A tool posture changing thin-walled five-axis milling chatter identification method and system

Abstract

This invention discloses a chatter recognition method and system for five-axis milling with variable tool pose in thin-walled milling. It solves the problem that existing chatter recognition methods often rely on a single path and suffer from unstable features, thus improving the accuracy of chatter recognition. The specific solution is as follows: A chatter recognition method for five-axis milling with variable tool pose in thin-walled milling includes conducting a milling experiment with variable tool pose and performing multi-channel signal acquisition and processing to obtain sample segments; performing multi-scale synchronous squeezing wavelet transform on each sample segment to map the one-dimensional time-domain signal into a two-dimensional time-frequency representation; constructing a network model oriented towards the time-frequency image; performing successive variational mode decomposition on each sample segment to decompose the original signal into several sets of intrinsic mode functions; constructing a multi-scale feature fusion attention network oriented towards the variational mode decomposition sequence; constructing a fusion network module, and outputting the fused global chatter feature vector and the final milling state recognition result.

Description

Technical Field

[0001] This invention relates to the field of thin-walled workpiece machining technology, and in particular to a method and system for identifying chatter in five-axis milling of thin-walled workpieces with variable tool position. 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] Thin-walled parts are widely used in the manufacturing of high-end equipment in the aerospace field. Due to their low structural stiffness, dense modalities, and dynamic changes with cutting position and attitude, they are prone to self-excited vibration, or chatter, during milling. Chatter not only reduces machining quality and accelerates tool wear, but can even damage machine tool accuracy in severe cases. Compared to integral rigid parts, thin-walled parts exhibit significant "time-varying system" characteristics during machining: as material is removed, local wall thickness, boundary conditions, and equivalent stiffness continuously change, making traditional methods based on fixed dynamic parameters, such as stable lobe diagrams, difficult to maintain effectiveness under all working conditions. Furthermore, thin-walled parts often employ five-axis variable-pose machining to ensure tool accessibility and surface quality. Changes in tool axis attitude further alter the cutting force direction and system coupling modes, making chatter boundaries and characteristics more complex. Therefore, to improve milling quality, it is necessary to detect chatter during the milling process of thin-walled parts for early prevention.

[0004] Currently, researchers have proposed various chatter detection techniques. Traditional chatter detection typically relies on artificial features such as spectral peaks, envelope spectra, time-frequency graph textures, or empirical thresholds. However, these methods are sensitive to changes in operating conditions, noise interference, and minor chatter (early chatter), and their generalization ability is limited. With the development of sensor and data acquisition technologies, multi-source signals such as acceleration, sound, and spindle current can more comprehensively characterize cutting dynamics, providing a foundation for data-driven chatter recognition. In recent years, feature construction methods based on time-frequency transformation and signal decomposition have achieved more accurate feature extraction from complex non-stationary cutting signals, showing significant potential in improving recognition accuracy and robustness.

[0005] However, chatter identification in thin-walled components still faces several challenges: 1) Existing flutter identification methods are mostly based on a single path of "data processing → classification", which makes it difficult to take into account both "global time-frequency energy evolution" and "coupling of various components". Under complex working conditions and noise interference, they are prone to problems such as unstable features and increased misjudgment.

[0006] 2) Research on the processing state of curved thin-walled parts is still at the stage of physical modeling and simulation. However, due to the complex structural stiffness coupling relationship and contact conditions, the processing process is highly random and it is difficult to accurately describe the processing process using physical methods.

[0007] 3) Most studies on chatter recognition focus on vertical, thin-walled plates, which do not require consideration of changes in tool position. However, in actual machining, most parts are curved and thin-walled, and the tool position changes constantly, leading to discrepancies between the results of chatter recognition and the actual situation, resulting in low recognition accuracy. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the purpose of this invention is to provide a chatter recognition method for thin-walled five-axis milling with variable tool pose. This method uses milling data with variable tool pose for training, enabling accurate and efficient chatter recognition under complex working conditions.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solution: A method for chatter identification in five-axis milling of thin-walled steel with variable tool pose includes the following: A variable tool position milling experiment was conducted, and multi-channel signal acquisition and processing were performed to obtain sample segments; For each sample segment, a multi-scale synchronous squeezing wavelet transform (MSST) is performed to map the one-dimensional time-domain signal into a two-dimensional time-frequency representation. A network model for time-frequency images is constructed, which takes the multi-scale synchronous squeezed wavelet transform time-frequency image as input and outputs the corresponding intermediate feature vector and branch prediction results. Successive variational mode decomposition (SVMD) is performed on each sample segment to decompose the original signal into a set of intrinsic mode functions (IMFs) to highlight the structural information of different frequency bands / modes. A multi-scale feature fusion attention network for variational mode decomposition sequences is constructed. The modal component sequence or its derived features output by variational mode decomposition are used as inputs, and the outputs are the intermediate feature vectors of the variational mode decomposition branches and the branch prediction results. A fusion network module is constructed, which takes the global time-frequency feature vector output by the multi-scale synchronous extrusion branch and the intermediate feature vector output by the variational mode decomposition branch as input, and outputs the fused global chatter feature vector and the final milling state recognition result.

