A method for constructing a pedestrian-vehicle collision injury prediction model and related products

Abstract

The application discloses a construction method of a pedestrian vehicle collision injury prediction model and related products. First, the input feature samples labeled with injury labels are preprocessed and split into head and leg collision injury sample subsets according to the injury types. Further, a training set, a verification set and a calibration set required for conditional coform prediction are divided. Initial proxy models corresponding to the head and the leg are obtained through type training. The calibration set is used to complete conditional coform prediction calibration and uncertainty screening, and a high-value candidate sample set is output. Through active learning, the initial prediction model is iteratively optimized to generate two types of special collision injury prediction models. The application can rely on data-driven proxy models to realize rapid injury evaluation, and screen samples to be supplemented based on uncertainty, so that the model focuses on learning the working conditions with insufficient credibility, and the prediction ability of the model for head and leg collision injuries is ensured while the development cost is controlled.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a method for constructing a pedestrian and vehicle collision damage prediction model and related products. Background Technology

[0002] Pedestrian protection is a crucial dimension for evaluating vehicle passive safety. Authoritative safety regulations such as the China New Car Assessment Program (C-NCAP) and the European New Car Assessment Program (EuroNCAP) set clear evaluation indicators and limits for vehicle pedestrian head and leg collision protection performance. In the traditional pedestrian collision performance development process, technicians must rely on computer-aided engineering (CAE) simulation software such as LS-DYNA to complete a series of complex preprocessing tasks, including building vehicle and pedestrian impactor models, geometry cleanup, mesh generation, material parameter definition, and contact relationship settings, before submitting for simulation. This entire process is not only cumbersome and highly dependent on manual labor, but a single collision simulation calculation can take several hours to several days. The traditional CAE simulation modeling and solution development method, due to its redundant and complex process and lengthy simulation cycle, has extremely low overall iteration efficiency and is difficult to adapt to the development needs of rapid iterative optimization of vehicle pedestrian protection performance.

[0003] Therefore, how to reduce the manual and time costs of pedestrian collision damage performance assessment and achieve efficient and accurate prediction and rapid iterative optimization of pedestrian collision damage conditions is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a method for constructing a pedestrian-vehicle collision damage prediction model and related products, which can reduce the manual and time costs of pedestrian collision damage performance assessment and achieve efficient and accurate prediction and rapid iterative optimization of pedestrian collision damage conditions.

[0005] The embodiments of this application disclose the following technical solutions: A method for constructing a pedestrian-vehicle collision damage prediction model, the method comprising: Obtain input feature samples labeled with damage tags corresponding to the collision damage types; the collision damage types include head collision damage and leg collision damage; The input feature samples and the corresponding injury labels are preprocessed to obtain a standardized dataset, and the standardized dataset is split into a subset of head impact injury samples and a subset of leg impact injury samples. The head collision injury sample subset and the leg collision injury sample subset are divided into training sets, validation sets, and calibration sets for conditional conformal prediction, respectively, for the head collision injury and the leg collision injury. A first surrogate model is trained based on the training and validation sets of the head impact injury subset to obtain a first initial prediction model, and a second surrogate model is trained based on the training and validation sets of the leg impact injury subset to obtain a second initial prediction model. Based on the calibration set of the head impact injury subset, the first initial prediction model is subjected to conditional conformal prediction calibration and uncertainty screening to obtain a first candidate sample set; based on the calibration set of the leg impact injury subset, the second initial prediction model is subjected to conditional conformal prediction calibration and uncertainty screening to obtain a second candidate sample set. Based on the first candidate sample set, the first initial prediction model is actively learned and iteratively optimized to obtain a head collision injury prediction model, and based on the second candidate sample set, the second initial prediction model is actively learned and iteratively optimized to obtain a leg collision injury prediction model.

[0006] In one possible implementation, the step of performing conditional conformal prediction calibration and uncertainty screening on the first initial prediction model based on the calibration set of the head impact injury subset to obtain a first candidate sample set includes: The input feature samples in the calibration set of the head collision injury subset are used as target samples in sequence. Obtain neighboring samples within the local neighborhood of the target sample, and input the target sample and each neighboring sample into the first initial prediction model to obtain the point prediction values ​​corresponding to the target sample and each neighboring sample; Conditional prediction intervals are generated for each neighboring sample based on the predicted point values ​​and conditional error quantiles of each neighboring sample. The cognitive uncertainty of the target sample is then calculated based on these conditional prediction intervals. The conditional error quantiles are obtained by training and fitting a quantile regression model using all predicted point values ​​and damage labels corresponding to all predicted point values ​​in the calibration set. The cognitive uncertainty is characterized by the dispersion of the prediction interval width within the local neighborhood of the target sample. The prediction interval width within the local neighborhood of the target sample is calculated from the conditional prediction intervals of each neighboring sample. A forward propagation analysis of the fluctuations of the input features of the target sample is performed to solve for the random uncertainty corresponding to the target sample. If the cognitive uncertainty is high, the random uncertainty is high, and / or the point prediction damage value of the target sample is greater than or equal to the damage critical threshold, then the input feature sample corresponding to the target sample is marked as a critical candidate sample; wherein, when the dispersion of the prediction interval width in the local neighborhood of the target sample is greater than the dispersion threshold, the cognitive uncertainty is determined to be high; when the random uncertainty is greater than the random uncertainty threshold, the random uncertainty is determined to be high. The first candidate sample set is obtained by summarizing all the key candidate samples.

[0007] In one possible implementation, the formula for calculating the conditional prediction interval is as follows: ; in, For the conditional prediction interval, The coverage condition is satisfied at any input x. Y is the damage label corresponding to the sample, α is the preset significance parameter; x is the sample; These are the predicted values ​​for the sample points; The conditional error quantile is given.

[0008] In one possible implementation, the step of actively learning and iteratively optimizing the first initial prediction model based on the first candidate sample set to obtain a head impact injury prediction model includes: Select the optimal supplementary sample points from the first candidate sample set; The optimal supplementary sample points carrying damage labels are incorporated into the training sample set to obtain the updated training sample set. The updated training sample set is used to iteratively optimize the parameters of the first initial prediction model. Based on each optimized first initial prediction model, the calibration set based on the head collision injury subset is used to perform conditional conformal prediction calibration and uncertainty screening on the first initial prediction model to obtain the first candidate sample set and subsequent steps, until the preset iteration stop condition is met.

[0009] In one possible implementation, the selection formula for the optimal supplementary sample points is as follows: ; in, The optimal supplementary sample point; The first candidate sample set or the second candidate sample set; The mean of the predictions made by the first initial prediction model or the second initial prediction model for the input condition x; The model predicts the variance; This is a weighting function used to enhance the sampling weights in regions of high uncertainty.

[0010] In one possible implementation, the preset iteration stopping condition is: the dispersion of the prediction interval width in the local neighborhood of the target sample is less than the dispersion threshold and the random uncertainty of the target sample is less than the random uncertainty threshold, or the improvement in prediction accuracy of the first initial prediction model optimized based on the updated training sample set is less than the accuracy threshold.

[0011] In one possible implementation, the first surrogate model is a model that combines PointNet++ with a multilayer perceptron, and the second surrogate model is a TabNet model. The training process of the surrogate model uses the Optuna framework to perform Bayesian hyperparameter optimization, and 4-fold cross-validation is used to evaluate the generalization performance of the surrogate model.

