A springback prediction method for automobile panels based on deep learning
Through the deep learning neural network model, the point cloud data of automobile cover parts is predicted, which solves the problem of difficult prediction of rebound volume distribution during stamping and forming, and achieves fast and accurate rebound prediction, reduces time and resource consumption, and improves prediction efficiency and accuracy.
Patent Information
- Application Number
- CN202111262930.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-25
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2041-10-25
AI Technical Summary
During the stamping and forming of automotive cover parts, the prior art is difficult to accurately predict the rebound distribution, resulting in frequent production defects, and manual deconstruction analysis is time-consuming and experience-dependent, so it is impossible to quickly deal with complex problems.
Using a deep learning-based method, a deep learning neural network model is built by acquiring and preprocessing the sample data of the automotive cover parts, a deep learning neural network model is used to extract structural features using point cloud data, and the model is trained to predict rebound values, reducing calculation time and resource consumption.
It realizes fast and accurate prediction of automotive cover parts rebound, reduces time and resource consumption, improves prediction efficiency and accuracy, and can construct training prediction models for cover parts at different locations, which is versatile.
Smart Images

Figure CN113987705B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of finite element analysis and artificial intelligence, and relates to a method for predicting the springback of automobile covering parts based on deep learning. Background Art
[0002] In the development of the national economy, the automobile industry is not only a national pillar industry, but also an important indicator of the national industrial level. At present, influenced by economic globalization, regional integration and cultural diversity, automobile models emerge in an endless stream and are updated more frequently. Designing and producing automobile products with dynamic and beautiful shapes, novel and unique styles quickly and well has become a key link for major automobile companies to accelerate their occupation of market segments and gain overall competitive advantages.
[0003] In the current mainstream automobile production process, the manufacture of body panels, which is an important part of automobile design and manufacturing, mostly adopts mature stamping technology. However, due to objective factors such as the complex structure of automobile panels and the many changes in spatial curved surfaces, various complex production defects such as springback, wear, large displacement, and large deformation often occur during the stamping process, which can easily affect the yield rate of the product. Therefore, it is particularly important to accurately judge the forming performance of automobile panels in the early mold design process in order to effectively predict the springback distribution of the panels after stamping.
[0004] At present, in order to solve this problem, many domestic automobile mold factories have introduced mold CAE analysis technology as a key technology into the mold development process, which has greatly improved the R&D efficiency and quality of automobile cover molds. Although CAE can predict all variables in the stamping process, under current technical conditions, it is still impossible to find the connection between these variables, and engineering analysts with a solid professional background and certain engineering practice experience are still required to conduct manual deconstruction analysis. In particular, as the complexity of the problem increases, it will be very time-consuming for engineers to carry out modeling, pre-processing, solving, and post-processing. In addition, when encountering similar engineering cases in the past, even if there are only slight changes in the geometric shape or set parameters, engineers cannot use the knowledge of previously completed simulation results, and often have to repeat this process several times.
[0005] Deep learning has a strong learning ability for high-order complex functions and has unique advantages in feature extraction, which can make fast and accurate predictions. Using deep learning technology to build a deep learning neural network prediction model and applying this model to the prediction of automotive cover springback can greatly reduce time costs and resource consumption. It is a feasible new idea with broad application prospects. Summary of the invention
[0006] In view of this, the purpose of the present invention is to provide a method for rapid prediction of the rebound of automobile covers based on deep learning. The method can highly extract the structural features of automobile covers, realize rapid and accurate prediction of the rebound of automobile covers, and greatly reduce the consumption of computing time and resources.
[0007] In order to achieve the above object, the present invention provides the following technical solutions:
[0008] A method for predicting the springback of automobile panels based on deep learning includes the following steps:
[0009] S1: Sample data acquisition: obtain the rebound sample data of automobile cover parts, and divide the sample data into simulation data and test data;
[0010] S2: Data preprocessing: preprocess the simulation data and test data respectively;
[0011] S3: Build a deep learning neural network prediction model based on sample data;
[0012] S4: Training deep learning neural network prediction model: training and evaluating the prediction model to obtain the best prediction model;
[0013] S5: Rebound prediction: Use the trained prediction model to perform rebound prediction on new samples.
[0014] Further, in step S1, the sample data is expressed in a point cloud format, and each point contains coordinate values, springback values, process parameters and material parameters; the sample data set is divided into a training set and a validation set; the sample data is divided into simulation data and test data; the simulation data is a set of springback value data corresponding to each grid node obtained by performing stamping springback simulation based on the actual cover structure and stamping process parameters; the test data is a set of coordinate data after springback obtained by on-site measurement after a stamping test with an on-site mold.
