A method, device and storage device for predicting surface roughness of a workpiece

Abstract

The application discloses a kind of processing workpiece surface roughness prediction method, equipment and storage equipment, method includes the following steps: selecting processing parameter and trying cutting, obtains cutting force;Obtain simulated cutting force;Processing experiment is carried out, and the process data of processing and the surface roughness numerical value of workpiece after processing are obtained;Install tool to obtain the chatter frequency range;According to process data, obtain traditional eigenvalue and chatter eigenvalue;Using AFM measures processing tool key parameter and calculates out theoretical roughness;Manufacture processing characteristic dataset;Processing characteristic dataset is handled, and processing preferred dataset is formed;With ResNet neural network model as framework, utilize preferred dataset to train and obtain surface roughness prediction model, and complete prediction.The application has beneficial effect: it can be online real-time state detection and processing quality prediction to numerical control processing, effectively improve production efficiency.

Description

Technical Field

[0001] This invention relates to the field of precision instrument processing, and in particular to a method, equipment, and storage device for predicting the surface roughness of processed workpieces. Background Technology

[0002] In the field of ultra-precision turning, surface roughness is a crucial aspect of part machining quality, and surface finish directly determines a part's performance in service. For precision parts, excessive surface roughness can lead to scrapping. In actual production, parts often require different surface accuracies depending on their intended use. Because machining is a complex, non-linear process, theoretical calculations are insufficient to obtain optimal process parameters; therefore, optimization based on experience is necessary.

[0003] Currently, the main approach is to inspect finished parts to feed back actual machining errors to the preparation stage for adjusting machining parameters. However, this method is inefficient and its effectiveness is limited. With the rise of machine learning, neural network prediction methods have been introduced into the traditional machining industry. Leveraging their powerful data fitting capabilities, these methods can predict quality before machining begins, improving production efficiency and yield rates to some extent. However, current surface roughness prediction models primarily consider machining parameters and cutting forces. Since cutting forces are often difficult to collect in actual machining, this limits the applicability of these prediction methods. Furthermore, in ultra-precision machining, due to microscale effects, relying on cutting forces and machining parameters to predict roughness fails to improve prediction accuracy and also hinders the improvement of production efficiency and part yield rates. Summary of the Invention

[0004] To further improve the accuracy of surface roughness prediction for precision instrument parts, this application provides a method, apparatus, and storage device for predicting the surface roughness of machined workpieces, wherein the method includes the following steps:

[0005] S1: Predetermine multiple sets of machining parameters for the cutting experiment of the part to be machined, install a force measuring instrument on the lathe, select machining parameters for trial cutting, and obtain the actual cutting force during machining; input the machining parameters into simulation software to obtain the simulated cutting force;

[0006] S2: Install a vibration sensor on the lathe and conduct a machining experiment with the machining parameters described in step S1 to obtain machining process data and the surface roughness value of the part after machining.

[0007] S3: Install the cutting tool on the lathe, obtain the modal frequency of the tool tip at the actual installation overhang, and determine the chatter frequency range;

[0008] S4: Using the processing data described in step S2 and the chatter frequency range described in step S3 as input, calculate the chatter feature value using the FI_VMD method; perform feature extraction on the processing data to obtain the traditional feature value of all processing data.

[0009] S5: Measure the cutting edge radius of the tool described in the experiment using an atomic force microscope, measure the tool tip radius using a metallographic microscope, and calculate the theoretical roughness of the machined part based on the machining parameters described in step S1.

[0010] S6: Using the machining parameters and simulated cutting force described in step S1, the traditional feature values ​​of the process data described in step S4, and the theoretical roughness described in step S5 as feature vectors; using the surface roughness value of the part after machining described in step S2 as a label, a machining feature dataset is created; the feature vectors in the machining feature dataset are normalized and feature selection is performed to obtain the feature with the highest correlation to the label, which is then used as the preferred feature to form a machining preferred dataset;

[0011] S7: Construct a surface roughness prediction model, train the surface roughness prediction model using the processing optimization dataset, and obtain the prediction model after training;

[0012] S8: Input new processing data and use the trained prediction model to obtain the surface roughness of the part to be processed.