[0010] The above-described method for chatter recognition in thin-walled five-axis milling with variable tool pose involves cutting the acquired multi-channel signal into sample segments by dividing them into fixed-length windows, and performing basic preprocessing on the sample segments, including DC removal, normalization / standardization, and outlier removal, to provide a unified input for subsequent feature transformation.

[0011] The above-described method for chatter recognition in thin-walled five-axis milling with variable tool pose involves acquiring multi-channel signals through sensors, including a speed sensor, a distance sensor, an acceleration sensor, and a microphone. The speed sensor and acceleration sensor are mounted on the workpiece, while the microphone and distance sensor are mounted on one side of the workpiece. The multi-channel signals include acceleration signals and sound signals. The spindle sensor and cutting depth corresponding to the multi-channel signals are obtained through the speed sensor and distance sensor, and the corresponding pose number is also obtained.

[0012] The chatter recognition method for thin-walled five-axis milling with variable tool pose, as described above, involves performing multi-scale synchronous squeezing wavelet transform on each sample segment, including the following: For channel signals To perform a continuous wavelet transform, first start with the mother wavelet. Construction Scale - Translation Wavelet Family:

[0013] Channel signal Projecting onto this wavelet family yields complex coefficients that vary with scale and time:

[0014] in, As a scale factor, For time translation, For the mother wavelet, Indicates conjugate. For channel signals, For integration variables; Instantaneous frequency estimation using the expression for continuous wavelet transform via fiber transform:

[0015] in, Indicates time translation The partial derivative operation, This indicates taking the imaginary part. These are the complex coefficients obtained from continuous wavelet transform; The expressions for continuous wavelet transform and the estimation expressions for instantaneous frequency are synchronously squeezed and redistributed to the frequency axis. superior:

[0016] in, Let be the Dirac function, representing the... The energy at that location is "squeezed" into the corresponding superior, These are commonly used scale weighting terms; Transforming multiple scale factors The synchronous squeezing wavelet transform results at different scales were obtained, and then the average of multiple sets of results was calculated to form the final MSST representation:

[0017] in, Configure the quantity according to the scale. For the first The complex coefficients of the continuous wavelet transform are calculated using group scale configuration. Estimate its corresponding instantaneous frequency.

[0018] The above-described method for chatter recognition in thin-walled five-axis milling with variable tool pose includes a network model oriented towards time-frequency images comprising a shallow feature extraction unit, a multi-level attention residual enhancement unit, and a feature dimension projection unit connected in sequence. The shallow feature extraction unit consists of a convolutional layer, a normalization layer, a ReLU activation layer, and a max pooling layer connected in sequence. The multi-stage attention residual enhancement unit includes four cascaded residual stages and corresponding CBAM attention modules. The feature dimension projection unit consists of an adaptive global average pooling layer and a fully connected layer connected in sequence.

[0019] The above-described method for chatter recognition in thin-walled five-axis milling with variable tool pose includes a CBAM attention module comprising a parallel channel attention submodule and a spatial attention submodule. The channel attention submodule focuses on mining features from multiple channels and reweighting the importance of different channels through dual-path feature aggregation using global average pooling and max pooling. The spatial attention submodule generates a spatial weight map by performing compressed convolution in the channel dimension, focusing on mining features from the time-frequency map. The adaptive global average pooling layer is used to compress the final residual feature map into a 512-dimensional global feature vector, and the fully connected layer acts as a projection operator to further map the 512-dimensional vector into a 64-dimensional compact feature representation.

[0020] The chatter identification method for thin-walled five-axis milling with variable tool pose, as described above, involves performing successive variational mode decomposition on each sample segment to decompose the original signal into a set of intrinsic mode functions, including the following: Channel signal Represented as several narrowband modes The superposition of modes, each revolving around a certain center frequency Concentrated; unlike variational mode decomposition, which solves all modes simultaneously in one go, successive variational mode decomposition solves the residuals by... Successive decomposition, solving for only one new mode at each step. and its center frequency And update the residuals until the stopping criterion is met.

