A vehicle pedestrian head injury prediction method and system based on transfer learning

Abstract

The application provides a vehicle pedestrian head injury prediction method and system based on transfer learning, comprising the following steps: S1: obtaining historical source domain samples, constructing an injury prediction point cloud neural network model, and pre-training the injury prediction point cloud neural network model using historical source domain data; S2: obtaining multiple target domain samples and simulation real HIC label values, and constructing a four-layer composite transfer total loss function; S3: according to a sampling strategy, multiple target domain samples are randomly taken out according to the simulation real HIC label values to form sampling samples; S4: predicting all sampling samples to generate HIC prediction values; S5: updating the hyperparameters of the injury prediction point cloud neural network model to obtain a performance prediction model; and S6: obtaining real measurement samples of a new target vehicle model, and using the performance prediction model to generate HIC prediction values. The application solves the problems of the prior art, such as serious dependence of evaluation of vehicle head protection performance on a large amount of labeled data, low generalization ability and low prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of automotive passive safety technology, and in particular to a method and system for predicting pedestrian head injuries in vehicles based on transfer learning. Background Technology

[0002] In the development of pedestrian protection systems for automobiles, the Hip Conformity (HIC) value is a key indicator for evaluating a vehicle's performance in protecting pedestrians' heads. Existing technologies for obtaining HIC values ​​mainly fall into three categories: real-world physical testing, finite element simulation, and traditional machine learning. In real-world physical testing, according to regulations, a pedestrian head impactor is used to strike the front end of the vehicle at a specified speed and angle. Acceleration-time curves during the collision are collected by an acceleration sensor installed within the impactor. The formula calculates the HIC value at the impact point. This method offers the highest accuracy and is the only standard method for regulatory certification. However, it is extremely costly, has a long testing cycle, and each grid point can only be tested once, making large-scale replication impossible. It is typically used only in the final certification stage and cannot be used extensively in the early stages of research and development.

[0003] In the finite element simulation method, after establishing a finite element model of the vehicle's front-end structure, the material, contact, and boundary conditions are defined, and a pedestrian head impactor model and regulatory conditions are set. Finally, the solution is submitted to perform nonlinear dynamic collision simulation, and the composite acceleration-time curve of the head module is extracted to calculate the HIC value. This method has high accuracy, good repeatability, and lower cost than physical testing. However, single-condition calculation takes several hours, and a complete evaluation of a vehicle model requires hundreds to thousands of conditions, with a total calculation time of several weeks. At the same time, the simulation accuracy is overly dependent on the accuracy of the finite element model and lacks versatility.

[0004] Traditional machine learning methods extract geometric and material features from finite element models to construct regression models, train them using a large amount of simulation data, and then input features from new operating conditions to predict HIC values. This method is fast and low-cost. However, it heavily relies on a large amount of labeled simulation data of the target vehicle model, cannot utilize the vast amount of accumulated historical data, and the model is only applicable to the specific structural configuration used during training, failing to generalize to new configurations after changes in structural parameters. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for predicting vehicle-pedestrian head injuries based on transfer learning. This method solves the problems in existing technologies, such as the heavy reliance on a large amount of labeled data for assessing the vehicle's head protection performance for pedestrians, the inability to utilize a large amount of historical data, and the low generalization ability and accuracy of the prediction model.

[0006] According to an embodiment of the present invention, a method for predicting head injuries to pedestrians in vehicles based on transfer learning includes: S1: Obtain historical source domain samples, construct the Xingbao head injury prediction point cloud neural network model, and pre-train the Xingbao head injury prediction point cloud neural network model using historical source domain data. The Xingbao head injury prediction point cloud neural network model includes a sampling module, an attention module, an encoding module, and a prediction module. S2: Obtain multiple target domain samples and corresponding simulated real HIC label values ​​of the pedestrian impact area at the front of the new target vehicle model, and construct a four-layer composite transfer total loss function; S3: Initialize the sampling strategy. According to the sampling strategy, the sampling module divides the target domain samples into multiple sampling spaces based on the simulated real HIC label values, and randomly selects multiple target domain samples from each sampling space to form the sampling samples. S4: Input all sampled samples into the prediction module to generate the corresponding model inference HIC prediction value; S5: Based on the model inference HIC prediction value and the total loss function of the four-layer composite transfer, update the hyperparameters of the Xingbao head injury prediction point cloud neural network model using the progressive unfreezing training method and repeat steps S2-S4 until the total loss function of the four-layer composite transfer is minimized, and obtain the performance prediction model. S6: Obtain measured samples of the front pedestrian impact areas of all vehicles of the new target model, and generate the model inference HIC prediction value corresponding to the front pedestrian impact area of ​​each vehicle based on the measured samples using the performance prediction model.

