A method and system for predicting surface roughness in high-speed milling considering chatter

Abstract

The present invention provides a method and system for predicting surface roughness of high-speed milling taking into account chatter, which relates to the technical field of milling processing. The present invention continuously collects force signals through a force sensor configured on the milling workbench, and continuously collects vibration signals through an acceleration sensor configured on the milling workbench, and then extracts multi-dimensional fine-grained features of the force signal and the vibration signal, thereby obtaining detailed information characterizing the surface roughness caused by the inherent process and chatter in the milling process, and then effectively fuses the multi-dimensional fine-grained features of the periodic signal and the chatter signal through a fine-grained feature fusion network to realize surface roughness prediction. Therefore, the present invention collects vibration signals and predicts surface roughness in combination with force signals, thereby reducing the situation where the surface roughness prediction is inaccurate due to the chatter phenomenon in high-speed milling processing.

Description

Technical Field

[0001] The present invention relates to the technical field of milling processing, in particular to a method and system for predicting surface roughness of high-speed milling considering chatter. Background Art

[0002] High-speed milling is a key subtractive technology in modern manufacturing and is widely used in the automotive, aviation, and aerospace industries. With the increasing demand for high quality and low cost of mechanical products, controlling machining quality has become particularly important. In actual production, surface roughness is often used as an evaluation indicator to measure machining quality. It not only affects the operating performance of the product, but also plays an important role in the overall service life of the product. At present, surface roughness measurement methods are usually performed offline after machining is completed, which brings some limitations, such as low efficiency, high cost, and scratches on the workpiece surface. Therefore, it is necessary to develop an online surface roughness prediction method, which is of great significance for actual production.

[0003] In recent years, many scholars have begun to study data-driven methods, that is, combining sensing technology and artificial intelligence technology to apply to online surface roughness prediction. For surface roughness prediction, vibration signals are currently monitored online during milling processing, and online surface roughness prediction is achieved by extracting features sensitive to surface roughness and training artificial neural networks. At present, there are also methods based on principal component analysis and grey correlation theory to pre-process the monitored noise, vibration and texture images, and use particle swarm optimization and least squares support vector machine to construct a multidimensional feature fusion model for surface roughness prediction. At present, there are also methods that use continuous wavelet transform to extract the time-frequency characteristics of force signals, and then apply convolutional neural networks and transfer learning to complete the surface roughness prediction task in the milling process of γ-TiAl alloy.

[0004] The aforementioned research has extensively explored data-driven methods, laying an important foundation for online surface roughness prediction during milling processes. The underlying principles are well established. However, few studies have examined online surface roughness prediction that considers chatter. In high-speed milling, chatter often occurs due to the attenuation of the damping effect, resulting in chatter marks on the machined surface. This significantly affects surface roughness and can lead to inaccurate surface roughness predictions. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the present invention aims to propose a method for predicting surface roughness in high-speed milling considering chatter, comprising:

[0006] Step 1: Perform a plane high-speed milling experiment on the workpiece on a milling table. Continuously collect force signals through the force sensor configured on the milling table, and continuously collect vibration signals through the acceleration sensor configured on the milling table.

[0007] Step 2: Use the continuously collected force signal and the continuously collected vibration signal as input samples, use the surface roughness corresponding to the input samples as output samples, and construct a training set based on the input samples and output samples;

[0008] Step 3: For the input samples in the training set, perform multidimensional fine-grained feature extraction on the force signal and vibration signal in the input samples and then splice them to obtain the multidimensional fine-grained feature components of the periodic signal and the multidimensional fine-grained feature components of the chatter signal;

[0009] Step 4: Input the multi-dimensional fine-grained feature components of the periodic signal and the multi-dimensional fine-grained feature components of the chatter signal into a fine-grained feature fusion network to obtain a predicted value of surface roughness. The fine-grained feature fusion network includes a DREF module, a FASM module, and a Bayesian prediction module.

[0010] Step 5: Based on the predicted values ​​of multiple surface roughness and their corresponding output samples, the parameters of the DREF module, FASM module, and Bayesian prediction module in the fine-grained feature fusion network are iteratively updated multiple times to obtain the trained fine-grained feature fusion network;

[0011] Step 6: Continuously collect the force signal and vibration signal to be predicted, and then execute steps 3-4 based on the trained fine-grained feature fusion network to obtain the surface roughness.

[0012] Optionally, step 3 specifically includes:

[0013] Step 3.1: Use the wavelet packet decomposition (WPD) method to reduce the noise of the vibration signal to obtain a reduced-noise vibration signal. Then, use successive variational mode decomposition (SVMD) to extract the first periodic signal and the first chatter signal from the reduced-noise vibration signal.

[0014] Step 3.2: Extract the second periodic signal and the second chatter signal from the force signal through successive variational mode decomposition (SVMD);

[0015] Step 3.3: Perform feature extraction on the first periodic signal, the first dither signal, the second periodic signal, and the second dither signal, respectively, to obtain multidimensional fine-grained features of the first periodic signal, the first dither signal, the second periodic signal, and the second dither signal, wherein the multidimensional fine-grained features include time domain features and frequency domain features.

