Vehicle door rigidity prediction method and device, vehicle and storage medium

Abstract

The invention relates to the technical field of vehicles, in particular to a vehicle door rigidity prediction method and device, a vehicle and a storage medium, and the method comprises the steps: obtaining a three-dimensional geometric model, material information and thickness distribution information of a to-be-predicted vehicle door; determining a loading point based on the three-dimensional geometric model, extracting a point cloud in a preset range by taking the loading point as a center, and marking geometric attributes, thickness attributes, gluing attributes and welding attributes of the point cloud according to the three-dimensional geometric model, the material information and the thickness distribution information; and according to the marked point cloud, performing voxelization processing on the point cloud to obtain a target voxel matrix, and inputting the target voxel matrix into a preset rigidity prediction model to obtain a rigidity value of the loading point. Therefore, the problem that vehicle door rigidity prediction depends on finite element simulation and is difficult to rapidly iterate is solved, voxelization input is constructed by fusing multi-dimensional structure information such as geometry, thickness, welding and gluing, and research and development efficiency and prediction precision are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicles, in particular to a method and device for predicting the rigidity of a vehicle door, a vehicle and a storage medium. BACKGROUND

[0002] Vehicle door rigidity prediction is a core link in the design and optimization of automobile structures, and its accuracy directly affects the realization of vehicle safety, NVH (Noise, Vibration, and Harshness) performance and lightweight targets.

[0003] Related technologies mainly rely on CAE (Computer-Aided Engineering) simulation methods, such as numerical simulation techniques based on FEA (Finite Element Analysis): a three-dimensional mesh model of the vehicle door is constructed, boundary conditions and loads are applied for iterative calculation, and finally the rigidity parameters are output; or structural mechanics performance prediction based on automobile body structure parameters: key parameters such as the geometry parameters (such as key dimensions, cross-sectional shape) of the body, material properties (such as elastic modulus, density) and connection methods are extracted, a parameterized model is constructed, and then data-driven algorithms (such as linear regression, support vector machines or shallow neural networks) are used to predict the overall structural mechanics performance indicators.

[0004] However, the vehicle door rigidity prediction based on the finite element analysis method has the following defects: the calculation amount is large, for a complex vehicle door structure, a large number of elements need to be divided, resulting in an extremely time-consuming calculation process and low iteration efficiency. In the optimization design process, each modification of the structure parameters requires re-performing finite element analysis, and the iteration speed is slow. The vehicle door rigidity prediction based on the body structure parameters has the following defects: the parameterization is mainly the cross-sectional size parameters of the key parts of the body structure, which is difficult to fully represent the spatial geometry of the structure; the material property parameters are mainly simple parameters such as elastic modulus or yield strength; the vehicle door rigidity is affected by multiple types and dimensions of characteristic factors such as the spatial shape of the vehicle door structure, material, thickness and boundary conditions, and there is a performance deficiency in the fusion of multiple dimensions and multiple factors, and the prediction model has poor generalization. SUMMARY

[0005] The present application provides a method and device for predicting the rigidity of a vehicle door, a vehicle and a storage medium to solve the problem of relying on finite element simulation for vehicle door rigidity prediction and difficulty in rapid iteration. By fusing multiple structural information such as geometry, thickness, welding and adhesion to construct voxelized input, high-precision rigidity prediction is achieved, and the prediction effect is improved.

[0006] The first aspect of the present application provides a method for predicting the rigidity of a vehicle door, comprising the following steps: acquire a three-dimensional geometric model, material information and thickness distribution information of a to-be-predicted vehicle door; determine a loading point based on the three-dimensional geometric model, extract a point cloud within a preset range centered on the loading point, and label geometric properties, thickness properties, adhesive properties and welding properties of the point cloud according to the three-dimensional geometric model, the material information and the thickness distribution information; According to the labeled point cloud, the point cloud is voxelized to obtain a target voxel matrix, and the target voxel matrix is input into a preset stiffness prediction model to obtain the stiffness value of the loading point.

[0007] Optionally, in some embodiments, the voxelization of the point cloud based on the labeled point cloud to obtain a target voxel matrix comprises: determine a three-dimensional cubic region within a preset spatial range centered on the loading point; based on a preset voxel resolution, the three-dimensional cubic region is divided into voxel units to form a voxel grid, and based on the voxel grid, a multi-channel voxel matrix is initialized, the multi-channel voxel matrix comprising: a geometric channel, a thickness channel, an adhesive channel and a welding point channel; traverse all the labeled point clouds, map the three-dimensional coordinates of the labeled point clouds to the corresponding voxel units in the voxel grid, and assign values to the corresponding voxel units according to the labeled point clouds to obtain the target voxel matrix.

[0008] Optionally, in some embodiments, before the target voxel matrix is input into the preset stiffness prediction model to obtain the stiffness value of the loading point, it comprises: acquire a training sample set; based on a preset division ratio, the training sample set is divided into a training set, a validation set and a test set; construct a target neural network, input the training set into the target neural network for training to obtain initial model parameters; based on the initial model parameters, input the validation set into the target neural network for performance evaluation, and adjust the initial model parameters according to the performance evaluation results until the joint loss function of the validation set converges to obtain optimal model parameters; based on the optimal model parameters, input the test set into the target neural network for model testing, and when the test result meets the preset requirements, obtain the preset stiffness prediction model.

[0009] Optionally, in some embodiments, the target neural network adopts a three-dimensional residual network comprising an initial convolutional layer, a plurality of three-dimensional residual layers and an output layer as a model architecture.

[0010] Optionally, in some embodiments, the obtaining the training sample set comprises: obtaining a three-dimensional geometric model of a test vehicle door; based on the three-dimensional geometric model of the test vehicle door, setting a plurality of test loading points in a preset region, applying a preset intensity of load to each test loading point to obtain a corresponding displacement, and obtaining a vehicle door surface stiffness value of the corresponding test loading point according to a ratio of the load of each test loading point and the corresponding displacement; extracting a test point cloud within a preset range centered on the test loading point, and labeling geometric properties, thickness properties, adhesive properties and welding properties of the test point cloud to obtain a labeled test point cloud, and combining the labeled test point cloud and the vehicle door surface stiffness value of the test loading point to form the training sample set.