[0007] Preferably, each set of target domain samples includes a global point cloud feature vector and material structure features; Methods for inputting all target domain samples into the Xingbao head injury prediction point cloud neural network model to generate corresponding model inference HIC prediction values ​​include: The attention module is used to extract cross-attention features from the global point cloud feature vector and the material structure features, respectively, to obtain the corresponding geometry-guided fusion features and material-guided fusion features. The geometry-guided fusion features and the material-guided fusion features are concatenated to obtain concatenated features. Then, the encoding module is used to reduce the dimensionality of the concatenated features, and the prediction module is used to make predictions to obtain the model inference HIC prediction value.

[0008] Preferably, the sampling strategy includes a range division and corresponding sampling weights, a contrast loss weight, a positive sample pair determination threshold, and a temperature coefficient. According to the defined range, all target domain samples are divided into multiple sampling spaces based on the simulated real HIC label values. Then, target domain samples are extracted proportionally according to the sampling weights corresponding to each sampling space to form a sampling sample.

[0009] Preferably, the total migration loss function of the four-layer composite layer is as follows:

[0010] Interval-adaptive weighted contrast loss: Loss in maintaining ranking: Conditional domain adversarial loss: KDE-weighted smoothed L1 supervised loss: in, , , and All are weighting coefficients, and B is the total number of samples in a single batch. For positive sample pair masks, For sample pairs L2 similarity, The cosine similarity of the sample pairs. To achieve the desired feature similarity, For conditional eigenvectors, These are the domain labels for each parameter in the conditional feature vector. For the model inference HIC prediction value of the i-th sample, Let be the simulated real HIC tag value of the i-th sample.

[0011] Preferably, the method for calculating the interval-based adaptive weighted contrast loss includes: All sampled samples are combined one-to-one to obtain multiple sample pairs. For each sample pair, the minimum value of the positive sample pair judgment threshold corresponding to the sampling space of the two sampled samples in the sample pair is used as the dynamic judgment threshold. The positive sample pair mask is calculated based on the model inference HIC prediction value and dynamic decision threshold of the two sampled samples in the sample pair. Based on the HIC values ​​of all samples in each hierarchical space, the QualityAttention network is used to generate the quality attention weights for each sample in that hierarchical space. Then, the sample adaptive weights for each sample are calculated based on the quality attention weights and the contrastive loss weights. Calculate the cosine similarity between the two samples in each sample pair, and then calculate the L2 similarity of the sample pair based on the cosine similarity and the temperature coefficient corresponding to the sampling space to which the two samples belong. The interval adaptive weighted contrast loss is calculated based on the L2 similarity of all sample pairs, the positive sample pair mask, and the sample adaptive weights.

[0012] Preferably, the sampled sample further includes corresponding real domain labels; The method for calculating the conditional domain adversarial loss includes: All model inference HIC prediction values ​​of the sampled samples are mapped to the dimension of the sampled samples to obtain the embedding vector. Then the embedding vector and all target domain samples are concatenated to obtain the conditional feature vector. Based on the preset gradient reversal strength, the conditional feature vector is flipped, and then the domain discriminator is used to determine the sample domain of the conditional feature vector to obtain the prediction domain probability corresponding to each sample. The conditional adversarial loss is calculated based on the predicted domain probabilities and corresponding ground truth domain labels of all sampled samples.

[0013] Preferably, the method of repeating steps S2-S4 using the progressive thawing training method includes: Set a first thawing boundary, a second thawing boundary, and a third thawing boundary. When the number of iterations is less than the first thawing boundary, freeze the sampling module, the attention module, and the prediction module. Adjust the hyperparameters of the encoding module according to the four-layer composite transfer total loss function, and then repeat steps S2-S4. When the number of iterations is greater than or equal to the first thawing boundary and less than the second thawing boundary, freeze the sampling module and the prediction module, adjust the hyperparameters of the attention module and the encoding module according to the total loss function of the four-layer composite transfer, synchronously increase the gradient reversal strength of the gradient reversal layer, and repeat steps S2-S4 after completing the parameter update. When the number of iterations is greater than or equal to the second unfreezing boundary and less than the third unfreezing boundary, the sampling module is frozen, the hyperparameters of the attention module, the encoding module and the prediction module are adjusted according to the four-layer composite transfer total loss function, and the weight coefficients corresponding to the interval adaptive weighted contrast loss are adjusted to the preset maximum value. Then, steps S2-S4 are repeated. When the number of iterations is greater than or equal to the third unfreezing boundary, adjust the hyperparameters of the sampling module, attention module, encoding module and prediction module with the goal of minimizing the total loss function of the four-layer composite transfer function, and then repeat steps S2-S4.

[0014] Preferably, the method for calculating the KDE-weighted smoothed L1 supervised loss includes: The KDE kernel density estimate is calculated based on the simulated real HIC label values ​​of the sampled samples, and then the kernel density estimate weights are calculated based on the KDE kernel density estimate. The smoothed L1 score is calculated based on the model inference HIC predicted value and the simulated real HIC label value of the sampled samples. Then, the KDE weighted smoothed L1 supervised loss is obtained by weighting the kernel density estimation weights of all sampled samples and the corresponding smoothed L1 scores.

