Passenger injury prediction model training method and device, passenger injury prediction method and device and electronic equipment

Abstract

The invention relates to a passenger injury prediction model training method, prediction method and device and electronic equipment, and the method comprises the steps: obtaining a training sample which comprises sample data and actual injury data corresponding to the sample data; the sample data are input into a passenger injury prediction model for passenger injury prediction, predicted injury data corresponding to the sample data are obtained, and the passenger injury prediction model is used for predicting injury data corresponding to multiple body parts; the total loss value of the predicted injury data relative to the actual injury data is determined through a total loss function, and the total loss function reflects that loss weights of loss functions of different body parts are different; and adjusting model parameters of the passenger injury prediction model based on the total loss value. According to the method, damage prediction is carried out on a plurality of body parts by adopting the same passenger damage prediction model, a plurality of prediction models do not need to be arranged according to the body parts needing to be predicted, model parameters can be shared, the occupied space is small, and the cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle collision safety, and in particular to an occupant injury prediction model training method, a prediction method, a device and an electronic device. Background Art

[0002] Occupant injury prediction is a technology that assesses the degree of injury that occupants may suffer in traffic accidents. It is widely used in automobile safety research and vehicle design. For example, it is used in the development of vehicle passive safety performance and the damage when a collision occurs under different paths planned by autonomous driving to calculate the path, etc. Occupant injury prediction includes the prediction of injuries to multiple parts of the body, such as the head, neck, chest, abdomen, upper limbs, lower limbs, etc. Among them, the head, neck and chest are generally important parts that cause casualties in accidents, so occupant injury prediction generally includes these three body parts.

[0003] In recent years, many studies have applied AI technology to occupant injury prediction. In related technologies, independent models are generally used to predict the damage to each body part, and then the degree of damage to each body part is output separately. For the prediction of damage to a single body part, the model has high accuracy, but it also has limitations. For example, each body part corresponds to a prediction model, and parameters between models are not shared, resulting in the model occupying a large storage space and increasing costs. Summary of the invention

[0004] One of the purposes of the present invention is to provide a method for training an occupant injury prediction model to solve the problem in the prior art that an independent model is used to predict the injury of each body part, resulting in the model occupying a large storage space, requiring training of multiple models, and high cost; a second purpose is to provide a method for predicting occupant injury; a third purpose is to provide a device for predicting occupant injury; and a fourth purpose is to provide an electronic device.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] A training method for an occupant injury prediction model, comprising:

[0007] Acquire a training sample, wherein the training sample includes sample data and actual damage data corresponding to the sample data, wherein the sample data includes at least one of sample collision parameters, sample restraint system parameters, and sample occupant state parameters;

[0008] Inputting the sample data into an occupant injury prediction model to perform occupant injury prediction, and obtaining predicted injury data corresponding to the sample data, wherein the occupant injury prediction model is used to predict injury data corresponding to multiple body parts respectively;

[0009] Determining the total loss value of the predicted injury data relative to the actual injury data through a total loss function, wherein the total loss function may reflect the difference in loss weights of loss functions of different body parts;

[0010] Based on the total loss value, model parameters of the occupant injury prediction model are adjusted.

[0011] According to the above technical means, the total loss function reflects the different loss weights of the loss functions of different body parts, balances the learning efficiency of each body part, synchronizes the learning efficiency of each body part with different learning difficulty, and ensures the learning accuracy of each body part. In this way, different body parts can use the same occupant injury prediction model for injury prediction, without setting up multiple prediction models according to the body parts to be predicted, and the model parameters can be shared, simplifying the prediction model, and the prediction model occupies less space, reducing costs.

[0012] As a possible implementation method, the training method also includes:

[0013] Determine, by means of a loss function for each body part, a loss value of the predicted damage data relative to the actual damage data for each body part;

[0014] Determine the loss weight of the loss function of each body part according to the loss value, wherein the greater the loss value, the greater the loss weight;

[0015] The total loss function is updated based on the loss weight, and the total loss function is the sum of the loss functions of various body parts.

[0016] According to the above technical means, the weight of the loss function of each body part can be adjusted, thereby adjusting the learning efficiency of different body parts and balancing the learning performance of each body part so that each body part can converge simultaneously or nearly simultaneously.

[0017] As a possible implementation, the total loss function is:

[0018]

[0019] in, is the loss function of the ith body part at the tth time step; w i (t) is the loss weight of the ith body part at the tth time step.

[0020] As a possible implementation method, the weight of the loss of each body part in the total loss function is adjusted by loss weight. During the learning process, more consideration is given to the loss of body parts with large loss weights. As a result, body parts that are difficult to learn have larger weights during model training, thereby improving their learning efficiency and balancing the learning rates of various body parts.

[0021] The loss weight is obtained by the following formula:

[0022]

[0023] Among them, β is a hyperparameter that controls sensitivity; k i (t-1) is the derivative of the loss function for the ith body part at the t-1th time step.

[0024] According to the above technical means, it is possible to dynamically obtain the loss weight of each body part and dynamically adjust the weight of the loss of each body part in the total loss function.

[0025] As a possible implementation, the method further includes:

[0026] Constructing a first loss function for each body part according to the predicted acceleration and the real acceleration, wherein the training sample also includes the real acceleration corresponding to the sample data;

[0027] Constructing a second loss function for each body part based on the predicted injury data and the actual injury data;

[0028] The total loss function is constructed according to the sum of the first loss functions and the sum of the second loss functions of each body part.

[0029] As a possible implementation manner, the step of inputting the sample data into an occupant injury prediction model to perform occupant injury prediction to obtain predicted injury data corresponding to the sample data includes:

[0030] Inputting the sample data into an encoder of the occupant injury prediction model, performing feature extraction through the encoder, and obtaining feature data;

[0031] Inputting the characteristic data into a decoder of the occupant injury prediction model, wherein the characteristic data are respectively decoded by using a multi-layer perceptron layer to obtain a representation of the predicted acceleration of each body part;

[0032] In the decoder, a linear projection layer is used to decode the predicted acceleration of each body part respectively to obtain the predicted injury data corresponding to each body part;

[0033] Determining the total loss value of the predicted damage data relative to the actual damage data through the total loss function includes:

[0034] Determining a first loss value of the predicted acceleration relative to the actual acceleration by summing the first loss functions;

[0035] Determining a second loss value of the predicted damage data relative to the actual damage data by summing the second loss functions;

[0036] The total loss value is determined based on the first loss value and the second loss value.

[0037] According to the above technical means, the total loss function is constructed through the first loss function and the second loss function. The first loss function is used to train the model to learn the semantics of acceleration / force that characterizes the injury mechanism. Then, the semantics of the injury mechanism of different parts are filtered and converged through the attention weight layer. After passing through the linear projection layer, the semantics are converted to injury data of different parts and with different dimensions, and the injury data is learned. Based on the characteristics of acceleration / force and the design of the second loss function, an effective semantic space can be quickly learned, the learning efficiency is improved, and the number of learning samples is reduced. The semantic features of acceleration / force of each body part are converged for learning, and the correlation between each body part is considered to improve the accuracy of predicting injury data.

[0038] As a possible implementation method, in the decoder, the predicted acceleration of each body part is decoded separately using the linear projection layer to obtain the predicted damage data corresponding to each body part, including:

[0039] In the decoder, an attention weight layer is used to weight the predicted acceleration of a body part and the related predicted accelerations of other body parts to update the predicted acceleration of each body part.

[0040] The updated predicted acceleration of each body part is decoded by a linear projection layer to obtain the predicted injury data corresponding to each body part.

[0041] According to the above technical means, by associating the predicted acceleration of each body part and predicting the injury data based on the predicted acceleration, it is possible to take into account the mutual influence between various body parts, thereby improving the accuracy of the predicted injury data.

[0042] As a possible implementation method, the total loss function is:

[0043] totalloss = α * totalloss aux +(1-α)totalloss main ;

[0044] Among them: totalloss aux is the sum of the first loss functions of multiple body parts, totalloss main (t) is the sum of the second loss functions of the multiple body parts, and α is a hyperparameter.

[0045] As a possible implementation method, the first loss function sum is expressed as follows:

[0046]

[0047] in, is the first loss function of the ith body part at the tth time step; num_task is the total number of body parts.

[0048] And / or, the second loss function sum is expressed in the following manner:

[0049]

[0050] in, is the loss function of the ith body part at the tth time step; w i (t) is the loss weight of the ith body part at the tth time step.

