Milling chatter prediction method and system based on time distribution and mixed attention

Abstract

The invention provides a milling chatter prediction method and system based on time distribution and mixed attention, and belongs to the field of milling state monitoring. The method comprises the following steps: collecting a multi-channel original time sequence signal in a milling process, and pre-processing the multi-channel original time sequence signal; inputting the preprocessed signal into a time sequence distribution convolution feature extraction module, and carrying out nonlinear feature extraction and dimension reduction to obtain a multi-channel time sequence feature sequence; inputting the multi-channel time sequence feature sequence into a global time sequence attention modeling unit, and outputting global dynamic evolution features; inputting the global dynamic evolution characteristics into a time dimension self-attention module and a channel attention module, and performing key time slice screening and multi-sensor channel importance re-calibration to obtain fusion characteristics; and inputting the fusion features into a classifier, and outputting the chatter probability in a future time window. According to the method, accurate capture and early warning of the chatter initiation precursor are realized, and an effective technical means is provided for active control of the stability of the machining process.

Description

Technical Field

[0001] This invention belongs to the field of milling machining condition monitoring technology, and particularly relates to a method and system for predicting milling chatter based on time distribution and hybrid attention. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Chatter is a common self-excited vibration phenomenon in metal cutting processes, posing a significant threat to machining efficiency, workpiece surface quality, and the lifespan of cutting tools and machine tools. For a long time, chatter suppression has relied on the experience of process engineers or offline mechanism model analysis. However, the actual machining environment is complex and variable, and traditional theoretical models such as regenerative chatter are difficult to accurately describe the real-world working conditions of multi-degree-of-freedom and nonlinear coupling, limiting their application value in real-time early warning and active control.

[0004] With the development of sensor technology and artificial intelligence, data-driven flutter detection methods have become a research hotspot. Methods based on convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been introduced into the field of flutter detection, achieving to some extent the automatic extraction of features from raw signals and temporal modeling. However, these methods still have significant limitations: CNNs excel at capturing local patterns but struggle to model dynamic evolutionary relationships over long time spans; while RNNs and their variants possess some temporal memory capabilities, they are prone to gradient vanishing or information forgetting problems when processing long sequences, and their serial computation method limits model efficiency and real-time performance.

[0005] In recent years, temporal modeling methods, represented by self-attention mechanisms, have demonstrated outstanding long-range dependency modeling capabilities in fields such as natural language processing and speech recognition, and have also provided new insights for flutter analysis. However, current attention-based methods mostly focus on current state recognition, directly employing general temporal modeling structures. This makes it difficult to effectively represent the transitional phase before flutter occurs, and they still suffer from insufficient feature representation in multi-channel signal collaborative modeling. Relying solely on general temporal modeling structures is insufficient to meet the requirements of feature selection and temporal modeling for early flutter prediction. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a milling chatter prediction method and system based on temporal distribution and hybrid attention. Using raw time-series signals from multiple sensors as input, a temporal distribution convolution module extracts features from each time window using shared weights, ensuring consistent representation of different time windows in the feature space. A prediction-driven global temporal attention modeling unit captures long-term dependencies, followed by time-dimension self-attention module to enhance time segment attention and channel attention module for channel recalibration. Finally, a fully connected classifier outputs the probability of chatter occurring within a future time window, thereby achieving early identification and prediction of chatter, ensuring milling stability and product quality.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a milling chatter prediction method based on time distribution and hybrid attention; A milling chatter prediction method based on temporal distribution and hybrid attention includes: Acquire multi-channel raw timing signals during the milling process and perform preprocessing; The preprocessed original time series signal is input into the time series distributed convolutional feature extraction module. A one-dimensional convolutional structure with shared weights is used to independently perform nonlinear feature extraction and dimensionality reduction for each time window to obtain a multi-channel time series feature sequence. The multi-channel temporal feature sequence is input into the global temporal attention modeling unit, and the long-range dependency between different historical time windows is modeled through the self-attention mechanism to output global dynamic evolution features. The global dynamic evolution features output by the global temporal attention modeling unit are sequentially input into the temporal dimension self-attention module and the channel attention module to perform key time segment screening and multi-sensor channel importance recalibration, respectively, to obtain the attention-enhanced fusion features. The fused features are input into the classifier, which outputs the probability of flutter occurring within a future time window.

