An online residual stress prediction method considering geometry-process-physical quantities
By combining deep learning networks with finite element simulation, a multi-channel one-dimensional vector dataset was established, enabling real-time visualization and accurate prediction of residual stress during the processing of thin-walled parts, thus solving the problems of insufficient real-time performance and accuracy in existing technologies.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
- Filing Date
- 2024-07-12
- Publication Date
- 2026-07-07
AI Technical Summary
Existing technologies struggle to achieve real-time and accurate prediction of residual stress during the machining of thin-walled parts, especially considering the lack of information on part geometry and machining processes, which leads to insufficient prediction accuracy.
By combining a deep learning network model with finite element simulation, a multi-channel one-dimensional vector dataset is established by measuring initial residual stress and cutting force data. The InncepU-net network is then trained for online prediction, enabling real-time visualization of three-dimensional residual stress during workpiece machining.
It enables real-time and accurate prediction of residual stress during the processing of thin-walled parts, reduces prediction time, improves the prediction accuracy and real-time performance of the model, and meets the needs of intelligent manufacturing.
Smart Images
Figure CN118821606B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of machining inspection, specifically, it relates to an online method for predicting residual stress considering geometric, process, and physical quantities. Background Technology
[0002] Thin-walled parts, due to their advantages of light weight and compact structure, are widely used in the aerospace field, such as skins, panels, and aero-engine blades. Their machining is primarily done through milling. However, the high material removal rate and low workpiece stiffness during the machining of thin-walled parts make them prone to deformation. This can severely affect subsequent assembly processes, further impacting performance, fatigue life, and safety. Related research indicates that initial residual stress and machining residual stress are crucial factors influencing machining deformation of thin-walled parts. During machining, material removal disrupts the initial residual stress equilibrium, and the internal stress rebalancing process leads to part deformation. Simultaneously, the high thermal load during cutting causes uneven plastic deformation, introducing new residual stresses and ultimately resulting in residual stress deformation. Therefore, predicting and controlling machining deformation hinges on predicting the distribution of the residual stress field.
[0003] Currently, methods for predicting or quantitatively evaluating residual stress mainly include experimental methods, analytical methods, finite element methods, and machine learning methods. However, experimental methods are costly in terms of equipment and time. Analytical methods involve parameters that are only applicable to specific machining conditions and make too many assumptions and simplifications about the physical state of the machining process, making it difficult to guarantee model accuracy. The evolution of the residual stress field during milling is time-varying, and finite element simulation results can only be calculated based on offline data. With the development of intelligent manufacturing, the real-time performance and accuracy of the above prediction methods are insufficient to meet the requirements.
[0004] In recent years, with the development of computer technology, big data, and cloud computing, machine learning methods using data mining have been increasingly widely applied in the manufacturing industry. Machine learning methods are very powerful in modeling complex data relationships, and therefore, there is a great deal of research on establishing the relationship between manufacturing processes and quality. At the same time, the timeliness and accuracy of machine learning can meet the needs of intelligent manufacturing. However, most current machine learning models, when predicting residual stress, only establish the relationship between physical quantities of the processing process and residual stress, without reflecting the stress field distribution of different geometric structures of the parts or the process technology information. Summary of the Invention
[0005] To address the problems existing in the prior art, this invention provides an online method for predicting residual stress that considers geometric, process, and physical quantities.
[0006] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:
[0007] A method for online prediction of residual stress considering geometric, technological, and physical quantities includes the following steps:
[0008] Step 1: Conduct a workpiece milling test, measure the initial residual stress field before machining, and collect cutting force data during machining. The cutting force data includes cutting force and machining trajectory data.
[0009] Step 2: Establish a finite element simulation model of the workpiece milling process. Input the initial residual stress field, cutting force and machining trajectory data before machining into the finite element simulation model of the workpiece milling process to obtain the residual stress field after machining.
[0010] Step 3: Extract the residual stress field before machining, cutting force, cutting layer, and residual stress field after machining to establish a multi-channel one-dimensional vector dataset;
[0011] Step 4: Build a deep learning network model and train the deep learning network model using a multi-channel one-dimensional vector dataset to obtain a stress field prediction model;
[0012] Step 5: Online acquisition of cutting force data during machining is input into the stress field prediction model, and the stress field prediction model outputs the prediction results of the residual stress field after machining.