[0021] The chatter recognition method for thin-walled five-axis milling with variable tool pose, as described above, includes the following steps in the classification feature learning and recognition process using a multi-scale feature fusion attention network: Multi-scale local feature encoding: Multi-scale local feature extraction units are used to perform weight-sharing feature mapping on multiple input sequence units; each sequence unit is sequentially passed through a multi-scale parallel convolution structure, using one-dimensional convolution kernels with different receptive scales to synchronously capture local flutter evolution features of different frequency bands, and combined with residual connection branches to ensure high-fidelity transmission of deep spatial features; The encoded feature vector sequence is fed into the cross-attention aggregation unit; by introducing a learnable global query vector into the network, the correlation weight between the global query vector and each sequence unit is calculated using the multi-head cross-attention mechanism, thereby realizing the dynamic weighted fusion of key chatter-sensitive information under complex milling conditions. The single feature vector obtained after aggregation is output as the intermediate feature vector of the variational mode decomposition branch, so that the subsequent feature fusion network can perform intermediate feature splicing and fusion. At the same time, the intermediate feature vector is fed into the classification prediction output unit of the variational mode decomposition branch. After passing through the random dropout layer to suppress overfitting, it is transformed into the target classification space through the linear mapping layer, and the preliminary flutter state prediction result for the variational mode decomposition branch is output.

[0022] The above-described method for chatter recognition in five-axis thin-wall milling with variable tool pose, wherein the fusion network module is constructed, taking the global time-frequency feature vector output by the multi-scale synchronous extrusion branch and the intermediate feature vector output by the variational mode decomposition branch as input, and outputting the fused global chatter feature vector and the final milling state recognition result, including the following: The heterogeneous features from different coding branches are preliminarily processed using the multi-source feature mapping and dimension alignment module; The mapped two feature vectors are fed into the bidirectional gated feature dynamic recalibration module to achieve dynamic filtering and enhancement of cross-modal information; The two recalibrated feature vectors are concatenated dimensionally to form a high-order fusion feature vector. The high-order fusion feature vector is then fed into the global classification output unit. After improving the model's generalization through a random dropout layer, the features are mapped to the chatter state space using a linear mapping layer. Finally, the milling chatter recognition result is output through the Softmax function.

[0023] Secondly, the present invention also discloses a chatter recognition system for thin-walled five-axis milling with variable tool pose, comprising a computing device configured to: A variable tool position milling experiment was conducted, and multi-channel signal acquisition and processing were performed to obtain sample segments; For each sample segment, a multi-scale synchronous squeezing wavelet transform is performed to map the one-dimensional time-domain signal into a two-dimensional time-frequency representation. A network model for time-frequency images is constructed, which takes the multi-scale synchronous squeezed wavelet transform time-frequency image as input and outputs the corresponding intermediate feature vector and branch prediction results. Successive variational mode decomposition is performed on each sample segment to decompose the original signal into a set of intrinsic mode functions, so as to highlight the structural information of different frequency bands / modes; A multi-scale feature fusion attention network for variational mode decomposition sequences is constructed. The modal component sequence or its derived features output by variational mode decomposition are used as inputs, and the outputs are the intermediate feature vectors of the variational mode decomposition branches and the branch prediction results. A fusion network module is constructed, which takes the global time-frequency feature vector output by the multi-scale synchronous extrusion branch and the intermediate feature vector output by the variational mode decomposition branch as input, and outputs the fused global chatter feature vector and the final milling state recognition result.

[0024] The beneficial effects of the present invention are as follows: 1) The method provided by this invention conducts milling experiments with variable tool pose and performs multi-channel signal acquisition and processing. The overall processing is divided into two paths: one path extracts time-frequency features from sample segments through multi-scale synchronous squeezing wavelet transform, and the other path extracts modal features from sample segments through successive variational mode decomposition. The two paths complement each other and cooperate in discrimination, which is no longer a single-path recognition. It effectively takes into account the expression of multi-size features and effectively improves the accuracy, stability and robustness of chatter recognition under complex noise background and working condition fluctuation conditions.

[0025] 2) The multi-channel signals obtained by this invention are not obtained through physical modeling and simulation, but through actual experiments. They can automatically mine and learn the dynamic response differences and potential correlation information corresponding to different tool pose changes from multi-source machining signals, thereby enhancing the adaptability to the variable working conditions of actual curved thin-walled parts. The multi-channel signals are obtained through multiple sensors, which are consistent with the characteristics of real objects such as curved thin-walled parts, taking into account the changes in tool pose.

[0026] 3) The overall method of this invention, compared with traditional methods that rely on physical modeling or single-path feature extraction, can more timely and accurately characterize changes in processing state through online signal recognition, reduce the reliance on complex mechanism modeling, and improve the engineering application value of the method in curved thin-walled parts and various variable pose processing scenarios. Attached Figure Description

[0027] 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.

[0028] Figure 1 This is a flowchart of a method for identifying chatter in thin-walled five-axis milling with variable tool position, according to one or more embodiments of the present invention.

