Motorized spindle thermal error modeling method based on digital twinning assisted deep transfer learning

Abstract

The invention discloses a motorized spindle thermal error modeling method based on digital twin-assisted deep transfer learning, which comprises the following steps: collecting motorized spindle physical system data, and forming a target domain data set by the motorized spindle physical system data; establishing an electric spindle thermal behavior digital twinborn simulation model; simulation data under different working conditions are generated by using the electric spindle thermal behavior digital twinborn simulation model, and the simulation data under different working conditions form a source domain data set; constructing a DGDAN model, and training the DGDAN model by using the target domain data set and the source domain data set to obtain a trained DGDAN model; and testing the trained DGDAN model on the target domain data set and the source domain data set under different working conditions, and outputting respective predicted thermal errors. According to the method, the problems that a traditional deep learning model cannot adapt to a thermal error prediction task under variable working conditions, and the prediction performance is poor due to the fact that an overfitting problem is likely to occur during scarce data modeling are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of thermal error compensation, and in particular to a method for thermal error modeling of an electric spindle based on digital twin-assisted deep transfer learning. Background Art

[0002] Thermal error modeling is a key means to ensure machining accuracy, optimize compensation strategies and improve the reliability of electric spindle systems in the field of intelligent manufacturing. However, this technology faces some major challenges: due to machining requirements and process limitations, electric spindles need to work under time-varying conditions. Due to the large differences in data distribution under variable conditions, the traditional deep learning model cannot adapt to the thermal error prediction task under variable conditions. In addition, when the electric spindle carries a tool for processing, it is difficult to collect thermal error data. Traditional deep learning models are prone to overfitting problems when modeling scarce data, resulting in poor prediction performance. Summary of the invention

[0003] In response to the above-mentioned defects, the present invention proposes an electric spindle thermal error modeling method based on digital twin assisted deep transfer learning, aiming to solve the problem that traditional deep learning models are unable to adapt to thermal error prediction tasks under variable working conditions, and are prone to overfitting problems when modeling with scarce data, resulting in poor prediction performance.

[0004] To achieve this object, the present invention adopts the following technical solutions:

[0005] The thermal error modeling method of the electric spindle based on digital twin assisted deep transfer learning includes the following steps:

[0006] Step S1: collecting the electric spindle physical system data, and forming the electric spindle physical system data into a target domain data set; wherein the electric spindle physical system data includes the temperature data and thermal elongation of the electric spindle;

[0007] Step S2: Establishing a digital twin simulation model of the thermal behavior of the electric spindle;

[0008] Step S3: using the digital twin simulation model of the thermal behavior of the electric spindle to generate simulation data under different working conditions, and forming a source domain data set with the simulation data under different working conditions;

[0009] Step S4: construct a distance-guided domain adversarial network (DGDAN) model, and train the DGDAN model using the target domain dataset and the source domain dataset to obtain a trained DGDAN model;

[0010] Step S5: Test the trained DGDAN model on the target domain dataset and the source domain dataset under different working conditions, and output the respective predicted thermal errors.

[0011] Preferably, in step S2, the following is specifically included:

[0012] Step S21: constructing a geometric model of the electric spindle;

[0013] Step S22: analyzing and calculating the thermal characteristics of the electric spindle physical system;

[0014] Step S23: discretizing the geometric model of the electric spindle into a finite element model, and adding preset boundary conditions and constraints;

[0015] Step S24: obtaining the temperature distribution of the electric spindle by numerically solving the finite element model;

[0016] Step S25: performing post-processing analysis on the temperature distribution and response of the electric spindle to obtain analysis results;

[0017] Step S26: Refine and adjust the finite element model according to the analysis results.

[0018] Preferably, in step S4, the distance-guided domain adversarial network (DGDAN) model includes an input layer, a convolutional attention long short-term memory network (C-ATLSTMN), a domain adversarial mechanism, and a maximum mean difference (MMD) metric;

[0019] Convolutional Attention Long Short-Term Memory Network (C-ATLSTMN) includes Convolutional Neural Network (CNN), Attention Mechanism and Long Short-Term Memory Network (LSTM);

[0020] The domain adversarial mechanism includes a feature extractor G f , domain discriminator G d , thermal error predictor G p .

