A method and system for predicting the surface roughness of ultra-precision turning by integrating vibration and process parameters
By integrating process parameters and high-frequency vibration characteristics into a deep learning model, the problem of insufficient surface roughness prediction accuracy in ultra-precision turning is solved, achieving high-precision and efficient surface roughness prediction, which is applicable to a variety of microstructures.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2026-05-11
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies cannot effectively integrate physical mechanisms and data-driven approaches in ultra-precision turning, resulting in insufficient accuracy in surface roughness prediction, especially in the turning of complex microstructures where errors are significant. Furthermore, traditional methods ignore the influence of high-frequency vibrations.
By extracting and constructing multi-source features, combining process and location, microstructure geometry, and high-frequency vibration amplitude features, a cross-attention deep learning model is used to predict surface roughness. High-frequency vibration features are forcibly fused with physical mechanisms to achieve high-precision prediction.
It achieves a prediction and classification accuracy of over 60% on various microstructure surfaces, which is significantly better than the baseline model that does not consider vibration characteristics, and has higher computational efficiency and physical interpretability.
Smart Images
Figure CN122165243B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of ultra-precision machining and intelligent manufacturing technology, and in particular to a method and system for predicting the surface roughness of ultra-precision turning by integrating vibration and process parameters. Background Technology
[0002] Surface roughness is one of the core indicators for evaluating the quality of ultra-precision turning, directly affecting the performance of optical components and functional surfaces of microstructures. Achieving high-precision prediction of surface roughness before or during machining is crucial for process optimization, quality control, and intelligent compensation.
[0003] Existing prediction methods are mainly divided into two categories:
[0004] Empirical / theoretical modeling methods: These methods rely on empirical formulas or theoretical models based on cutting parameters, including feed rate, depth of cut, and spindle speed. However, these methods severely neglect the high-frequency vibrations generated dynamically during machining, which are a direct cause of surface texture deterioration and increased roughness. Therefore, their prediction accuracy is limited, especially in the machining of complex microstructures where errors are significant.
[0005] Purely data-driven methods: These methods directly utilize process parameters or sensor signals, employing machine learning or deep learning models for prediction. These methods have significant drawbacks:
[0006] Features are disconnected from physical mechanisms: The sensor signals are often simply piled up without targeted feature engineering based on physical mechanisms (such as the influence of vibration on surface formation), resulting in poor model interpretability and weak generalization ability.
[0007] Ignoring the "weak correlation and strong causality" characteristic: Key physical quantities such as high-frequency vibration signals may exhibit weak correlation in statistics due to their instantaneous nature and scale difference with the average roughness space. They are easily mistakenly eliminated by traditional feature screening methods (such as simply relying on the Pearson coefficient), thus losing key causal information.
[0008] Therefore, there is an urgent need for a high-precision surface roughness prediction method that can deeply integrate physical mechanisms and data-driven approaches, especially one that can effectively incorporate and process dynamic process features such as vibration. Summary of the Invention
[0009] The purpose of this invention is to provide a method and system for predicting the surface roughness of ultra-precision turning by integrating vibration and process parameters, thereby solving the problems mentioned in the background art.
[0010] To achieve the above objectives, this invention provides a method for predicting the surface roughness of ultra-precision turning by integrating vibration and process parameters, comprising the following steps:
[0011] S1. Perform multi-source feature extraction and construction to obtain and construct three sets of input features;
[0012] S2. Preprocess the three sets of input features constructed;
[0013] S3. Simultaneously input the three preprocessed sets of input features and labels into the deep learning prediction model for training to obtain the trained deep learning prediction model.
[0014] S4. Input three sets of input features into the trained deep learning prediction model and output a surface roughness distribution prediction map.
[0015] The three sets of input features in S1 are as follows:
[0016] The first group consists of process and position characteristics, including the radial position of the tool, circumferential angle, cutting axis displacement, velocity, and acceleration.
[0017] The second group consists of microstructural geometric features, including the first-order gradient of the surface in the circumferential direction, the first-order gradient in the radial direction, and the second-order gradient in the radial direction.
