Typical part working condition adaptability application data generation method

Abstract

The invention provides a typical part working condition adaptability application data generation method, and belongs to the technical field of cutting machining. According to the method, time-varying signals, non-time-varying signals and cutting residual stress values in the test process are collected, the non-time-varying signals are represented in a unified mode, self-adaptive feature extraction is conducted on time-varying signal data, a feature extraction mode of low-dimensional representation of high-dimensional data is designed, then the processed non-time-varying signals and the time-varying signals are fused, and the cutting residual stress values are obtained. According to the method, the residual stress is used as input of a CNN cutting residual stress prediction model, a cutting residual stress value is used as model output, a complex mapping relation between acquired data and final residual stress under a time-varying working condition is established, and online prediction of the cutting residual stress is realized. According to the method, the nonlinear relationship between each working condition factor feature and the cutting residual stress is expressed by using the convolutional neural network structure, and the cutting residual stress can be quickly predicted based on the machining process data.

Description

Technical Field

[0001] The present invention belongs to the technical field of cutting processing, and in particular relates to a method for generating application data of adaptability of typical part working conditions. Background Art

[0002] Accurately predicting residual stress during machining is crucial for ensuring the surface quality of typical thin-walled parts and reducing losses caused by machining deformation. During machining, the magnitude of residual stress, a key application metric for adaptability to the working conditions, can vary due to adjustments to factors such as workpiece material properties, workpiece structural characteristics, actual cutting parameters, tool geometry, and the machine tool's inherent characteristics. These factors are coupled to each other and exhibit a nonlinear relationship with residual stress.

[0003] Currently, there are two main methods for predicting cutting residual stress adaptability data: empirical methods performed after machining, and analytical methods and finite element methods (FEMs) performed before machining. Empirical methods involve repeated attempts to correct geometric errors caused by cutting residual stress after machining. They directly consider machining parameters such as tool parameters, process parameters, and cutting fluid parameters, establish a mapping relationship between machining parameters and cutting residual stress, and use these parameters to predict cutting residual stress. However, this method cannot clearly reveal the impact of machining parameters on residual stress and relies primarily on manual adjustment based on experience. This method is inefficient and cannot be used for online residual stress prediction. Analytical and FEMs simulate the cutting forces and heat during the cutting process using simulation models such as finite element or analytical methods before machining. These methods are also not suitable for online residual stress prediction. Furthermore, these methods are performed under ideal machining conditions, leading to certain deviations between the simulation model and the actual machining process due to uncertainties in the machining process. Furthermore, efficiency is a shortcoming that simulation models cannot address, as each calculation requires a significant amount of time. Summary of the Invention

[0004] The present invention aims to solve the technical problems of existing cutting residual stress prediction methods that cannot predict residual stress online and are inefficient. From a data-driven perspective, a method for generating application data that is adaptive to typical part working conditions is proposed. Specifically, a data-driven cutting residual stress prediction method under time-varying working conditions is proposed. By establishing a correlation between the real data of various coupled working conditions during the machining process and the cutting residual stress, a theoretical basis is provided for subsequent online optimization and control of the cutting residual stress, thereby improving the pass rate of workpieces such as thin-walled parts in one-step machining. Taking into account the real-time nature of the data-driven prediction method, the present invention overcomes the deficiency of existing residual stress prediction methods that cannot consider changes in the machining process. A convolutional neural network (CNN) is used to learn the mapping relationship between the time-varying working condition data and the cutting residual stress during the actual machining process. The machining process data is input into a trained model to quickly generate application data that is adaptive to the working conditions, that is, to achieve online prediction of the cutting residual stress. Then, parameters can be adjusted before the next process to achieve online control of the cutting residual stress.

[0005] To achieve the above objectives, the technical solutions provided by the present invention are:

[0006] A method for generating application data of typical part working condition adaptability is characterized in that it includes the following steps:

[0007] Step 1: Conduct cutting tests and data collection: Collect monitoring data, working condition data, and cutting residual stress values corresponding to the cutting process under time-varying working conditions. The monitoring data includes cutting force signals and vibration signals, and the working condition data includes spindle speed, cutting depth, cutting width, and feed rate.

