A prediction method for residual stress on the surface layer of superalloy milling based on multi-source signals

Abstract

The present invention belongs to the technical field of precision milling machining. The present invention discloses a method for predicting the residual stress on the surface layer of superalloy milling based on multi-source signals. The method includes: obtaining the residual stress data in the depth direction of the superalloy surface; constructing a mathematical model for the gradient distribution curve of the residual stress on the surface layer of the superalloy, and fitting to obtain the optimal values of the model parameters; constructing a Gaussian process regression model, using the key features of the multi-source signals as inputs and the optimal values of the model parameters as the theoretical outputs, and training the Gaussian process regression model; inputting the key features of any group of multi-source signals into the trained Gaussian process regression model to obtain the corresponding model parameters. The present invention uses the key features of the multi-source signals as inputs, thereby accurately predicting the gradient distribution of the residual stress on the surface layer of the workpiece, solving the problem of being unable to predict the gradient distribution of the residual stress on the surface layer of superalloy workpieces in real time, and providing the possibility for online monitoring of the residual stress on the surface layer of superalloy workpieces.

Description

Technical Field

[0001] The present invention belongs to the technical field of precision milling machining, and particularly relates to a method for predicting the surface residual stress of nickel-based superalloy milling based on multi-source signals. Background Technique

[0002] With the development of the new generation of aerospace major equipment towards the direction of large-scale and integral structure, the application ratio of high-performance lightweight alloy materials such as titanium alloy and nickel-based superalloy has been significantly improved in the field of aviation manufacturing. Among them, GH4169G nickel-based superalloy has become the core manufacturing material for key components such as thin-walled blades of aero engines due to its excellent mechanical properties, high-temperature resistance, oxidation resistance and corrosion resistance. However, during the milling process of such materials, the machining surface layer will generate residual stress under the action of thermal-mechanical coupling, and the magnitude and distribution state of the residual stress will directly affect the key performance indicators such as the fatigue strength and dimensional accuracy of nickel-based superalloy workpieces, thus threatening the service performance of the equipment. To ensure the performance requirements of aerospace devices, on the premise of maintaining the integrity of nickel-based superalloy workpieces, accurately predicting the gradient distribution of the surface residual stress of nickel-based superalloy workpieces has become an indispensable technical link in the manufacturing process.

[0003] At present, most of the prediction methods based on machine learning prediction models are used to characterize the gradient distribution of the surface residual stress of nickel-based superalloy workpieces. However, the prediction methods based on machine learning prediction models cannot reflect the dynamic changes of the surface residual stress during the milling process of nickel-based superalloy workpieces, so the gradient distribution of the surface residual stress of nickel-based superalloy workpieces cannot be predicted in real time. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for predicting the surface residual stress of nickel-based superalloy milling based on multi-source signals, taking the key features extracted from multi-source signals as inputs, so as to accurately predict the gradient distribution of the surface residual stress of nickel-based superalloy workpieces, providing the possibility for online monitoring of the surface residual stress of nickel-based superalloy workpieces.

[0005] The present invention adopts the following technical solutions:

[0006] A method for predicting the surface residual stress of nickel-based superalloy milling based on multi-source signals, comprising the following steps:

[0007] Obtain the residual stress data along the depth direction of the workpiece surface layer by layer after nickel-based superalloy milling;

[0008] Construct a mathematical model of the gradient distribution curve of the surface residual stress of nickel-based superalloy, input the residual stress data into the mathematical model, and fit to obtain the optimal value of the model parameters of the mathematical model;

[0009] Construct a Gaussian process regression model, use the key features of multi-source signals as the input of the Gaussian process regression model, and use the optimal values of the model parameters as the theoretical output of the Gaussian process regression model. Train the Gaussian process regression model to obtain a trained Gaussian process regression model; the multi-source signals include cutting force signals, cutting vibration signals, cutting noise signals, and cutting temperature signals during the milling process of superalloys.

[0010] Input the key features of any set of multi-source signals into the trained Gaussian process regression model to obtain the model parameters corresponding to the key features of any set of multi-source signals.

