A vehicle collision occupant injury prediction method, device, equipment, medium and product

Abstract

This application discloses a method, apparatus, device, medium, and product for predicting occupant injuries in automotive collisions, relating to the field of automotive technology. The method includes: acquiring a first dataset, a second dataset, and operating parameters of a target large-angle seat; training a first deep neural network model on the first dataset to obtain a second training model; wherein the second training model includes a shared feature extraction network and a task prediction network; with the parameters of the shared feature extraction network frozen, training the parameters of the task prediction network a second time on the second dataset to obtain an injury prediction model; inputting the operating parameters of the target large-angle seat into the injury prediction model to obtain an occupant injury prediction result; wherein the occupant injury prediction result includes predicted values ​​corresponding to occupant injury indicators. This method can achieve accurate prediction of occupant injury conditions.

Description

Technical Field

[0001] This application relates to the field of automotive technology, and in particular to a method, device, equipment, medium, and product for predicting occupant injuries in a car collision. Background Technology

[0002] With the development of intelligent cockpits, seats with large angles (such as zero-gravity seats and reclining seats) have become increasingly popular due to their ability to provide a more comfortable driving and riding experience. However, the occupant dynamic response of these seats in a collision differs significantly from that of a standard sitting posture; for example, the backrest angle typically exceeds 45°, and in some cases even reaches over 60°, leading to a decrease in the accuracy of traditional predictive models trained on standard sitting posture data. Therefore, how to accurately predict occupant injury in scenarios involving large-angle seats has become a crucial issue for ensuring vehicle passive safety performance and promoting the safety upgrade of intelligent cockpits.

[0003] Currently, existing technologies face many limitations in high-angle seat scenarios, such as the scarcity of real-world crash test data for high-angle seats, high costs, and strict ethical restrictions, making it impossible to accurately predict occupant injuries.

[0004] Therefore, there is an urgent need for a method that can accurately predict occupant injuries. Summary of the Invention

[0005] This application provides a method, device, equipment, medium, and product for predicting occupant injuries in automobile collisions, which can achieve accurate prediction of occupant injury conditions.

[0006] To achieve the above objectives, this application adopts the following technical solution:

[0007] In a first aspect, this application provides a method for predicting occupant injuries in a car collision, including:

[0008] Obtain the operating parameters of the first dataset, the second dataset, and the target large-angle seat; wherein, the first dataset is occupant collision data under standard sitting posture conditions, the second dataset is occupant collision data under large-angle seat conditions, and the operating parameters include at least one of the following: seat back angle, collision pulse waveform, seat belt parameters, and airbag parameters.

[0009] Based on the first dataset, the first deep neural network model is trained for the first time to obtain the second training model. The second training model includes a shared feature extraction network and a task prediction network. With the parameters of the shared feature extraction network frozen, the parameters of the task prediction network are trained for the second time based on the second dataset to obtain the damage prediction model. The working condition parameters of the target large-angle seat are input into the damage prediction model to obtain the occupant damage prediction result. The occupant damage prediction result includes the predicted value corresponding to the occupant damage index.

[0010] In some possible implementations, the sample data of the first dataset comes from occupant response data verified by physical collision tests or finite element simulations under standard sitting postures; the sample data of the second dataset comes from finite element simulation data under large-angle seat conditions where the backrest angle is greater than the angle threshold; the number of samples in the second dataset is less than the number of samples in the first dataset.

[0011] In some possible implementations, the first training uses a composite loss function to optimize the first deep neural network model; the composite loss function includes task loss and domain alignment loss; the task loss is calculated based on the difference between the predicted and actual values ​​of the occupant injury index output by the first deep neural network model; the domain alignment loss is calculated based on the distribution difference between the first dataset and the second dataset in the feature space, and is used to drive the shared feature extraction network to extract domain-invariant features.

[0012] In some possible implementations, a shared feature extraction network is trained to extract domain-invariant features between standard sitting posture and large-angle seat conditions, and a task prediction network is used to map the domain-invariant features to predicted values ​​of occupant injury indicators.

[0013] In some possible implementations, occupant injury indicators include at least one of the following: head injury criteria, chest compression, chest viscosity index, lumbar spine stress, and pelvic acceleration.

[0014] In some possible implementations, after obtaining the occupant injury prediction results, the following is also included:

[0015] By performing forward inference on the injury prediction model, multiple predicted values ​​corresponding to the injury indicators of each occupant are obtained.

[0016] For each occupant injury index, the variance of the occupant injury index is determined based on the multiple predicted values ​​corresponding to that occupant injury index.

[0017] Based on the variance, determine the confidence level corresponding to each occupant injury index.

[0018] Secondly, this application provides a vehicle collision occupant injury prediction device, comprising:

[0019] The acquisition module is used to acquire the first dataset, the second dataset, and the operating parameters of the target large-angle seat; wherein, the first dataset is occupant collision data under standard sitting posture conditions, the second dataset is occupant collision data under large-angle seat conditions, and the operating parameters include at least one of the following: seat back angle, collision pulse waveform, seat belt parameters, and airbag parameters.

[0020] The first training module is used to train the first deep neural network model for the first time based on the first dataset to obtain the second training model; wherein the second training model includes a shared feature extraction network and a task prediction network;

[0021] The second training module is used to train the parameters of the task prediction network a second time based on the second dataset, with the parameters of the shared feature extraction network frozen, to obtain the damage prediction model.

[0022] The prediction module is used to input the working condition parameters of the target large-angle seat into the damage prediction model to obtain the occupant damage prediction results; among which, the occupant damage prediction results include the predicted values ​​corresponding to the occupant damage indicators.

[0023] Thirdly, this application provides a computing device, including a memory and a processor;

[0024] The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of the first aspects.

[0025] Fourthly, this application provides a computer-readable storage medium for storing a computer program for performing the method as described in any one of the first aspects.

[0026] Fifthly, this application provides a computer program product comprising one or more computer instructions, wherein when the computer instructions are executed by a computer, the computer performs the method as described in any one of the first aspects.