[0015] On the other hand, according to embodiments of the present invention, a vehicle-pedestrian head injury prediction system based on transfer learning is also provided. This system uses the aforementioned vehicle-pedestrian head injury prediction method based on transfer learning, and includes: The data acquisition module is used to acquire multiple target domain samples and corresponding simulated real HIC label values ​​of the pedestrian impact area at the front of the vehicle. The model building module is used to build the Xingbao head injury prediction point cloud neural network model and a four-layer composite transfer total loss function, and to pre-train the Xingbao head injury prediction point cloud neural network model. The model training module is used to predict target domain samples using the Xingbao head injury prediction point cloud neural network model, generate corresponding model inference HIC prediction values, and then repeat steps S2-S4 using the progressive unfreezing training method based on the model inference HIC prediction values ​​until the total loss function of the four-layer composite transfer is minimized, thus obtaining the performance prediction model.

[0016] Preferably, the point cloud neural network model for predicting head injuries includes a sampling module, an attention module, an encoding module, and a prediction module; The sampling module is used to divide the target domain samples into multiple sampling spaces according to the simulated real HIC label values, and randomly select multiple target domain samples from each sampling space to form a sampling sample. The prediction module is used to input all sampled samples into the prediction module to generate the corresponding model inference HIC prediction value.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention uses historical source domain samples to pre-train the point cloud neural network model for predicting head injury in HIC, making full use of historical data. Then, a small number of target domain samples of the newly acquired target vehicle are subjected to full-impact mesh second-level batch inference prediction to achieve rapid data detection and improve engineering efficiency. At the same time, the model uses the built-in KDE weighted smoothing L1 supervised loss to alleviate the problem of long-tail data distribution in HIC engineering. No manual additional weighting of samples is required, and the model prediction stability is significantly improved in small sample scenarios.

[0018] (2) The present invention performs hierarchical sampling of the target domain samples to form sampling samples to train the point cloud neural network model for predicting head injury, thus completely solving the engineering defects of traditional random sampling models that focus on low-risk areas and whose prediction results cannot support rectification decisions. Attached Figure Description

[0019] Figure 1 This is a diagram illustrating the method for constructing the prediction model according to an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] like Figure 1 As shown in the figure, this invention proposes a method for predicting head injuries to pedestrians and vehicles based on transfer learning, including: S1: Obtain historical source domain samples, construct the Xingbao head injury prediction point cloud neural network model, and pre-train the Xingbao head injury prediction point cloud neural network model using historical source domain data. The Xingbao head injury prediction point cloud neural network model includes a sampling module, an attention module, an encoding module, and a prediction module. Historical source domain samples were constructed using complete simulation data of pedestrian impact areas at the front of various other vehicle models. Each historical source domain sample contains a global point cloud feature vector, material structure features, and a corresponding real domain label (source domain). The global point cloud feature vector is a 1024-dimensional global geometric feature vector output after three-layer downsampling and abstraction of 8192 three-dimensional points in the front collision area of ​​the vehicle. It contains all geometric information affecting the head collision response, such as the curvature of the hood surface, the direction of the reinforcing ribs, and the A-pillar tilt angle. The material structure features are composed of four types of engineering parameters: impact coordinates, thickness sequence, age grouping, and material properties.

[0022] Subsequently, a point cloud neural network model for predicting head injuries was constructed, and historical source domain samples were used to pre-train the model. The model includes a sampling module, an attention module, an encoding module, and a prediction module.

[0023] S2: Obtain multiple target domain samples and corresponding simulated real HIC label values ​​of the pedestrian impact area at the front of the new target vehicle model, and construct a four-layer composite transfer total loss function; Similar to the historical source domain samples, multiple target domain samples from the front pedestrian impact area of ​​the current vehicle model are collected, along with their corresponding simulated real HIC label values ​​and real domain labels (target domain). Simultaneously, a four-layer composite transfer total loss function is constructed for the pedestrian head injury prediction point cloud neural network model, as follows: The total loss function for the four-layer composite migration is as follows:

[0024] in, , , and All are weighting coefficients. The KDE-weighted smoothed L1 supervised loss is used to constrain the absolute prediction accuracy of HIC. To maintain ranking, a loss is used to constrain the clustering of features among samples within the same risk interval. This is a conditional domain adversarial loss used to constrain the alignment of the source and target domains with damage level features. An interval-adaptive weighted contrast loss is used to constrain the clustering of features among samples in the same risk interval.

[0025] S2: Initialize the sampling strategy. According to the sampling strategy, the sampling module divides the target domain samples into multiple sampling spaces based on the simulated real HIC label values, and randomly selects multiple target domain samples from each sampling space to form the sampling samples. Based on NCAP's rating criteria for HIC values, this invention designs a sampling strategy as shown in Table 1.