[0011] Furthermore, the process of extracting key features includes:

[0012] Extract the time-domain features and frequency-domain features of the multi-source signals respectively;

[0013] Normalize the time-domain features and frequency-domain features to obtain normalized features;

[0014] Reduce the dimensionality of the normalized features to obtain key features.

[0015] Furthermore, the mathematical model is constructed based on the exponentially decaying cosine function and is expressed as:

[0016] ,

[0017] In the formula, is the residual stress at a depth of from the workpiece surface, is the amplitude constant of the mathematical model, is the damping coefficient of the mathematical model, is the damping frequency of the mathematical model, is the phase angle of the mathematical model, and are both constants. Among them, , , and are all model parameters to be optimized.

[0018] Furthermore, inputting the residual stress data into the mathematical model and fitting to obtain the optimal values of the model parameters of the mathematical model includes:

[0019] Input the residual stress data into the mathematical model and initialize the model parameters;

[0020] Initialize the firefly algorithm and perform fitting calculations on the model parameters to obtain the correlation coefficient;

[0021] At the last N iteration of the firefly algorithm, if the change value of the correlation coefficient is less than , the fitting is successful, and the current model parameters are selected as the optimal values of the model parameters; otherwise, the model parameters are updated and the fitting calculation is performed again.

[0022] Furthermore, the mean function adopted by the Gaussian process regression model is as follows:

[0023] ,

[0024] In the formula, is the th group of key feature vectors, is the mean of the th group of key feature vectors, C is a constant.

[0025] Furthermore, the covariance function adopted by the Gaussian process regression model is an isotropic rational quadratic covariance function, as shown in the following formula:

[0026] ,

[0027] In the formula, is the th group of key feature vectors, is the th group of key feature vectors, is the covariance matrix of the th group of key feature vectors and the th group of key feature vectors, is the standard deviation of the key features, is the shape parameter, is the characteristic length scale, is the th group of key feature vectors and the th group of key feature vectors, and

[0028] The beneficial effects of the present invention are as follows: First, the present invention fits the residual stress data by establishing a mathematical model of the residual stress gradient distribution curve on the surface layer of the superalloy, and obtains the optimal values of the model parameters. Then, using the Gaussian process regression model, the relationship between the key features of the multi-source signals and the optimal values of the model parameters is constructed, thereby establishing the correlation relationship among the key features - the optimal values of the model parameters - the residual stress gradient distribution curve on the workpiece surface layer. The method of the present invention can directly predict the gradient distribution of the residual stress on the surface layer of the superalloy by using the multi-source signals in the milling process of the superalloy, so as to achieve the purpose of on-line monitoring of the residual stress on the surface layer of the superalloy during the processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is the flowchart of the method of the present invention;

[0030] Figure 2 It is the time-domain diagram of the cutting force signal in the present invention;

[0031] Figure 3 It is the frequency-domain diagram of the cutting force signal in the present invention;

[0032] Figure 4 It is the schematic diagram of the main component contribution rate and the cumulative contribution rate of the main components in the present invention;

[0033] Figure 5 It is the schematic diagram of the fitting effect of the mathematical model in the present invention;

[0034] Figure 6 It is the schematic diagram of the model parameter optimization iteration process of the mathematical model in the present invention;

[0035] Figure 7 It is the schematic diagram of the predicted curve of the surface residual stress distribution in the feed direction of the 5th group of data sets in the present invention;

[0036] Figure 8 It is the schematic diagram of the predicted curve of the surface residual stress distribution in the direction perpendicular to the feed direction of the 5th group of data sets in the present invention;

[0037] Figure 9 It is the schematic diagram of the predicted curve of the surface residual stress distribution in the feed direction of the 18th group of data sets in the present invention;

[0038] Figure 10 It is the schematic diagram of the predicted curve of the surface residual stress distribution in the direction perpendicular to the feed direction of the 18th group of data sets in the present invention;

[0039] Figure 11 It is the schematic diagram of the predicted curve of the surface residual stress distribution in the feed direction of the 23rd group of data sets in the present invention;

[0040] Figure 12 It is the schematic diagram of the predicted curve of the surface residual stress distribution in the direction perpendicular to the feed direction of the 23rd group of data sets in the present invention. Detailed implementation manners

[0041] The present invention will be described in detail below with reference to the drawings and embodiments.