[0029] Figure 2 This is a schematic diagram of the network model oriented towards time-frequency images in a variable tool pose thin-walled five-axis milling chatter recognition method according to one or more embodiments of the present invention.

[0030] Figure 3 This is a schematic diagram of the multi-scale feature fusion attention network in a variable tool pose thin-walled five-axis milling chatter recognition method according to one or more embodiments of the present invention.

[0031] Figure 4 This is a schematic diagram of the fusion network module in a variable tool pose thin-walled five-axis milling chatter recognition method according to one or more embodiments of the present invention.

[0032] Figure 5 This is a schematic diagram of the installation positions of some sensors and workpieces in a variable tool position thin-walled five-axis milling chatter recognition method according to one or more embodiments of the present invention.

[0033] Figure 6 These are experimental images of multiple datasets collected in a variable tool pose thin-walled five-axis milling chatter recognition method according to one or more embodiments of the present invention.

[0034] Figure 7 This is a result image of the multi-scale synchronous extrusion wavelet transform (MSST) in a chatter identification method for thin-walled five-axis milling with variable tool pose according to one or more embodiments of the present invention.

[0035] Figure 8 This is a diagram showing the training (confusion matrix) results of a deep learning network in a chatter recognition method for thin-walled five-axis milling with variable tool pose according to one or more embodiments of the present invention.

[0036] Figure 9 This invention relates to a method for identifying chatter in thin-walled five-axis milling with variable tool pose, based on one or more embodiments. The method involves performing successive variational mode decomposition on each sample segment, decomposing the original signal into several sets of intrinsic mode functions to highlight the structural information of different frequency bands / modes.

[0037] Figure 10This is a training result diagram of a multi-scale feature fusion attention network for SVMD decomposition sequence constructed in a chatter recognition method for thin-walled five-axis milling with variable tool pose according to one or more embodiments of the present invention.

[0038] Figure 11 This is a training result diagram of the fusion network module constructed in a variable tool pose thin-walled five-axis milling chatter recognition method according to one or more embodiments of the present invention. Detailed Implementation

[0039] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0040] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless otherwise expressly indicated by the invention, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. As introduced in the background section, most existing chatter identification methods follow a single path of data processing → classification, which leads to unstable features under complex working conditions. To address these technical issues, this invention proposes a chatter identification method for thin-walled five-axis milling with variable tool pose.

[0041] Example 1 In a typical embodiment of the present invention, reference is made to Figure 1 As shown, a chatter identification method for thin-walled five-axis milling with variable tool pose includes the following: Step 1: Variable Tool Pose Milling Experiment and Multi-channel Signal Acquisition and Processing A variable tool position milling experiment was conducted, acquiring multi-channel signals using sensors including a speed sensor, a distance sensor, an acceleration sensor, and a microphone. The speed sensor and acceleration sensor were mounted on the workpiece for reference. Figure 5 As shown, the workpiece mounting direction and the tool axis are spaced at an angle. A microphone and distance sensor are mounted on one side of the workpiece. During machining, all sensors synchronously acquire multi-channel sensing signals, including acceleration and sound signals; and obtain the corresponding operating condition information for each signal segment, including spindle speed, depth of cut, and pose number (the pose number is obtained from the tool's encoder, which is connected to the control unit). (Refer to...) Figure 6 As shown, the original dataset is formed, referencing... Figure 7As shown, a comparison chart of frequencies under different tool poses with the same milling parameters can be obtained.

[0042] The continuously acquired signal is cut into fixed-length windows (with overlapping sliding windows) to obtain sample segments; and the sample segments are preprocessed, including DC removal, normalization / standardization, and outlier removal, to provide a unified input for the feature transformation of the two subsequent branches.

[0043] Step 2: Perform multi-scale synchronous squeeze wavelet transform (MSST) on each sample segment obtained in Step 1 to map the one-dimensional time-domain signal into a two-dimensional time-frequency representation.

[0044] During the MSST process: For channel signals To perform a continuous wavelet transform, first start with the mother wavelet. Construction Scale - Translation Wavelet Family:

[0045] Channel signal Projecting onto this wavelet family yields complex coefficients that vary with scale and time:

[0046] in, As a scale factor, For time translation, For the mother wavelet, Indicates conjugate. For channel signals, For integration variables; Instantaneous frequency estimation using the expression for continuous wavelet transform via fiber transform:

[0047] in, Indicates time translation The partial derivative operation, This indicates taking the imaginary part. These are the complex coefficients obtained from continuous wavelet transform; The expressions for continuous wavelet transform and the estimation expressions for instantaneous frequency are synchronously squeezed and redistributed to the frequency axis. superior:

[0048] in, Let be the Dirac function, representing the... The energy at that location is "squeezed" into the corresponding superior, These are commonly used scale weighting terms; Transforming multiple scale factors The synchronous squeezing wavelet transform results at different scales were obtained, and then the average of multiple sets of results was calculated to form the final MSST representation:

[0049] in, Configure the quantity according to the scale. For the first The complex coefficients of the continuous wavelet transform are calculated using group scale configuration. Its corresponding instantaneous frequency estimation; as Increased frequency energy usually results in more concentrated time-frequency energy, clearer texture and energy bands, which is beneficial for subsequent convolutional neural networks (CNNs) to learn flutter-related features.