[0016] Step 3.4: Perform vector concatenation on the multidimensional fine-grained features of the first periodic signal and the multidimensional fine-grained features of the second periodic signal to obtain a multidimensional fine-grained feature component of the periodic signal. Perform vector concatenation on the multidimensional fine-grained features of the first chatter signal and the multidimensional fine-grained features of the second chatter signal to obtain a multidimensional fine-grained feature component of the chatter signal.

[0017] Optionally, step 4 specifically includes:

[0018] Step 4.1: Based on the DREF module in the fine-grained feature fusion network, feature extraction and fusion are performed on the multi-dimensional fine-grained feature components of the periodic signal and the multi-dimensional fine-grained feature components of the chatter signal to obtain fused features;

[0019] Step 4.2: Process the fused features through SENet and BiLSTM in the FASM module to obtain high-level features;

[0020] Step 4.3: Input the high-level features into the Bayesian prediction module to obtain the predicted value of surface roughness.

[0021] Optionally, step 4.1 specifically includes:

[0022] Step 4.1.1: Perform deep feature extraction on the multidimensional fine-grained feature components of the periodic signal to obtain deep features of the periodic signal. Specifically, perform a convolution operation, batch normalization (BN), ReLU, maximum pooling, a convolution operation, batch normalization (BN), ReLU, and maximum pooling on the multidimensional fine-grained feature components of the periodic signal to obtain a first deep feature. Pass the multidimensional fine-grained feature components of the periodic signal through a fully connected layer to obtain a second deep feature. Perform a residual connection between the first deep feature and the second deep feature by adding them to obtain the deep feature of the periodic signal.

[0023] Step 4.1.2: Process the deep features of the periodic signal through the gating mechanism composed of sigmoid layer and tanh layer to obtain the first gating weight matrix , which is specifically achieved through the following formula:

[0024] ;

[0025] in, μ and ν To learn the parameters, is the depth feature of the periodic signal, represents element-wise multiplication, Represents the sigmoid layer;

[0026] The first gate weight matrix is ​​processed by three one-dimensional convolutions with different receptive fields Perform convolution to obtain first-scale features, second-scale features, and third-scale features, then concatenate the first-scale features, second-scale features, and third-scale features to obtain first multi-scale features. Input the first multi-scale features into the FC layer, where the first multi-scale features undergo ReLU activation function and sigmoid activation function to obtain first adaptive weights.

[0027] Step 4.1.3: Perform deep feature extraction on the multidimensional fine-grained feature components of the chatter signal to obtain deep features of the chatter signal. Specifically, perform a convolution operation, batch normalization (BN), ReLU, max pooling, convolution operation, batch normalization (BN), ReLU, and max pooling on the multidimensional fine-grained feature components of the chatter signal to obtain a third deep feature. Pass the multidimensional fine-grained feature components of the chatter signal through a fully connected layer to obtain a fourth deep feature. The third and fourth deep features are residually connected by addition to obtain the deep feature of the chatter signal.

[0028] Step 4.1.4: Process the deep features of the chatter signal through the gating mechanism composed of sigmoid layer and tanh layer to obtain the second gating weight matrix , which is specifically achieved through the following formula:

[0029] ;

[0030] in, is the depth feature of the chatter signal;

[0031] The second gate weight matrix is ​​adjusted by three one-dimensional convolutions with different receptive fields. Perform convolution to obtain fourth-scale features, fifth-scale features, and sixth-fifth-scale features, and then concatenate the fourth-scale features, fifth-scale features, and sixth-fifth-scale features to obtain second multi-scale features. Input the second multi-scale features into the FC layer. In the FC layer, the second multi-scale features are activated by ReLU and sigmoid functions to obtain second adaptive weights.

[0032] Step 4.1.5: Use the first adaptive weight and the second adaptive weight to perform weighted summation on the depth features of the periodic signal and the depth features of the chatter signal to obtain the fusion feature, which is specifically achieved by the following formula:

[0033] ;

[0034] in, F represents the fusion feature, represents the first adaptive weight, represents the deep features of periodic signals, represents the second adaptive weight, represents the depth feature of the chatter signal, Indicates addition of elements.

[0035] Optionally, step 4.2 specifically includes:

[0036] Step 4.2.1: In SENet, the fused features are subjected to global pooling, ReLU activation function, and sigmoid activation function to obtain channel attention weights. The channel attention weights and the fused features are then multiplied by the scale operation to obtain recalibrated fused features.

[0037] Step 4.2.2: Process the recalibrated fused features through BiLSTM to obtain high-level features.

[0038] Optionally, step 4.3 specifically includes:

[0039] In the Bayesian prediction module, high-level features are converted into one-dimensional vectors through the compression layer. The one-dimensional vector is passed through the ReLU activation function in the fully connected layer to obtain a vector after enhanced nonlinearity. The enhanced nonlinear vector is processed by approximate Bayesian dropout to obtain the initial surface roughness. The same input sample is repeatedly input into the fine-grained feature fusion network through Monte Carlo random sampling to obtain multiple initial surface roughnesses. The average of all initial surface roughnesses is calculated to obtain the predicted value of surface roughness.

