Multi-channel online monitoring device and method for chatter in robotic milling

Abstract

This invention relates to the field of robot machining performance analysis technology, and particularly to a multi-channel online monitoring device and method for chatter in robot milling. The invention employs successive multiple variational mode decomposition to decompose the acquired three-channel acceleration signal into multiple sets of intrinsic mode functions with mode alignment characteristics, thereby extracting the common frequency components between channels. This method not only adaptively decomposes the signal but also effectively captures the physical correlations between different channels. The invention proposes a dimensionless chatter discrimination index—frequency ratio—for automatically extracting chatter components from the decomposed signal. Subsequently, the energy ratio sequence between the chatter signal and the milling signal is used as input to train a deep convolutional neural network model to achieve chatter detection. This method avoids manually setting chatter index thresholds and improves the generalization ability of chatter detection.

Description

Technical Field

[0001] This invention relates to the field of robot machining performance analysis technology, and in particular to a multi-channel online monitoring device and method for chatter in robot milling. Background Technology

[0002] In recent years, robotic milling technology for large structural components has been increasingly widely used in industries such as aerospace and automotive manufacturing. Compared with traditional machine tools, robots have significant advantages such as lower cost and higher flexibility when machining large and complex parts. However, due to the relatively weak rigidity of the milling robot body, chatter is prone to occur during milling, severely limiting machining quality and production efficiency. Therefore, timely and accurate online chatter monitoring is crucial for achieving high-precision and high-efficiency machining. Traditional methods based on single-channel signals require manually setting chatter thresholds for state classification, which is often complex and time-consuming under varying machining conditions. Furthermore, due to the nonlinearity of the milling robot structure, vibration signals in the x, y, and z directions will couple when chatter occurs. Treating the signals in each direction as independent components ignores their physical correlation, thus affecting the performance of chatter monitoring. Therefore, avoiding manually setting chatter thresholds and utilizing the physical correlation between multi-channel signals have become key technical bottlenecks restricting high-precision chatter monitoring in robotic milling.

[0003] The paper "K. Yang, G. Wang, Y. Dong, Q. Zhang, L. Sang, Early chatteridentification based on an optimized variational mode decomposition, Mechanical Systems and Signal Processing, 115 (2019) 238-254" discloses a chatter monitoring method based on optimized variational mode decomposition. This method processes the cutting force signal using optimized variational mode decomposition and extracts chatter components using a simulated annealing algorithm. Chatter monitoring is achieved by calculating sample entropy and approximate entropy. However, the determination of the machining state still relies on manual observation of the changes in sample entropy and approximate entropy, making fully automated monitoring difficult to achieve.

[0004] The paper "S. Chen, Z. Liang, Y. Du, Z. Gao, H. Zheng, Z. Liu, T. Qiu, X. Wang, Early chatter identification based on optimized VMD with multi-band information fusion and compression method in robotic milling process, Chinese Journal of Aeronautics, 37 (2024) 464-484" discloses a novel chatter identification method based on multi-band information fusion and compression technology. This method utilizes the mayfly optimization algorithm to obtain globally optimal variational mode decomposition parameters and uses the image cross-entropy as a chatter indicator to identify chatter. However, this method only considers acceleration signals in a single direction, ignoring the physical correlation between signals in different directions, thus reducing the performance of chatter monitoring. Summary of the Invention

[0005] To address the limitations of existing methods, this invention provides a multi-channel online monitoring device, method, and system for chatter in robotic milling, which solves the problems of manually setting chatter thresholds and ignoring the physical correlation between multiple channels.