[0050] Step 3: Construct a network model for time-frequency images, refer to... Figure 2 As shown, taking the MSST time-frequency image as input, the output is the intermediate feature vector and branch prediction results. The network model for time-frequency images includes a deep neural network model (ResNet18) backbone network and an attention mechanism module (Convolutional Block Attention Module, CBAM). The network model for time-frequency images includes sequentially connected shallow feature extraction units, multi-order attention residual enhancement units, and feature dimension projection units. The shallow feature extraction unit consists of a convolutional layer, a normalization layer, a ReLU (Rectified Linear Unique Function) activation layer, and a max pooling layer connected in sequence; wherein, the convolutional layer employs... Convolutional kernels are used to perform preliminary spatial downsampling and feature channel mapping on the input three-channel processed image; The multi-level attention residual enhancement unit comprises four cascaded residual stages (Layer 1 to Layer 4) and their respective CBAM attention modules. The CBAM attention module further includes parallel channel attention sub-modules and spatial attention sub-modules. The channel attention sub-module focuses on mining features from multiple channels and reweighting the importance of different channels through dual-path feature aggregation using global average pooling and max pooling. The spatial attention sub-module generates a spatial weight map by performing compressed convolution in the channel dimension, focusing on mining features from the time-frequency map. The feature dimension projection unit consists of an adaptive global average pooling layer and a fully connected layer connected sequentially. The adaptive global average pooling layer compresses the final residual feature map into a 512-dimensional global feature vector. The fully connected layer acts as a projection operator, further mapping the 512-dimensional vector into a compact 64-dimensional feature representation. This approach can reduce the computational complexity and overfitting risk of the subsequent tremor recognition classifier while ensuring feature robustness. Figure 7This is the result image after Multi-Scale Synchronous Squeezed Wavelet Transform (MSST). It is used as input to train the deep learning network proposed in step three, resulting in a confusion matrix. Figure 8 .

[0051] Step 4: Perform successive variational mode decomposition (SVMD) on each sample segment obtained in Step 1 to decompose the original signal into a set of several intrinsic mode functions (IMFs) to highlight the structural information of different frequency bands / modes.

[0052] SVMD is an adaptive decomposition method improved upon variational mode decomposition (VMD), specifically representing the channel signal as several narrowband modes. The superposition of modes, each revolving around a certain center frequency Concentrated; unlike VMD, which solves all modes simultaneously in one go, SVMD addresses the residuals by... Successive decomposition, solving for only one new mode at each step. and its center frequency The residuals are updated until the stopping criterion is met. For non-stationary, aliased noise signals such as milling signals, SVMD achieves adaptive decomposition by successively extracting narrowband modal components from the residual signal. Compared with one-time decomposition methods, it has less dependence on parameters such as the number of modes, is more robust to noise and operating condition fluctuations, and is less prone to mode aliasing. At the same time, it can more clearly separate the tooth-related components such as harmonics, structural resonances, and chatter in the frequency band. Step four mainly involves decomposing the signal. Figure 9 This is a visualization of the decomposition results of stable and flutter signals, respectively. The specific decomposition data is used as input into the deep learning model in step five.

[0053] Step 5: Construct a multi-scale feature fusion attention network for SVMD decomposition sequences. The network takes the modal component sequences output by SVMD or their derived features (such as statistical, energy, and frequency domain features) as input, and outputs the intermediate feature vectors of the SVMD branches and the branch prediction results. (Reference) Figure 3 As shown, the multi-scale feature fusion attention network sequentially comprises a multi-scale local feature extraction unit, a cross-attention aggregation unit, and a classification prediction output unit. The specific classification feature learning and recognition process of the multi-scale feature fusion attention network is as follows: 5-1) Local multi-scale feature encoding: Multi-scale local feature extraction units are used to perform weight-sharing feature mapping on multiple input sequence units; each sequence unit is sequentially passed through a multi-scale parallel convolution structure, using one-dimensional convolution kernels of different receptive field scales (such as convolution kernels of lengths 9, 19 and 39) to synchronously capture local flutter evolution features of different frequency bands, and combined with residual connection branches to ensure high-fidelity transmission of deep spatial features, thereby independently encoding each sequence unit into a high-dimensional feature vector.