[0011] The second aspect embodiment of the application provides a vehicle door stiffness prediction device, comprising: an obtaining module configured to obtain a three-dimensional geometric model, material information and thickness distribution information of a vehicle door to be predicted; a labeling module configured to determine a loading point based on the three-dimensional geometric model, extract a point cloud within a preset range centered on the loading point, and label geometric properties, thickness properties, adhesive properties and welding properties of the point cloud according to the three-dimensional geometric model, the material information and the thickness distribution information; a prediction module configured to perform voxelization processing on the labeled point cloud to obtain a target voxel matrix, and input the target voxel matrix into a preset stiffness prediction model to obtain a stiffness value of the loading point.

[0012] Optionally, in some embodiments, the prediction module comprises: a determination unit configured to determine a three-dimensional cubic region within a preset spatial range centered on the loading point; a first division unit configured to divide the three-dimensional cubic region into voxel units based on a preset voxel resolution to form a voxel grid, and initialize a multi-channel voxel matrix based on the voxel grid, the multi-channel voxel matrix comprising a geometric channel, a thickness channel, an adhesive channel and a welding point channel; a first generation unit configured to traverse all labeled point clouds, map three-dimensional coordinates of the labeled point clouds to corresponding voxel units in the voxel grid, and assign values to the corresponding voxel units according to the labeled point clouds to obtain the target voxel matrix.

[0013] Optionally, in some embodiments, before the target voxel matrix is input into the preset stiffness prediction model to obtain the stiffness value of the loading point, the prediction module comprises: a first obtaining unit configured to obtain a training sample set; a second dividing unit, configured to divide the training sample set into a training set, a validation set and a test set based on a preset division ratio; a training unit, configured to construct a target neural network, and input the training set into the target neural network for training to obtain initial model parameters; a validation unit, configured to input the validation set into the target neural network based on the initial model parameters for performance evaluation, and adjust the initial model parameters according to the performance evaluation result until a joint loss function of the validation set converges, to obtain optimal model parameters; a test unit, configured to input the test set into the target neural network based on the optimal model parameters for model testing, and obtain the preset stiffness prediction model when a test result meets a preset requirement.

[0014] Optionally, in some embodiments, the target neural network adopts a three-dimensional residual network including an initial convolutional layer, a plurality of three-dimensional residual layers and an output layer as a model architecture.

[0015] Optionally, in some embodiments, the obtaining of the training sample set comprises: a second obtaining unit, configured to obtain a three-dimensional geometric model of a test vehicle door; a second generating unit, configured to set a plurality of test loading points in a preset area based on the three-dimensional geometric model of the test vehicle door, apply a load of a preset intensity to each test loading point, obtain a corresponding displacement, and obtain a vehicle door surface stiffness value of the corresponding test loading point according to a ratio of the load of each test loading point to the corresponding displacement; a third generating unit, configured to extract a test point cloud in a preset range with the test loading point as a center, label geometric properties, thickness properties, adhesive properties and welding properties of the test point cloud, obtain a labeled test point cloud, and combine the labeled test point cloud and the vehicle door surface stiffness value of the test loading point to form the training sample set.

[0016] The third aspect of the embodiments of the present application provides a vehicle, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor executes the program to implement the prediction method of the vehicle door stiffness as described in the above embodiments.

[0017] The fourth aspect of the embodiments of the present application provides a computer readable storage medium having a computer program stored thereon, and the program is executed by a processor to implement the prediction method of the vehicle door stiffness as described in the above embodiments.

[0018] Thus, by acquiring the three-dimensional geometric model, material information and thickness distribution information of the to-be-predicted vehicle door, determining the loading point based on the three-dimensional geometric model, extracting the point cloud in a preset range centered on the loading point, and labeling the geometric attribute, thickness attribute, adhesive attribute and welding attribute of the point cloud according to the three-dimensional geometric model, material information and thickness distribution information, the target voxel matrix is obtained by voxelizing the point cloud according to the labeled point cloud, and the stiffness value of the loading point is obtained by inputting the target voxel matrix into the preset stiffness prediction model. Thus, the problem that the vehicle door stiffness prediction relies on finite element simulation and is difficult to quickly iterate is solved, and the research and development efficiency and prediction accuracy are significantly improved by fusing multi-dimensional structural information such as geometry, thickness, welding and adhesion to construct voxelized input.

[0019] Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and / or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0020] The above and / or additional aspects and advantages of the application will become apparent and be readily understood by considering the following detailed description, including the accompanying drawings, in which: Figure 1 Flowchart of the prediction method of the vehicle door stiffness provided according to an embodiment of the application; Figure 2 Flowchart of the point cloud data attribute labeling provided according to an embodiment of the application; Figure 3 Flowchart of the multi-channel voxelization assignment provided according to an embodiment of the application; Figure 4 Schematic diagram of the initial layer provided according to an embodiment of the application; Figure 5 Schematic diagram of the residual block provided according to an embodiment of the application; Figure 6 Schematic diagram of the residual layer and residual block provided according to an embodiment of the application; Figure 7 Schematic diagram of the Resnet network architecture provided according to an embodiment of the application; Figure 8 Block schematic diagram of the prediction device of the vehicle door stiffness provided according to an embodiment of the application; Figure 9 Structural schematic diagram of the vehicle provided according to an embodiment of the application. DETAILED DESCRIPTION