[0026] Table 1: Sampling Strategy

[0027] The sampling module divides the target domain samples into 5 sampling spaces based on the true HIC value corresponding to the target domain samples and the HIC value range of the sampling strategy. Each sampling space includes multiple target domain samples. Multiple target domain samples are randomly selected from the corresponding sampling space according to the sampling weight to ensure that the samples in each sampling space participate in the model training in a balanced manner. If the number of samples in a certain interval is insufficient, sampling with replacement is used to supplement them. The extracted target domain samples together constitute the sampling sample set.

[0028] The core purpose of stratified sampling is to solve this long-tailed distribution problem, enabling the model to fully learn the characteristics of high-risk samples at the tail. To this end, differentiated sampling weights are assigned to each interval. For the extremely high-risk interval at the tail of the distribution, although the absolute number of samples is small, by assigning appropriate sampling weights, it is ensured that tail samples appear in every training batch, thereby significantly increasing the participation frequency of tail samples.

[0029] S3: Input all sampled samples into the prediction module to generate the corresponding model inference HIC prediction value; The attention module contains two sets of symmetrical CrossAttention layers, enabling bidirectional interaction between 3D geometric features and material stiffness features, thus overcoming the shortcomings of general networks in separating geometric and material information. First-path cross-attention: using global point cloud feature vectors For the query vector Query, material structure features As keys and values, output geometry-guided fusion features. ; Second-path cross-attention: Independently encode and reduce the dimensionality of impact coordinates, thickness sequence, age grouping, and material properties. Impact coordinates: 3D coordinates are mapped to 64D features via a linear layer. ; Thickness sequence: Global thickness sequence compressed into 32-dimensional features ; Age grouping: embedding layer mapping 8-dimensional features ; Material properties: After encoding 15-dimensional material parameters point by point, global average pooling is used to output 256-dimensional material features. ; By combining the four types of structural features, the material structural features are obtained:

[0030] Based on material structural characteristics For Query, global point cloud feature vector As Key / Value, output material-oriented fusion features .

[0031] By splicing together the geometry-guided fusion feature and the material-guided fusion feature, a spliced ​​feature is obtained. Used for subsequent loss calculation; equipped with two-layer multi-head self-attention modules to enhance the internal features of structured parameters and strengthen the ability to represent high-stiffness hard point features such as hinges.

[0032] Then, a feature encoding module is concatenated (two layers of BatchNorm and Dropout regularized fully connected layers for dimensionality reduction), finally outputting the single-dimensional model inference HIC prediction value. .

[0033] S4: Based on the model inference HIC prediction value and the total loss function of the four-layer composite transfer, update the hyperparameters of the Xingbao head injury prediction point cloud neural network model using the progressive unfreezing training method and repeat steps S2-S4 until the total loss function of the four-layer composite transfer is minimized, thus obtaining the performance prediction model.

[0034] The four-layer composite transfer total loss function includes KDE weighted smoothing L1 supervised loss, interval adaptive weighted contrast loss, ranking preservation loss, and conditional domain adversarial loss. The above four losses are calculated based on the model inference HIC predicted value and the simulated real HIC label value.

[0035] (1) KDE-weighted smoothed L1 supervised loss The KDE-weighted smoothed L1 supervised loss in the four-layer composite migration total loss function is adapted to the long-tailed distribution characteristics of engineering HIC. This loss has the characteristics of high sample density and low weight in low-risk areas, and low sample density and amplified weight in high-risk areas. The specific loss function is as follows:

[0036] Where B represents the total number of samples in a single batch. The weights for kernel density estimation of the i-th sample are determined by the pedestrian head injury index value, with higher values ​​resulting in greater weights. This encourages the model to focus more on high-risk hard-point structural areas such as A-pillars and windshield wiper mounting points during training. Let i be the true HIC value of the i-th sample. This represents the smoothed L1 loss function, which takes a squared form when the prediction error is small and a linear form when the error is large.

[0037] Kernel density estimation weights:

[0038]

[0039] in, The kernel density estimate is represented by the probability density function y constructed based on all true HIC values ​​in the current target domain samples. The value at the position is larger in the low-risk area with dense samples (HIC<650), and the weight is reduced; while it is smaller in the high-risk area with sparse samples (HIC≥1350), and the weight is amplified. This represents the total number of samples in the target domain. This represents the kernel density bandwidth parameter, which controls the smoothness of the density estimation. If it is too large, it will be unable to distinguish the differences between high and low risk intervals; if it is too small, the estimation will be unstable. Indicates the first The HIC label value of each sampled sample is used to construct the true HIC value of all samples for the overall density distribution; This represents the kernel function form of the Gaussian kernel, used to calculate the contribution of each sample to the density estimate.

[0040] Smoothing L1 piecewise function:

[0041] in Indicates the prediction error. The deviation between the model-predicted HIC value and the actual HIC value is: a positive error indicates an overestimation of the risk of head injury, and a negative error indicates an underestimation of the risk. The adaptive smoothing threshold is dynamically determined by the true HIC mean of all samples in the current sample (usually 5% to 10% of the mean). It determines the boundary by which the loss function switches from the squared form to the linear form, allowing the loss function to adopt a more lenient error tolerance for high-risk areas (where the HIC mean is higher).

