Multimodal fusion method for prediction of collision plastic deformation point cloud sequence

Abstract

The present invention discloses a method for predicting a collision plastic deformation point cloud sequence of multimodal fusion, which belongs to the technical field of structural collision deformation point cloud prediction, and comprises the following steps: S1: constructing a multimodal fusion prediction agent model for a structural collision deformation process; S2: constructing a data set and performing preprocessing, dividing it into a training set and a verification set; S3: using the training set to train the multimodal fusion prediction agent model for the structural collision deformation process, and using the verification set to verify the model; S4: inputting the geometric parameters of the structure, the material design parameters and the initial finite element mesh node point cloud of the structure determined by the geometric parameters, and predicting the collision resistance response data and the structural collision deformation point cloud data.

Description

Technical Field

[0001] The present invention belongs to the technical field of structural collision deformation point cloud prediction, and in particular relates to a multi-modal fusion collision plastic deformation point cloud sequence prediction method. Background Art

[0002] The optimization problem of structural crashworthiness under collision deformation mode constraints requires the rapid construction of individual deformation mode constraints for the population. However, the time cost of obtaining structural collision deformation data through finite element simulation is too high and difficult to apply in practice. Therefore, it is necessary to use a proxy model to quickly predict the collision deformation data of thin-walled structures.

[0003] There are still great challenges in combining the finite element mesh node point cloud sequence of the structural collision deformation process with the structural crashworthiness optimization process. The key lies in the slow speed of finite element simulation solution.

[0004] Usually, in the process of optimizing structural crashworthiness under the constraints of collision deformation modes, it is necessary to solve the structural crashworthiness response and collision deformation representation thousands of times. If the finite element method is used for simulation and solution, the time cost and computing resources consumed are incalculable. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a multi-modal fusion collision plastic deformation point cloud sequence prediction method, to construct an agent model that can simultaneously complete the structural crashworthiness response and structural collision deformation representation prediction, and to accelerate the structural crashworthiness optimization process.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A multi-modal fusion collision plastic deformation point cloud sequence prediction method includes the following steps:

[0008] S1: Construct a multi-modal fusion prediction agent model for the structural collision deformation process;

[0009] S2: Build and preprocess the dataset, and divide it into training set and validation set;

[0010] S3: using the training set to train the multimodal fusion prediction agent model of the structural collision deformation process, and using the validation set to validate the model;

[0011] S4: Input the geometric parameters of the structure, material design parameters and the initial finite element mesh node point cloud of the structure determined by the geometric parameters, and predict the crashworthiness response data and structural collision deformation point cloud data.

[0012] Furthermore, the multimodal fusion prediction agent model of the structural collision deformation process includes a multimodal feature embedding and fusion layer, a feature extraction layer, a crashworthiness response prediction layer, and a structural collision deformation point cloud prediction layer;

[0013] The multimodal feature embedding and fusion layer includes a feature embedding module and a feature fusion module; the feature embedding module consists of a point cloud slice merging PPM module and a fully connected layer MLP module, which is used to uniformly map data of different modes to a dimensional space of size C; the feature fusion module is used to amplify the channel of variable features through feature amplification operations and align them with point cloud features; the amplified features are added to the point cloud features to achieve feature fusion and output to the feature extraction layer; the design variable embedding features before amplification are sent to the crashworthiness response prediction layer;

[0014] The feature extraction layer is composed of an encoder module, a decoder module and a Bottle Neck module, which are used to extract the spatial information features of the entire point cloud from local to global, and send them to the crashworthiness response prediction layer and the structural collision deformation point cloud prediction layer respectively;

[0015] The crashworthiness response prediction layer is used to predict crashworthiness response indicators, including energy absorption SEA response and maximum peak force PCF response;

[0016] The structure collision deformation point cloud prediction layer is used to predict the structure collision deformation point cloud.

[0017] Furthermore, in the feature extraction layer, the encoder module is composed of three point cloud sliding window attention PSAB modules and two PPM modules stacked alternately, the decoder module is symmetrical with the encoder module, and is composed of three PSAB and two point cloud pixel rearrangement operation PPS modules stacked alternately, and the Bottle Neck module is composed of "PPM-PSAB-PPS" in series; the PSAB module is regarded as a point cloud adaptive attention feature extraction module, and the size of the fused features processed by PSAB remains unchanged; the PPM module is regarded as a downsampling layer in the encoder module, and its function is to merge the fused features transmitted from the upper layer within a certain range of adjacent layers, and thereby obtain the features of a higher layer as the high-level abstraction of the current features; the PPS module is regarded as an upsampling layer, and its function is to restore the abstract high-level features to the corresponding underlying feature dimensions;

[0018] The features output by the first PSAB module of the encoder module are not only input to the next PPM module, but also fused with the features output by the last PSAB module of the decoder module as the output of the entire feature extraction layer;

[0019] The features output by each PPM module in the encoder are not only passed to the next layer, but also fused with the features output by the PSAB module in the previous layer corresponding to its symmetric PPS in the decoder module as the input of the symmetric PPS module;

[0020] After entering the feature extraction layer, the fused features extract features at multiple scales in the encoder module. The size of the fused features gradually changes from [C, N / 4] to [4C, N / 16]. At the same time, the fused features of each scale are retained. The features passing through the encoding layer are first subjected to the last downsampling feature extraction, attention feature extraction, and feature recovery process in the Bottle Neck module, which are completed by the PPM, PSAB, and PPS modules respectively. The size of the features passing through the Bottle Neck remains unchanged.

[0021] In the decoder module, the size of the fused features is gradually restored from [4C, N / 16] to [C, N / 4], and in the process is fused with the features of the same scale retained from the encoder module.

[0022] Furthermore, in the structural collision deformation point cloud prediction layer, the features first pass through a PPS module, and the feature size is restored from [C, N / 4] to [t×3, N], where t represents the number of time steps contained in the predicted structural collision deformation point cloud sequence. Finally, a PSAB module is used to fine-tune the final prediction output.

[0023] Furthermore, in the crashworthiness response prediction layer, the feature size input to this layer is consistent with the feature size output by the feature embedding and fusion layer, both of which are [C, N / 4], where C is the size of the embedded feature vector and N is the number of nodes contained in the point cloud; the input feature is first compressed into a feature vector through a feature pooling layer, and the size is uniformly changed from [C, N / 4] to [1, 256]; the initial embedded variable feature obtained from the feature embedding module is mapped from a feature vector of size C to a feature vector space of size 256 through a feature alignment module; then, the two parts of the feature are fused, and the fused feature is finally obtained through a unified crashworthiness response prediction layer to obtain the final predicted crashworthiness response;

[0024] The feature alignment module consists of two MLP layers and the corresponding GELU activation function;

[0025] The feature pooling layer includes a one-dimensional convolution layer and an adaptive average pooling layer (AdaptiveAvgPool1D) as well as transpose and flatten operations. The adaptive average pooling layer adaptively adjusts the size of the pooling kernel according to the size of the input feature tensor so that the dimension of the output feature meets the preset size.

