A numerical control milling chatter online monitoring method based on residual connection and joint attention mechanism

Abstract

The application discloses a numerical control milling chatter online monitoring method based on residual connection and joint attention mechanism, which comprises the following steps: 1) collecting three-direction cutting forces and three-direction accelerations of a tool in a numerical control milling system; 2) processing the three-direction cutting forces and three-direction accelerations to obtain frequency domain features of the cutting forces and cutting vibrations; 3) based on the joint attention mechanism of the residual connection, adaptively fusing the frequency domain features of the cutting forces and cutting vibrations respectively to obtain a cutting force feature vector and a cutting vibration feature vector; 4) constructing a machining state classification model; and 5) inputting the cutting force feature vector and the cutting vibration feature vector into the machining state classification model to obtain a current machining state, thereby realizing numerical control milling chatter monitoring. The application realizes numerical control milling chatter online monitoring by using a normalized exponential function to probabilistically classify the machining state, and the implementation method is convenient and effective, and the accuracy of multi-axis cutting chatter state monitoring is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of mechanical manufacturing process monitoring and control, specifically a method for online monitoring of chatter in CNC milling based on residual connection and joint attention mechanisms. Background Technology

[0002] Complex irregular curved surface parts are widely used in high-end equipment manufacturing fields such as energy and power, and their surface quality has a significant impact on the performance of high-end equipment in these fields. Complex irregular curved surface parts have intricate surface features and are often machined using long overhanging tools in multi-axis cutting. These tools have relatively weak rigidity, and the material removal rate during multi-axis cutting varies over time. The drastic fluctuations in cutting force and acceleration can easily lead to chatter during machining, damaging the surface quality of complex irregular curved surface parts and hindering the improvement of high-end equipment performance. Therefore, researching an online chatter monitoring method for CNC milling based on a joint attention mechanism using residual connections is of great significance for high-quality machining of complex irregular curved surface parts and for promoting the development of high-end manufacturing in fields such as energy and power.

[0003] Existing chatter identification methods have the following drawbacks: 1) They use signal thresholds as the standard for chatter identification. The determination of the threshold depends on machining parameters and human knowledge, resulting in low robustness of chatter monitoring. 2) They only consider a single cutting force signal, leading to low accuracy in cutting chatter identification. Summary of the Invention

[0004] The purpose of this invention is to provide an online monitoring method for chatter in CNC milling based on residual connectivity and joint attention mechanisms, comprising the following steps:

[0005] 1) Collect the three-dimensional cutting force and three-dimensional acceleration of the tool in the CNC milling system;

[0006] 2) Process the triaxial cutting force and triaxial acceleration to obtain the frequency domain characteristics of the cutting force and cutting vibration;

[0007] 3) Based on the joint attention mechanism of residual connection, the frequency domain features of cutting force and cutting vibration are adaptively fused to obtain the cutting force feature vector and the cutting vibration feature vector respectively;

[0008] 4) Construct a processing status classification model;

[0009] 5) Input the cutting force feature vector and the cutting vibration feature vector into the machining state classification model to obtain the current machining state and realize the monitoring of chatter in CNC milling.

[0010] Furthermore, the triaxial cutting force and triaxial acceleration are acquired by a force gauge and an acceleration sensor, respectively.

[0011] Furthermore, the steps for processing the triaxial cutting force and triaxial acceleration include:

[0012] 2.1) By using a sliding window, the triaxial cutting force and triaxial acceleration are divided into several segments of the original signal sequence; the original signal sequence includes the triaxial cutting force signal sequence and the triaxial acceleration signal sequence;

[0013] 2.2) Perform Fast Fourier Transform on each segment of the original signal sequence to obtain the triaxial cutting force and triaxial acceleration after Fast Fourier Transform;

[0014] 2.3) The triaxial cutting force and triaxial acceleration after the fast Fourier transform are standardized to obtain the standardized triaxial cutting force and triaxial acceleration;

[0015] 2.4) The three-dimensional cutting forces after standardization are fused to obtain the frequency domain characteristics of the cutting forces;

[0016] The frequency domain characteristics of the cutting vibration are obtained by fusing the standardized triaxial accelerations.