[0008] Step 2: Unified characterization of working condition data information: The collected working condition data is uniformly characterized;

[0009] Step 3, monitoring data processing, includes:

[0010] Sub-step 3.1, data preprocessing: noise reduction and standardization of the collected monitoring data;

[0011] Sub-step 3.2, extraction and dimensionality reduction of main features: Use deep learning methods to perform adaptive feature extraction on the preprocessed data to obtain a deep coding feature set. Then, the preprocessed data is processed in the time domain, frequency domain, and wavelet packet energy feature extraction to obtain a time domain feature set, a frequency domain feature set, and a wavelet packet energy feature set, respectively. Feature extraction is performed on the feature data of the obtained deep coding feature set and the time domain feature set, frequency domain feature set, and wavelet packet energy feature set, and dimensionality reduction is performed on the extracted feature data.

[0012] Step 4: Fusion of monitoring data and working condition data: The working condition data that has undergone unified characterization processing is fused with the monitoring data that has undergone dimensionality reduction processing to obtain fused feature data. The fused feature data and the corresponding collected cutting residual stress values form a residual stress data set;

[0013] Step 5, constructing a CNN cutting residual stress prediction model: using the residual stress dataset to construct a CNN-based cutting residual stress prediction model, and training and testing the cutting residual stress prediction model to obtain a final cutting residual stress prediction model;

[0014] Step 6, cutting residual stress prediction: Using the final cutting residual stress prediction model, the cutting residual stress prediction value is obtained according to the monitoring data and working condition data during the cutting process under time-varying working conditions.

[0015] Furthermore, in sub-step 3.1, a wavelet threshold decomposition denoising method is used to perform noise reduction processing.

[0016] Furthermore, in sub-step 3.1, noise reduction is performed by using the VisuShrink threshold algorithm and a shrinkage factor and a soft threshold method.

[0017] Furthermore, in sub-step 3.1, the Z-score algorithm is used for standardization.

[0018] Furthermore, in sub-step 3.2, a deep encoding feature set is obtained by inputting the preprocessed data into a CNN-VAE-like variational autoencoder adaptive feature extraction network.

[0019] Furthermore, in sub-step 3.2, the RFR method is used to perform feature screening on the feature data of the depth coding feature set and the time domain feature set, the frequency domain feature set, and the wavelet packet energy feature set.

[0020] Furthermore, in sub-step 3.2, the PCA method is used to perform dimensionality reduction on the selected feature data.

[0021] Furthermore, in step 4, the Concat algorithm is used to fuse and calculate to obtain fused feature data.

[0022] Furthermore, in step 5, the Adam algorithm is used to optimize the training process of the cutting residual stress prediction model.

[0023] The advantages of the present invention are:

[0024] The invention discloses a method for generating adaptive application data for typical part working conditions, which collects time-varying signals such as cutting force signals and vibration signals, non-time-varying signals such as spindle speed and cutting depth, and actual cutting residual stress values during the test processing process, uniformly characterizes and processes the non-time-varying signals, performs adaptive feature extraction on the time-varying signal data, designs a feature extraction method for low-dimensional representation of high-dimensional data, and then fuses the processed non-time-varying signals with the time-varying signals as the input of a CNN cutting residual stress prediction model, and outputs the cutting residual stress values as the model output. This establishes a complex mapping relationship between the collected data under time-varying working conditions and the final residual stress, thereby realizing the generation of adaptive application data for working conditions, i.e., online prediction of cutting residual stress. Therefore, the invention can use the data in the processing process as input, objectively mine the key features within the data in a data-driven manner, and use the convolutional neural network structure to express the nonlinear relationship between the characteristics of various working condition factors and cutting residual stress. The processing process data can be input into the model, and the residual stress prediction results can be quickly output, providing effective guidance for online optimization of cutting parameters, timely adjustment of process plans, and control of residual stress in the actual cutting process, thereby improving production efficiency and processing quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The features and advantages of the present invention will become more readily understood through the following description with reference to the accompanying drawings, which are not drawn to scale and in which some features are exaggerated or minimized to show details of particular components.

[0026] Figure 1 is a flow chart of a method for generating application data of typical part working condition adaptability according to the present invention;

[0027] Figure 2 : is a structural diagram of the CNN-VAE type variational autoencoder adaptive feature extraction network in the method of the present invention;

[0028] Figure 3 4 is a structural diagram of a variational autoencoder feature extraction module included in the CNN-VAE variational autoencoder adaptive feature extraction network in the method of the present invention;

[0029] Figure 4 1 is a structural diagram of a variational autoencoder computing module included in the CNN-VAE variational autoencoder adaptive feature extraction network in the method of the present invention;

[0030] Figure 5 1 is a diagram of the conversion process of time-frequency domain conversion in the method of the present invention, (a) is time domain data, (b) is frequency domain data;

[0031] Figure 6 is a schematic diagram of wavelet decomposition in the method of the present invention;