[0027] As can be seen from the above technical solution, this application has at least the following beneficial effects:

[0028] In this application, by acquiring the first dataset, the second dataset, and the working parameters of the target large-angle seat, the high cost of large-angle seat experiments is fully considered. By using simulation data as the primary source and experimental data as a supplement, the pain point of scarce target domain data is addressed, and the authenticity of the data is ensured through experimental verification, providing high-quality data support for subsequent model training and prediction. Furthermore, based on the first dataset, the first deep neural network model can be trained for the first time to obtain a second training model including a shared feature extraction network and a task prediction network. Sufficient data allows the model to avoid training instability caused by random initial parameters, improving the model's basic predictive ability and convergence speed. This approach preserves transferable knowledge for subsequent second training, avoiding the risk of model overfitting in small-sample scenarios. Furthermore, with the parameters of the shared feature extraction network frozen, the parameters of the task prediction network are trained a second time based on the second dataset to obtain the damage prediction model. This allows for performance improvement with only a small amount of the second dataset, reducing reliance on its massive size while ensuring the accuracy and stability of the damage prediction model under large-angle seat conditions. By inputting the operating parameters of the target large-angle seat into the damage prediction model, occupant injury prediction results are obtained, enabling virtual assessment of the safety performance of large-angle seats without the need for expensive physical collision tests. This solution, by introducing both the first and second datasets, ensures data authenticity and sufficiency. Moreover, the inclusion of both the first and second training sessions guarantees the accuracy and stability of the damage prediction model, ultimately achieving accurate prediction of occupant injury.

[0029] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0030] Figure 1 An application environment diagram for a method for predicting occupant injuries in a car collision provided in this application embodiment;

[0031] Figure 2A flowchart illustrating a method for predicting occupant injuries in a car collision, provided as an embodiment of this application;

[0032] Figure 3 A structural diagram of a vehicle collision occupant injury prediction device provided in an embodiment of this application;

[0033] Figure 4 This is a schematic diagram of a computing device provided in an embodiment of this application. Detailed Implementation

[0034] The terms "first," "second," and "third," etc., used in this application specification and accompanying drawings are used to distinguish different objects, not to limit a specific order.

[0035] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0036] To ensure clarity and conciseness in the description of the following embodiments, a brief introduction to the related technologies is given first:

[0037] Currently, with the rapid development of smart cockpit technology, large-angle seats (such as zero-gravity seats and reclining seats) have received increasing attention due to their ability to provide a more comfortable driving experience. These seats typically have backrest angles exceeding 45°, with some even reaching over 60°. Their unique seating posture results in a significant difference in the occupant's dynamic response during a collision compared to the traditional standard sitting posture. Therefore, accurately predicting occupant injury in large-angle seat scenarios has become a crucial issue for ensuring vehicle passive safety performance and promoting the safety upgrade of smart cockpits.

[0038] Currently, occupant collision injury prediction mainly relies on two technical approaches: one is the traditional method based on physical experiments, which uses physical dummies such as Hybrid III and THOR to conduct real-vehicle collision or slide tests, collecting injury data such as head injury criteria and chest compression. The other is the simulation prediction method based on computer-aided engineering, which uses software such as LS-DYNA and MADYMO to build finite element or multibody dynamics models, and combines them with human models such as THUMS and GHBMC for parametric simulation, generating a large amount of injury-related data for model training. In addition, some solutions employ machine learning models such as support vector regression and traditional neural networks to establish injury prediction relationships based on existing experimental or simulation data, assisting in safety performance assessment.

[0039] However, existing technologies face numerous limitations in high-angle seating scenarios, hindering accurate prediction of occupant injuries. Firstly, real-world crash tests for high-angle seating are extremely costly and subject to strict ethical constraints. Furthermore, simulation data consumes significant computational resources, resulting in scarce target domain data (typically less than 5% of standard seating data). Direct model training is prone to severe overfitting and poor generalization. Secondly, traditional prediction models do not consider the data distribution differences between standard and high-angle seating positions, relying solely on single-domain data for training. This prevents knowledge transfer across different scenarios, leading to a significant drop in prediction accuracy in high-angle scenarios. Thirdly, traditional data augmentation methods only increase the amount of data, failing to address the fundamental distribution differences between the source and target domains and lacking effective domain adaptation mechanisms, making it difficult to adapt to the unique dynamic response patterns of occupants in high-angle seating situations. Fourthly, some models do not quantify the uncertainty of prediction results, failing to provide reliable confidence levels for engineering decisions. These combined problems make existing technologies insufficient to meet the demand for accurate injury prediction in high-angle seating scenarios, necessitating a method capable of accurately predicting occupant injuries.

[0040] In view of this, embodiments of this application provide a method for predicting occupant injuries in a car collision. To make the technical solution of this application clearer and easier to understand, the application scenarios of the technical solution are described below with reference to the accompanying drawings. Figure 1 As shown, this figure is an application environment diagram provided by an embodiment of this application.

[0041] In this application environment, server 104 receives a first dataset, a second dataset, and operational parameters of the target large-angle seat uploaded by terminal 102. The first dataset contains occupant collision data under standard seating conditions, and the second dataset contains occupant collision data under large-angle seat conditions. The operational parameters include at least one of the following: seat back angle, collision pulse waveform, seatbelt parameters, and airbag parameters. After confirming receipt, server 104 sends a successful reception message to terminal 102. Subsequently, server 104 performs its first training on a first deep neural network model based on the first dataset, obtaining a second training model that includes a shared feature extraction network and a task prediction network, and synchronizes the training completion status to terminal 102. Next, according to the parameter freeze command issued by terminal 102, server 104 performs a second training on the parameters of the task prediction network based on the second dataset, while fixing the parameters of the shared feature extraction network, generating a damage prediction model. Afterward, server 104 pushes a notification to terminal 102 indicating that the model training is complete, along with performance metrics. Finally, server 104 calls the damage prediction model to calculate the received target large-angle seat condition parameters, obtains the occupant damage prediction result including the predicted value of occupant damage index, and feeds it back to terminal 102 to complete the entire interaction and data processing process.

[0042] To make the technical solution of this application clearer and easier to understand, the following describes a method for predicting occupant injuries in a car collision, based on the above application scenarios. Figure 2 As shown in the figure, this is a flowchart illustrating a method for predicting occupant injuries in a car collision according to an embodiment of this application. The method for predicting occupant injuries in a car collision includes:

[0043] S201. Obtain the operating parameters of the first dataset, the second dataset, and the target large-angle seat.

[0044] The first dataset contains occupant collision data under standard sitting posture conditions, while the second dataset contains occupant collision data under large-angle seat conditions.

[0045] The standard seating posture is an occupant seating posture that conforms to industry standards, with a typical backrest angle of 23°, and is the benchmark condition for automobile crash safety testing.

[0046] The large-angle seat condition is the occupant sitting posture condition where the seat back angle is greater than a preset angle threshold, such as 40°, and the threshold for zero-gravity seats can be 60°; under this condition, the occupant's dynamic response during a collision is significantly different from the standard sitting posture.