[0008] As a further technical solution, multi-channel raw timing signals during the milling process are acquired and preprocessed, including: Triaxial acceleration signals and acoustic emission signals are acquired during the milling process using accelerometers and acoustic emission sensors; Windowing slicing is performed using a fixed window length and sliding step size, and the sliced ​​window data is normalized. Construct a prediction time alignment window, map the historical window to the future prediction window in time, and label the signal segments according to the set prediction time interval.

[0009] As a further technical solution, the preprocessed original time-series signal is input into the time-series distributed convolutional feature extraction module. A one-dimensional convolutional structure with shared weights is used to independently perform nonlinear feature extraction and dimensionality reduction for each time window, resulting in a multi-channel time-series feature sequence, including: The preprocessed original temporal signal is input into the temporal distributed convolutional feature extraction module in a time-distributed manner; The temporal distribution convolutional feature extraction module adopts a parameter-shared one-dimensional convolutional neural network structure to perform convolution operations on each time window in the input sequence, extract nonlinear features and perform dimensionality compression. The high-dimensional original signal of each time window is mapped to a low-dimensional feature vector, and the feature vectors of all windows are arranged in chronological order to form the multi-channel time-series feature sequence.

[0010] As a further technical solution, the multi-channel temporal feature sequence is input into a global temporal attention modeling unit. A self-attention mechanism is used to model the long-range dependencies between different historical time windows, outputting global dynamic evolution features, including: The multi-channel temporal features are input into a prediction-driven global temporal attention modeling unit. Through a parallel temporal attention computing mechanism, the correlation between the multi-channel signals in the time dimension is modeled to capture the long-term dynamic characteristics in the flutter evolution process. In the process of temporal attention modeling, the probability of flutter occurring within the future prediction time window is used as a supervision signal. A prediction target constraint mechanism is introduced in the attention weight generation stage to modulate the attention distribution and output global dynamic evolution features.

[0011] As a further technical solution, the time-dimensional self-attention module is used to calculate the relative importance weights between different time segments within the same sensor channel, and to filter out the key time segments that are most sensitive to flutter prediction.

[0012] As a further technical solution, the channel attention module is used to calculate the importance weights between different sensor channels within the same time segment under the constraint of key time segments, so as to realize the collaborative and adaptive recalibration of multiple sensor channels.

[0013] A second aspect of the present invention provides a milling chatter prediction system based on time distribution and hybrid attention.

[0014] A milling chatter prediction system based on temporal distribution and hybrid attention includes: The data acquisition and preprocessing module is configured to: acquire multi-channel raw timing signals during the milling process and perform preprocessing; The temporal feature sequence extraction module is configured to: input the preprocessed original temporal signal into the temporal distributed convolutional feature extraction module, and use a one-dimensional convolutional structure with shared weights to independently perform nonlinear feature extraction and dimensionality reduction for each time window to obtain a multi-channel temporal feature sequence; The global dynamic evolution feature extraction module is configured to: input the multi-channel temporal feature sequence into the global temporal attention modeling unit, model the long-range dependency relationship between different historical time windows through the self-attention mechanism, and output the global dynamic evolution features; The fusion feature acquisition module is configured to: input the global dynamic evolution features output by the global temporal attention modeling unit into the temporal dimension self-attention module and the channel attention module in sequence, and perform key time segment screening and multi-sensor channel importance recalibration respectively to obtain the attention-enhanced fusion features; The flutter prediction output module is configured to input the fused features into a classifier and output the probability of flutter occurring within a future time window.