[0032] Figure 7 Schematic diagram of the top ten features with the greatest importance in the method of the present invention, namely, depth coding features, time domain features, frequency domain features, and wavelet packet energy features, (a) represents the importance of depth coding features, (b) represents the importance of time domain features, (c) represents the importance of frequency domain features, and (d) represents the importance of wavelet packet energy features;

[0033] Figure 8 It is the result of normalized data processing of monitoring data in the example by the method of the present invention;

[0034] Figure 9 : are the feature extraction results of the method of the present invention in the example, (a) is the depth coding feature extraction result, (b) is the time domain feature extraction result, (c) is the frequency domain feature extraction result, and (d) is the wavelet packet energy feature extraction result;

[0035] Figure 10 It is a feature set obtained by subjecting the feature data of the depth coding feature set, the time domain feature set, the frequency domain feature set and the wavelet packet energy feature set in the example of the method of the present invention to feature extraction and dimensionality reduction processing;

[0036] Figure 11 is a comparison chart of the cutting residual stress prediction value and the true value of the test set in the example of the method of the present invention;

[0037] Figure 12 It is the loss function change curve during the model training process in the example of the method of the present invention. DETAILED DESCRIPTION

[0038] The present invention will be described in detail below with reference to the accompanying drawings by means of exemplary embodiments of the present invention. It should be noted that the following detailed description of the present invention is only for the purpose of illustration and is not intended to limit the present invention.

[0039] The present invention provides a method for generating application data of working condition adaptability of typical parts, specifically a method for predicting cutting residual stress under time-varying working conditions based on data driven by the cutting process under time-varying working conditions. The method is used to generate application data of working condition adaptability of typical parts such as thin-walled parts based on data from the cutting process under time-varying working conditions, that is, to predict cutting residual stress online, so as to optimize cutting parameters online and control residual stress.

[0040] During the cutting process, cutting residual stress is affected by many aspects. The residual stress prediction problem has different characteristics in terms of the spatiotemporal correlation between cutting residual stress and working condition data, the relationship between working condition data and residual stress, the diversity and coupling of working condition influencing factors, etc., which brings difficulties to residual stress prediction.

[0041] By monitoring the time-varying working condition data during processing and revealing the correlation between it and the cutting residual stress, the residual stress can be predicted and controlled in real time, thereby providing a more intuitive understanding of the prediction of residual stress to a certain extent. When predicting the cutting residual stress, the present invention uses both non-time-varying working condition factors (working condition signals) and time-varying working condition factors (monitoring signals) to predict the cutting residual stress. This method has several assumptions:

[0042] (1) In the present study, the tool hardness is greater than the workpiece hardness, and the tool is replaced in each working condition, ignoring the effect of tool wear on surface residual stress;

[0043] (2) The research material of the present invention is titanium alloy, a typical aviation material. The initial residual stress of the matrix after heat treatment can be ignored, and only the residual stress introduced during the cutting process is targeted;

[0044] (3) This invention primarily studies residual stress during side milling. Since residual stress acquisition is time-consuming, typically requiring approximately six minutes for a single location, this invention measures the residual stress at a point on the tool path relative to the acceleration sensor.

[0045] Based on the above assumptions, the problem of residual stress prediction in thin-walled workpiece cutting based on CNN can be formalized as follows:

[0046] Input: Input = (Con, S), where Input represents the influencing factors related to cutting residual stress, Con represents the non-time-varying working condition factor information, and S represents the time-varying working condition factor information for machining process monitoring.

[0047] The non-time-varying working condition factor information Con includes process parameter information, workpiece information and tool information; the time-varying working condition factors monitored during the machining process can be expressed as: S = {S1, S2, ..., S n}, where S i Represents the factors that affect a specific time-varying working condition, such as three-dimensional milling force signal, vibration signal, etc.; n represents the number of signal types.

[0048] Output: Output = (δ xx , δ yy ), where Output is the final predicted residual stress, δ xx , δ yy They refer to the residual stresses along the two directions of the workpiece respectively. The purpose is to use the neural network to construct a mapping function Y(·) from input to output, so that Output = Y(Input).

[0049] Reference Figure 1, as an exemplary embodiment of the present invention, a method for generating typical part working condition adaptability application data includes the following steps:

[0050] Step S1, performing cutting tests and data collection: collecting monitoring data, working condition data, and cutting residual stress values corresponding to the cutting process under time-varying working conditions. The monitoring data includes cutting force signals and vibration signals, and the working condition data includes spindle speed, cutting depth, cutting width, and feed rate.