[0013] A storage device that stores instructions and data for implementing the aforementioned method for predicting the surface roughness of a machined workpiece.

[0014] A workpiece surface roughness prediction device includes: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the workpiece surface roughness prediction method.

[0015] The beneficial effects provided by this invention are: the model constructed by this invention can perform online real-time status detection and machining quality prediction for CNC machining, effectively improving production efficiency. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0017] Figure 2 (a) and 2(b) are simulation diagrams of the ABAQUS turning process and comparison diagrams of simulated force and actual force in an embodiment of the present invention;

[0018] Figure 3 This is a flowchart of the data filtering process in an embodiment of the present invention;

[0019] Figure 4 This is a waveform diagram of vibration acceleration data collected by an experimental group in an embodiment of the present invention;

[0020] Figure 5 for Figure 4 The waveform of the vibration acceleration data after filtering;

[0021] Figure 6 (a) and (b) are respectively the iterative graph of the accuracy of the prediction model and the comparison graph of the predicted value and the true value of the prediction model obtained by training the preferred dataset in the embodiment of the present invention.

[0022] Figure 7 This is a schematic diagram of the hardware device of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0024] Please refer to Figure 1 , Figure 1 This is a schematic diagram of the method flow of the present invention; a method, equipment, and storage device for predicting the surface roughness of a machined workpiece, wherein the method includes the following steps:

[0025] S1: Predetermine multiple sets of machining parameters for the cutting experiment of the part to be machined, install a force measuring instrument on the lathe, select machining parameters for trial cutting, and obtain the actual cutting force during machining; input the machining parameters into simulation software to obtain the simulated cutting force;

[0026] It should be noted that in this embodiment of the invention, ultra-precision turning is the object and the machining feature of the part is a boss; in addition, in some other embodiments, the machining feature of the part can be a plane, a curved surface, etc.; this is not intended to limit it.

[0027] It should be noted that the machining parameters include spindle speed, depth of cut, and feed rate. The parameters of the material being cut are measured, the cutting force is obtained using the simulation software ABAQUS, and the simulation results are calibrated based on trial cutting experiments.

[0028] As one embodiment, under the condition that the fixture, tooling, machine tool rigidity, workpiece material, and cooling conditions remain constant, the machining parameters for turning operations on the ultra-precision milling and turning machine tool include spindle speed (400–3000 r / min), feed rate (0.001–0.009 mm / min), and depth of cut (0.001–0.006 mm). In this embodiment, using spindle speed, feed rate, and depth of cut as three factors, a three-factor, six-level data acquisition experiment was designed according to orthogonal experimental design, with a total of 7 groups, each group machining 36 bosses, for a total of 252 machining operations. One set of parameters was randomly selected, and machining began with a force gauge installed, obtaining 36 force data points during the machining process.

[0029] In this embodiment, the material being processed is Ti-6Al-4V titanium alloy. The material parameters for this embodiment are shown in Table 1, obtained through actual measurement and literature review.

[0030] Table 1 Material Parameters

[0031]

[0032] Since the cutting width is usually much larger than the cutting thickness during the cutting process, the entire cutting process can be simplified into a two-dimensional plane strain problem. This reduces the computational load and saves computational resources and time while ensuring calculation accuracy. An orthogonal cutting finite element model of Ti-6Al-4V was established using the finite element software Abaqus / Explicit, as follows: Figure 2 As shown. The entire model is divided into four parts: workpiece base, separation zone, cut layer (chips), and tool. The workpiece uses four-node thermo-coupled plane strain quadrilateral elements (CPE4RT). The mesh of the cut layer is refined, and the mesh direction is set to an angle of 45° with the X-axis to avoid mesh distortion when the tool contacts the chip layer. The bottom and left side of the workpiece are completely fixed. The tool is set as a rigid body with a rake angle of 0° and a clearance angle of 7°. During the simulation, the tool moves along the negative X-axis at a constant cutting speed V. The cutting speed is:

[0033]

[0034] Where N is the rotational speed and R is the machining radius. The friction between the tool surface and the chip follows Coulomb's law of friction, and the coefficient of friction μ is fixed at 0.1.