[0017] Furthermore, the triaxial cutting force and triaxial acceleration after the fast Fourier transform are shown below:

[0018]

[0019] Where x(n) is the time-domain information of the original signal; It represents the Fourier amplitude; n = 0, 1, ..., N-1, where N is the length of the signal sequence. k is a positive integer; i is an imaginary number;

[0020] Furthermore, the standardized triaxial cutting forces and triaxial accelerations are shown below:

[0021]

[0022] Where x represents the triaxial cutting force and triaxial acceleration after fast Fourier transform, z' represents the triaxial cutting force and triaxial acceleration after standardization, and μ and σ are the mean and standard deviation.

[0023] Furthermore, the steps for adaptively fusing the frequency domain characteristics of cutting force and cutting vibration separately include:

[0024] 3.1) Compress the input data X into a vector z∈R using global average pooling. C×1 The input data X is the frequency domain characteristic of the cutting force or the frequency domain characteristic of the cutting vibration.

[0025] The j-th component of vector z is shown below:

[0026]

[0027] In the formula, j = 1, 2, ..., C, where C is the number of channels; N is the length of the signal sequence; x j (n) is the input data for channel j;

[0028] 3.2) By utilizing the joint features of the input data activated by the convolutional layer, we obtain:

[0029]

[0030] In the formula, For joint features; This indicates multiplication by components;

[0031] Among them, the spatial location weight u and the channel features They are shown below:

[0032]

[0033]

[0034] In the formula, σ(·) represents the activation function; w3 and b3 represent the weights and biases, respectively.

[0035] Among them, the weights of the convolutional layer channels As shown below:

[0036]

[0037] In the formula, w1 and w2 represent the weights of the two-channel convolutional layer, b1 and b2 represent the corresponding biases, δ(·) represents the modified linear unit, σ(·) represents the activation function, and z represents the elements of the vector z.

[0038] 3.3) Establish the relationship between the joint features and the input data using residual connections to obtain the feature vector Y, i.e.:

[0039]

[0040] In the formula, when X is the frequency domain characteristic of cutting force, Y represents the characteristic vector of cutting force; when X is the frequency domain characteristic of cutting vibration, Y represents the characteristic vector of cutting vibration.

[0041] Furthermore, the modified linear unit is a ReLU unit.

[0042] Furthermore, the activation function is the Sigmoid function.

[0043] Furthermore, the processing state classification model is a classifier containing only one fully connected layer;

[0044] The loss function FL of the processing state classification model is as follows:

[0045]

[0046] In the formula, α m For coefficients, To represent the true processing state of the training samples, The model predicts the processing state; M is the number of training samples; r is the focusing parameter, r>0.

[0047] Furthermore, the processing state includes a stable state and a chattering state.

[0048] It is worth noting that this invention performs Fast Fourier Transform on the acquired triaxial cutting force and triaxial acceleration signals, standardizes the data according to standard scores, and uses two independent convolutional neural networks to extract the frequency domain features of cutting force and cutting vibration respectively. A joint attention mechanism with residual connections is established to adaptively fuse features from different sensors, forming two three-channel acceleration and cutting force spectrum data sets. The feature maps are converted into one-dimensional feature vectors using global average pooling. The dataset partitioning is optimized using the Focal Loss function, and various machining states are probabilistically classified using a normalized exponential function. Ultimately, online monitoring of chatter in CNC milling based on a joint attention mechanism with residual connections is achieved.

[0049] The technical effects of this invention are undeniable. This invention utilizes the joint attention mechanism of residual connections to adaptively fuse the features of different sensors, and performs probabilistic classification of the machining state through a normalized exponential function to achieve online monitoring of chatter in CNC milling. The implementation method is convenient and effective, and significantly improves the accuracy of chatter monitoring in multi-axis cutting. Attached Figure Description

[0050] Figure 1 This is a flowchart of the online monitoring method for chatter in CNC milling.

[0051] Figure 2 This is a diagram of a two-stream convolutional neural network.

[0052] Figure 3 This is a schematic diagram of the joint attention mechanism for residual connections. Here, C is the number of channels, N is the length of the feature vector components, ReLU is the corrected linear unit, Softmax is the normalized exponential function, and Sigmoid is the binary activation function.

[0053] Figure 4 This is a cutting force signal diagram of a stable cutting state according to an embodiment of the present invention.