[0051] Step S2, unified characterization of working condition data information: performing unified characterization processing on the collected working condition data;

[0052] Step S3, monitoring data processing, includes:

[0053] Sub-step S3.1, data preprocessing: performing noise reduction and standardization on the collected monitoring data;

[0054] Sub-step S3.2, extraction and dimensionality reduction of main features: Use deep learning methods to perform adaptive feature extraction on the preprocessed data to obtain a deep coding feature set, then subject the preprocessed data to time domain processing, frequency domain processing, and wavelet packet energy feature extraction to obtain a time domain feature set, a frequency domain feature set, and a wavelet packet energy feature set, respectively; perform feature extraction on the feature data of the obtained deep coding feature set and the time domain feature set, frequency domain feature set, and wavelet packet energy feature set, and perform dimensionality reduction on the extracted feature data;

[0055] Step S4, fusion of monitoring data and working condition data: fusing the working condition data after unified characterization processing with the monitoring data after dimensionality reduction processing to obtain fused feature data, and the fused feature data and the collected corresponding cutting residual stress values form a residual stress data set;

[0056] Step S5, constructing a CNN cutting residual stress prediction model: constructing a CNN-based cutting residual stress prediction model using the residual stress dataset, and training and testing the cutting residual stress prediction model to obtain a final cutting residual stress prediction model;

[0057] Step S6, cutting residual stress prediction: using the final cutting residual stress prediction model, the cutting residual stress prediction value is obtained according to the monitoring data and working condition data during the cutting process under time-varying working conditions.

[0058] In step S1, cutting tests can be performed on multiple workpieces. Force sensors such as Kistle force sensors and acceleration sensors can be used to collect and store monitoring data and real-time monitoring signals during the processing process. After the processing is completed, a dedicated residual stress analyzer is used to perform an X-ray diffraction experiment to obtain the true value of the cutting residual stress.

[0059] The cutting process signal data collected by the sensor during machining includes cutting force data and vibration data in three directions, including the idle travel phase (when the tool has not yet cut into the workpiece) and the idle travel phase (when the tool is withdrawn). Therefore, the data must first be preprocessed to clean it and avoid interference from irrelevant information. The present invention primarily performs two types of data preprocessing: noise reduction on the original signal and data standardization. In particular, sub-step S3.1 can use wavelet threshold decomposition denoising to perform noise reduction.

[0060] When calculating wavelet threshold denoising, there are currently three commonly used methods, namely Sqtwolog threshold (fixed threshold), VisuShrink threshold and VisuShrink threshold + shrinkage factor. The corresponding threshold function construction methods can be roughly divided into three types: hard threshold method, soft threshold method, and soft and hard threshold compromise method. In order to determine the most effective wavelet threshold denoising method, the present invention takes the time-varying data in the thin-walled parts processing data set - 5000 continuous sampling point data in the milling force signal as an example. On the basis of determining the wavelet basis (db10 wavelet), the number of decomposition layers (j=5), and the adjustment coefficient (a=0, 0.5, 1), and excluding the influence of subjective factors, the noise reduction performance of 9 groups of different threshold function combinations is compared from both qualitative and quantitative perspectives. Finally, the threshold selection is VisuShrink threshold + shrinkage factor and soft threshold method. The final processed milling force signal data has a relatively better filtering effect.

[0061] Considering the potential for positive and negative variations in the time-varying operating condition data being processed, this paper specifically uses the Z-score method to normalize this data, unifying the dimensions and magnitudes while preserving the inherent properties of the original data and retaining valuable information. Z-score normalization, also known as standard deviation normalization, normalizes the data to a mean of 0 and a standard deviation of 1.

[0062] It should be understood that the above preprocessing method is not intended to limit the present invention, and other algorithms may also be used to implement it.

[0063] In an exemplary embodiment of the present invention, in sub-step S3.2, the deep coding feature set is obtained by inputting the pre-processed data into a CNN-VAE type variational autoencoder adaptive feature extraction network. Figure 2 As shown in the figure, the variational autoencoder (VAE) is a neural network model for unsupervised learning. The present invention builds a variational autoencoder-like network based on a convolutional neural network, reconstructs the loss function of the self-supervised model, and extracts the latent variables of the preprocessed data as part of the input features for the residual stress model prediction.

[0064] The variational autoencoder feature extraction network consists of two parts, one of which is the convolutional encoding network, such as Figure 3 As shown in , it is responsible for performing convolution operations on preprocessed data to obtain preliminary extracted shallow and deep features, and then input the extracted features into Figure 4 The computing module shown in is another part of the variational self-encoder feature extraction network. The input purpose is to construct a loss function to help train the convolutional coding network.