[0021] Embodiments of the present application are described below in detail, examples of which are shown in the drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0022] A prediction method, device, vehicle and storage medium for the rigidity of a vehicle door are described below with reference to the drawings. In order to solve the problem that the prediction of the rigidity of a vehicle door relies on finite element simulation and is difficult to iterate quickly, the present application provides a prediction method for the rigidity of a vehicle door. In the method, a three-dimensional geometric model, material information and thickness distribution information of a vehicle door to be predicted are obtained, a loading point is determined based on the three-dimensional geometric model, a point cloud in a preset range is extracted with the loading point as the center, and the geometric properties, thickness properties, adhesive properties and welding properties of the point cloud are labeled according to the three-dimensional geometric model, the material information and the thickness distribution information. The rigidity value of the loading point is obtained by inputting the target voxel matrix obtained by voxelizing the point cloud after labeling into a preset rigidity prediction model. Thus, the problem that the prediction of the rigidity of a vehicle door relies on finite element simulation and is difficult to iterate quickly is solved, and the research and development efficiency and prediction accuracy are significantly improved by fusing multi-dimensional structural information such as geometry, thickness, welding and adhesion to construct voxelized input.

[0023] Specifically, Figure 1 A flowchart of a prediction method for the rigidity of a vehicle door provided by an embodiment of the present application is shown in FIG. 1.

[0024] As Figure 1 shown, the prediction method for the rigidity of a vehicle door includes the following steps: In step S101, a three-dimensional geometric model, material information and thickness distribution information of a vehicle door to be predicted are obtained.

[0025] The three-dimensional geometric model is the shape and structure expression of the vehicle door in space, and usually exists in the form of CAD or mesh; the material information includes the types of materials used by each component of the vehicle door and the mechanical properties (such as elastic modulus, density, etc.) thereof; and the thickness distribution information refers to the actual thickness values of the shell elements in different regions of the vehicle door outer panel, reinforcement panel, etc., reflecting the local rigidity characteristics of the structure.

[0026] Specifically, the three-dimensional geometric model of the vehicle door to be predicted is obtained first, and the corresponding material types and thickness data of each structural region are collected synchronously, which constitutes the basis for subsequent voxelization processing and multi-channel feature coding, ensuring that the model can fully perceive the spatial configuration and physical properties of the vehicle door.

[0027] In step S102, a loading point is determined based on the three-dimensional geometric model, a point cloud in a preset range is extracted with the loading point as the center, and geometric properties, thickness properties, adhesive properties and welding properties of the point cloud are labeled according to the three-dimensional geometric model, material information and thickness distribution information.

[0028] Specifically, the embodiment of the application first determines one or more loading points (usually located in the stiffness concern area of the door surface) based on the three-dimensional geometric model of the to-be-predicted door, wherein the loading points can be selected by a user or preset by a system, which is not limited here; then a local point cloud is extracted within a preset spatial range (such as 200mmx200mmx200mm) with each loading point as the center; then, in combination with the material information and thickness distribution information attached to the original model, the unit type to which each point in the point cloud belongs is identified, and its thickness value, whether it is a welding point, and the adhesive material category and other properties are labeled, forming a labeled point cloud with rich structural semantic information.

[0029] In the process of point cloud processing of the data sample, the embodiment of the application extracts point cloud data in a 200mmx200mmx200mm three-dimensional spatial range with each surface stiffness indenter loading point as the center using a neighborhood search method, and considers unit type, thickness and material in point cloud data attribute labeling. The thickness corresponding to the shell unit is the actual thickness of the door outer plate and the reinforcing plate. The adhesive and the welding point are solid elements in the finite element model. The adhesive type label is labeled as 0, 1 and 2 according to the adhesive material type. The welding point type attribute label represents whether it is a welding point with 0 and 1 respectively. For example, Figure 2 The attribute labeling logic block diagram for a certain sample point cloud data, wherein different materials of the adhesive type are represented in a binary manner.

[0030] In step S103, the point cloud is voxelized to obtain a target voxel matrix according to the labeled point cloud, and the target voxel matrix is input into a preset stiffness prediction model to obtain the stiffness value of the loading point.

[0031] Further, in some embodiments, in combination with Figure 3 As shown in the figure, the point cloud is voxelized to obtain a target voxel matrix according to the labeled point cloud, including: determining a three-dimensional cubic region in a preset spatial range with the loading point as the center; dividing the three-dimensional cubic region into voxel units based on a preset voxel resolution to form a voxel grid, initializing a multi-channel voxel matrix based on the voxel grid, the multi-channel voxel matrix including: a geometric channel, a thickness channel, an adhesive channel and a welding point channel; traversing all the labeled point clouds, mapping the three-dimensional coordinates of the labeled point clouds to the corresponding voxel units in the voxel grid, and assigning values to the corresponding voxel units according to the labeled point clouds to obtain the target voxel matrix.

[0032] The voxelization processing refers to a process of converting continuous three-dimensional point cloud data into a regular three-dimensional grid structure (i.e., a voxel grid), wherein each voxel (three-dimensional pixel) can carry one or more attribute values. In the embodiments of the present application, the target voxel matrix is a multi-channel three-dimensional array, and each channel encodes geometric existence, thickness, adhesive type and welding point information, which is used as an input of a deep learning model.

[0033] Specifically, the embodiments of the present application first determine a preset spatial range (such as 200mmx200mmx200mm) centered on the loading point as a voxelization region; then, according to a set voxel resolution (such as 5mm or 10mm), the region is divided into voxel units to form a voxel grid; then, a multi-channel voxel matrix is initialized, including a geometric channel, a thickness channel, an adhesive channel and a welding point channel; subsequently, all labeled point cloud data is traversed, the three-dimensional coordinates of each point are mapped to the corresponding voxel position, and the corresponding channel is valued according to the labeled attribute (for example, the geometric channel is set to 1, the thickness channel is filled with the actual thickness value, and the adhesive / welding point channel is valued according to the category or binary encoding), and finally the target voxel matrix is formed and input into a preset stiffness prediction model to output the stiffness value of the loading point.