[0012] In one possible implementation, the method further includes: Obtain the collision input features to be evaluated; If the collision damage type of the collision input feature to be evaluated is the head collision damage, then the collision input feature to be evaluated is input into the head collision damage prediction model to predict the head collision damage, and the head point prediction damage result is obtained. Based on the head point prediction damage result, conformal prediction calculation, cognitive uncertainty calculation and feature fluctuation positive propagation analysis are performed on the collision input feature to be evaluated to obtain the conditional prediction interval, cognitive uncertainty rating and random uncertainty rating. If the collision damage type of the collision input feature to be evaluated is the leg collision damage, then the collision input feature to be evaluated is input into the leg collision damage prediction model to predict the leg collision damage, and the leg point prediction damage result is obtained. Based on the leg point prediction damage result, conformal prediction calculation, cognitive uncertainty calculation and feature fluctuation positive propagation analysis are performed on the collision input feature to be evaluated to obtain the conditional prediction interval, cognitive uncertainty rating and random uncertainty rating.

[0013] An apparatus for constructing a pedestrian-vehicle collision damage prediction model, the apparatus comprising: The sample acquisition unit is used to acquire input feature samples labeled with damage tags corresponding to the collision damage types; the collision damage types include head collision damage and leg collision damage. The data preprocessing unit is used to preprocess the input feature samples and the corresponding damage labels of the input feature samples to obtain a standardized dataset. The dataset splitting unit is used to split the standardized dataset into a subset of head impact injury samples and a subset of leg impact injury samples; The dataset partitioning unit is used to partition the head collision injury sample subset and the leg collision injury sample subset respectively to obtain the training set, validation set and calibration set for conditional conformal prediction corresponding to the head collision injury and the leg collision injury respectively. The first model training unit is used to train the first surrogate model based on the training set and validation set of the head collision injury subset to obtain the first initial prediction model. The second model training unit is used to train the second surrogate model based on the training set and validation set of the leg collision injury subset to obtain the second initial prediction model. The first candidate set generation unit is used to perform conditional conformal prediction calibration and uncertainty screening on the first initial prediction model based on the calibration set of the head collision injury subset to obtain the first candidate sample set; The second candidate set generation unit is used to perform conditional conformal prediction calibration and uncertainty screening on the second initial prediction model based on the calibration set of the leg collision injury subset to obtain a second candidate sample set; The first iterative optimization unit is used to actively learn and iteratively optimize the first initial prediction model based on the first candidate sample set to obtain a head collision injury prediction model. The second iterative optimization unit is used to actively learn and iteratively optimize the second initial prediction model based on the second candidate sample set to obtain a leg collision injury prediction model.

[0014] A device for constructing a pedestrian and vehicle collision damage prediction model includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for constructing the pedestrian and vehicle collision damage prediction model as described above.

[0015] Compared with the prior art, this application has the following advantages: This application provides a method for constructing a pedestrian and vehicle collision damage prediction model and related products. Specifically, when implementing the method for constructing a pedestrian and vehicle collision damage prediction model provided in this application, input feature samples with damage labels are first obtained. After data preprocessing, the samples are split into corresponding sample subsets according to head and leg collision damage types. Then, training sets, validation sets, and calibration sets for conditional conformal prediction are respectively divided into two sample subsets. Initial surrogate models corresponding to head and leg collision damage are generated by training the training and validation sets separately. Then, conditional conformal prediction calibration is performed using the calibration set. Deterministic condition screening is carried out by combining cognitive uncertainty and random uncertainty to obtain corresponding candidate sample sets. Finally, the candidate sample sets are used to drive the initial prediction model to complete active learning and iterative optimization, outputting head collision damage prediction models and leg collision damage prediction models respectively. This application no longer relies on complex simulation modeling and solution processes, but uses data-driven proxy models to achieve damage prediction. At the same time, it abandons the random sampling mode and relies on quantitative dual uncertainty indicators to selectively screen high-value working conditions, allowing model iteration to focus on samples with insufficient prediction credibility. This reduces labor costs and time expenditure, and improves the reproducibility and training efficiency of model optimization, thereby better meeting the development needs of rapid iterative optimization of vehicle and pedestrian protection performance. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings used in the description of the embodiment or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating a method for constructing a pedestrian-vehicle collision damage prediction model, as provided in an embodiment of this application. Figure 2 A flowchart illustrating a method for generating a candidate sample set, as provided in an embodiment of this application; Figure 3 A flowchart illustrating an iterative optimization method for a prediction model provided in this application embodiment; Figure 4 This is a schematic diagram of a device for constructing a pedestrian and vehicle collision damage prediction model, provided in an embodiment of this application. Detailed Implementation

[0018] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0019] Pedestrian protection is an important evaluation dimension of vehicle passive safety. Regulations such as C-NCAP and Euro NCAP have clear requirements for pedestrian head and leg collision performance. In the traditional development process, engineers need to create vehicle and pedestrian impactor models in CAE software such as LS-DYNA, complete preprocessing such as geometry cleaning, mesh generation, material definition, and contact settings, and then submit the solution. The drawback is that a single collision simulation can take several hours to several days. As a result, the traditional CAE simulation method will seriously restrict the design iteration efficiency of the vehicle pedestrian protection scheme due to the excessively long simulation calculation cycle.

[0020] To address this issue, this application provides a method and related products for constructing a pedestrian-vehicle collision damage prediction model. It constructs corresponding sample subsets for head and leg collision damage, and divides these subsets into training, validation, and calibration sets for conditional conformal prediction. Initial surrogate models for each damage type are then trained. The calibration sets are used to perform conditional conformal prediction calibration, and a candidate sample set is output based on uncertainty screening. Active learning and iterative optimization are then implemented using these candidate sample sets, with high-value samples selectively added to update the model. This application eliminates the need for repeated complex pre-simulation processing and lengthy simulations, reducing the cost of pedestrian collision performance evaluation while maintaining prediction reliability, and effectively improving the iterative optimization efficiency of vehicle-pedestrian protection performance.

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] See Figure 1 The figure is a flowchart of a method for constructing a pedestrian-vehicle collision damage prediction model according to an embodiment of this application. Figure 1 As shown, the method for constructing the pedestrian-vehicle collision damage prediction model may include steps S101-S106: S101: Obtain input feature samples labeled with damage tags corresponding to the collision damage type.

[0023] To train models separately for two different types of pedestrian head and leg collision injuries, we first acquire input feature samples labeled with damage tags corresponding to each collision injury type. These collision injury types include head collision injuries and leg collision injuries. Each acquired input feature sample carries a true damage label, with each label corresponding one-to-one with the collision input feature. This forms the initial data foundation required for model training. Subsequently, the samples can be grouped according to the damage type to distinguish between head and leg collision injury data, avoiding data mixing and interference between different collision conditions. This ensures that the subsequent proxy model can learn the feature-damage mapping relationship corresponding to head and leg collisions respectively, providing a reliable data source for subsequent dataset splitting, model training, and uncertainty assessment.

[0024] Specifically, when the collision injury type is head collision injury, the input feature sample includes the shell element geometry features, material parameters, thickness distribution, impact point coordinates, and point cloud data around the impact point (i.e., point cloud data within a 50mm radius around the impact point), and the damage label corresponding to the input feature sample is the HIC15 value; when the collision injury type is leg collision injury, the input feature sample includes the shell element geometry features, material parameters, thickness distribution, and impact point coordinates, and the damage label corresponding to the input feature sample is tibial acceleration and knee flexion angle.

[0025] S102: Perform data preprocessing on the input feature samples and the damage labels corresponding to the input feature samples to obtain a standardized dataset, and split the standardized dataset into a subset of head collision damage samples and a subset of leg collision damage samples.