[0021] Preferably, in step S4, the following sub-steps are specifically included:

[0022] Step S41: Divide the source domain dataset and the target domain dataset into a training set and a test set respectively, and construct an input matrix;

[0023] Step S42: Through the feature extractor G f The input matrices of the source domain dataset and the target domain dataset are extracted as features f s and f t ;

[0024] Step S43: Set the feature f s and f t Input into the maximum mean difference (MMD) metric, measure and reduce the probability distribution of the two domains, and output the maximum mean difference loss function Loss MMD ;

[0025] Step S44: The feature f s and f t Input to the domain discriminator G d In the process, the two domain features are confused, the domain invariant features of the two domains are learned, and the domain discrimination loss function Loss is output. Domain ;

[0026] Step S45: Set the feature f s and f t Input to the attention mechanism and get the output of the attention mechanism;

[0027] Step S46: Input the output of the attention mechanism to the thermal error predictor G p In the output, the thermal error prediction value of the source domain is output, and the loss function Loss of the source domain is calculated source .

[0028] Preferably, in step S4, the following sub-steps are also included:

[0029] The output of the attention mechanism is input to the thermal error predictor G p In the process of , when there are scarce thermal error samples in the target domain, the target domain thermal error prediction values ​​with the same number as the source domain thermal error prediction values ​​are output, and the loss function Loss of the target domain is calculated. target .

[0030] Preferably, in step S4, the following sub-steps are also included:

[0031] According to the maximum mean difference loss function Loss MMD , domain discrimination loss function Loss Domain , Loss function of the source domain source And the loss function Loss of the target domain target , calculate the total loss function Loss, the specific calculation formula is as follows:

[0032] Loss=Loss MMD +Loss Domain +Loss source +Loss target .

[0033] The technical solution provided by the embodiments of the present application may have the following beneficial effects:

[0034] In this solution, the electric spindle physical system data is used as the target domain, and the simulation data generated by the electric spindle thermal behavior digital twin simulation model under different working conditions is used as the source domain. The data sets of the source domain and the target domain are used to build and train the distance-guided domain adversarial network model. The distance-guided domain adversarial network model can realize the prediction of thermal errors in the source domain and the target domain. This solution combines the digital twin technology with the model of deep transfer learning, that is, the distance-guided domain adversarial network model, which can solve the problem of data distribution differences caused by variable working conditions, thereby adapting to the thermal error prediction task under variable working conditions. And when the thermal error data in the target domain is scarce, the distance-guided domain adversarial network model can achieve accurate prediction of thermal errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flowchart of the steps of the electric spindle thermal error modeling method based on digital twin assisted deep transfer learning. DETAILED DESCRIPTION

[0036] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions from beginning to end. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0037] The thermal error modeling method of the electric spindle based on digital twin assisted deep transfer learning includes the following steps:

[0038] Step S1: collecting the electric spindle physical system data, and forming the electric spindle physical system data into a target domain data set; wherein the electric spindle physical system data includes the temperature data and thermal elongation of the electric spindle;

[0039] Step S2: Establishing a digital twin simulation model of the thermal behavior of the electric spindle;

[0040] Step S3: using the digital twin simulation model of the thermal behavior of the electric spindle to generate simulation data under different working conditions, and forming a source domain data set with the simulation data under different working conditions;

[0041] Step S4: construct a distance-guided domain adversarial network (DGDAN) model, and train the DGDAN model using the target domain dataset and the source domain dataset to obtain a trained DGDAN model;

[0042] Step S5: Test the trained DGDAN model on the target domain dataset and the source domain dataset under different working conditions, and output the respective predicted thermal errors.