[0054] Figure 5 This is a diagram of the cutting force signal under chattering state according to an embodiment of the present invention. Detailed Implementation

[0055] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0056] Example 1:

[0057] A method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism includes the following steps:

[0058] 1) Collect the three-dimensional cutting force and three-dimensional acceleration of the tool in the CNC milling system;

[0059] 2) Process the triaxial cutting force and triaxial acceleration to obtain the frequency domain characteristics of the cutting force and cutting vibration;

[0060] 3) Based on the joint attention mechanism of residual connection, the frequency domain features of cutting force and cutting vibration are adaptively fused to obtain the cutting force feature vector and the cutting vibration feature vector respectively;

[0061] 4) Construct a processing status classification model;

[0062] 5) Input the cutting force feature vector and the cutting vibration feature vector into the machining state classification model to obtain the current machining state and realize the monitoring of chatter in CNC milling.

[0063] The triaxial cutting force and triaxial acceleration are obtained by force measuring instrument and acceleration sensor, respectively.

[0064] The steps for processing the triaxial cutting force and triaxial acceleration include:

[0065] 2.1) By using a sliding window, the triaxial cutting force and triaxial acceleration are divided into several segments of the original signal sequence; the original signal sequence includes the triaxial cutting force signal sequence and the triaxial acceleration signal sequence;

[0066] 2.2) Perform Fast Fourier Transform on each segment of the original signal sequence to obtain the triaxial cutting force and triaxial acceleration after Fast Fourier Transform;

[0067] 2.3) The triaxial cutting force and triaxial acceleration after the fast Fourier transform are standardized to obtain the standardized triaxial cutting force and triaxial acceleration;

[0068] 2.4) The three-dimensional cutting forces after standardization are fused to obtain the frequency domain characteristics of the cutting forces;

[0069] The frequency domain characteristics of the cutting vibration are obtained by fusing the standardized triaxial accelerations.

[0070] The triaxial cutting forces and triaxial accelerations after the Fast Fourier Transform are shown below:

[0071]

[0072] Where x(n) is the time-domain information of the original signal; It represents the Fourier amplitude; n = 0, 1, ..., N-1, where N is the length of the signal sequence. k is a positive integer. i is an imaginary number;

[0073] The standardized triaxial cutting forces and triaxial accelerations are shown below:

[0074]

[0075] Where x represents the triaxial cutting force and triaxial acceleration after fast Fourier transform, z' represents the triaxial cutting force and triaxial acceleration after standardization, and μ and σ are the mean and standard deviation.

[0076] The steps for adaptively fusing the frequency domain characteristics of cutting force and cutting vibration separately include:

[0077] 3.1) Compress the input data X into a vector z∈R using global average pooling. C×1 The input data X is the frequency domain characteristic of the cutting force or the frequency domain characteristic of the cutting vibration.

[0078] The j-th component of vector z is shown below:

[0079]

[0080] In the formula, j = 1, 2, ..., C, where C is the number of channels; N is the length of the signal sequence; x j (n) is the input data for channel j;

[0081] 3.2) By utilizing the joint features of the input data activated by the convolutional layer, we obtain:

[0082]

[0083] In the formula, For joint features; This indicates multiplication by components;

[0084] Among them, the spatial location weight u and the channel features They are shown below:

[0085]

[0086]

[0087] In the formula, σ(·) represents the activation function; w3 and b3 represent the weights and biases, respectively.

[0088] Among them, the weights of the convolutional layer channels As shown below:

[0089]

[0090] In the formula, w1 and w2 represent the weights of the two-channel convolutional layer, b1 and b2 represent the corresponding biases, δ(·) represents the modified linear unit, σ(·) represents the activation function, and z represents the elements of the vector z.

[0091] 3.3) Establish the relationship between the joint features and the input data using residual connections to obtain the feature vector Y, i.e.:

[0092]

[0093] In the formula, when X is the frequency domain characteristic of cutting force, Y represents the characteristic vector of cutting force; when X is the frequency domain characteristic of cutting vibration, Y represents the characteristic vector of cutting vibration.

[0094] The modified linear unit is a ReLU unit.

[0095] The activation function is the Sigmoid function.

[0096] The processing state classification model is a classifier containing only one fully connected layer;

[0097] The loss function FL of the processing state classification model is as follows:

[0098]

[0099] In the formula, α m For coefficients, To represent the true processing state of the training samples, The model predicts the processing state; M is the number of training samples; r is the focusing parameter, r>0.

[0100] The processing states include a stable state and a chattering state.