[0026] To eliminate issues such as missing data, inconsistent units, and large differences in numerical ranges in the original input feature samples, and to reduce the adverse effects of outlier data on the subsequent model training convergence and prediction accuracy, while simultaneously achieving separate management of head and leg collision damage data, a standardized dataset can be obtained by preprocessing the input feature samples and their corresponding damage labels. Through data preprocessing, outlier removal, missing value imputation, and feature normalization are performed, unifying the numerical distribution of all sample features to form a high-quality standardized dataset. Then, based on the collision damage type of the samples, the standardized dataset is split into a head collision damage sample subset and a leg collision damage sample subset, isolating the data from these two different collision mechanisms. This prevents interference between features caused by mixed training of head and leg collision samples, facilitating independent dataset partitioning, surrogate model training, and uncertainty filtering for the two subsets, ensuring the independence of model mapping relationship learning under different collision damage scenarios.

[0027] The data preprocessing includes: data cleaning, feature normalization / standardization, point cloud format conversion, and dataset partitioning.

[0028] S103: Divide the head collision injury sample subset and the leg collision injury sample subset into separate sets to obtain the training set, validation set, and calibration set for conditional conformal prediction corresponding to the head collision injury and the leg collision injury, respectively.

[0029] To accommodate the multiple needs of surrogate model training and parameter tuning, generalization performance evaluation, and conditional conformal prediction uncertainty calibration, the head collision damage sample subset and the leg collision damage sample subset are divided into training sets, validation sets, and calibration sets for conditional conformal prediction corresponding to the head collision damage and the leg collision damage, respectively.

[0030] The training set primarily serves to learn model parameters and drives the surrogate model to fit the mapping relationship between input features and collision damage labels. The validation set is used to monitor the generalization performance of the model during training, assist in hyperparameter tuning, and avoid model overfitting. The specially allocated calibration set does not participate in model parameter training and is used for quantile regression fitting, cognitive uncertainty, and random uncertainty quantification in the subsequent conditional conformal prediction stage.

[0031] By independently splitting each subset of head and leg injuries into three sets of datasets with different purposes, we can ensure that the data for model training and validation are not leaked from each other, and provide a dedicated data base for uncertainty screening. This avoids data leakage problems caused by direct reuse of training data and ensures the reliability of subsequent conditional conformal prediction calibration results and candidate sample screening results.

[0032] S104: Train a first surrogate model based on the training and validation sets of the head impact injury subset to obtain a first initial prediction model, and train a second surrogate model based on the training and validation sets of the leg impact injury subset to obtain a second initial prediction model.

[0033] To overcome the technical bottlenecks of traditional pedestrian collision simulation, which relies on CAE modeling, has long solution time, and low iteration efficiency, and to adapt to the completely different characteristics and damage mechanisms of head and leg collision conditions, a first surrogate model can be trained based on the training and validation sets of the head collision damage subset to obtain a first initial prediction model for head collision damage prediction. A second surrogate model can be independently trained based on the training and validation sets of the leg collision damage subset to obtain a second initial prediction model for leg collision damage prediction.

[0034] Compared to the approach of uniformly fitting multiple collision scenarios with a single model, this application adopts a strategy of independently modeling damage types. This strategy enables the targeted learning of the feature-damage mapping rules specific to head and leg collisions, effectively avoiding model fitting bias and learning confusion caused by differences in feature distribution between the two scenarios. Simultaneously, iterative fitting of model parameters is performed using a dedicated training set, and the model's generalization ability is verified in real time using a validation set to select the optimal model hyperparameters. This effectively suppresses overfitting and underfitting issues, rapidly constructing a reliable initial surrogate model adapted to the corresponding collision scenario. This provides a high-precision basic prediction model for subsequent uncertainty calibration, high-value sample selection, and active learning iterative optimization, completely replacing the traditional time-consuming and laborious full-process CAE simulation calculation, and achieving rapid preliminary prediction of collision damage.

[0035] In one possible implementation, the first surrogate model is a model that combines PointNet++ with a multilayer perceptron, and the second surrogate model is a TabNet model. The training process of the surrogate model uses the Optuna framework to perform Bayesian hyperparameter optimization, and 4-fold cross-validation is used to evaluate the generalization performance of the surrogate model.

[0036] Specifically, to adapt to the data characteristics of head collision injury samples, which simultaneously contain point cloud geometric information and structured numerical features around the impact point, the first surrogate model adopts a model that integrates PointNet++ and multilayer perceptron. PointNet++ is responsible for extracting the complex point cloud geometric features around the impact point, while the multilayer perceptron completes the mapping learning of structured features such as thickness, coordinates, and material parameters, thereby realizing the fusion modeling of geometric and numerical features. Leg collision injury samples are mainly characterized by structured tabular features, so the TabNet model is selected, relying on its built-in feature selection mechanism to explore the contribution relationship of different input features to leg collision injury.

[0037] In one possible implementation, the training process of the surrogate model can incorporate the Optuna framework for Bayesian hyperparameter optimization, automatically searching for the optimal combination of network hyperparameters and reducing subjective bias caused by manual parameter tuning. Simultaneously, the training process of the surrogate model employs a 4-fold cross-validation method to perform grouped training and testing on the dataset, evaluating the generalization performance of the surrogate model from multiple perspectives, reducing the evaluation randomness caused by a single dataset partition, avoiding model overfitting, and ensuring that the initially trained prediction model possesses stable and reliable predictive capabilities.

[0038] S105: Based on the calibration set of the head impact injury subset, the first initial prediction model is subjected to conditional conformal prediction calibration and uncertainty screening to obtain a first candidate sample set; based on the calibration set of the leg impact injury subset, the second initial prediction model is subjected to conditional conformal prediction calibration and uncertainty screening to obtain a second candidate sample set.

[0039] In order to accurately identify high-value labeled work condition samples in the active learning process, avoid blindly selecting samples and wasting labeling resources, and at the same time take into account the reliability of the model output results, the first initial prediction model can be calibrated for conditional conformal prediction and uncertainty screening based on the calibration set of the head collision injury subset to obtain the first candidate sample set, and the second initial prediction model can be calibrated for conditional conformal prediction and uncertainty screening based on the calibration set of the leg collision injury subset to obtain the second candidate sample set.

[0040] By using a calibration set to perform conditional conformal prediction calibration, the conditional error quantiles can be calculated, providing a basis for subsequent calculations of the conditional prediction interval and ensuring that the model prediction results have a controllable confidence coverage. This uncertainty screening does not rely solely on a single-dimensional indicator; instead, it combines two types of indicators—cognitive uncertainty representing model fitting defects and random uncertainty representing inherent discrete tolerances of input features—to comprehensively evaluate the risk level of samples. Samples with high cognitive uncertainty, high random uncertainty, or those in critical damage threshold conditions are identified as high-value candidate samples. The candidate sample set selected in this way focuses on conditions where model predictions are unreliable, discarding a large number of ordinary samples with low information gain. This reduces the workload of subsequent sample labeling and provides a high-quality sample source for subsequent active learning iterative optimization, preventing the limitation of model improvement due to one-sided sample selection, and laying the foundation for subsequent model performance iterative optimization.

[0041] S106: Based on the first candidate sample set, the first initial prediction model is actively learned and iteratively optimized to obtain a head collision injury prediction model, and based on the second candidate sample set, the second initial prediction model is actively learned and iteratively optimized to obtain a leg collision injury prediction model.

[0042] In order to fully explore the gain effect of high-value working condition samples in the candidate sample set and further make up for the deficiency of the initial prediction model in predictive ability under high uncertainty working conditions, the first initial prediction model can be actively learned and iteratively optimized based on the first candidate sample set to obtain a head collision injury prediction model, and the second initial prediction model can be actively learned and iteratively optimized based on the second candidate sample set to obtain a leg collision injury prediction model.