[0051] The present application also provides a method for predicting occupant injury, including:

[0052] Obtaining one or more of collision parameters, restraint system parameters of a vehicle, and state parameters of an occupant in the vehicle under a collision scenario;

[0053] One or more of the collision parameters, the restraint system parameters and the occupant state parameters are input into an occupant injury prediction model to obtain predicted injury data of the occupant in the collision scenario. The occupant injury prediction model is obtained through the training method described above.

[0054] According to the above technical means, the occupant injury prediction model has high accuracy in predicting occupant injuries in collision scenarios.

[0055] As a possible implementation, the occupant injury prediction model includes an encoder and a decoder, and inputting one or more of the collision parameters, the restraint system parameters, and the occupant state parameters into the occupant injury prediction model to obtain predicted injury data of the occupant in the collision scenario includes:

[0056] Input one or more of the collision parameters, the restraint system parameters, and the occupant state parameters into an occupant injury prediction model, and extract features through the encoder to obtain feature data;

[0057] Inputting the characteristic data into a decoder of the occupant injury prediction model, wherein a multi-layer perceptron layer is used in the decoder to respectively decode and calculate the characteristic data to obtain predicted accelerations of various body parts;

[0058] In the decoder, a linear projection layer is used to decode and calculate the predicted acceleration of each body part respectively to obtain the predicted injury data of the occupant in the collision scenario.

[0059] As a possible implementation, in the decoder, a linear projection layer is used to decode and calculate the predicted acceleration of each body part respectively to obtain the predicted injury data of the occupant in the collision scenario, including:

[0060] In the decoder, an attention weight layer is used to filter the predicted acceleration of each body part, and the predicted acceleration of the body part and the related predicted accelerations of other body parts are weightedly aggregated to update the predicted acceleration of each body part;

[0061] The updated predicted acceleration is decoded using a linear projection layer to obtain predicted injury data of the occupant in the collision scenario.

[0062] The embodiment of the present application also provides an occupant injury prediction model training device, comprising:

[0063] A first acquisition module, acquiring training samples, wherein the training samples include sample data and actual damage data corresponding to the sample data, wherein the sample data includes at least one of sample collision parameters, sample restraint system parameters, and sample occupant state parameters;

[0064] A prediction model, used to predict occupant injury according to the input sample data, and obtain predicted injury data corresponding to the sample data, wherein the occupant injury prediction model is used to predict injury data corresponding to multiple body parts respectively;

[0065] A training module is used to determine the total loss value of the predicted injury data relative to the actual injury data through a total loss function, wherein the total loss function can reflect the different loss weights of the loss functions of different body parts; and based on the total loss value, adjust the model parameters of the occupant injury prediction model to obtain the occupant injury prediction model.

[0066] The embodiment of the present application also provides an occupant injury prediction device, comprising:

[0067] A second acquisition module, used to acquire one or more of collision parameters, restraint system parameters of the vehicle, and state parameters of passengers in the vehicle under the current collision scenario;

[0068] The occupant injury prediction model is used to predict occupant injury based on the input collision parameters, one or more of the restraint system parameters and the occupant state parameters to obtain predicted injury data of the occupant in the collision scenario.

[0069] The embodiment of the present application also provides an electronic device, including a processor and a memory communicatively connected to the processor;

[0070] The memory is used to store computer-executable instructions;

[0071] The processor executes the computer-executable instructions stored in the memory to implement the method described above.

[0072] An embodiment of the present application also provides a computer-readable storage medium / computer program product, wherein the computer-readable storage medium stores computer execution instructions / the computer program product includes computer execution instructions, and when the computer execution instructions are executed by a processor, they are used to implement the method described above.

[0073] Beneficial effects of the present invention:

[0074] (1) By setting the occupant injury prediction model as a multi-task model (one task for each body part), it is possible to predict injuries to different body parts of the occupant. However, for the multi-task prediction model, due to the different learning difficulties of different tasks, the prediction model is difficult to train, and after the prediction model is trained, the prediction result is not accurate. Based on this, the present application dynamically allocates the loss weight of the loss function of each body part in the total loss function to balance the learning efficiency of the model in each task, avoiding the situation where the loss value of the difficult-to-learn task is large after the easy-to-learn task converges, resulting in low accuracy. This achieves the use of the same occupant injury prediction model for injury prediction of different body parts, and the model parameters are shared, which reduces the setting of the prediction model, reduces the memory usage, and reduces the cost.

[0075] (2) In a further embodiment, the total loss function is constructed by a first loss function and a second loss function. The first loss function is used to train the model to learn the semantics of acceleration / force that characterizes the injury mechanism. The semantics of the injury mechanism of different parts are then filtered and aggregated through an attention weight layer. After passing through a linear projection layer, the semantics are converted to injury data of different parts and with different dimensions to learn the injury data. The training model designed based on the characteristics of acceleration / force and the second loss function can quickly learn the injury data, improve learning efficiency, and reduce the number of learning samples. The semantic features of acceleration / force of each body part are aggregated for learning, taking into account the correlation between each body part to improve the accuracy of predicting injury data.

[0076] (3) In combination with the occupant injury prediction model architecture and training method of the embodiment of the present application, during the training process, the risk of overfitting of the occupant injury prediction model can be reduced and the generalization ability of the occupant injury prediction model can be improved. At the same time, a single occupant injury prediction model covers the prediction of all damage parts, which significantly improves the reasoning speed of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 A schematic diagram of the framework of an encoder of the occupant injury prediction model provided by the present invention;

[0078] Figure 2 A schematic diagram of a decoder of the occupant injury prediction model provided by the present invention;

[0079] Figure 3 A schematic flow chart of a method for training an occupant injury prediction model provided by an embodiment of the present invention;

[0080] Figure 4 A schematic diagram of the loss weights of the loss functions of various body parts during the learning process of the occupant injury prediction model provided by the present invention;

[0081] Figure 5 A schematic flow chart of a method for training an occupant injury prediction model provided by another embodiment of the present invention;

[0082] Figure 6 for Figure 5 A schematic flow chart of a further embodiment of S507;

[0083] Figure 7 This is a schematic diagram of the test effect after the occupant injury prediction model provided by the present invention is trained;

[0084] Figure 8 A schematic flow chart of the occupant injury prediction method provided by the present invention;

[0085] Fig. 9 A schematic diagram of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0086] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention, not for limiting the scope of protection of the present invention.

[0087] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0088] Occupant injury prediction is a technology that assesses the degree of injury that occupants may suffer in traffic accidents. It is widely used in automobile safety research and vehicle design. For example, it is used in the development of vehicle passive safety performance and the damage when a collision occurs under different paths planned by autonomous driving to calculate the path, etc. Occupant injury prediction includes the prediction of injuries to multiple parts of the body, such as the head, neck, chest, abdomen, upper limbs, lower limbs, etc. Among them, the head, neck and chest are generally important parts that cause casualties in accidents, so occupant injury prediction generally includes these three body parts.

[0089] In recent years, many studies have applied AI technology to occupant injury prediction. In related technologies, independent models are generally used to predict the damage to each body part, and then the degree of damage to each body part is output separately. For the prediction of damage to a single body part, the model has high accuracy, but it also has limitations. For example, each body part corresponds to a prediction model, and parameters between models are not shared, resulting in the model occupying a large storage space and increasing costs.

[0090] Based on this, an embodiment of the present application proposes a method for predicting injuries to multiple body parts of an occupant through a prediction model. The following is an explanation from the perspectives of the architecture of the occupant injury prediction model, the training method of the occupant injury prediction model, and the occupant injury prediction method, with specific reference to the following embodiments.

[0091] It should be noted that a complete prediction model requires processes such as building data, constructing models, and training models.

[0092] In the embodiment of the present application, data is constructed in the following manner:

[0093] First, a high-precision finite element model or multi-rigid body constraint system model of the driver and passengers is established.

[0094] Secondly, random sampling is performed on the space formed by collision parameters (the collision parameters in this embodiment include collision waveforms), restraint system state parameters and occupant state parameters to obtain sample collision parameters, sample restraint system parameters and sample occupant state parameters, and a simulation matrix table is generated.