[0043] The thermal error modeling method of the electric spindle based on digital twin assisted deep transfer learning in this scheme is as follows: Figure 1As shown in the figure, the first step is to collect the electric spindle physical system data and form the electric spindle physical system data into a target domain data set; wherein, the electric spindle physical system data includes the temperature data and thermal elongation of the electric spindle. Specifically, by arranging temperature sensors and position sensors at key positions of the electric spindle, the temperature sensor is used to collect the temperature data of the electric spindle, and the position sensor is used to measure the thermal elongation of the electric spindle, and the thermal error is calculated based on the physical measurement. The second step is to establish a digital twin simulation model of the thermal behavior of the electric spindle. Specifically, the establishment of the digital twin simulation model of the thermal behavior of the electric spindle includes the following key sub-steps: establishing the geometric model of the electric spindle; analyzing and calculating the thermal characteristics of the electric spindle; meshing and constraints; solving and post-processing, and fine-tuning the finite element model. The third step is to use the digital twin simulation model of the thermal behavior of the electric spindle to generate simulation data under different working conditions, and form the simulation data under different working conditions into a source domain data set. Specifically, the digital twin simulation model of the thermal behavior of the electric spindle can obtain the temperature field simulation data under different working conditions. According to the temperature field simulation data, the position of the key temperature measurement point is determined, and the temperature sensor is arranged at the corresponding position of the electric spindle physical system to ensure that the temperature measurement points on the digital twin simulation model of the thermal behavior of the electric spindle have the same position distribution as the temperature sensor in the electric spindle physical system. The fourth step is to construct a distance-guided domain adversarial network (DGDAN) model, and train the DGDAN model using the target domain data set and the source domain data set to obtain the trained DGDAN model. Specifically, the distance-guided domain adversarial network (DGDAN) model includes an input layer, a convolutional attention long short-term memory network (C-ATLSTMN), a domain adversarial mechanism, and domain adaptation guided by the maximum mean difference (MMD) metric. The distance-guided domain adversarial network aligns the distribution of the source domain data set and the target domain data set, learns the domain-invariant features of the two domains, and can realize thermal error modeling when the thermal error data in the target domain is scarce, so that the model established in the source domain can generalize and adapt to the prediction task in the target domain. The fifth step is to test the trained DGDAN model on the target domain data set and the source domain data set under different working conditions, and output the respective predicted thermal errors. Specifically, by calculating the predicted thermal errors on the target domain data set and the source domain data set under different working conditions, and comparing them with the actual measured values, the prediction performance of the trained DGDAN model can be tested. In this embodiment, the performance of the model is tested by multiple migration tasks. For example, the migration task can be to use the simulation data set under the virtual working condition S = 4000rpm of the digital twin model as the source domain to establish the model, and the data set collected under the working condition S = 10000rpm of the physical system as the target domain, and there is no thermal error data in the data set, and the model established in the source domain is migrated to the target domain for prediction. Through a large number of migration task prediction results, it can be seen that the trained DGDAN model has good generalization performance and prediction accuracy.

[0044] In this solution, the electric spindle physical system data is used as the target domain, and the simulation data generated by the electric spindle thermal behavior digital twin simulation model under different working conditions is used as the source domain. The data sets of the source domain and the target domain are used to build and train the distance-guided domain adversarial network model. The distance-guided domain adversarial network model can realize the prediction of thermal errors in the source domain and the target domain. This solution combines the digital twin technology with the model of deep transfer learning, that is, the distance-guided domain adversarial network model, which can solve the problem of data distribution differences caused by variable working conditions, thereby adapting to the thermal error prediction task under variable working conditions. And when the thermal error data in the target domain is scarce, the distance-guided domain adversarial network model can achieve accurate prediction of thermal errors.

[0045] Preferably, in step S2, the following is specifically included:

[0046] Step S21: constructing a geometric model of the electric spindle;

[0047] Step S22: analyzing and calculating the thermal characteristics of the electric spindle physical system;

[0048] Step S23: discretizing the geometric model of the electric spindle into a finite element model, and adding preset boundary conditions and constraints;

[0049] Step S24: obtaining the temperature distribution of the electric spindle by numerically solving the finite element model;

[0050] Step S25: performing post-processing analysis on the temperature distribution and response of the electric spindle to obtain analysis results;

[0051] Step S26: Refine and adjust the finite element model according to the analysis results.