[0043] This step expands the existing training data with candidate samples obtained through uncertainty screening. It continuously updates model parameters using newly added high-value samples, repeatedly executing a closed-loop process of model prediction, uncertainty screening, sample expansion, and parameter updates. This continuously strengthens the model's ability to learn high-risk, highly discrete collision conditions, rather than indiscriminately expanding all condition samples, effectively reducing the manpower and time costs of sample annotation. Furthermore, during the iteration process, a preset stopping condition is used to determine whether to end the optimization. While considering training overhead, the model continuously improves its prediction accuracy and generalization ability for head and leg collision injuries. Finally, it outputs iteratively optimized head and leg collision injury prediction models, resulting in prediction models that can be used for inference in the actual development stage, thus meeting the engineering development needs for rapid and reliable assessment of pedestrian and vehicle collision injuries.

[0044] Based on the content of S101-S106, we first distinguish between two core collision injury scenarios—head and leg—and construct labeled sample sets. Through data preprocessing and refined dataset partitioning, we construct dedicated data subsets for model training, performance verification, and conformal calibration. On this basis, we independently train dedicated surrogate prediction models for different injury types. Simultaneously, in the calibration and screening stage, we integrate the dual dimensions of cognitive uncertainty and random uncertainty caused by input feature fluctuations, combining conditional conformal prediction to complete model calibration and high-value candidate sample screening. This overcomes the shortcomings of insufficient randomness and one-sided screening dimensions in single-uncertainty screening. Finally, through an active learning iterative optimization strategy, we continuously focus on updating and optimizing the model for high-uncertainty, high-risk injury scenarios, eliminating the need for repeated, time-consuming simulation solutions. While comprehensively covering pedestrian head and leg collision injury prediction scenarios, this significantly shortens the performance iteration cycle and substantially improves the overall efficiency and prediction accuracy of pedestrian protection performance development.

[0045] See Figure 2 , Figure 2 This is a flowchart of a method for generating a candidate sample set provided in an embodiment of this application. Accordingly, in step S105, the first candidate sample set is obtained by performing conditional conformal prediction calibration and uncertainty screening on the first initial prediction model based on the calibration set of the head collision injury subset. This can be specifically implemented through steps S201-S206: S201: The input feature samples in the calibration set of the head collision injury subset are used as target samples in sequence.

[0046] In order to fully perform uncertainty assessment and candidate sample discrimination for each input feature sample in the calibration set, the input feature samples in the calibration set of the head collision injury subset can be used as target samples in sequence.

[0047] By employing a sample-by-sample traversal approach, all samples in the calibration set can be considered for evaluation. For each target sample, neighboring samples within its local neighborhood are searched, thereby completing a full set of calculations for subsequent conditional prediction intervals, cognitive uncertainty, and stochastic uncertainty. This avoids the omission of some high-risk operating conditions due to partial sampling, ensuring that no operating condition samples with high annotation value are missed. It achieves full coverage screening of the calibration set samples, guaranteeing the integrity of subsequent uncertainty screening results, and preventing the generation of a candidate sample set from missing key samples due to incomplete sample traversal. This lays the foundation for providing sufficient and reliable candidate samples for subsequent active learning.

[0048] S202: Obtain neighboring samples in the local neighborhood of the target sample, and input the target sample and each neighboring sample into the first initial prediction model to obtain the point prediction values ​​corresponding to the target sample and each neighboring sample.

[0049] In order to accurately quantify the model prediction stability of the target sample in the local feature space and effectively explore the prediction fluctuation characteristics of the model in similar working conditions, we can obtain the neighboring samples in the local neighborhood of the target sample, and input the target sample and each neighboring sample into the first initial prediction model to obtain the point prediction values ​​corresponding to the target sample and each neighboring sample.

[0050] Since the prediction results of a single target sample can only reflect the prediction output of a single working condition, they cannot reflect the prediction bias and dispersion characteristics of the model in a cluster of similar working conditions, making it difficult to support the subsequent calculation of cognitive uncertainty. Therefore, by retrieving local neighbor samples in the feature space of the target sample, a set of sample neighborhoods with similar working conditions is constructed. These neighborhoods are then uniformly input into the initial prediction model to complete batch predictions, obtaining the point prediction results of the entire set of neighborhood samples. This method can rely on multiple sets of prediction data in the local neighborhood to objectively reflect the model's fitting stability and learning defects in that working condition region. This provides sufficient and effective data support for subsequent calculation of the dispersion of the conditional prediction interval and accurate quantification of cognitive uncertainty, ensuring that the uncertainty evaluation results closely match the model's true prediction characteristics.

[0051] For example, assuming the target sample is a pedestrian head collision, its input features include the geometric features of the shell unit in the impact area, material parameters, thickness distribution, impact point coordinates, and point cloud data within a 50mm radius around the impact point, with the corresponding damage label being the HIC15 value. The key structured parameters of this target sample are: material equivalent stiffness 12000 N / m, pedestrian head collision velocity 40 km / h, head mass 4.5 kg, and collision angle 18°. By searching the feature space distance, the first neighboring sample can be obtained: the shell unit geometric features in the impact area are similar, the same material parameters are used, the thickness distribution is similar, the impact point coordinates have small deviations, and the point cloud distribution within 50mm around the impact point is similar. The structured parameters are: material equivalent stiffness 11800 N / m, velocity 39.5 km / h, and head mass 4.5 kg. The first neighboring sample has a weight of 45 kg and a collision angle of 17.6°. The second neighboring sample has similar geometric features of the shell units in the impact area, uses the same material parameters, has similar thickness distribution, has small deviation in the coordinates of the impact point, and has a similar point cloud distribution pattern in the 50 mm area around the impact point. The structured parameters are: material equivalent stiffness 12200 N / m, velocity 40.2 km / h, head mass 4.52 kg, and collision angle 18.3°. The third neighboring sample has similar geometric features of the shell units in the impact area, uses the same material parameters, has similar thickness distribution, has small deviation in the coordinates of the impact point, and has a similar point cloud distribution pattern in the 50 mm area around the impact point. The structured parameters are: material equivalent stiffness 11900 N / m, velocity 39.8 km / h, head mass 4.48 kg, and collision angle 17.9°. These three samples are the neighboring samples in the local neighborhood of the target sample.

[0052] S203: Generate conditional prediction intervals for each neighboring sample based on the point prediction values ​​and conditional error quantiles corresponding to each neighboring sample, and calculate the cognitive uncertainty of the target sample based on the conditional prediction intervals corresponding to each neighboring sample.

[0053] To accurately construct reliable prediction intervals by combining the distribution patterns of model prediction errors, and to quantify the cognitive uncertainty of the model itself by relying on the prediction fluctuation characteristics of local operating conditions, conditional prediction intervals can be generated for each neighboring sample based on the point prediction values ​​and conditional error quantiles corresponding to each neighboring sample. The conditional error quantiles are not fixed empirical values, but are obtained by training and fitting a quantile regression model using the model prediction values ​​of all samples in the calibration set and the actual damage labels. This ensures that the generated conditional prediction intervals closely match the actual error distribution characteristics of the operating condition, guaranteeing accurate confidence coverage and effectively correcting systematic prediction biases in the model.

[0054] Based on this, the cognitive uncertainty corresponding to the target sample is quantitatively characterized by solving the interval width of the condition prediction interval of each neighboring sample and statistically analyzing the dispersion of all interval widths in the local neighborhood of the target sample.

[0055] This method, by observing the fluctuation differences of multiple prediction intervals in a local neighborhood, intuitively reflects the inherent defects of the model in that feature region, such as insufficient fitting and incomplete learning. Unlike the one-sided evaluation of single-point prediction error, it can comprehensively and stably quantify the model's cognitive uncertainty, providing a reliable basis for judging model uncertainty for subsequent high-value sample selection.