[0095] Then, the multi-finite element model or the rigid body restraint system model is updated according to the simulation matrix table, the simulation calculation is completed and the calculation results are processed, and the input (feature) is the sample collision parameter, the sample restraint system state parameter and the sample occupant state parameter, and the output is the real injury data representing the injury value of each body part. Multiple groups of sample collision parameters (the collision parameters in this embodiment include collision waveforms), sample restraint system state parameters and sample occupant state parameters form sample data, and the real injury data of each body part obtained by simulating each group of sample data is the real injury data corresponding to the sample data, and all sample data and the real injury data corresponding to the sample data form training samples.

[0096] In the embodiment of the present application, the model is constructed in the following manner:

[0097] The model constructed in the embodiment of the present application is an occupant injury prediction model, which includes an encoder and a decoder. The input parameters (sample data) are encoded by the encoder, and the encoded feature data is transmitted to the decoder. The decoder decodes the feature data to obtain the predicted injury data of the occupant.

[0098] like Figure 1 As shown, in this embodiment, the encoder includes an embedding layer, a linear projection layer, a multi-layer perceptron MLP_E1, a multi-layer perceptron MLP_E2, a hidden layer, a sequence feature extractor and a feature dimension splicing layer. Optionally, for the occupant injury prediction model, the input sample data includes sample collision parameters (also called sequence variables), sample restraint system parameters (also called category variables) and sample occupant state parameters (also called numerical variables). Among them, the sample restraint system parameters are input into the embedding layer, spliced ​​in the feature dimension splicing layer, and then input into the multi-layer perceptron MLP_E1 to obtain the hidden layer representation of the sample restraint system parameters. The sample occupant state parameters are directly input into the multi-layer perceptron MLP_E2 to obtain the hidden layer representation of the sample restraint system parameters. Each sample collision parameter is input into the linear projection layer respectively, and added in the feature dimension, and sent to the sequence feature extractor to obtain the hidden layer representation of the sample collision parameter, which contains the characteristics of the time series in the time dimension. Finally, the categorical variable hidden layer representation, the numerical variable hidden layer representation and the sequence variable hidden layer representation are concatenated in the feature dimension to obtain the final hidden layer representation H of the encoder, that is, the feature data output by the encoder. Optionally, the feature data is a data array.

[0099] Optionally, the embedding layer is a linear layer without bias, representing the categorical variable as a high-dimensional vector space.

[0100] Optionally, the sequence feature extractor is a bidirectional LSTM. The sequence feature extractor can be a recurrent neural network (RNN, GRU, LSTM), a convolutional neural network (CNN), or a Transformer based on an attention mechanism.

[0101] Optionally, MLP_E1 and MLP_E2 are both composed of a linear input layer, a hidden layer and a linear output layer, wherein the hidden layer is composed of a linear projection layer, a batch normalization layer and an activation function layer.

[0102] like Figure 2 As shown, in this embodiment, the decoder includes multiple injury prediction units (multiple heads) corresponding to multiple body parts, each injury prediction unit is connected to the encoder to obtain the feature data output by the encoder (sharing the same set of feature data), and each injury prediction unit is responsible for predicting the injury of a body part. The number of each injury prediction unit is determined according to the number of body parts that need to be predicted. Taking the prediction of the injury of the three parts of the head, neck and chest as an example, the decoder includes three injury prediction units, namely the head injury prediction unit head1, the neck injury prediction unit head2 and the chest injury prediction unit head3, and each injury prediction unit includes a multi-layer perceptron layer MLP_D, an attention weight layer G and a linear projection layer Project. Among them, the multi-layer perceptron layer MLP_D of each injury prediction unit converts the hidden layer representation H of the encoder into a semantic space Z suitable for a specific body part, and the attention weight layer filters and converges the semantic spaces of different body parts, aiming to associate different body parts and solve the problem of conflict between injury data of different body parts at the system level. The linear projection layer Project converts the converged semantic representation into the semantic space Y required by the body part, and realizes the prediction of injury data through the semantic control Y. Therefore, the embodiment of the present application uses an occupant injury prediction model to learn the same set of input parameters to obtain predicted injury data for different parts of the body. There is no need to establish multiple independent models to learn each body part separately. The model parameters and input parameters are shared to reduce memory usage.

[0103] Optionally, MLP_D consists of a linear input layer, two hidden layers and a linear output layer, where the hidden layer consists of a linear projection layer, a batch normalization layer and an activation function layer. The batch normalization layer is used to accelerate the convergence speed of the model and avoid overfitting. The activation function layer performs nonlinear transformation to improve the fitting ability of the model. Both the input and output layers are linear layers.

[0104] Optionally, the attention weight layer G is composed of a linear projection layer and a Softmax layer to achieve selective convergence on different body parts, and the Softmax layer maps the attention weights to the interval [0,1].

[0105] In this embodiment, the decoder realizes injury prediction for different body parts through multiple injury prediction units. Each injury prediction unit has an independent nonlinear conversion layer (multi-layer perceptron layer). This layer, combined with the design of the first loss function, can learn more semantics of acceleration / force that characterize the injury mechanism. The semantics of acceleration / force of different body parts are then filtered and converged through the attention weight layer. After passing through the linear projection layer, the semantics are converted to injury data of different body parts with different dimensions, which can make the model training effect better and the model prediction results more accurate.

[0106] Based on the above occupant injury prediction model, the present application embodiment proposes a training method for the occupant injury prediction model, such as Figure 3 As shown, Figure 3 A training method for an occupant injury prediction model provided in an embodiment of the present application is run in the hardware environment described in each of the above embodiments. The method includes:

[0107] S301, obtaining a training sample, where the training sample includes sample data and actual damage data corresponding to the sample data, where the sample data includes at least one of a sample collision parameter, a sample restraint system parameter, and a sample occupant state parameter;

[0108] Optionally, the training environment of the occupant injury prediction model can be a model training device, a car, or a terminal. The following is an example of running on a model training device, that is, the occupant injury prediction model is loaded on the model training device to train the occupant injury prediction model.

[0109] The training samples in this embodiment are output through simulation of the above-mentioned multi-finite element model or rigid body constraint system model, and the sample data in the training samples correspond one-to-one to the actual damage data. Optionally, the sample data includes N samples.

[0110] S302, inputting the sample data into the occupant injury prediction model to perform occupant injury prediction, and obtaining predicted injury data corresponding to the sample data;

[0111] The occupant injury prediction model includes an encoder and a decoder. The encoder performs feature extraction on sample data to extract feature data, and the decoder decodes the feature data to obtain semantic features corresponding to each body part. The language features of each body part are analyzed and calculated respectively to obtain predicted injury data corresponding to the sample data (including predicted injury data corresponding to each body part).

[0112] S303, determining the total loss value of the predicted damage data relative to the actual damage data through a total loss function;

[0113] S304: Adjust model parameters of the occupant injury prediction model based on the total loss value.

[0114] At the beginning of training (the first time step), the predicted damage data is calculated by forward propagation based on the initial model parameters of the occupant damage prediction model (the initial value is generally obtained by random initialization). There is a large deviation between the predicted damage data and the actual damage data. The total loss function is calculated, and the model parameters are adjusted by back propagation so that the predicted damage data predicted by the occupant damage prediction model tends to the real damage data. The specific implementation method is: determine the total loss value (deviation) of the predicted damage data relative to the actual damage data through the total loss function; calculate the partial derivative of the loss function for each parameter to obtain the gradient of the loss function with respect to each parameter (this process is calculated layer by layer from back to front through the chain rule), and use the optimization algorithm (such as gradient descent) to update the model parameters according to the gradient. Repeat the above steps to update the predicted damage data of the occupant damage prediction model until the total loss value is less than or equal to the preset value, or when the loss function converges (no longer decreases), it is determined that the predicted damage data is close to the actual damage data, then the occupant damage prediction model training is completed, and the input parameters in the collision scenario are predicted with the model parameters at the end of the training, which is close to the actual damage situation.

[0115] In this embodiment, the occupant injury prediction model is used to predict the injury data corresponding to multiple body parts. Since each body part has different indicator labels representing the injury status, the injury data of each body part needs to be predicted separately. Therefore, each body part has its own loss function, and the total loss function is the sum of the loss functions corresponding to each body part, that is, the total loss function is a composite loss function.