[0040] A high-speed milling surface roughness prediction system considering chatter is used to implement the high-speed milling surface roughness prediction method considering chatter, comprising a data acquisition module, a model training module, a data processing module and a surface roughness prediction module;

[0041] The data acquisition module is used to continuously collect force signals and vibration signals, and transmit the force signals and vibration signals to the model training module;

[0042] The model training module is used to use the continuously collected force signal and the continuously collected vibration signal as input samples, and the surface roughness corresponding to the input samples as output samples, to construct a training set based on the input samples and the output samples, and to transmit the training set input samples to the data processing module; the model training module is also used to receive the predicted value of the surface roughness, and based on the multiple predicted values ​​of the surface roughness and their corresponding output samples, to iteratively update the parameters in the fine-grained feature fusion network multiple times;

[0043] The data processing module is used to receive input samples, perform multi-dimensional fine-grained feature extraction and splicing on the force signal and vibration signal in the input samples, and transmit the obtained multi-dimensional fine-grained feature components of the periodic signal and the multi-dimensional fine-grained feature components of the chatter signal to the surface roughness prediction module;

[0044] The surface roughness prediction module is used to receive the multidimensional fine-grained feature components of the periodic signal and the multidimensional fine-grained feature components of the chatter signal, and input the multidimensional fine-grained feature components of the periodic signal and the multidimensional fine-grained feature components of the chatter signal into a fine-grained feature fusion network to obtain a predicted value of the surface roughness, and transmit the predicted value of the surface roughness to the model training module.

[0045] The beneficial effects of adopting the above technical solution are:

[0046] The present invention continuously collects force signals through a force sensor configured on the milling workbench, and continuously collects vibration signals through an acceleration sensor configured on the milling workbench, and then extracts multi-dimensional fine-grained features of the force signal and the vibration signal, thereby obtaining detailed information characterizing the surface roughness caused by the inherent process and chattering during the milling process, and then effectively fuses the multi-dimensional fine-grained features of the periodic signal and the chattering signal through a fine-grained feature fusion network to achieve surface roughness prediction. Therefore, the present invention collects vibration signals and predicts surface roughness in combination with force signals, thereby reducing the situation where inaccurate surface roughness prediction is caused by chattering in high-speed milling processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of a flow chart of a method for predicting surface roughness in high-speed milling considering chatter in an embodiment of the present invention;

[0048] Figure 2 This is a diagram of a workpiece processing experiment in an embodiment of the present invention;

[0049] Figure 3 Schematic diagram of the structure of a fine-grained feature fusion network in an embodiment of the present invention;

[0050] Figure 4 Graph showing prediction results of a high-speed milling surface roughness prediction method considering chatter in an embodiment of the present invention;

[0051] Figure 5 Schematic diagram of the structure of a high-speed milling surface roughness prediction system considering chatter in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0053] In view of the problems existing in the prior art, the present invention provides a method for predicting surface roughness of high-speed milling considering chatter, so as to improve the accuracy of surface roughness prediction, thereby bringing significant economic benefits to the manufacturing industry by reducing the scrap rate and the number of downtimes. Figure 1 , which may include the following steps:

[0054] Step 1: On the milling table, perform a plane high-speed milling experiment on the workpiece. The force sensor configured on the milling table continuously collects force signals, and the acceleration sensor configured on the milling table continuously collects vibration signals. Combined with Figure 2 In the specific implementation process of the present invention, a plane high-speed milling experiment is carried out in a machining center through a computer programming control system and a control handle. Data is collected through force sensors and acceleration sensors and output to a data acquisition system. The data acquisition system outputs multi-source signal data, namely force signals and vibration signals.

[0055] In the specific implementation process of the present invention, a series of Al6061 plane high-speed milling experiments under different processing parameters were carried out on the three-axis vertical machining center CNC-400DBS-S3. The force sensor was installed on the workbench and connected to the charge amplifier to collect x Direction and y The accelerometer is fixed to the workpiece surface to collect vibration signals. All collected signals are sampled at a frequency of 20kHz by the DH5956 data acquisition system and transmitted to a computer for storage and processing.

[0056] The workpiece is an Al6061 block of 140 × 140 × 15 mm, and the cutting tool is an uncoated three-edge flat-bottom milling cutter with a diameter of 4 mm. A constant feed rate of 200 mm / min is used along the x High-speed milling is performed in the same direction. The spindle speed and depth of cut vary, corresponding to different milling conditions, including stable, transitional, and severe chatter. This is shown in Table 1, where spindle speed is in rpm and depth of cut is in mm. After machining, the machined surface is observed using a microscope to assess the milling condition.