[0042] (2) Interval adaptive weighted contrast loss To address the limitations of traditional uniform threshold contrast loss in distinguishing between high- and low-risk samples and insufficient constraint on high-stiffness hard point features, this invention constructs an interval-based adaptive weighted contrast loss:

[0043] This invention combines all sampled samples one-to-one to obtain multiple sample pairs. For each sample pair, a positive sample pair determination threshold is determined based on the positive sample pair determination threshold of the sampling space in which the two sampled samples are located. The minimum value of the positive sample pair determination threshold is used as the dynamic determination threshold.

[0044] in, It indicates that the i-th sample is in the first position. The threshold for determining positive sample pairs in a sampling space.

[0045] Calculate the positive sample pair mask based on the model inference HIC prediction value and dynamic decision threshold of the two sampled samples in the sample pair:

[0046] in, Let be the true HIC value of the i-th sample in the sample pair.

[0047] A value of 1 indicates a sample pair ( A sample pair is identified as a positive sample pair (the two samples have similar degrees of head injury), and a value of 0 indicates a negative sample pair. This mask is used in the contrast loss to control which sample pairs need to be brought closer in terms of feature distance.

[0048] Sample adaptive attention weights:

[0049] in This represents the quality attention weight. The quality attention network outputs sample-level fine-grained weights. The higher the model's inference HIC prediction value, the closer this weight is to 1. For example, a sample with HIC=1600 receives a larger attention weight than a sample with HIC=1400. In contrastive loss, high-risk hard point features are preferentially aggregated. No. The contrast loss weights for each sampling space reflect the differences in the basic attention given to samples in different sampling spaces in the contrast loss.

[0050] The QualityAttention network is a lightweight sample reweighting subnetwork embedded in the contrastive loss calculation branch. It is used to adaptively generate attention weights for each sample that monotonically increase with the model's inference HIC prediction value. The network structure is a four-layer fully connected network: the input layer receives the normalized model inference HIC prediction value (scalar, 1-dimensional), extracts features through two fully connected layers (1→64→32, ReLU activation), and the third layer adopts a parallel fusion structure of main path and side path skip connections (32→16, element-wise addition) to retain the absolute value information of the predicted HIC. The output layer outputs weights after being fully connected (16→1) and activated by Sigmoid. The weights output by this network are multiplied by the basic contrastive loss weights of the risk interval to which the sample belongs to obtain the final weighting coefficient of the sample in the contrastive loss, so that the sample with a high model inference HIC prediction value receives a larger gradient contribution.

[0051] Different temperature coefficients are configured for each sampling space, with even smaller temperature coefficients configured for extremely high-risk areas. If we amplify the sample similarity discriminative power, then the similarity after L2 normalization is as follows:

[0052] in Indicates sample pairs ( , The feature similarity is the cosine similarity between two sampled samples in a 1024-dimensional feature space. The higher the similarity, the closer the head injury response characteristics of the vehicle structures corresponding to the two sampled samples are. Indicates the sampled sample and sampling samples The global point cloud feature vector; This means taking the smaller of the temperature coefficients of the sampling spaces to which the two samples belong. The temperature coefficient controls the sharpness of the similarity transformation into a probability distribution. A smaller temperature coefficient is used in the high-risk interval to amplify the subtle similarity differences between high-risk samples.

[0053] If there are no matching positive sample pairs in a batch, the three samples with the smallest predicted HIC difference are automatically selected to force the construction of sample pairs to prevent loss and failure, ensure that the features of high-risk hard points of the same type are close to each other, and modify the local structure so that the predicted change trend is consistent with engineering understanding.

[0054] (3) Loss of ranking maintenance The feature space distance exhibits monotonicity consistent with the actual HIC value difference, eliminating the defect of "slight structural changes leading to drastic jumps in HIC predictions" and ensuring the effectiveness of horizontal comparisons of lightweight multi-scheme approaches. Calculate the true HIC normalized difference of the sampled samples:

[0055] Target similarity:

[0056] Feature normalized similarity:

[0057] After excluding its own samples, the MSE loss is used to constrain the matching degree between the two:

[0058] in This indicates the degree of difference in the true HIC values ​​between sample pairs, normalized to the 0~1 range; This represents the maximum true HIC difference within a batch, ensuring that the loss value is unaffected by changes in the order of magnitude of HIC values ​​in the batch samples. This represents the target similarity, which is the expected feature similarity defined based on the true HIC value. The closer the HIC values ​​of two sampled samples are, the closer their target similarity is to 1. This represents the actual feature similarity, specifically the cosine similarity between the 1024-dimensional feature vectors actually output by the model. The ranking maintenance loss is a constraint that keeps the feature similarity output by the model monotonically consistent with the difference in HIC values. This ensures that after the engine hood structure is fine-tuned, the rise and fall trend of the HIC prediction value is in line with engineering expectations, and there will be no anti-physical phenomena such as "the HIC prediction decreases when the plate thickness increases".