[0116] Since the indicator labels representing the damage of each body part of the occupant are different, for example, the indicator label of the degree of head injury is acceleration, the neck is the neck injury index Nij, and the chest is the compression (or displacement). The damage to the head mainly depends on the acceleration, which is linear, and this feature is relatively easy to learn. The damage to the chest mainly depends on the force of the seat belt and airbag, as well as the intensity of the collision waveform. The mapping relationship is relatively linear, and the model learning efficiency is high. However, as the part connecting the chest and the head, the neck has a complex force situation and is greatly affected by the posture of the human body. The model learning efficiency is low, which leads to different learning efficiency (speed) of different body parts. It can be seen that the learning difficulty of the damage of each body part is different. When the same prediction model is used to train the damage degree of multiple body parts, the convergence speed is different, which will affect the learning accuracy of some body parts. Therefore, in the related technology, independent models are often used to train the damage prediction of each body part.

[0117] In the embodiment of the present application, the total loss function reflects that the loss weights of the loss functions of different body parts are different, and the loss weights balance the weights of the losses of each body part, thereby balancing the learning efficiency of different body parts. The greater the loss weight, the higher the learning difficulty, that is, the learning efficiency of each body part is adjusted by the loss weight. For body parts that are more difficult to learn, the loss weight of the part is increased to improve the learning efficiency of the part, so that the convergence speeds of different body parts are close and the learning accuracy is high, so that different body parts use the same occupant injury prediction model to predict the injury situation.

[0118] As an example, the loss weights of the loss functions of different body parts can be set values, which can be specifically set according to the difficulty of learning each body part. For example, if the injury data of the head is easy to learn, then the loss weight of the loss function of the head is set to be small; if the injury data of the neck is difficult to learn, then the loss weight of the loss function of the neck is set to be large, and so on.

[0119] As an example, the loss weight of the loss function of each body part is adjusted by the loss value determined by the loss function in the previous time step during the model training process, that is, the method also includes: determining the loss value of the predicted injury data of each body part relative to the actual injury data through the loss function of each body part; determining the loss weight of the loss function of each body part according to the loss value, the larger the loss value, the larger the loss weight; updating the total loss function based on the loss weight.

[0120] The total loss function is the sum of the loss functions of all body parts that need to be predicted. Updating the total loss function based on the loss weight actually updates the loss function of each body part in the total loss function and its loss weight. The total loss function in the embodiment of the present application and the loss weights of the loss functions of different body parts are explained below by formula expressions.

[0121] As an example, the above total loss function can be expressed by the following formula:

[0122]

[0123] in, is the loss function of the ith body part at the tth time step; w i (t) is the loss weight of the ith body part at the tth time step.

[0124] As an example, if the prediction of injury data is a regression problem, the root mean square loss function is selected. Therefore, the loss function corresponding to each body part can be expressed by the following formula:

[0125]

[0126] Where N is the number of samples, Y j , are the predicted damage data and actual damage data of the jth sample respectively.

[0127] As an example, if the prediction of injury data is a classification problem (such as predicting the injury level of each body part), the loss function for each body part can be a cross entropy loss function.

[0128] Optionally, the loss weight of the loss function for each body part is calculated by the following formula:

[0129]

[0130] Among them, k i (t-1) reflects the performance index of the ith body part at the t-1 time step; β is a hyperparameter for controlling sensitivity. In this embodiment, the value of β is 1.5. The larger β is, the closer the probability of tasks (body parts) with higher probabilities is to 1. β is used to adjust the loss weight of each body part to the performance index k i The larger the β is, the more sensitive it is. Its value is usually around 1.0. Hyperparameter optimization is performed based on the overall performance of the model.

[0131] Optionally, k i The larger the value of (t-1), the better the performance of the model. For classification problems, the accuracy can be selected as the performance indicator, and for regression problems, the inverse of the mean square error can be selected as the performance indicator. This embodiment takes the regression task as an example. i (t-1) can be expressed by the following formula:

[0132]

[0133] Among them, k i The larger (t-1), the exponential function exp(β*k i The larger the result of (t-1), the larger the loss weight assigned, which ensures that the model gives higher loss weights to difficult tasks. After softmax normalization, the normalized loss weight w is obtained. i (t), ensuring that the sum of the loss weights of all body parts is equal to 1, which has better interpretability and increases the stability of the calculation.

[0134] For example: (1) In the t-1th time step, the occupant injury prediction model predicts the occupant injury of the sample data through forward propagation to obtain the predicted injury data corresponding to the sample data. (2) The loss value of the predicted injury data of the head relative to the actual injury data is calculated as X1, the loss value of the predicted injury data of the neck relative to the actual injury data is calculated as X2, and the loss value of the predicted injury data of the chest relative to the actual injury data is calculated as X3. The total loss value of the predicted injury data relative to the actual injury data determined by the total loss function is X=X1+X2+X3 (when the model is just beginning to be trained, which is also the first time the data is scanned, the loss weights of each body part are the same or close, and the sum of the loss weights is 1); the learning performance index K1(t-1) of the head at the t-1 time step is determined according to the inverse of X1, and then the loss weight w1(t) of the head is determined; the learning performance index K2(t-1) of the neck at the t-1 time step is determined according to the inverse of X2, and then the loss weight w2(t) of the neck is determined; the learning performance index K3(t-1) of the chest at the t-1 time step of the head is determined according to the inverse of X3, and then the loss weight w3(t) of the chest is determined. (3) Calculate the total loss function. After adjusting the model parameters of the occupant injury prediction model through back propagation, the occupant injury prediction model predicts the occupant injury of the sample data in the tth time step to obtain the predicted injury data corresponding to the sample data. For example, the loss function of the head determines the loss value of the predicted injury data of the head relative to the actual injury data as X11, the loss value of the predicted injury data of the neck relative to the actual injury data as X21, and the loss value of the predicted injury data of the chest relative to the actual injury data as X31. The total loss function loss value is X=w1(t)*X11+w2(t)*X21+w3(t)*X31, where w1(t)+w2(t)+w3(t) is equal to 1. Repeat steps (1), (2), and (3) until the total loss value is less than or equal to the preset value, or the total loss value converges (no longer decreases).

[0135] In this embodiment, the contribution of different body parts to the total loss value is adjusted by loss weights, thereby affecting the total loss function so that the model will not be biased towards a specific body part, especially when the difficulty or data volume of the body parts is unbalanced. The contribution of different body parts to the total loss value is adjusted by loss weights to balance the learning efficiency of each body part. And the loss weight directly affects the gradient of the loss of each body part. By adjusting the loss weights, the dominant role of easy-to-train tasks in model training can be reduced, thereby avoiding the model from being biased towards easy-to-train body parts; it can help the model better learn more difficult tasks.

[0136] like Figure 4As shown, during the training process of the occupant injury prediction model in the embodiment of the present application, the weight of the head (w1) is the smallest and the weight of the neck (w3) is the highest, indicating that the head is easiest to learn useful features, while the neck is the most difficult to learn. As the component connecting the head and the chest, the neck has a more complex force condition and is highly nonlinear. In the early and middle stages of a collision, the head is in a relatively free state of motion, and the force analysis is relatively simple, mainly depending on acceleration, and it is easier to learn useful features.

[0137] Optionally, in each time step, the loss value determined for each body part is different, and therefore the loss weight is also different. For example, at the beginning of training, the loss value is large, so the loss weight is increased accordingly. During the training process, the loss value gradually decreases, or the rate of decrease is accelerated, so the loss weight is adjusted accordingly to adjust the convergence speed of each body part so that the convergence speed of each body part is close or the same. In this way, it can be ensured that each body part can be learned at the same time, and the learning accuracy of each body part is high.

[0138] In the embodiment of the present application, the loss weights of the loss functions of different body parts can be reflected by the design of the total loss function, and the learning efficiency of each body part can be balanced, so that the learning efficiency of each body part with different learning difficulty is synchronized, and the learning accuracy of each body part is ensured. In this way, different body parts can use the same occupant injury prediction model for injury prediction, without setting multiple prediction models according to the body parts to be predicted, and the model parameters can be shared, simplifying the prediction model, and the prediction model occupies less space, reducing costs.

[0139] Figure 5 This is a flow chart of a training method for a passenger loss prediction model proposed in another embodiment of the present application. This embodiment is based on the above Figures 1 to 4 In the embodiment shown, based on the above embodiment, the training method of the occupant loss prediction model of this embodiment can reduce the number of learning samples, improve learning efficiency, and can associate the mutual influence between various body parts, so that the learning accuracy is higher. Figure 5 As shown, the training method includes:

[0140] S501, obtaining a training sample, where the training sample includes sample data, actual damage data corresponding to the sample data, and real acceleration corresponding to the sample data, and the sample data includes at least one of sample collision parameters, sample restraint system parameters, and sample occupant state parameters;

[0141] In this embodiment, the method for obtaining training samples is the same as the method for obtaining training samples in the above embodiment, and will not be described in detail here.