[0013] Further improvements to optimize the technical solution include:
[0014] In step 1, the material is removed in layers during the milling process of the workpiece; the measured initial residual stress field of the part before machining includes residual stress field data at different positions and depths of the part.
[0015] In step 2, the mesh of the finite element simulation model for milling the workpiece is consistent with the actual cutting depth of the workpiece in the depth direction, that is, the number of mesh layers in the depth direction is equal to the number of layers removed from the workpiece during cutting.
[0016] The specific method of step 2 is as follows: Abaqus simulation software is used to establish a finite element simulation model of workpiece milling. The number of layers l to be removed from the workpiece and the number of finite element mesh nodes m in each layer are set. The input cutting force data is averaged according to the number of mesh nodes in the removal layer to obtain the cutting force data of each mesh node in the removal layer. The cutting force data is decomposed into three-dimensional cutting force data in spatial direction. The birth and death element technique is used to apply the three-dimensional cutting force data to each mesh node according to the tool trajectory in the current analysis step, and the applied cutting force is deactivated in the next analysis step, and so on. Finally, the dynamic milling process simulation is realized in the static analysis step, and the residual stress value of each mesh layer is obtained. The residual stress value of the mesh is mapped to each mesh node to obtain the residual stress field after machining.
[0017] In step 3, Python programming is used to extract the grid cells of each removal layer and establish a cell set E for a multi-channel one-dimensional vector dataset. seti i = 1 to l; Traverse the unit set, extract the pre-machining residual stress value, the triaxial cutting force value applied during machining simulation, the current cutting layer number, and the post-machining residual stress value data for each node in the unit set; Save the extracted data as an n×m dataset, where n is the number of channels, and the channels include: pre-machining residual stress value channel s j j = 1 ~ l, three-dimensional cutting force channel f k k = 1~3, current cutting layer information channel c l residual stress value channel m after processing sx x = 1 to l, total number of channels n = j + k + 1 + x, s j Each channel contains m data points, s j It reflects the geometric and stress field information of the part before processing; f k Each channel contains m data points, f k This reflects the cutting force information experienced by each node of each mesh layer during the machining process; c l The channel contains m data points, reflecting the current cutting layer information; m sx Each channel contains m data points, m sx This reflects the predicted results of the residual stress field under the combined effects of geometry, process, and physical quantities.
[0018] In step 4, the deep learning network model is based on the U-net network model, with the convolutional layers replaced by Inception modules for deeper feature extraction, forming the InncepU-net network. The InncepU-net network processes the input dataset as follows: input data - two Inception modules - pooling layer - two Inception modules - pooling layer - two Inception modules - pooling layer - two Inception modules - pooling layer, thus completing the data encoding stage; then through convolutional layer - two transposed Inception modules - transposed pooling layer - two transposed Inception modules - transposed pooling layer - two transposed Inception modules - transposed pooling layer - two transposed Inception modules - transposed pooling layer - two transposed Inception modules - fully connected layer, thus completing the data decoding stage and outputting the predicted data.
[0019] The structure of the Inception module is as follows:
[0020] Inception(X) = Concat(Conv) 1×1 (X),Conv3×3 (X),Conv 5×5 (X),MaxPool 3×3 (X))
[0021] Among them, Conv 1×1 (X) can be represented as:
[0022]
[0023] Conv 3×3 (X) can be represented as:
[0024]
[0025] Conv 5×5 (X) can be represented as:
[0026]
[0027] Maxpool 3×3 (X) can be represented as:
[0028] MaxPool 3x3 (X)(i,j,k)=f m (max m,n X(3i+m,3j+n,k)+b(k))
[0029] Among them, f c and f m These are the convolutional activation function and the max pooling activation function, respectively, and b(k) is the bias term. The training steps of IncepU-net network are as follows: (1) Read the dataset data and divide it into training set and test set according to the ratio; (2) Initialize the model hyperparameters, which include: learning rate, batch size, initial learning rate and number of training periods; (3) Read the training set data, extract the data features through the deep learning stage, restore and calculate the predicted value through the decoding stage, and then calculate the training loss; (4) Adjust the parameters using the Adam optimizer, and then update the parameters in the network through the backpropagation algorithm; (5) Read the test set data, calculate the predicted value through forward propagation, and then calculate the model test loss; (6) Repeat (3)-(4) until the number of iterations meets the preset maximum number of iterations.