[0026] The MLP layer is a single-layer fully connected layer and does not contain an activation function. Its input is in the form of a tensor, and its weight is only effective for the last dimension of the tensor, which can be expressed as follows:

[0027] MLP(X)=Xw+b

[0028] Where X∈R […，c] Represents the input feature tensor with the last dimension size c, w∈R c×d represents the network weight of size c×d, d represents the size of the output dimension; b∈R d Represents the bias vector of the MLP layer, with a size of d;

[0029] The function of the Sigmoid function is to limit the value of the model output to between [0, 1]. Its mathematical expression is as follows:

[0030]

[0031] Denormalize the predicted crashworthiness response values:

[0032] x′=x×x max

[0033] Where x represents the input non-denormalized crashworthiness response, x′ is the denormalized crashworthiness response value, and x max Indicates the maximum value of the crashworthiness response value in the structural collision analysis data set;

[0034] The expression of the GELU function is as follows:

[0035]

[0036] Furthermore, the point cloud slice merging PPM module specifically includes:

[0037] For point cloud features with an input size of [C, N], the input features are first divided into N / d slices (Patch) of size d, where d is the downsampling factor of the point cloud features, and then the features in the slices are merged. At this time, the size of the features becomes [C×d, N / d]; finally, a fully connected layer (MLP) is used to adjust the dimensions of the merged slice features and map them to the preset dimensional space. The final output feature size is [C′, N / d], where C′ is the preset feature dimensional space size; the PPM module processing function is as follows:

[0038] PPM(X)=MLP(X′)

[0039]

[0040] Among them, X represents the point cloud feature with an input size of [C, N], mod(c, C) represents the modulus of c to C, It means rounding down, and d is the downsampling ratio.

[0041] Furthermore, the point cloud pixel rearrangement operation PPS module specifically includes:

[0042] For point cloud features with an input size of [C, N], a fully connected layer MLP is first used to adjust the feature dimension so that it is mapped to a dimensional space of size C′×d, where C′ represents the feature dimension size of the desired final output and d is the upsampling factor of the point cloud feature; finally, the adjusted features are rearranged to expand the extra dimensions and finally obtain features of size [N×d, C′]; the PPS module processing function is as follows:

[0043]

[0044] X′=MLP(X)

[0045] Where X∈R C×N Represents the feature tensor of the input before resizing, X′∈R C′×Nd It represents the feature tensor after feature dimension adjustment, and mod(n, d) represents the modulo of n to d.

[0046] Furthermore, the point cloud sliding window attention PSAB module calculates the self-attention weight once before and after moving the window. For each feature tensor with an input size of [C, N], PSAB divides it according to the size of the window, and then calculates its local self-attention weight. Then, the window is moved with half the window size as the step size, and the self-attention weight within the window is calculated again.

[0047] The PSAB module is a two-stage self-attention feature extraction module, including layer normalization (LayerNorm, LN), feed forward neural network (Feed Forward Network, FFN), window-based multi-head self-attention module (Window-based Multi-head Self-Attention, W-MSA) and sliding window-based multi-head self-attention module (Shified Window-based Multi-head Self-Attention, SW-MSA);

[0048] The LN layer is used to remap the input features to a normal distribution. The mathematical expression of the LN layer is as follows:

[0049]

[0050] Where x represents the input feature, E[x] is the mean of x, Var(x) is the variance of x, ∈ is a constant, and γ and β are learnable affine transformation parameters;

[0051] First, in W-MSA, the features are divided according to the pre-set window size, and the self-attention weight in each window is calculated respectively. Then, the window is moved, and the self-attention weight is calculated again in the moved window, that is, SW-MSA.

[0052] The calculation process of self-attention weight of W-MSA and SW-MSA is the same. For a feature tensor of size in a window, the MSA module performs a linear transformation on the feature through an MLP layer to obtain the embedded feature vectors Query, Key and Value, whose sizes are consistent with the input feature tensor. Then the similarity between Query and Key is calculated, and the output of the similarity is adjusted by an adjustment coefficient Scale and Softmax function to obtain the final attention weight (AttentionWeight); the obtained attention weight is multiplied by Value to obtain the final output; the mathematical expression of this process is as follows:

[0053]

[0054]

[0055] Where Q, K, and V correspond to Query, Key, and Value, respectively; B is a learnable bias whose size is consistent with AttentionWeight; d is the feature dimension size of Query and Key; exp(·) refers to an exponential function with base e;

[0056] The FFN layer consists of two MLP layers and a GELU nonlinear activation function. Its function is to increase the nonlinear fitting ability of the entire PSAB. FFN maps the features to a higher dimensional space, adjusts them using a nonlinear activation function, and then maps them back to the original dimensional space. The mathematical expression of FFN is as follows:

[0057] FFN(x)=MLP(GeLU(MLP(x)))

[0058] Where x∈R c×n For the input features, the first MLP layer maps x to a dimensional space of size 4×c, and the second MLP layer maps the features from the dimensional space of 4×c back to a dimensional space of size c; the entire calculation process of PSAB is described by the following formula:

[0059]

[0060]

[0061]

[0062]

[0063] Among them, x l-1 Represents the input feature tensor, x l+1 Represents the feature tensor of the final output of the entire PSAB.

[0064] Further, the training process of the model in step S3 is as follows:

[0065] (1) First, initialize the model, determine the maximum number of training epochs, initialize the learning rate lr, the model saving interval n, and the batch training size bs; divide the samples of the data set into many subsets of the same size, each subset contains bs samples, and use the data of one subset each time to load the model for forward and backward propagation to calculate the gradient, so as to complete the update iteration of the network model parameters;

[0066] The dataset is divided into a training set and a validation set in a ratio of 80% and 20%, and the validation set does not participate in model training. The training has a total of epoch rounds. In each round of training, the model is trained with the training set, and the validation set is used to verify the model effect. In this process, the model parameters with the best performance on the validation set loss are saved, and the network parameters are saved every n training rounds until the final trained network model is obtained after cyclic epoch rounds of training.

[0067] In each round of training, each small batch of data is preprocessed and then input into the proxy model to obtain the prediction results. After calculating the loss function, the Consine algorithm is used to adjust the learning rate, and the parameters of the proxy model are updated based on the AdamW algorithm. This process is repeated continuously for all batches.

[0068] Furthermore, a loss function is constructed based on the mean square error loss (MSE Loss), which is expressed as follows:

[0069] L n (r p , r t , p p , p t )=λMSE(r p , r t )+(1-λ)MSE(p p , p t )

[0070]

[0071] Among them, r p , r t denote the predicted and true crashworthiness response vectors, respectively, and p p , p t They represent the predicted and real structural collision deformation point cloud sequences respectively, λ is the weighting coefficient, and n represents the size of the input x, y.

[0072] Furthermore, a learning rate adjustment strategy based on cosine annealing is used in training, which is expressed as follows:

[0073]

[0074] ε=α 3 lr, 0<α<1

[0075] Where lr represents the initial learning rate during the entire training process, lr′ represents the adjusted learning rate, k represents the kth round of training, epoch represents the total number of training rounds, ε represents the preset lower limit of the learning rate, and α is the adjustment coefficient;

[0076] The AdamW network parameter optimization method is used for parameter optimization. For a data set containing l data samples, a batch training size of bs, and a training round of epoch, the training task needs to be performed. Network parameter update. When the network parameters are updated for the tth time, the parameter update process is as follows:

[0077]

[0078] mt =β1m t-1 +(1-β1)g t

[0079]

[0080]

[0081]

[0082]

[0083] Among them, θ t represents the network parameters after the tth update; g t represents the gradient of the loss function to the network parameters at the tth iteration; η t represents the learning rate at the tth iteration; mt is the first-order momentum, m t＝0 =0;v t represents the second-order momentum, v t＝0 =0; represents the first-order moment of momentum after correction of the offset; Represents the second-order momentum after correcting the offset; α is a constant; ∈ is a constant; β1 and β2 are adjustment coefficients; λ is the network weight attenuation coefficient.