[0054] 5-2) Dynamic aggregation of global correlation features: The encoded feature vector sequence is fed into the cross-attention aggregation unit; by introducing a learnable global query vector into the network, the correlation weight between the global query vector and each sequence unit is calculated using the multi-head cross-attention mechanism, so as to realize the dynamic weighted fusion of key chatter sensitive information under complex milling conditions, thereby compressing the feature vector sequence into a single milling state feature vector with global representativeness.

[0055] 5-3) SVMD Branch Features and Preliminary Prediction Output: The single feature vector obtained after aggregation is output as the intermediate feature vector of this SVMD branch for subsequent feature fusion network to perform intermediate feature splicing and fusion. Simultaneously, this intermediate feature vector is fed into the classification prediction output unit of this branch, passing through a random dropout layer to suppress overfitting, and then transformed to the target classification space via a linear mapping layer, outputting the preliminary flutter state prediction result for this single branch. (Refer to...) Figure 10 The image shows the training results after constructing a multi-scale feature fusion attention network for SVMD decomposition sequences.

[0056] Step Six: Construct the converged network module, refer to Figure 4 As shown, the global time-frequency feature vector output from the MSST branch in step three and the intermediate feature vector output from the SVMD branch in step five are used as inputs, and the outputs are the fused global chatter feature vector and the final milling state recognition result. The fusion network module sequentially includes a feature dimension alignment unit, a bidirectional gated feature recalibration unit, and a global classification output unit. The specific feature fusion and state recognition process of the network is as follows: 6-1) The multi-source feature mapping and dimension alignment module is used to perform preliminary processing on heterogeneous features from different coding branches. The MSST feature vector and the SVMD intermediate feature vector are respectively normalized through their own independent layer normalization layers to eliminate the magnitude differences caused by different signal processing methods. Subsequently, a linear projection layer is used to map the heterogeneous features to a unified implicit feature space, laying the foundation for subsequent feature interaction.

[0057] 6-2) The mapped two feature vectors are sent to the bidirectional gated feature dynamic recalibration module to realize dynamic filtering and enhancement of cross-modal information.

[0058] Gated weight generation: Using SVMD branch features, a small gated network (containing linear mapping and Sigmoid activation function) is used to generate weight vectors for MSST features; at the same time, weight vectors for SVMD features are generated synchronously using MSST branch features.

[0059] Feature cross-mapping: By using element-wise multiplication, the gating weights generated from one branch are used to nonlinearly recalibrate the features of another branch. This mechanism allows the network to adaptively suppress feature components severely affected by noise and enhance chatter-sensitive features with high discriminative power, depending on the complexity of the milling process.

[0060] 6-3) The two recalibrated feature vectors are concatenated dimensionally to form a higher-order fused feature vector.

[0061] Intermediate feature fusion: The concatenated vector is fed into the fusion classification head, and the deep nonlinear correlation between cross-branch features is further explored through the multilayer perceptron structure, thereby aggregating them into a single global fusion feature vector with strong discriminative ability.

[0062] Final classification: The global feature vector is synchronously fed into the global classification output unit. After improving the model's generalization ability through a random dropout layer, a linear mapping layer is used to map the features to the chatter state space. The final milling chatter recognition result is output through the Softmax (normalized exponential function) function. Figure 11 As shown, the training results after constructing the fusion network module are presented.

[0063] The method provided in this embodiment is trained and modeled based on a variable tool pose machining experimental dataset. This enables the model to automatically mine and learn the dynamic response differences and potential correlation information corresponding to different tool pose changes from multi-source machining signals, thereby enhancing its adaptability to the variable working conditions of actual curved thin-walled parts. Simultaneously, a dual-branch fusion framework is adopted. One branch focuses on characterizing the global time-frequency energy evolution characteristics of the machining signal, while the other branch focuses on depicting the local structural features and coupling relationships of each modal component after signal decomposition. These two branches complement and coordinate the discrimination, effectively improving the accuracy, stability, and robustness of chatter identification under complex noise backgrounds and fluctuating working conditions. Compared to traditional methods that rely on physical modeling or single-path feature extraction, this method can more timely and accurately characterize machining state changes through online signal recognition, reducing reliance on complex mechanism modeling and enhancing the engineering application value of the method in curved thin-walled parts and various variable pose machining scenarios.