[0047] Operating parameters include at least one of the following: seat back angle, collision pulse waveform, seat belt parameters, and airbag parameters. These parameters refer to the important configuration and environmental parameters of the seat at large angles to be evaluated or predicted in a collision scenario, and are the direct input for damage prediction.

[0048] The sample data of the first dataset comes from occupant response data verified by physical collision tests or finite element simulations under standard sitting postures, which can provide occupant dynamic response knowledge that is applicable across working conditions; the sample data of the second dataset comes from finite element simulation data under the working condition of large-angle seats with backrest angles greater than the angle threshold. It belongs to the small sample target domain data and is used to adapt the model to the unique collision response mode of large-angle seats; the number of samples in the second dataset is less than the number of samples in the first dataset.

[0049] Physical crash tests are standardized tests conducted using physical dummies such as Hybrid III and THOR in laboratory sliding stages or real vehicle crash environments to obtain occupant response data under real crash scenarios.

[0050] Finite element simulation verification is based on experimentally verified human body models such as THUMS and GHBMC, and virtual collision analysis is performed using simulation tools such as LS-DYNA and MADYMO, which can generate multi-condition data in batches.

[0051] Occupant response data is data reflecting the risk of injury to occupants during a collision. It includes physical quantities such as acceleration, displacement, and force of dummies or simulation models, and is the basis for extracting damage index labels.

[0052] The angle threshold is the critical backrest angle that distinguishes between a standard sitting posture and a large-angle seat (e.g., set to 40°). If the backrest angle exceeds this value, it is determined to be a large-angle seat condition.

[0053] Seat belt parameters are important configurations related to seat belt performance, including force limiter rating, pretensioner action time and pretension amount, webbing stiffness, etc.

[0054] Airbag parameters are key configurations related to the working characteristics of airbags, including vent size, ignition time, and gas generator mass flow rate.

[0055] The collision pulse waveform is a curve of acceleration changing with time during the collision process. It is usually discretized into a time-acceleration sequence and is an important parameter characterizing the collision intensity and characteristics.

[0056] Optionally, outlier detection and removal, missing value handling, and standardization of all numerical input features after processing can be performed on the original data of the first and second datasets. The mean and standard deviation required for standardization are calculated only from the first dataset and then uniformly applied to the first and second datasets. Finally, feature-level distribution alignment is introduced based on the maximum mean difference method.

[0057] For example, basic training data can be collected through two channels. First, a first dataset is obtained, which is based on standard sitting posture conditions (such as a backrest angle of 23°) and generated through laboratory physical collision tests (using physical dummies) or high-precision finite element simulations (using human simulation models) verified by real tests, ensuring the authenticity and reliability of the data. Second, a second dataset is obtained, which targets large-angle seat conditions with backrest angles greater than a preset angle threshold (such as 40°). Considering the high cost and ethical restrictions of physical tests, it is mainly generated through finite element simulations, and its sample size is much smaller than that of the first dataset, usually 1% to 5% of the first dataset, exhibiting typical small sample characteristics. At the same time, the working parameters of the target large-angle seat to be predicted can be collected. These parameters need to cover at least one of the following: seat backrest angle, collision pulse waveform (such as a curve with a peak value of 35g-45g and a duration of 80-120ms), seat belt parameters (such as force limit level of 2kN-8kN), and airbag parameters (such as vent area and ignition time), to ensure that the parameters can comprehensively characterize the target collision scenario. The entire acquisition process needs to take into account both the integrity and adaptability of the data. The first dataset provides a general knowledge base for the model, the second dataset is used for the model's target domain adaptation, and the working condition parameters provide specific scenario inputs for the final damage prediction. The three together support high-precision prediction of large-angle seat occupant damage under small sample conditions.

[0058] It should be noted that the domain distribution alignment of the data in the first and second datasets can be achieved using Maximum Mean Discrepancy (MMD). This involves measuring the difference in probability distributions between the first and second datasets in the reproducing kernel Hilbert space, and then using a Gaussian kernel function to map high-dimensional features to a more separable space, thus achieving effective alignment of cross-domain feature distributions. The alignment expression can be represented as:

[0059]

[0060] in, For the first dataset, For the second dataset, The square of the MMD distance between the first and second datasets. The number of samples in the first dataset. The number of samples in the second dataset. For the first dataset, the first One sample, For the first dataset, the first One sample, For the second dataset, the first One sample, For the second dataset, the first One sample, A kernel function (such as a Gaussian kernel) is used to calculate the similarity between two samples in the RKHS. This represents the average similarity between samples in the first dataset. This represents the negative twice mean of the similarity between the first and second datasets. This represents the average similarity between samples in the second dataset.

[0061] S202. Based on the first dataset, the first deep neural network model is trained for the first time to obtain the second training model.

[0062] Among them, the first deep neural network model is an initial network model tailored for domain adaptive transfer learning. It has the ability to transfer knowledge across working conditions and serves as the basic architecture for subsequent training. It includes a shared feature extraction network, a task prediction network, and a domain discriminator.

[0063] The second training model includes a shared feature extraction network and a task prediction network. It is an intermediate model obtained after the first training. It retains the parameters of the trained shared feature extraction network and task prediction network and has the basic ability to extract domain-invariant features and map impairment indicators. It can be directly used for subsequent fine-tuning of the target domain.

[0064] The first training refers to domain adversarial pre-training based on the first dataset. Its goal is to enable the model to learn the common features of standard sitting posture and large-angle seat conditions, while simultaneously reducing the data distribution differences between the two domains, thus laying the groundwork for subsequent target domain adaptation. The process of the first training can be represented as follows: Where X is the preprocessed input feature vector, This is the final output of the entire model, representing the prediction result obtained after the input feature vector X has been processed by the complete model. To share the feature extraction layer, Output layer for specific tasks.

[0065] The first training iteration uses a composite loss function to optimize the first deep neural network model. The composite loss function includes task loss, domain classification loss, and domain alignment loss. The task loss is calculated based on the difference between the predicted and actual values ​​of the occupant injury index output by the first deep neural network model.

[0066] The shared feature extraction network is trained to extract domain-invariant features between standard sitting posture and large-angle seat conditions. It is an important component of the second training model and is responsible for extracting low-level features that are common across different conditions (such as energy absorption and load transfer path) from the input features. It is unaffected by differences in sitting posture and is the key to achieving knowledge transfer.

[0067] The task prediction network is used to map domain-invariant features to predicted values ​​of occupant injury indicators. It is the output layer component of the second training model. It receives the domain-invariant features output by the shared feature extraction network and maps them to specific predicted values ​​of occupant injury indicators, thus completing the transformation between features and injury.