[0084] The beneficial effect of the present invention is that the present invention constructs an agent model that can simultaneously complete the prediction of structural crashworthiness response and structural collision deformation representation, thereby accelerating the structural crashworthiness optimization process.

[0085] Other advantages, objectives and features of the present invention will be described in the following description and will be apparent to those skilled in the art to some extent, or those skilled in the art may be taught from the practice of the present invention. The objectives and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] In order to make the purpose, technical solution and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration:

[0087] Figure 1 This is a flow chart of the method for predicting collision plastic deformation point cloud sequence by multi-modal fusion according to the present invention;

[0088] Figure 2 It is a schematic diagram of a multi-modal information proxy model for the structure crashworthiness response and collision deformation point cloud representation of the present invention;

[0089] Figure 3 Schematic diagram of the sliding window self-attention mechanism in PSAB;

[0090] Figure 4 The PSAB architecture and multi-head self-attention calculation flow chart, (a) is the overall architecture diagram of PSAB, (b) is the schematic diagram of the Multi-head SelfAttention calculation flow chart;

[0091] Figure 5 Schematic diagram of the agent model training process. DETAILED DESCRIPTION

[0092] like Figure 1 As shown, the present invention provides a multi-modal fusion collision plastic deformation point cloud sequence prediction method. After constructing the collision deformation point cloud prediction layer, the predicted crashworthiness response and structural collision deformation point cloud sequence can be obtained. The regression loss of the predicted value and the true value is calculated, and this process is continuously repeated until the regression loss converges, and then the model training process can be considered to be completed.

[0093] In the present invention, a multimodal fusion prediction agent model for the collision deformation process of thin-walled structures is constructed, which mainly includes several structures: point cloud slicing merging (Point Patch Merging, PPM), point cloud sliding window attention module (Point Swin Attention Block, PSAB), point cloud pixel shuffle operation (Point Pixel Shuffle, PPS) and fully connected layer (Multilayer Perceptron, MLP), and these structures constitute the feature embedding and fusion layer, feature extraction layer and decision layer.

[0094] like Figure 2 As shown in the figure, the entire proxy model takes the structural design variables and the initial finite element mesh point cloud of the structure generated by them as input. They are first sent to the feature embedding and fusion layer. The point cloud and the design variables belong to data of different modes. Therefore, they will be uniformly mapped to the same feature dimension space in the feature embedding module.

[0095] The feature embedding module consists of a PPM and an MLP module. Data of different modalities will be uniformly mapped to a dimensional space of size C. The point cloud will be mapped from [3, N] to [C, N / 4], where N represents the number of points in the point cloud; and the structural design variable will be mapped from [1, dim] to [1, C]. Here, the downsampling factor of PPM is 4, that is, d=4, the input dimension of the MLP layer is the structural design variable dimension dim, and the output dimension is the embedded feature vector size C. In this embodiment, C=32. Wherein, dim represents the original dimension size of the structural design variable. However, since there are still differences between the embedded point cloud features and the variable features, the feature fusion module amplifies the channels of the variable features through a feature amplification operation and then aligns them with the point cloud features. The feature amplification layer consists of a one-dimensional convolutional layer, whose output channel number is equal to N / 4, the convolution kernel size is 3, the convolution step size is 1, and the zero padding size is 1. The amplified features are added to the point cloud features to achieve feature fusion and output to the feature extraction layer. At the same time, the design variable embedding features before amplification will be retained and sent to the crashworthiness response prediction layer.

[0096] The fused features obtained through the feature embedding and fusion layers will be sent to the feature extraction layer, which is mainly composed of an encoder module, a decoder module, and a Bottle Neck module. The encoder is composed of three PSAB and two PPM modules stacked alternately, and the decoder is symmetrical with the encoder, consisting of three PSAB and two PPS modules stacked alternately. The Bottle Neck module is composed of "PPM-PSAB-PPS" in series. The feature extraction layer of such an encoder-decoder configuration refers to the multi-scale pyramid structure model in the field of image vision, so that the entire model can extract the spatial information features of the entire point cloud from local to global.

[0097] Among them, the PSAB module is a point cloud adaptive attention feature extraction module inspired by Swin Transformer. The size of the fused features processed by PSAB remains unchanged; the PPM module can be regarded as a downsampling layer in the encoder module, which is used to merge the fused features from the previous layer within a certain range, and thus obtain the features of a higher layer as the high-level abstraction of the current features. The PPS module can be regarded as the inverse operation of PPM, that is, the upsampling layer, which has the opposite function of PPM, that is, to restore the abstract high-level features to the corresponding underlying feature dimensions.

[0098] The features output by the first PSAB module of the encoder module will not only be input to the next PPM module, but also be fused with the features output by the last PSAB module of the decoder module as the output of the entire feature extraction layer; the features output by each PPM module in the encoder will not only be passed to the next layer, but also be fused with the features output by the previous PSAB module corresponding to its symmetrical PPS in the decoder module as the input of the symmetrical PPS module. Such connections across modules are generally called skip connections, which were first proposed in residual networks (Residual Networks, ResNets). Its original purpose was to reduce the gradient vanishing and network degradation problems caused by the increase in network depth, speed up network training, etc., and some related studies have found that skip connections can also fuse the local high-frequency information features extracted by the underlying network with the global features extracted by the high-level network, further improving the prediction performance of the network model.

[0099] After entering the feature extraction layer, the fused features will extract features at multiple scales in the encoder module. The size of the fused features will gradually change from [C, N / 4] to [4C, N / 16]. At the same time, the fused features of each scale will be retained. The features that have passed the encoding layer will first undergo the last downsampling feature extraction, attention feature extraction, and feature recovery process in the Bottle Neck module, which will be completed by the PPM, PSAB, and PPS modules respectively. The size of the features that have passed the Bottle Neck remains unchanged. In the decoder module, the size of the fused features will gradually recover from [4C, N / 16] to [C, N / 4], and in this process, it will be fused with the features of the same scale retained from the encoder to ensure that the gradient information can be better transmitted to the underlying network during the back propagation process. The features that have passed the feature extraction layer will be sent to the crashworthiness response prediction layer and the structural collision deformation point cloud prediction layer of the decision layer respectively.

[0100] In the structural collision deformation point cloud prediction layer, the feature will first pass through a PPS module, and the feature size will be restored from [C, N / 4] to [t×3, N], where t refers to the number of time steps contained in the predicted structural collision deformation point cloud sequence. In this embodiment, t=5 and N=1024. Finally, a PSAB module is used to fine-tune the final prediction output to further improve the accuracy of the prediction. Among them, the default parameters of the PPS module are set to: the upsampling ratio is 4, that is, d=4; the output feature dimension size is N=1024; the default parameters of the PSAB module are set to: the number of heads of the multi-head self-attention mechanism is 8, and the self-attention weight window size is 4.

[0101] In the crashworthiness response prediction layer, the feature size of the input layer is consistent with the feature size of the feature embedding and fusion layer output, both of which are [C, N / 4], where C is the size of the embedded feature vector and N is the number of nodes contained in the point cloud. The input features will first pass through a feature pooling layer (Feature Pooling), where the features will be compressed into a feature vector, and the size will be uniformly changed from [C, N / 4] to [1, 256]; and the initial embedded variable features obtained from the feature embedding module will pass through a feature alignment module (Feature Alignment) because the dimension is C, and it will be mapped from a feature vector of size C to a feature vector space of size 256. Afterwards, these two parts of the features are fused so that the underlying embedded feature information can also better participate in the prediction of the crashworthiness response. The fused features will eventually obtain the final predicted crashworthiness response through a unified crashworthiness response prediction layer. The feature alignment module includes two MLP layers and the corresponding GELU activation function; the feature pooling layer includes a one-dimensional convolution layer and an adaptive average pooling layer (AdaptiveAvgPool 1D) as well as transpose and flatten operations. The adaptive average pooling layer adaptively adjusts the size of the pooling kernel according to the size of the input feature tensor so that the dimension of the output feature meets the preset size; the crashworthiness response prediction layer includes two MLP layers, a GELU activation function, and the final Sigmoid activation function.