[0065] The convolutional coding network processes preprocessed data in two stages: Stage 1 and Stage 2. Stages 1 and 2 consist of a convolution up module (increasing the number of output channels) and a convolution down module (reducing the number of output channels), respectively. Each module consists of a convolutional layer, a batch normalization layer, an activation function layer, and a residual connection. The output of the preprocessed data after Stage 1 is called shallow features. After processing in Stage 2, these shallow features become deep features. Mutual information is calculated between the shallow and deep features and the original preprocessed data to construct a loss function. The deep features are then input into the computation module. After global max pooling, the deep features are input into a fully connected layer to obtain two latent variables (the mean vector and the variance vector). The KL divergence of the mean and variance vectors with the normal distribution is calculated to construct the loss function. Normal noise is then added to the obtained mean and variance vectors using a reparameterization method. The sampled reparameterized vector is the final output feature vector, i.e., the deep coding feature set.

[0066] In addition to the use of CNN-VAE to obtain deep encoding feature sets as described above, other deep learning methods can also be used for adaptive feature extraction, such as Transformer and Encoder-decoder structures.

[0067] In order to fully analyze and evaluate the test data, a feature extraction method for different time-varying working conditions is established to enrich the feature set of the final input fitting model. The traditional time-varying data feature extraction method is used to extract the time domain features of the milling force signal and vibration signal in the different experimental groups of data, and a time domain feature matrix consisting of 6×4 time domain feature vectors is obtained. First, the time domain signal is converted into a frequency domain signal using Fourier transform. The conversion process is as follows: Figure 5 shown.

[0068] The traditional data feature extraction method is used to extract the frequency domain features of different experimental data. 12 frequency domain features are extracted for each signal data. There are four types of monitoring data, namely milling data in three different directions and vibration data. Therefore, a frequency domain feature matrix consisting of 12×4 frequency domain feature vectors is obtained. The wavelet packet energy feature extraction method is used to decompose the input signal based on the determination of the wavelet basis (db1 wavelet) and the number of decomposition layers (3) for different experimental data. Figure 6As shown, a wavelet packet energy feature matrix consisting of 8×4 energy feature vectors is obtained.

[0069] After the above work, four sets of feature sets were extracted (the number of deep coding feature sets, time domain feature sets, frequency domain features, and wavelet packet energy feature sets were 16, 24, 48, and 32, respectively). A total of 120 sets of features are needed for the subsequent cutting residual stress prediction model. For the subsequent cutting residual stress prediction model, if the input feature dimension is too high, it will affect the prediction performance of the model, so further screening and optimization are required. In some embodiments of the present invention, random forest (RFR) is used to perform feature screening on the feature data of the deep coding feature set and the time domain feature set, frequency domain feature set, and wavelet packet energy feature set.

[0070] RFR is an algorithm that integrates multiple trees using the concept of ensemble learning. Its basic unit is the decision tree. It uses random forests to select multiple feature sets extracted from time-varying data. RFR provides a metric vector for the fitted numerical attribute columns, with the sum of their importance being 100. The specific implementation process is shown in Table 1 below.

[0071] Table 1 Feature selection process

[0072]

[0073]

[0074] The number of subtrees was determined to be 10,000, and feature selection was performed on the four feature sets obtained above (the number of deep coding features, time domain features, frequency domain features, and wavelet packet energy features were 16, 24, 48, and 32, respectively). RFR was used to perform regression variable importance analysis. In order to evaluate the importance of the features, the average impurity attenuation of 10,000 decision trees was calculated, and the top ten main features in terms of importance were retained to complete the preliminary screening. The screening results are shown as follows: Figure 7 shown.

[0075] Figure 7 The horizontal axis in each figure represents the importance of the feature, and the vertical axis represents the name of the feature. (a) In the depth coding feature importance diagram, the 16 extracted depth features are named from A to P. Here, only the first 10 features are used according to the importance ranking. (b) In the time domain feature importance diagram, to make the image clearer, the six extracted features are named from t1 to t6. For example, "t5_1" represents the margin factor of a directional force. The vertical axes in (c) and (d) are both represented by numbers, representing the corresponding frequency domain features and wavelet packet energy features.