[0015] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of the milling chatter prediction method based on time distribution and hybrid attention as described in the first aspect of the present invention.

[0016] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the milling chatter prediction method based on time distribution and hybrid attention as described in the first aspect of the present invention.

[0017] The fifth aspect of the present invention provides a computer program product, including a computer program / instruction that, when executed by a processor, implements the steps in the milling chatter prediction method based on time distribution and hybrid attention described in the first aspect of the present invention.

[0018] The above one or more technical solutions have the following beneficial effects: (1) This invention uses a temporally distributed convolutional feature extraction module to process the original multi-channel signal. Multiple consecutive time slices are processed through parallel computation with shared weights, achieving both nonlinear dimensionality reduction and redundant information removal of the high-dimensional original signal, while ensuring the physical consistency of features across different time windows on the time axis. By designing a prediction-driven global temporal attention modeling unit, with the probability of flutter occurrence within the future prediction time window as the supervision target, a prediction target constraint mechanism is introduced into the attention weight calculation. Differentiated constraints are applied to the attention allocation of historical time segments through weight modulation factors, prompting the model to actively focus on time segments that have a key impact on flutter evolution. This effectively solves the gradient vanishing and information forgetting problems in long sequence modeling, and can accurately capture the long-term dynamic evolution characteristics of flutter from its inception, growth to full development.

[0019] (2) This invention constructs a dual feature recalibration mechanism through the synergistic effect of the time-dimensional self-attention module and the channel attention module. On the one hand, it filters out key time segments that are sensitive to flutter, highlighting the weak nonlinear characteristics of flutter precursors; on the other hand, it adaptively and dynamically adjusts the weights of multi-sensor channels, fully explores the complementary information of different channels, avoids interference from irrelevant channels in non-critical periods, and significantly improves the robustness and anti-interference ability of the model to complex working conditions.

[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is a flowchart of the method in the first embodiment.

[0023] Figure 2 This is a flowchart of the reconstruction feature attention mechanism in the first embodiment.

[0024] Figure 3 This is a system structure diagram of the second embodiment. Detailed Implementation

[0025] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0026] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0027] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0028] Example 1 This embodiment discloses a milling chatter prediction method based on temporal distribution and hybrid attention. By collecting signals from multiple sensors and processing them in a windowed manner, features are extracted using a temporal distribution convolution module, and long-term dependencies are captured by a prediction-driven global temporal attention modeling unit. Features are recalibrated using a dual attention module of time and channel, and finally, the chatter probability of the future time window is output. This invention can achieve early warning of the early stage of chatter and has the advantages of high prediction accuracy and strong real-time performance.

[0029] Specifically, such as Figure 1 As shown, the milling chatter prediction method based on temporal distribution and hybrid attention includes: Step S1: Acquire multi-channel raw timing signals during the milling process and perform preprocessing.

[0030] In this embodiment, in order to comprehensively capture the vibration and acoustic emission characteristics during the milling process, a multi-sensor collaborative acquisition scheme is adopted: a three-axis accelerometer and an acoustic emission sensor are respectively arranged at the spindle box of the milling machine tool and the workpiece fixture. The three-axis accelerometer is used to collect vibration acceleration signals in the X, Y, and Z directions, and the acoustic emission sensor is used to collect acoustic emission signals generated by material fracture, friction, etc. during the milling process.

[0031] After data acquisition, the raw signals were initially screened to remove abnormal abrupt changes caused by loose sensor installation, electromagnetic interference, etc., and to remove non-machining state signal segments generated during the tool feed and retraction phases at both ends. Only valid signals from the stable cutting phase were retained to avoid interference from invalid signals on model training and prediction. Since the triaxial acceleration signals and acoustic emission signals have different dimensions, and the numerical ranges of signals from different channels vary significantly, global normalization was performed on the data from each sliced ​​window to eliminate the impact of these dimensional differences on model training and improve the model's numerical stability.