[0068] The composite loss function is the optimization objective function for the first training. It consists of a weighted average of task loss, domain classification loss, and domain alignment loss, taking into account both the model's damage prediction accuracy and cross-domain adaptability, ensuring that the trained model can both learn accurately and transfer knowledge.

[0069] Task loss is an important component of the composite loss function. It is calculated based on the difference between the predicted values ​​of occupant injury indicators output by the model and the true values ​​in the first dataset, and is used to drive the model to improve the accuracy of injury prediction.

[0070] Domain classification loss measures the ability of a domain discriminator to distinguish feature sources (first dataset or second dataset), and can be calculated using binary cross-entropy loss.

[0071] Domain alignment loss is the cross-domain adaptation component of the composite loss function. It is calculated based on the distributional differences between the first and second datasets in the feature space, such as using MMD to calculate the distributional differences. This is used to force the shared feature extraction network to learn domain-invariant features, thereby reducing the data differences between the two domains.

[0072] Occupant injury indices are the outputs that the model needs to predict, including head injury criteria (HIC15), chest compression, chest viscosity index, lumbar spine stress, pelvic acceleration, etc., which are important physical quantities for assessing collision safety.

[0073] Domain-invariant features are those shared by both standard sitting posture and large-angle seat conditions and are unaffected by differences in sitting posture, such as the timing of collision pulses and the energy absorption characteristics of constraint systems. They are important carriers for realizing cross-condition knowledge transfer.

[0074] For example, based on a pre-acquired first dataset (standard sitting posture, 10... 4The model was trained using two datasets: a first dataset (of similar magnitude) and a second dataset (large-angle seats, small sample size). The first dataset served as the primary training data, providing the foundation for the model to learn general knowledge. The second dataset, while not used in the primary training, was used to calculate the domain alignment loss for cross-domain adaptation. Subsequently, the initial deep neural network model was trained for the first time based on the first dataset. This determined the parameters of the shared feature extraction network and the task prediction network under the first dataset. This first training was essentially domain adversarial pre-training. During the training process, a composite loss function was used for optimization. On the one hand, the task loss, based on the difference between the predicted value and the true value in the first dataset, drove the model to improve the prediction accuracy of occupant injury indicators, ensuring accurate model learning. On the other hand, the domain alignment loss, based on the feature distribution differences between the first and second datasets, such as the MMD (Maximum Mean Discrepancy) distance, forced the shared feature extraction network to focus on learning domain-invariant features, such as energy transfer paths and constraint system response patterns, ignoring the interference caused by posture differences. This was further driven by the domain classification loss. (·) Generate domain-invariant features to ensure that the model can be transferred.

[0075] The goal of the entire training process is to obtain a second training model, which consists of two main components: first, a well-trained shared feature extraction network that can stably extract domain-invariant features common to both standard sitting posture and large-angle seat conditions; and second, a well-adaptive task prediction network that can accurately map the extracted domain-invariant features to predicted values ​​of occupant injury indicators such as head injury criteria and chest compression, laying the foundation for subsequent fine-tuning based on a small sample second dataset, and ultimately achieving the goal of using sufficient source domain knowledge to compensate for the scarcity of target domain data.

[0076] Optionally, the shared feature extraction network can employ a parallel multi-branch structure. Each branch typically contains sequential one-dimensional convolutional layers (1D-CNN), batch normalization layers, modified linear unit activation functions (ReLU), and max pooling layers. The feature maps extracted by each branch are adaptively fused using a channel attention mechanism (such as the SENet module) or simple channel concatenation to form a unified domain-invariant feature Z. Each branch uses convolutional kernels of different sizes (3, 5, 7) to extract local, mid-range, and global temporal features of the collision pulse signal.

[0077] Specifically, a multi-layer perceptron (MLP) or a one-dimensional convolutional neural network can be used: The MLP scheme consists of 3 to 5 fully connected layers, with the number of neurons in each layer decreasing (e.g., [input dimension, 128, 64, 32]). Each fully connected layer is followed by a ReLU activation function and a dropout layer (with a dropout rate of 0.2~0.5) to prevent overfitting. The 1D-CNN scheme is particularly suitable for processing temporal signal features such as collision pulses. It contains several one-dimensional convolutional layers and pooling layers to extract local temporal patterns, and finally connects to the fully connected layers through a flattening layer. Each branch contains a convolutional layer, a batch normalization layer, a ReLU activation function, and a max-pooling layer. Finally, a channel attention mechanism is used to achieve adaptive fusion of multi-scale features.

[0078] Specifically, a gradient reversal layer is introduced after the shared feature extraction layer. When the gradient is passed to the reversal layer, it is processed with negative weights, and the processed gradient is backpropagated to the feature extraction layer. The formula can be expressed as:

[0079]

[0080] in, For the gradient to be passed to the previous layer, I The gradient received from the subsequent layer is represented by λ, which is the gradient inversion weight. λ=0 indicates no gradient propagation, meaning the feature extractor does not perform domain adaptation; λ>0 indicates gradient inversion, with larger λ values ​​indicating stronger inversion. This design enables the feature extractor to learn domain-invariant features that are confusing to the domain discriminator, thus effectively improving the model's domain adaptation capability. The shared layer outputs a low-dimensional, high-level general feature vector Z=G. I (X1); X1 is the input data, G I (·) for The mapping function of the feature extractor Z The extracted domain-invariant features are used for main tasks (such as classification) and domain discrimination tasks.

[0081] The task prediction network maps the general feature vector Z to the specific damage index prediction value. It is responsible for learning how to adapt the general features to the damage response of a specific working condition (especially the target domain). It usually consists of 1 to 2 fully connected layers. For multi-task learning (predicting multiple occupant damage indices at the same time), a multi-branch output structure can be adopted, with each branch corresponding to a damage index and sharing the features preceding it.

[0082] During the initial training of the first deep neural network model, domain adversarial pre-training is performed by optimizing a composite objective function that includes task loss, domain classification loss, and domain alignment loss.

[0083]

[0084] in, For the total loss, For mission losses, For domain classification loss, Let be the domain alignment loss, where a represents the task loss weight, b represents the domain classification loss weight, and c represents the domain alignment loss weight, and 1 = a + b + c.

[0085] The task loss uses smoothed L1 loss, which is insensitive to outliers (mean squared error loss can be used for regression tasks), the domain classification loss uses cross-entropy loss, and the domain alignment loss is the MMD distance.

[0086] 1) Task loss (L) task ): Using smooth L1 The loss function, when the error is small, behaves as mean squared error to ensure gradient stability, and when the error is large, it behaves as absolute value error to resist outlier interference.