[0030] The deep learning network model is based on the CNN kernel in PyTorch and uses mean squared error as the loss function for the entire network.
[0031]
[0032] Where MSE is the mean squared error, B s It is the batch size, d sP represents the true stress value. s Predict stress values for the model. Evaluate the prediction model using the following evaluation function:
[0033]
[0034] Where Eva is the absolute mean error, N is the number of samples, and d predictedi d represents the predicted value. measuredi It is the actual value.
[0035] Compared with the prior art, the present invention has the following beneficial effects:
[0036] This invention proposes an online prediction method for residual stress that considers geometry, process, and physical quantities. This method differs from existing residual stress prediction models. This invention provides an accurate representation of the geometric state and process information of the workpiece during machining by using a prediction dataset, thereby realizing the online prediction of the three-dimensional residual stress field during machining.
[0037] This invention enables real-time visualization of machined workpieces in a virtual environment and prediction of the three-dimensional residual stress field of the workpiece during the machining process, thereby realizing a digital twin of the geometric and physical quantities of the machining process. Attached Figure Description
[0038] Figure 1 A flowchart of the online prediction method for residual stress considering geometric, process, and physical quantities in this invention;
[0039] Figure 2 This is a structural diagram of the InncepU-net deep learning model proposed in this invention;
[0040] Figure 3 These are processing test and initial residual stress measurement diagrams from embodiments of the present invention;
[0041] Figure 4 This is a diagram illustrating the method for creating the dataset in this embodiment;
[0042] Figure 5 This is a diagram showing the training results of the deep learning network model in this embodiment. Detailed Implementation
[0043] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.
[0044] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.
[0045] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.
[0046] like Figure 1 As shown, the present invention provides an online method for predicting residual stress considering geometric, technological, and physical quantities, comprising the following steps:
[0047] Step 1: Conduct a workpiece milling test, measure the initial residual stress field before machining, and collect cutting force data during machining. The cutting force data includes cutting force and machining trajectory data.
[0048] Step 2: Establish a finite element simulation model of the workpiece milling process. Input the initial residual stress field, cutting force and machining trajectory data before machining into the finite element simulation model of the workpiece milling process to obtain the residual stress field after machining.
[0049] Step 3: Extract the residual stress field before machining, cutting force, cutting layer, and residual stress field after machining to establish a multi-channel one-dimensional vector dataset;
[0050] Step 4: Build a deep learning network model and train the deep learning network model using a multi-channel one-dimensional vector dataset to obtain a stress field prediction model;
[0051] Step 5: Online acquisition of cutting force data during machining is input into the stress field prediction model, and the stress field prediction model outputs the prediction results of the residual stress field after machining.
[0052] This embodiment focuses on a thin-walled skin structure component. The blank dimensions are 120mm × 75mm × 5mm, and the material is TA15 titanium alloy. It is machined on a WFL-M35 machine tool. Figure 3As shown in (a) above. The cutting tool is a Walter four-flute carbide end mill. The cutting force was measured using a Kistler rotary force gauge.
[0053] Milling removes material according to defined cutting layers, and the cutting force value after each layer is machined is recorded. The collected cutting force data has two uses: one is to be input into the prediction model to predict the residual stress state during the machining of that layer; the other is to be input into the finite element simulation model to build the finite element simulation model.
[0054] The residual stress of the part before finishing was measured using the ultrasonic method, such as... Figure 2 As shown in (b) above, the residual stress value is relatively small because the part has undergone stress relief and aging treatment after rough machining. Therefore, the average value of multiple residual stress measurements is calculated, and then a random value within the range of ±1MPa is imported into the finite element simulation model to reconstruct the stress field, thereby obtaining the residual stress field distribution state before machining.