[0018] The third group consists of high-frequency vibration amplitude features, which are extracted from the three-dimensional cutting force signal to represent the vibration amplitude of specific frequency components.
[0019] Preferably, the extraction of high-frequency vibration amplitude features includes:
[0020] S11. Perform high-pass filtering on the original cutting force signal to separate the high-frequency vibration signal component;
[0021] S12. For the separated high-frequency vibration signal component, within a preset number of key frequency neighborhood windows, calculate half of the difference between the peak value and the valley value of the high-frequency vibration signal component in the window, and use it as the approximate vibration amplitude of the frequency component.
[0022] S13. Screen the amplitudes of the same frequency of cutting forces in different directions that have strong correlations, and retain only the amplitude characteristics of representative directions to reduce the feature dimension.
[0023] Preferably, S2 includes:
[0024] S21. Divide the machined surface into a two-dimensional grid;
[0025] S22. Organize the twelve feature values corresponding to each grid into a sample, construct a two-dimensional array of input features for all samples, and output the label as the feature.
[0026] Preferably, in step S21, for planar microstructures, the radial division is performed every 100 μm, and the circumferential division is performed according to the principle of equal arc length; for cylindrical microstructures, they are first unfolded along the circumferential direction, and then the radial division is performed every 100 μm, and the circumferential division is performed according to the principle of equal arc length.
[0027] Preferably, the label in S22 is the measured surface roughness value.
[0028] Preferably, the deep learning prediction model in S3 employs a cross-attention network, whose attention mechanism is used to capture the long-range spatial dependencies of the two-dimensional feature map in both radial and circumferential dimensions, which is consistent with the directional physical mechanism of surface roughness formation.
[0029] Preferably, when the deep learning prediction model in S3 is trained, it uses the radial quadratic gradient features and high-frequency vibration amplitude features as key inputs.
[0030] Preferably, in the training process, the deep learning prediction model in S3 outputs a predicted surface roughness value based on the preprocessed input features, and calculates the loss between the predicted surface roughness value and the label, continuously iterating to reduce the loss and obtain the optimal deep learning prediction model.
[0031] The system integrates vibration and process parameter prediction methods for ultra-precision turning roughness, including a multi-source data acquisition and feature construction module, a feature spatialization processing module, an intelligent surface roughness prediction module, and a quality assessment and output module.
[0032] Therefore, the present invention employs the above-mentioned method and system for predicting the surface roughness of ultra-precision turning by integrating vibration and process parameters, and has the following beneficial effects:
[0033] (1) This method achieves a model that is closer to the physical essence by forcibly integrating the high-frequency vibration amplitude features ignored by traditional methods and combining them with geometric gradient features that have physical significance. Experiments show that the prediction and classification accuracy on various microstructure surfaces reaches more than 60%, which is significantly better than the baseline model that does not consider vibration features.
[0034] (2) The feature selection of this method, including vibration amplitude, radial quadratic gradient, etc., and model selection (such as cross attention) are all derived from the in-depth analysis of the formation mechanism of the machined surface (vibration copying, geometric interference). The model decision process has a clear physical correspondence.
[0035] (3) This method creatively proposes the “frequency window amplitude extraction method” and the “feature retention principle based on physical causality” for high-frequency vibration signals, which effectively solves the problem that dynamic process signals are difficult to be directly used for spatial quality prediction, and transforms the statistical “weak correlation” characteristics into the key to improving model performance.
[0036] (4) By comparing various attention mechanisms (CCNet, GCNet, DANet, EMANet), this method selects the cross-interference attention model that best matches the processing mechanism in terms of direction, which has higher computational efficiency while ensuring high accuracy.
[0037] (5) This method and system have been experimentally verified and can be applied to the surface roughness prediction of various types of microstructure turning, such as planes and cylinders.
[0038] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0039] Figure 1 This is a schematic diagram of 12 input features of the ultra-precision turning microstructure of the present invention, which integrates vibration and process parameter ultra-precision turning roughness prediction method and system.