[0101] Example 2:

[0102] See Figures 1 to 5 A method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism includes the following steps:

[0103] 1) Collect the three-dimensional cutting force and three-dimensional acceleration of the tool in the CNC milling system;

[0104] 2) Process the triaxial cutting force and triaxial acceleration to obtain the frequency domain characteristics of the cutting force and cutting vibration;

[0105] 3) Based on the joint attention mechanism of residual connection, the frequency domain features of cutting force and cutting vibration are adaptively fused to obtain the cutting force feature vector and the cutting vibration feature vector respectively;

[0106] 4) Construct a processing status classification model;

[0107] 5) Input the cutting force feature vector and the cutting vibration feature vector into the machining state classification model to obtain the current machining state and realize the monitoring of chatter in CNC milling.

[0108] Example 3:

[0109] A method for online monitoring of chatter in CNC milling based on residual connection and joint attention mechanism, with the same technical content as in Embodiment 2, further wherein the triaxial cutting force and triaxial acceleration are acquired by force measuring instrument and acceleration sensor, respectively.

[0110] Example 4:

[0111] A method for online monitoring of chatter in CNC milling based on residual connection and joint attention mechanism, with technical content the same as any one of embodiments 2-3, further comprising the following steps for processing the triaxial cutting force and triaxial acceleration:

[0112] 2.1) By using a sliding window, the triaxial cutting force and triaxial acceleration are divided into several segments of the original signal sequence; the original signal sequence includes the triaxial cutting force signal sequence and the triaxial acceleration signal sequence;

[0113] 2.2) Perform Fast Fourier Transform on each segment of the original signal sequence to obtain the triaxial cutting force and triaxial acceleration after Fast Fourier Transform;

[0114] 2.3) The triaxial cutting force and triaxial acceleration after the fast Fourier transform are standardized to obtain the standardized triaxial cutting force and triaxial acceleration;

[0115] 2.4) The three-dimensional cutting forces after standardization are fused to obtain the frequency domain characteristics of the cutting forces;

[0116] The frequency domain characteristics of the cutting vibration are obtained by fusing the standardized triaxial accelerations.

[0117] Example 5:

[0118] A method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism, with technical content the same as any one of embodiments 2-4, further wherein the triaxial cutting force and triaxial acceleration after fast Fourier transform are shown below:

[0119]

[0120] Where x(n) is the time-domain information of the original signal; It is the Fourier amplitude; k = 0, 1, ..., N-1, where N is the length of the signal sequence.

[0121] Example 6:

[0122] A method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism, with technical content the same as any one of embodiments 2-5, further wherein the standardized triaxial cutting force and triaxial acceleration are shown below:

[0123]

[0124] Where x represents the triaxial cutting force and triaxial acceleration after fast Fourier transform, z represents the triaxial cutting force and triaxial acceleration after standardization, and μ and σ are the mean and standard deviation.

[0125] Example 7:

[0126] A method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism, with technical content the same as any one of embodiments 2-6, further comprising the step of adaptively fusing the frequency domain characteristics of cutting force and cutting vibration respectively, including:

[0127] 3.1) Compress the input data X into a vector z∈R using global average pooling. C×1 The input data X is the frequency domain characteristic of the cutting force or the frequency domain characteristic of the cutting vibration.

[0128] The i-th component of vector z is shown below:

[0129]

[0130] In the formula, i = 1, 2, ..., C, and C is the number of channels;

[0131] 3.2) By utilizing the joint features of the input data activated by the convolutional layer, we obtain:

[0132]

[0133] In the formula, For joint features; This indicates multiplication by components;

[0134] Among them, the spatial location weight u and the channel features They are shown below:

[0135]

[0136]

[0137] In the formula, σ(·) represents the activation function; w3 and b3 represent the weights and biases, respectively.

[0138] Among them, the weights of the convolutional layer channels As shown below:

[0139]

[0140] In the formula, w1 and w2 represent the weights of the two-channel convolutional layer, b1 and b2 represent the corresponding biases, δ(·) represents the modified linear unit, and σ(·) represents the activation function.

[0141] 3.3) Establish the relationship between the joint features and the input data using residual connections to obtain the feature vector Y, i.e.:

[0142]

[0143] In the formula, when X is the frequency domain characteristic of cutting force, Y represents the characteristic vector of cutting force; when X is the frequency domain characteristic of cutting vibration, Y represents the characteristic vector of cutting vibration.

[0144] Example 8:

[0145] A method for online monitoring of chatter in CNC milling based on residual connection and joint attention mechanism, with the same technical content as any one of embodiments 2-7, further wherein the modified linear unit is a ReLU unit.

