolution operation. The convolution kernel slides along the channel dimension to calculate and output the weight vector of several adjacent channels.

[0016] In the above technical solution, in step 4, fusion is performed by element-level addition, and the fusion weight vector is input into the Sigmoid activation function for normalization. The weight values ​​are mapped to the [0,1] interval to obtain the final weight coefficients of each feature channel. The weight coefficient of the chatter-related sensitive channel is close to 1, indicating that the channel contains key time-frequency features that are highly related to milling chatter. The model assigns extremely high attention weights to it to achieve adaptive enhancement of chatter-sensitive features, thereby improving the accuracy and reliability of chatter recognition. The weight coefficients of redundant or interference channels are close to 0.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention proposes a milling chatter monitoring method based on feature fusion and a squeeze-excitation attention mechanism. It achieves fine characterization of multi-scale time-frequency features of milling vibration signals through continuous wavelet transform, making it suitable for non-stationary vibration signal analysis. The multi-scale feature fusion mechanism can simultaneously capture local abrupt changes and overall time-frequency structural features during chatter evolution. The proposed triple squeeze-excitation attention mechanism significantly enhances the model's ability to focus on chatter-sensitive features, improving the accuracy and robustness of chatter identification. This method requires no manual setting of feature parameters and thresholds, is highly adaptable, and suitable for online monitoring under complex milling conditions, showing promising engineering application prospects. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention.

[0019] Figure 2 The time-frequency energy evolution and time-frequency diagram are shown, where a represents the milling stabilization stage, b represents the chatter incubation stage, and c represents the chatter outbreak stage.

[0020] Figure 3 A block diagram of the Triple SE triple squeeze-stimulated attention mechanism.

[0021] Figure 4 Triple SE-CNN network flowchart.

[0022] Figure 5 The workpiece machining surface and machining signal waveform diagram, where a represents the milling stabilization stage, b represents the chatter incubation stage, and c represents the chatter outbreak stage.

[0023] Figure 6 The evolution curves of the loss function on the training and validation sets and the change curve of the AUC index on the validation set are shown, where a is the loss curve and b is the AUC index change curve.

[0024] Figure 7 Multi-scale feature fusion flowchart.

[0025] Figure 8 Visualization results of chatter identification at each stage, where a represents the milling stabilization stage, b represents the chatter incubation stage, and c represents the chatter outbreak stage.

[0026] Figure 9 This is the evolution curve of the loss components for normal samples and synthetic abnormal samples during the training process. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0028] The milling chatter monitoring method described in the following embodiments is performed on a high-speed CNC machine, RodersTech RFM760. A three-tooth ball end mill is used to machine an Inconel 718 rectangular thin-walled workpiece. The machine tool spindle speed is 10400 r / min, the feed rate is 1555 mm / min, the axial depth of cut is 0.2 mm, the radial cutting width is 0.125 mm, the feed per pass is 0.001 mm, and dry cutting with cooling is used.

[0029] A method for monitoring milling chatter based on feature fusion and squeeze-excitation attention mechanism includes the following steps: like Figure 1As shown, in step 1, Kistler 8636C triaxial vibration accelerometers are installed in the tool feed direction (X-axis), spindle radial direction (Y-axis), and spindle axial direction (Z-axis) of the CNC milling machine. The signals output by the sensors are conditioned by a Kistler 5019A multichannel charge amplifier and then connected to an NI DAQ data acquisition card. The vibration response generated in the spindle-tool-workpiece system during the milling process is due to the interaction between the tool and the workpiece at a sampling frequency of 50 kHz. The milling vibration signal is a one-dimensional non-stationary vibration signal. The entire milling process is carried out under dry cutting and cooling conditions. The data acquisition of the milling process is completed to construct an experimental dataset. The experimental conditions are shown in Table 1.