[0056] S204: Perform a forward propagation analysis of the input features of the target sample to obtain the random uncertainty corresponding to the target sample.

[0057] In order to fully reflect the objective disturbances inherent in the input features during engineering development and avoid the lack of dimension caused by relying solely on the model's neighborhood dispersion for uncertainty evaluation, a positive propagation analysis of the fluctuations of the input features of the target sample can be performed to solve for the random uncertainty corresponding to the target sample.

[0058] In actual pedestrian collision testing, manufacturing tolerances of components, assembly differences, and fluctuations in test parameters can all cause random fluctuations in input features within a reasonable range. This type of uncertainty originates from the working conditions themselves and cannot be eliminated by optimizing model parameters. This step uses the original input features of the target sample as a benchmark and constructs multiple sets of perturbation inputs based on a preset fluctuation range for the perturbable feature dimensions. These perturbation inputs are then fed into the initial prediction model to complete forward inference, capturing the output fluctuation amplitude generated by the transmission of input feature fluctuations to the damage prediction results. This allows for the quantitative calculation of random uncertainty.

[0059] This computational logic is independent of the cognitive uncertainty that characterizes the model fitting defects. It incorporates the inherent random disturbances of the working conditions into the evaluation system, so that the subsequent candidate sample screening takes into account both the risk factors at the model level and the working conditions themselves, thereby improving the completeness of the sample screening and preventing the omission of working condition samples with high annotation value due to their sensitivity to input disturbances.

[0060] S205: If the cognitive uncertainty is high, the random uncertainty is high, and / or the point prediction damage value of the target sample is greater than or equal to the damage critical threshold, then the input feature sample corresponding to the target sample is marked as a critical candidate sample.

[0061] In order to select high-value working condition samples that maximize the gain of model iteration optimization and avoid unnecessary consumption of labeling resources due to indiscriminate selection of ordinary samples, if the cognitive uncertainty is high, the random uncertainty is high and / or the point prediction damage value of the target sample is greater than or equal to the damage critical threshold, then the input feature sample corresponding to the target sample is marked as a critical candidate sample.

[0062] Among them, the dispersion of the prediction interval width within the local neighborhood of the target sample is used as the basis for judging cognitive uncertainty. When the dispersion exceeds the preset dispersion threshold, it means that the model's fitting effect in the feature region is unstable, and the cognitive uncertainty is judged to be high. The calculation result obtained by the positive propagation analysis of feature fluctuations is used as the basis for judging random uncertainty. When its value exceeds the random uncertainty threshold, it means that the working condition is more sensitive to the disturbance of input parameters, and the working condition itself has a high risk.

[0063] This application employs a multi-condition "OR" decision logic, capturing samples with high cognitive uncertainty due to insufficient model learning, identifying samples with high random uncertainty due to inherent input perturbations, and simultaneously considering hazardous conditions that reach critical damage thresholds—a key focus in engineering development. All three types of samples are uniformly identified as key candidate samples. Compared to single-indicator screening methods, this multi-dimensional joint decision-making approach can fully explore different types of high-gain samples, preventing the omission of certain risk conditions and providing a high-quality sample source for subsequent active learning iterations and optimizations, maximizing model performance improvement with limited annotation costs.

[0064] S206: Summarize all key candidate samples to obtain the first candidate sample set.

[0065] To aggregate and integrate the various high-value samples obtained through multi-dimensional uncertainty discrimination and screening, a first candidate sample set is obtained by summarizing all key candidate samples. The preceding steps have already identified risky operating condition samples with high cognitive uncertainty, high random uncertainty, or those at the critical damage threshold from the calibration set. This aggregation operation gathers all labeled key candidate samples together, eliminating a large number of ordinary operating condition samples with low information gain, forming a first candidate sample set that centrally contains high-risk samples. This candidate sample set no longer contains a large number of inefficient samples, and the samples as a whole have high training gain, which can be directly used for subsequent active learning iterative optimization. This reduces the training overhead caused by invalid samples participating in iterations and avoids labeling a large number of low-value samples, saving labeling costs. At the same time, it fully retains samples of various risky operating conditions, ensuring that subsequent model optimization can specifically supplement the learning capabilities of weak operating conditions, providing reliable and centralized sample data support for the iterative updates of the initial prediction model.

[0066] In one possible implementation, the critical damage threshold can be set to 1000, corresponding to the HIC15 safety evaluation critical index for pedestrian head collisions.

[0067] Steps S201-S206 iterate through all samples in the calibration set as target samples, and perform conditional conformal prediction calculations using neighboring samples in the local neighborhood. The dispersion of the conditional prediction interval width of neighboring samples is used to quantify cognitive uncertainty, and random uncertainty is obtained by combining the positive propagation analysis of the fluctuations in the input features of the target samples. Then, a multi-dimensional conditional discrimination is performed by integrating cognitive uncertainty, random uncertainty, and key damage thresholds, identifying key candidate samples with high annotation gain and summarizing them into the first candidate sample set. This achieves joint screening of conditional conformal prediction calibration and dual-dimensional uncertainty, ensuring the reliability of the coverage of the conditional prediction interval while taking into account both the cognitive uncertainty brought by the model itself and the random uncertainty brought by the input feature tolerance. This avoids the one-sidedness of single-index screening, provides a high-value sample source for subsequent active learning iteration and optimization, reduces the overhead of invalid sample annotation, and helps the model focus on optimizing prediction performance under high-risk and highly unreliable conditions.

[0068] It should be noted that the generation process of the second candidate sample set is similar to that of the first candidate sample set, and will not be repeated here. The difference lies in that the dataset used for the screening operation is a subset of leg collision injury samples, and the prediction model used is replaced by the second initial prediction model. Furthermore, leg collision injury conditions have independent critical thresholds: tibial acceleration threshold and knee flexion angle threshold are set separately. When the predicted tibial acceleration of a sample point is greater than or equal to the tibial acceleration threshold, and / or the knee flexion angle is greater than or equal to the knee flexion angle threshold, the sample is also determined to be a high-risk condition and marked as a critical candidate sample. By adapting the evaluation thresholds specific to leg conditions, the uncertainty screening under leg collision conditions is ensured to conform to engineering safety assessment standards, guaranteeing the effectiveness of the generated second candidate sample set.

[0069] In one possible implementation, the formula for calculating the conditional prediction interval is as follows: ; in, For the conditional prediction interval, The coverage condition is satisfied at any input x. Y is the damage label corresponding to the sample, α is the preset significance parameter; x is the sample; These are the predicted values ​​for the sample points; The conditional error quantile is given.

[0070] See Figure 3 , Figure 3This application provides a flowchart of an iterative optimization method for a prediction model. Accordingly, step S106, which involves actively learning and iteratively optimizing the first initial prediction model based on the first candidate sample set to obtain a head collision injury prediction model, can be specifically implemented through steps S301-S303. S301: Select the optimal supplementary sample points from the first candidate sample set.

[0071] To maximize the model's prediction accuracy and generalization ability within limited sample labeling costs and training iterations, and to avoid resource redundancy and low iteration efficiency caused by indiscriminately inputting all high-uncertainty samples from the first candidate sample set into the model training set, optimal supplementary sample points can be further screened within the first candidate sample set. Although the first candidate sample set obtained through the aforementioned uncertainty screening consists of high-value risk condition samples with high information gain, the improvement effects of different samples on model optimization vary significantly. Some samples have high condition redundancy and limited marginal gain on model performance. Therefore, using the optimal supplementary sample point screening formula, all key candidate samples in the first candidate sample set are evaluated and ranked a second time. This comprehensively considers the magnitude of cognitive uncertainty, random uncertainty, and damage risk level of the samples, accurately selecting the optimal supplementary sample points that have the most significant effect on model correction and condition completion in the current iteration round.