[0087]

[0088] in, Task loss is a part of the composite loss function, used to measure the difference between the model's predicted value and the true value, driving the model to improve prediction accuracy; The smoothing L1 loss function is a loss function that is more robust to outliers. It behaves as L2 loss (smoothing) when the residuals are small, and as L1 loss (avoiding gradient explosion) when the residuals are large. The number of samples in the first dataset. For the first The true value of each sample For the first The model prediction value for each sample.

[0089] Optional, smooth L1 Loss is defined as:

[0090]

[0091] in, The smoothing coefficient can be set to 1, and x is the input vector.

[0092] 2) Domain classification loss (L) domain Adversarial training is achieved through a gradient inversion layer, and the domain discriminator uses cross-entropy loss.

[0093]

[0094] in, Domain Classification Loss measures the ability of a domain discriminator to distinguish the source of features (the first dataset or the second dataset). The number of samples in the first dataset. The number of samples in the second dataset. For the first The true label of each sample The first output of the model The probability that a sample is classified as category 1.

[0095] This loss is essentially a binary cross-entropy loss, which aims to make the domain discriminator identify, as accurately as possible, whether a sample belongs to the first dataset (i.e., the source domain) or the second dataset (i.e., the target domain). In domain adversarial training, the model's feature extractor will try to minimize this loss in order to confuse the domain discriminator and thus make the extracted features domain invariant.

[0096] 3) Domain alignment loss: Based on the maximum mean difference theory, minimize the feature distribution difference between the first and second datasets.

[0097]

[0098] in, Domain alignment loss (MMD Loss) is the cross-domain adaptation component in the composite loss function. It is used to force the shared feature extraction network to learn domain-invariant features, thereby reducing the data distribution differences between the first and second datasets.

[0099] In adaptive weight adjustment strategies, an adaptive weight adjustment strategy can be adopted:

[0100]

[0101] in, This is the scaling factor, where p represents the training progress, achieving a smooth transition from domain-focused alignment to task-focused learning. The weight configuration scheme is as follows: Early training (p < 0.3): focus on distribution alignment, setting [a = 0.3, b = 0.4, c = 0.3]; Mid-training (0.3 ≤ p < 0.7): balance optimization, calculating the composite objective function... At the initial stage, the task loss weight 'a', the domain classification loss weight 'b', and the domain alignment loss weight 'c' can be set to [a=0.5, b=0.3, c=0.2]. In the later stages of training (p≥0.7): the focus is on task performance, and the composite objective function is calculated... At that time, the task loss weight a, the domain classification loss weight b, and the domain alignment loss weight c can be set to [a=0.7, b=0.2, c=0.1].

[0102] In the training configuration and optimization strategy, the adaptive moment estimation optimizer is used, with an initial learning rate set to 1e-3 and a weight decay coefficient set to 0.01. A learning rate decay strategy is employed, linearly increasing to 1e-3 during the warm-up phase (the first 10 epochs). During the stable training phase (10-400 epochs), cosine annealing scheduling is used, with the learning rate lr calculated as lr = lrmin + 0.5. (lrmax lrmin) (1+cos(π Tcur / Tmax)) decreases, where lrmax=1e 3, lrmin=1e 5. Tcur is the current epoch number, Tmax=390; during the fine-tuning stage (400-500 epochs), it is fixed at 1e-5; the batch size is set according to the amount of data (e.g., 128 or 256); the dropout ratio is 0.3 in the hidden layer and 0.5 in the input layer; the gradient norm threshold is set to 1.0.

[0103] In the training process monitoring and early stopping mechanism, real-time monitoring is carried out. The main monitoring indicators include validation set loss and the changing trend of each loss component. Auxiliary monitoring indicators include feature distribution similarity and domain classification accuracy. An adaptive early stopping mechanism is adopted, such as a base patience value of 15 epochs, automatically extending the patience value when the validation loss enters the plateau period, and saving the model parameters with the lowest validation loss.

[0104] After pre-training, a comprehensive evaluation was performed on the first dataset. Through the refined pre-training strategy described above, the model successfully learned general feature representations across different operating conditions, laying a solid foundation for the subsequent fine-tuning stage. This design ensures the robustness of the feature representations while providing an effective feature space for domain adaptation, thus improving the prediction performance under large-angle seating conditions.

[0105] S203. With the parameters of the shared feature extraction network frozen, the parameters of the task prediction network are trained a second time based on the second dataset to obtain the damage prediction model.

[0106] Among them, parameter freezing is an important operation in model training. It refers to fixing the existing parameters of the shared feature extraction network so that they are no longer updated in subsequent training. The purpose is to preserve the general feature extraction capabilities learned in the pre-training stage and avoid forgetting the source domain knowledge.

[0107] The second training phase, also known as the fine-tuning phase, is a targeted training process based on a small sample second dataset. The goal is to adapt the task prediction network to the unique dynamic response of large-angle seats and improve the accuracy of target domain damage prediction while retaining general features.

[0108] The damage prediction model is the final usable model obtained after the second training. It integrates general knowledge from the source domain with the unique rules of the target domain. It can receive input parameters of seat conditions at large angles and output accurate prediction results of occupant damage indicators.

[0109] For example, based on two pre-completed foundational preparations (obtaining a mature shared feature extraction network capable of extracting cross-condition domain invariant features through pre-training and acquiring a second dataset), a second training (fine-tuning) is initiated. First, a parameter freeze operation can be performed to fix all parameters of the shared feature extraction network, preventing modification during subsequent training. This preserves the general knowledge learned from the massive source domain data (standard sitting posture) during pre-training, such as collision load transfer and the underlying laws of occupant dynamic response, avoiding knowledge forgetting due to the scarcity of target domain data. Next, using the second dataset as the primary training data, only the parameters of the task prediction network are updated and optimized. During training, the shared feature extraction network continuously outputs domain-invariant features, while the task prediction network focuses on learning the mapping relationship between general features and large-angle seat damage indicators, adapting to the unique dynamic responses of occupants under large-angle seats (such as differences in force distribution caused by body posture). In terms of training configuration, a very small learning rate, a small batch size (e.g., 32), and early stopping can be used to avoid overfitting due to small sample data.

[0110] The goal of the entire process is to achieve accurate model adaptation to large-angle seat conditions by freezing the general feature layer and fine-tuning the task output layer, while making full use of source domain knowledge. This results in a damage prediction model that balances versatility and specificity, thus solving the problem of insufficient model accuracy under small sample target domain data.