[0035] In this embodiment, the simulation uses the Johnson-Cook constitutive model, which is specifically in the form of:

[0036]

[0037] In the formula, σ is the flow stress, and ε is the equivalent plastic strain. For equivalent plastic strain rate, For reference, the equivalent plastic strain rate is usually taken as 1 or other experimentally determined value, where T is the temperature. m T is the melting point of the material. r Let T be the reference temperature. A, B, n, C, and m represent the initial yield strength, hardening modulus, stress hardening exponent, strain rate hardening coefficient, and thermal softening coefficient, respectively. Clearly, when T > T0... m When the temperature-dependent term is less than 0, the model fails. The Johnson-Cook damage model is specifically in the form of...

[0038]

[0039]

[0040] In the formula σ p σ represents hydrostatic pressure. m Indicates Mises stress; D i (i = 1, 2, 3, 4, 5) are five JC damage parameters; ε f The equivalent strain is the failure strain; Δε is the equivalent plastic strain increment. When w = 1, the material begins to fail.

[0041] In this embodiment, Figure 2 This is a comparison chart of the average cutting force obtained from actual machining and the cutting force obtained from simulation calculation. The simulation error is within 10%, which has a high confidence level. The simulated force can be used to replace the actual cutting force for surface roughness prediction.

[0042] S2: Install a vibration sensor on the lathe and conduct a machining experiment with the machining parameters described in step S1 to obtain machining process data and the surface roughness value of the part after machining.

[0043] It should be noted that the machining process data includes workpiece vibration data and spindle current data.

[0044] Before the processing data is used for calculations later, it needs to be preprocessed. The preprocessing methods in this application are, in order: spectral analysis and adaptive filtering.

[0045] The spectrum analysis involves performing a discrete Fourier transform on the processing data to obtain the amplitude spectrum of the processing data.

[0046] The adaptive filtering involves obtaining an amplitude threshold based on the amplitude spectrum and applying weighted filtering to each frequency component after the discrete Fourier transform.

[0047] As an embodiment of step S2, the machining process data of ultra-precision turned parts includes tool vibration data and spindle current data; in this embodiment, the tool vibration data is collected by an external sensor, and the spindle current signal is derived from inside the machine tool.

[0048] Vibration data acquisition specifically involves: installing a vibration sensor at the tool holder, setting the sampling rate to 20kHz, and acquiring vibration signals in real time during workpiece machining on the machine tool. Data transmission is achieved via a data acquisition card, a data acquisition chassis, and a host computer, ultimately obtaining real-time vibration acceleration data (A) in the X, Y, and Z axes of the Cartesian coordinate system during machining. x A y A z ).

[0049] The spindle current data acquisition process involves: exporting the data from inside the machine tool using the machine tool control software, filtering the data, and then saving the spindle current data.

[0050] After data acquisition is completed, the acquired processing data undergoes data preprocessing; specifically, data preprocessing includes sequential spectral analysis and adaptive filtering; such as... Figure 3 As shown, spectrum analysis involves performing a discrete Fourier transform on the processing data to obtain the amplitude spectrum of the processing data; adaptive filtering involves obtaining the amplitude threshold based on the amplitude spectrum and applying weighted filtering to each frequency component after the discrete Fourier transform.

[0051] In this embodiment, the discrete Fourier transform is performed on the collected tool vibration data and spindle current data in the spectrum analysis section. The formula is as follows:

[0052]

[0053] Where X(n) represents the acquired data, n is the step length, and N is the total length of the acquired data; X(k) represents the frequency components obtained after the Discrete Fourier Transform, and k is the sampling frequency of the Discrete Fourier Transform (k = 0, 1, 2, ..., N-1); the amplitude of each frequency component in the signal is calculated by the following formula:

[0054]

[0055]

[0056]

[0057] Where A(k) is the amplitude of each frequency component; with frequency as the independent variable and the amplitude corresponding to that frequency component as the dependent variable, the resulting frequency function is the amplitude spectrum.

[0058] In the adaptive filtering section, the amplitudes of all frequency components in the obtained amplitude spectrum are sorted, and the median is taken as the amplitude threshold A0. The amplitude of each frequency component X(k) is compared with the amplitude threshold A0. When A(k) > A0, X(k) = X(k); when 0.5 < A(k) < A0, ... In other cases, X(k) = 0.