[0072] This screening method can accurately identify the core gain samples in each iteration, eliminate redundant samples with low gain in the candidate set, and achieve precise empowerment of sample labeling and model training. It ensures that each iteration of optimization can efficiently make up for the model's fitting defects, greatly improve the efficiency and pertinence of active learning iteration optimization, and provide optimal data support for subsequent model parameter updates and performance iteration convergence.

[0073] In one possible implementation, the selection formula for the optimal supplementary sample points is as follows: ; in, The optimal supplementary sample point; The first candidate sample set or the second candidate sample set; The mean of the predictions made by the first initial prediction model or the second initial prediction model for the input condition x; The model predicts the variance; This is a weighting function used to enhance the sampling weights in regions of high uncertainty.

[0074] S302: The optimal supplementary sample points carrying damage labels are incorporated into the training sample set to obtain the updated training sample set.

[0075] To enable the model to learn the real-world operating conditions corresponding to the optimal supplementary sample points and compensate for the insufficient coverage of the original training sample set under highly unreliable operating conditions, the optimal supplementary sample points carrying damage labels are incorporated into the training sample set, resulting in an updated training sample set. The optimal supplementary sample points themselves belong to high-gain operating conditions identified within the dataset; only by introducing their real damage labels can the model correct the original network parameters through supervised learning.

[0076] By merging labeled samples into the original training sample set, the training data is expanded in a targeted manner, rather than randomly adding ordinary working condition samples. This allows the updated training sample set to focus on engineering working conditions with high cognitive uncertainty, high random uncertainty, or those at the critical threshold of damage. This enriches the training set's coverage of risky working conditions and provides a complete label-feature pairing data foundation for subsequent model parameter iteration and optimization, helping the model to improve the prediction effect of vulnerable working conditions in a targeted manner.

[0077] S303: The updated training sample set is used to iteratively optimize the parameters of the first initial prediction model, and step S105 and subsequent steps are executed repeatedly based on the first initial prediction model after each optimization until the preset iteration stop condition is met.

[0078] In order to continuously enhance the model’s ability to learn from collision conditions with high uncertainty and give full play to the data gain brought by the newly added optimal supplementary sample points, the updated training sample set can be used to iteratively optimize the parameters of the first initial prediction model. Step S105 and subsequent steps are executed repeatedly based on the first initial prediction model after each optimization until the preset iteration stop condition is met.

[0079] After the model parameters are updated, the model's fit in the feature space changes, and the original uncertainty assessment results are no longer applicable. Therefore, it is necessary to re-traverse the calibration set samples, complete the entire process of finding neighboring samples, generating conditional conformal intervals, calculating cognitive uncertainty and random uncertainty, and identifying key candidate samples, and start a new round of closed-loop iteration.

[0080] During the iteration process, sample expansion and parameter updates are continuously performed to make up for the model's learning shortcomings in high-risk working conditions. When the preset iteration stopping condition is reached, the closed-loop process is terminated to avoid meaningless continuous training that consumes computing power. Under controllable training overhead, the prediction accuracy and generalization performance of the head collision injury prediction model are continuously improved.

[0081] In one possible implementation, the preset iteration stopping condition is: the dispersion of the prediction interval width in the local neighborhood of the target sample is less than the dispersion threshold and the random uncertainty of the target sample is less than the random uncertainty threshold, or the improvement in prediction accuracy of the first initial prediction model optimized based on the updated training sample set is less than the accuracy threshold.

[0082] The closed-loop active learning mechanism in steps S301-S303 can fully leverage the advantages of uncertainty screening, prioritize supplementing working condition samples with weak model identification capabilities, maximize sample gain with limited labeling costs, effectively improve the fitting effect of the initial model in unreliable working condition regions, and enhance the overall generalization ability and engineering practical value of the model.

[0083] It should be noted that the construction process of the leg collision injury prediction model is similar to that of the head collision injury prediction model. Both models complete the corresponding sample subset partitioning, surrogate model training, conditional conformal prediction calibration, candidate sample set generation based on dual uncertainty index, and active learning closed-loop iterative optimization. The only difference between the two models is the input dataset and the network structure of the selected surrogate model. The processing logic of other processes such as neighborhood sample retrieval, uncertainty calculation, sample screening and determination, and iteration stopping conditions are the same, and will not be elaborated here.

[0084] In one possible implementation, the method further includes: Obtain the collision input features to be evaluated; If the collision damage type of the collision input feature to be evaluated is the head collision damage, then the collision input feature to be evaluated is input into the head collision damage prediction model to predict the head collision damage, and the head point prediction damage result is obtained. Based on the head point prediction damage result, conformal prediction calculation, cognitive uncertainty calculation and feature fluctuation positive propagation analysis are performed on the collision input feature to be evaluated to obtain the conditional prediction interval, cognitive uncertainty rating and random uncertainty rating. If the collision damage type of the collision input feature to be evaluated is the leg collision damage, then the collision input feature to be evaluated is input into the leg collision damage prediction model to predict the leg collision damage, and the leg point prediction damage result is obtained. Based on the leg point prediction damage result, conformal prediction calculation, cognitive uncertainty calculation and feature fluctuation positive propagation analysis are performed on the collision input feature to be evaluated to obtain the conditional prediction interval, cognitive uncertainty rating and random uncertainty rating.

[0085] Specifically, to complete pedestrian collision damage assessment under actual engineering conditions and output confidence information corresponding to the prediction results, this embodiment adds an inference evaluation step after the model construction and training are completed: It acquires the collision input features to be evaluated in the actual engineering scenario, distributes them to the corresponding trained prediction model according to their collision damage type, and the model outputs point-predicted damage results. Based on this, conformal prediction calculation is performed to obtain a conditional prediction interval with statistical coverage, representing the credible range of damage prediction results. Then, cognitive uncertainty calculation is completed through neighborhood sample analysis in a high-dimensional feature space, outputting a cognitive uncertainty rating to reflect the model's generalization defects for this condition. Furthermore, a positive propagation analysis of feature fluctuations is performed on the collision input features to be evaluated to obtain a stochastic uncertainty rating, representing the prediction fluctuations caused by the perturbation of the input features themselves. Finally, it simultaneously outputs point-predicted damage results, conditional prediction intervals, and two types of uncertainty ratings. This not only provides a damage value estimate but also provides developers with intuitive feedback on the reliability of the prediction results, facilitating engineers to comprehensively assess collision damage risks by combining confidence information. This overcomes the shortcomings of traditional proxy models that only output single-point prediction values ​​and lack confidence references, better serving the analysis of vehicle and pedestrian protection performance.

[0086] It should be noted that the execution logic of conformal prediction calculation, cognitive uncertainty calculation, and feature fluctuation forward propagation analysis for leg collision injuries is the same as that for head collision injuries, only the input objects and the prediction models called are different, which will not be elaborated here.