[0057] Table 1 Processing parameters

[0058] Serial number Spindle speed Cutting depth Milling status 1 15000 0.1 Stablize 2 15000 0.2 Stablize 3 15000 0.5 transition 4 15000 0.6 transition 5 15000 0.9 Severe vibration 6 15600 0.1 Stablize 7 15600 0.2 Stablize 8 15600 0.5 transition 9 15600 0.7 Severe vibration 10 15600 1.1 Severe vibration 11 16200 0.2 Stablize 12 16200 0.4 transition 13 16200 0.5 transition 14 16200 0.7 Severe vibration 15 16200 1.2 Severe vibration

[0059] Step 2: Use the continuously collected force signal and the continuously collected vibration signal as input samples, use the surface roughness corresponding to the input samples as output samples, and construct a training set based on the input samples and output samples;

[0060] Step 2.1: Slide a sliding window over the continuously collected force signals and vibration signals to obtain multiple force signals and vibration signals corresponding to the force signals. The force signals and vibration signals corresponding to the force signals constitute input samples.

[0061] In the specific implementation process, the size and step size of the sliding window are both set to 1s.

[0062] Step 2.2: Continuously measure the surface roughness Ra of the workpiece using a roughness meter to determine the surface roughness corresponding to the input sample, and use the surface roughness corresponding to the input sample as the output sample;

[0063] In a specific implementation process, the surface roughness of the workpiece is measured three times, and the average value of the three measurements is calculated as the final surface roughness, thereby determining the surface roughness corresponding to the input sample.

[0064] Step 2.3: The input sample and the output sample constitute a training sample, and multiple training samples constitute a sample set. The sample set is divided into a training set and a test set according to a preset ratio. Specifically, 70% of the sample set is used as a training set and 30% is used as a test set. The training set is used for model training, and the test set is used for model testing.

[0065] Step 3: For the input samples in the training set, perform multidimensional fine-grained feature extraction on the force signal and vibration signal in the input samples and then splice them to obtain the multidimensional fine-grained feature components of the periodic signal and the multidimensional fine-grained feature components of the chatter signal;

[0066] Step 3.1: Use the wavelet packet decomposition (WPD) method to reduce the noise of the vibration signal to obtain a reduced-noise vibration signal. Then, use successive variational mode decomposition (SVMD) to extract the first periodic signal and the first chatter signal from the reduced-noise vibration signal.

[0067] Because force signals are relatively stable, while vibration signals are easily affected by environmental noise, we first use wavelet packet decomposition (WPD) to reduce the noise of the vibration signal. WPD is an improvement on WT, decomposing the input signal into multiple frequency bands, effectively filtering out most noise.

[0068] Step 3.2: Extract the second periodic signal and the second chatter signal from the force signal through successive variational mode decomposition (SVMD);

[0069] Step 3.3: Perform feature extraction on the first periodic signal, the first dither signal, the second periodic signal, and the second dither signal, respectively, to obtain multidimensional fine-grained features of the first periodic signal, the first dither signal, the second periodic signal, and the second dither signal, wherein the multidimensional fine-grained features include time domain features and frequency domain features.

[0070] Among them, the time domain features include standard deviation, peak value, root square amplitude, absolute average amplitude, root mean square, peak-to-peak value, kurtosis and kurtosis factor, and the frequency domain feature is the average frequency.

[0071] Among them, the multidimensional fine-grained features of the periodic signal mainly reflect the periodic fluctuation of the surface quality caused by the inherent milling process, while the multidimensional fine-grained features of the chatter signal reveal the non-periodic changes in the surface quality caused by the chatter effect.

[0072] Step 3.4: Normalize and vector-concatenate the multidimensional fine-grained features of the first periodic signal and the multidimensional fine-grained features of the second periodic signal to obtain the multidimensional fine-grained feature components of the periodic signal. Normalize and vector-concatenate the multidimensional fine-grained features of the first chatter signal and the multidimensional fine-grained features of the second chatter signal to obtain the multidimensional fine-grained feature components of the chatter signal.

[0073] In order to balance the impact of each feature on the model, normalization is performed to eliminate the dimensional differences between them.

[0074] Step 4: Combine Figure 3 , the multi-dimensional fine-grained feature components of the periodic signal and the multi-dimensional fine-grained feature components of the chatter signal are input into a fine-grained feature fusion network to obtain a predicted value of the surface roughness, wherein the fine-grained feature fusion network includes a DREF module, a FASM module and a Bayesian prediction module;

[0075] Step 4.1: Based on the DREF module in the fine-grained feature fusion network, feature extraction and fusion are performed on the multi-dimensional fine-grained feature components of the periodic signal and the multi-dimensional fine-grained feature components of the chatter signal to obtain fused features;

[0076] Among them, the DREF module is designed to deeply explore the coupling relationship between the multi-dimensional fine-grained features of periodic signal components and chatter signal components in surface roughness prediction. It consists of two parts: deep feature extraction and feature fusion.

[0077] Step 4.1.1: Perform deep feature extraction on the multidimensional fine-grained feature components of the periodic signal to obtain the deep features of the periodic signal. Specifically, perform convolution operation, batch normalization BN, ReLU, maximum pooling, convolution operation, batch normalization BN, ReLU, and maximum pooling on the multidimensional fine-grained feature components of the periodic signal to ensure the stability and efficiency of feature extraction to obtain the first deep feature. Pass the multidimensional fine-grained feature components of the periodic signal through the fully connected layer to obtain the second deep feature. Perform residual connection between the first deep feature and the second deep feature by addition to obtain the deep feature of the periodic signal, thereby preventing the gradient from disappearing.