[0042] Currently, most use machine learning prediction models to characterize the gradient distribution of the surface residual stress of superalloys. However, the input parameters of the machine learning prediction model are mostly static process parameters, which cannot reflect the dynamic changes of the surface residual stress during the milling process of superalloys. Therefore, it is impossible to predict the gradient distribution of the surface residual stress of superalloys in real time. The distribution of the residual stress generated during the cutting process is the main reason for the deformation and scrapping of thin-walled parts after machining exceeding the tolerance. Therefore, to meet the needs of intelligent manufacturing, it is urgent to study the online prediction of the surface residual stress distribution.

[0043] The present invention provides a method for predicting the residual stress on the surface layer of a superalloy during milling based on multi-source signals, as Figure 1 shown, which includes the following steps:

[0044] S110, layer by layer, obtain the residual stress data along the depth direction of the workpiece surface after the superalloy milling process.

[0045] First, design a full-factorial experiment with 3 factors and 3 levels. Using a milling process experiment plan with 27 groups of different milling parameters, mill the workpiece, which is a superalloy workpiece. Collect the multi-source signals monitored during the milling process to obtain 27 groups of multi-source signals. The multi-source signals include cutting force signals, cutting vibration signals, cutting noise signals, and cutting temperature signals during the superalloy milling process.

[0046] Specifically:

[0047] The cutting force signal is collected by a Kistler 9257B type dynamometer table, the sampling frequency is set to 10000Hz, and the cutting force signal is amplified by a Kistler 5080 multi-channel charge amplifier to form cutting force data in three directions along the x-axis, y-axis, and z-axis; the cutting vibration signal is collected by an acceleration sensor, and the sampling frequency is set to 25600Hz; the cutting noise signal is collected by an acoustic test microphone sensor, and the sampling frequency is set to 25600Hz; the cutting temperature signal is collected by an infrared thermal imager, and the temperature data is recorded every 0.2 seconds. Among them, the cutting force signal, cutting vibration signal, cutting noise signal, and cutting temperature signal are all generated jointly by the workpiece and the tool during the processing. Table 1 shows the processing parameter levels.

[0048] Table 1 Processing parameter levels

[0049]

[0050] Secondly, respectively extract the time-domain features and frequency-domain features of the 27 groups of multi-source signals. Normalize the time-domain features and frequency-domain features of the 27 groups of multi-source signals, and then perform dimensionality reduction processing. Each group obtains several key features, and 10 key features are selected from the several key features of each group to form a key feature set, and a total of 27 groups of key feature sets are obtained. Among them, Figure 2 is the time-domain diagram of the cutting force signal. In the figure, the abscissa is time, and the ordinate is the cutting force signal, Figure 3 is the frequency-domain diagram of the cutting force signal. In the figure, the abscissa is frequency, and the ordinate is energy.

[0051] Specifically:

[0052] Unify the characterization of multi-source signals collected by various sensors during the milling process of the workpiece, and then perform preprocessing and feature extraction on the multi-source signals.

[0053] Filter the multi-source signals. Use low-pass, high-pass, or band-pass filters to filter the cutting force signal, cutting vibration signal, and cutting noise signal to improve the signal stability; on this basis, perform noise reduction and detrending processing through the wavelet transform algorithm.

[0054] Extract features from the multi-source signals after noise reduction and detrending processing. Through time-domain processing and frequency-domain processing, time-domain features and frequency-domain features are obtained respectively. The time-domain features include 13 time-domain features such as maximum value, minimum value, root mean square value, average value, variance, peak-to-peak value, absolute average value, skewness value, kurtosis value, margin factor, waveform factor, pulse factor, and peak factor; the frequency-domain features include 5 frequency-domain features such as center frequency, average frequency, root mean square frequency, frequency variance, and band energy.

[0055] Normalize the extracted time-domain features and frequency-domain features to obtain normalized features. The present invention uses the maximum-minimum normalization method. The purpose of normalization is to convert data with different dimensions to the same dimension, so that the signal features can be compared and calculated on the same scale, in order to improve the accuracy and stability of the prediction results.