[0076] Time-varying data during machining is a key factor affecting cutting residual stress. As the amount of collected information increases, the number of monitoring signal features also increases, resulting in feature information redundancy. Directly selecting some features randomly will result in a loss of the value of the monitoring information. Therefore, it is necessary to perform dimensionality reduction on the selected feature data to reduce the feature dimension, improve the generalization of the model, and simplify the model complexity. Principal Component Analysis (PCA) can be used to fuse the features filtered by RFR, maximizing the value of the original information and more comprehensively characterizing the intrinsic information of the data. This avoids the weakening of the generalization of the prediction model caused by an excessive number of features or randomly selected features.

[0077] The implementation process is shown in Table 2 below:

[0078] Table 2 PCA implementation method

[0079]

[0080]

[0081] Among them, the m value is set to 2, that is, in each group, only the first two coordinate axes containing the vast majority of the variance are retained. The above four feature sets (depth coding feature set, time domain feature set, frequency domain feature set, and wavelet packet energy feature set) are subjected to dimensionality reduction processing respectively to obtain four new feature sets. These four feature sets are combined together to form the final selected feature set, which is used to obtain the feature set of cutting residual stress together with the uniformly characterized working condition data.

[0082] Then, in step S4, the Concat algorithm can be used to fuse the working condition data that has been uniformly characterized and the monitoring data that has been reduced in dimension, and the fused feature data can be calculated to serve as the input of the cutting residual stress prediction model.

[0083] For step S5, in view of the strong coupling between the time-varying working condition data characteristics and the cutting residual stress in the cutting process, the present invention utilizes the excellent nonlinear modeling ability of the convolutional neural network to learn the complex mapping relationship between the feature set extracted based on the working condition data and the cutting residual stress, and constructs a CNN-based cutting residual stress prediction model, and finally realizes the prediction of cutting residual stress based on the processing process monitoring data.

[0084] The basic steps of the algorithm are shown in Table 3 below:

[0085] Table 3 Algorithm calculation process

[0086]

[0087]

[0088]

[0089] Regarding model training and optimization, the reverse training principle of CNN network models is similar to the Back Propagation Through Time (BPTT) algorithm. This algorithm uses a cost function as the ultimate optimization objective, calculates the gradient using the cost function, then backpropagates and uses an optimizer to learn network parameters, gradually optimizing the deviation between the output value and the true value. This article will not elaborate on this. The Adam algorithm uses very little memory resources and can be used efficiently and reliably as the gradient adaptive optimization algorithm used in this article.

[0090] After obtaining the final cutting residual stress prediction model, the cutting residual stress prediction value can be obtained according to the monitoring data and working condition data during the actual cutting process.

[0091] As described above, the method provided by the present invention uses in-process monitoring data as input, designs a multi-element spatiotemporal feature extraction network capable of reflecting changes in the data structure, reduces the high-dimensional data to correspond to cutting residual stress, and establishes a complex correlation between the monitoring data and residual stress during the process. This method, based on real-world data from the process, can capture the impact of variable factors such as cutting force and chatter on residual stress during the actual process. It also provides a theoretical basis for subsequent online optimization and control of residual stress, allowing timely adjustment of process parameters before the next step to ensure workpiece quality.

[0092] The present invention adopts a method that combines data-driven and experimental methods. Based on the residual stress data set obtained from the experiment, the key features within the data are objectively mined in a data-driven manner. The convolutional neural network structure is used to express the nonlinear relationship between the characteristics of various working condition factors and the cutting residual stress. The cutting residual stress is predicted online efficiently, and then within an acceptable range, certain guidance is provided for timely adjustment of process parameters to regulate the cutting residual stress. The method establishes a complex mapping relationship between the monitoring data and the final residual stress by monitoring various signals during the processing, such as cutting force signals and vibration signals. First, a feature extraction network that can extract features reflecting changes in processing data is designed, and a feature extraction method for low-dimensional representation of high-dimensional data is found to achieve adaptive extraction of sensitive features of cutting residual stress; then a cutting residual stress prediction model based on a convolutional neural network is constructed to quickly achieve the generation of working condition adaptability application data, that is, online prediction of cutting residual stress.

[0093] The method for generating typical part working condition adaptability application data provided by the present invention will now be further described with reference to examples.

[0094] By side milling titanium alloys under different working conditions, the authors collected working condition data, monitoring data, and residual stress data generated during the process to construct a residual stress dataset. The proposed method was then used to mine the complex mapping relationship between time-varying data and cutting residual stress. The experimental workpiece was a thin-walled Ti6Al4V titanium alloy, a typical aviation material, measuring 100 × 100 × 5 mm with a width-to-thickness ratio of 20.