[0059] The frequency components are recombine, and then the frequency domain signal is converted into a time domain signal by inverse Fourier transform. The formula for inverse Fourier transform is:

[0060]

[0061] Taking vibration data as an example, the collected vibration data is as follows: Figure 4 As shown in the figure, from top to bottom, they represent the vibration signals along the X, Y, and Z axes of a Cartesian coordinate system. The filtered vibration data is as follows. Figure 5 As shown, the Y-axis noise data is significantly suppressed, and in this embodiment, the cutting force in the Y-axis direction is almost zero. The Y-axis vibration is mainly caused by installation error and measurement device error. This filtering method can effectively suppress the introduction of such errors.

[0062] After filtering, feature values ​​of the processing data are calculated through feature extraction. Feature extraction involves calculating time-domain features, frequency-domain features, and wavelet packet features of the processing data. Time-domain features include maximum, minimum, mean, range, root mean square, variance, skewness, kurtosis, kurtosis factor, waveform factor, impulse factor, and margin factor. Frequency-domain features include spectral density, energy spectral density, and power spectral density. Wavelet packet features are the skewness and kurtosis of the eight node coefficients within the 0-20kHz range. A total of 312 features are extracted for vibration acceleration in the X, Y, and Z axes and for the U, V, and W phase currents.

[0063] In this embodiment, the finished product is a multi-level boss, and the actual roughness value of the end face of the machined part is obtained by measuring with a white light interferometer.

[0064] S3: Install the cutting tool on the lathe, obtain the modal frequency of the tool tip at the actual installation overhang, and determine the chatter frequency range;

[0065] As an embodiment of step S3, in this embodiment, the tool is a split diamond turning tool. The diamond turning tool is combined with the tool holder and mounted on the tool post. Then, an acceleration sensor is fixed to the tool holder, and a hammer impact test is used to obtain the input force and vibration acceleration data at the tool tip. In this embodiment, the magnitude of the input force x(t) is measured by a force sensor mounted on the modal hammer, and the amplitude of the vibration acceleration y(t) is measured by a vibration sensor mounted on the tool post. By changing the magnitude of the input force x(t) in the hammer impact test, the corresponding vibration acceleration amplitude y(t) can be obtained. Then, Fourier transforms are performed on them to obtain the frequency domain distributions X(ω) and Y(ω) of the input force and vibration acceleration; the ratio of Y(ω) to X(ω) is the frequency response function of the machining system. That is, the formula for calculating the frequency response function is:

[0066]

[0067] Since the frequency response function is a complex number, it can be rewritten as:

[0068]

[0069] Where A(ω) is the amplitude-frequency characteristic of the machining system. Let j be the phase frequency characteristic of the machining system, j be an imaginary number, and e be a natural constant.

[0070] Plot the amplitude-frequency response of the machining system with frequency on the x-axis and amplitude y(t) on the y-axis. The x-axis at the peak value is the natural frequency of the specific mode of the machining system.

[0071] In this embodiment, it was determined that the vibration around 987Hz was the main contributor, and the frequency range in which flutter occurred was finally selected as 800-1100Hz.

[0072] S4: Using the processing data described in step S2 and the chatter frequency range described in step S3 as input, calculate the chatter feature value using the FI_VMD method; perform feature extraction on the processing data to obtain the traditional feature value of all processing data.

[0073] It should be noted that the feature extraction involves calculating the time-domain features, frequency-domain features, and wavelet packet features of the processing data.

[0074] The time-domain features include maximum value, minimum value, mean, range, root mean square value, variance, skewness, kurtosis, kurtosis factor, waveform factor, impulse factor, and margin factor.

[0075] The frequency domain features include spectral density, energy spectral density, and power spectral density.

[0076] The wavelet packet features are the skewness and kurtosis of the eight node coefficients within the range of 0-2 kHz.

[0077] It should be noted that the flutter feature extraction in step S4 uses the FI_VMD method. FI_VMD is based on variational mode decomposition (VMD), which decomposes a multi-component signal into multiple single-component amplitude-modulated and frequency-modulated signals at one time. This avoids the endpoint effect and spurious component problems encountered in the iteration process, and quickly locates the components in the frequency range of interest, so as to realize rapid flutter detection and quantitative analysis.