[0030] Table 1 Experimental conditions for PHM2010 ; Step 2: Since the milling vibration signal has obvious non-stationary characteristics and its frequency components change continuously with time, the milling vibration signal is preprocessed as follows: Outlier removal employs the 3σ criterion to detect and remove outliers, impulse interference, and abnormal acquisition values ​​from the signal. The signal mean μ and standard deviation σ are calculated, retaining data within the interval [μ-3σ, μ+3σ]. Outliers outside this interval are replaced with interpolated adjacent means to avoid interfering with time-frequency characteristics. Detrending and baseline correction utilize the least squares method to fit the trend term and subtract it from the original signal, eliminating low-frequency trend shifts caused by sensor temperature drift and machine tool spindle rotation errors. The expression is: ,in, This is a milling vibration signal. p(t) To fit the trend polynomial, The signal is detrended. Multi-band filtering and noise reduction are performed using a Butterworth bandpass filter with a passband of 10 Hz to 10 kHz to filter out low-frequency noise from machine tool idling and high-frequency electromagnetic interference. Wavelet thresholding is then applied using a db4 wavelet basis and a 4-level decomposition to soft-threshold shrink high-frequency detail coefficients, further suppressing non-flutter-related noise. Signal normalization is performed on the denoised signal using min-max normalization, mapping the amplitude to the [0,1] interval to unify the signal amplitude under different operating conditions and sensor sensitivities. The expression is: ,in, The preprocessed signal This is the signal after noise reduction. This represents the maximum value of the signal after noise reduction. To minimize the noise-reduced signal, the signal is segmented and aligned. Continuous signals are segmented according to a fixed time window length (e.g., 0.2s / window), with a 50% overlap between windows to ensure that flutter transient features are not lost. Each signal segment is padded or truncated to a fixed length required for the model input, ensuring consistent dimensions in the converted two-dimensional time-frequency plot.

[0031] The preprocessed milling vibration signal was converted from a one-dimensional non-stationary vibration signal to a two-dimensional time-frequency image using continuous wavelet transform (CWT). The acquired vibration signal can be divided into three typical stages: milling stabilization, chatter incubation, and chatter burst. Figure 5 As shown, during the milling stabilization phase... Figure 5 In section a: the vibration acceleration amplitude is stable around 0 m / s², with minimal fluctuations; only small-amplitude random vibrations exist, resulting in a smooth processing procedure; the flutter incubation stage is as follows: Figure 5 In section b: The vibration amplitude increases significantly to approximately ±15 m / s², and the time-domain waveform exhibits periodic oscillations with the amplitude gradually increasing over time, representing an early warning stage of flutter; the flutter outbreak stage is as follows... Figure 5 c: The vibration amplitude increases sharply, with the peak value exceeding ±20 m / s², exhibiting violent and large-amplitude oscillations. Self-excited vibration fully erupts, indicating a severely abnormal machining condition. This is used to characterize the distribution characteristics of vibration energy in the time and frequency domains during milling. The obtained time-frequency image serves as input data for subsequent feature extraction and chatter identification. The time-frequency energy evolution and time-frequency image are shown below. Figure 2 As shown, where: The wavelet transform is expressed as: dt= dt; Where x(t) is the acquired milling vibration signal, a is the scale factor, which determines the size and location of the time-frequency window, b is the translation factor, which determines the position of the time-frequency window in the time domain, and t is the time variable. (t) is the wavelet basis function, also called the mother wavelet, and its expression is: (t)= ,b 0,a .

[0032] like Figure 4 and Figure 7As shown, in step 3, the two-dimensional time-frequency image is input into a convolutional neural network model containing four parallel convolutional layers. By using four parallel convolutional layers (conv1, conv2, conv3, and conv4), features at different scales are extracted from the input two-dimensional time-frequency image. After each convolutional layer, the ReLU activation function is used as the activation value of each layer's features. The extracted features at different scales from the four layers are first fused into multi-scale features, and then the fused features are fed into a Triple SE Block attention network to complete channel adaptive weighting, enhance the attention weight of the features, and make the network more sensitive to the flutter regions in the time-frequency image. The flutter regions receive more attention, resulting in multi-scale fused features.