[0146] Example 9:

[0147] A method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism, with the same technical content as any one of embodiments 2-8, further wherein the activation function is the Sigmoid function.

[0148] Example 10:

[0149] A method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism, with the same technical content as any one of embodiments 2-9, further wherein the machining state classification model is a classifier containing only one fully connected layer;

[0150] The loss function FL of the processing state classification model is as follows:

[0151]

[0152] In the formula, α m For coefficients, To represent the true processing state of the training samples, The model predicts the processing state; M is the number of training samples; r is the focusing parameter, r>0.

[0153] Example 11:

[0154] A method for online monitoring of chatter in CNC milling based on residual connection and joint attention mechanism, the technical content of which is the same as any one of embodiments 2-10, further wherein the machining state includes a stable state and a chatter state.

[0155] Example 12:

[0156] A method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism, with technical content the same as any one of embodiments 2-11, further wherein the training process of the machining state classification model is as follows:

[0157] 1) Collect the triaxial cutting force and triaxial acceleration of the tool in multiple CNC milling systems, and obtain the corresponding actual machining state;

[0158] 2) Process the triaxial cutting force and triaxial acceleration to obtain the frequency domain characteristics of the cutting force and cutting vibration;

[0159] 3) Based on the joint attention mechanism of residual connection, the frequency domain features of cutting force and cutting vibration are adaptively fused to obtain multiple sets of cutting force feature vectors and cutting vibration feature vectors;

[0160] 4) Divide the cutting force feature vector, cutting vibration feature vector, and actual machining state into training set and test set; where the cutting force feature vector and cutting vibration feature vector are inputs, and the actual machining state is output;

[0161] 5) Train the processing state classification model using the training set;

[0162] 6) Test the processing state classification model using the test set and determine the classification accuracy of the processing state classification model. If the classification accuracy is less than the preset threshold (e.g., 90%), return to step 1.

[0163] Example 13:

[0164] A method for online monitoring of chatter in CNC milling based on residual connection and joint attention mechanism, with the same technical content as any one of embodiments 2-12, further wherein the frequency domain characteristics of cutting force and cutting vibration are obtained through two independent convolutional neural networks.

[0165] Example 14:

[0166] A method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism is proposed. First, the collected X / Y / Z triaxial cutting force and X / Y / Z triaxial acceleration signals are subjected to Fast Fourier Transform (FFT). Data standardization is performed using the Z-score method. A dual-stream convolutional neural network is used to extract the frequency domain features of cutting force and cutting vibration, respectively. A joint attention mechanism based on residual connectivity is established to adaptively fuse features from different sensors. The feature map is converted into a one-dimensional feature vector using global average pooling. The dataset partitioning is optimized using the Focal Loss function. Various machining states are probabilistically classified using a normalized exponential function to achieve chatter monitoring in CNC milling.

[0167] The specific steps of the method are as follows:

[0168] Step 1 Sensor signal preprocessing

[0169] Vibration and cutting force signals in the X, Y, and Z directions during the cutting process were collected using an accelerometer and a force gauge. To obtain frequency domain information, the original signals were subjected to a fast Fourier transform according to formula (1):

[0170]

[0171] Where x(n) is the time-domain information of the original signal, It is the Fourier amplitude, k = 0, 1, ..., N-1, where N is the length of the signal sequence.

[0172] The acquired multi-source signals are divided into q segments of length m×1 using a sliding window segmentation method. Each one-dimensional time series is then subjected to a Fast Fourier Transform to obtain a one-dimensional frequency domain sequence of length l×1. The data is standardized using the Z-score standard scoring method according to formula (2):

[0173]

[0174] Where x represents the components of the input feature vector, z represents the standardized data, and μ and σ are the mean and standard deviation of each component.

[0175] The frequency domain sequences of the three channels of cutting vibration and cutting force are fused into acceleration signal feature vectors and cutting force signal feature vectors of size l×3, respectively.

[0176] Step 2: Sensor signal feature extraction based on two-stream convolutional neural network

[0177] To ensure that the features of the cutting force sensor signal and the acceleration sensor signal are not confused, a two-stream convolutional neural network based on a one-dimensional convolutional neural network is established to extract features from the cutting force sensor signal and the acceleration sensor signal respectively.

[0178] Step 3 utilizes the joint attention mechanism of residual connections to fuse sensor signal features.