[0034] In the actual execution process, first, the point cloud data in the 200mmx200mmx200mm three-dimensional space range around the loading center of the door surface stiffness is calculated, the maximum and minimum values of the point cloud inner point coordinates in the x, y and z directions of the three-dimensional space are calculated, (x min , y min , z min ) and (x max , y max , z max ) constitute the diagonal line of the cuboid voxelization region. In this way, the division region of the voxelization is determined.

[0035] However, the embodiments of the present application can set a voxel resolution, divide the voxelization region into voxel units, and the voxel unit size is (a, b, c). According to the difference between the maximum and minimum coordinate values in a single direction, the length of the voxel unit in the same direction is divided to obtain the number of voxel blocks in a single direction. The number of voxel blocks in three directions is obtained by calculation.

[0036] Thirdly, the shell unit geometric channel voxel unit valuation is completed. It is necessary to determine whether any point marked as a shell unit in the door point cloud data is located in a certain voxel unit space. The determination method is as follows: (1) set all voxel unit values to 0; (2) traverse the point cloud coordinates of the points marked as shell units and the same part, and take a certain point coordinate as (x , , ), then take , and The numerical composition after rounding up (N) a N b N c (3) The coordinates are ( , , H i The point is located in voxel space, at the Nth position in the x-direction. a , Nth in y direction b The Nth direction in the z direction c Within the voxel unit space, the value of the voxel unit is normalized to 1.

[0037] The method for defining various types of data voxel values ​​in this application embodiment is as follows: (1) Definition of thickness channel value: based on H i The thickness of the shell element corresponding to the point is defined as the value of the thickness channel matrix position corresponding to the matrix position of the voxel element at that point.

[0038] (2) Numerical definition of welding and bonding channels: based on H i Check if the point is a solid element. If the point is a solid element, define the value of the welding channel matrix position corresponding to the voxel element at that point as 1. Then check H. i The elastic modulus of the material label is calculated, with amplitudes of 2 and 3 for the two adhesive types, respectively. The values ​​are represented in binary.

[0039] Finally, in this embodiment of the application, the formed target voxel matrix is ​​input into a preset stiffness prediction model, and the stiffness value of the loading point is output.

[0040] Optionally, in some embodiments, before inputting the target voxel matrix into a preset stiffness prediction model to obtain the stiffness value of the loading point, the method includes: acquiring a training sample set; dividing the training sample set into a training set, a validation set, and a test set based on a preset partitioning ratio; constructing a target neural network, inputting the training set into the target neural network for training to obtain initial model parameters; based on the initial model parameters, inputting the validation set into the target neural network for performance evaluation, and adjusting the initial model parameters according to the performance evaluation results until the joint loss function of the validation set converges to obtain the optimal model parameters; based on the optimal model parameters, inputting the test set into the target neural network for model testing, and obtaining the preset stiffness prediction model when the test results meet preset requirements.

[0041] Furthermore, in some embodiments, the target neural network employs a three-dimensional residual network as its model architecture, comprising an initial convolutional layer, multiple three-dimensional residual layers, and an output layer.

[0042] Specifically, before using the target voxel matrix to predict the stiffness, the model needs to be constructed and trained: first, a training sample set composed of labeled point clouds and corresponding stiffness values is obtained, and is divided into a training set, a validation set and a test set according to a preset ratio; then an architecture of a three-dimensional residual network as a target neural network is constructed, which sequentially includes an initial 3D convolution layer, multiple three-dimensional residual layers (including group normalization and ReLU activation) and an output layer; the training set is input into the network for iterative training to obtain initial model parameters; the validation set is used to evaluate the model performance, and the hyperparameters or structure are adjusted according to the joint loss function until the validation loss converges, to obtain optimal model parameters; finally, the model prediction accuracy is verified on the test set, and when the preset indicators are met, the model is confirmed as a usable preset stiffness prediction model.

[0043] In actual execution, the door stiffness prediction is a typical 3D structure analysis task, which needs to capture complex geometric features and mechanical relationships from three-dimensional voxel data, so in order to realize the door surface stiffness prediction, the 3D deep learning model ResNet can be used as the target neural network model.

[0044] ResNet3D solves the training degradation problem of deep network through residual connection, and the core lies in the design of residual block (ResidualBlock). ResNet3D includes an initial layer, a residual layer and an output layer.

[0045] (1) Initial layer: mainly includes a convolution layer and a pooling layer, as shown in Figure 4 .

[0046] Among them, the convolution layer: uses a 7x7x7 3D convolution kernel, the input channel number is 3, the output channel number is 64, the stride is 2, the padding is 3, and the bias term (bias=False) is not used. Then a group normalization layer is followed, the group number is 8, and the normalized data after convolution is processed to stabilize the training process. Then a ReLU activation function is used to make the data have nonlinear characteristics and enhance the expression ability of the model.

[0047] Pooling layer: uses a 3x3x3 maximum pooling kernel, the stride is 2, and the padding is 1. Further reduce the spatial size of the data, reduce the computational complexity, while retaining the main features and key information of the data, and prevent overfitting.

[0048] (2) Residual layer: the model includes four residual layers (layer1, layer2, layer3, layer4), each residual layer is composed of multiple residual blocks (ResidualBlock3D), as shown in Figure 5 .

[0049] The first convolution branch: convolution layer 1 (conv1): 3D convolution kernel of 3x3x3, input channel number is the input channel number of the residual block, output channel number is the output channel number of the residual block, and the stride is determined according to whether it is the first residual block of the residual layer and whether it needs to be down-sampled (usually the stride of the first residual block is 1, and if the residual layer needs to be down-sampled, the stride is 2). The padding is 1, and no bias term is used (bias=False). It is used to extract and convert the features of the input data.

[0050] Group normalization layer 1 (gn1): the number of groups is 8, and the normalized data after convolution is processed to stabilize the training process and accelerate convergence.