[0032] In addition, considering the timing characteristics of the milling signal, a fixed window length and sliding step size are used to slice the effective signal, and a time mapping relationship between the historical window and the future prediction window is constructed. According to the set prediction time interval, each signal segment is labeled with three tags: stable state, transition state, and chatter state.

[0033] Step S2: The preprocessed original time series signal is input into the time series distributed convolutional feature extraction module. A one-dimensional convolutional structure with shared weights is used to independently perform nonlinear feature extraction and dimensionality reduction for each time window to obtain a multi-channel time series feature sequence.

[0034] The preprocessed standardized time-series signal is input into the time-series distributed convolutional feature extraction module, whose core is a parameter-shared one-dimensional convolutional neural network structure. This parameter-shared one-dimensional convolutional neural network performs synchronous convolution operations on the multi-channel signal within each time window of the input. The convolution kernel slides along the time dimension of the signal, capturing local correlation features between adjacent sampling points point by point, while simultaneously fusing complementary information from the multi-channel signals to generate an initial nonlinear feature map. Subsequently, the convolution output is nonlinearly transformed using the ReLU activation function to filter out effective features sensitive to flutter states and suppress the propagation of invalid noise features.

[0035] By utilizing the convolution stride, feature dimension compression is achieved, mapping each original high-dimensional signal window into a low-dimensional feature vector.

[0036] After feature extraction and dimensionality compression for all time windows, the low-dimensional feature vectors corresponding to each window are arranged sequentially according to the original time order to form a multi-channel temporal feature sequence. This multi-channel temporal feature sequence not only preserves the temporal evolution of the original signal but also obtains deep features that can characterize the flutter state through nonlinear feature extraction. At the same time, it effectively reduces the data dimensionality, providing high-quality input data for the subsequent prediction-driven global temporal attention modeling unit to capture long-term dependencies.

[0037] Step S3: Input the multi-channel temporal feature sequence into the global temporal attention modeling unit, model the long-range dependency relationship between different historical time windows through the self-attention mechanism, and output the global dynamic evolution features.

[0038] The output multi-channel temporal feature sequence is input into the prediction-driven global temporal attention modeling unit. This unit takes a self-attention-based parallel temporal modeling structure as its core and models the correlation between multiple historical time windows simultaneously through a parallel temporal attention computing mechanism. This efficiently captures the long-term dynamic evolution characteristics of flutter from its inception, growth to full development, and avoids the gradient vanishing and information forgetting problems in long sequence modeling.

[0039] To ensure the model focuses on key historical segments crucial to future flutter evolution, a prediction target constraint mechanism is introduced during the attention weight generation stage. The probability of flutter occurrence in the next three windows is used as a monitoring signal to construct a weight modulation factor based on the prediction target. The specific implementation process is as follows: First, the basic self-attention weights under unconstrained conditions are calculated, reflecting the inherent correlation strength between historical windows. Then, the label information of the future prediction windows is converted into constraint signals and fused with the basic self-attention weights for modulation. This differentiates the attention allocation for different historical time segments, assigning higher weights to historical windows with strong indicative significance for future flutter occurrences and reducing the weight proportion of irrelevant or interfering segments, thereby reconstructing the attention distribution. Specifically: First, define the input multi-channel time-series feature sequence as follows: ,in Represents batch size. This represents the number of historical time windows. The feature dimension is used. The input is mapped to three vectors through transformation:

[0040] Next, the basic association strength under unconstrained conditions is calculated to obtain the original attention score matrix. It reflects the inherent correlation strength between different historical event windows under unsupervised signal intervention, and the calculation formula is:

[0041] in, These represent the indices of the time windows. ; is the dimension of the key vector, and the scaling factor is used to prevent the gradient vanishing problem caused by excessively large dot product results.