[0095] In the experiment, the full factor experimental method was used to study the influence of spindle speed, feed rate, and radial depth of cut on cutting residual stress. The parameter settings are shown in Table 4 below:

[0096] Table 4 Process parameter settings

[0097]

[0098] The non-time-varying working condition factors in this sample data are spindle speed n, feed speed v f , axial cutting depth a p , radial cutting depth a e , so the time-invariant operating characteristic vector Con of this sample is:

[0099]

[0100] Combined with the combination of process parameters, the following operating condition factor feature vector is constructed:

[0101] Con D1 =[14001500.25]

[0102] Con D2 =[14002500.45]

[0103] Con D3 =[220010000.35]

[0104] Con D4 =[30001000.25]

[0105] Then, the time-varying working condition data preprocessing and feature extraction steps are carried out as follows:

[0106] Step 1: For the three-dimensional milling force data and vibration data in the four data sets D1-D4, the VisuShrink threshold + shrinkage factor and soft threshold method are used to perform noise reduction and filtering on the time-varying working condition data. The processed data are normalized using the regularization method proposed in this paper. After data processing using the Z-score normalization method, the visualization results of the four data sets are as follows: Figure 8 shown.

[0107] Step 2: Based on data preprocessing, feature extraction is performed on the data set. Due to the defects of poor robustness and high time cost of traditional feature extraction, the adaptive and data-driven characteristics of convolutional neural networks are utilized to perform self-supervised learning through the improved VAE variational autoencoder, train the high-dimensional feature extraction network, and decouple the data to obtain the VAE latent feature set; the time domain feature set and frequency domain feature set of the data are obtained by time domain calculation and frequency domain calculation respectively; the wavelet packet energy feature extraction method is used to extract the frequency domain features of the three-dimensional milling force and vibration in the preprocessed data. Based on the determination of the wavelet basis (db1 wavelet) and the number of decomposition layers (3), the input signal vector is decomposed to extract the energy feature set of the data. The visualization results of the four groups of data are shown as follows: Figure 9 shown.

[0108] The third step: based on the convolutional neural network, the variational autoencoder working condition feature extraction is carried out. Under the premise of determining the number of subtrees (10,000), the four sets of feature sets obtained (the number of deep coding features, time domain features, frequency domain features, and wavelet packet energy features are 16, 24, 48, and 32 respectively) are selected. The importance of regression variables is analyzed using RFR, and the importance of features is evaluated by calculating the average impurity attenuation of the decision tree. The top ten main features are retained to complete the preliminary screening. After that, the PCA method is used to reduce the dimension of the 4×10 features obtained and compress the size of the feature quantity to 4×2 features. These four sets of feature sets are combined together and used to participate in the training of the regression model. The visualization results of the four groups of data D1-D4 are as follows: Figure 10 shown.

[0109] Step 4: Based on the prediction of cutting residual stress under time-varying working conditions driven by CNN, the non-time-varying working condition feature vector and the time-varying working condition feature vector are fused using the Concat algorithm to obtain the joint feature set F all As the input of the prediction model, the output is the predicted cutting residual stress. The actual measured cutting residual stress is used as the label, and the cross-validation method is used to verify the accuracy of the cutting residual stress prediction model. The collected processing data set is divided into a training set and a test set in a ratio of 9:1. The full name of the test set is not included in the training; then the model is trained based on the cutting residual stress prediction method. The initial learning rate of the model training is set to 0.001, the batch size is set to 100, and the number of training iterations is set to 100. The Adam algorithm is used to optimize the gradient descent process. The predicted value of the cutting residual stress of the test set is compared with the true value, as shown in the figure. Figure 11 shown.

[0110] According to the CNN prediction model established above, the convergence of the loss function curve during training and the test results are as follows: Figure 12As shown in the loss function curve, the X-axis is the number of iterations, and the Y-axis is the loss function value of the training set. Figure 12 It can be seen that as the training process continues, the loss function of the training set continues to decrease and tends to be stable, that is, the error between the predicted value and the actual value gradually shrinks and finally converges, and the loss values converge to 0.044760817 and 0.013503408 respectively, which shows the effectiveness of the model proposed in this invention.