[0056] Use the principal component analysis method for data dimensionality reduction to determine the first 10 key features. Excessive features will cause redundancy in feature data, resulting in a decrease in the accuracy of the mathematical model and unable to meet the prediction requirements. Since there is no significant correlation between these redundant features and the surface residual stress, they need to be removed. The present invention uses the principal component analysis method to reduce the dimension of the normalized features and construct new low-dimensional features. The purpose of principal component analysis is to transform the data from the original feature space (high-dimensional space) to a new feature space (low-dimensional space). The new feature space is composed of linear combinations of the original features. Its advantage is that it can automatically assign weights to the extracted feature set by analyzing the correlation between features, so as to obtain key features, reduce the feature dimension, enhance the generalization ability of the mathematical model, simplify the complexity of the mathematical model, and be able to ensure the value of the original information, more comprehensively represent the internal information of the data, thereby improving the prediction accuracy and prediction efficiency of the mathematical model. According to the standard of principal component analysis, when the cumulative contribution rate of each principal component exceeds 90%, these principal components can fully represent all the information of the original features. After calculation, it is found that when accumulating to the 10th principal component, the cumulative contribution rate exceeds 90%. Therefore, the first 10 principal components are determined as the key feature set for surface residual stress prediction. The principal component contribution rate and principal component cumulative contribution rate in the present invention are as Figure 4 shown.

[0057] Finally, the X-ray diffraction method, combined with the method of removing materials layer by layer through electrolytic polishing, was used to measure the milling surface of the workpiece after milling and its residual stress along the surface depth direction. A portable X-ray stress measuring instrument (DS-21P) was used to test the residual stress on the milling surface of the workpiece. Since the X-ray diffraction method can only measure the surface stress of the workpiece, an electrolytic polishing instrument was needed to gradually corrode the surface of the workpiece to measure the residual stress data along the depth direction of the machined surface of the workpiece, and 27 groups of residual stress data sets were obtained. The residual stress includes the residual stress in the feed direction and the residual stress perpendicular to the feed direction.

[0058] Specifically:

[0059] The workpiece was corroded and polished along the depth direction of the machined surface of the workpiece. The thickness of each polishing was 5 - 20 μm. After each polishing, the residual stress was measured using the X-ray diffraction method until the residual stress attenuated and tended to be stable before stopping the measurement.

[0060] S120, construct a mathematical model for the residual stress gradient distribution curve of the superalloy surface layer, input the residual stress data into the mathematical model, and fit to obtain the optimal values of the model parameters of the mathematical model.

[0061] The 27 groups of residual stress data sets were sequentially input into the mathematical model for fitting to obtain 27 groups of optimal values of the model parameters.

[0062] The fitting process is as follows: Input the residual stress data into the mathematical model and initialize the model parameters; Initialize the firefly algorithm and perform fitting calculations on the model parameters to obtain the correlation coefficient; In the last N iteration of the firefly algorithm, if the change value of the correlation coefficient is less than , then the fitting is successful, and the current model parameters are selected as the optimal values of the model parameters; otherwise, update the model parameters and re-perform the fitting calculation. Among them, N is the number of the last several iterations of the firefly algorithm. In the present invention, N takes 50, is the threshold value of the change value of the correlation coefficient. In the present invention, takes 0.01.

[0063] Specifically:

[0064] By observing and analyzing the residual stress data measured in each group of experiments, it can be seen that the gradient distribution of the residual stress on the workpiece surface layer shows a trend of first decreasing, then increasing, and finally approaching the matrix stress. That is, as the depth of the workpiece processing surface continuously increases, the residual stress first gradually increases, then gradually decreases, and finally approaches the matrix stress. Therefore, the gradient distribution curve of the residual stress on the workpiece surface layer is overall spoon-shaped. Therefore, the distribution characteristics can be described by establishing a mathematical model of the gradient distribution curve of the residual stress on the superalloy surface layer. When fitting different residual stress gradient distribution curves using traditional polynomial fitting, the polynomial order needs to be adjusted, and there are too many undetermined coefficients and they are not fixed. However, the exponential decay cosine function has fixed undetermined coefficients. Therefore, it is more suitable as a mathematical model for describing the gradient distribution curve of the residual stress on the superalloy surface layer. The mathematical model of the gradient distribution curve of the residual stress on the superalloy surface layer is shown as follows:

[0065] ,

[0066] In the formula, is the residual stress at a depth of from the workpiece surface, is the amplitude constant of the mathematical model, is the damping coefficient of the mathematical model, is the damping frequency of the mathematical model, is the phase angle of the mathematical model, and are both constants, = 1000, = 100, where, , , and are all model parameters to be optimized.