[0033] Step 4: To address the issues of channel importance imbalance and redundant feature interference in the fused multi-scale features, the multi-scale fused features are input into the trained Triple SE-CNN (the core of which is to adaptively weight the feature channels through three structurally independent and functionally complementary parallel branches, accurately enhancing flutter-related sensitive features and effectively suppressing irrelevant redundant or interfering features; its structural diagram is shown below). Figure 3As shown, it consists of three parallel branches. These three branches use fully connected layers, 1×1 convolutions, and one-dimensional convolutions to perform dimensionality reduction and then dimensionality increase operations on the channels, respectively. The size of the dimensionality-increased feature map is the same as the input. When using convolutional neural networks to extract image features, the relationships between adjacent pixels are closer, and one-dimensional convolutions are very effective in capturing local features. They can identify patterns and relationships between consecutive elements in the input sequence, enhancing the sensitivity of the attention mechanism to local information and thus improving overall performance. The triple-squeeze-excited attention mechanism outputs a weight vector containing global, local, and adjacent dimensions, breaking the limitations of traditional single-branch attention modeling. Finally, after weighting the feature maps from the three branches, a sigmoid function is applied to fix the value range at (0, 1), resulting in a weight between 0 and 1 for each channel of the input feature layer. This yields the final weight coefficients for each feature channel – the weight coefficient for the flutter-sensitive channel approaches 1 (indicating that this channel contains key time-frequency features highly correlated with milling flutter; the model assigns it extremely high attention weights, achieving adaptive enhancement of flutter-sensitive features, thereby improving the accuracy and reliability of flutter recognition), while the weight coefficients for redundant or interfering channels approach 0. These weight coefficients are then multiplied with the original multi-scale fusion features at the channel level to complete the weighting process, ultimately achieving significant enhancement of flutter-sensitive features and effective suppression of redundant and interfering information, providing highly discriminative feature inputs for accurate subsequent flutter state discrimination. The triple-squeeze-driven attention mechanism includes three parallel feature weighting branches: a channel global weight modeling branch based on a fully connected layer; a channel local correlation modeling branch based on 1×1 convolution; and an adjacent channel association feature modeling branch based on one-dimensional convolution.

[0034] Furthermore, the first layer, based on the channel-wide weight modeling branch of the fully connected layer: for the input features Perform global average pooling: ; Activated by two fully connected layers and ReLU: ; in, This is the weight matrix of the first fully connected layer. This is the weight matrix for the second fully connected layer.

[0035] Global weight modeling, with the fully connected branch acting as the "global perceptron" in the triple mechanism, plays a crucial role in capturing the global correlation between feature channels, uncovering cross-channel global weight patterns, and avoiding weight allocation biases caused by limited local perspectives. Its specific workflow consists of two steps: First, global average pooling is performed on the input multi-scale fused features, compressing the two-dimensional features of each channel into a one-dimensional value. This value comprehensively reflects the global feature information of the corresponding channel, achieving a "squeeze" of the feature channels while preserving their global statistical characteristics. Then, the pooled one-dimensional features are input into two fully connected layers. Through parameter learning in the fully connected layers, the global dependencies between different channels are fully explored, outputting a global weight vector consistent with the number of input channels, thus completing the modeling of the global channel weights. The advantage of this branch lies in its ability to grasp the overall contribution of each channel to the flutter feature, ensuring that sensitive channels are given sufficient attention globally.

[0036] Second layer: Using 1×1 convolution to model local channel relationships: ; The convolutional branch—local dimensional weight modeling—serves as a "local correlator" in the triple mechanism, focusing on the local correlations of feature channels to compensate for the insufficient capture of local channel correlations by the fully connected branch. This branch uses point convolution, without changing the spatial size of the input features, performing convolution operations only on the channel dimension: by setting a 1×1 convolution kernel matching the number of input channels, it locally weights and combines the features of each channel, effectively capturing the local dependencies between adjacent channels while reducing the model's computational complexity. Unlike the global modeling of the fully connected branch, the 1×1 convolutional branch emphasizes local interactions between channels, accurately identifying channel combinations sensitive to flutter within a local area, further refining the weight allocation, and avoiding the weakening of locally sensitive features caused by global modeling, thus achieving complementarity between global and local approaches.

[0037] The third layer: One-dimensional convolutional branch – adjacent dimension weights are modeled, and one-dimensional convolution is performed along the channel dimension: ; The one-dimensional convolution branch, acting as the "neighbor catcher" in the triple mechanism, aims to accurately capture the local correlation features between adjacent channels, further refining the local allocation accuracy of channel weights and forming a triple progressive modeling approach of "global-local-neighbor". This branch employs one-dimensional convolution operations, with the convolution kernel sliding along the channel dimension for computation, focusing on the feature interaction relationships between several adjacent channels. By setting an appropriate convolution kernel size, it can effectively capture the feature correlations between adjacent channels, uncovering flutter-sensitive features that are only manifested in the interaction between adjacent channels. These features are often difficult to capture by fully connected branches and 1×1 convolution branches. Furthermore, the lightweight nature of one-dimensional convolution can further improve the accuracy of weight allocation without significantly increasing computational cost, providing more detailed support for enhancing flutter-sensitive features.