[0102] The multilayer perceptron is a classic fully connected network (Multilayer Perceptron, MLP). The MLP layers included in all the model methods proposed in this embodiment refer to single-layer fully connected layers and do not include activation functions. In the classic multilayer perceptron, its input is usually a one-dimensional vector. Different from this, the input of the MLP layer included in the model proposed in this embodiment is in the form of a tensor, and its weight is only effective for the last dimension of the tensor. Therefore, the MLP layers included in the model can be expressed using formula (1):

[0103] MLP(X)=Xw+b (1)

[0104] Where X∈R […，c] Represents the input feature tensor with the last dimension size c, w∈R c×d represents the network weight of size c×d, d represents the size of the output dimension; b∈R d Represents the bias vector of the MLP layer, with a size of d.

[0105] The function of the Sigmoid function is to limit the value of the model output to between [0, 1]. Its mathematical expression is shown in formula (2).

[0106]

[0107] Since the predicted crashworthiness response value is between [0, 1], it needs to be denormalized. The denormalization is calculated using formula (3):

[0108] x′=x×x max (3)

[0109] Where x represents the input non-denormalized crashworthiness response, x′ is the denormalized crashworthiness response value, and x max It represents the maximum value of the crashworthiness response value in the structural collision analysis data set. The expression of the GELU function is shown in formula (4).

[0110]

[0111] The merging operation (Point Patch Merging, PPM) applicable to point cloud data can be regarded as a one-dimensional patch merging operation. For point cloud features with an input size of [C, N], the input features are first divided into N / d slices (Patch) of size d, where d is the downsampling factor of the point cloud features, and then the features in the slices are merged. At this time, the size of the features becomes [C×d, N / d]; finally, a fully connected layer (MLP) is used to adjust the dimensions of the merged slice features and map them to a preset dimensional space. The final output feature size is [C′, N / d], where C′ is the preset feature dimensional space size. In this embodiment, unless otherwise specified, all PPM default parameter settings are as follows: the downsampling factor is 2, that is, d=2; the output feature dimension size defaults to twice the input feature dimension size, that is, C′=2C. This process can be expressed by equations (5) and (6).

[0112] PPM(X)=MLP(X′) (5)

[0113]

[0114] Among them, X represents the point cloud feature with an input size of [C, N], mod(c, C) represents the modulus of c to C, It means rounding down, and d is the downsampling ratio.

[0115] The present invention proposes a point cloud pixel reordering operation (Point Pixel Shuffle, PPS) suitable for point cloud data. PPS can be regarded as a one-dimensional form of Pixel Shuffle operation. For point cloud features with an input size of [C, N], a fully connected layer (MLP) is first used to adjust the feature dimension so that it is mapped to a dimensional space of size C′×d, where C′ represents the feature dimension size of the desired final output and d is the upsampling factor of the point cloud feature; finally, the adjusted features are rearranged to expand the extra dimensions, and finally a feature of size [N×d, C′] is obtained. This process can be expressed by equations (7) and (8):

[0116]

[0117] X′=MLP(X) (8)

[0118] Where X∈R C×N Represents the feature tensor of the input before resizing, X′∈R C′×Nd represents the feature tensor after feature dimension adjustment, and mod(n, d) represents the modulus of n over d. In this embodiment, unless otherwise specified, all PPS default parameter settings are as follows: the upsampling ratio is 2, that is, d=2, and the output feature dimension size defaults to twice the input feature dimension size, that is, C′=2C.

[0119] The Shifted Window based Multi-head Self-Attention (SW-MSA) mechanism originated from Swin Transformer and was originally used to deal with image recognition problems. It mainly solves the defect of high computational complexity of the ViT method based on the standard Transformer architecture when calculating global attention. Swin Transformer splits the feature map, approximates global attention with the attention combination of local windows, and realizes the exchange of information between windows by sliding windows, which effectively reduces the complexity of calculating self-attention weights while retaining as much global information as possible.

[0120] The present invention proposes a sliding window self-attention module (Point SwinAttention Block, PSAB) suitable for point cloud data. Figure 3As shown in the figure, each PSAB will calculate the self-attention weight once before and after moving the window. For each feature tensor with an input size of [C, N], PSAB will divide it according to the size of the window, and then calculate its local self-attention weight. Then, it will move the window with half the window size as the step size, and calculate the self-attention weight within the window again. This can enable information exchange between adjacent windows and reduce computational complexity.

[0121] The overall structure of PSAB is as follows Figure 4 As shown in (a) in the figure, it is a two-stage self-attention feature extraction module, which mainly includes layer normalization (LN), feed forward neural network (FFN), window-based multi-head self-attention module (W-MSA) and shifted window-based multi-head self-attention module (SW-MSA).

[0122] The role of the LN layer is to remap the input features to a normal distribution. This approach allows the feature values ​​to fall within a relatively sensitive interval of the model activation function, so that during the training process of the model, the gradient in the gradient return increases, avoiding the gradient vanishing phenomenon caused by the network being too deep. The mathematical expression of the LN layer is shown in formula (9):

[0123]

[0124] Where x represents the input feature, E[x] is the mean of x, Var(x) is the variance of x, and ∈ is a constant, usually 1×10 -5 , γ and β are learnable affine transformation parameters.

[0125] The W-MSA module and the SW-MSA module correspond to Figure 3 The self-attention calculation module before and after the moving window in the PSAB. In the entire PSAB calculation process, the features are first divided according to the pre-set window size in W-MSA, and the self-attention weights in each window are calculated respectively. Then the window is moved, and the self-attention weights are calculated again in the moved window, that is, SW-MSA. The moving window makes the coverage of the two attention weights overlap, and the calculated local attention weights can approximate the global attention weights while maintaining a low computational complexity.

[0126] The self-attention weight calculation process of W-MSA and SW-MSA is the same, such as Figure 4As shown in (b), for a feature tensor of size in a window, the MSA module will perform a linear transformation of the features through an MLP layer to obtain embedded feature vectors Query, Key and Value, whose sizes are consistent with the input feature tensor. In this embodiment, a new dimension is added to facilitate matrix operations. Then the similarity between Query and Key is calculated, and the output of the similarity is adjusted by an adjustment coefficient Scale and a Softmax function to obtain the final attention weight. The obtained attention weight is multiplied by Value to obtain the final output. The mathematical expression of this process is shown in formula (10):

[0127]

[0128]

[0129] Among them, Q, K, V correspond to Query, Key and Value respectively; B is a learnable bias, whose size is consistent with AttentionWeight; d is the feature dimension size of Query and Key. Figure 4 (b) refers to the first dimension of the feature tensor size, i.e., C; exp(·) refers to the exponential function with e as the base.