[0142] The physical essence of the injury to each part of the body is that it is caused by force, and the force is proportional to the acceleration, so the acceleration label is easy to learn. Based on this consideration, if the acceleration is learned first, and then the acceleration is used as an intermediate variable, the injury to each part of the body is mapped through the intermediate variable, the number of learning samples can be reduced, and the learning accuracy is high. In order to obtain the intermediate variable-acceleration, consider that the output of the multi-layer perception machine layer corresponding to each body part includes acceleration that can characterize the injury mechanism. Therefore, in an embodiment of the present application, the acquired training samples include sample data, actual injury data corresponding to the sample data, and true acceleration corresponding to the sample data, and the acceleration label is learned by adding a loss function.

[0143] S502, constructing a first loss function for each body part according to the predicted acceleration and the actual acceleration;

[0144] The first loss function in the embodiment of the present application is constructed based on the predicted acceleration and the actual acceleration. The occupant injury prediction model is trained to obtain a characterization label of the predicted acceleration, so that the model can predict the injury data through the predicted acceleration.

[0145] As an example, the sum of the first loss functions for multiple body parts is:

[0146]

[0147] Among them: totalloss aux is the sum of the first loss functions of multiple body parts, is the first loss function of the ith body part at the tth time step; num_task is the total number of body parts.

[0148] Optionally, in this embodiment, the first loss function of each body part can be expressed by the following formula:

[0149]

[0150] Where N is the number of samples, Z j , are the predicted acceleration and true acceleration of the jth sample respectively.

[0151] S503, constructing a second loss function for each body part according to the predicted injury data and the actual injury data;

[0152] The injury index of each body part of the occupant is not acceleration, and the injury index of different body parts is different. For example, the index of the head is acceleration, the index of the neck is the neck injury index, and the index of the chest is compression. Through the first loss function, only the acceleration label can be learned, and the predicted injury data of each body part cannot be obtained. Therefore, it is necessary to learn the injury index of each body part through the second loss function. It can be understood that the second loss function is the main loss function and should play a leading role, and the first loss function is an auxiliary loss function.

[0153] In the embodiment of the present application, the second loss function is based on the predicted injury data and actual injury data of each body part, and the injury data represents the injury index of each body part.

[0154] As an example, the sum of the second loss functions for multiple body parts is:

[0155]

[0156] Among them, totalloss main (t) is the sum of the second loss functions of multiple body parts, is the second loss function of the ith body part at the tth time step; w i (k) is the loss weight of the ith body part at the tth time step.

[0157] Among them, w i (t) is obtained in the same way as above Figure 3 In the embodiment shown, w i The method of obtaining (t) is the same. For details, please refer to the above Figure 3 The embodiments are not described in detail here.

[0158] Optionally, the second loss function of each body part can be expressed by the following formula:

[0159]

[0160] Where N is the number of samples, Y j , are the predicted damage data and actual damage data of the jth sample respectively.

[0161] It should be noted that, in this embodiment, the second loss function of each body part is Figure 3 The loss function is the same for each body part shown.

[0162] S504, constructing a total loss function according to the sum of the first loss functions and the sum of the second loss functions of each body part.

[0163] In this embodiment, the loss function of each body part includes the above-mentioned first loss function and second loss function, so that the total loss function composed of multiple body parts is composed of the sum of the first loss functions and the sum of the second loss functions.

[0164] In this embodiment, the sum of the first loss function and the sum of the second loss function of each body part constructs a total loss function, and the total loss value of the predicted injury data relative to the actual injury data is determined by the total loss function.

[0165] As an example, the total loss function can be expressed by the following formula:

[0166] totalloss = α * totalloss aux +(1-α)totalloss main ;

[0167] Among them, totalloss aux is the sum of the first loss function; totalloss main (t) is the sum of the second loss function; α is a hyperparameter used to adjust the weight of the sum of the first loss function and the sum of the second loss function, where 0<α<0.5, optionally, α=0.15. Based on the above-constructed total loss function, the training process of the occupant injury prediction model in the embodiment of the present application is as follows:

[0168] S505, inputting the sample data into an encoder of the occupant injury prediction model, performing feature extraction through the encoder, and obtaining feature data;

[0169] The occupant injury prediction model includes an encoder and a decoder, wherein the encoder is used for feature extraction to convert sample data extraction into feature data required by the model so as to facilitate the operation of the model.

[0170] It should be noted that, in this embodiment, the principle and process of the encoder extracting features from sample data can be referred to the above Figure 1 The description will not be repeated here.

[0171] S506, inputting the characteristic data into a decoder of the occupant injury prediction model, wherein the characteristic data are respectively decoded by using a multi-layer perceptron layer to obtain predicted accelerations of various body parts;

[0172] Combination Figure 2 As shown, the feature data is the hidden layer representation H output by the encoder. After the hidden layer representation H is input into the decoder, the feature data is decoded by the multi-layer perceptron layer of the decoder and output as follows: Figure 2Among Z1, Z2 and Z3, Z1 represents the predicted acceleration related to body part 1, Z2 represents the predicted acceleration related to body part 2, and Z3 represents the predicted acceleration related to body part 3.

[0173] S507, in the decoder, using a linear projection layer to decode the predicted acceleration of each body part respectively, to obtain predicted injury data corresponding to each body part;

[0174] Continue to refer Figure 3 ,After the multi-layer perceptron layer of the decoder outputs Z1, Z2 and Z3, Z1, Z2 and Z3 are respectively input to the linear projection layers corresponding to each body part, decoding and calculation are performed in the linear projection layers, and the predicted injury data (Y1, Y2 and Y3) corresponding to each body part are obtained.

[0175] In the embodiment of the present application, the acceleration characteristics are first quickly predicted, and then the injury data corresponding to each body part is mapped based on the acceleration characteristics. This can effectively reduce the sample learning data, converge quickly, and have high accuracy.

[0176] S508, determining a first loss value of the predicted acceleration relative to the actual acceleration by summing the first loss function;

[0177] S509, determining a second loss value of the predicted damage data relative to the actual damage data by summing the second loss function;

[0178] S510: Determine a total loss value based on the first loss value and the second loss value.

[0179] As an example, the total loss value is the sum of the first loss value and the second loss value; or, the total loss value is determined based on the first loss value and the first loss function weight, the second loss value and the second loss function weight, as shown in the above expression of the total loss function.

[0180] S511: Adjust the model parameters of the occupant injury prediction model based on the total loss value.

[0181] Optionally, in this step, the specific implementation method of adjusting the model parameters of the occupant injury prediction model based on the total loss value is the same as that described above. Figure 2 The embodiments shown are the same, and specific details may be referred to the above embodiments, which will not be described again.

[0182] The total loss function of the embodiment of the present application includes the sum of the first loss function and the sum of the second loss function. The predicted acceleration that characterizes the injury mechanism is obtained by adding the design of the sum of the first loss function, and the final predicted injury data is obtained by the design of the sum of the second loss function. The predicted acceleration is obtained based on the predicted injury data through the predicted acceleration mapping, and the predicted acceleration improves the learning efficiency of the model, thereby reducing the number of learning samples, and the learning accuracy is high.

[0183] Furthermore, in the related art, each body part is predicted using an independent prediction model. However, as a system, different body parts of the human body are physically (mechanically) correlated, that is, there is a correlation between the various body parts of the occupant. The technical solution of predicting each part separately has poor correlation between body parts, and the model cannot learn the correlation characteristics of various parts of the human body. Therefore, independent models generally do not have correlation between different body parts, resulting in inaccurate predicted injury levels. For example: in a collision accident, the head penetrates the airbag, causing the head acceleration to be very high. The force on the head will be transmitted to the neck, resulting in high force on the neck, thereby causing high neck injury. When learning the degree of neck injury, this correlation is usually difficult to learn at the same time in an independent model. Based on this consideration, the embodiment of the present application combines the setting of the above-mentioned first loss function to learn the predicted acceleration, and then associates the predicted acceleration of each body part, and combines the predicted acceleration of each body part to calculate the injury value of each body part to ensure the correlation of each part.