[0052] In this embodiment, in step S21, geometric modeling is a process of describing the shape, size and other features of the electric spindle parts. The shape and size of the electric spindle are converted into a digital geometric model suitable for computer processing through computer-aided design (CAD) software. When establishing a geometric model, it is necessary to consider the working principle and internal structure of the electric spindle, which includes understanding the assembly relationship between various components, such as bearing components, built-in motors and housings. In order to accurately capture the geometry of each component, including size and spatial direction, the size and position of each part must be carefully analyzed and measured. In step S22, by calculating and analyzing the thermal characteristics of the electric spindle physical system, the key factors affecting the accuracy of the digital twin model of the thermal behavior of the electric spindle can be determined. The thermal characteristics of the electric spindle physical system include the heat sources of the built-in motor and the front and rear bearing components of the electric spindle, as well as the convection heat transfer coefficient inside the electric spindle system. In step S23, since different components have different material properties, such as thermal conductivity, specific heat capacity and thermal expansion coefficient, these affect the thermal behavior of the system. In order to accurately simulate the thermal behavior of the electric spindle under actual working conditions, it is necessary to add appropriate boundary conditions and constraints to the finite element model, including defining thermal conductivity, or limiting a certain degree of freedom to represent fixed support or contact with other components. By introducing these boundary conditions and constraints, the finite element model can accurately describe the thermal interactions and thermal behaviors within the physical system of the electric spindle. In step S24, by numerically solving the finite element model, the temperature distribution and response of the electric spindle can be accurately simulated. During the solution process, iterative calculations and solutions to the linear equations are performed. In step S25, the post-processing analysis includes visualizing the temperature distribution of the electric spindle system, extracting insights into temperature changes within the electric spindle assembly, analyzing the spindle deformation caused by thermal expansion, and capturing transient information related to thermal behavior. In step S26, by refining and adjusting the finite element model, the accuracy of the model can be improved and the actual thermal behavior of the electric spindle can be better reflected.

[0053] Preferably, in step S4, the distance-guided domain adversarial network (DGDAN) model includes an input layer, a convolutional attention long short-term memory network (C-ATLSTMN), a domain adversarial mechanism, and a maximum mean difference (MMD) metric;

[0054] Convolutional Attention Long Short-Term Memory Network (C-ATLSTMN) includes Convolutional Neural Network (CNN), Attention Mechanism and Long Short-Term Memory Network (LSTM);

[0055] The domain adversarial mechanism includes a feature extractor G f , domain discriminator G d , thermal error predictor G p .

[0056] In this embodiment, the distance-guided domain adversarial network aligns the distribution of the source domain dataset and the target domain dataset, learns the domain-invariant features of the two domains, and can realize thermal error modeling when the thermal error data in the target domain is scarce, so that the model established in the source domain can be generalized and adapted to the prediction tasks in the target domain.

[0057] In the Convolutional Attention Long Short-Term Memory Network (C-ATLSTMN), the use of convolutional neural networks (CNNs) can extract the input temperature features and effectively capture the local features of the temperature data, thereby achieving the learning and extraction of spatial information and eliminating the high collinearity of multi-dimensional temperature data. The introduction of the attention mechanism can automatically focus on and emphasize the information most relevant to the thermal error in the temperature data when processing the input temperature features, improve the model's expressiveness and performance, and enable the network to focus more on processing important information related to the thermal error prediction task. The use of long short-term memory networks (LSTMs) can effectively capture and process long-term dependencies in sequences and avoid the gradient vanishing problem in traditional RNNs.

[0058] In the domain adversarial mechanism, the feature extractor G f It is replaced by the convolutional layer in the Convolutional Attention Long Short-Term Memory Network (C-ATLSTMN) to extract the features of the source domain and the target domain; the domain discriminator G d Used to confuse the features of the two domains and learn the domain-invariant features of the two domains; thermal error predictor G p It is replaced by a long short-term memory network (LSTM) to output the source domain thermal error prediction value.