[0051] The second convolution branch: convolution layer 2 (conv2): 3D convolution kernel of 3x3x3, input channel number and output channel number are both the output channel number of the residual block, stride is 1, padding is 1, and no bias term is used (bias=False). It is used to further process the features and extract deeper feature information.

[0052] Group normalization layer 2 (gn2): the number of groups is 8, and the normalized data after convolution is processed.

[0053] Jump connection and addition: if the input channel number of the residual block is not equal to the output channel number, or needs to be down-sampled (i.e. the stride of the residual layer is not 1), a down-sampling module (downsample) will be used. The module includes a 1x1x1 3D convolution layer (used to adjust the channel number) and a group normalization layer (the number of groups is 8), which adjusts the input data so that its channel number and spatial size match the output of the second convolution branch.

[0054] The output of the second convolution branch is added to the adjusted (if necessary) input data to realize the jump connection, so that the network can learn the residual mapping between the input and the output, and alleviate the gradient disappearance problem in the deep network.

[0055] The residual layer and the residual block are as shown in Figure 6 , where the dashed line indicates down-sampling, and the solid line indicates no down-sampling. The number of residual blocks in each residual layer is as follows: layer1: contains 2 residual blocks, input channel number is 64, output channel number is 64, and stride is 1. Because the data channel number after the initial convolution layer and the pooling layer is 64, and the residual layer does not need to be down-sampled to reduce the spatial size, the stride and channel number of the residual block remain the same, mainly for feature extraction and nonlinear transformation of the data.

[0056] layer2: contains 2 residual blocks, the first residual block has 64 input channels and 128 output channels, and the stride is 2 (down-sampling, reducing the spatial size of data); the second residual block has 128 input channels and 128 output channels, and the stride is 1. The residual layer reduces the spatial size of data by half through the down-sampling operation of the first residual block, and increases the channel number to 128, which is used to extract higher-level features.

[0057] layer3: contains 2 residual blocks, the first residual block has 128 input channels and 256 output channels, and the stride is 2 (down-sampling); the second residual block has 256 input channels and 256 output channels, and the stride is 1. Further reduce the spatial size of data, increase the channel number, and extract deeper-level features.

[0058] layer4: contains 2 residual blocks, the first residual block has 256 input channels and 512 output channels, and the stride is 2 (down-sampling); the second residual block has 512 input channels and 512 output channels, and the stride is 1. Continue to reduce the spatial size of data and extract the highest-level features.

[0059] (3) The Resnet network architecture constructed in the embodiment of the present application is as shown in Figure 7 , wherein the output layer includes: adaptive average pooling layer (avgpool): performing adaptive average pooling on the feature map after the residual layer, and outputting a size of (1, 1, 1). Regardless of the size of the input feature map, this layer can convert it into a fixed-size output, which is convenient for subsequent fully connected layers to process.

[0060] flatten operation: flattening the three-dimensional data after pooling into a one-dimensional vector, so as to be input into the fully connected layer.

[0061] fully connected layer (head): containing a ReLU activation function and a linear layer. The input feature number of the linear layer is 256 (determined by the previous network structure), and the output feature number is 1, which is used to output a single door stiffness prediction value.

[0062] The training and verification of the model in the embodiment of the present application include the following aspects: (1) Loss function.

[0063] Main loss: mean square error (MSE) is used as the regression loss.

[0064] ; wherein: is the true stiffness value, is the predicted stiffness value, and N is the number of samples.

[0065] Auxiliary loss: To improve the generalization ability of the model, MAE is used as an auxiliary index.

[0066] .

[0067] (2) Optimization strategy.

[0068] Optimizer: The Adam optimizer is used in the embodiments of the application, combined with a dynamic learning rate decay strategy to accelerate convergence.

[0069] ; wherein: represents the parameter (such as weight or bias) at the tthiteration; represents the updated parameter at the (t+1) thiteration; a is the learning rate, which controls the step size of each update; m t is the first moment estimate (i.e., the momentum term), which is an exponentially weighted average of the gradients gt, calculated as: ; wherein, is the decay rate of the first moment estimate, usually set to around 0.9, is the gradient at the tthiteration.

[0070] v t is the second moment estimate (i.e., the RMSProp term), which is an exponentially weighted average of the squared gradients g t 2 , calculated as: ; wherein, β2 is the decay rate of the second moment estimate, usually set to around 0.999; is a very small constant to avoid division by zero, usually set to .

[0071] Regularization: Dropout is used in the fully connected layer.

[0072] (3) Validation method.

[0073] Cross-validation: Five-fold cross-validation is used to evaluate the stability of the model. By randomly dividing the dataset into 5 non-overlapping subsets, each time selecting 1 subset as the test set and the remaining 4 as the training set, repeating the experiment, and finally combining all the results to evaluate the performance of the model.

[0074] Therefore, the embodiment of the present application ensures that the rigidity prediction model has good generalization ability and prediction stability on the basis of fully learning the multi-dimensional structural characteristics of the vehicle door; and the three-dimensional residual network architecture effectively improves the modeling capability for complex geometry and material distribution, thereby providing reliable technical support for subsequent efficient and accurate rigidity prediction.

[0075] Optionally, in some embodiments, the training sample set is obtained by: obtaining a three-dimensional geometric model of a test vehicle door; setting a plurality of test loading points in a preset region based on the three-dimensional geometric model of the test vehicle door, applying a preset intensity of load to each test loading point, obtaining a corresponding displacement, and obtaining a vehicle door surface rigidity value of the corresponding test loading point according to the ratio of the load of each test loading point to the corresponding displacement; extracting a test point cloud in a preset range centered on the test loading point, and labeling the geometric attribute, thickness attribute, adhesive attribute and welding attribute of the test point cloud to obtain a labeled test point cloud; and combining the labeled test point cloud and the vehicle door surface rigidity value of the test loading point to form the training sample set.