[0078] Step 4.1.2: Process the deep features of the periodic signal through the gating mechanism composed of sigmoid layer and tanh layer to obtain the first gating weight matrix , which is specifically achieved through the following formula:

[0079] ;

[0080] in, μ and ν To learn the parameters, is the depth feature of the periodic signal, represents element-wise multiplication, Represents the sigmoid layer;

[0081] The first gate weight matrix is ​​processed by three one-dimensional convolutions with different receptive fields Perform convolution to obtain first-scale features, second-scale features, and third-scale features, and then concatenate the first-scale features, second-scale features, and third-scale features to obtain first multi-scale features to enhance nonlinear expression capabilities. Input the first multi-scale features into the FC layer, where the first multi-scale features undergo ReLU activation function and sigmoid activation function to obtain first adaptive weights.

[0082] Step 4.1.3: Perform deep feature extraction on the multidimensional fine-grained feature components of the chatter signal to obtain deep features of the chatter signal. Specifically, perform a convolution operation, batch normalization (BN), ReLU, max pooling, convolution operation, batch normalization (BN), ReLU, and max pooling on the multidimensional fine-grained feature components of the chatter signal to obtain a third deep feature. Pass the multidimensional fine-grained feature components of the chatter signal through a fully connected layer to obtain a fourth deep feature. The third and fourth deep features are residually connected by addition to obtain the deep feature of the chatter signal.

[0083] Step 4.1.4: Process the deep features of the chatter signal through the gating mechanism composed of sigmoid layer and tanh layer to obtain the second gating weight matrix , which is specifically achieved through the following formula:

[0084] ;

[0085] in, is the depth feature of the chatter signal;

[0086] The second gate weight matrix is ​​adjusted by three one-dimensional convolutions with different receptive fields. Perform convolution to obtain fourth-scale features, fifth-scale features, and sixth-fifth-scale features, and then concatenate the fourth-scale features, fifth-scale features, and sixth-fifth-scale features to obtain second multi-scale features. Input the second multi-scale features into the FC layer. In the FC layer, the second multi-scale features are activated by ReLU and sigmoid functions to obtain second adaptive weights.

[0087] Step 4.1.5: Use the first adaptive weight and the second adaptive weight to perform weighted summation on the depth features of the periodic signal and the depth features of the chatter signal to obtain the fusion feature, which is specifically achieved by the following formula:

[0088] ;

[0089] in, F represents the fusion feature, represents the first adaptive weight, represents the deep features of periodic signals, represents the second adaptive weight, represents the depth feature of the chatter signal, Indicates addition.

[0090] Step 4.2: Process the fused features through SENet and BiLSTM in the FASM module to obtain high-level features;

[0091] Step 4.2.1: In SENet, the fused features are subjected to global pooling, ReLU activation function, and sigmoid activation function to obtain channel attention weights. The channel attention weights and the fused features are then multiplied by the scale operation to obtain recalibrated fused features.

[0092] SENet can be understood as consisting of Squeeze, Excitation, and Scale operations, which suppress the influence of irrelevant or redundant information by focusing on different parts of the fused features. In the Squeeze operation, global information of the input features is embedded into the network through global pooling. Next, the Excitation operation establishes two FC layers, each with a ReLU activation function and a sigmoid activation function. These two FC layers learn the dependencies between channels and generate channel attention weights. The Scale operation performs a dot product of the obtained channel attention weights with the fused features, outputting recalibrated fused features.

[0093] Step 4.2.2: Process the recalibrated fused features through BiLSTM to obtain high-level features.

[0094] The BiLSTM is the final component of the FASM module, consisting of two layers of 50 neurons. Each BiLSTM layer contains a forward LSTM and a backward LSTM, which store the long-term temporal information of the fused features. LSTM is a special recurrent neural network that addresses the vanishing and exploding gradient problems faced by traditional RNNs when processing long sequences by introducing gating units. The gating unit consists of a forget gate, an input gate, an output gate, and a memory cell. The forget gate is used to clear out no longer useful information from the historical state, the input gate is used to determine new information that can be added to the memory cell, and the output gate is used to control the output of the memory cell information at the current time step.

[0095] Step 4.3: Input the high-level features into the Bayesian prediction module to obtain the predicted value of surface roughness.

[0096] In the Bayesian prediction module, high-level features are converted into one-dimensional vectors through the compression layer. The one-dimensional vectors are then activated by the ReLU activation function in the fully connected layer to enhance the nonlinear expression ability of the model. The number of neurons is set to 30 to obtain the vector after enhanced nonlinearity. The enhanced nonlinearity vector is processed by approximate Bayesian dropout to obtain the initial surface roughness. The same input sample is repeatedly input into the fine-grained feature fusion network through Monte Carlo random sampling to obtain multiple initial surface roughnesses. The average of all initial surface roughnesses is calculated to obtain the predicted value of the surface roughness.