[0078] As one embodiment of step S4, the Fast Variational Mode Decomposition (FI-VMD) method comprises the following steps:

[0079] 1) For signal s k (t) Perform a Fourier transform to calculate the maximum frequency f of its frequency spectrum. k ;

[0080] 2) Given {K = 1; α = 3500 - 500n, stα ≥ 500} and initialize w 1 =f k The modal u is calculated and output based on VMD. k (t) and modal center frequency w k ;

[0081] 3) Based on formula s k+1 (t)=s k (t)-u k (t) Update the input signal, when Then proceed to the next iteration;

[0082] Where S 1 (t) = f, where f is the original input signal; [a, b] is the frequency band to which the flutter signal belongs.

[0083] Using energy difference ratio (ERD) as an evaluation metric offers higher sensitivity compared to traditional metrics (ER and RER), and it is less susceptible to noise and external influences, exhibiting strong robustness. It is an important criterion for online flutter detection, and its definition is as follows:

[0084]

[0085]

[0086]

[0087] ERD = RER - ER

[0088] When ERD > Θ, the system is determined to be fluttering.

[0089] In this embodiment, ERD is used as a feature input into the neural network for training, and the magnitude of the threshold Θ is not a concern.

[0090] S5: Measure the cutting edge radius of the tool described in the experiment using an atomic force microscope, measure the tool tip radius using a metallographic microscope, and calculate the theoretical roughness of the machined part based on the machining parameters described in step S1.

[0091] As an embodiment of step S5, a total of 7 cutting tools are used in this embodiment. Their tool parameters are similar. The key parameters of the tools are measured in sequence, including the cutting edge radius, rake angle, clearance angle and tool tip radius. The cutting edge radius refers to the arc of the main cutting edge formed by the intersection of the rake face and the main clearance face. Usually, this value is small, at the nanometer level. Therefore, the surface morphology near the main cutting edge can be measured by atomic force microscopy, and then the cutting edge radius can be calculated by curve fitting.

[0092] In this embodiment, the rake angle, clearance angle, and blade tip radius can be measured using a metallographic microscope. The blade tip radius is approximately 1 mm, and the rake angle and clearance angle are 0° and 7°, respectively.

[0093] After measuring the key tool parameters, and combining them with the machining parameters described in step S1, the theoretical surface roughness (R) can be calculated. a The calculation formula is:

[0094]

[0095] Where H and E are Vickers hardness and Young's modulus, and k2 is calculated according to the following formula:

[0096] when hour

[0097] when hour

[0098] When r ∈ When r > 0.9 mm, k1 = 12, when r ∈ When ≤0.9mm, k1=10; h min Calculate using the following formula:

[0099] h min =cr n

[0100] Where r n Where c is the cutting edge radius, and c is a constant coefficient between 0.3 and 0.4; in this embodiment, c = 0.35; h Dmax Calculate using the following formula:

[0101]

[0102] Where r ∈ where is the tool tip radius, f is the feed rate, and ap It is the cutting depth.

[0103] S6: Using the machining parameters and simulated cutting force described in step S1, the traditional feature values ​​of the process data described in step S4, and the theoretical roughness described in step S5 as feature vectors; using the surface roughness value of the part after machining described in step S2 as a label, a machining feature dataset is created; the feature vectors in the machining feature dataset are normalized and feature selection is performed to obtain the feature with the highest correlation to the label, which is then used as the preferred feature to form a machining preferred dataset;

[0104] It should be noted that the specific method of feature selection in step S6 is to remove redundant features by using a scatter matrix diagram and select features with high relevance to the label by using correlation coefficient analysis to obtain the preferred features, thereby reducing the number of features and ensuring that the filtered features and label data have the highest relevance.

[0105] The correlation coefficient analysis method involves calculating the correlation coefficient between each feature in the stable processing feature dataset and the label using a correlation coefficient calculation formula. Feature X i The formula for calculating the correlation coefficient with label C is:

[0106]

[0107] in, and C and X respectively i The mean of R, where N′ is the sample size, and R0 is the mean of the sample size. xc X represents i The degree of linear correlation between C and C.