[0087] See Figure 4 , Figure 4 This is a schematic diagram of a device for constructing a pedestrian and vehicle collision damage prediction model, provided in an embodiment of this application. Figure 4 As shown, the apparatus for constructing the pedestrian-vehicle collision damage prediction model includes: The sample acquisition unit 401 is used to acquire input feature samples labeled with damage tags corresponding to the collision damage types; the collision damage types include head collision damage and leg collision damage. Data preprocessing unit 402 is used to preprocess the input feature samples and the damage labels corresponding to the input feature samples to obtain a standardized dataset; The dataset splitting unit 403 is used to split the standardized dataset into a subset of head impact injury samples and a subset of leg impact injury samples; The dataset partitioning unit 404 is used to partition the head collision injury sample subset and the leg collision injury sample subset respectively to obtain the training set, validation set and calibration set for conditional conformal prediction corresponding to the head collision injury and the leg collision injury respectively. The first model training unit 4051 is used to train the first surrogate model based on the training set and validation set of the head collision injury subset to obtain the first initial prediction model. The second model training unit 4052 is used to train the second surrogate model based on the training set and validation set of the leg collision injury subset to obtain the second initial prediction model. The first candidate set generation unit 4061 is used to perform conditional conformal prediction calibration and uncertainty screening on the first initial prediction model based on the calibration set of the head collision injury subset to obtain the first candidate sample set. The second candidate set generation unit 4062 is used to perform conditional conformal prediction calibration and uncertainty screening on the second initial prediction model based on the calibration set of the leg collision injury subset to obtain a second candidate sample set; The first iterative optimization unit 4071 is used to actively learn and iteratively optimize the first initial prediction model based on the first candidate sample set to obtain a head collision injury prediction model. The second iterative optimization unit 4072 is used to actively learn and iteratively optimize the second initial prediction model based on the second candidate sample set to obtain a leg collision injury prediction model.

[0088] In one possible implementation, the first candidate set generation unit 4061 is specifically used for: The input feature samples in the calibration set of the head collision injury subset are used as target samples in sequence. Obtain neighboring samples within the local neighborhood of the target sample, and input the target sample and each neighboring sample into the first initial prediction model to obtain the point prediction values ​​corresponding to the target sample and each neighboring sample; Conditional prediction intervals are generated for each neighboring sample based on the predicted point values ​​and conditional error quantiles of each neighboring sample. The cognitive uncertainty of the target sample is then calculated based on these conditional prediction intervals. The conditional error quantiles are obtained by training and fitting a quantile regression model using all predicted point values ​​and damage labels corresponding to all predicted point values ​​in the calibration set. The cognitive uncertainty is characterized by the dispersion of the prediction interval width within the local neighborhood of the target sample. The prediction interval width within the local neighborhood of the target sample is calculated from the conditional prediction intervals of each neighboring sample. A forward propagation analysis of the fluctuations of the input features of the target sample is performed to solve for the random uncertainty corresponding to the target sample. If the cognitive uncertainty is high, the random uncertainty is high, and / or the point prediction damage value of the target sample is greater than or equal to the damage critical threshold, then the input feature sample corresponding to the target sample is marked as a critical candidate sample; wherein, when the dispersion of the prediction interval width in the local neighborhood of the target sample is greater than the dispersion threshold, the cognitive uncertainty is determined to be high; when the random uncertainty is greater than the random uncertainty threshold, the random uncertainty is determined to be high. The first candidate sample set is obtained by summarizing all the key candidate samples.

[0089] In one possible implementation, the formula for calculating the conditional prediction interval is as follows: ; in, For the conditional prediction interval, The coverage condition is satisfied at any input x. Y is the damage label corresponding to the sample, α is the preset significance parameter; x is the sample; These are the predicted values ​​for the sample points; The conditional error quantile is given.

[0090] In one possible implementation, the first iterative optimization unit 4071 is specifically used for: Select the optimal supplementary sample points from the first candidate sample set; The optimal supplementary sample points carrying damage labels are incorporated into the training sample set to obtain the updated training sample set. The updated training sample set is used to iteratively optimize the parameters of the first initial prediction model. Based on each optimized first initial prediction model, the calibration set based on the head collision injury subset is used to perform conditional conformal prediction calibration and uncertainty screening on the first initial prediction model to obtain the first candidate sample set and subsequent steps, until the preset iteration stop condition is met.

[0091] In one possible implementation, the selection formula for the optimal supplementary sample points is as follows: ; in, The optimal supplementary sample point; The first candidate sample set or the second candidate sample set; The mean of the predictions made by the first initial prediction model or the second initial prediction model for the input condition x; The model predicts the variance; This is a weighting function used to enhance the sampling weights in regions of high uncertainty.

[0092] In one possible implementation, the preset iteration stopping condition is: the dispersion of the prediction interval width in the local neighborhood of the target sample is less than the dispersion threshold and the random uncertainty of the target sample is less than the random uncertainty threshold, or the improvement in prediction accuracy of the first initial prediction model optimized based on the updated training sample set is less than the accuracy threshold.

[0093] In one possible implementation, the first surrogate model is a model that combines PointNet++ with a multilayer perceptron, and the second surrogate model is a TabNet model. The training process of the surrogate model uses the Optuna framework to perform Bayesian hyperparameter optimization, and 4-fold cross-validation is used to evaluate the generalization performance of the surrogate model.

[0094] In one possible implementation, the device further includes: The feature acquisition unit is used to acquire the collision input features to be evaluated; The head injury prediction unit, if the collision injury type of the collision input feature to be evaluated is the head collision injury, is used to input the collision input feature to be evaluated into the head collision injury prediction model to predict the head collision injury and obtain the head point prediction injury result. The first analysis unit is used to perform conformal prediction calculation, cognitive uncertainty calculation and feature fluctuation forward propagation analysis on the collision input features to be evaluated based on the head point predicted damage results, and to obtain the conditional prediction interval, cognitive uncertainty rating and random uncertainty rating. The leg injury prediction unit is used to input the collision input feature to be evaluated into the leg collision injury prediction model to predict the leg collision injury if the collision injury type of the collision input feature to be evaluated is the leg collision injury, and to obtain the leg point prediction injury result. The second analysis unit is used to perform conformal prediction calculation, cognitive uncertainty calculation, and feature fluctuation forward propagation analysis on the collision input features to be evaluated based on the leg point predicted damage results, so as to obtain the conditional prediction interval, cognitive uncertainty rating, and random uncertainty rating.

[0095] In addition, this application embodiment also provides a device for constructing a pedestrian and vehicle collision damage prediction model, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the pedestrian and vehicle collision damage prediction model construction method as described above.

[0096] In addition, this application embodiment also provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the pedestrian and vehicle collision damage prediction model construction method as described above.

[0097] This application's embodiments acquire input feature samples with damage labels and, through data preprocessing and dataset partitioning, obtain training sets, validation sets, and calibration sets for conditional conformal prediction. Initial proxy models corresponding to the head and legs are trained on these sets. Then, based on the calibration sets, conditional conformal prediction calibration and uncertainty screening are performed to obtain candidate sample sets. Finally, active learning iterative optimization is performed on these candidate sample sets to obtain the final collision damage prediction model. This eliminates the need for repeated complex CAE preprocessing and time-consuming simulations, reducing manual overhead and shortening performance evaluation time. Furthermore, by leveraging conditional conformal prediction and uncertainty-driven active learning strategies to prioritize high-uncertainty conditions for iteration, it can adapt to the development needs of rapid iterative optimization of vehicle and pedestrian protection performance while ensuring prediction reliability.

[0098] The foregoing provides a detailed description of the method for constructing a pedestrian and vehicle collision damage prediction model and related products provided in this application. The various embodiments in the specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0099] It should be understood that in this application, "and / or" is used to describe the relationship between related objects, indicating that there can be three relationships. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the related objects before and after it are in an "or" relationship.