[0076] The training sample set is a set of input-output data pairs for training the machine learning model, wherein the input is a local point cloud data centered on the test loading point and labeled with geometric, thickness, adhesive and welding attributes, and the output is a vehicle door surface rigidity value at the loading point obtained by finite element simulation.

[0077] Specifically, the construction process of the training sample set includes: first, obtaining a three-dimensional geometric model of a test vehicle door; then arranging a plurality of test loading points in a preset region (for example, a vehicle door outer panel attention region) of the vehicle door, and applying a fixed load to each point, calculating the corresponding displacement by simulation, and then obtaining the rigidity value; then, a local point cloud of a fixed size is extracted centered on each loading point, and the unit type, thickness, adhesive material type and welding point existence in the local point cloud are labeled; finally, the labeled point cloud and its corresponding rigidity value are paired to form a complete training sample.

[0078] In actual execution, the embodiment of the present application can collect vehicle door three-dimensional geometric models covering various vehicle models, different vehicle door structure forms (stiffening plate distribution forms), various materials (different adhesive materials) and different thickness distribution schemes, construct a vehicle door surface rigidity finite element analysis model, set the loading points to be uniformly distributed in the vehicle door surface rigidity attention region, and value the loading point interval according to the data sample amount requirement, so as to ensure that the surface rigidity sample amount participating in the training is not less than 2000 groups. After completing the surface rigidity finite element analysis of each loading point, the (deformation, force) data of the loading position are extracted as the output data of the vehicle door surface rigidity prediction model. For example, the front door surface rigidity loading point interval of a certain B-class vehicle is valued at 50mm-100mm.

[0079] Thus, the key multi-dimensional physical information affecting the stiffness is ensured to be contained in the sample, thereby providing a high-quality and high-consistency training basis for a subsequent stiffness prediction model based on voxelization and deep learning.

[0080] To sum up, the method for predicting the stiffness of the vehicle door according to the embodiments of the present application firstly acquires the three-dimensional geometric model, material information, thickness distribution and working condition of the vehicle door; then extracts the local point cloud with the loading point as the center and labels the element type, thickness, adhesive and welding point attributes in the local point cloud; next voxelizes the point cloud according to a preset resolution to generate a multi-channel voxel matrix containing geometric, thickness, adhesive and welding information, obtains the stiffness value of the corresponding loading point through finite element simulation, constructs the voxel matrix-stiffness value sample pair, and trains based on a three-dimensional convolution residual network to finally obtain a stiffness model that can be directly used for prediction.

[0081] The embodiments of the present application simplify the vehicle door stiffness analysis process. In actual application, only the geometric data, material type, structure thickness and loading working condition of the vehicle door need to be input, and the trained deep learning model can quickly output the stiffness prediction result, without the tedious finite element pre-processing steps such as mesh division, welding point modeling, boundary and load setting. At the same time, the inference speed of the neural network is much higher than the finite element iterative solution process, which greatly improves the calculation efficiency, and is particularly suitable for rapid evaluation of multiple schemes in the vehicle design stage.

[0082] In addition, the embodiments of the present application have stronger adaptability and generalization ability. The voxelization expression can naturally retain the curvature and other geometric characteristics of the vehicle door outer panel, without the need for artificial definition or extraction of key structure parameters, avoiding subjective bias; and it does not depend on prior division of the vehicle door structure type, and therefore can be applied to vehicle body parts with complex shapes and irregular structures. This data-driven approach more comprehensively reflects the characteristics of the real vehicle body structure, so that the stiffness prediction model remains highly robust and widely applicable when facing frequent changes in design schemes.

[0083] According to the method for predicting the stiffness of the vehicle door provided by the embodiments of the present application, the three-dimensional geometric model, material information and thickness distribution information of the vehicle door to be predicted are acquired, the loading point is determined based on the three-dimensional geometric model, the point cloud in a preset range is extracted with the loading point as the center, the geometric attributes, thickness attributes, adhesive attributes and welding attributes of the point cloud are labeled according to the three-dimensional geometric model, material information and thickness distribution information, the target voxel matrix is obtained by voxelizing the labeled point cloud, and the stiffness value of the loading point is obtained by inputting the target voxel matrix into the preset stiffness prediction model. Thus, the problem that the vehicle door stiffness prediction relies on finite element simulation and is difficult to iterate quickly is solved, and the research and development efficiency and prediction accuracy are significantly improved by constructing the voxelization input by fusing multi-dimensional structure information such as geometry, thickness, welding and adhesion.

[0084] Secondly, a device for predicting stiffness of a vehicle door according to an embodiment of the present application is described with reference to the accompanying drawings.

[0085] Figure 8 is a block schematic diagram of the device for predicting stiffness of a vehicle door according to an embodiment of the present application.

[0086] As shown in Figure 8 , the device for predicting stiffness of a vehicle door 10 comprises an acquisition module 100, a labeling module 200 and a prediction module 300.

[0087] The acquisition module 100 is configured to acquire a three-dimensional geometric model, material information and thickness distribution information of a vehicle door to be predicted.

[0088] The labeling module 200 is configured to determine a loading point based on the three-dimensional geometric model, extract point clouds within a preset range centered on the loading point, and label geometric properties, thickness properties, adhesive properties and welding properties of the point clouds according to the three-dimensional geometric model, the material information and the thickness distribution information. The prediction module 300 is configured to perform voxelization processing on the labeled point clouds to obtain a target voxel matrix, and input the target voxel matrix into a preset stiffness prediction model to obtain a stiffness value of the loading point.

[0089] Optionally, in some embodiments, the prediction module comprises a determination unit, a first division unit and a first generation unit.

[0090] The determination unit is configured to determine a three-dimensional cubic region within a preset spatial range centered on the loading point.

[0091] The first division unit is configured to divide the three-dimensional cubic region into voxel units based on a preset voxel resolution to form a voxel grid, and initialize a multi-channel voxel matrix based on the voxel grid, wherein the multi-channel voxel matrix comprises a geometric channel, a thickness channel, an adhesive channel and a welding point channel.