[0097] Step 5: Based on the predicted values ​​of multiple surface roughness and their corresponding output samples, the parameters of the DREF module, FASM module, and Bayesian prediction module in the fine-grained feature fusion network are iteratively updated multiple times to obtain the trained fine-grained feature fusion network;

[0098] Specifically, the mean square error between the predicted value of surface roughness and the output sample is calculated to obtain the loss function value. Based on the loss function value, the parameters of the DREF module, FASM module, and Bayesian prediction module in the fine-grained feature fusion network are updated through the Adam optimizer, and the next training sample in the training set is obtained. The algorithm returns to step 3 and continues until the number of parameter updates reaches the preset number of iterations, resulting in a trained fine-grained feature fusion network.

[0099] During the training process, model training was performed based on the Tensorflow 1.14.0 framework with Keras 2.2.4, the learning rate and batch size were set to 0.001 and 64 respectively, and the preset number of iterations was 30.

[0100] Combine Figure 4 The predicted values ​​obtained by the present invention are highly consistent with the actual values ​​measured by the roughness meter, which fully demonstrates its potential and value in industrial applications. In order to verify the performance advantages of the proposed method in surface roughness prediction, five classic deep learning methods were selected for comparison, including CapsNet, DCNN, BiLSTM, BiGRU and ResNet. The evaluation index is R 2 , RMSE, MAPE and MAE. The comparison results are shown in Table 2. Obviously, the proposed method obtains the best results in surface roughness prediction. 2 The improvement was 6.36%-24.52%, the RMSE was reduced by 50.7%-70.19%, the MAPE was reduced by 51.97%%-80.5%, and the MAE was reduced by 55.08%-70.84%.

[0101] Table 2 Comparison results with typical deep learning methods

[0102] method <![CDATA[R 2 ]]> RMSE MAPE(%) MAE The present invention 0.9812 0.0493 6.5347 0.0345 CapsNet 0.8734 0.1279 19.0277 0.1003 DCNN 0.8364 0.1454 33.5193 0.1089 BiLSTM 0.9225 0.1 13.8407 0.0768 BiGRU 0.9081 0.1090 13.6052 0.0824 ResNet 0.7880 0.1654 29.0784 0.1183

[0103] Step 6: Continuously collect the force signal and vibration signal to be predicted, and then execute steps 3-4 based on the trained fine-grained feature fusion network to obtain the surface roughness. Specifically, perform multi-dimensional fine-grained feature extraction and splicing on the force signal and the vibration signal corresponding to the force signal to obtain the multi-dimensional fine-grained feature components of the periodic signal and the multi-dimensional fine-grained feature components of the flutter signal. The multi-dimensional fine-grained feature components of the periodic signal and the multi-dimensional fine-grained feature components of the flutter signal are input into the trained fine-grained feature fusion network to obtain the surface roughness.

[0104] A high-speed milling surface roughness prediction system considering chatter is used to implement a high-speed milling surface roughness prediction method considering chatter, combined with Figure 5 , including data acquisition module, model training module, data processing module and surface roughness prediction module;

[0105] The data acquisition module is used to continuously collect force signals and vibration signals, and transmit the force signals and vibration signals to the model training module;

[0106] The model training module is used to use the continuously collected force signal and the continuously collected vibration signal as input samples, and the surface roughness corresponding to the input samples as output samples, to construct a training set based on the input samples and the output samples, and to transmit the training set input samples to the data processing module; the model training module is also used to receive the predicted value of the surface roughness, and based on the multiple predicted values ​​of the surface roughness and their corresponding output samples, to iteratively update the parameters in the fine-grained feature fusion network multiple times;

[0107] The data processing module is used to receive input samples, perform multi-dimensional fine-grained feature extraction and splicing on the force signal and vibration signal in the input samples, and transmit the obtained multi-dimensional fine-grained feature components of the periodic signal and the multi-dimensional fine-grained feature components of the chatter signal to the surface roughness prediction module;

[0108] The surface roughness prediction module is used to receive the multidimensional fine-grained feature components of the periodic signal and the multidimensional fine-grained feature components of the chatter signal, and input the multidimensional fine-grained feature components of the periodic signal and the multidimensional fine-grained feature components of the chatter signal into a fine-grained feature fusion network to obtain a predicted value of the surface roughness, and transmit the predicted value of the surface roughness to the model training module.

[0109] It should be noted that the interaction process between the data processing module, the surface roughness prediction module and the model training module only exists in the training process of the fine-grained feature fusion network. In the process of completing the specific prediction by the fine-grained feature fusion network training, combined with step 6, specifically, step 6 continuously collects the force signal to be predicted and the vibration signal to be predicted through the data acquisition module, and then the data acquisition module transmits the force signal to be predicted and the vibration signal to be predicted to the data processing module. The data processing module receives the force signal to be predicted and the vibration signal to be predicted, performs multi-dimensional fine-grained feature extraction and splicing on the force signal and the vibration signal, obtains the multi-dimensional fine-grained feature components of the periodic signal and the multi-dimensional fine-grained feature components of the flutter signal, and transmits the multi-dimensional fine-grained feature components of the periodic signal and the multi-dimensional fine-grained feature components of the flutter signal to the surface roughness prediction module. After processing by the surface roughness prediction module, the surface roughness is obtained.