[0059] The maximum mean difference (MMD) metric can minimize the difference in data distribution between the two domains and align the data distribution of the two domains.

[0060] Preferably, in step S4, the following sub-steps are specifically included:

[0061] Step S41: Divide the source domain dataset and the target domain dataset into a training set and a test set respectively, and construct an input matrix;

[0062] Step S42: Through the feature extractor G f The input matrices of the source domain dataset and the target domain dataset are extracted as features f s and f t ;

[0063] Step S43: Set the feature f s and f t Input into the maximum mean difference (MMD) metric, measure and reduce the probability distribution of the two domains, and output the maximum mean difference loss function Loss MMD ;

[0064] Step S44: The feature f sand f t Input to the domain discriminator G d In the process, the two domain features are confused, the domain invariant features of the two domains are learned, and the domain discrimination loss function Loss is output. Domain ;

[0065] Step S45: Set the feature f s and f t Input to the attention mechanism and get the output of the attention mechanism;

[0066] Step S46: Input the output of the attention mechanism to the thermal error predictor G p In the output, the thermal error prediction value of the source domain is output, and the loss function Loss of the source domain is calculated source .

[0067] In this embodiment, in the input layer, the source domain and target domain data sets are divided into training sets and test sets respectively to construct the input matrix. The one-dimensional CNN acts as a feature extractor G of the domain adversarial mechanism. f , extract the abstract features of high-dimensional temperature, eliminate the high collinearity of temperature data, and extract the input matrices of the source domain and the target domain as features f s and f t On the one hand, the feature f s and f t Input into the MMD metric, measure and reduce the probability distribution of the two domains, and output the maximum mean difference loss function Loss MMD At the same time, the feature f s and f t It is then input into the domain discriminator G d In the process, the two domain features are confused, the domain invariant features of the two domains are learned, and the domain discrimination loss function Loss is output. Domain On the other hand, the features f of the source domain and the target domain s and f t The input is fed into the attention mechanism to enhance the attention to different parts of the input temperature data. Further explanation, by calculating the importance weight of each input temperature element, attention is focused on the temperature elements related to the current task or timing. The output of the attention mechanism is further fed into the thermal error predictor G p , thermal error predictor G p It is composed of LSTM and can process time series data. Then, it outputs the source domain thermal error prediction value and calculates the loss function Loss of the source domain. source , through back-propagation to update the parameters of the DGDAN model.

[0068] Preferably, in step S4, the following sub-steps are also included:

[0069] The output of the attention mechanism is input to the thermal error predictor G p In the process of , when there are scarce thermal error samples in the target domain, the target domain thermal error prediction values ​​with the same number as the source domain thermal error prediction values ​​are output, and the loss function Loss of the target domain is calculated. target .

[0070] In this embodiment, in actual processing, due to the processing materials, requirements, process restrictions, and measurement method limitations, it is difficult to collect thermal error data of the physical system, and only limited or unusable thermal error samples can be collected. When there are scarce thermal error samples in the target domain, the thermal error predictor G p It can output the thermal error prediction value of the target domain, so that the model established in the source domain can be generalized and adapted to the prediction task in the target domain.

[0071] Preferably, in step S4, the following sub-steps are also included:

[0072] According to the maximum mean difference loss function Loss MMD , domain discrimination loss function Loss Domain , Loss function of the source domain source And the loss function Loss of the target domain target , calculate the total loss function Loss, the specific calculation formula is as follows:

[0073] Loss=Loss MMD +Loss Domain +Loss source +Loss target .

[0074] In this embodiment, by calculating the total loss function Loss, it is helpful to test the prediction performance of the DGDAN model and further optimize the parameters of the DGDAN model. It is further explained that the smaller the total loss function Loss is, the better the prediction performance of the DGDAN model is.