[0064] Example 2 This embodiment discloses a chatter recognition system for thin-walled milling with variable tool pose, including a computing device configured as follows: A variable tool position milling experiment was conducted, and multi-channel signal acquisition and processing were performed to obtain sample segments; For each sample segment, a multi-scale synchronous squeezing wavelet transform is performed to map the one-dimensional time-domain signal into a two-dimensional time-frequency representation. A network model for time-frequency images is constructed, which takes the multi-scale synchronous squeezed wavelet transform time-frequency image as input and outputs the corresponding intermediate feature vector and branch prediction results. Successive variational mode decomposition is performed on each sample segment to decompose the original signal into a set of intrinsic mode functions, so as to highlight the structural information of different frequency bands / modes; A multi-scale feature fusion attention network for variational mode decomposition sequences is constructed. The modal component sequence or its derived features output by variational mode decomposition are used as inputs, and the outputs are the intermediate feature vectors of the variational mode decomposition branches and the branch prediction results. A fusion network module is constructed, which takes the global time-frequency feature vector output by the multi-scale synchronous extrusion branch and the intermediate feature vector output by the variational mode decomposition branch as input, and outputs the fused global chatter feature vector and the final milling state recognition result.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A variable tool pose thin-walled five-axis milling chatter identification method, characterized in that, Includes the following: A variable tool position milling experiment was conducted, and multi-channel signal acquisition and processing were performed to obtain sample segments; For each sample segment, a multi-scale synchronous squeezing wavelet transform is performed to map the one-dimensional time-domain signal into a two-dimensional time-frequency representation. A network model for time-frequency images is constructed, which takes the multi-scale synchronous squeezed wavelet transform time-frequency image as input and outputs the corresponding intermediate feature vector and branch prediction results. Successive variational mode decomposition is performed on each sample segment to decompose the original signal into a set of intrinsic mode functions, so as to highlight the structural information of different frequency bands / modes; A multi-scale feature fusion attention network for variational mode decomposition sequences is constructed. The modal component sequence or its derived features output by variational mode decomposition are used as inputs, and the outputs are the intermediate feature vectors of the variational mode decomposition branches and the branch prediction results. A fusion network module is constructed, which takes the global time-frequency feature vector output by the multi-scale synchronous extrusion branch and the intermediate feature vector output by the variational mode decomposition branch as input, and outputs the fused global chatter feature vector and the final milling state recognition result.

2. The tool posture varying thin-walled five-axis milling chatter identification method according to claim 1, characterized in that, The acquired multi-channel signal is cut into sample segments by a fixed-length window, and the sample segments are preprocessed, including DC removal, normalization / standardization, and outlier removal, to provide a unified input for subsequent feature transformation.

3. The tool posture varying thin-walled five-axis milling chatter identification method according to claim 1, characterized in that, The multi-channel signals are acquired by sensors, including a speed sensor, a distance sensor, an acceleration sensor, and a microphone. The speed sensor and acceleration sensor are mounted on the workpiece, and the microphone and distance sensor are mounted on one side of the workpiece. The multi-channel signals include acceleration signals and sound signals. The spindle sensor and cutting depth corresponding to the multi-channel signals are obtained through the speed sensor and distance sensor, and the corresponding pose number is obtained.

4. The tool posture varying thin-walled five-axis milling chatter identification method according to claim 1, characterized in that, The process of performing multi-scale synchronous squeezing wavelet transform on each sample segment includes the following: To the channel signal The continuous wavelet transform is first constructed from a mother wavelet The scale-shifted wavelet family is constructed: The channel signal is projected onto the wavelet family, resulting in complex coefficients that vary with scale and time: wherein, is a scale factor, is a time shift, is a mother wavelet, denotes a conjugate, is a channel signal, is an integration variable; Instantaneous frequency estimation using the expression for continuous wavelet transform via fiber transform: wherein denotes the partial derivative operation with respect to the time translation denotes the partial derivative operation with respect to the time translation denotes taking the imaginary part, are complex coefficients obtained by the continuous wavelet transform; Synchronously squeezing and redistributing the expression of the continuous wavelet transform and the estimated expression of the instantaneous frequency to the frequency axis Top: wherein, is the Dirac function, representing the "squashing" of the energy at to the corresponding , is the usual scale weight term; Transforming multiple scale factors The results of the synchronous extrusion wavelet transform at different scales are obtained, and then a plurality of results are averaged to form the final MSST representation: wherein, is a number of scale configurations, is a first group of scale configurations computes complex coefficients of a continuous wavelet transform, is its corresponding instantaneous frequency estimate.

5. The chatter identification method for thin-walled five-axis milling with variable tool pose according to claim 1, characterized in that, The network model for time-frequency images includes a shallow feature extraction unit, a multi-level attention residual enhancement unit, and a feature dimension projection unit connected in sequence. The shallow feature extraction unit consists of a convolutional layer, a normalization layer, a ReLU activation layer, and a max pooling layer connected in sequence. The multi-stage attention residual enhancement unit includes four cascaded residual stages and corresponding CBAM attention modules. The feature dimension projection unit consists of an adaptive global average pooling layer and a fully connected layer connected in sequence.