[0040] Figure 2 This is a schematic diagram of high-frequency filtering processing of vibration signals in the ultra-precision turning roughness prediction method and system that integrates vibration and process parameters of the present invention.
[0041] Figure 3 This is a schematic diagram illustrating the vibration signal amplitude extraction of the ultra-precision turning roughness prediction method and system that integrates vibration and process parameters according to the present invention.
[0042] Figure 4 This is a schematic diagram of the Pearson correlation coefficient of the vibration signal in the ultra-precision turning roughness prediction method and system that integrates vibration and process parameters of the present invention.
[0043] Figure 5 This is a diagram illustrating the overall framework of the deep learning prediction model for surface roughness in the ultra-precision turning roughness prediction method and system that integrates vibration and process parameters, as presented in this invention.
[0044] Figure 6 This is a CCNet attention model framework diagram of the ultra-precision turning roughness prediction method and system that integrates vibration and process parameters of the present invention.
[0045] Figure 7 This is a GCNet attention model framework diagram of the ultra-precision turning roughness prediction method and system that integrates vibration and process parameters of the present invention.
[0046] Figure 8 This is a DANet attention model framework diagram of the ultra-precision turning roughness prediction method and system that integrates vibration and process parameters of the present invention.
[0047] Figure 9 This is a diagram of the EMANet attention model framework for the ultra-precision turning roughness prediction method and system that integrates vibration and process parameters of the present invention. Detailed Implementation
[0048] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0049] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0050] Please see Figures 1-9 This invention provides a method for predicting the roughness of ultra-precision turning by integrating vibration and process parameters. Based on an in-depth analysis of the mechanism by which vibration affects surface formation, it directionally extracts and fuses high-frequency vibration amplitude features and process geometric features, and uses a deep learning model with direction-aware capabilities for modeling. The method includes the following steps:
[0051] S1. Perform multi-source feature extraction and construction to obtain and construct three sets of input features. The three sets of input features are as follows:
[0052] The first group consists of process and positional characteristics, including the radial position of the tool, circumferential angle, cutting axis displacement, speed, and acceleration.
[0053] The second group consists of microstructure geometric features, including the first-order gradient of the surface in the circumferential direction, the first-order gradient in the radial direction, and the second-order gradient in the radial direction. Among them, the second-order gradient in the radial direction refers to the second derivative of the surface morphology of the microstructure to be processed, which is used to characterize the local rate of change of surface curvature.
[0054] The third group consists of high-frequency vibration amplitude features, which are extracted from the triaxial cutting force signal, using vibration amplitudes of specific frequency components as representative features. High-frequency vibration amplitude features refer to the vibration signal components above 10Hz extracted from the triaxial cutting force signal during ultra-precision turning through high-pass filtering, and further extracted as specific frequency amplitudes.
[0055] The extraction of high-frequency vibration amplitude characteristics includes the following steps:
[0056] S11. High-pass filtering is applied to the original cutting force signal to separate the high-frequency vibration signal component.
[0057] S12. For the separated high-frequency vibration signal component, within a preset number of key frequency neighborhood windows, calculate half of the difference between the peak and valley values of the high-frequency vibration signal component in the window, and use it as the approximate vibration amplitude of the frequency component.
[0058] S13. Screen the amplitudes of the same frequency of cutting forces in different directions that have strong correlations, and retain only the amplitude characteristics of representative directions to reduce the feature dimension.
[0059] S2. Preprocess the three sets of input features. This includes:
[0060] S21. Divide the machined surface into a two-dimensional mesh. For planar microstructures, divide radially every 100μm and circumferentially according to the principle of equal arc length; for cylindrical microstructures, first unfold along the circumferential direction, then divide radially every 100μm and circumferentially according to the principle of equal arc length.
[0061] S22. Organize the twelve feature values corresponding to each grid into a sample, construct a two-dimensional array of input features of all samples, and output the measured surface roughness value as the output feature.
[0062] S3. Simultaneously input the three sets of preprocessed input features and labels into the deep learning prediction model for training to obtain the trained deep learning prediction model.