[0111] It should be noted that the goal of this step is to adapt the model to the unique dynamics of large-angle seating conditions, while avoiding overfitting due to the scarcity of data in the target domain. To achieve this goal, a systematic parameter freezing and progressive unfreezing strategy is adopted. First, the parameters of the pre-trained shared feature extraction network are completely frozen to ensure that they are not updated in subsequent training. This measure effectively prevents the forgetting of knowledge from the first dataset and forces the model to fully utilize the learned general feature representations as a foundation.

[0112] Building upon this foundation, a phased, refined training strategy was implemented. In the initial phase, only the second dataset was used to retrain the parameters of the task prediction network, focusing on optimizing the mapping relationship from general features to specific damage indicators. Subsequently, a layered, gradual unfreezing strategy was adopted, gradually releasing the parameters of high-level feature layers for targeted fine-tuning. Finally, full parameter fine-tuning was performed with a very small learning rate, achieving a comprehensive improvement in the model's capabilities.

[0113] The training configuration uses a significantly reduced learning rate (10). -4 Up to 10 -5 (On a larger scale), effectively controlling the magnitude of parameter updates and avoiding overfitting on limited data. Simultaneously, the training cycle is shortened accordingly, and early stopping is implemented by closely monitoring the performance of the second dataset validation set, terminating training promptly when the model performance reaches its optimal level.

[0114] This training strategy preserves the learning outcomes for general feature representations while adapting to the characteristics of the second dataset. It protects knowledge from the first dataset through parameter freezing, achieves a smooth transition through gradual unfreezing, and ensures generalization performance through regularization constraints, ultimately achieving excellent prediction accuracy and stability on the scarce second dataset.

[0115] S204. Input the working parameters of the target large-angle seat into the damage prediction model to obtain the occupant damage prediction results.

[0116] Among them, the occupant injury prediction results include the predicted values ​​corresponding to the occupant injury indicators. These are quantitative results reflecting the risk of occupant injury in a collision, and are the specific predicted values ​​of each occupant injury indicator, providing a direct decision-making basis for the optimization of seats and restraint systems.

[0117] Occupant injury indicators include at least one of the following: head injury criteria, chest compression, chest viscosity index, lumbar spine stress, and pelvic acceleration. These are key physical quantities for assessing crash safety and are used to quantify the degree of injury to critical parts of the occupant's head, chest, lumbar spine, and pelvis.

[0118] The head injury criterion is an injury index calculated based on the change of head acceleration over time. It focuses on reflecting the impact injury risk to the head during a collision and is one of the indicators for collision safety assessment.

[0119] Chest compression is the maximum amount of compression deformation of an occupant's chest during a collision. It is directly related to the risk of injury to the thoracic bones and internal organs, and is an important parameter for assessing chest protection performance.

[0120] The chest viscosity index is a comprehensive injury indicator that combines chest compression amount and compression velocity, providing a more complete reflection of the dynamic damage to the chest during a collision.

[0121] Lumbar spine stress is the longitudinal force exerted on the lumbar spine of an occupant during a collision. It is used to assess the risk of injury such as fractures and dislocations of the lumbar spine, and is especially suitable for assessing the stress characteristics of the lumbar spine under large-angle seats.

[0122] Pelvic acceleration is the peak acceleration of the occupant's pelvis during a collision. It reflects the risk of impact injury to the pelvis and lower limbs and is an important supplementary indicator for a comprehensive assessment of occupant safety.

[0123] For example, precise acquisition of the operating parameters of the target large-angle seat can be achieved through engineering design documents, simulation model configuration, or actual measurement data collection. These parameters include seat posture parameters (such as backrest angle and H-point position), collision condition parameters (such as collision pulse waveform and equivalent collision velocity), seat belt parameters (such as force limiter level and pretensioner action time), and airbag parameters (such as vent size and ignition time), ensuring that the parameters comprehensively and realistically recreate the target collision scenario. Furthermore, the processed operating parameters are input into the trained damage prediction model according to the required format (such as 52-dimensional feature vector and JSON format). This model has been tested using the first dataset (standard sitting posture, 10...). 4 The model was pre-trained with a dataset of 10¹ to 10² magnitudes to master general collision dynamics knowledge. It was then fine-tuned with a second dataset (large-angle seats, 10¹ to 10² magnitude data) to adapt to a unique response mode, enabling it to quickly process input parameters and complete calculations. Finally, the damage prediction model outputs the prediction results of at least one occupant injury index, including head injury criteria, chest compression, chest viscosity index, lumbar force, and pelvic acceleration. These are the specific quantitative values ​​corresponding to each index, such as HIC15=856 and lumbar force=3.1kN.

[0124] The entire process does not rely on time-consuming finite element simulation (traditional single-condition simulation takes 48 hours) or expensive physical tests. It can complete the prediction in just a few seconds. Engineers can directly judge the safety of the current seat and restraint system design based on the prediction results, quickly iterate and optimize parameters, significantly shorten the development cycle and reduce R&D costs.

[0125] Furthermore, forward reasoning is performed on the injury prediction model to obtain multiple predicted values ​​corresponding to each occupant injury index; for each occupant injury index, the variance of the occupant injury index is determined based on the multiple predicted values ​​corresponding to that occupant injury index; based on the variance, the confidence level corresponding to each occupant injury index is determined.

[0126] Forward inference refers to the process of inputting input parameters (such as the operating parameters of a large-angle seat) into the model, and outputting the prediction result through feature operations and mapping of each layer of the neural network, without the need for backpropagation to update parameters.

[0127] Multiple predictions are obtained by using the Monte Carlo Dropout method, which randomly discards some neurons by repeatedly enabling the Dropout layer during the model inference stage. The same set of input parameters is subjected to forward inference a preset number of times (e.g., 100 times), resulting in a preset number of different prediction results for the same damage index, which are used to characterize the prediction uncertainty.

[0128] Variance is a statistical measure of the dispersion of data. It is calculated from multiple predicted values ​​of the same damage index. The larger the variance, the greater the fluctuation and uncertainty of the prediction results, and vice versa.

[0129] Confidence levels are prediction reliability ratings based on variance partitioning. They can be divided into three levels: high confidence, medium confidence, and low confidence. These levels are used to intuitively inform engineers of the reliability of the prediction results and support engineering decisions.