[0110] In summary, this paper extracts multidimensional, fine-grained features from force and vibration signals based on SVMD and WPD, effectively characterizing detailed information about surface roughness caused by inherent process and chatter effects during milling. A novel fine-grained feature fusion network is constructed to effectively fuse the multidimensional, fine-grained features of periodic and chatter signals, enabling intelligent surface roughness prediction.

[0111] The advantage of this invention is that it can improve the accuracy of surface roughness prediction during milling, providing an effective solution to the problem of surface roughness prediction in real-world scenarios where chatter may occur. This will effectively ensure the stability of surface quality during machining, thereby reducing scrap rates and downtime.

[0112] The above description is merely an illustration of the preferred embodiments of the present disclosure and the technical principles employed. Those skilled in the art should understand that the scope of the invention encompassed by the embodiments of the present disclosure is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the aforementioned inventive concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A method for predicting surface roughness in high-speed milling considering chatter, characterized in that: include: Step 1: Perform a plane high-speed milling experiment on the workpiece on a milling table. Continuously collect force signals through the force sensor configured on the milling table, and continuously collect vibration signals through the acceleration sensor configured on the milling table. Step 2: Use the continuously collected force signal and the continuously collected vibration signal as input samples, use the surface roughness corresponding to the input samples as output samples, and construct a training set based on the input samples and output samples; Step 3: For the input samples in the training set, perform multidimensional fine-grained feature extraction on the force signal and vibration signal in the input samples and then splice them to obtain the multidimensional fine-grained feature components of the periodic signal and the multidimensional fine-grained feature components of the chatter signal; Step 4: Input the multi-dimensional fine-grained feature components of the periodic signal and the multi-dimensional fine-grained feature components of the chatter signal into a fine-grained feature fusion network to obtain a predicted value of surface roughness. The fine-grained feature fusion network includes a DREF module, a FASM module, and a Bayesian prediction module. Step 5: Based on the predicted values ​​of multiple surface roughness and their corresponding output samples, the parameters of the DREF module, FASM module, and Bayesian prediction module in the fine-grained feature fusion network are iteratively updated multiple times to obtain the trained fine-grained feature fusion network; Step 6: Continuously collect the force signal and vibration signal to be predicted, and then execute steps 3-4 based on the trained fine-grained feature fusion network to obtain the surface roughness; Among them, step 3 specifically includes: Step 3.1: Use the wavelet packet decomposition (WPD) method to reduce the noise of the vibration signal to obtain a reduced-noise vibration signal. Then, use successive variational mode decomposition (SVMD) to extract the first periodic signal and the first chatter signal from the reduced-noise vibration signal. Step 3.2: Extract the second periodic signal and the second chatter signal from the force signal through successive variational mode decomposition (SVMD); Step 3.3: Perform feature extraction on the first periodic signal, the first dither signal, the second periodic signal, and the second dither signal, respectively, to obtain multidimensional fine-grained features of the first periodic signal, the first dither signal, the second periodic signal, and the second dither signal, wherein the multidimensional fine-grained features include time domain features and frequency domain features. Step 3.4: Perform vector concatenation on the multidimensional fine-grained features of the first periodic signal and the multidimensional fine-grained features of the second periodic signal to obtain a multidimensional fine-grained feature component of the periodic signal. Perform vector concatenation on the multidimensional fine-grained features of the first chatter signal and the multidimensional fine-grained features of the second chatter signal to obtain a multidimensional fine-grained feature component of the chatter signal. Among them, step 4 specifically includes: Step 4.1: Based on the DREF module in the fine-grained feature fusion network, feature extraction and fusion are performed on the multi-dimensional fine-grained feature components of the periodic signal and the multi-dimensional fine-grained feature components of the chatter signal to obtain fused features; Step 4.2: Process the fused features through SENet and BiLSTM in the FASM module to obtain high-level features; Step 4.3: Input the high-level features into the Bayesian prediction module to obtain the predicted value of surface roughness.