[0059] (4) Conditional domain adversarial loss Unlike traditional global unconstrained domain alignment, this invention introduces predicted HIC values ​​as conditional information to achieve fine-grained feature matching across vehicle models within the same risk range. Feature alignment is performed only on cross-vehicle samples with the same HIC damage level, enabling accurate transfer of historical vehicle model knowledge.

[0060] Map the model inference HIC predictions to an embedding vector of the same dimension as the backbone output. Conditional features are obtained by concatenating the global point cloud feature vectors:

[0061] in, This represents a vector concatenation operation, which joins two vectors end-to-end along the feature dimension; The predicted HIC embedding vector is generated by mapping the predicted HIC scalar value to an embedding vector of the same dimension as the feature vector through the embedding layer. The embedding layer is a two-layer fully connected network (with non-linear activation). The input is a 1-dimensional HIC scalar value, and the output is a 1024-dimensional vector. The purpose of this embedding vector is to inject the severity of head injury as conditional information into the domain discriminator so that domain alignment is performed only under the same injury level.

[0062] Binary cross-entropy as domain adversarial loss:

[0063] in The domain adversarial loss value is represented by the binary cross-entropy measure of the classification error of the domain discriminator. This loss drives the backbone network to extract domain-invariant features through the gradient inversion layer, so that the structural features of the same damage level in historical models and new models are aligned in the feature space. This represents the binary cross-entropy loss function, a standard binary classification loss function used to measure the difference between the domain label predicted by the domain discriminator and the true domain label. The domain discriminator is a binary classification neural network. Its input is conditional features, and its output is a scalar between 0 and 1 (representing the probability that a sample comes from the target domain). If the discriminator cannot accurately distinguish whether a sample belongs to the source domain or the target domain, it means that the backbone network has extracted domain-invariant features. This indicates a gradient reversal layer. During forward propagation, the output is identical, and during backward propagation, the gradient is multiplied by a negative coefficient. This operation causes the domain discriminator and the backbone network to engage in an adversarial game, driving the features output by the backbone network to gradually become insensitive to domain differences. This represents the true domain label, with source domain samples taking 0 and target domain samples taking 1, used for training the supervised domain discriminator.

[0064] The joint loss is calculated based on the four-layer composite transfer total loss function. If it is not the minimum value, the hyperparameters of the Xingbao head injury prediction point cloud neural network model are adjusted according to the preset parameters. Then, steps S2-S4 are repeated to iterate the Xingbao head injury prediction point cloud neural network model. During the iteration process, this invention adopts a progressive unfreezing training method for iteration. The specific process is as follows: (1) Set multiple unfreeze boundary rounds. Generally, the first unfreeze boundary has 10 iterations, the second unfreeze boundary has 30 iterations, and the third unfreeze boundary has 100 iterations. The network parameters are gradually opened in four stages.

[0065] (2) Stage 1 (number of iterations < 10) The sampling module, attention module, and prediction module are frozen, and only the encoding module is trained. The hyperparameters of the encoding module are adjusted according to the total loss function of the four-layer composite transfer. At this time, the domain adversarial weights are extremely low, and priority is given to ensuring the accuracy of the basic grid regression of the new vehicle.

[0066] (2) Stage 2 (10 ≤ number of iterations < 30) The sampling and prediction modules are frozen, and the hyperparameters of the attention and encoding modules are adjusted according to the total loss function of the four-layer composite transfer. At the same time, the gradient reversal intensity of the gradient reversal layer is gradually increased, and coarse-grained feature alignment across vehicle models is carried out. After the parameter update is completed, the next iteration is performed.

[0067] (3) Stage 3 (30≤number of iterations<100) The frozen sampling module adjusts the hyperparameters of the attention module, encoding module, and prediction module according to the four-layer composite transfer total loss function, while adjusting the weights corresponding to the contrast loss to the preset maximum value, thus finely constraining the feature discrimination of high and low risk samples.

[0068] (4) Stage 4 (iteration count ≥ 100) Unfreeze all modules and adjust the hyperparameters of the sampling module, attention module, encoding module, and prediction module with the goal of minimizing the total loss function of the four-layer composite migration. Balance the four loss weights and loss functions, taking into account both the general collision patterns of historical models and the unique engine compartment structure characteristics of new cars.

[0069] When the total loss function of the four-layer composite transfer is minimized, the iteration stops. To avoid overfitting, a maximum number of iterations (greater than 100) can be set. When the maximum number of iterations is reached, the iteration stops when the total loss function of the four-layer composite transfer is considered to be minimized. At this point, the performance prediction model applicable to the current vehicle model can be obtained.

[0070] S6: Obtain measured samples of the front pedestrian impact area of ​​all vehicles of the new target model, and generate the model inference HIC prediction value corresponding to each vehicle front pedestrian impact area using the performance prediction model based on the measured samples. Based on the model inference HIC prediction value, improve the structure and parameters of the front pedestrian impact area of ​​vehicles with model inference HIC prediction values ​​greater than the preset value. Then repeat the above steps until the overall vehicle front protection performance meets the regulatory requirements.