[0006] In a first aspect, embodiments of the present invention provide a multi-channel online monitoring device for chatter in robotic milling, the multi-channel online monitoring device for chatter in robotic milling includes a robotic milling machine, a three-axis accelerometer, a data acquisition device, and a computer, wherein... The robotic milling equipment includes an industrial robot, a spindle, a tool holder, and cutting tools, used to mill workpieces to generate vibration signals; The triaxial accelerometer is attached to the spindle of the robot milling equipment and connected to the data acquisition device. It is used to detect the vibration signal generated during the milling of the workpiece and transmit the vibration signal to the data acquisition device. The data acquisition device is connected to the triaxial accelerometer and the computer, and is used to acquire the vibration signal transmitted by the triaxial accelerometer and transmit the data to the computer. The computer includes a data processing module and a flutter monitoring module, which are connected to the data acquisition device. The data processing module processes the vibration signal transmitted by the data acquisition device and transmits the features obtained from processing the vibration signal to the flutter monitoring module for identifying the processing status.

[0007] Secondly, the present invention also provides a multi-channel online monitoring method for chatter in robot milling, the multi-channel online monitoring method for chatter in robot milling includes: Step 101: Collect the robot's acceleration signals in the x, y and z directions under different processing conditions, and divide the signals into three states based on the number and amplitude of the chatter components in the frequency domain acceleration signals and the processing surface quality: stable, transitional and chattering. Step 102: Use the sliding window method to segment the acceleration signal to generate sample data required for model training, and divide the sample data into training set, validation set and test set. The validation set has the same processing conditions as the training set, while the test set is obtained under different processing conditions. Step 103: Perform successive multivariable mode decomposition on the original acceleration signal to obtain multiple signal components with different center frequencies. At the same time, use the frequency ratio index to identify the flutter components and combine them to form a flutter signal. Step 104: Calculate the energy ratio sequence between the chatter signal and the milling signal using the sliding window method; Step 105: Construct a deep convolutional neural network model consisting of Inception, a convolutional block attention module (CBAM), a residual network (ResNet), and a classification module; Step 106: Use the energy ratio sequence as the input to the deep convolutional neural network model and the processing state as the output to train the model. Evaluate the model performance on the validation set and test set to determine the optimal hyperparameters and construct the flutter monitoring module. Step 107: Acceleration signals in three directions are detected using a triaxial accelerometer, and the detected signals are collected using a data acquisition device. Next, the chatter component is extracted and the energy ratio sequence is calculated using the data processing module comprised of steps 103 and 104. Finally, the processing status is identified using the chatter monitoring module constructed in step 106, achieving online monitoring of chatter.

[0008] Thirdly, the present invention also provides a multi-channel online monitoring system for chatter in robot milling, wherein the multi-channel online monitoring system for chatter in robot milling includes any of the multi-channel online monitoring devices for chatter in robot milling described in the present invention.

[0009] The multi-channel online monitoring method for chatter in robotic milling provided in this application has at least the following advantages compared with the prior art: 1. This invention employs successive multiple variational mode decomposition to decompose the acquired three-channel acceleration signal into multiple sets of intrinsic mode functions with mode alignment characteristics, thereby extracting the common frequency components between channels. This method not only adaptively decomposes the signal but also effectively captures the physical correlation between different channels.

[0010] 2. This invention proposes a dimensionless chatter discrimination index—frequency ratio—for automatically extracting chatter components from the decomposed signal. Subsequently, the energy ratio sequence between the chatter signal and the milling signal is used as input to train a deep convolutional neural network model to achieve chatter detection. This method avoids manually setting chatter index thresholds and improves the generalization ability of chatter detection.

[0011] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart of the multi-channel online monitoring method for chatter in robot milling provided by the present invention; Figure 2 A schematic diagram of the multi-channel online monitoring device for chatter in robotic milling provided by the present invention; Figure 3 Frequency domain acceleration signals and surface views of the machining side under different machining conditions are provided by this invention; Figure 4 The present invention provides the time-domain waveforms of the acceleration signal in three channels, the spectrum diagrams, and the component spectrum diagrams obtained by successive multivariable mode decomposition. Figure 5 This is a structural diagram of the deep convolutional neural network model proposed in this invention; Figure 6 A schematic diagram of the loss and accuracy curves of the deep convolutional neural network model provided by this invention on the training and validation sets in three scenarios; Figure 7 A schematic diagram of the confusion matrix on the verification and test sets of the deep convolutional neural network model provided by the present invention in three scenarios; Figure 8 This diagram illustrates the accuracy comparison of the proposed method and the use of a single-channel acceleration signal as model input on a test set. Detailed Implementation