[0055] A finite element simulation model of milling a skinned test piece was established using Abaqus simulation software. The material cutting layers were divided into 8 layers (l=8), with 1380 nodes per layer (m=1380). The collected cutting force data was averaged according to the number of nodes from which the material was removed to obtain the cutting force data for each node. The birth and death element technique was used to apply the cutting force data in three directions to the mesh nodes according to the tool path, and the applied cutting forces were deactivated in the next analysis step to achieve dynamic simulation of the milling process. In the simulation results file, Python programming was used to extract the residual stress values of each mesh layer, map the residual stress values to the mesh nodes, extract the pre-machining and post-machining residual stress values for each mesh node, and then read the cutting force data and cutting layer data to finally form the training and validation dataset for the machine learning model. The process is as follows: Figure 4 As shown (data dimension is 20×1380).
[0056] A total of 1000 machining data points were collected from finite element simulation models and machining experiments, including cutting force, cutting layer, initial residual stress field, and post-machining residual stress field. The existing sample size was insufficient for effective model training; therefore, the dataset was expanded to 40,000 data points using data rotation, Gaussian noise, and data augmentation. The sample data was divided into training and test sets according to a 70% and 30% split. The selection of hyperparameters for the deep learning framework is crucial for achieving good predictive performance. Key hyperparameters include learning rate, batch size, and number of iterations; specific model parameters are shown in Table 1. The model was trained and tested on a workstation, with a training time of approximately 3 days and 14 hours.
[0057] Table 1 Configuration parameters of the InncepU-net network
[0058]
[0059] Model training results are as follows Figure 5 As shown in the figure, the proposed deep learning framework performs well in predicting the evolution of residual stress fields. The calculated average error of the residual stress prediction at each grid node is 3.38 MPa. The figure also shows that the error decreases significantly with increasing training set and training periods. Furthermore, the prediction time of the model was verified: the finite element simulation time for one model is approximately 800 s. The IncepU-net network model requires an average prediction time of 0.017 s, which is 99.97% less than the finite element simulation prediction. The prediction time of the proposed model is much shorter than that of the finite element simulation, therefore, this model can meet the requirements for online prediction of residual stress during processing.
[0060] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.
Claims
1. A method for online prediction of residual stress considering geometric, technological, and physical quantities, characterized in that, Includes the following steps: Step 1: Conduct a workpiece milling test, measure the initial residual stress field before machining, and collect cutting force data during machining. The cutting force data includes cutting force and machining trajectory data. Step 2: Establish a finite element simulation model of the workpiece milling process. Input the initial residual stress field, cutting force and machining trajectory data before machining into the finite element simulation model of the workpiece milling process to obtain the residual stress field after machining. Step 3: Extract the residual stress field before machining, cutting force, cutting layer, and residual stress field after machining to establish a multi-channel one-dimensional vector dataset; Step 4: Build a deep learning network model and train the deep learning network model using a multi-channel one-dimensional vector dataset to obtain a stress field prediction model; Step 5: Online acquisition of cutting force data during machining is input into the stress field prediction model. The stress field prediction model outputs the predicted results of the residual stress field after machining. In step 3, Python programming is used to extract the grid cells of each removal layer and establish a cell set E for a multi-channel one-dimensional vector dataset. seti i = 1~1; Traverse the element set, extract the pre-machining residual stress value, the triaxial cutting force value applied during machining simulation, the current cutting layer number, and the post-machining residual stress value data for each node in the element set; save the extracted data as an n×m dataset, where n is the number of channels and m is the number of finite element mesh nodes per layer, and the channels include: pre-machining residual stress value channel. , Three-dimensional cutting force channel , Current cutting layer information channel c l residual stress value channel after processing , Total number of channels , Each channel contains m data points. It reflects the geometric and stress field information of the part before processing; Each channel contains m data points. This reflects the cutting force information experienced by each node of each mesh layer during the machining process; c l The channel contains m data points, reflecting the current cutting layer information; Each channel contains m data points. This reflects the predicted results of the residual stress field under the combined effects of geometry, process, and physical quantities.