[0067] Based on the obtained residual stress data, the mathematical model of the gradient distribution curve of the residual stress on the superalloy surface layer is used to fit the residual stress data. The fitting process is regarded as an optimization process of the model parameters, so that the fitted curve is closest to the gradient distribution of the residual stress on the workpiece surface layer after milling. At the same time, the firefly algorithm is selected to optimize the parameters of the mathematical model, and the correlation coefficient is used as the optimization objective function to obtain the optimal values of the model parameters.

[0068] Among them, for the model parameters to be optimized, their optimization space should cover all residual stress data. Therefore, by analyzing the measured residual stress data and combining the characteristics of the exponential decay cosine function, it is determined that has a value range of (0, 10], has a value range of [0, 1], has a value range of (0, 10], The value range of

[0069] is [0, 2π], and the correlation coefficient can be expressed as the following formula:

[0070] In the formula, R 2 is the correlation coefficient, is the fitting value of the exponential decay cosine function, is the measured value of the residual stress, is the average value of the measured values of the residual stress.

[0071] When using the firefly algorithm to solve the optimal values of the model parameters, the parameter settings of the firefly algorithm are as follows: the population size is set to 300, the number of iterations is set to 200, the light intensity absorption factor is set to 1, the maximum attraction constant is set to 1, and the position update constant is set to 0.2.

[0072] Figure 5 is the schematic diagram of the fitting effect of the mathematical model in the present invention, Figure 6 is the schematic diagram of the model parameter optimization iteration process of the mathematical model in the present invention. It can be seen that the firefly algorithm has converged after about 50 iterations. Finally, the correlation coefficient is 0.986. The correlation coefficient in the feed direction is distributed between 88.3% and 99.8%, and the correlation coefficient perpendicular to the feed direction is distributed between 86.0% and 99.5%, indicating that the mathematical model based on the exponential decay cosine function has high accuracy and can effectively characterize the surface residual stress gradient distribution of the workpiece after milling.

[0073] S130. Construct a Gaussian process regression model. Use the key features of the multi-source signal as the input of the Gaussian process regression model, and the optimal values of the model parameters as the theoretical output of the Gaussian process regression model. Train the Gaussian process regression model to obtain the trained Gaussian process regression model; the multi-source signal includes the cutting force signal, cutting vibration signal, cutting noise signal, and cutting temperature signal during the milling process of superalloy.

[0074] Among them, for the regression problem, Gaussian noise needs to be considered, is the standard deviation of the Gaussian noise signal.

[0075] The mean function adopted by the Gaussian process regression model is as follows:

[0076] where

[0077] In the formula, is the th group of key feature vectors, is the mean of the th group of key feature vectors,C is a constant.

[0078] The covariance function adopted by the Gaussian process regression model is the isotropic rational quadratic covariance function, as shown in the following formula:

[0079] ,

[0080] In the formula, is the th group of key feature vectors, is the th group of key feature vectors, is the covariance matrix of the th group of key feature vectors and the th group of key feature vectors, is the standard deviation of the key features, is the shape parameter, is the characteristic length scale, is the th group of key feature vectors and the th group of key feature vectors, and the Euclidean distance therebetween.

[0081] The performance of the Gaussian process regression model is affected by the hyperparameters . Initialize the hyperparameters of the Gaussian process regression model, where c is initialized to 1, is set to 1, has a value range of (0, 2], is the standard deviation of the Gaussian noise signal, has a value range of [0.01, 10], is the standard deviation of the key features, has a value range of [0.01, 10]. Then, calculate the hyperparameters using the maximum likelihood estimation method. Finally, obtain the optimal hyperparameters using the gradient descent algorithm.

[0082] Specifically:

[0083] According to the same milling parameters, correspond the 27 groups of key feature sets and the optimal values of the 27 groups of model parameters to obtain 27 groups of data sets. Each data set includes the key feature set and the optimal value of the model parameters under the same milling parameters.