[0108] In this embodiment, the feature vector is normalized using (0, 1), and the specific formula is as follows:

[0109]

[0110] Where Max and Min are the maximum and minimum values ​​of the corresponding feature in all feature data, respectively.

[0111] In this embodiment, the threshold for the correlation coefficient is set to 0.7, and the number of preferred features obtained by the final screening is 121, which greatly reduces the complexity of the dataset compared to the 317 features extracted in step S3.

[0112] S7: Construct a surface roughness prediction model, train the surface roughness prediction model using the processing optimization dataset, and obtain the prediction model after training;

[0113] It should be noted that the Residual Neural Network (ResNet) model incorporates residual units into the neural network through a short-circuit mechanism, effectively solving the problem of gradient vanishing or exploding in deep neural networks, and enabling deep network training.

[0114] The specific way to incorporate the error propagation characteristic is as follows: the error propagation characteristic is added to the neural nodes of the output layer, and the weights and biases of each node in the prediction model are determined by backpropagation of the error in the backpropagation neural network.

[0115] The specific way to integrate the convolutional layer is as follows: the convolutional layer is added between the hidden layer and the output layer. Through pooling operations, on the one hand, the features and parameters are reduced, thereby simplifying the computational complexity of the convolutional network, and on the other hand, the invariance of features (rotation, translation, scaling, etc.) is maintained.

[0116] S8: Input new processing data and use the trained prediction model to obtain the surface roughness of the part to be processed.

[0117] ResNet demonstrates outstanding performance in deep learning by utilizing shortcut connections to address the degradation problem in deep networks. Compared to ordinary networks, it adds a shortcut mechanism between every two or three layers, enabling deeper networks to function effectively through residual learning. The ResNet model architecture mimics the 8-layer structure of AlexNet, dividing ResNet into eight building layers. A building layer can contain one or more network layers and one or more building blocks (such as ResNet building blocks). The first building layer consists of one ordinary convolutional layer and a max-pooling layer. The second building layer consists of three residual modules. The third, fourth, and fifth building layers all begin with downsampling residual modules, followed by three, five, and two residual modules respectively.

[0118] Taking the training of a surface roughness prediction model using a stable processing optimization dataset as an example, the stable processing optimization dataset is divided into a training set and a test set. In this embodiment, the training set accounts for 80% of the sample size, and the test set accounts for 20%. In some embodiments, the sample ratio of the training set and the test set can be adjusted according to the number of data samples, for example: the training set accounts for 70% of the sample size, and the test set accounts for 30%; or the training set accounts for 75% of the sample size, and the test set accounts for 25%. In this embodiment, the number of training set samples in the optimization dataset is 252 × 0.80 = 202, and the number of test set samples is 252 × 0.20 = 50.

[0119] The feature data selected from the preferred dataset used for training is input into the ResNet neural network to obtain the final neural network prediction model. The model is then tested using feature data from the test set, and the results are compared with the labeled data to determine the model's accuracy.

[0120] Based on the surface roughness prediction model in this embodiment, the predicted value is compared with the actual measured surface roughness, i.e., the label value, and the percentage error is calculated. i .

[0121] In this embodiment, a prediction model is obtained by training a preferred dataset in the manner described above. For example... Figure 6 As shown in (a), the vertical axis represents the prediction accuracy 1 - error. i The horizontal axis represents the training period; in the test set, the degree of fit between the predicted and actual roughness values ​​is as follows: Figure 6 As shown in (b), it can be seen that the model trained using the preferred dataset has a better accuracy.

[0122] It should be noted that the manufacturing process is a nonlinear and complex process, making it difficult to predict the quality of the process. Further expansion of the dataset and improvement of data quality, as well as optimization of the model algorithm, are needed to break through the current bottleneck in prediction accuracy.

[0123] Please see Figure 7 , Figure 7 This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: a workpiece surface roughness prediction device 401, a processor 402, and a storage device 403.

[0124] A workpiece surface roughness prediction device 401: The workpiece surface roughness prediction device 401 implements the workpiece surface roughness prediction method.

[0125] Processor 402: The processor 402 loads and executes the instructions and data in the storage device 403 to implement the method for predicting the surface roughness of a machined workpiece.