[0092] The first generation unit is configured to traverse all the labeled point clouds, map three-dimensional coordinates of the labeled point clouds to corresponding voxel units in the voxel grid, and assign values to the corresponding voxel units according to the labeled point clouds to obtain the target voxel matrix.

[0093] Optionally, in some embodiments, before the target voxel matrix is input into the preset stiffness prediction model to obtain the stiffness value of the loading point, the prediction module 300 comprises a first acquisition unit, a second division unit, a training unit, a verification unit and a testing unit.

[0094] The first acquisition unit is configured to acquire a training sample set.

[0095] The second dividing unit is configured to divide the training sample set into a training set, a validation set and a test set based on a preset division ratio.

[0096] The training unit is configured to construct a target neural network, input the training set into the target neural network for training to obtain initial model parameters.

[0097] The verification unit is configured to input the validation set into the target neural network for performance evaluation based on the initial model parameters, and adjust the initial model parameters until the joint loss function of the validation set converges according to the performance evaluation result to obtain optimal model parameters.

[0098] The test unit is configured to input the test set into the target neural network for model testing based on the optimal model parameters, and obtain a preset stiffness prediction model when the test result meets a preset requirement.

[0099] Optionally, in some embodiments, the target neural network adopts a three-dimensional residual network including an initial convolutional layer, a plurality of three-dimensional residual layers and an output layer as a model architecture.

[0100] Optionally, in some embodiments, the training sample set is obtained by a second obtaining unit, a second generating unit and a third generating unit.

[0101] The second obtaining unit is configured to obtain a three-dimensional geometric model of a test vehicle door.

[0102] The second generating unit is configured to set a plurality of test loading points in a preset area based on the three-dimensional geometric model of the test vehicle door, apply a preset strength of load to each test loading point to obtain a corresponding displacement, and obtain a vehicle door surface stiffness value of the corresponding test loading point according to a ratio of the load of each test loading point to the corresponding displacement.

[0103] The third generating unit is configured to extract a test point cloud in a preset range centered on the test loading point, and label geometric properties, thickness properties, adhesive properties and welding properties of the test point cloud to obtain a labeled test point cloud, and combine the labeled test point cloud and the vehicle door surface stiffness value of the test loading point to form the training sample set.

[0104] It should be noted that the foregoing explanation and description of the embodiment of the method for predicting the stiffness of the vehicle door also applies to the embodiment of the device for predicting the stiffness of the vehicle door, which will not be described herein again.

[0105] According to the vehicle door stiffness prediction device provided by the embodiment of the present application, the three-dimensional geometric model, material information and thickness distribution information of the vehicle door to be predicted are obtained, the loading point is determined based on the three-dimensional geometric model, the point cloud in the preset range is extracted with the loading point as the center, the geometric attribute, thickness attribute, adhesive attribute and welding attribute of the point cloud are labeled according to the three-dimensional geometric model, material information and thickness distribution information, the voxelization processing is performed on the labeled point cloud to obtain the target voxel matrix, and the target voxel matrix is input into the preset stiffness prediction model to obtain the stiffness value of the loading point. Therefore, the problem that the vehicle door stiffness prediction depends on finite element simulation and is difficult to quickly iterate is solved, the voxelization input is constructed by fusing multi-dimensional structural information such as geometry, thickness, welding and adhesion, and the research and development efficiency and prediction accuracy are significantly improved.

[0106] Figure 9 A structural schematic diagram of a vehicle is provided for the embodiments of the present application. The vehicle can include: The memory 901, the processor 902 and the computer program stored in the memory 901 and executable on the processor 902.

[0107] The processor 902 implements the vehicle door stiffness prediction method provided in the above embodiments when executing the program.

[0108] Further, the vehicle further includes: The communication interface 903 is used for communication between the memory 901 and the processor 902.

[0109] The memory 901 is used to store the computer program executable on the processor 902.

[0110] The memory 901 can include a high-speed RAM (Random Access Memory, random access memory) memory, and can also include a non-volatile memory, such as at least one disk memory.

[0111] If the memory 901, the processor 902 and the communication interface 903 are independently implemented, the communication interface 903, the memory 901 and the processor 902 can be connected to each other through a bus and complete the communication between each other. The bus can be an ISA (Industry Standard Architecture, industry standard architecture) bus, a PCI (Peripheral Component, peripheral component interconnect) bus or an EISA (Extended Industry Standard Architecture, extended industry standard architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 9 In the figure, only one thick line is used to represent, but it does not mean that there is only one bus or one type of bus.

[0112] Optionally, if the memory 901, the processor 902 and the communication interface 903 are integrated on a chip, the memory 901, the processor 902 and the communication interface 903 can complete mutual communication through an internal interface.

[0113] The processor 902 can be a CPU (Central Processing Unit, central processor), or an ASIC (Application Specific Integrated Circuit, specific integrated circuit), or one or more integrated circuits configured to implement the embodiments of the present application.

[0114] The embodiments of the present application also provide a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the above-mentioned method for predicting vehicle door stiffness.

[0115] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms is not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0116] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "N" is at least two, such as two, three, etc., unless otherwise explicitly specified.

[0117] Any process or method descriptions in flow charts or otherwise described herein can be understood as representing code modules, segments, or portions of code that include one or more executable instructions for implementing the specified logic functions (or steps) and / or can be implemented as hardware or as software stored on a computer readable medium. The preferred embodiments of the present application include additional implementations, in which the order of steps can be different, including substantially simultaneous or reverse order, depending on the functionality involved, which should be understood by those skilled in the art.

[0118] It should be understood that portions of the application can be implemented in hardware, software, firmware, or combinations thereof. In the above embodiments, the N steps or methods can be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. As such, if implemented in hardware, and in another embodiment, any of the following technologies, known in the art, or their combinations, can be used to implement: discrete logic circuitry having logic gates for implementing logic functions on data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays, field programmable gate arrays, etc.