[0015] Further, the simulation data is collected through the following steps:
[0016] Sort out the types of geometric models of automotive panels, limit the range of process parameters, select commonly used materials and limit the thickness range of sheet materials, use orthogonal tables to arrange test plans and calculate test results;
[0017] The stamping springback process of automobile cover parts is simulated and analyzed, and a high-density three-dimensional point cloud of the cover parts before and after stamping is directly output. The point cloud is expressed in the form of a point set {Pi|i=1,…,N}), and each point Pi is a vector containing the spatial coordinates (x, y, z) of the point and the springback value. Pi is also the connection point of the finite element analysis grid (usually called a node).
[0018] Further, the test data is collected through the following steps:
[0019] After the stamping process of the automobile cover, when the stress of the automobile cover is released without clamping, the coordinate set measured by the three-dimensional laser scanner is the on-site test scanning data. The test data is expressed in the form of a point cloud; the test data is expressed in the form of a point set {Qi|i=1,…,N}), and each point Qi is a vector containing the spatial coordinates (x, y, z) of the point and the rebound value.
[0020] Further, step S2 specifically includes the following steps:
[0021] S21: sorting and screening invalid data, outlier data and missing data in the simulation data one by one; filtering the test data to remove duplicate data points;
[0022] S22: Encode discrete parameters;
[0023] S23: Sampling the farthest point of the point cloud data to ensure the consistency of the input format.
[0024] Further, step S3 specifically includes the following steps:
[0025] S31: Using the farthest point sampling method, N data points that can represent the geometric features of the model are extracted for each sample;
[0026] S32: input all parameters of all data points;
[0027] S33: Connect the data points to the fully connected layer for dimension increase;
[0028] S34: extracting features from the data points after dimension upgrade;
[0029] S35: reducing the dimension of the data points after feature extraction;
[0030] S36: Output the predicted rebound value.
[0031] Further, step S4 specifically includes the following steps:
[0032] The root mean square error of the rebound parameter is used as the loss function, and the Adam optimization algorithm is used to iteratively optimize the neural network. The optimization goal is to minimize the loss function until the loss function of the training sample data set no longer decreases, and the training is completed.
[0033] Using unsupervised learning methods, all the data in the training set are imported into the neural network in batches to obtain the optimal hyperparameters of the deep learning neural network;
[0034] The deep learning neural network expands the point cloud data after convolution and pooling processing into a vector form, and outputs the rebound value of the automobile cover;
[0035] Calling the validation set to test the deep learning neural network to determine whether the deep learning neural network meets the target conditions, and if so, outputting the deep learning neural network; if not, resetting the hyperparameters of the deep learning neural network until the deep learning neural network meets the target conditions;
[0036] The target conditions include: the rebound prediction accuracy of the deep neural network is not lower than the recognition threshold; the value range of the recognition threshold is 90% to 100%.
[0037] Furthermore, in step S3, point cloud coordinate parameters, process parameters, and material parameters are expressed as parameter vectors, which are used as input parameters of the network together with the grid point coordinates in the calculation space, and are expressed as (x, y, z, f1, f2, f3...fn), wherein (x, y, z) are the coordinates of the grid points in the calculation space, (f1, f2,...fn) are the process parameter vectors of the automobile cover material, and the output layer is a single vector rebound value.
[0038] The beneficial effect of the present invention is that compared with the prior art, the present invention can reduce the number of data points and time consumption as much as possible to improve efficiency while ensuring the accuracy of rebound feature extraction and rebound parameter prediction. Compared with the prior art, the multi-layer perceptron (MLP) neural network model constructed by the present invention can characterize more complex nonlinear relationships, improve the resolution of the rebound characteristics of automobile covering parts, and help to accurately identify the rebound characteristics. The present invention aims to construct and train neural networks for a series of automobile covering parts derived from the same benchmark automobile covering parts, and is highly targeted. For example, different prediction models are trained for covering parts in different positions of the same automobile engine hood, side door, chassis, etc. Thereby, the rebound of the same series of automobile covering parts can be predicted quickly and accurately. The sample data of the present invention is three-dimensional point cloud data, which can better characterize geometric features than the previous two-dimensional data of pictures. Applying deep learning in the field of automobile covering stamping does not rely on manual experience to achieve rapid prediction of rebound values. The shape characteristics, process parameters and springback values of automobile covers are considered together, and the influence of shape characteristics on springback is fully considered. This solves the difficulty that the model can only be applied to covers of the same shape, and realizes the universality of the prediction model.