2. The method for online prediction of residual stress considering geometric, technological, and physical quantities according to claim 1, characterized in that, In step 1, the material is removed in layers during the milling process of the workpiece; the measured initial residual stress field of the part before machining includes residual stress field data at different positions and depths of the part.
3. The method for online prediction of residual stress considering geometric, technological, and physical quantities according to claim 2, characterized in that, In step 2, the mesh of the finite element simulation model for milling the workpiece is consistent with the actual cutting depth of the workpiece in the depth direction, that is, the number of mesh layers in the depth direction is equal to the number of layers removed from the workpiece during cutting.
4. The method for online prediction of residual stress considering geometric, technological, and physical quantities according to claim 3, characterized in that, The specific method of step 2 is as follows: Abaqus simulation software is used to establish a finite element simulation model of workpiece milling. The number of layers l to be removed from the workpiece and the number of finite element mesh nodes m in each layer are set. The input cutting force data is averaged according to the number of mesh nodes in the removal layer to obtain the cutting force data of each mesh node in the removal layer. The cutting force data is decomposed into three-dimensional cutting force data in spatial direction. The birth and death element technique is used to apply the three-dimensional cutting force data to each mesh node according to the tool trajectory in the current analysis step. The applied cutting force is deactivated in the next analysis step, and so on. Finally, the dynamic milling process simulation is realized in the static analysis step, and the residual stress value of each mesh layer is obtained. The residual stress value of the mesh is mapped to each mesh node to obtain the residual stress field after machining.
5. The method for online prediction of residual stress considering geometric, technological, and physical quantities according to claim 4, characterized in that, In step 4, the deep learning network model is based on the U-net network model, with the convolutional layers replaced by Inception modules for deeper feature extraction, forming the InncepU-net network. The InncepU-net network processes the input dataset as follows: input data - two Inception modules - pooling layer - two Inception modules - pooling layer - two Inception modules - pooling layer - two Inception modules - pooling layer, thus completing the data encoding stage; then through convolutional layer - two transposed Inception modules - transposed pooling layer - two transposed Inception modules - transposed pooling layer - two transposed Inception modules - transposed pooling layer - two transposed Inception modules - transposed pooling layer - two transposed Inception modules - fully connected layer, thus completing the data decoding stage and outputting the predicted data.
6. The method for online prediction of residual stress considering geometric, technological, and physical quantities according to claim 5, characterized in that, The structure of the Inception module is as follows: ； Among them, Conv 1×1 (X) can be represented as: ； Conv 3×3 (X) can be represented as: ； Conv 5×5 (X) can be represented as: ； Maxpool 3×3 (X) can be represented as: ； Among them, f c and f m These are the convolutional activation function and the max-pooling activation function, respectively, and b(k) is the bias term.
7. The method for online prediction of residual stress considering geometric, technological, and physical quantities according to claim 6, characterized in that, The training steps of IncepU-net network are as follows: (1) Read the dataset data and divide it into training set and test set according to the ratio; (2) Initialize the model hyperparameters, including learning rate, batch size, initial learning rate and number of training periods; (3) Read the training set data, extract the data features through the encoding stage, restore and calculate the predicted value through the decoding stage, and then calculate the training loss; (4) Adjust the parameters using the Adam optimizer, and then update the parameters in the network through the backpropagation algorithm; (5) Read the test set data, calculate the predicted value through forward propagation, and then calculate the model test loss; (6) Repeat (3)-(4) until the number of iterations meets the preset maximum number of iterations.
8. The method for online prediction of residual stress considering geometric, technological, and physical quantities according to claim 7, characterized in that, The deep learning network model is based on the CNN kernel in PyTorch and uses mean squared error as the loss function for the entire network. ； Where MSE is the mean squared error, B s It is the batch size, d s P represents the true stress value. s The predicted stress values are evaluated using the following evaluation function: ； Where Eva is the absolute mean error, N is the number of samples, and d predictedi d represents the predicted value. measuredi It is the actual value.