[0014] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected. Example

[0015] like Figure 1 As shown, the multi-channel online monitoring method for chatter in robot milling provided by the present invention includes the following steps: Step 101: Collect the robot's acceleration signals in the x, y and z directions under different processing conditions, and divide the signals into three states based on the number and amplitude of the chatter components in the frequency domain acceleration signals, combined with the processing surface quality: stable, transitional and chattering.

[0016] Specifically, this invention conducted a series of milling experiments on different materials, cutting tools, and various robot poses. The robot used in the robotic milling equipment 200 was a KUKA KR210. The workpiece materials were aluminum alloy 7050 and nickel-aluminum bronze, both with dimensions of 120×80×100mm. Cutting tool 1 was used to machine aluminum alloy 7050, while cutting tools 2 and 3 were used to machine NiAB. All three cutting tools were carbide flat-bottom end mills, and their detailed parameters are listed in Table 1. Specific milling experiment parameters are detailed in Tables 2 and 3, and the joint angle parameters for each pose in the experiment are listed in Table 4. The feed rate was set to 300mm / min.

[0017] A schematic diagram of a multi-channel online monitoring device for chatter in robotic milling is shown below. Figure 2 As shown. During the milling process, a triaxial accelerometer 201 (PCB 356A24) is used to detect acceleration signals in the x, y, and z directions. The detected signals are collected by a data acquisition device 202 (LMS-SCADAS-Mobile-05), with a sampling frequency set to 6400Hz. The triaxial accelerometer 201 and the data acquisition device 202 are connected via a 034G05 cable. Finally, the data acquisition device 202 transmits the collected signal data to a computer 203 for storage and analysis via a network cable.

[0018] After milling, the machining state is determined based on the number and amplitude of chatter components in the frequency domain acceleration signal, combined with the surface finish, and each set of cutting parameters is labeled. The frequency domain acceleration signals and surface finishes of aluminum alloy 7050 and nickel-aluminum bronze under different machining states are shown below. Figure 3 As shown. Figure 3(a)-(f) represent the side view and frequency domain acceleration signal obtained from machining aluminum alloy 7050 in experiments 1, 3 and 6, respectively. Figure 3 (g)-(l) represent the side views and frequency domain acceleration signals obtained from machining nickel-aluminum bronze in experiments 70, 74, and 71, respectively. The dashed lines in the figures represent the spindle frequency and its harmonics. It can be observed that for different workpiece materials and tool combinations, the surface damage characteristics caused by chatter and the variation law of the frequency domain acceleration signal show similar trends. Therefore, the machining state is divided into three categories: stable state, transitional state, and chatter state. In the stable machining state, the workpiece surface is uniform and flat, with no obvious tool marks, and no chatter component is detected in the frequency domain acceleration signal. When machining enters the transitional state, irregular tool marks begin to appear on the workpiece surface, and the machining quality decreases significantly; at the same time, a small number of chatter components appear in the frequency domain acceleration signal, indicating that the energy of the periodic component begins to transfer to the chatter component. When chatter develops fully, the tool mark depth on the workpiece surface increases significantly, the number of chatter components in the frequency domain acceleration signal increases significantly, and the proportion of their amplitude relative to the periodic component further increases. These differences lead to a richer feature diversity in the milling signal, which in turn affects the accuracy of chatter monitoring. Throughout the experiment, a super depth-of-field microscope was used to observe the processed surface, and LMS software was used to process the collected acceleration signals.

[0019] Table 1. Geometric parameters of the cutting tool

[0020] Table 2 Milling experiment using tool 1 and aluminum alloy 7050 workpiece (radial depth of cut 3mm)

[0021]

[0022] Table 3 Milling experiments using tools 2 and 3 and nickel-aluminum bronze (radial depth of cut 8 mm).