[0071] On the other hand, embodiments of the present invention also provide a vehicle-pedestrian head injury prediction system based on transfer learning. This system uses the above-mentioned vehicle-pedestrian head injury prediction method based on transfer learning, including: The data acquisition module is used to acquire historical source domain samples and multiple target domain samples of the front pedestrian impact area of ​​the new target vehicle model. The model building module is used to construct the Xingbao head injury prediction point cloud neural network model and a four-layer composite transfer total loss function, and to pre-train the Xingbao head injury prediction point cloud neural network model using historical source domain data; the Xingbao head injury prediction point cloud neural network model includes a sampling module, an attention module, an encoding module and a prediction module; The sampling module is used to divide the target domain samples into multiple sampling spaces according to the simulated real HIC label values, and randomly select multiple target domain samples from each sampling space to form a sampling sample. The prediction module is used to input all sampled samples into the prediction module to generate the corresponding model inference HIC prediction value; The model training module is used to predict target domain samples using the Xingbao head injury prediction point cloud neural network model, generate corresponding model inference HIC prediction values, and then repeat steps S2-S4 using the progressive unfreezing training method based on the model inference HIC prediction values ​​until the total loss function of the four-layer composite transfer is minimized, thus obtaining the performance prediction model.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting head injuries to pedestrians and vehicles based on transfer learning, characterized in that: include: S1: Obtain historical source domain samples, construct the Xingbao head injury prediction point cloud neural network model, and pre-train the Xingbao head injury prediction point cloud neural network model using historical source domain data. The Xingbao head injury prediction point cloud neural network model includes a sampling module, an attention module, an encoding module, and a prediction module. S2: Obtain multiple target domain samples and corresponding simulated real HIC label values ​​of the pedestrian impact area at the front of the new target vehicle model, and construct a four-layer composite transfer total loss function; S3: Initialize the sampling strategy. According to the sampling strategy, the sampling module divides the target domain samples into multiple sampling spaces based on the simulated real HIC label values, and randomly selects multiple target domain samples from each sampling space to form the sampling samples. S4: Input all sampled samples into the prediction module to generate the corresponding model inference HIC prediction value; S5: Based on the model inference HIC prediction value and the total loss function of the four-layer composite transfer, update the hyperparameters of the Xingbao head injury prediction point cloud neural network model using the progressive unfreezing training method and repeat steps S2-S4 until the total loss function of the four-layer composite transfer is minimized, and obtain the performance prediction model. S6: Obtain measured samples of the front pedestrian impact areas of all vehicles of the new target model, and generate the model inference HIC prediction value corresponding to the front pedestrian impact area of ​​each vehicle based on the measured samples using the performance prediction model.

2. The method for predicting head injuries to vehicles and pedestrians based on transfer learning as described in claim 1, characterized in that: Each set of target domain samples includes global point cloud feature vectors and material structure features; Methods for inputting all target domain samples into the Xingbao head injury prediction point cloud neural network model to generate corresponding model inference HIC prediction values ​​include: The attention module is used to extract cross-attention features from the global point cloud feature vector and the material structure features, respectively, to obtain the corresponding geometry-guided fusion features and material-guided fusion features. The geometry-guided fusion features and the material-guided fusion features are concatenated to obtain concatenated features. Then, the encoding module is used to reduce the dimensionality of the concatenated features, and the prediction module is used to make predictions to obtain the model inference HIC prediction value.

3. The method for predicting head injuries to pedestrians and vehicles based on transfer learning as described in claim 1, characterized in that: The sampling strategy includes the division range and the corresponding sampling weight, the contrast loss weight, the positive sample pair determination threshold, and the temperature coefficient. According to the defined range, all target domain samples are divided into multiple sampling spaces based on the simulated real HIC label values. Then, target domain samples are extracted proportionally according to the sampling weights corresponding to each sampling space to form a sampling sample.

4. The method for predicting head injuries to vehicles and pedestrians based on transfer learning as described in claim 2, characterized in that: The total loss function for the four-layer composite migration is as follows: Interval-adaptive weighted contrast loss: Loss in maintaining ranking: Conditional domain adversarial loss: KDE-weighted smoothed L1 supervised loss: in, , , and All are dynamic weighted coefficients for migration loss, where B is the total number of samples in a single batch. For positive sample pair masks, For sample pairs L2 similarity, The cosine similarity of the sample pairs. To achieve the desired feature similarity, For conditional eigenvectors, These are the domain labels for each parameter in the conditional feature vector. For the model inference HIC prediction value of the i-th sample, Let be the simulated real HIC tag value of the i-th sample.