[0042] Subsequently, a prediction target constraint mechanism is introduced to fuse and modulate the weights. This utilizes the flutter occurrence probability label within the future prediction time window. As a supervisory signal, a prediction loss function is constructed as a constraint signal. The calculation formula is:

[0043] During the model training process, constraint signals Gradient information is generated using the backpropagation algorithm, and the linear transformation matrix is ​​then processed. and The parameters are updated and modulated. This mechanism forces the model to reconstruct the distribution of attention weights learned, and the final attention weights are obtained after normalization using the Softmax function.

[0044]

[0045] Attention weights modulated by the predicted target constraint It can adaptively differentiate the allocation of historical time segments, that is, give higher weight values ​​to historical windows that have a strong causal relationship with the occurrence of future flutter, while reducing the weight ratio of irrelevant or noisy segments.

[0046] Finally, the reconstructed attention weights are utilized. value vector Perform weighted aggregation to output global dynamic evolution features that include long-range dependencies.

[0047]

[0048] in, It represents a global dynamic evolution characteristic.

[0049] Through the parallel self-attention computation with prediction target constraints described above, the model completes long-range dependency modeling of multi-channel temporal feature sequences. It integrates the local deep features of each historical window and enhances the contribution of key time segments, ultimately outputting global dynamic evolution features. These features not only preserve the evolutionary patterns of the temporal dimension but also encapsulate key information strongly correlated with future flutter states, providing high-quality feature support for subsequent attention enhancement and final prediction.

[0050] Step S4: The global dynamic evolution features output by the global temporal attention modeling unit are sequentially input into the time dimension self-attention module and the channel attention module to perform key time segment screening and multi-sensor channel importance recalibration, respectively, to obtain the attention-enhanced fusion features.

[0051] like Figure 2 As shown, the global dynamic evolution features are input into the time-dimensional self-attention module. This module, for each sensor channel, uses the feature vector of a historical time window as the computational unit, models the dependencies between windows through a self-attention mechanism, and outputs the attention weight for each time window. Windows that strongly indicate the initiation and evolution of flutter are assigned higher weights, while irrelevant or low-contribution windows are suppressed. Subsequently, based on these weights, the features of each time window are weighted and fused to filter out the feature information of key time segments.

[0052] Specifically, the global dynamic evolution features output from the previous level are denoted as the input tensor. ,in Represents batch size. This represents the number of historical time windows. The feature dimension is used. Normalization is performed using the Softmax function to obtain the attention weight matrix in the time dimension. Furthermore, regarding the attention weight matrix... The global importance score vector for each time window is obtained by averaging along the row dimension or by max pooling. This scoring vector represents the contribution of each historical time window to flutter prediction:

[0053] Window segments with higher scores were identified as key time segments with strong indicative significance for the initiation and evolution of flutter. By weighting calculations to highlight the feature representation of key segments while suppressing irrelevant or low-contribution window weights, time-enhanced features were obtained after time filtering. :

[0054] Output This will serve as the input to the subsequent channel attention module, providing it with feature constraints enhanced by the time dimension.

[0055] Feature tensors filtered by the time dimension Global pooling is used to compress the data into... To capture the nonlinear interactions between channels, a gating mechanism with two fully connected layers is used to generate channel weights. Then, a normalized channel weight vector is generated using a sigmoid activation function. The generated channel weights Input features Perform channel-based multiplication to recalibrate the original features:

[0056] In this process, channel features containing key flutter information are given high weight, while channel features containing noise or redundant information are suppressed, thereby achieving synergistic enhancement of multi-channel information.

[0057] Furthermore, the features filtered by the time-dimension self-attention module are input into the channel attention module. For each time segment, especially the filtered key time segments, the importance weights of different sensor channels are calculated. By modeling the feature complementarity and correlation between channels, channels that can effectively characterize flutter state are given high weights, while channels with large noise interference or low feature contribution have reduced weights. Subsequently, based on the dynamically generated channel weights, the features of each channel are adaptively weighted and fused to achieve synergistic enhancement of multi-channel information and avoid interference from irrelevant channels in non-critical time periods.