[0126] Storage device 403: The storage device 403 stores instructions and data; the storage device 403 is used to implement the method for predicting the surface roughness of a machined workpiece.

[0127] The beneficial effects of this invention are as follows: The construction method of this invention obtains cutting force through simulation calculation, replacing the actual force, reducing the difficulty of practical application. Furthermore, by adding features extracted by the FI_VMD method and features calculated by improved theoretical roughness to traditional data features, the dimensions of the input feature vector are enriched, correspondingly improving the neural network model and enhancing its learning ability and efficiency under multi-dimensional feature vectors. Ultimately, a neural network model with higher prediction accuracy is obtained through training. The prediction model constructed using this invention can perform online real-time status detection and surface roughness prediction for ultra-precision turning, effectively improving production efficiency and reducing the use of measuring instruments, especially force gauges. Simultaneously, the prediction model constructed by this invention can guide part processing to obtain the desired machined surface. In addition, with the expansion of the machining feature dataset, more machining features sensitive to changes in surface roughness can be extracted, improving the efficiency and accuracy of model prediction. Furthermore, the prediction model constructed by this invention is easy to integrate into CNC machine tool systems, facilitating operator use.

[0128] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the surface roughness of a machined workpiece, characterized in that: Includes the following steps: S1: Predetermine multiple sets of machining parameters for the cutting experiment of the part to be machined, install a force measuring instrument on the lathe, select machining parameters for trial cutting, and obtain the actual cutting force during machining; input the machining parameters into simulation software to obtain the simulated cutting force; S2: Install a vibration sensor on the lathe and conduct a machining experiment with the machining parameters described in step S1 to obtain machining process data and the surface roughness value of the part after machining. S3: Install the cutting tool on the lathe, obtain the modal frequency of the tool tip at the actual installation overhang, and determine the chatter frequency range; S4: Using the processing data described in step S2 and the chatter frequency range described in step S3 as input, calculate the chatter feature value using the FI_VMD method; perform feature extraction on the processing data to obtain the traditional feature value of all processing data. S5: Measure the cutting edge radius of the tool described in the experiment using an atomic force microscope, measure the tool tip radius using a metallographic microscope, and calculate the theoretical roughness of the machined part based on the machining parameters described in step S1. S6: Using the machining parameters and simulated cutting force described in step S1, the traditional feature values ​​of the process data described in step S4, and the theoretical roughness described in step S5 as feature vectors; using the surface roughness value of the part after machining described in step S2 as a label, a machining feature dataset is created; the feature vectors in the machining feature dataset are normalized and feature selection is performed to obtain the feature with the highest correlation to the label, which is then used as the preferred feature to form a machining preferred dataset; S7: Construct a surface roughness prediction model, train the surface roughness prediction model using the processing optimization dataset, and obtain the prediction model after training; In step S7, the surface roughness prediction model uses the ResNet network framework; S8: Input new processing data and use the trained prediction model to obtain the surface roughness of the part to be processed.

2. The method for predicting the surface roughness of a machined workpiece as described in claim 1, characterized in that: The machining parameters mentioned in step S1 include spindle speed, depth of cut, and feed rate.

3. The method for predicting the surface roughness of a machined workpiece as described in claim 1, characterized in that: Before the processing data described in step S4 is used as input parameters, it is preprocessed by spectrum analysis and adaptive filtering in sequence.

4. The method for predicting the surface roughness of a machined workpiece as described in claim 1, characterized in that: The feature extraction in step S4 specifically involves calculating the time-domain features, frequency-domain features, and wavelet packet features of the processing data.

5. The method for predicting the surface roughness of a machined workpiece as described in claim 1, characterized in that: The feature selection in step S6 specifically refers to: removing redundant features through a scatter matrix diagram and selecting features with high relevance to the label through correlation coefficient analysis to obtain the preferred features.

6. A storage device, characterized in that: The storage device stores instructions and data for implementing any of the workpiece surface roughness prediction methods described in claims 1 to 5.

7. A device for predicting the surface roughness of a machined workpiece, characterized in that: include: A processor and a storage device; the processor loads and executes instructions and data in the storage device to implement any one of the workpiece surface roughness prediction methods according to claims 1 to 5.

Images