[0184] For example, see Figure 6 Based on the above embodiment, the specific implementation of step S507 may be:

[0185] S601, in the decoder, filtering the predicted acceleration of each body part using an attention weight layer;

[0186] The embodiment of the present application filters the predicted acceleration of each body part through the attention weight layer to obtain the predicted acceleration corresponding to each body part and avoid noise interference.

[0187] S602, updating the predicted acceleration of each body part according to the predicted acceleration of each body part and the related predicted accelerations of other body parts;

[0188] S603, using a linear projection layer to decode the updated predicted acceleration of each body part respectively, to obtain predicted injury data corresponding to each body part.

[0189] This application is based on the correlation between various body parts, considers the impact of each part on other parts, and combines the relevant predicted accelerations of other body parts to predict the injury data of the body part itself.

[0190] by Figure 2 For example, Z1 represents the predicted acceleration associated with body part 1 (such as the head), Z2 represents the predicted acceleration associated with body part 2 (such as the neck), and Z3 represents the predicted acceleration associated with body part 3 (such as the chest). In some embodiments, Z1 can be used to map the predicted injury data of body part 1, Z2 can be used to map the predicted injury data of body part 2, and Z3 can be used to map the predicted injury data of body part 3. In this embodiment, Z1 and the related Z2 and Z3 are used to map the predicted injury data of body part 1; Z2 and the related Z1 and Z3 are used to map the predicted injury data of body part 2; and Z3 and the related Z1 and Z2 are used to map the predicted injury data of body part 2.

[0191] As an example, the predicted acceleration of the body part itself and the related predicted acceleration of other body parts can be allocated based on weights, and optionally, the weight of the body part itself is the largest, the weight of other body parts directly connected to the body part itself is small, and the weight of other body parts indirectly connected is the smallest. For example, when calculating the predicted injury data of the head, the weight of Z1 of the head is the largest, the weight of Z2 of the neck is small, and the weight of Z3 of the chest is the smallest.

[0192] The embodiment of the present application adopts a prediction model to predict the injury data of multiple body parts, and based on the correlation between each body part, when predicting the injury data, the relevant data of each body part are combined for prediction, so that the prediction result is more accurate and closer to the actual injury value, thereby accelerating the model learning speed.

[0193] It is understandable that single-task model training learns features and patterns from the data of a single task. With a limited amount of training data, the model is more likely to overfit the noise and specific patterns in the data and cannot learn more general features. Multi-task learning can share the feature representations of multiple tasks, better capture the common features in the data, and force the model to learn features that are useful for all tasks, thereby reducing the risk of model overfitting. At the same time, different tasks have their own constraints, which play a regularization effect and also alleviate overfitting to a certain extent. Therefore, the multi-task (multiple body parts) learning model training of the embodiment of the present application has high learning efficiency, and can meet the prediction of injury data of multiple body parts in the same model. The model parameters are shared, which reduces the memory usage and simplifies the model training.

[0194] In this embodiment, step S501 and steps S502 to S504 are not limited to the described action sequence, and step S501 and steps S502 to S504 may be performed in other sequences or simultaneously.

[0195] Optionally, to ensure that the magnitude of the losses of different body parts is comparable, the damage data (labels) of the training data of each body part are normalized and mapped to the interval [0,1]. That is, after obtaining the predicted damage data, the predicted damage data of each body part is normalized and mapped to the interval [0,1].

[0196] Optionally, in this embodiment, adaptive moment estimation (ADAM) is used as an optimizer to optimize the trained model. In order to prevent overfitting, L2 regularization, node random dropout and early stopping are used.

[0197] Optionally, after the occupant injury prediction model is trained through the embodiment of the present application, 4 samples are randomly selected from the test set for visual verification. The effect diagram of the predicted injury data output by the occupant injury prediction model is as follows: Figure 7 As shown in the figure, the solid line is the actual injury value, and the dotted line is the predicted value of the model. Among the four test samples, whether from the maximum value of the curve or the overlap of the curve, the head, neck and chest obtain good accuracy at the same time or nearly at the same time. Therefore, using the same occupant injury prediction model to predict injuries to the head, neck and chest can obtain high-precision prediction results for each body part.

[0198] The occupant injury prediction model is trained by the training method of the embodiment of the present application, so that the occupant injury prediction model can predict multiple body parts, without the need to set up multiple models separately to predict each body part, and based on the same occupant injury prediction model, the injury data of each body part is mapped by setting the intermediate variable acceleration, and each body part can be associated through the intermediate variable, taking into account the mutual influence of each body part when it is stressed, so that the accuracy of the predicted injury data is higher. Through the above model architecture and training method, the model can learn a more general feature representation, reduce the risk of model overfitting, and improve the generalization ability of the model. At the same time, a single model covers the prediction of all injury parts, which significantly improves the reasoning speed of the model and reduces the memory overhead of the model.

[0199] Based on the above training method, an occupant injury prediction model is obtained, and the occupant injury prediction model is used to predict occupant injuries in collision scenarios. The present application also provides an occupant injury prediction method, such as Figure 8 As shown, the method includes:

[0200] S801, obtaining one or more of collision parameters, vehicle restraint system parameters, and vehicle occupant status parameters under a collision scenario;

[0201] The occupant injury prediction method of the embodiment of the present application can be applied to various collision scenarios, including collision scenarios of various vehicles, such as cars, airplanes, etc.

[0202] It should be noted that the parameters and proportions to be obtained can be adjusted accordingly in different vehicles. For cars, collision parameters, vehicle restraint system parameters, and vehicle occupant status parameters can be obtained for prediction. For ships or trains, collision parameters and vehicle occupant status parameters can be obtained for prediction.

[0203] In this embodiment, the collision parameters (or collected in the form of waveforms, that is, collision waveforms) include but are not limited to: in the current collision scenario, the acceleration of the vehicle in the X, Y, and Z directions and the rotational acceleration around the X, Y, and Z axes, a total of 6 sequence data. It should be noted that in actual applications, one or more of these 6 sequence data can be selected as features to participate in the prediction, but the acceleration in the X direction is a necessary parameter. Optionally, except that the Z-direction acceleration information has a small contribution and may bring noise to the model, the more other vectors are selected, the higher the prediction accuracy.

[0204] The state parameters of the restraint system include but are not limited to the seat belt state, airbag state, layout space and other parameters that mainly affect occupant injuries. The seat belt state refers to whether the seat belt is fastened, and the airbag state refers to whether the airbag is detonated, which are categorical variables. The layout space refers to the seat belt point parameters and the occupant survival space parameters. The seat belt point parameters refer to the X, Y, and Z distances from the seat belt fixing point (seat belt ring, buckle, anchor point, retractor) to the seat R point. The occupant survival space parameters include the distance from the steering wheel / instrument to the occupant, the distance from the knee to the instrument, and other parameters that determine the occupant's survival space in a collision, which are numerical variables.

[0205] Occupant status parameters include but are not limited to parameters that affect occupant tolerance, such as height, weight, and age, and are numerical variables.

[0206] As an example, collision parameters can be obtained through finite element model simulation, or through other multi-rigid body dynamics models, or through table lookup / interpolation of table data pre-stored on the vehicle side, or by inputting the key parameters of the current scene into another pre-trained model to obtain a predicted waveform; the seat belt status in the restraint system parameters can be obtained from the buckle status, the airbag status parameters can be obtained from the airbag controller, the layout space parameters are fixed for each vehicle model and can be obtained from a storage device; the occupant status parameters can be obtained from the user's pre-input and stored in a storage device, or they can be dynamically obtained based on the personnel status detection system (Driver Monitor System, DMS).

[0207] As an example, after obtaining collision parameters in a collision scenario, restraint system parameters of a vehicle, and occupant status parameters in the vehicle, the numerical variables in the collision parameters, restraint system, and occupant status parameters are preprocessed.

[0208] In this embodiment, normalization preprocessing is used to adjust the numerical range of the data to a specific vector, thereby improving the convergence speed of the algorithm.

[0209] Optionally, normalize to MinMax normalization:

[0210]

[0211] Among them, X min , X max are the minimum and maximum values ​​of the data X corresponding to the collision parameter, the restraint system and the personnel status parameter, respectively. norm is the normalized data. If during the inference process, the range of the input parameter exceeds [X min , X max ], then convert the input: X = max(min(X, X max ), X min ).