[0111] In summary, the present invention analyzes the bending moment signal data collected by sensors during milling processing, compares the filtering and noise reduction functions of different threshold combinations and threshold functions using quantitative and qualitative analysis, and performs preprocessing of the monitored data to unify the dimensions and reduce noise, thereby improving the data quality. Using an autoencoder network instead of the traditional feature extraction method, a variational autoencoder network structure based on a convolutional neural network is designed. Leveraging its unsupervised learning advantages, it adaptively extracts features reflecting the changes in the working condition data from the raw data. Observing the decoupled feature distribution of the raw data, it is found to conform to the set standard normal distribution, indicating that the feature extraction model performs well. Leveraging the advantages of convolutional neural networks in processing the complex correlation between time-varying working condition data and cutting residual stress, a convolutional neural network structure is designed. Its input is a sequence of working condition feature data (time-varying and non-time-varying), and its output is the cutting residual stress value under this sequence, ensuring the effectiveness of the model in practical applications. Experimental comparison results show that the proposed working condition adaptability application data generation method can establish the correlation between monitoring data and cutting residual stress during the machining process, providing effective guidance for online optimization of cutting parameters and control of residual stress in actual milling processes.

[0112] Finally, it should be noted that the features mentioned and / or illustrated in the above description of the exemplary embodiments of the present invention may be incorporated into one or more other embodiments in the same or similar manner, combined with features in other embodiments, or substituted for corresponding features in other implementations. The technical solutions obtained by such combination or substitution shall also be deemed to be included in the scope of protection of the present invention.

Claims

1. A method for generating typical part working condition adaptability application data, characterized in that: The following steps are involved: Step 1: Conduct cutting tests and data collection: Collect monitoring data, working condition data, and cutting residual stress values corresponding to the cutting process under time-varying working conditions. The monitoring data includes cutting force signals and vibration signals, and the working condition data includes spindle speed, cutting depth, cutting width, and feed rate. Step 2: Unified characterization of working condition data information: The collected working condition data is uniformly characterized; Step 3, monitoring data processing, includes: Sub-step 3.1, data preprocessing: noise reduction and standardization of the collected monitoring data; Sub-step 3.2, extraction and dimensionality reduction of main features: Use deep learning methods to perform adaptive feature extraction on the preprocessed data to obtain a deep coding feature set. Then, the preprocessed data is processed in the time domain, frequency domain, and wavelet packet energy feature extraction to obtain a time domain feature set, a frequency domain feature set, and a wavelet packet energy feature set, respectively. Feature extraction is performed on the feature data of the obtained deep coding feature set and the time domain feature set, frequency domain feature set, and wavelet packet energy feature set, and dimensionality reduction is performed on the extracted feature data. Step 4, fusion of monitoring data and working condition data: fusing the working condition data after unified characterization processing with the monitoring data after dimensionality reduction processing to obtain fused feature data, which together with the collected corresponding cutting residual stress values constitute a residual stress data set; Step 5, constructing a CNN cutting residual stress prediction model: using the residual stress dataset to construct a CNN-based cutting residual stress prediction model, and training and testing the cutting residual stress prediction model to obtain a final cutting residual stress prediction model; Step 6, cutting residual stress prediction: Using the final cutting residual stress prediction model, the cutting residual stress prediction value is obtained according to the monitoring data and working condition data during the cutting process under time-varying working conditions.

2. The method for generating typical part working condition adaptability application data according to claim 1, characterized in that: In the sub-step 3.1, a wavelet threshold decomposition denoising method is used to perform noise reduction processing.

3. The method for generating typical part working condition adaptability application data according to claim 2, characterized in that: In the sub-step 3.1, noise reduction is performed by using the VisuShrink threshold algorithm and a shrinkage factor and a soft threshold method.

4. The method for generating typical part working condition adaptability application data according to claim 2 or 3, characterized in that: In the sub-step 3.1, the Z-score algorithm is used for standardization.

5. The method for generating typical part working condition adaptability application data according to claim 1 or 2, characterized in that: In the sub-step 3.2, a deep encoding feature set is obtained by inputting the preprocessed data into a CNN-VAE type variational autoencoder adaptive feature extraction network.

6. The method for generating typical part working condition adaptability application data according to claim 1 or 2, characterized in that: In the sub-step 3.2, the RFR method is used to perform feature screening on the feature data of the depth coding feature set and the time domain feature set, the frequency domain feature set, and the wavelet packet energy feature set.

7. The method for generating typical part working condition adaptability application data according to claim 6, characterized in that: In the sub-step 3.2, the PCA method is used to perform dimensionality reduction processing on the selected feature data.

8. The method for generating typical part working condition adaptability application data according to claim 1 or 2, characterized in that: In step 4, the Concat algorithm is used to fuse and calculate to obtain fused feature data.

9. The method for generating typical part working condition adaptability application data according to claim 1 or 2, characterized in that: In step 5, the Adam algorithm is used to optimize the training process of the cutting residual stress prediction model.

Images