[0058] After two stages of processing—key segment selection in the time dimension self-attention module and channel recalibration in the channel attention module—the attention-enhanced fusion feature is finally obtained. This fusion feature not only incorporates key temporal information strongly correlated with flutter prediction but also achieves optimized collaboration of multiple sensor channels, effectively improving the feature's discriminability and robustness.

[0059] Step S5: Input the fused features into the classifier and output the probability of flutter occurring within the future time window.

[0060] Global mean pooling is applied to the attention-enhanced fused features to aggregate feature information from multiple time windows, resulting in a globally fused feature. This globally fused feature is then input into a fully connected classifier containing multiple layers of neurons. The ReLU activation function enhances the nonlinear fitting capability, and the output is finally converted into a probability distribution using a Softmax function. The class corresponding to the highest probability is selected as the predicted state, and an alert is triggered when the probability of the tremor class exceeds a set threshold.

[0061] Example 2 This embodiment discloses a milling chatter prediction system based on time distribution and hybrid attention; like Figure 3 As shown, a milling chatter prediction system based on temporal distribution and hybrid attention includes: The data acquisition and preprocessing module is configured to: acquire multi-channel raw timing signals during the milling process and perform preprocessing; The temporal feature sequence extraction module is configured to: input the preprocessed original temporal signal into the temporal distributed convolutional feature extraction module, and use a one-dimensional convolutional structure with shared weights to independently perform nonlinear feature extraction and dimensionality reduction for each time window to obtain a multi-channel temporal feature sequence; The global dynamic evolution feature extraction module is configured to: input the multi-channel temporal feature sequence into the global temporal attention modeling unit, model the long-range dependency relationship between different historical time windows through the self-attention mechanism, and output the global dynamic evolution features; The fusion feature acquisition module is configured to: input the global dynamic evolution features output by the global temporal attention modeling unit into the temporal dimension self-attention module and the channel attention module in sequence, and perform key time segment screening and multi-sensor channel importance recalibration respectively to obtain the attention-enhanced fusion features; The flutter prediction output module is configured to input the fused features into a classifier and output the probability of flutter occurring within a future time window.

[0062] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0063] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps in the milling chatter prediction method based on time distribution and hybrid attention as described in Example 1.

[0064] Example 4 The purpose of this embodiment is to provide an electronic device.

[0065] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the milling chatter prediction method based on time distribution and hybrid attention as described in Embodiment 1.

[0066] Example 5 Embodiment 5 of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps in the milling chatter prediction method based on time distribution and hybrid attention as described in Embodiment 1.

[0067] The steps and methods involved in the apparatuses of Embodiments 2, 3, 4, and 5 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0068] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0069] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for milling chatter prediction based on time distribution and hybrid attention, characterized in that, The method comprises the following steps: Collecting multi-channel original time sequence signals in the milling process and preprocessing; Inputting the preprocessed original time sequence signals into a time sequence distribution convolution feature extraction module, independently performing nonlinear feature extraction and dimension reduction on each time window by adopting a one-dimensional convolution structure with shared weights, and obtaining a multi-channel time sequence feature sequence; Inputting the multi-channel time sequence feature sequence into a global time sequence attention modeling unit, modeling the long-range dependence relationship between different historical time windows by a self-attention mechanism, and outputting global dynamic evolution features; Inputting the global dynamic evolution features output by the global time sequence attention modeling unit into a time dimension self-attention module and a channel attention module in sequence, respectively performing key time segment screening and multi-sensor channel importance re-labeling, and obtaining attention-enhanced fusion features; Inputting the fusion features into a classifier, and outputting the probability of vibration occurring in the future time window.

2. The time distribution and hybrid attention based milling chatter prediction method of claim 1, wherein, Collecting multi-channel original time sequence signals in the milling process and preprocessing, comprising: Collecting three-axis acceleration signals and acoustic emission signals in the milling process by using acceleration sensors and acoustic emission sensors; Windowing and slicing are performed by using a fixed window length and a sliding step, and the sliced window data is normalized; A prediction time alignment window is constructed, the historical window and the future prediction window are time-mapped, and the signal segments are labeled according to the set prediction time interval.