[0023] Table 4. Robot joint angles corresponding to different poses in the milling experiment. A

[0024] Step 102: Use the sliding window method to segment the acceleration signal to generate sample data required for model training, and divide the sample data into training set, validation set and test set. The validation set has the same processing conditions as the training set, while the test set is obtained under different processing conditions.

[0025] Specifically, all experimental data were segmented using a sliding window method to generate samples. The window size was set to 1 second, and the step size was 0.1 seconds, resulting in a total of 7881 samples. The experimental data was divided into three groups: groups 1-26 (D1), 27-52 (D2), and 53-78 (D3). Cross-validation was used to construct three experimental scenarios, where any two data groups were divided into training and validation sets in an 8:2 ratio, with the remaining group serving as the test set. The data partitioning for the three scenarios is shown in Table 5. In each scenario, the training and validation sets corresponded to the same processing conditions, while the training and test sets corresponded to different processing conditions.

[0026] Table 5 Dataset partitioning for different scenarios

[0027] Step 103: Perform successive multivariable mode decomposition on the original acceleration signal to obtain multiple signal components with different center frequencies. At the same time, use the frequency ratio index to identify the flutter components and combine them to form a flutter signal. Specifically, the acceleration signal with a duration of one second in sequence number 46. S 46 For example, its time-domain waveforms in the three channels, the corresponding spectrum diagrams, and the component spectra after successive multivariable mode decomposition are shown below. Figure 4 As shown in (a)-(c), the decomposed signal mainly consists of periodic and dizzy components, while noise components have been significantly filtered out during the decomposition process. The signal components corresponding to each channel maintain mode alignment characteristics, indicating that successive multivariable mode decomposition can preserve the physical correlation between different channels. To effectively extract dizzy components from the decomposed signal, the frequency ratio is selected as the evaluation index for dizzy components. The frequency ratio represents the value of dividing the remainder of the signal component's main frequency by the spindle fundamental frequency, and its expression is as follows: (1) In MATLAB, mod represents the modulo operation. f o Indicates the dominant frequency of the signal component. f s This indicates the fundamental frequency of the spindle. The present invention will... FR ≤0.05 or FR Signal components with a value ≥0.95 are defined as periodic components, and the remaining components are defined as flutter components.

[0028] Step 104: Calculate the energy ratio sequence between the chatter signal and the milling signal using the sliding window method; Specifically, the length of the sliding window is set to 320, thus compressing 6400 original data points into 20 data points. Within each sliding window, the energy of the chatter signal and the milling signal is calculated separately. (2) in, , and These represent the flutter signals in the x, y, and z directions, respectively. , and This indicates the milling signal in the corresponding direction. , , , , and These are the corresponding energy values. Finally, the energies of the flutter signals in the three directions are summed to obtain the total energy of the flutter signals: (3) The same method is also applied to milling signals to obtain their total energy: (4) The energy ratio is defined as the ratio of the total energy of the chatter signal to the total energy of the milling signal: (5) The energy ratio sequence is expressed as: (6) Step 105: Construct a deep convolutional neural network model consisting of Inception, a convolutional block attention module (CBAM), a residual network (ResNet), and a classification module; Specifically, the structure of a deep convolutional neural network model is as follows: Figure 5 As shown, all convolutional kernels are 1D, and the stride is uniformly set to 1. Figure 5 As shown in (a), the Inception structure consists of multiple parallel branches, and the outputs of each branch are ultimately fused through a concatenation operation. Each branch has a different receptive field to capture different characteristics of the input features. The first branch uses 1×1 convolutions for feature dimension compression and efficient feature transformation; the second branch consists of concatenated 1×1 and 3×1 convolutions, aiming to extract medium-scale local features; the third branch combines 1×1 and 5×1 convolutions to expand the receptive field and capture more comprehensive feature information; the fourth branch consists of 3×1 max pooling layers and 1×1 convolutions, which helps to maintain local invariance and enhance feature diversity.