[0212] S802, inputting one or more of the collision parameters, the restraint system parameters, and the occupant state parameters into an occupant injury prediction model to obtain predicted injury data of the occupant in the collision scenario;

[0213] One or more of the pre-processed collision parameters, restraint system and occupant status parameter data are input into a pre-trained occupant injury prediction model to obtain predicted injury data of different parts of the occupant.

[0214] In this step, the parameters input into the occupant injury prediction model are consistent with the parameters obtained. For example, if the collision parameters are obtained, the collision load is input into the occupant injury prediction model; if the collision parameters and the restraint system parameters are obtained, both the collision parameters and the restraint system parameters are input into the occupant injury prediction model; if the collision parameters, the restraint system parameters and the occupant status parameters are obtained, then the collision parameters, the restraint system parameters and the occupant status parameters are all input into the occupant injury prediction model.

[0215] In this embodiment, the occupant injury prediction model is trained using the above-mentioned training method. Therefore, based on the occupant injury prediction module of the embodiment of the present application, the predicted injury data of various body parts of the occupant can be obtained, and the predicted injury data has high accuracy and a higher prediction effect.

[0216] As an example, based on the above embodiment, the specific implementation of step S802 may be:

[0217] The collision parameters, restraint system parameters and occupant status parameters are input into the occupant injury prediction model, and feature data is obtained by extracting features through an encoder;

[0218] The occupant injury prediction model includes an encoder and a decoder. The encoder is used to extract features to convert collision parameters, restraint system parameters and occupant status parameters into feature data required by the model, so that the model can operate.

[0219] It should be noted that, in this embodiment, the principle and process of the encoder extracting features of the collision parameters, restraint system parameters and occupant status parameters can be referred to above. Figure 2 The illustrated embodiments will not be described in detail here.

[0220] The characteristic data is input into the decoder of the occupant injury prediction model. In the decoder, a multi-layer perceptron layer is used to decode and calculate the characteristic data respectively to obtain the predicted acceleration of each body part;

[0221] In the decoder, a linear projection layer is used to decode and calculate the predicted acceleration of each body part respectively to obtain the predicted injury data of the occupant in the collision scenario.

[0222] In this embodiment, based on the occupant injury prediction model, the acceleration of each body part is learned through the first loss function training, and the output of the multi-layer perceptron layer of the decoder includes the predicted acceleration. Therefore, the predicted acceleration of each body part can be obtained by calculating the feature data through the multi-layer perceptron layer of the encoder. Then, the predicted acceleration is used to map the predicted injury data of each body part, thereby improving the prediction efficiency, and the predicted injury data has high accuracy.

[0223] Furthermore, in some embodiments, the output of the multi-layer perceptron layer based on the occupant injury prediction model includes predicted acceleration. In this embodiment, after the multi-layer perceptron layer corresponding to each body part outputs predicted acceleration, the predicted acceleration of each body part can be combined to map the predicted injury data of each body part. As described below:

[0224] In the decoder, an attention weight layer is used to filter the predicted acceleration of each body part;

[0225] updating the predicted acceleration of each body part based on the predicted acceleration of each body part and the associated predicted accelerations of other body parts;

[0226] The updated predicted acceleration is decoded using a linear projection layer to obtain the predicted injury data of the occupants in the collision scenario.

[0227] In this embodiment, the process of processing the feature data by the occupant injury prediction model is the same as the process of processing the sample data in the above model training process. For details, please refer to the above Figure 5 and Figure 6 The description of the feature data processing in the embodiment will not be repeated here.

[0228] The damage data predicted by the occupant damage prediction model of this embodiment has high accuracy.

[0229] It should be understood that although the steps in the flowchart are executed sequentially as indicated by the arrows, these steps are not necessarily executed sequentially in the specified order. Unless otherwise clearly stated herein, these steps may be executed in other orders. Moreover, a step in the flowchart of each embodiment may be decomposed into multiple steps, which may not necessarily be executed at the same time, and may be executed simultaneously or alternately with at least a portion of other steps.

[0230] The present application also provides an occupant injury prediction model training device, which includes:

[0231] A first acquisition module acquires a training sample, wherein the training sample includes sample data and actual damage data corresponding to the sample data, wherein the sample data includes at least one of a sample collision parameter, a sample restraint system parameter, and a sample occupant state parameter;

[0232] A prediction model is used to predict occupant injury based on the input sample data to obtain predicted injury data corresponding to the sample data, wherein the occupant injury prediction model is used to predict injury data corresponding to multiple body parts respectively;

[0233] The training module is used to determine the total loss value of the predicted injury data relative to the actual injury data through the total loss function, where the total loss function reflects that the loss weights of the loss functions of different body parts are different; and based on the total loss value, adjust the model parameters of the occupant injury prediction model to obtain the occupant injury prediction model.

[0234] As an example, the training module includes:

[0235] The first determination module is used to determine the loss value of the predicted injury data of each body part relative to the actual injury data through the loss function of each body part; and determine the loss weight of the loss function of each body part according to the loss value, wherein the larger the loss value, the greater the loss weight.

[0236] The update module is used to update the total loss function according to the loss weight, and the total loss function is the sum of the loss functions of each body part.

[0237] During the training process, the loss weights of each loss function of the total loss function are adjusted based on the learning loss of each learning step to balance the learning efficiency of each body part.

[0238] As an example, the training module also includes a construction module for constructing a first loss function for each body part based on the predicted acceleration and the actual acceleration; for constructing a second loss function for each body part based on the predicted injury data and the actual injury data; and for constructing a total loss function based on the sum of the first loss functions of each body part and the sum of the second loss functions.

[0239] As an example, the prediction model includes:

[0240] The encoder is used to extract features from the input sample data to obtain feature data;

[0241] A decoder, used to process the input feature data, optionally, the decoder includes a multi-layer perceptron layer and a linear projection layer, the multi-layer perceptron layer is used to decode the feature data respectively to obtain the representation of the predicted acceleration of each body part; the linear projection layer is used to decode the predicted acceleration of each body part respectively to obtain the predicted injury data corresponding to each body part;

[0242] As an example, the decoder further includes: an attention weight layer for weighted aggregation of the predicted acceleration of the body part and the related predicted accelerations of other body parts to update the predicted acceleration of each body part;

[0243] The linear projection layer is also used to decode the updated predicted acceleration of each body part respectively to obtain the predicted injury data corresponding to each body part.

[0244] The implementation scheme of the device is similar to the implementation scheme recorded in the above method. The occupant injury prediction device provided can implement the training method of the occupant injury prediction model described above. Therefore, the occupant injury prediction device has the same effect as the above embodiment and will not be repeated here.

[0245] The embodiment of the present application further provides an occupant injury prediction device, which includes an occupant injury prediction model trained by an occupant injury prediction model training method, and is applied to an automobile or an electronic device on the automobile, and includes:

[0246] A second acquisition module is used to acquire collision parameters, restraint system parameters of the vehicle and occupant status parameters in the vehicle under the current collision scenario;

[0247] The occupant injury prediction model is used to predict occupant injuries based on input collision parameters, restraint system parameters and occupant status parameters to obtain the predicted injury data of the occupant in the current collision scenario.

[0248] As an example, the occupant injury prediction model includes:

[0249] An encoder is used to extract features from the input collision parameters, restraint system parameters of the vehicle, and occupant status parameters in the vehicle under the current collision scenario to obtain feature data;

[0250] A decoder, used to process the input feature data, optionally, the decoder includes a multi-layer perceptron layer and a linear projection layer, the multi-layer perceptron layer is used to decode the feature data respectively to obtain the representation of the predicted acceleration of each body part; the linear projection layer is used to decode the predicted acceleration of each body part respectively to obtain the predicted injury data corresponding to each body part;

[0251] As an example, the decoder further includes: an attention weight layer for weighted aggregation of the predicted acceleration of the body part and the related predicted accelerations of other body parts to update the predicted acceleration of each body part;

[0252] The linear projection layer is also used to decode the updated predicted acceleration of each body part respectively to obtain the predicted injury data corresponding to each body part.

[0253] The implementation scheme of the device is similar to the implementation scheme recorded in the above method. The provided occupant injury prediction device can implement the occupant injury prediction method described above. Therefore, the occupant injury prediction device has the same effect as the above embodiment and will not be repeated here.