[0130] For example, based on completing two basic preparations in advance (acquiring a well-trained injury prediction model and obtaining the working parameters of the target large-angle seat), multiple rounds of forward inference are then initiated. This can be achieved using the Monte Carlo Dropout method, keeping the Dropout layer enabled during inference (used to prevent overfitting during training and to quantify uncertainty during inference). The same set of working parameters is repeatedly inferred multiple times (e.g., 100 times). Each inference produces slight differences due to the random dropping of neurons, ultimately yielding multiple predicted values ​​for each occupant injury indicator (e.g., HIC15, lumbar spine stress, etc.). Next, for each occupant injury indicator, the variance is calculated based on its multiple predicted values. This can be solved using statistical formulas. The variance directly reflects the dispersion of the prediction results; a smaller variance indicates a more concentrated set of inference results and a more reliable prediction. Finally, confidence levels are assigned based on the variance. A score of 0.05 corresponds to a high confidence level, indicating that the prediction results are reliable and can be directly used for design decisions. variance 0.1 corresponds to a medium confidence level, indicating that engineering experience is needed to assist in the judgment; variance... A score of 0.1 indicates low confidence; it is recommended to supplement with a second dataset or optimize the parameters and then re-predict.

[0131] For example, integrating Monte Carlo Dropout methods for uncertainty quantification. The mean and variance of the prediction results are calculated through T forward propagations:

[0132]

[0133] Where E is the mean of the prediction results, T is the number of propagations, and t is the current number of propagations. (X) represents the predicted value of the prediction model, and Var represents the variance of the prediction result.

[0134] The entire process does not require additional physical experiments or time-consuming simulations. It can achieve predictive reliability assessment through multiple rounds of reasoning and statistical calculations. This provides engineers with accurate damage prediction results and clarifies the credibility boundaries of the results, avoiding design risks caused by blindly relying on model predictions and greatly improving the scientific nature and safety of engineering decisions.

[0135] Based on the above, the vehicle collision occupant injury prediction method, by acquiring the first dataset, the second dataset, and the working parameters of the target large-angle seat, fully considers the high cost of large-angle seat experiments. By primarily using simulation data and supplementing it with experimental data, it addresses the pain point of scarce data in the target domain and ensures data authenticity through experimental verification, providing high-quality data support for subsequent model training and prediction. Furthermore, based on the first dataset, the first deep neural network model can be trained for the first time to obtain a second training model including a shared feature extraction network and a task prediction network. Sufficient data allows the model to avoid training instability caused by random initial parameters, improving the model's basic predictive ability. While improving convergence speed, this approach also preserves transferable knowledge for subsequent second training, avoiding the risk of overfitting in small sample scenarios. Furthermore, with the parameters of the shared feature extraction network frozen, the parameters of the task prediction network are trained a second time based on the second dataset to obtain the damage prediction model. This allows for performance improvement with only a small amount of the second dataset, reducing reliance on its massive size while ensuring the accuracy and stability of the damage prediction model under large-angle seat conditions. By inputting the operating parameters of the target large-angle seat into the damage prediction model, occupant injury prediction results are obtained, enabling virtual assessment of the safety performance of large-angle seats without the need for expensive physical collision tests. This solution, by introducing both the first and second datasets, ensures data authenticity and sufficiency. Moreover, the inclusion of both the first and second training sessions guarantees the accuracy and stability of the damage prediction model, ultimately achieving accurate prediction of occupant injury.

[0136] The above text combined Figures 1 to 2 The method for predicting occupant injury in a car collision provided in this application has been described in detail. The apparatus and equipment provided in this application will be described below with reference to the accompanying drawings.

[0137] This application also provides a vehicle collision occupant injury prediction device, such as... Figure 3 As shown in the figure, this is a schematic diagram of a vehicle collision occupant injury prediction device 300 provided in an embodiment of this application. The device includes:

[0138] The acquisition module 301 is used to acquire the first dataset, the second dataset, and the working condition parameters of the target large-angle seat; wherein, the first dataset is occupant collision data under standard sitting posture conditions, the second dataset is occupant collision data under large-angle seat conditions, and the working condition parameters include at least one of the following: seat back angle, collision pulse waveform, seat belt parameters, and airbag parameters.

[0139] The first training module 302 is used to perform a first training on the first deep neural network model based on the first dataset to obtain a second training model; wherein the second training model includes a shared feature extraction network and a task prediction network;

[0140] The second training module 303 is used to train the parameters of the task prediction network a second time based on the second dataset when the parameters of the shared feature extraction network are frozen, so as to obtain the damage prediction model.

[0141] The prediction module 304 is used to input the working condition parameters of the target large-angle seat into the damage prediction model to obtain the occupant damage prediction result; wherein, the occupant damage prediction result includes the predicted value corresponding to the occupant damage index.

[0142] In some possible implementations, the sample data of the first dataset comes from occupant response data verified by physical collision tests or finite element simulations under standard sitting postures; the sample data of the second dataset comes from finite element simulation data under large-angle seat conditions where the backrest angle is greater than the angle threshold; the number of samples in the second dataset is less than the number of samples in the first dataset.

[0143] In some possible implementations, the first training uses a composite loss function to optimize the first deep neural network model; the composite loss function includes task loss and domain alignment loss; the task loss is calculated based on the difference between the predicted and actual values ​​of the occupant injury index output by the first deep neural network model; the domain alignment loss is calculated based on the distribution difference between the first dataset and the second dataset in the feature space, and is used to drive the shared feature extraction network to extract domain-invariant features.

[0144] In some possible implementations, a shared feature extraction network is trained to extract domain-invariant features between standard sitting posture and large-angle seat conditions, and a task prediction network is used to map the domain-invariant features to predicted values ​​of occupant injury indicators.

[0145] In some possible implementations, occupant injury indicators include at least one of the following: head injury criteria, chest compression, chest viscosity index, lumbar spine stress, and pelvic acceleration.

[0146] In some possible implementations, the vehicle collision occupant injury prediction device 300 also includes:

[0147] The inference module is used to perform forward inference on the injury prediction model to obtain multiple predicted values ​​corresponding to each occupant injury index; for each occupant injury index, the variance of the occupant injury index is determined based on the multiple predicted values ​​corresponding to the occupant injury index; and the confidence level corresponding to each occupant injury index is determined based on the variance.

[0148] The vehicle collision occupant injury prediction device according to the embodiments of this application can correspond to the execution of the method described in the embodiments of this application, and the other operations and / or functions of each module / unit of the vehicle collision occupant injury prediction device are respectively for the purpose of implementing Figure 2 For the sake of brevity, the corresponding processes of each method in the illustrated embodiments will not be described in detail here.

[0149] This application also provides a computing device. For example... Figure 4 As shown in the figure, this is a schematic diagram of a computing device provided in an embodiment of this application. The computing device 400 includes a bus 401, a processor 402, a communication interface 403, and a memory 404. The processor 402, the memory 404, and the communication interface 403 communicate with each other via the bus 401.