[0084] Randomly select the 5th data set, the 18th data set, and the 23rd data set in the 27 groups of data sets as the validation set, and the remaining 24 groups of data sets as the training set. Use the key feature set of each data set in the training set as the input and the optimal value of the model parameters as the output to train the Gaussian process regression model and establish the mapping relationship between the key features and the optimal values of the model parameters.

[0085] Taking the key feature set of each dataset in the validation set as the input of the trained Gaussian process regression model, the model parameters can be predicted. Inputting the model parameters into the mathematical model, the surface residual stress distribution curves predicted for the 5th dataset in the validation set and the experimental measurement results are as Figure 7 and Figure 8 shown. The surface residual stress distribution curves predicted for the 18th dataset in the validation set and the experimental measurement results are as Figure 9 and Figure 10 shown. The surface residual stress distribution curves predicted for the 23rd dataset in the validation set and the experimental measurement results are as Figure 11 and Figure 12 shown. It can be seen from the figure that the prediction method of the present invention can accurately predict the surface residual stress distribution of superalloys.

[0086] As shown in Table 2, in the table, No. 5, No. 18 and No. 23 are the validation sets. Among them, No. 5 is the 5th dataset, No. 18 is the 18th dataset, and No. 23 is the 23rd dataset. σ represents the residual stress, σ x represents the surface residual stress in the feed direction, σ y represents the surface residual stress perpendicular to the feed direction, R 2 represents the correlation coefficient. The highest correlation coefficient can reach 97.62%, the lowest correlation coefficient reaches 91.16%, and the average value is 94.39%.

[0087] Table 2 Correlation Coefficients for Predicting the Surface Residual Stress Gradient Distribution of Milled Superalloys

[0088]

[0089] The isotropic rational quadratic covariance function adopted in the present invention introduces a shape parameter, and the shape parameter can regulate the multi-scale characteristics of the covariance function. Essentially, the isotropic rational quadratic covariance function adopted in the present invention can be regarded as the superposition of multiple squared exponential covariance functions with different scales. When , the isotropic rational quadratic covariance function can degenerate into a squared exponential covariance function. Generally, the value range is (0, 2]. The isotropic rational quadratic covariance function adopted in the present invention can effectively balance the global and local features, enhance the anti-interference ability of local fluctuations, and is suitable for data modeling with complex dynamic changes.

[0090] To achieve the prediction of the residual stress gradient distribution on the surface layer of superalloy during milling, the present invention establishes the mapping relationship between key features and model parameters based on the Gaussian process regression model, thereby establishing the correlation relationship among the key features - the optimal values of model parameters - the residual stress gradient distribution curve on the workpiece surface layer. The Gaussian process regression model is a generalization of the multivariate Gaussian distribution. Since the Gaussian process regression model is for quantitative modeling, its prediction accuracy is not affected by Gaussian noise. Therefore, the Gaussian process regression model has better prediction performance than other traditional machine learning models.

[0091] S140, Input the key features of any set of multi-source signals into the trained Gaussian process regression model to obtain the model parameters corresponding to the key features of any set of multi-source signals.

[0092] Input the model parameters corresponding to the key features of any set of multi-source signals into the mathematical model to obtain the residual stress gradient distribution curve on the workpiece surface layer corresponding to the key features of any set of multi-source signals.