[0254] Each module in the above-mentioned multi-modal occupant injury prediction device can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0255] The present application also provides an electronic device, such as Fig. 9 As shown, the electronic device includes a processor 901 and a memory 902 that is communicatively connected to the processor 901, and the memory 902 includes a memory bar and a non-volatile storage medium;

[0256] The memory 902 is used to store computer executable instructions;

[0257] The processor 901 executes the computer-executable instructions stored in the memory 902 to implement the training method and the prediction method in the above-mentioned embodiments.

[0258] The electronic device also includes a communication interface 903 and a CAN bus 904, wherein the processor is used to provide computing power and control capabilities, and can be a GPU, CPU, NPU, MCU, FPGA, etc. The storage device includes an internal memory and a non-volatile memory. The non-volatile memory stores a computer program that implements the above method. The internal memory provides an environment for program startup and operation. The communication interface is used to communicate with an external terminal by wire or wireless.

[0259] In this embodiment, the electronic device may be an electronic device in a car, or may be a car. Alternatively, the electronic device is a computer device, and the device may be a vehicle-mounted terminal, a cloud platform, or a mobile terminal.

[0260] An embodiment of the present application also provides a computer-readable storage medium / computer program product, wherein the computer-readable storage medium stores computer execution instructions / the computer program product includes computer execution instructions, and when the computer execution instructions are executed by a processor, they are used to implement the methods provided in the various embodiments described above.

[0261] The above embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Any equivalent substitution or change made by a person skilled in the art based on the present invention is within the protection scope of the present invention.

Claims

1. A training method for an occupant injury prediction model, characterized in that: include: Acquire a training sample, wherein the training sample includes sample data and actual damage data corresponding to the sample data, wherein the sample data includes at least one of sample collision parameters, sample restraint system parameters, and sample occupant state parameters; Inputting the sample data into an occupant injury prediction model to perform occupant injury prediction, and obtaining predicted injury data corresponding to the sample data, wherein the occupant injury prediction model is used to predict injury data corresponding to multiple body parts respectively; Determining the total loss value of the predicted injury data relative to the actual injury data through a total loss function, wherein the total loss function may reflect the difference in loss weights of loss functions of different body parts; Based on the total loss value, model parameters of the occupant injury prediction model are adjusted.

2. The training method according to claim 1, characterized in that: The training method further comprises: Determine, by means of a loss function for each body part, a loss value of the predicted damage data relative to the actual damage data for each body part; Determine the loss weight of the loss function of each body part according to the loss value, wherein the greater the loss value, the greater the loss weight; The total loss function is updated based on the loss weight, and the total loss function is the sum of the loss functions of various body parts.

3. The training method according to claim 1 or 2, characterized in that: The total loss function is: in, is the loss function of the ith body part at the tth time step; w i (t) is the loss weight of the ith body part at the tth time step.

4. The training method according to claim 3, characterized in that: The loss weight is obtained by the following formula: Among them, β is a hyperparameter that controls sensitivity; k i (t-1) is the derivative of the loss function for the ith body part at the t-1th time step.

5. The training method according to claim 1, characterized in that: The method further comprises: Constructing a first loss function for each body part according to the predicted acceleration and the real acceleration, wherein the training sample also includes the real acceleration corresponding to the sample data; Constructing a second loss function for each body part based on the predicted injury data and the actual injury data; The total loss function is constructed according to the sum of the first loss functions and the sum of the second loss functions of each body part.

6. The training method according to claim 5, characterized in that: The step of inputting the sample data into an occupant injury prediction model to perform occupant injury prediction to obtain predicted injury data corresponding to the sample data includes: Inputting the sample data into an encoder of the occupant injury prediction model, performing feature extraction through the encoder, and obtaining feature data; Inputting the characteristic data into a decoder of the occupant injury prediction model, wherein the characteristic data are respectively decoded by using a multi-layer perceptron layer to obtain a representation of the predicted acceleration of each body part; In the decoder, a linear projection layer is used to decode the predicted acceleration of each body part respectively to obtain the predicted injury data corresponding to each body part; Determining the total loss value of the predicted damage data relative to the actual damage data through the total loss function includes: Determining a first loss value of the predicted acceleration relative to the actual acceleration by summing the first loss functions; Determining a second loss value of the predicted damage data relative to the actual damage data by summing the second loss functions; The total loss value is determined based on the first loss value and the second loss value.

7. The training method according to claim 6, characterized in that: In the decoder, the predicted acceleration of each body part is decoded by a linear projection layer to obtain the predicted injury data corresponding to each body part, including: In the decoder, an attention weight layer is used to weight the predicted acceleration of a body part and the related predicted accelerations of other body parts to update the predicted acceleration of each body part. The updated predicted acceleration of each body part is decoded by a linear projection layer to obtain the predicted injury data corresponding to each body part.

8. The training method according to any one of claims 5 to 7, characterized in that: The total loss function is: totalloss=α*totalloss aux +(1-α)totalloss main ; Among them: totalloss aux is the sum of the first loss functions of multiple body parts, totalloss main (t) is the sum of the second loss functions of the multiple body parts, and α is a hyperparameter.

9. The training method according to claim 8, characterized in that: The first loss function sum is expressed as follows: in, is the first loss function of the ith body part at the tth time step; num_task is the total number of body parts And / or, the second loss function sum is expressed in the following manner: in, is the loss function of the ith body part at the tth time step; w i (t) is the loss weight of the ith body part at the tth time step.

10. A method for predicting occupant injury, characterized in that: include: Obtaining one or more types of collision parameters, restraint system parameters of a vehicle, and state parameters of passengers in the vehicle under a collision scenario; One or more of the collision parameters, the restraint system parameters and the occupant status parameters are input into an occupant injury prediction model to obtain predicted injury data of the occupant in the collision scenario, wherein the occupant injury prediction model is obtained by the training method described in any one of claims 1 to 9.

11. The method according to claim 10, characterized in that The occupant injury prediction model comprises an encoder and a decoder, wherein one or more of the collision parameters, the restraint system parameters, and the occupant state parameters are input into the occupant injury prediction model to obtain predicted injury data of the occupant in the collision scenario, including: Input one or more of the collision parameters, the restraint system parameters, and the occupant state parameters into an occupant injury prediction model, and extract features through the encoder to obtain feature data; Inputting the characteristic data into a decoder of the occupant injury prediction model, wherein a multi-layer perceptron layer is used in the decoder to respectively decode and calculate the characteristic data to obtain predicted accelerations of various body parts; In the decoder, a linear projection layer is used to decode and calculate the predicted acceleration of each body part respectively to obtain the predicted injury data of the occupant in the collision scenario.

12. The method according to claim 11, characterized in that In the decoder, the predicted acceleration of each body part is decoded and calculated using a linear projection layer to obtain the predicted injury data of the occupant in the collision scenario, including: In the decoder, an attention weight layer is used to filter the predicted acceleration of each body part, and the predicted acceleration of the body part and the related predicted accelerations of other body parts are weightedly aggregated to update the predicted acceleration of each body part; The updated predicted acceleration is decoded using a linear projection layer to obtain predicted injury data of the occupant in the collision scenario.

13. A occupant injury prediction model training device, characterized in that: include: A first acquisition module, acquiring training samples, wherein the training samples include sample data and actual damage data corresponding to the sample data, wherein the sample data includes at least one of sample collision parameters, sample restraint system parameters, and sample occupant state parameters; A prediction model, used to predict occupant injury according to the input sample data, and obtain predicted injury data corresponding to the sample data, wherein the prediction model is used to predict injury data corresponding to multiple body parts respectively; A training module is used to determine the total loss value of the predicted injury data relative to the actual injury data through a total loss function, wherein the total loss function can reflect the different loss weights of loss functions of different body parts; and based on the total loss value, adjust the model parameters of the prediction model to obtain an occupant injury prediction model.

14. An occupant injury prediction device, characterized in that: include: A second acquisition module, used to acquire collision parameters in a current collision scenario, restraint system parameters of a vehicle, and one or more types of occupant status parameters in the vehicle; The occupant injury prediction model is used to predict occupant injury based on one or more of the input collision parameters, the restraint system parameters and the occupant status parameters to obtain predicted injury data of the occupant in the collision scenario.

15. An electronic device, characterized in that: comprising a processor and a memory in communication with the processor; The memory is used to store computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the method according to any one of claims 10 to 12.

16. A computer-readable storage medium / computer program product, characterized in that: The computer-readable storage medium stores computer-executable instructions / the computer program product includes computer-executable instructions, and when the computer-executable instructions are executed by a processor, they are used to implement the method according to any one of claims 1 to 12.

Images