[0029] The goal of CBAM is to introduce an adaptive attention mechanism into convolutional neural networks without significantly increasing computational overhead, thereby enhancing the representational power of extracted flutter features. For example... Figure 5 As shown in (b), the feature map generated by Inception As input, CBAM sequentially generates one-dimensional channel attention maps. And one-dimensional attention mapping ,in L Indicates the length of one-dimensional data. W This represents the number of feature maps. The entire attention process can be represented as: (7) in This indicates element-wise multiplication.

[0030] like Figure 5 As shown in (c), ResNet consists of two parts: the main path and the shortcut path. In the main path, the input features... Redundant information is removed using the Inception and CBAM modules to obtain multi-level features. In the shortcut path, the input features are passed directly to the output without any additional transformation. The final output of ResNet... The result is obtained by element-wise addition of the convolutional output of the main path and the identity mapping of the shortcut path, followed by the Leaky ReLU activation function. The entire process can be represented as: (8) in This indicates element-wise addition.

[0031] The classification module is used for processing high-level features from ResNet. H Process and output the processing status. For example... Figure 5 As shown in (d), this module consists of an average pooling layer and a random deactivation layer. The average pooling operation is applied to each feature map, pooling layers of size 1× L × M Feature matrix H Compressed to a size of M eigenvector of ×1 y , represented as This process can be represented as: (9) in, h j,i In the characteristic matrix H The Middle j The feature map, the first i The feature values ​​at each position.

[0032] Step 106: Use the energy ratio sequence as the input of the deep convolutional neural network model and the processing state as the output to train the model, and evaluate the model performance on the validation set and test set to determine the optimal hyperparameters and construct the flutter monitoring module 205. Specifically, the model ultimately employs the cross-entropy loss function and the Adam optimizer, with a learning rate of 0.001, a random kill rate of 0.1, a batch size of 64, and 50 training epochs. The loss versus accuracy curves for the training and validation sets in the three scenarios are shown below. Figure 6 As shown in the figure, in all three scenarios, the model's accuracy on the training set stabilized after a certain number of rounds, reaching 97.67%, 97.41%, and 98.74%, respectively. The loss value also stabilized, gradually decreasing from a high initial value to 0.0624, 0.0696, and 0.0401. Meanwhile, the model's accuracy and loss on the validation set were essentially consistent with those on the training set. The trends in accuracy and loss indicate that the proposed model possesses high accuracy and strong generalization ability. To further demonstrate the classification performance of the proposed model, confusion matrices for the three scenarios were plotted, as shown in the figure. Figure 7 As shown in the figure. The results indicate that the trained model exhibits strong classification ability in all three scenarios.

[0033] Step 107: Acceleration signals in three directions are detected using a triaxial accelerometer 201, and the detected signals are collected by a data acquisition device 202. Next, the chatter component is extracted and the energy ratio sequence is calculated by the data processing module 204, which consists of steps 103 and 104. Finally, the processing status is identified by the chatter monitoring module 205 constructed in step 106, achieving online monitoring of chatter.

[0034] Specifically, firstly, a triaxial accelerometer 201 is attached to the robot's spindle to detect acceleration signals in the x, y, and z directions. The detected signals are collected by a data acquisition device 202 at a sampling frequency of 6400Hz. Next, the data acquisition device 202 transmits the collected signal data to a computer 203 via a network cable for storage and analysis. When the computer 203 receives the acceleration signal, the data processing module 204, composed of steps 103 and 104, extracts the chatter component and calculates the energy ratio sequence. Finally, using the energy ratio sequence as input, the chatter monitoring module 205, constructed in step 106, identifies the processing status, achieving online chatter monitoring.

[0035] To verify that the proposed method can improve the accuracy of flutter monitoring by utilizing the physical correlation between different channels, this invention processes the acceleration signals in the x, y, and z directions separately through successive multivariable mode decomposition, and uses the resulting energy ratio sequence as input to train a deep convolutional neural network model. The results are compared with the method proposed in this invention, such as... Figure 8 As shown, the proposed method achieved the highest monitoring accuracy in all scenarios, indicating that considering the physical correlation of multiple channels can improve the accuracy of flutter monitoring.