5. The method for predicting head injuries to vehicles and pedestrians based on transfer learning as described in claim 4, characterized in that: The calculation methods for interval adaptive weighted contrast loss include: All sampled samples are combined one-to-one to obtain multiple sample pairs. For each sample pair, the minimum value of the positive sample pair judgment threshold corresponding to the sampling space of the two sampled samples in the sample pair is used as the dynamic judgment threshold. The positive sample pair mask is calculated based on the model inference HIC prediction value and dynamic decision threshold of the two sampled samples in the sample pair. Based on the HIC values ​​of all samples in each hierarchical space, the QualityAttention network is used to generate the quality attention weights for each sample in that hierarchical space. Then, the sample adaptive weights for each sample are calculated based on the quality attention weights and the contrastive loss weights. Calculate the cosine similarity between the two samples in each sample pair, and then calculate the L2 similarity of the sample pair based on the cosine similarity and the temperature coefficient corresponding to the sampling space to which the two samples belong. The interval adaptive weighted contrast loss is calculated based on the L2 similarity of all sample pairs, the positive sample pair mask, and the sample adaptive weights.

6. The method for predicting head injuries to vehicles and pedestrians based on transfer learning as described in claim 4, characterized in that: The sampled data also includes the corresponding real domain labels; The method for calculating the conditional domain adversarial loss includes: All model inference HIC prediction values ​​of the sampled samples are mapped to the dimension of the sampled samples to obtain the embedding vector. Then the embedding vector and all target domain samples are concatenated to obtain the conditional feature vector. Based on the preset gradient reversal strength, the conditional feature vector is flipped, and then the domain discriminator is used to determine the sample domain of the conditional feature vector to obtain the prediction domain probability corresponding to each sample. The conditional adversarial loss is calculated based on the predicted domain probabilities and corresponding ground truth domain labels of all sampled samples.

7. The method for predicting head injuries to vehicles and pedestrians based on transfer learning as described in claim 1, characterized in that: The method of repeating steps S2-S4 using the progressive thawing training method includes: Set a first thawing boundary, a second thawing boundary, and a third thawing boundary. When the number of iterations is less than the first thawing boundary, freeze the sampling module, the attention module, and the prediction module. Adjust the hyperparameters of the encoding module according to the four-layer composite transfer total loss function, and then repeat steps S2-S4. When the number of iterations is greater than or equal to the first thawing boundary and less than the second thawing boundary, freeze the sampling module and the prediction module, adjust the hyperparameters of the attention module and the encoding module according to the total loss function of the four-layer composite transfer, synchronously increase the gradient reversal strength of the gradient reversal layer, and repeat steps S2-S4 after completing the parameter update. When the number of iterations is greater than or equal to the second unfreezing boundary and less than the third unfreezing boundary, the sampling module is frozen, the hyperparameters of the attention module, the encoding module and the prediction module are adjusted according to the four-layer composite transfer total loss function, and the weight coefficients corresponding to the interval adaptive weighted contrast loss are adjusted to the preset maximum value. Then, steps S2-S4 are repeated. When the number of iterations is greater than or equal to the third unfreezing boundary, adjust the hyperparameters of the sampling module, attention module, encoding module and prediction module with the goal of minimizing the total loss function of the four-layer composite transfer function, and then repeat steps S2-S4.

8. The method for predicting head injuries to vehicles and pedestrians based on transfer learning as described in claim 4, characterized in that: The calculation methods for KDE-weighted smoothed L1 supervised loss include: The KDE kernel density estimate is calculated based on the simulated real HIC label values ​​of the sampled samples, and then the kernel density estimate weights are calculated based on the KDE kernel density estimate. The smoothed L1 score is calculated based on the model inference HIC predicted value and the simulated real HIC label value of the sampled samples. Then, the KDE weighted smoothed L1 supervised loss is obtained by weighting the kernel density estimation weights of all sampled samples and the corresponding smoothed L1 scores.

9. A vehicle-pedestrian head injury prediction system based on transfer learning, characterized in that: The system uses a vehicle-pedestrian head injury prediction method based on transfer learning as described in any one of claims 1-8, comprising: The data acquisition module is used to acquire historical source domain samples and multiple target domain samples of the front pedestrian impact area of ​​the new target vehicle model. The model building module is used to build the Xingbao head injury prediction point cloud neural network model and a four-layer composite transfer total loss function, and to pre-train the Xingbao head injury prediction point cloud neural network model using historical source domain data. The model training module is used to predict target domain samples using the Xingbao head injury prediction point cloud neural network model, generate corresponding model inference HIC prediction values, and then repeat steps S2-S4 using the progressive unfreezing training method based on the model inference HIC prediction values ​​until the total loss function of the four-layer composite transfer is minimized, thus obtaining the performance prediction model.

10. The vehicle-pedestrian head injury prediction system based on transfer learning as described in claim 9, characterized in that: The Xingbao head injury prediction point cloud neural network model includes a sampling module, an attention module, an encoding module, and a prediction module; The sampling module is used to divide the target domain samples into multiple sampling spaces according to the simulated real HIC label values, and randomly select multiple target domain samples from each sampling space to form a sampling sample. The prediction module is used to input all sampled samples into the prediction module to generate the corresponding model inference HIC prediction value.

Images