[0038] Furthermore, all images processed in step 2 are converted to 256. The image size is 256 pixels, consistent with the input layer of Triple SE-CNN. During model training, a batch size of 8 was used, the Adam optimizer was selected, the learning rate was set to 0.001, and the weight decay coefficient was set to 0.00001 to prevent overfitting. Furthermore, the training epochs for Triple SE-CNN were set to 200, and the cross-entropy loss function was used.

[0039] like Figure 6 As shown, the loss and accuracy curves of Triple SE-CNN on the training and validation sets are displayed. It can be clearly seen that Triple SE-CNN begins to converge at the 40th iteration, and the accuracy curve tends to stabilize at the 100th iteration. At the same time, the loss and accuracy of the training and validation sets are in good agreement, and there is no overfitting phenomenon. Finally, at the 200th epoch, the accuracy of Triple SE-CNN on the training set is 99.1%. Figure 6 In the figure, 'a' represents the loss curves of normal and abnormal samples during training. As can be seen from the figure, the synthesized abnormal samples maintain a high fit with the normal samples during training, and the loss curve tends to stabilize when the 100th iteration is reached.

[0040] The Triple SE-CNN model provided by this invention exhibits excellent convergence characteristics and generalization performance during training. Figure 6 and Figure 9 The depicted loss function and accuracy metric curves were obtained through data augmentation operations. Figure 6The evolution curves of the loss function on the training and validation sets, as well as the change curve of the AUC index on the validation set, show that the model training process exhibits clear phased characteristics: In the early stage of training (0 to 40 epochs), both the training loss and the validation loss decrease rapidly, and the model efficiently learns the basic feature patterns in the data; in the 40 to 100 epoch stage, the rate of loss decrease slows down significantly, the curves become flat, and the model enters the stage of refined feature learning and gradually transitions to a convergent state; when the training epochs exceed 100 epochs, both the training loss and the validation loss enter a stable plateau period, with no obvious fluctuations or rebounds, and the model parameters tend to stabilize. Meanwhile, the evolution trends of training loss and validation loss remained highly synchronized throughout the training process, with the difference between them consistently remaining at a low level. The typical overfitting characteristics of validation loss significantly deviating from training loss or rebounding later were not observed. Correspondingly, the validation set AUC index rose rapidly in the early stages of training, gradually decreasing its growth rate between 20 and 40 epochs, and then stabilizing at a high level above 0.99 after 40 epochs. When training reached the 200th epoch, the model's classification accuracy on the training set reached 99.1%, validating the model's excellent feature learning and classification capabilities. Furthermore, regarding... Figure 9 Analysis of the loss component evolution curves for normal samples and synthetic abnormal samples during training shows that the loss of both normal and abnormal samples on the training and validation sets exhibits highly consistent decreasing and converging behavior: In the early stages of training, the loss values ​​of both types of samples decrease rapidly and synchronously, as the model simultaneously learns the inherent pattern features of normal samples and the deviation features of synthetic abnormal samples, without any learning lag; as the number of training epochs increases, the rate of decrease in the loss of both types of samples slows down synchronously, and the curves gradually flatten out. By the 100th epoch, all loss curves enter a stable plateau period without significant fluctuations; and throughout the entire training process, the loss values ​​of normal and abnormal samples maintain a small difference, with no significant overfitting or underfitting of a single type of sample. This indicates that synthetic abnormal samples can effectively assist the model in learning abnormal patterns, and the model's fitting process for normal and abnormal samples has good synchronicity and balance, without overfitting or underfitting of a single type of sample, providing reliable model support for anomaly detection tasks in industrial scenarios.

[0041] The flutter recognition results of the Triple SE-CNN model were visualized to obtain flutter recognition heatmaps for each stage, as shown below. Figure 8As shown. Each group, from top to bottom, contains the flutter input image and the flutter localization heatmap. The heatmap clearly shows that the Triple SE-CNN model can accurately locate key feature regions highly correlated with flutter in the time-frequency image: In the milling stabilization stage, the model's focus area is dispersed and has low brightness, with no obvious bright areas, highly consistent with the time-frequency characteristics of uniform vibration energy distribution and no strong feature concentration in the stable state; in the flutter incubation stage, the model can automatically lock onto weak energy accumulation areas in the time-frequency image, forming local bright responses, effectively capturing the weak disturbance features in the early stages of flutter, verifying the model's early flutter recognition ability; in the flutter outbreak stage, the model exhibits a strong bright response to high-energy concentration areas in the time-frequency image, with attention weights highly concentrated on the core feature area of ​​flutter, fully demonstrating the weighted strengthening effect of the Triple SE attention mechanism on flutter-related features and the effective suppression of redundant background information. The time-domain signals of the three stages and... Figure 5 The three stages correspond to each other.