[0119] Those skilled in the art of the present technology can understand that all or part of the steps carried out by the above-mentioned embodiment methods can be completed by programs instructing related hardware, and the programs can be stored in a computer readable storage medium. When the program is executed, it includes one of the steps of the method embodiment or a combination thereof.

[0120] Although the embodiments of the present application have been shown and described above, it should be understood that the above-mentioned embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-mentioned embodiments within the scope of the present application.

Claims

1. A method of predicting a stiffness of a vehicle door, characterized by, The method comprises the following steps: obtaining a three-dimensional geometric model, material information and thickness distribution information of a vehicle door to be predicted; determining a loading point based on the three-dimensional geometric model, extracting a point cloud within a preset range centered on the loading point, and labeling geometric properties, thickness properties, adhesive properties and welding properties of the point cloud according to the three-dimensional geometric model, the material information and the thickness distribution information; performing voxelization processing on the labeled point cloud to obtain a target voxel matrix, and inputting the target voxel matrix into a preset stiffness prediction model to obtain a stiffness value of the loading point.

2. The method of claim 1, wherein, The voxelization processing on the labeled point cloud to obtain the target voxel matrix comprises: determining a three-dimensional cubic region within a preset spatial range centered on the loading point; dividing the three-dimensional cubic region into voxel units based on a preset voxel resolution to form a voxel grid, initializing a multi-channel voxel matrix based on the voxel grid, and the multi-channel voxel matrix comprises a geometric channel, a thickness channel, an adhesive channel and a welding point channel; traversing all the labeled point clouds, mapping three-dimensional coordinates of the labeled point clouds to corresponding voxel units in the voxel grid, and assigning values to the corresponding voxel units according to the labeled point clouds to obtain the target voxel matrix.

3. The method of claim 1, wherein, Before inputting the target voxel matrix into the preset stiffness prediction model to obtain the stiffness value of the loading point, the method comprises: obtaining a training sample set; dividing the training sample set into a training set, a validation set and a test set based on a preset division ratio; constructing a target neural network, inputting the training set into the target neural network for training to obtain initial model parameters; based on the initial model parameters, inputting the validation set into the target neural network for performance evaluation, and adjusting the initial model parameters according to the performance evaluation results until the joint loss function of the validation set converges to obtain optimal model parameters; based on the optimal model parameters, inputting the test set into the target neural network for model testing, and obtaining the preset stiffness prediction model when the test result meets the preset requirements.

4. The method of claim 1, wherein, The target neural network adopts a three-dimensional residual network comprising an initial convolutional layer, a plurality of three-dimensional residual layers and an output layer as a model architecture.

5. The method of claim 1, wherein, The method of obtaining a training sample set comprises: obtaining a three-dimensional geometric model of a test vehicle door; based on the three-dimensional geometric model of the test vehicle door, setting a plurality of test loading points in a preset region, applying a load of a preset intensity to each test loading point to obtain a corresponding displacement, and obtaining a vehicle door surface stiffness value of each test loading point according to a ratio of the load to the corresponding displacement of each test loading point; extracting a test point cloud within a preset range centered on the test loading point, and labeling geometric properties, thickness properties, adhesive properties and welding properties of the test point cloud to obtain labeled test point clouds, and combining the labeled test point clouds and the vehicle door surface stiffness value of the test loading point to form the training sample set.

6. A device for predicting the rigidity of a vehicle door, characterized by The method comprises: an obtaining module configured to obtain a three-dimensional geometric model, material information and thickness distribution information of a vehicle door to be predicted; The labeling module is configured to determine a loading point based on the three-dimensional geometric model, extract a point cloud within a preset range centered on the loading point, and label geometric properties, thickness properties, adhesive properties, and welding properties of the point cloud according to the three-dimensional geometric model, the material information, and the thickness distribution information. The prediction module is configured to perform voxelization processing on the labeled point cloud to obtain a target voxel matrix, and input the target voxel matrix into a preset stiffness prediction model to obtain a stiffness value of the loading point.

7. The apparatus of claim 6, wherein, The prediction module includes: A determination unit configured to determine a three-dimensional cubic region within a preset spatial range centered on the loading point; A first division unit configured to divide the three-dimensional cubic region into voxel units based on a preset voxel resolution to form a voxel grid, and initialize a multi-channel voxel matrix based on the voxel grid, the multi-channel voxel matrix including a geometric channel, a thickness channel, an adhesive channel, and a welding point channel; A first generation unit configured to traverse all the labeled point clouds, map three-dimensional coordinates of the labeled point clouds to corresponding voxel units in the voxel grid, and assign values to the corresponding voxel units according to the labeled point clouds to obtain the target voxel matrix.

8. The apparatus of claim 7, wherein, Before the target voxel matrix is input into the preset stiffness prediction model to obtain the stiffness value of the loading point, the prediction module includes: A first acquisition unit configured to acquire a training sample set; A second division unit configured to divide the training sample set into a training set, a validation set, and a test set based on a preset division ratio; A training unit configured to construct a target neural network, input the training set into the target neural network for training to obtain initial model parameters; A validation unit configured to input the validation set into the target neural network based on the initial model parameters for performance evaluation, and adjust the initial model parameters according to the performance evaluation results until a joint loss function of the validation set converges to obtain optimal model parameters; A test unit configured to input the test set into the target neural network based on the optimal model parameters for model testing, and obtain the preset stiffness prediction model when the test result meets a preset requirement.

9. A vehicle characterized by comprising: It includes: A memory, a processor, and a computer program stored on the memory and executable on the processor, the processor executing the program to implement the prediction method of the stiffness of the vehicle door according to any one of claims 1-5.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the prediction method of the stiffness of the vehicle door according to any one of claims 1-5.

Images