[0130] The FFN layer is mainly composed of two MLP layers and a GELU nonlinear activation function. Its function is to increase the nonlinear fitting ability of the entire PSAB. Although the self-attention mechanism is applied in W-MSA and SW-MSA, it is still a linear transformation in fact. Therefore, the FFN layer is designed in the standard Transformer architecture to improve the overall nonlinear fitting ability of the model. FFN maps the features to a higher dimensional space first, adjusts them using a nonlinear activation function, and then maps them back to the original dimensional space. The mathematical expression of FFN is shown in formula (12).

[0131] FFN(x)=MLP(GeLU(MLP(x))) (12)

[0132] Where x∈R c×n For the input features, the first MLP layer maps x to a dimensional space of size 4×c, and the second MLP layer maps the features from the dimensional space of 4×c back to a dimensional space of size c. In summary, Figure 4 The entire calculation process of PSAB in (a) can be described by equations (13) to (16).

[0133]

[0134]

[0135]

[0136]

[0137] Among them, x l-1 Represents the input feature tensor, x l+1 Indicates the feature tensor of the final output of the entire PSAB. In this embodiment, unless otherwise specified, the default parameter settings of all PSABs are as follows: The number of heads of the multi-head self-attention mechanism is 8, that is, when calculating a certain window Figure 4 When performing self-attention as shown in (b) in the figure, a linear transformation is simultaneously performed to generate 8 sets of the same Q, K, and V for calculating the self-attention weights. Finally, a fully connected layer is used to map these 8 sets of self-attention weights back to the dimensional space of the original feature vector. In addition, the window size for calculating the self-attention weights defaults to 4.

[0138] The training process of the entire model is as follows Figure 5 As shown, first, we need to initialize the model, determine the maximum number of training epochs, initialize the learning rate lr, the model saving interval n, and the batch training size bs.

[0139] Since the data set is too large to be loaded into the system memory at once, this embodiment uses a small batch training method, that is, the samples of the data set are divided into many subsets of the same size, each subset contains bs samples, and the data of a subset is used to load the model for forward and backward propagation to calculate the gradient to complete the update iteration of the network model parameters. This method is called a small batch training method. Compared with loading the entire data set for training at one time, the small batch training method can make the model have stronger generalization performance.

[0140] Secondly, the data set is divided into a training set and a validation set in a ratio of 80% and 20%, and the validation set does not participate in model training. There are a total of epochs in the training. In each round of training, the model is trained with the training set, and the validation set is used to verify the model effect. In this process, the model parameters with the best performance on the validation set loss are saved, and the network parameters are saved every n training rounds until the final trained network model is obtained after the epoch training.

[0141] In each round of training, each small batch of data is preprocessed and then input into the proxy model to obtain the prediction results. After calculating the loss function, the Consine algorithm is used to adjust the learning rate, and the parameters of the proxy model are updated based on the AdamW algorithm. This process is repeated continuously for all batches.

[0142] In this embodiment, the structural design variables of the thin-walled circular tube include length L, diameter D, and wall thickness T; crashworthiness response indicators, maximum peak force PCF and energy absorption SEA; the initial finite element mesh node point cloud of the thin-walled circular tube, the size of which is [3, 1024]; the structural collision deformation point cloud sequence corresponding to the initial point cloud of the thin-walled circular tube at 5 randomly sampled moments, the size of which is [5, 3, 1024].

[0143] These data are all original data obtained from finite element simulation. Since the main action range of most activation functions in neural networks is between -1 and 1, the original data needs to be appropriately scaled to ensure that the output of the neural network can cover the distribution range of the data. At the same time, due to the different dimensions and sizes between different data variables, the phenomenon of "large numbers eating small numbers" is prone to occur. Therefore, it is also necessary to eliminate the impact of dimensions on the data by scaling the data. This process is called normalization. It should be noted that the distribution of data samples should remain consistent before and after normalization. The following normalization method is used for this data set:

[0144] For the design variable X = [L, D, T], normalize it according to formula (17):

[0145]

[0146] Among them, X′ represents the normalized design variable vector, X max and X min Respectively represent the upper and lower limits of the range of the design variables in the data set. Since the axial direction of the thin-walled circular tube is in the Y-axis direction, in order to maintain the difference between the samples before and after normalization, the initial point cloud and collision deformation point cloud sequence corresponding to the design variables are uniformly normalized using equations (18) to (20):

[0147] Nod y ′=Nod y / L max (18)

[0148] Nod x,z′ =Nod x,z / r (19)

[0149]

[0150] Among them, Nod y and Nod y ′ refers to the Y-axis coordinates of the point cloud before and after normalization, Nod x,z and Nod x,z ′ refers to the X-axis and Z-axis coordinates of the point cloud before and after normalization, L max Indicates the maximum design length, D maxRepresents the maximum design thin-walled circular tube diameter. For the crashworthiness response index Y = [PCF, SEA], the normalization operation is performed as shown in formula (21):

[0151] Y′=Y / Y max (twenty one)

[0152] Among them, Y and Y′ represent the crashworthiness response index before and after normalization, respectively. max Represents the maximum value of the crashworthiness response in the data set.

[0153] The loss function is a function used to measure the difference between the prediction output by the model and the true label. It determines the final learning effect of the deep neural network and plays a vital role in the learning process of the deep neural network.

[0154] The constructed multimodal information proxy model of structural crashworthiness response and collision deformation point cloud representation has the output of the structure's crashworthiness response during the collision process and the corresponding collision deformation point cloud sequence. Therefore, the closer the output result is to the real result, the better the fitting effect of the proxy model. To this end, a loss function for training the proxy model proposed in this chapter is constructed based on the mean square error loss (MSE Loss), and its mathematical expressions are shown in Equations (22) and (23):

[0155] L n (r p , r t , p p , p t )=λMSE(r p , r t )+(1-λ)MSE(p p , p t ) (twenty two)

[0156]

[0157] Among them, r p , r t denote the predicted and true crashworthiness response vectors, respectively, and p p , p t They represent the predicted and real structural collision deformation point cloud sequences respectively, λ is a weighting coefficient, and its value is between 0 and 1. In this embodiment, it is 0.5. In the formula, n represents the size of input x, y.

[0158] In the process of deep neural network training, in the early stage of training, due to the large prediction error of the network, the gradient of the loss function calculated is also large. At this time, if the learning rate is too small, the training time will be prolonged. In the later stage of training, the network gradually converges, and accordingly, the prediction error also becomes smaller. At this time, the gradient of the calculated loss function becomes smaller. At this time, if a large learning rate is maintained, the model may fall into a local optimal solution or overfitting. If a fixed learning rate is maintained throughout the training process, the above problems will occur. Therefore, in the past, when training deep neural networks, it is often necessary to adjust the learning rate parameters multiple times to find a suitable learning rate for model training, which often requires a lot of effort in this process.

[0159] In order to solve the above problems, and speed up the network training process while avoiding the problem of falling into local optimum or overfitting, many strategies for adaptively adjusting the learning rate have been proposed, such as learning rate adjustment strategies according to equal training intervals, learning rate adjustment strategies according to linear decay, and learning rate adjustment strategies according to exponential decay. In this embodiment, a learning rate adjustment strategy based on cosine annealing (Cosine) is used. This method uses a cosine function to adjust the learning rate. The learning rate decreases slowly in the early and late stages of training, and accelerates in the middle of training, which is consistent with the trend of gradients becoming smaller during deep neural network training. It is a widely used learning rate adjustment strategy. Its mathematical expression is shown in formula (24).

[0160]

[0161] ε=α 3 lr, 0<α<1 (25)

[0162] Wherein, lr represents the initial learning rate in the entire training process, lr′ represents the adjusted learning rate, k represents the kth round of training, epoch represents the total number of training rounds, ε represents the preset lower limit of the learning rate, which is determined by formula (25) in this embodiment, and α is the adjustment coefficient, which is 0.1 in this embodiment.