[0042] In summary, the core advantage of the triple squeeze-incentivized attention mechanism lies in "triple modeling and synergistic complementarity": the fully connected branch ensures the rationality of global weights, the 1×1 convolutional branch makes up for the lack of local correlations, and the one-dimensional convolutional branch captures the fine correlation between adjacent channels. The three work together to achieve comprehensive and precise adaptive adjustment of feature channel weights, effectively solve the problem of uneven channel importance in multi-scale fusion features, and significantly improve the discriminative ability of features.

[0043] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for monitoring milling chatter using feature fusion and squeeze-excitation attention mechanism, characterized in that, Includes the following steps: Step 1: During the milling process, collect milling vibration signals in real time; Step 2: Continuous wavelet transform is used to perform multi-scale time-frequency analysis on the preprocessed milling vibration signal, converting the one-dimensional non-stationary vibration signal into a two-dimensional time-frequency image, and extracting the time-domain signals of the three typical stages of milling stabilization, chatter incubation, and chatter outbreak. Step 3: Input the two-dimensional time-frequency image into a convolutional neural network model containing multiple parallel convolutional layers, extract multi-scale features from the two-dimensional time-frequency image, fuse the multi-scale features to obtain multi-scale fused features; Step 4: Input the multi-scale fusion features into the triple squeeze-stimulated attention mechanism. The triple squeeze-stimulated attention mechanism outputs a weight vector containing global, local, and adjacent dimensions. The weight vectors are fused to obtain a fusion weight vector that comprehensively reflects the importance of each channel. After normalization, weighting is performed to identify flutter. Among them, the adjacent channel association feature modeling branch based on one-dimensional convolution adopts one-dimensional convolution operation. The convolution kernel slides along the channel dimension to calculate and output the weight vector of several adjacent channels. The triple squeeze-incentivized attention mechanism includes three parallel feature weighting branches: a channel global weight modeling branch based on a fully connected layer, a channel local correlation modeling branch based on 1×1 convolution, and an adjacent channel association feature modeling branch based on one-dimensional convolution. The channel global weight modeling branch based on the fully connected layer performs global average pooling on the input multi-scale fusion features, inputs the pooled one-dimensional features into two-level fully connected layers, and outputs a global weight vector consistent with the number of input channels through parameter learning of the fully connected layers. The channel local correlation modeling branch based on 1×1 convolution locally weights the features of each channel by setting a 1×1 convolution kernel that matches the number of input channels, and outputs a local weight vector.

2. The milling chatter monitoring method according to claim 1, characterized in that, In step 2, the preprocessing includes outlier removal, detrending, bandpass wavelet denoising, normalization, and segment alignment.

3. The milling chatter monitoring method according to claim 1, characterized in that, The wavelet transform is expressed as: dt= dt; Where x(t) is the acquired milling vibration signal, a is the scale factor, which determines the size and location of the time-frequency window, b is the translation factor, which determines the position of the time-frequency window in the time domain, and t is the time variable. For the complete wavelet basis conjugation with normalized coefficients, To remove the normalized coefficients of the mother wavelet conjugate, (t) is the wavelet basis function, also known as the mother wavelet.

4. The milling chatter monitoring method according to claim 3, characterized in that, The expression for the mother wavelet is: (t)= ,b 0,a 。 5. The milling chatter monitoring method according to claim 1, characterized in that, In step 3, features at different scales are extracted from the input two-dimensional time-frequency image by using four parallel convolutional layers: conv1, conv2, conv3, and conv4. After each convolutional layer, the ReLU activation function is used as the activation value of each layer's features. The extracted features at different scales from the four layers are then fed into the Triple SE Block attention network to increase the attention weights of the features.

6. The milling chatter monitoring method according to claim 1, characterized in that, In step 4, fusion is performed by element-level addition, and the fused weight vector is input into the Sigmoid activation function for normalization, mapping the weight values ​​to the [0,1] interval to obtain the final weight coefficients of each feature channel.

Images