[0063] The deep learning prediction model employs a cross-attention network, whose attention mechanism is used to capture the long-range spatial dependencies of two-dimensional feature maps in both radial and circumferential dimensions, which is highly consistent with the directional physical mechanism of surface roughness formation.
[0064] During training, the deep learning prediction model uses radial quadratic gradient features and high-frequency vibration amplitude features as key inputs. In the training process, the deep learning prediction model outputs predicted surface roughness values based on the preprocessed input features, calculates the loss by comparing the predicted surface roughness values with the labels, and iteratively reduces the loss to obtain the optimal deep learning prediction model.
[0065] S4. Input three sets of input features into the trained deep learning prediction model and output a surface roughness distribution prediction map.
[0066] A system integrating vibration and process parameter prediction methods for ultra-precision turning roughness includes a multi-source data acquisition and feature construction module, a feature spatialization processing module, a surface roughness intelligent prediction module, and a quality assessment and output module. Specifically, the multi-source data acquisition and feature construction module acquires and constructs three sets of input features in step S1; the feature spatialization processing module preprocesses the constructed three sets of input features in step S2; the surface roughness intelligent prediction module incorporates a deep learning prediction model for model training in step S3; and the quality assessment and output module outputs a predicted surface roughness distribution map in step S4.
[0067] Example 1
[0068] S1, Feature Extraction.
[0069] First, based on the processing procedure, 12 specific input features are extracted, such as... Figure 1 As shown, it includes:
[0070] F1: Workpiece radial position; F2: Workpiece circumferential position; F3: Cutting axis position; F4: Cutting axis velocity; F5: Cutting axis acceleration; F6: Microstructure circumferential gradient; F7: Microstructure radial primary gradient; F8: Tool tip radial spatial distance; F9: Radial secondary gradient; F10: Main cutting force 268Hz vibration amplitude; F11: Main cutting force 305Hz vibration amplitude; F12: Main cutting force 50Hz harmonic vibration amplitude.
[0071] The vibration amplitude characteristics (F10-F12) are extracted from the cutting force signal through the following steps:
[0072] (1) High-pass filtering is applied to the original signal to preserve high-frequency components, such as... Figure 2 .
[0073] (2) Extract vibration signals near five main frequencies (268Hz, 305Hz, 50Hz, 200Hz, 300Hz).
[0074] (3) Calculate (maximum value - minimum value) / 2 within the sliding window as the approximate amplitude for that frequency, such as... Figure 3 .
[0075] (4) Calculate the Pearson correlation coefficient among all vibration amplitude characteristics, and select representative characteristics, such as Figure 4 .
[0076] S2, Data Preprocessing and Organization.
[0077] The machined surface is divided into a two-dimensional mesh. For planar microstructures, the radial mesh is divided at 100 μm intervals, and the circumferential mesh is divided according to the principle of equal arc length. For cylindrical microstructures, they need to be unfolded along the circumferential direction first.
[0078] The twelve feature values corresponding to each grid are organized into a sample. The input features of all samples are constructed into a two-dimensional array. The output features are the measured surface roughness values within the grid (an RMS value is calculated for every 25 μm and divided into 16 categories).
[0079] S3, Model Training and Prediction.
[0080] The processed dataset is then input into the deep learning model for training. The model framework is as follows: Figure 5 As shown, an improved ResNet50 is used for downsampling (without compressing features), outputting 512-dimensional features, which are then input into the attention model.
[0081] This embodiment compares CCNet (such as...) Figure 6 As shown), GCNet (as shown) Figure 7 As shown), DANet (as shown) Figure 8 As shown), EMANet (as shown) Figure 9 (As shown) Four attention models. Figure 6 As shown, CCNet achieved the highest accuracy (0.34) in the test predicting a 200μm wavelength planar sinusoidal wave surface, and was therefore selected as the final model.
[0082] The selected CCNet model is trained for 30 rounds to obtain the final deep learning prediction model.