[0039] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:
[0041] Figure 1 A schematic flow chart of a method for predicting the springback of an automobile cover based on deep learning according to an embodiment of the present invention;
[0042] Figure 2 Schematic diagram of deep learning network for automotive cover springback prediction method based on deep learning. DETAILED DESCRIPTION
[0043] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.
[0044] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.
[0045] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.
[0046] See also Figure 1-2 The embodiment of the present application provides a method for predicting the rebound of automobile covering parts based on deep learning, which can simply and quickly predict the rebound value of automobile covering parts, and the obtained prediction results have sufficient reliability.
[0047] Step S1: Sample data acquisition: obtain sample data of automobile covering parts and collect data for preprocessing;
[0048] S11: The three-dimensional neural network is trained based on the point cloud data samples to complete the deep learning of the nonlinear mapping from the three-dimensional point cloud data to the rebound value; the three-dimensional point cloud samples are the coordinate point sets of the automobile cover model, and the three-dimensional point cloud samples include: a training set and a verification set;
[0049] S12: Collect samples of the rebound value of the automobile cover. In the embodiment of the present application, the samples are divided into two types: simulation calculation data and field test scanning data;
[0050] S13: Simulation data acquisition: Use the finite element analysis software AutoForm to simulate the forming process of the automobile cover and calculate the springback results;
[0051] a) Test condition design: sort out the types of automotive panel geometric models, limit the range of process parameters, select commonly used materials and limit the thickness range of sheet materials, use orthogonal tables to scientifically arrange test plans and calculate test results. The orthogonal test method has the characteristics of a small number of experiments required to complete the test requirements and uniform distribution of data points.
[0052] b) Simulation calculation: Use simulation software (Autoform) to simulate and analyze the stamping springback process of automobile cover parts, and directly output high-density three-dimensional point cloud of cover parts before and after stamping. The point cloud can be expressed in the form of point set {Pi|i=1,…,N}), and each point Pi is a vector containing the spatial coordinates (x,y,z) of the point and the value of the springback. Pi is also the connection point of the finite element analysis grid (usually called node).
[0053] S14: Field test scanning data acquisition;
[0054] a) After the stamping process of the automobile cover, when the automobile cover is not clamped by a fixture and the stress is released, the coordinate set measured by the 3D laser scanner is the field test scanning data, and the data is expressed in the form of point cloud.
[0055] b) The data is also expressed in the form of a point set {Qi|i=1, ..., N}), where each point Qi is a vector containing the spatial coordinates (x, y, z) of the point and the rebound value.
[0056] S15: Each cover has different geometric models, different material parameters, different thickness materials, and different process parameters corresponding to different sample data.
[0057] S16: As the types of automobile covers are complex and diverse, models of the same type are usually put together to train a prediction model to reduce the complexity of the network and improve the prediction accuracy.
[0058] S17: Each sample has corresponding label information, and the label information in the present invention is the size of the rebound value.
[0059] Step S2: Data preprocessing: preprocessing the simulation data and the test data respectively;
[0060] S21: Simulation data processing:
[0061] Use Python and other development languages to import CSV files and sort and filter invalid data, outlier data, and missing data one by one.
[0062] In order to make the rebound value more accurate, the data can be processed, for example, by removing simulation singularities.
[0063] Autoform can only output the rebound value of the triangular mesh. Before model training, the triangular face nodes need to be converted into node nodes.
[0064] S22: Field test scan data processing: Data filtering: The original collected point cloud data often contains a large number of hash points and isolated points, and duplicate data points are removed. Compare with the original model and extract the rebound value.
[0065] S23: For discrete variables, they are coded digitally, for example, material 1, 2, 3…
[0066] S24: Sample the farthest point of the point cloud data to ensure the consistency of the input scale. The input parameter format is: input(x,y,z,f1,f2…fn), and the output parameter format is (S), where x,y,z are point cloud coordinate values, and f1,f2,fn are material, process and other parameter values. S is the coordinate rebound value.
[0067] S25: Split the data sample into training set and validation set.
[0068] Step S3: Building a deep learning neural network prediction model;
[0069] S31: The farthest point sampling method is used to extract N data points that can represent the geometric features of the model for each sample to ensure the consistency of the input parameter format.
[0070] S32: Input the coordinate values of N data points, process parameters and other M data as input parameters into the neural network.