[0150] Bus 401 can be a standard bus for interconnecting peripheral components or an extended industry standard structure bus, etc. Buses can be divided into address buses, data buses, control buses, etc. For ease of representation, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0151] Processor 402 can be any one or more of processors such as central processing unit, graphics processing unit, microprocessor or digital signal processor.

[0152] Communication interface 403 is used for communication with external devices.

[0153] Memory 404 may include volatile memory, such as random access memory. Memory 404 may also include non-volatile memory, such as read-only memory, flash memory, hard disk drive, or solid-state drive.

[0154] The memory 404 stores executable code, and the processor 402 executes the executable code to perform the aforementioned vehicle collision occupant injury prediction method.

[0155] Specifically, in achieving Figure 3 In the case of the illustrated embodiment, and Figure 3 When the modules or units of the vehicle collision occupant injury prediction device described in the embodiments are implemented by software, the following steps are performed: Figure 3 The software or program code required for the functions of each module / unit can be partially or entirely stored in memory 404. Processor 402 executes the program code corresponding to each unit stored in memory 404 to execute the aforementioned vehicle collision occupant injury prediction method.

[0156] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium capable of being stored by a computing device, or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct a computing device to execute the aforementioned method for predicting occupant injuries in a car collision.

[0157] This application also provides a computer program product comprising one or more computer instructions. When the computer instructions are loaded and executed on a computing device, all or part of the processes or functions described in this application are generated.

[0158] The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, or data center to another website, computer, or data center via wired (e.g., coaxial cable, fiber optic) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0159] When the computer program product is executed by a computer, the computer performs any of the aforementioned methods for predicting occupant injuries in a car collision. The computer program product can be a software installation package; when any of the aforementioned methods for predicting occupant injuries in a car collision needs to be used, the computer program product can be downloaded and executed on the computer.

[0160] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0161] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be covered within the scope of protection of this application.

Claims

1. A method for predicting occupant injuries in a car collision, characterized in that, The method includes: Acquire a first dataset, a second dataset, and the operating parameters of the target large-angle seat; wherein, the first dataset is occupant collision data under standard sitting posture conditions, and the second dataset is occupant collision data under large-angle seat conditions, and the operating parameters include at least one of seat back angle, collision pulse waveform, seat belt parameters, and airbag parameters; the sample data of the first dataset comes from occupant response data verified by physical collision tests or finite element simulations under standard sitting posture conditions; the sample data of the second dataset comes from finite element simulation data under large-angle seat conditions where the backrest angle is greater than the angle threshold; the number of samples in the second dataset is less than the number of samples in the first dataset. Based on the first dataset, a first deep neural network model is trained for the first time to obtain a second training model. The second training model includes a shared feature extraction network and a task prediction network. The first training uses a composite loss function to optimize the first deep neural network model. The composite loss function includes task loss and domain alignment loss. The task loss is calculated based on the difference between the predicted and actual values ​​of the occupant injury index output by the first deep neural network model. The domain alignment loss is calculated based on the distribution differences between the first and second datasets in the feature space, and is used to drive the shared feature extraction network to extract domain-invariant features. The shared feature extraction network is trained to extract domain-invariant features between standard sitting posture and large-angle seat conditions. The task prediction network maps the domain-invariant features to the predicted values ​​of the occupant injury index. With the parameters of the shared feature extraction network frozen, the parameters of the task prediction network are trained a second time based on the second dataset to obtain the damage prediction model. The working parameters of the target large-angle seat are input into the damage prediction model to obtain the occupant damage prediction result; wherein, the occupant damage prediction result includes the predicted value corresponding to the occupant damage index. After obtaining the occupant injury prediction results, the following is also included: The damage prediction model is forward inferenced to obtain multiple predicted values ​​corresponding to each occupant's damage index; For each occupant injury index, the variance of the occupant injury index is determined based on the multiple predicted values ​​corresponding to that occupant injury index. Based on the variance, determine the confidence level corresponding to each occupant injury index.

2. The method according to claim 1, characterized in that, The occupant injury indicators include at least one of the following: head injury criteria, chest compression, chest viscosity index, lumbar spine stress, and pelvic acceleration.

3. A vehicle collision occupant injury prediction device, characterized in that, The device includes: The acquisition module is used to acquire a first dataset, a second dataset, and the operating parameters of the target large-angle seat. The first dataset contains occupant collision data under standard seating conditions, and the second dataset contains occupant collision data under large-angle seat conditions. The operating parameters include at least one of the following: seat back angle, collision pulse waveform, seatbelt parameters, and airbag parameters. The sample data of the first dataset originates from occupant response data verified by physical collision tests or finite element simulations under standard seating conditions. The sample data of the second dataset originates from finite element simulation data under large-angle seat conditions where the backrest angle exceeds a threshold. The number of samples in the second dataset is less than the number of samples in the first dataset. A first training module is used to perform a first training on a first deep neural network model based on the first dataset to obtain a second training model. The second training model includes a shared feature extraction network and a task prediction network. The first training uses a composite loss function to optimize the first deep neural network model. The composite loss function includes a task loss and a domain alignment loss. The task loss is calculated based on the difference between the predicted and actual values ​​of the occupant injury index output by the first deep neural network model. The domain alignment loss is calculated based on the distribution differences between the first and second datasets in the feature space, and is used to drive the shared feature extraction network to extract domain-invariant features. The shared feature extraction network is trained to extract domain-invariant features between standard sitting posture and large-angle seat conditions. The task prediction network maps the domain-invariant features to the predicted values ​​of the occupant injury index. The second training module is used to train the parameters of the task prediction network a second time based on the second dataset when the parameters of the shared feature extraction network are frozen, so as to obtain the damage prediction model. The prediction module is used to input the working condition parameters of the target large-angle seat into the damage prediction model to obtain occupant damage prediction results. The occupant damage prediction results include predicted values ​​corresponding to occupant damage indicators. After obtaining the occupant damage prediction results, the module further includes: performing forward inference on the damage prediction model to obtain multiple predicted values ​​corresponding to each occupant damage indicator; for each occupant damage indicator, determining the variance of the occupant damage indicator based on the multiple predicted values; and determining the confidence level corresponding to each occupant damage indicator based on the variance.

4. A computing device, characterized in that, Including memory and processor; The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of claims 1 to 2.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program for performing the method as described in any one of claims 1 to 2.

6. A computer program product, characterized in that, The computer program product includes one or more computer instructions, which, when executed by a computer, perform the method as described in any one of claims 1 to 2.

Images