[0179] Combining the input feature map X = [x1, x2, ..., x C ], x i ∈R N×1 X is compressed into a vector z∈R by global average pooling. C×1 The i-th component of z is calculated according to formula (3):

[0180]

[0181] Where i = 1, 2, ..., C, and C is the number of channels.

[0182] Two 1×1 convolutional layers are used, and the channel weights are calculated according to formula (4):

[0183]

[0184] Where w1 and w2 represent the weights of the two-channel convolutional layer, b1 and b2 represent the corresponding biases, δ(·) represents the modified linear unit ReLU, and σ(·) represents the activation function Sigmoid.

[0185] The weights are assigned to each channel by multiplying each component, and the channel characteristics are calculated according to formula (5):

[0186]

[0187] in, This indicates multiplication by components.

[0188] To obtain the spatial location information of each feature map, local features are extracted using a 3×1 convolution method based on the attention mechanism, and the weights of the spatial locations are calculated according to formula (6):

[0189]

[0190] Activate the joint features according to formula (7):

[0191]

[0192] The relationship between the joint features and the original features is established using formula (8) with residual connection:

[0193]

[0194] Step 4: Online monitoring of CNC milling machining status

[0195] The feature map processed by the joint attention mechanism of residual connection is converted into a one-dimensional feature vector by global average pooling, and the feature vector is then input into a classifier containing only one fully connected layer.

[0196] The processing state is probabilistically classified using the normalized exponential function Softmax, and the dataset is partitioned and optimized using the Focal Loss function to obtain the model parameters. The Focal Loss function is calculated according to formula (9):

[0197]

[0198] Where, α m To increase the coefficient of the weight of misclassified samples, To collect the actual cutting state of the sample, This represents the cutting state predicted by the model.

[0199] Ultimately, online monitoring of chatter conditions in CNC milling was achieved.

[0200] Example 15:

[0201] A method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism.

[0202] In multi-axis machining of complex irregular curved surfaces, drastic fluctuations in cutting force and acceleration can easily lead to chatter, hindering the improvement of surface quality. To address this problem, an online chatter monitoring method for CNC milling based on residual connection and joint attention mechanisms is proposed. The overall process is shown in the attached figure. Figure 1 As shown.

[0203] Step 1: Acceleration and cutting force signals in the X, Y, and Z directions during the cutting process are collected using an accelerometer and a force gauge. The signals are then processed using a fast Fourier transform according to formula (1) to obtain frequency domain information. The data is then standardized according to formula (2), and the frequency domain sequences of the three channels of acceleration and cutting force are fused into acceleration signal feature vectors and cutting force signal feature vectors, which are used as inputs to the deep learning network.

[0204] Step 2, establish the following based on a one-dimensional convolutional neural network: Figure 2 The convolutional neural network with a dual-stream structure shown extracts features from the signals of the cutting force sensor and the acceleration sensor, respectively.

[0205] Step 3, combined with the appendix Figure 3The joint attention mechanism diagram of residual connection compresses the feature map into a feature vector through global average pooling. The feature vector elements are calculated according to formula (3), the channel weights are calculated according to formula (4), the channel features are calculated according to formula (5), the spatial position weights are calculated according to formula (6), the joint features are activated according to formula (7), and the relationship between the joint features and the original features is established according to formula (8).

[0206] Step 4: Input the feature vector into a classifier containing only one fully connected layer, and divide the dataset using the Focal Loss loss function according to formula (9), where α m =0.25, r=2, using a normalized exponential function to classify the machining state probabilistically, thus realizing the monitoring of chatter in CNC milling. Figure 4 This is the cutting force signal in a steady state. Figure 5 This is the cutting force signal during chattering.

[0207] Compared with chatter identification methods based on a single sensor, such as cutting vibration signal or cutting force signal, the single-signal-based method only employs a single-stream convolutional neural network and does not consider the attention mechanism of residual connections. Prediction accuracy is calculated as the ratio of the number of samples that correctly predict cutting stability and chatter states to the total number of test samples. The accuracies of chatter identification methods based on a single cutting vibration signal and a single cutting force signal are 93.77% and 96.03%, respectively. The proposed chatter monitoring method combining cutting force and cutting vibration characteristics achieves an accuracy of 98.68%, which is higher than chatter monitoring methods that only consider cutting vibration signal or cutting force signal.