3. The time distribution and hybrid attention based milling chatter prediction method of claim 1, wherein, Inputting the preprocessed original time sequence signals into a time sequence distribution convolution feature extraction module, independently performing nonlinear feature extraction and dimension reduction on each time window by adopting a one-dimensional convolution structure with shared weights, and obtaining a multi-channel time sequence feature sequence, comprising: The preprocessed original time sequence signals are input into the time sequence distribution convolution feature extraction module in a time distribution manner; The time sequence distribution convolution feature extraction module adopts a one-dimensional convolutional neural network structure with shared parameters to perform convolution operation on each time window in the input sequence, extract nonlinear features, and perform dimension compression; Each time window high-dimensional original signal is mapped into a low-dimensional feature vector, and all window feature vectors are arranged in time sequence to form the multi-channel time sequence feature sequence.

4. The time distribution and hybrid attention based milling chatter prediction method of claim 1, wherein, Inputting the multi-channel time sequence feature sequence into a global time sequence attention modeling unit, modeling the long-range dependence relationship between different historical time windows by a self-attention mechanism, and outputting global dynamic evolution features, comprising: The multi-channel time sequence features are input into a prediction-driven global time sequence attention modeling unit, the correlation between the multi-channel signals in the time dimension is modeled by a parallel time sequence attention computing mechanism, and the long-term dynamic characteristics in the vibration evolution process are captured; In the time sequence attention modeling process, the vibration occurrence probability in the future prediction time window is used as a supervision signal, a prediction target constraint mechanism is introduced in the attention weight generation stage to modulate the attention distribution, and global dynamic evolution features are output.

5. The time distribution and hybrid attention based milling chatter prediction method of claim 1, wherein, The time dimension self-attention module is used to calculate the relative importance weight between different time segments in the same sensor channel, and the key time segment most sensitive to vibration prediction is screened out.

6. The time distribution and hybrid attention based milling chatter prediction method of claim 1, wherein, The channel attention module is used to calculate the importance weight between different sensor channels in the same time segment under the constraint of key time segments, and realize the cooperation and adaptive re-labeling of multi-sensor channels.

7. A milling chatter prediction system based on time distribution and hybrid attention, characterized in that, The method comprises the following steps: A data acquisition and preprocessing module is configured to acquire multi-channel original time sequence signals in the milling process and perform preprocessing; A time sequence feature sequence extraction module is configured to input the preprocessed original time sequence signals into a time distribution convolution feature extraction module, independently perform nonlinear feature extraction and dimension reduction on each time window using a one-dimensional convolution structure with shared weights, and obtain a multi-channel time sequence feature sequence; A global dynamic evolution feature extraction module is configured to input the multi-channel time sequence feature sequence into a global time sequence attention modeling unit, model the long-range dependency between different historical time windows through a self-attention mechanism, and output global dynamic evolution features; A fusion feature acquisition module is configured to input the global dynamic evolution features output by the global time sequence attention modeling unit into a time dimension self-attention module and a channel attention module in sequence, respectively perform key time segment screening and multi-sensor channel importance re-labeling, and obtain attention-enhanced fusion features; A chatter prediction output module is configured to input the fusion features into a classifier and output the probability of chatter occurring in a future time window.

8. A computer-readable storage medium having stored thereon a program, characterized in that, The program is executed by the processor to realize the steps in the milling chatter prediction method based on time distribution and hybrid attention according to any one of claims 1-6.

9. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized by The processor executes the program to realize the steps in the milling chatter prediction method based on time distribution and hybrid attention according to any one of claims 1-6.

10. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instruction is executed by the processor to realize the steps in the milling chatter prediction method based on time distribution and hybrid attention according to any one of claims 1-6.

Images