[0100] It should also be noted that, in this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for constructing a pedestrian-vehicle collision damage prediction model, characterized in that, The method includes: Obtain input feature samples labeled with damage tags corresponding to the collision damage types; the collision damage types include head collision damage and leg collision damage; The input feature samples and the corresponding injury labels are preprocessed to obtain a standardized dataset, and the standardized dataset is split into a subset of head impact injury samples and a subset of leg impact injury samples. The head collision injury sample subset and the leg collision injury sample subset are divided into training sets, validation sets, and calibration sets for conditional conformal prediction, respectively, for the head collision injury and the leg collision injury. A first surrogate model is trained based on the training and validation sets of the head impact injury subset to obtain a first initial prediction model, and a second surrogate model is trained based on the training and validation sets of the leg impact injury subset to obtain a second initial prediction model. Based on the calibration set of the head impact injury subset, the first initial prediction model is subjected to conditional conformal prediction calibration and uncertainty screening to obtain a first candidate sample set; based on the calibration set of the leg impact injury subset, the second initial prediction model is subjected to conditional conformal prediction calibration and uncertainty screening to obtain a second candidate sample set. Based on the first candidate sample set, the first initial prediction model is actively learned and iteratively optimized to obtain a head collision injury prediction model, and based on the second candidate sample set, the second initial prediction model is actively learned and iteratively optimized to obtain a leg collision injury prediction model.

2. The method according to claim 1, characterized in that, The conditional conformal prediction calibration and uncertainty screening of the first initial prediction model based on the calibration set of the head impact injury subset yields a first candidate sample set, including: The input feature samples in the calibration set of the head collision injury subset are used as target samples in sequence. Obtain neighboring samples within the local neighborhood of the target sample, and input the target sample and each neighboring sample into the first initial prediction model to obtain the point prediction values ​​corresponding to the target sample and each neighboring sample; Conditional prediction intervals are generated for each neighboring sample based on the predicted point values ​​and conditional error quantiles of each neighboring sample. The cognitive uncertainty of the target sample is then calculated based on these conditional prediction intervals. The conditional error quantiles are obtained by training and fitting a quantile regression model using all predicted point values ​​and damage labels corresponding to all predicted point values ​​in the calibration set. The cognitive uncertainty is characterized by the dispersion of the prediction interval width within the local neighborhood of the target sample. The prediction interval width within the local neighborhood of the target sample is calculated from the conditional prediction intervals of each neighboring sample. A forward propagation analysis of the fluctuations of the input features of the target sample is performed to solve for the random uncertainty corresponding to the target sample. If the cognitive uncertainty is high, the random uncertainty is high, and / or the point prediction damage value of the target sample is greater than or equal to the damage critical threshold, then the input feature sample corresponding to the target sample is marked as a critical candidate sample; wherein, when the dispersion of the prediction interval width in the local neighborhood of the target sample is greater than the dispersion threshold, the cognitive uncertainty is determined to be high; when the random uncertainty is greater than the random uncertainty threshold, the random uncertainty is determined to be high. The first candidate sample set is obtained by summarizing all the key candidate samples.

3. The method according to claim 2, characterized in that, The formula for calculating the conditional prediction interval is as follows: ； in, For the conditional prediction interval, The coverage condition is satisfied at any input x. Y is the damage label corresponding to the sample, α is the preset significance parameter; x is the sample; These are the predicted values ​​for the sample points; The conditional error quantile is given.

4. The method according to claim 1, characterized in that, The step of actively learning and iteratively optimizing the first initial prediction model based on the first candidate sample set to obtain a head impact injury prediction model includes: Select the optimal supplementary sample points from the first candidate sample set; The optimal supplementary sample points carrying damage labels are incorporated into the training sample set to obtain the updated training sample set. The updated training sample set is used to iteratively optimize the parameters of the first initial prediction model. Based on each optimized first initial prediction model, the calibration set based on the head collision injury subset is used to perform conditional conformal prediction calibration and uncertainty screening on the first initial prediction model to obtain the first candidate sample set and subsequent steps, until the preset iteration stop condition is met.

5. The method according to claim 4, characterized in that, The formula for selecting the optimal supplementary sample points is as follows: ； in, The optimal supplementary sample point; The first candidate sample set or the second candidate sample set; The mean of the predictions made by the first initial prediction model or the second initial prediction model for the input condition x; The model predicts the variance; This is a weighting function used to enhance the sampling weights in regions of high uncertainty.

6. The method according to claim 4, characterized in that, The preset iteration stopping condition is: the dispersion of the prediction interval width in the local neighborhood of the target sample is less than the dispersion threshold and the random uncertainty of the target sample is less than the random uncertainty threshold, or the improvement in prediction accuracy of the first initial prediction model optimized based on the updated training sample set is less than the accuracy threshold.

7. The method according to claim 1, characterized in that, The first surrogate model is a model that combines PointNet++ and a multilayer perceptron, and the second surrogate model is a TabNet model. The training process of the surrogate models uses the Optuna framework to perform Bayesian hyperparameter optimization, and 4-fold cross-validation is used to evaluate the generalization performance of the surrogate models.

8. The method according to claim 1, characterized in that, The method further includes: Obtain the collision input features to be evaluated; If the collision damage type of the collision input feature to be evaluated is the head collision damage, then the collision input feature to be evaluated is input into the head collision damage prediction model to predict the head collision damage, and the head point prediction damage result is obtained. Based on the head point prediction damage result, conformal prediction calculation, cognitive uncertainty calculation and feature fluctuation positive propagation analysis are performed on the collision input feature to be evaluated to obtain the conditional prediction interval, cognitive uncertainty rating and random uncertainty rating. If the collision damage type of the collision input feature to be evaluated is the leg collision damage, then the collision input feature to be evaluated is input into the leg collision damage prediction model to predict the leg collision damage, and the leg point prediction damage result is obtained. Based on the leg point prediction damage result, conformal prediction calculation, cognitive uncertainty calculation and feature fluctuation positive propagation analysis are performed on the collision input feature to be evaluated to obtain the conditional prediction interval, cognitive uncertainty rating and random uncertainty rating.

9. A device for constructing a pedestrian and vehicle collision damage prediction model, characterized in that, The device includes: The sample acquisition unit is used to acquire input feature samples labeled with damage tags corresponding to the collision damage types; the collision damage types include head collision damage and leg collision damage. The data preprocessing unit is used to preprocess the input feature samples and the corresponding damage labels of the input feature samples to obtain a standardized dataset. The dataset splitting unit is used to split the standardized dataset into a subset of head impact injury samples and a subset of leg impact injury samples; The dataset partitioning unit is used to partition the head collision injury sample subset and the leg collision injury sample subset respectively to obtain the training set, validation set and calibration set for conditional conformal prediction corresponding to the head collision injury and the leg collision injury respectively. The first model training unit is used to train the first surrogate model based on the training set and validation set of the head collision injury subset to obtain the first initial prediction model. The second model training unit is used to train the second surrogate model based on the training set and validation set of the leg collision injury subset to obtain the second initial prediction model. The first candidate set generation unit is used to perform conditional conformal prediction calibration and uncertainty screening on the first initial prediction model based on the calibration set of the head collision injury subset to obtain the first candidate sample set; The second candidate set generation unit is used to perform conditional conformal prediction calibration and uncertainty screening on the second initial prediction model based on the calibration set of the leg collision injury subset to obtain a second candidate sample set; The first iterative optimization unit is used to actively learn and iteratively optimize the first initial prediction model based on the first candidate sample set to obtain a head collision injury prediction model. The second iterative optimization unit is used to actively learn and iteratively optimize the second initial prediction model based on the second candidate sample set to obtain a leg collision injury prediction model.

10. A device for constructing a pedestrian and vehicle collision damage prediction model, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method for constructing a pedestrian-vehicle collision damage prediction model as described in any one of claims 1-8.

Images