[0083] S4. Input three sets of input features into the final deep learning prediction model and output a surface roughness distribution prediction map.
[0084] Therefore, the present invention adopts the above-mentioned method and system for predicting the surface roughness of ultra-precision turning by integrating vibration and process parameters. By forcibly integrating the high-frequency vibration amplitude characteristics ignored by traditional methods and combining them with geometric gradient characteristics with physical significance, it achieves a model that is closer to the physical essence and is applicable to the surface roughness prediction of turning various types of microstructures such as planes and cylinders.
[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for predicting the surface roughness of ultra-precision turning by integrating vibration and process parameters, characterized in that, Includes the following steps: S1. Perform multi-source feature extraction and construction to obtain and construct three sets of input features; S2. Preprocess the three sets of input features constructed; S3. Simultaneously input the three preprocessed sets of input features and labels into the deep learning prediction model for training to obtain the trained deep learning prediction model. S4. Input three sets of input features into the trained deep learning prediction model and output a surface roughness distribution prediction map. The three sets of input features in S1 are as follows: The first group consists of process and position characteristics, including the radial position of the tool, circumferential angle, cutting axis displacement, velocity, and acceleration. The second group consists of microstructural geometric features, including the first-order gradient of the surface in the circumferential direction, the first-order gradient in the radial direction, and the second-order gradient in the radial direction. The third group consists of high-frequency vibration amplitude features, which extract the vibration amplitude of specific frequency components from the three-dimensional cutting force signal as representative features. Extraction of high-frequency vibration amplitude features includes: S11. Perform high-pass filtering on the original cutting force signal to separate the high-frequency vibration signal component; S12. For the separated high-frequency vibration signal component, within a preset number of key frequency neighborhood windows, calculate half of the difference between the peak value and the valley value of the high-frequency vibration signal component in the window, and use it as the approximate vibration amplitude of the frequency component. S13. Screen the amplitude values of the same frequency of cutting forces in different directions that have strong correlations, and retain only the amplitude characteristics of representative directions to reduce the feature dimension. The deep learning prediction model in S3 uses a cross-attention network, whose attention mechanism is used to capture the long-range spatial dependencies of two-dimensional feature maps in the radial and circumferential dimensions, which is consistent with the directional physical mechanism of surface roughness formation. When the deep learning prediction model in S3 is trained, it uses the radial quadratic gradient features and high-frequency vibration amplitude features as key inputs. Here, the radial quadratic gradient refers to the second derivative of the surface morphology of the microstructure to be processed in the radial direction.
2. The method for predicting the surface roughness of ultra-precision turning by integrating vibration and process parameters according to claim 1, characterized in that, S2 includes: S21. Divide the machined surface into a two-dimensional grid; S22. Organize the twelve feature values corresponding to each grid into a sample, construct a two-dimensional array of input features for all samples, and output the label as the feature.
3. The method for predicting the surface roughness of ultra-precision turning by integrating vibration and process parameters according to claim 2, characterized in that: In S21, for planar microstructures, the radial direction is divided every 100 μm, and the circumferential direction is divided according to the principle of equal arc length; for cylindrical microstructures, they are first unfolded along the circumferential direction, and then divided radially every 100 μm, and the circumferential direction is divided according to the principle of equal arc length.
4. The method for predicting the surface roughness of ultra-precision turning by integrating vibration and process parameters according to claim 3, characterized in that: The label in S22 is the measured surface roughness value.
5. The method for predicting the surface roughness of ultra-precision turning by integrating vibration and process parameters according to claim 4, characterized in that: During the training process, the deep learning prediction model in S3 outputs predicted surface roughness values based on the preprocessed input features, and calculates the loss between the predicted surface roughness values and the labels, continuously iterating to reduce the loss and obtain the optimal deep learning prediction model.
6. A system applied to the ultra-precision turning roughness prediction method integrating vibration and process parameters as described in claim 5, characterized in that: It includes a multi-source data acquisition and feature construction module, a feature spatialization processing module, a surface roughness intelligent prediction module, and a quality assessment and output module.