2. The method for predicting surface roughness in high-speed milling considering chatter according to claim 1, characterized in that: Step 4.1 specifically includes: Step 4.1.1: Perform deep feature extraction on the multidimensional fine-grained feature components of the periodic signal to obtain deep features of the periodic signal. Specifically, perform a convolution operation, batch normalization (BN), ReLU, maximum pooling, a convolution operation, batch normalization (BN), ReLU, and maximum pooling on the multidimensional fine-grained feature components of the periodic signal to obtain a first deep feature. Pass the multidimensional fine-grained feature components of the periodic signal through a fully connected layer to obtain a second deep feature. Perform a residual connection between the first deep feature and the second deep feature by adding them to obtain the deep feature of the periodic signal. Step 4.1.2: Process the deep features of the periodic signal through the gating mechanism composed of sigmoid layers and tanh layers to obtain the first gating weight matrix ξ1, which is specifically implemented by the following formula: Among them, μ and ν are learning parameters, f1 is the deep feature of the periodic signal, Represents element-wise multiplication, and σ represents the sigmoid layer; The first gating weight matrix ξ1 is convolved by three one-dimensional convolutions with different receptive fields to obtain first-scale features, second-scale features, and third-scale features. The first-scale features, second-scale features, and third-scale features are then concatenated to obtain first multi-scale features. The first multi-scale features are input into the FC layer. In the FC layer, the first multi-scale features are activated by ReLU and sigmoid functions to obtain first adaptive weights. Step 4.1.3: Perform deep feature extraction on the multidimensional fine-grained feature components of the chatter signal to obtain deep features of the chatter signal. Specifically, perform a convolution operation, batch normalization (BN), ReLU, max pooling, convolution operation, batch normalization (BN), ReLU, and max pooling on the multidimensional fine-grained feature components of the chatter signal to obtain a third deep feature. Pass the multidimensional fine-grained feature components of the chatter signal through a fully connected layer to obtain a fourth deep feature. The third and fourth deep features are residually connected by addition to obtain the deep feature of the chatter signal. Step 4.1.4: Process the deep features of the chatter signal through a gating mechanism consisting of a sigmoid layer and a tanh layer to obtain the second gating weight matrix ξ2, which is implemented using the following formula: Where f2 is the depth feature of the chatter signal; The second gated weight matrix ξ2 is convolved by one-dimensional convolution of three different receptive fields to obtain the fourth scale feature, the fifth scale feature and the sixth and fifth scale features, and then the fourth scale feature, the fifth scale feature and the sixth and fifth scale features are concatenated to obtain the second multi-scale feature, which is input into the FC layer. In the FC layer, the second multi-scale feature is activated by the ReLU activation function and the sigmoid activation function to obtain the second adaptive weight; Step 4.1.5: Use the first adaptive weight and the second adaptive weight to perform weighted summation on the depth features of the periodic signal and the depth features of the chatter signal to obtain the fusion feature, which is specifically achieved by the following formula: Among them, F represents the fusion feature, ω p represents the first adaptive weight, F p Represents the deep features of periodic signals, ω c represents the second adaptive weight, F c represents the depth feature of the chatter signal, Indicates addition of elements.

3. The method for predicting surface roughness in high-speed milling considering chatter according to claim 1, characterized in that: Step 4.2 specifically includes: Step 4.2.1: In SENet, the fused features are subjected to global pooling, ReLU activation function, and sigmoid activation function to obtain channel attention weights. The channel attention weights and the fused features are then multiplied by the scale operation to obtain recalibrated fused features. Step 4.2.2: Process the recalibrated fused features through BiLSTM to obtain high-level features.

4. The method for predicting surface roughness in high-speed milling considering chatter according to claim 1, characterized in that: Step 4.3 specifically includes: In the Bayesian prediction module, high-level features are converted into one-dimensional vectors through the compression layer. The one-dimensional vector is passed through the ReLU activation function in the fully connected layer to obtain a vector after enhanced nonlinearity. The enhanced nonlinear vector is processed by approximate Bayesian dropout to obtain the initial surface roughness. The same input sample is repeatedly input into the fine-grained feature fusion network through Monte Carlo random sampling to obtain multiple initial surface roughnesses. The average of all initial surface roughnesses is calculated to obtain the predicted value of surface roughness.

5. A high-speed milling surface roughness prediction system considering chatter, characterized in that: A method for predicting surface roughness of high-speed milling considering chatter as described in claim 1, comprising a data acquisition module, a model training module, a data processing module and a surface roughness prediction module; The data acquisition module is used to continuously collect force signals and vibration signals, and transmit the force signals and vibration signals to the model training module; The model training module is used to use the continuously collected force signal and the continuously collected vibration signal as input samples, and the surface roughness corresponding to the input samples as output samples, to construct a training set based on the input samples and the output samples, and to transmit the training set input samples to the data processing module; the model training module is also used to receive the predicted value of the surface roughness, and based on the multiple predicted values ​​of the surface roughness and their corresponding output samples, to iteratively update the parameters in the fine-grained feature fusion network multiple times; The data processing module is used to receive input samples, perform multi-dimensional fine-grained feature extraction and splicing on the force signal and vibration signal in the input samples, and transmit the obtained multi-dimensional fine-grained feature components of the periodic signal and the multi-dimensional fine-grained feature components of the chatter signal to the surface roughness prediction module; The surface roughness prediction module is used to receive the multidimensional fine-grained feature components of the periodic signal and the multidimensional fine-grained feature components of the chatter signal, and input the multidimensional fine-grained feature components of the periodic signal and the multidimensional fine-grained feature components of the chatter signal into a fine-grained feature fusion network to obtain a predicted value of the surface roughness, and transmit the predicted value of the surface roughness to the model training module.

Images