[0075] In addition, each functional unit in each embodiment of the present invention may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0076] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. The electric spindle thermal error modeling method based on digital twin assisted deep transfer learning is characterized by: The following steps are involved: Step S1: collecting the electric spindle physical system data, and forming the electric spindle physical system data into a target domain data set; wherein the electric spindle physical system data includes the temperature data and thermal elongation of the electric spindle; Step S2: Establishing a digital twin simulation model of the thermal behavior of the electric spindle; Step S3: using the digital twin simulation model of the thermal behavior of the electric spindle to generate simulation data under different working conditions, and forming a source domain data set with the simulation data under different working conditions; Step S4: construct a distance-guided domain adversarial network (DGDAN) model, and train the DGDAN model using the target domain dataset and the source domain dataset to obtain a trained DGDAN model; Step S5: Test the trained DGDAN model on the target domain dataset and the source domain dataset under different working conditions, and output the respective predicted thermal errors.

2. The method for modeling thermal error of an electric spindle based on digital twin-assisted deep transfer learning according to claim 1, characterized in that: In step S2, the following is specifically included: Step S21: constructing a geometric model of the electric spindle; Step S22: analyzing and calculating the thermal characteristics of the electric spindle physical system; Step S23: discretizing the geometric model of the electric spindle into a finite element model, and adding preset boundary conditions and constraints; Step S24: obtaining the temperature distribution of the electric spindle by numerically solving the finite element model; Step S25: performing post-processing analysis on the temperature distribution and response of the electric spindle to obtain analysis results; Step S26: Refine and adjust the finite element model according to the analysis results.

3. The method for modeling thermal error of an electric spindle based on digital twin-assisted deep transfer learning according to claim 1, characterized in that: In step S4, the distance-guided domain adversarial network (DGDAN) model includes an input layer, a convolutional attention long short-term memory network (C-ATLSTMN), a domain adversarial mechanism, and a maximum mean difference (MMD) metric; Convolutional Attention Long Short-Term Memory Network (C-ATLSTMN) includes Convolutional Neural Network (CNN), Attention Mechanism and Long Short-Term Memory Network (LSTM); The domain adversarial mechanism includes a feature extractor G f , domain discriminator G d , thermal error predictor G p .

4. The method for modeling thermal error of an electric spindle based on digital twin-assisted deep transfer learning according to claim 3 is characterized in that: In step S4, the following sub-steps are specifically included: Step S41: Divide the source domain dataset and the target domain dataset into a training set and a test set respectively, and construct an input matrix; Step S42: Through the feature extractor G f The input matrices of the source domain dataset and the target domain dataset are extracted as features f s and f t ; Step S43: Set the feature f s and f t Input into the maximum mean difference (MMD) metric, measure and reduce the probability distribution of the two domains, and output the maximum mean difference loss function Loss MMD ; Step S44: The feature f s and f t Input to the domain discriminator G d In the process, the two domain features are confused, the domain invariant features of the two domains are learned, and the domain discrimination loss function Loss is output. Domain ; Step S45: Set the feature f s and f t Input to the attention mechanism and get the output of the attention mechanism; Step S46: Input the output of the attention mechanism to the thermal error predictor G p In the output, the thermal error prediction value of the source domain is output, and the loss function Loss of the source domain is calculated source .

5. The method for modeling thermal error of an electric spindle based on digital twin-assisted deep transfer learning according to claim 4 is characterized in that: In step S4, the following sub-steps are also included: The output of the attention mechanism is input to the thermal error predictor G p In the process of , when there are scarce thermal error samples in the target domain, the target domain thermal error prediction values ​​with the same number as the source domain thermal error prediction values ​​are output, and the loss function Loss of the target domain is calculated. target .

6. The method for modeling thermal errors of electric spindles based on digital twin-assisted deep transfer learning according to claim 5 is characterized in that: In step S4, the following sub-steps are also included: According to the maximum mean difference loss function Loss MMD , domain discrimination loss function Loss Domain , Loss function of the source domain source And the loss function Loss of the target domain target , calculate the total loss function Loss, the specific calculation formula is as follows: Loss=Loss MMD +Loss Domain +Loss source +Loss target 。

Images