[0093] In summary, the prediction method for the residual stress gradient distribution on the surface layer of superalloy based on multi-source signals proposed by the present invention first fits the residual stress data by establishing a mathematical model of the residual stress gradient distribution curve on the surface layer of superalloy to obtain the optimal values of model parameters, and then uses the Gaussian process regression model to construct the relationship between the key features of multi-source signals and the optimal values of model parameters, thereby establishing the correlation relationship among the key features - the optimal values of model parameters - the residual stress gradient distribution curve on the workpiece surface layer, providing a basis for online monitoring of the residual stress gradient distribution on the workpiece surface layer. Verified by experiments, the prediction accuracy rate of the present invention is above 0.9, and the highest reaches 0.976, which can meet the monitoring requirements for the residual stress on the machined surface layer of superalloy workpieces during production and manufacturing. Compared with the prior art, the present invention improves the prediction accuracy of the residual stress on the machined surface layer of workpieces; the average running time of the prediction process is 5 seconds, meeting the requirements of online monitoring, greatly reducing the prediction time of traditional residual stress prediction methods (such as the finite element method), and improving the efficiency. The distribution of the residual stress on the machined surface layer of workpieces is the main reason for the out-of-tolerance deformation of workpieces after machining. Since the present invention can predict the distribution of the residual stress on the workpiece surface layer during the machining process, when the residual stress on the surface layer exceeds the expected range, the process parameters can be changed in time to adjust the distribution of the residual stress, avoiding the scrapping of workpieces caused by the out-of-tolerance deformation of workpieces after machining, saving the machining cost, providing a solution for using online data to monitor the residual stress on the surface layer of superalloy, making it possible to online monitor the residual stress on the surface layer of superalloy, and meeting the requirements of intelligent manufacturing.

Claims

1. A prediction method for the residual stress on the surface layer of superalloy milling based on multi-source signals, characterized in that, Including the following steps: Obtain the residual stress data along the depth direction of the workpiece surface layer by layer after the milling process of superalloy; Construct a mathematical model for the residual stress gradient distribution curve of the superalloy surface layer, input the residual stress data into the mathematical model, and fit to obtain the optimal values of the model parameters of the mathematical model; Construct a Gaussian process regression model, use the key features of multi-source signals as the input of the Gaussian process regression model, and the optimal values of the model parameters as the theoretical output of the Gaussian process regression model, and train the Gaussian process regression model to obtain a trained Gaussian process regression model; the multi-source signals include cutting force signals, cutting vibration signals, cutting noise signals, and cutting temperature signals during the milling process of superalloy; Input the key features of any group of multi-source signals into the trained Gaussian process regression model to obtain the model parameters corresponding to the key features of any group of multi-source signals.

2. The method for predicting the surface residual stress of superalloy milling based on multi-source signals according to claim 1, wherein The extraction process of the key features includes: Extract the time-domain features and frequency-domain features of the multi-source signals respectively; Perform normalization processing on the time-domain features and frequency-domain features to obtain normalized features; Reduce the dimension of the normalized features to obtain key features.

3. A method for predicting the surface residual stress of superalloy milling based on multi-source signals according to claim 2, characterized in that, The mathematical model is constructed based on the exponential decay cosine function and is expressed as: ， In the formula, is the residual stress at a depth of from the surface of the workpiece, is the amplitude constant of the mathematical model, is the damping coefficient of the mathematical model, is the damping frequency of the mathematical model, is the phase angle of the mathematical model, and are both constants. Among them, , , and are all model parameters to be optimized.

4. A method for predicting the surface residual stress of superalloy milling based on multi-source signals according to claim 3, characterized in that, Inputting the residual stress data into the mathematical model and fitting to obtain the optimal values of the model parameters of the mathematical model includes: Input the residual stress data into the mathematical model and initialize the model parameters; Initialize the firefly algorithm and perform fitting calculation on the model parameters to obtain the correlation coefficient; At the end of the firefly algorithm N in the last iteration, if the change value of the correlation coefficient is less than , the fitting is successful, and the current model parameters are selected as the optimal values of the model parameters; otherwise, the model parameters are updated and the fitting calculation is performed again.

5. A method for predicting the surface residual stress of superalloy milling based on multi-source signals according to claim 4, characterized in that The mean function adopted by the Gaussian process regression model is as follows: ， In the formula, is the group of key feature vectors, is the mean value of the group of key feature vectors, C is a constant.

6. A method for predicting the residual stress on the surface layer of a superalloy during milling based on multi-source signals according to claim 5, characterized in that, The covariance function adopted by the Gaussian process regression model is an isotropic rational quadratic covariance function, as shown in the following formula: ， Wherein, is the group of key feature vectors, is the group of key feature vectors, is the covariance matrix of the group of key feature vectors and the group of key feature vectors, is the standard deviation of the key features, is the shape parameter, is the feature length scale, is the Euclidean distance between the group of key feature vectors and the group of key feature vectors.

Images