[0071] S33: Each data point is connected to a fully connected layer, and the dimension is increased to 1024 feature vectors in turn.
[0072] S34: Perform the maximum max pooling operation on the N×1024 features to eliminate the disorder of the point cloud and obtain 1024 vectors.
[0073] S35: Concatenate N×64 and 1024 features to obtain the local feature value of each point
[0074] S36: The concatenated N×1088 features are fully connected to finally obtain 10000×1 features, which are the rebound values of 10000 data points.
[0075] Step S4: training the deep learning neural network prediction model: training and evaluating the prediction model to obtain the best prediction model;
[0076] S41: Using unsupervised learning, import all the data in the training set into the neural network in batches. Obtain the optimal hyperparameters of the deep learning neural network;
[0077] S42: The deep learning neural network expands the point cloud data after convolution and pooling processing into a vector form, and outputs the rebound value of the automobile cover.
[0078] S43: calling the validation set to test the deep learning neural network to determine whether the deep learning neural network meets the target conditions, and if so, outputting the deep learning neural network; if not, resetting the hyperparameters of the deep learning neural network until the deep learning neural network meets the target conditions;
[0079] S44: The target conditions include: the rebound prediction accuracy of the three-dimensional neural network is not lower than the recognition threshold; the recognition threshold value ranges from 90% to 100%.
[0080] Step S5: Rebound prediction: The model data to be predicted is used as a new sample, the input parameters are input into the prediction model, the structural hyperparameters of the optimal prediction model are loaded, and the corresponding rebound prediction value is obtained.
[0081] In a specific implementation of a method for predicting the rebound of an automobile cover based on deep learning, the automobile cover refers to a certain automobile axle housing part. The method comprises the following steps
[0082] 1. Bridge shell data set collection.
[0083] a) A total of 1230 samples of bridge housings with different geometric structures, different materials, and different process parameters were selected. The simulation samples and test samples were placed in the same file.
[0084] b) The bridge shell simulation data and the field test scanning data constitute the entire sample data set, and the sample data set is used as training data and is randomly divided into a training set and a verification set according to an 8:2 ratio.
[0085] 2. Data Preprocessing
[0086] a) Only the rebound value of the triangular mesh can be output from the autoform. Find the coordinates of the center point of the triangular mesh, use the rebound value of this triangular mesh as the rebound value of the center point coordinates, and output the rebound value in the format of (x, y, z, S).
[0087] b) The test data is a set of coordinates of the stamped bridge shell parts measured by a 3D laser scanner. The coordinate difference is the springback value after alignment and comparison with the original model. The output format of the result after comparison is (x, y, z, S).
[0088] c) Screen the output simulation and test point cloud data to remove redundant, duplicate, and bad data points.
[0089] d) Using the farthest point sampling method, 10,000 data points are extracted for each sample. 10,000 data points are sufficient to characterize the geometric features of the bridge housing parts. The coordinate values (x, y, z), material parameter 1, material thickness parameter 2, process parameter blank holding force 3 and other 6 data of the 10,000 data points are used as input parameters and stored in pkl format for later use.
[0090] 3. Construction of deep learning neural network prediction model
[0091] a) Deep learning neural networks are compiled and trained based on the TensorFlow platform.
[0092] b) Each of the 10,000 extracted data points is connected to a fully connected layer. After passing through multiple layers of fully connected layers, the data is upgraded to 128, 256, 512, and 1024 feature vectors respectively.
[0093] c) Perform the maximum max pooling operation on the 10000×1024 features to eliminate the disorder of the point cloud and obtain 1024 vectors.
[0094] d) Copy the above 1024 vectors 10,000 times and concatenate them with the 10,000×64 features to obtain 10,000×1088 features, which are the local eigenvalues of each point.
[0095] e) Perform multiple full-connection connections on the concatenated 10000×1088 features, and finally obtain 10000×1 features, which is the rebound value of 10,000 data points
[0096] 4. Deep learning neural network prediction model training
[0097] a) Store all training set data in pkl format and import them into the neural network in batches. After training, the optimal prediction model is obtained and all weight parameters in the optimal prediction module are stored in the file weight.h5.
[0098] b) Using the Adam optimization algorithm, the neural network is iteratively optimized. The optimization goal is to minimize the loss function. The initial learning rate is set to 1×10-3 until the loss function of the training set no longer decreases and the training is completed.