6. The chatter identification method for thin-walled five-axis milling with variable tool pose according to claim 5, characterized in that, The CBAM attention module includes a parallel channel attention submodule and a spatial attention submodule. The channel attention submodule focuses on mining features from multiple channels and reweighting the importance of different channels through dual-path feature aggregation using global average pooling and max pooling. The spatial attention submodule generates a spatial weight map by performing compressed convolution in the channel dimension, focusing on mining features from the time-frequency map. The adaptive global average pooling layer is used to compress the final residual feature map into a 512-dimensional global feature vector, and the fully connected layer acts as a projection operator to further map the 512-dimensional vector into a 64-dimensional compact feature representation.

7. The chatter identification method for thin-walled five-axis milling with variable tool pose according to claim 1, characterized in that, The step of performing successive variational mode decomposition on each sample segment, decomposing the original signal into a set of several intrinsic mode functions, includes the following: Channel signal Represented as several narrowband modes The superposition of modes, each revolving around a certain center frequency Concentrated; unlike variational mode decomposition, which solves all modes simultaneously in one go, successive variational mode decomposition solves the residuals by... Successive decomposition, solving for only one new mode at each step. and its center frequency And update the residuals until the stopping criterion is met.

8. The chatter identification method for thin-walled five-axis milling with variable tool position according to claim 1, characterized in that, The classification feature learning and recognition process of the multi-scale feature fusion attention network includes the following: Multi-scale local feature encoding: Multi-scale local feature extraction units are used to perform weight-sharing feature mapping on multiple input sequence units; each sequence unit is sequentially passed through a multi-scale parallel convolution structure, using one-dimensional convolution kernels with different receptive scales to synchronously capture local flutter evolution features of different frequency bands, and combined with residual connection branches to ensure high-fidelity transmission of deep spatial features; The encoded feature vector sequence is fed into the cross-attention aggregation unit; by introducing a learnable global query vector into the network, the correlation weight between the global query vector and each sequence unit is calculated using the multi-head cross-attention mechanism, thereby realizing the dynamic weighted fusion of key chatter-sensitive information under complex milling conditions. The single feature vector obtained after aggregation is output as the intermediate feature vector of the variational mode decomposition branch, so that the subsequent feature fusion network can perform intermediate feature splicing and fusion. At the same time, the intermediate feature vector is fed into the classification prediction output unit of the variational mode decomposition branch. After passing through the random dropout layer to suppress overfitting, it is transformed into the target classification space through the linear mapping layer, and the preliminary flutter state prediction result for the variational mode decomposition branch is output.

9. The chatter identification method for thin-walled five-axis milling with variable tool pose according to claim 1, characterized in that, The fusion network module takes the global time-frequency feature vector output by the multi-scale synchronous extrusion branch and the intermediate feature vector output by the variational mode decomposition branch as input, and outputs the fused global chatter feature vector and the final milling state recognition result, including the following: The heterogeneous features from different coding branches are preliminarily processed using the multi-source feature mapping and dimension alignment module; The mapped two feature vectors are fed into the bidirectional gated feature dynamic recalibration module to achieve dynamic filtering and enhancement of cross-modal information; The two recalibrated feature vectors are concatenated dimensionally to form a high-order fusion feature vector. The high-order fusion feature vector is then fed into the global classification output unit. After improving the model's generalization through a random dropout layer, the features are mapped to the chatter state space using a linear mapping layer. Finally, the milling chatter recognition result is output through the Softmax function.

10. A chatter recognition system for thin-walled five-axis milling with variable tool position, characterized in that, Includes a computing device, which is configured as follows: A variable tool position milling experiment was conducted, and multi-channel signal acquisition and processing were performed to obtain sample segments; For each sample segment, a multi-scale synchronous squeezing wavelet transform is performed to map the one-dimensional time-domain signal into a two-dimensional time-frequency representation. A network model for time-frequency images is constructed, which takes the multi-scale synchronous squeezed wavelet transform time-frequency image as input and outputs the corresponding intermediate feature vector and branch prediction results. Successive variational mode decomposition is performed on each sample segment to decompose the original signal into a set of intrinsic mode functions, so as to highlight the structural information of different frequency bands / modes; A multi-scale feature fusion attention network for variational mode decomposition sequences is constructed. The modal component sequence or its derived features output by variational mode decomposition are used as inputs, and the outputs are the intermediate feature vectors of the variational mode decomposition branches and the branch prediction results. A fusion network module is constructed, which takes the global time-frequency feature vector output by the multi-scale synchronous extrusion branch and the intermediate feature vector output by the variational mode decomposition branch as input, and outputs the fused global chatter feature vector and the final milling state recognition result.

Images