[0163] The AdamW optimization algorithm solves the problem of poor generalization performance caused by insufficient regularization in Adam by changing the gradient calculation strategy and the weight update method. In the AdamW algorithm, for a data set containing l data samples, a batch training size of bs, and a training round of epoch, it is necessary to perform Network parameter update. When the network parameters are updated for the tth time, the parameter update process is as shown in equation (26) to equation (31).

[0164]

[0165] mt =β1m t-1 +(1-β1)g t (27)

[0166]

[0167]

[0168]

[0169]

[0170] Among them, θ t represents the network parameters after the tth update; g t represents the gradient of the loss function to the network parameters at the tth iteration; η t represents the learning rate at the tth iteration, which is determined by formula (24); m t is the first-order angular momentum, m t＝0 =0;v t represents the second-order momentum, v t＝0 =0; represents the first-order moment of momentum after correction of the offset; Represents the second-order momentum after the offset correction; α is a constant, usually 0.001; ∈ is a constant, usually used for numerical stability, the default value is 1×10 -8 ; β1 and β2 are adjustment coefficients, and their values ​​are between the interval (0, 1). In this embodiment, they are 0.9 and 0.999 respectively; λ is the network weight attenuation coefficient, and the default value is 0.01.

[0171] The evaluation function is a function used to measure the final effect and performance of the model after training. It is an important indicator for quantifying the error between the model prediction result and the true label. The requirements and application scenarios of different tasks determine the selection of the evaluation function. In this embodiment, the proxy model actually performs a regression task, and its performance depends on the final fitting error. In this embodiment, the relative error (Relative Error, RE) and the mean relative error (MeanRelative Error, MRE) are used to evaluate the errors of the crashworthiness response and the structural collision deformation point cloud sequence output by the proxy model, respectively. The mathematical expressions are shown in Equations (32) and (33):

[0172]

[0173]

[0174] Among them, δ r and δ pThey represent the relative error of the crashworthiness response and the average relative error of the collision deformation point cloud sequence, respectively. n represents the input p p , p t It is not difficult to see that the smaller the error, the better the fitting effect of the proxy model.

[0175] In order to comprehensively test the effect of the proposed multimodal fusion prediction model for the collision deformation process of thin-walled structures, three experiments were designed based on the thin-walled circular tube collision analysis dataset to test the fitting effect of the model under different conditions. The prediction effect of the model in the structural collision analysis dataset of thin-walled circular tubes, the prediction effect in the case of missing some point cloud data, and the prediction effect after replacing some key modules of the model were tested. The specific settings and purposes of each experiment are introduced below. To ensure that the training effect is not affected by the training hyperparameters, the hyperparameters of the training model are kept consistent in all experiments.

[0176] Experimental design for comprehensive performance verification of the proxy model: After preprocessing the data set according to the input and output of the model, the entire model is trained according to the training process, and the prediction effect of the proxy model is evaluated based on the mentioned evaluation indicators. The purpose of this experiment is mainly to comprehensively evaluate the model's prediction effect on the crashworthiness response of thin-walled circular tubes and the deformation point cloud sequence during the collision process. It will also serve as a reference for the next two experiments to explore the factors that affect the model's prediction effect in terms of data and model structure.

[0177] Experimental design for proxy model performance test under incomplete data conditions: This experiment will study the prediction effect of the proxy model under the condition of missing partial structural collision deformation data. In this experiment, two data missing conditions are set, namely the condition of missing Y-axis collision deformation point cloud data and the condition of missing Z-axis collision deformation point cloud data, to study the influence of each dimension of point cloud data on the final prediction result. Since the constructed thin-walled circular tube receives axial impact along the negative direction of the Y-axis, the missing X-axis data and the missing Z-axis data are the same for the point cloud data, so only one of the two missing conditions needs to be studied. The experiment is arranged as follows: for a certain data missing condition, the missing point cloud axial coordinate data is set to 0, keeping the input and output dimensions of the proxy model unchanged, and when evaluating the model effect, the error of the missing dimensional data is not calculated.

[0178] Experimental design for testing the importance of key modules of the proxy model: This experiment will study the impact of the proxy model's prediction accuracy when some key modules are missing, in order to explore the contribution of each module to the final recognition effect. This experiment will mainly study the impact of three modules, Point Patch Merging (PPM), Point Swin Attention Block (PSAB), and Point Pixel Shuffle (PPS), on the model's prediction accuracy.

[0179] The specific arrangement of the experiment is as follows: the proxy model is trained in the absence of the three modules PPM, PSAB, and PPS. In order to ensure that the changes in input and output due to the missing modules will not affect other modules, a simple network layer will be used to replace it in this experiment to ensure that the input and output forms before and after the replacement are consistent with the original model.

[0180] For the PPM module, a set of serially connected one-dimensional convolutional layers and one-dimensional maximum pooling layers (Max Pooling) are used to replace it, and its corresponding hyperparameter settings are adjusted to ensure that the input and output dimension sizes of the module remain consistent before and after the replacement.

[0181] For the PSAB module, the window-based multi-head self-attention module (Window-basedMulti-head Self-Attention, W-MSA) and the shifted window-based multi-head self-attention module (ShiftedWindow-basedMulti-headSelf-Attention, SW-MSA) were mainly replaced. The simple network model used for replacement is a one-dimensional convolution layer with a convolution kernel size of 1 and a step size of 1. The number of convolution kernel channels remains consistent with the feature dimension size that needs to be output in the original model.

[0182] For the PPS module, this embodiment uses a set of serially connected one-dimensional convolutional layers and one-dimensional upsampling layers (UpSample) to replace it, and adjusts its corresponding hyperparameter settings to ensure that the input and output dimension sizes of the module remain consistent before and after the replacement.

[0183] This experiment will replace the corresponding modules separately in three cases, train the proxy model separately, and detect the changes in prediction accuracy caused by module replacement.

[0184] After 4000 rounds of iterative training on the training set, the model's loss function value tends to converge. By observing the Loss curve, it can be found that as the training progresses, the model's loss on the training set and the validation set shows a fluctuating downward trend. As the number of training iterations increases, the rate of loss decreases as well. This is due to the use of the Cosine learning rate adjustment method, i.e., formula (24). The model's loss change on the training set fluctuates less, while it fluctuates more on the validation set. This phenomenon is common in deep learning because the adjustment of model parameters is based on the gradient of the model parameter's loss function on the training set. The validation set does not participate in the optimization of model parameters. Therefore, under normal circumstances, the loss fluctuation of the validation set will be greater than that of the training set. It is for this reason that the model's loss function on the validation set is often greater than its loss function on the training set.

[0185] This experimental example calculates the relative error between the prediction results of the proxy model on the training set and the validation set and the finite element simulation results, and compares its average calculation time with the time consumed by the finite element simulation. From the results, it can be seen that the relative error between the crashworthiness response and the structural collision deformation point cloud sequence output by the proxy model and the results of the finite element simulation solution can be controlled within 5%, whether in the training set or the validation set, indicating that the proposed proxy model has high accuracy and can accurately fit the numerical relationship between the structural design variables and the crashworthiness response and the structural collision deformation point cloud sequence. The solution speed of the proxy model is significantly better than that of the finite element solution, and its solution speed is about 8169 times that of the finite element solution speed.