[0099] 5. Data prediction
[0100] a) Store the new sample input parameters in pkl format, extract all weight parameters in the weight.h5 file, input them into the prediction model, and obtain the rebound result of the new model.
[0101] 6. Evaluation Indicators
[0102] Evaluation indicators: Four indicators are set to evaluate the prediction performance: root mean square error (RMSE), mean absolute error (MAE), mean relative error (MRE), and goodness of fit (R2).
[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.
Claims
1. A method for predicting the springback of automobile panels based on deep learning, characterized in that: The following steps are involved: S1: Sample data acquisition: Obtain the rebound sample data of automobile cover parts, and divide the sample data into simulation data and test data; the sample data is in point cloud format, and each point contains coordinate values, rebound values, process parameters and material parameters; the sample data set is divided into training set and verification set; the sample data is divided into simulation data and test data; the simulation data is a set of rebound value data corresponding to each grid node obtained by stamping rebound simulation based on the actual cover part structure and stamping process parameters; the test data is a set of coordinate data after springback obtained by on-site measurement after stamping test with on-site mold; The simulation data is collected through the following steps: Sort out the types of geometric models of automotive panels, limit the range of process parameters, select commonly used materials and limit the thickness range of sheet materials, use orthogonal tables to arrange test plans and calculate test results; The stamping springback process of automobile cover parts is simulated and analyzed, and high-density three-dimensional point clouds before and after stamping of the cover parts are directly output. The point cloud is represented in the form of a point set {Pi | i = 1, ..., N}). Each point Pi is a vector containing the spatial coordinates (x, y, z) of the point and the springback value. Pi is also the connection point of the finite element analysis grid. The test data is collected through the following steps: After the stamping process of the automobile cover, without clamping, the automobile cover is released from stress. The coordinate set measured by the 3D laser scanner is the field test scanning data. The test data is expressed in the form of point cloud; the test data is expressed in the form of point set {Qi | i = 1, ..., N}), and each point Qi is a vector containing the spatial coordinates (x, y, z) of the point and the value of the springback; S2: Data preprocessing: preprocess the simulation data and test data respectively; S3: Build a deep learning neural network prediction model based on sample data; specifically include the following steps: S31: Using the farthest point sampling method, N data points that can represent the geometric features of the model are extracted for each sample; S32: Input all parameters of all data points; S33: Connect the data points to the fully connected layer for dimension increase; S34: extracting features from the data points after dimension upgrade; S35: reducing the dimension of the data points after feature extraction; S36: output predicted rebound value; S4: Training deep learning neural network prediction model: training and evaluating the prediction model to obtain the best prediction model; S5: Rebound prediction: Use the trained prediction model to perform rebound prediction on new samples.
2. The method for predicting the springback of automobile cover parts based on deep learning according to claim 1 is characterized in that: Step S2 specifically includes the following steps: S21: sorting and screening invalid data, outlier data and missing data in the simulation data one by one; filtering the test data to remove duplicate data points; S22: Encode discrete parameters; S23: Sampling the farthest point of the point cloud data to ensure the consistency of the input format.
3. The method for predicting the springback of automobile cover parts based on deep learning according to claim 1 is characterized in that: Step S4 specifically includes: The root mean square error of the rebound parameter is used as the loss function, and the Adam optimization algorithm is used to iteratively optimize the neural network. The optimization goal is to minimize the loss function until the loss function of the training sample data set no longer decreases, and the training is completed. Using unsupervised learning methods, all the data in the training set are imported into the neural network in batches to obtain the optimal hyperparameters of the deep learning neural network; The deep learning neural network expands the point cloud data after convolution and pooling processing into a vector form, and outputs the rebound value of the automobile cover; Calling the validation set to test the deep learning neural network to determine whether the deep learning neural network meets the target conditions, and if so, outputting the deep learning neural network; if not, resetting the hyperparameters of the deep learning neural network until the deep learning neural network meets the target conditions; The target conditions include: the rebound prediction accuracy of the deep neural network is not lower than the recognition threshold; the value range of the recognition threshold is 90% to 100%.
4. The method for predicting the springback of automobile cover parts based on deep learning according to claim 1, characterized in that: In step S3, the point cloud coordinate parameters, process parameters, and material parameters are expressed as parameter vectors, which are used as input parameters of the network together with the grid point coordinates in the calculation space, expressed as (x, y, z, f1, f2, f3...fn), where (x, y, z) are the coordinates of the grid points in the calculation space, (f1, f2,...fn) are the material and process parameter vectors of the automobile cover; the output layer is a single vector rebound value.