[0208] The online monitoring method for chatter in CNC milling based on the joint attention mechanism of residual connection in this invention can realize chatter status monitoring in multi-axis milling of irregular curved surfaces. The method is convenient to implement and can effectively improve the accuracy of chatter status monitoring.

Claims

1. A method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism, characterized in that, Includes the following steps: Step 1) Collect the three-dimensional cutting force and three-dimensional acceleration of the tool in the CNC milling system; Step 2) Process the triaxial cutting force and triaxial acceleration to obtain the frequency domain characteristics of the cutting force and cutting vibration; Step 3) Based on the joint attention mechanism of residual connection, adaptively fuse the frequency domain features of cutting force and cutting vibration respectively to obtain the cutting force feature vector and the cutting vibration feature vector; Step 4) Construct a processing status classification model; Step 5) Input the cutting force feature vector and the cutting vibration feature vector into the machining state classification model to obtain the current machining state and realize the monitoring of chatter in CNC milling; The steps for adaptively fusing the frequency domain characteristics of cutting force and cutting vibration separately include: Step 3.1) Use global average pooling to process the input data. Compressed into vector The input data This refers to the frequency domain characteristics of cutting force or cutting vibration. Where, vector The j-th component is shown below: （3） In the formula, j = 1, 2, ..., C, where C is the number of channels and N is the length of the signal sequence; This is the input data for channel j; Step 3.2) Activate the joint features of the input data using convolutional layers to obtain: （4） In the formula, For joint features; This indicates multiplication by components; Among them, the weight of spatial location and channel characteristics They are shown below: （5） （6） In the formula, Indicates the activation function; , Indicates weights and biases; Among them, the weights of the convolutional layer channels As shown below: （7） In the formula, and These represent the weights of the two channels of the convolutional layer. and These represent the corresponding biases. Indicates a modified linear unit; Indicates the activation function; Representing vectors Element; Step 3.3) Establish the relationship between the joint features and the input data using residual connections to obtain the feature vector Y, i.e.: （8） In the formula, when X is the frequency domain characteristic of the cutting force, Y represents the characteristic vector of the cutting force; when X is the frequency domain characteristic of the cutting vibration, Y represents the characteristic vector of the cutting vibration. The processing state classification model is a classifier containing only one fully connected layer; The loss function FL of the processing state classification model is as follows: （9） In the formula, For coefficients, To represent the true processing state of the training samples, The model predicts the processing state; M is the number of training samples; r is the focusing parameter, r>

0.

2. The method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism according to claim 1, characterized in that, The triaxial cutting force and triaxial acceleration are obtained by force measuring instrument and acceleration sensor, respectively.

3. The method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism according to claim 1, characterized in that, The steps for processing the triaxial cutting force and triaxial acceleration include: Step 2.1) Using a sliding window, divide the triaxial cutting force and triaxial acceleration into several segments of the original signal sequence; the original signal sequence includes the triaxial cutting force signal sequence and the triaxial acceleration signal sequence; Step 2.2) Perform Fast Fourier Transform on each segment of the original signal sequence to obtain the triaxial cutting force and triaxial acceleration after Fast Fourier Transform; Step 2.3) Standardize the triaxial cutting force and triaxial acceleration after the fast Fourier transform to obtain the standardized triaxial cutting force and triaxial acceleration; Step 2.4) The standardized three-dimensional cutting forces are fused to obtain the frequency domain characteristics of the cutting forces; The frequency domain characteristics of the cutting vibration are obtained by fusing the standardized triaxial accelerations.

4. The method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism according to claim 3, characterized in that, The triaxial cutting forces and triaxial accelerations after the Fast Fourier Transform are shown below: （1） in, It is the time-domain information of the original signal; It is the Fourier amplitude; n=0, 1, …, N-1, where N is the length of the signal sequence; k is a positive integer; i is an imaginary number.

5. The method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism according to claim 3, characterized in that, The standardized triaxial cutting forces and triaxial accelerations are shown below: （2） Where x represents the triaxial cutting force and triaxial acceleration after the fast Fourier transform. This represents the standardized triaxial cutting force and triaxial acceleration. and These are the mean and standard deviation.

6. The method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism according to claim 1, characterized in that, The modified linear unit is a ReLU unit.

7. The method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism according to claim 1, characterized in that, The activation function is the Sigmoid function.

8. The method for online monitoring of chatter in CNC milling based on residual connectivity and joint attention mechanism according to claim 1, characterized in that, The processing states include a stable state and a chattering state.

Images