[0186] From the experimental results, it can be seen that for the structural collision deformation point cloud sequence, the model has a good prediction effect and has a high degree of similarity with the real data obtained by finite element simulation. The predicted value fits the actual value very well, indicating that the proposed proxy model can fit the numerical relationship between the structural design variables and the required crashworthiness response and the structural collision deformation point cloud sequence with high accuracy.

[0187] To sum up the above results, for the structural optimization problem, a large number of solution calculations are required to obtain the optimal structural design variables. The proposed proxy model can quickly solve the required crashworthiness response values ​​and the corresponding structural collision deformation point cloud sequence while ensuring sufficient accuracy. It is suitable for the structural crashworthiness optimization problem under the constraints of the collision deformation mode to be solved.

[0188] The relative error of the predicted value of the proxy model on the thin-walled circular tube structure collision analysis data set when the Y-axis and Z-axis point cloud coordinate data are missing respectively. It can be found that in the case of missing Y-axis data, the relative error of the proxy model prediction result has slightly decreased; while in the case of missing Z-axis data, the relative error of the proxy model prediction result has slightly increased. Combined with the case of no missing data, it can be found that among the various errors, the relative error of the prediction for the Y-axis data is the highest among all items, that is, the prediction error of the Y-axis data from the collision axis contributes the largest relative error of the prediction. A reasonable explanation is that the lack of Y-axis data reduces the difficulty of the proxy model fitting task, thereby reducing its prediction error relatively. When the Z-axis data is missing, the proportion of Y-axis data increases relatively, thereby increasing the difficulty of fitting the proxy model, and the corresponding prediction error also increases relatively.

[0189] Based on the above results, it can be considered that a certain degree of data loss will not affect the prediction accuracy of the proxy model, indicating that the proxy model proposed in this chapter has a certain robustness. In addition, it can be seen that the prediction of the axial collision displacement of the thin-walled structure by the proxy model is the main source of error, which provides a reference for the improvement of the proxy model.

[0190] The relative error of the proxy model's prediction of each response without each key module shows that whether replacing the PPM, PPS or PSAB module, the impact on the relative error of the prediction of SEA is the smallest, while the impact on the point cloud coordinates in the collision direction, that is, the Y-axis coordinates of the collision deformation point cloud sequence, is the largest. This shows that reducing the complexity of the proxy model will increase the relative error of its output prediction. The degree of influence of the replacement of the key module on the prediction items of the proxy model is from large to small, namely, the Y-axis coordinates of the collision deformation point cloud sequence, the X-axis coordinates of the collision deformation point cloud sequence, the Z-axis coordinates of the collision deformation point cloud sequence, PCF and SEA.

[0191] Comprehensively observing the relative errors of various predictions of the proxy model, it can be found that after replacing the PSAB module, the impact on the relative errors of various predictions is the greatest, indicating that this module has the greatest impact on the prediction performance of the entire proxy model; followed by the PPS module and the PPM module. After replacing these two modules respectively, the changes in the relative errors of the proxy model's predictions are not much different, indicating that these two modules have similar effects on the prediction performance of the entire proxy model, but the impact of replacing the PPS module is relatively larger. This may be because the location of the PPS module is relatively closer to the output end of the proxy model, and therefore, it has a greater impact on the output of the proxy model.