[0036] To verify that the proposed method can achieve high accuracy in chatter monitoring under different machining conditions by avoiding manual setting of chatter threshold, a series of milling experiments were conducted under different robot poses, cutting parameters, workpiece materials and tool conditions. Figure 6 and Figure 7 The results show that the method proposed in this invention exhibits strong classification ability in all three scenarios, indicating that the flutter monitoring method proposed in this invention has high accuracy and generalization ability.

Claims

1. A multi-channel online monitoring method for chatter in robotic milling, characterized in that, include: Step 101: Collect the robot's acceleration signals in the x, y, and z directions under different processing conditions, and divide the acceleration signals into three states: stable, transitional, and chattering; Step 102: Use the sliding window method to segment the acceleration signal to generate sample data required for model training, and divide the sample data into training set, validation set and test set. The validation set has the same processing conditions as the training set, while the test set is obtained under different processing conditions. Step 103: Perform successive multivariable mode decomposition on the original acceleration signal to obtain multiple signal components with different center frequencies. At the same time, use the frequency ratio index to identify the flutter components and combine them to form a flutter signal. Step 104: Calculate the energy ratio sequence between the chatter signal and the milling signal using the sliding window method; Step 105: Construct a deep convolutional neural network model consisting of Inception, a convolutional block attention module (CBAM), a residual network (ResNet), and a classification module; Step 106: Use the energy ratio sequence as the input to the deep convolutional neural network model and the processing state as the output to train the model. Evaluate the model performance on the validation set and test set to determine the optimal hyperparameters and construct the flutter monitoring module. Step 107: Use a triaxial accelerometer to detect acceleration signals in three directions and collect the detected signals through a data acquisition device; extract the chatter component and calculate the energy ratio sequence through the data processing module composed of steps 103 and 104; finally, identify the processing status through the chatter monitoring module constructed in step 106 to realize online monitoring of chatter.

2. The method according to claim 1, characterized in that, Acceleration signals are classified based on the number and amplitude of chatter components in the frequency domain acceleration signal, combined with the surface quality of the processed area.

3. The method according to claim 1, characterized in that, In step 103, the energy of the flutter signals in the three directions is summed to obtain the total energy of the flutter signals.

4. The method according to claim 1, characterized in that, In step 105, all convolutional kernels are 1D with a uniform stride.

5. The method according to claim 1, characterized in that, The Inception structure consists of multiple parallel branches, and the outputs of each branch are ultimately merged through a concatenation operation.

6. The method according to claim 5, characterized in that, Each branch has a different receptive field, used to capture different characteristics of the input features.

7. The method according to claim 1, characterized in that, ResNet consists of two parts: the main path and the shortcut path. In the main path, the input features are processed by the Inception and CBAM modules to remove redundant information and obtain multi-level features. In the shortcut path, the input features are directly passed to the output without any additional transformation.

8. A multi-channel online monitoring device for chatter in robotic milling, characterized in that, include: Robotic milling equipment, three-axis accelerometer, data acquisition equipment, and computer, among which... The robotic milling equipment includes an industrial robot, a spindle, a tool holder, and cutting tools, used to mill workpieces to generate vibration signals; The triaxial accelerometer is attached to the spindle of the robot milling equipment and connected to the data acquisition device. It is used to detect the vibration signal generated during the milling of the workpiece and transmit the vibration signal to the data acquisition device. The data acquisition device is connected to the triaxial accelerometer and the computer, and is used to acquire the vibration signal transmitted by the triaxial accelerometer and transmit the data to the computer. The computer includes a data processing module and a flutter monitoring module, which are connected to the data acquisition device. The data processing module processes the vibration signal transmitted by the data acquisition device and transmits the features obtained from processing the vibration signal to the flutter monitoring module for identifying the processing status.

Images