[0192] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A multi-modal fusion collision plastic deformation point cloud sequence prediction method, characterized by: The following steps are involved: S1: constructing a multimodal fusion prediction agent model for a structural collision deformation process; the multimodal fusion prediction agent model for a structural collision deformation process comprises a multimodal feature embedding and fusion layer, a feature extraction layer, a crashworthiness response prediction layer, and a structural collision deformation point cloud prediction layer; The multimodal feature embedding and fusion layer includes a feature embedding module and a feature fusion module; The feature embedding module is composed of a point cloud slice merging PPM module and a fully connected layer MLP module, which is used to uniformly map data of different modes to a dimensional space of size C; the feature fusion module is used to amplify the channel of variable features through feature amplification operation and align them with point cloud features; the amplified features are added to the point cloud features to achieve feature fusion and output to the feature extraction layer; the design variable embedding features before amplification are sent to the crashworthiness response prediction layer; The feature extraction layer is composed of an encoder module, a decoder module and a Bottle Neck module, which are used to extract the spatial information features of the entire point cloud from local to global, and send them to the crashworthiness response prediction layer and the structural collision deformation point cloud prediction layer respectively; The crashworthiness response prediction layer is used to predict crashworthiness response indicators, including energy absorption SEA response and maximum peak force PCF response; The structure collision deformation point cloud prediction layer is used to predict the structure collision deformation point cloud; S2: Build and preprocess the dataset, and divide it into training set and validation set; S3: using the training set to train the multimodal fusion prediction agent model of the structural collision deformation process, and using the validation set to validate the model; S4: Input the geometric parameters of the structure, material design parameters and the initial finite element mesh node point cloud of the structure determined by the geometric parameters, and predict the crashworthiness response data and structural collision deformation point cloud data; In the feature extraction layer, the encoder module is composed of three point cloud sliding window attention PSAB modules and two PPM modules stacked alternately, the decoder module is symmetrical with the encoder module, and is composed of three PSAB and two point cloud pixel rearrangement operation PPS modules stacked alternately, and the Bottle Neck module is composed of "PPM-PSAB-PPS" in series; the PSAB module is regarded as a point cloud adaptive attention feature extraction module, and the size of the fused features processed by PSAB remains unchanged; the PPM module is regarded as a downsampling layer in the encoder module, and its function is to merge the fused features transmitted from the upper layer within a certain range of adjacent layers, and thereby obtain a higher-level feature as a high-level abstraction of the current feature; the PPS module is regarded as an upsampling layer, and its function is to restore the abstract high-level features to the corresponding underlying feature dimensions; The features output by the first PSAB module of the encoder module are not only input to the next PPM module, but also fused with the features output by the last PSAB module of the decoder module as the output of the entire feature extraction layer; The features output by each PPM module in the encoder are not only passed to the next layer, but also fused with the features output by the PSAB module in the previous layer corresponding to its symmetric PPS in the decoder module as the input of the symmetric PPS module; After entering the feature extraction layer, the fused features extract features at multiple scales in the encoder module. The size of the fused features gradually changes from [C, N / 4] to [4C, N / 16]. At the same time, the fused features of each scale are retained. The features passing through the encoding layer are first subjected to the last downsampling feature extraction, attention feature extraction, and feature recovery process in the Bottle Neck module, which are completed by the PPM, PSAB, and PPS modules respectively. The size of the features passing through the Bottle Neck remains unchanged. In the decoder module, the size of the fused features is gradually restored from [4C, N / 16] to [C, N / 4], and in this process is fused with the features of the same scale retained from the encoder module; In the structural collision deformation point cloud prediction layer, the features first pass through a PPS module, and the feature size is restored from [C, N / 4] to [t×3, / V], where t represents the number of time steps contained in the predicted structural collision deformation point cloud sequence. Finally, a PSAB module is used to fine-tune the final prediction output; In the crashworthiness response prediction layer, the feature size of the input layer is consistent with the feature size of the feature embedding and fusion layer output, both of which are [C, N / 4], where C is the size of the embedded feature vector and N is the number of nodes contained in the point cloud; the input feature is first compressed into a feature vector through a feature pooling layer, and the size is uniformly changed from [C, N / 4] to [1, 256]; the initial embedded variable feature obtained from the feature embedding module is mapped from a feature vector of size C to a feature vector space of size 256 through a feature alignment module; Afterwards, the two features are fused, and the fused features are finally used to obtain the final predicted crashworthiness response through a unified crashworthiness response prediction layer; The feature alignment module consists of two MLP layers and the corresponding GELU activation function; The feature pooling layer includes a one-dimensional convolution layer and an adaptive average pooling layer as well as transposition and flattening operations. The adaptive average pooling layer adaptively adjusts the size of the pooling kernel according to the size of the input feature tensor so that the dimension of the output feature meets the preset size. The MLP layer is a single-layer fully connected layer and does not contain an activation function. Its input is in the form of a tensor, and its weight is only effective for the last dimension of the tensor, which can be expressed as follows: MLP(X)=Xw+b Where X∈R […，c] Represents the input feature tensor with the last dimension size c, w∈R c×d represents the network weight of size c×d, d represents the size of the output dimension; b∈R d Represents the bias vector of the MLP layer, with a size of d; The function of the Sigmoid function is to limit the value of the model output to between [0, 1]. Its mathematical expression is as follows: Denormalize the predicted crashworthiness response values: x′=x×x max Where x represents the input non-denormalized crashworthiness response, x′ is the denormalized crashworthiness response value, and x max Indicates the maximum value of the crashworthiness response value in the structural collision analysis data set; The expression of the GELU function is as follows: The point cloud slice merging PPM module specifically includes: For point cloud features with an input size of [C, / V], the input features are first divided into N / d slices of size d, where d is the downsampling factor of the point cloud features, and then the features in the slices are merged. At this time, the size of the feature becomes [C×dN / d]; finally, a fully connected layer is used to adjust the dimension of the merged slice features and map them to the preset dimensional space. The final output feature size is [C′, N / d], where C′ is the preset feature dimensional space size; the PPM module processing function is as follows: PPM(X)=MLP(X′) Among them, X represents the point cloud feature with an input size of [C, N], mod(c, C) represents the modulus of c to C, It means rounding down, and d is the downsampling ratio; The point cloud pixel rearrangement operation PPS module specifically includes: For point cloud features with an input size of [C, N], a fully connected layer MLP is first used to adjust the feature dimension so that it is mapped to a dimensional space of size C′×d, where C′ represents the feature dimension size of the desired final output and d is the upsampling factor of the point cloud feature; finally, the adjusted features are rearranged to expand the extra dimensions and finally obtain features of size [N×d, C′]; the PPS module processing function is as follows: X′=MLP(X) Where X∈R C×N Represents the feature tensor of the input before resizing, X′∈R C′×Nd represents the feature tensor after feature dimension adjustment, mod(n, d) represents the modulus of n to d; The point cloud sliding window attention PSAB module calculates the self-attention weight once before and after moving the window. For each feature tensor with an input size of [C, N], PSAB divides it according to the size of the window, and then calculates its local self-attention weight. After that, it moves the window with a step size of half the window size, and calculates the self-attention weight within the window again. The PSAB module is a two-stage self-attention feature extraction module, including layer normalization LN, feed-forward neural network FFN, window-based multi-head self-attention module W-MSA and sliding window-based multi-head self-attention module SW-MSA; The LN layer is used to remap the input features to a normal distribution. The mathematical expression of the LN layer is as follows: Where x represents the input feature, E[x] is the mean of x, Var(x) is the variance of x, ∈ is a constant, and γ and β are learnable affine transformation parameters; First, in W-MSA, the features are divided according to the pre-set window size, and the self-attention weight in each window is calculated respectively. Then, the window is moved, and the self-attention weight is calculated again in the moved window, that is, SW-MSA. The calculation process of self-attention weight of W-MSA and SW-MSA is the same. For a feature tensor of size in a window, the MSA module performs a linear transformation on the feature through an MLP layer to obtain the embedded feature vectors Query, Key and Value, whose sizes are consistent with the input feature tensor. Then the similarity between Query and Key is calculated, and the output of the similarity is adjusted by an adjustment coefficient Scale and Softmax function to obtain the final attention weight; the obtained attention weight is multiplied by Value to obtain the final output; the mathematical expression of this process is as follows: Among them, Q, K, V correspond to Query, Key and Value respectively; B is a learnable bias, whose size is consistent with AttentionWeight; d is the feature dimension size of Query and Key, which refers to the first dimension of the feature tensor size, that is, C in Figure 4-4-b; exp(·) refers to an exponential function with base e; The FFN layer consists of two MLP layers and a GELU nonlinear activation function. Its function is to increase the nonlinear fitting ability of the entire PSAB. FFN maps the features to a higher dimensional space, adjusts them using a nonlinear activation function, and then maps them back to the original dimensional space. The mathematical expression of FFN is as follows: FFN(x)=MLP(GeLU(MLP(x))) Where x∈R c×n For the input features, the first MLP layer maps x to a dimensional space of size 4×c, and the second MLP layer maps the features from the dimensional space of 4×c back to a dimensional space of size c; the entire calculation process of PSAB is described by the following formula: Among them, x l-1 Represents the input feature tensor, x l+1 Represents the feature tensor of the final output of the entire PSAB; The training process of the model in step S3 is as follows: (1) First, initialize the model, determine the maximum number of training epochs, initialize the learning rate lr, the model saving interval n, and the batch training size bs; divide the samples of the data set into many subsets of the same size, each subset contains bs samples, and use the data of one subset each time to load the model for forward and backward propagation to calculate the gradient, so as to complete the update iteration of the network model parameters; The dataset is divided into a training set and a validation set in a ratio of 80% and 20%, and the validation set does not participate in model training. The training has a total of epoch rounds. In each round of training, the model is trained with the training set, and the validation set is used to verify the model effect. In this process, the model parameters with the best performance on the validation set loss are saved, and the network parameters are saved every n training rounds until the final trained network model is obtained after cyclic epoch rounds of training. In each round of training, each small batch of data is preprocessed and then input into the proxy model to obtain the prediction result. After calculating the loss function, the Consine algorithm is used to adjust the learning rate, and the parameters of the proxy model are updated based on the AdamW algorithm. This process is repeated over all batches. The loss function is constructed based on the mean square error loss, which is expressed as follows: L n (r p ,r t ,p p ,p t )=λMSE(r p ,r t )+(1-λ)MSE(p p ,p t ) Among them, r p , r t denote the predicted and true crashworthiness response vectors, respectively, and p p , p t Respectively represent the predicted and real structural collision deformation point cloud sequences, λ is the weighting coefficient, and n represents the size of input x, y; The learning rate adjustment strategy based on cosine annealing is used in training, which is expressed as follows: e=a 3 lr,0<α<1 Where lr represents the initial learning rate during the entire training process, lr′ represents the adjusted learning rate, k represents the kth round of training, epoch represents the total number of training rounds, ε represents the preset lower limit of the learning rate, and α is the adjustment coefficient; Use the AdamW network parameter optimization method to optimize parameters. For a data set containing / data samples, a batch training size of bs, and a training round of epoch, you need to perform Network parameter update; at the tth network parameter update, there is the following parameter update process: m t =β1m t-1 +(1-β1)g t Among them, θ t represents the network parameters after the tth update; g t represents the gradient of the loss function to the network parameters at the tth iteration; η t represents the learning rate at the tth iteration; m t is the first-order angular momentum, m t＝0 =0;v t represents the second-order angular momentum, ν t＝0 =0; represents the first-order moment of momentum after correction of the offset; Represents the second-order momentum after correcting the offset; α is a constant; ∈ is a constant; β1 and β2 are adjustment coefficients; λ is the network weight attenuation coefficient.

Images