An occupant injury prediction method, device and vehicle

Abstract

The application claims a passenger injury prediction method, device and vehicle, comprising: acquiring a crash waveform; pre-extracting a waveform characteristic value according to the crash waveform; inputting the crash waveform, the waveform characteristic value, driver and passenger information and restraint system information to obtain a real injury curve Y pre； Finally, the injury values of different parts of the human body are calculated to determine the injury level of the driver and passenger. The application pre-extracts the waveform characteristic value significantly related to human body injury as prior knowledge, and inputs the typical characteristics of the driver and passenger, the typical characteristics of the restraint system and the crash waveform. An injury prediction algorithm based on an encoder-decoder architecture is adopted. The encoder fuses the scalar characteristics into the vector characteristics, and the decoder maps the input hidden state to the injury of the human body. Thus, high-precision prediction of the passenger injury is realized with a small sample size.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of vehicle collision, and particularly relates to the field of occupant injury prediction in vehicle collision. BACKGROUND

[0002] Firstly, path planning in the automatic driving risk state needs to predict the injury of the occupant when a collision occurs in different planned paths in real time to calculate the cost of the path; secondly, the triggering logic of the pre-crash system is based on the collision risk and the personnel injury risk, when the collision risk and the personnel injury risk are both high, the highest level of pre-crash system is triggered to realize the early protection of the occupant safety; finally, engineers need to perform finite element simulation of occupant injury according to the collision waveform in the vehicle collision safety development process to realize the evaluation of the waveform, which is very time-consuming. If the given collision waveform can get the injury of the occupant in real time through the algorithm, the above problems will be solved.

[0003] Currently, the occupant injury prediction algorithm is mainly based on two methods. One is based on real traffic accident data, using regression, support vector machine, random forest and other machine learning algorithms, taking the collision scene parameters (relative speed, overlap rate, collision angle, collision target mass), driver and passenger information (gender, age, size, sitting posture, etc.) and restraint system state information (whether wearing a seat belt, whether the airbag is triggered, etc.) as input, and the personnel injury level as output, to train the prediction model. This method has problems such as difficult data collection and many default values. For example, the vehicle speed, overlap rate, collision angle and other parameters depend on the reconstruction after the accident, which is time-consuming and has low accuracy; most of the traffic accident data is low-level occupant injury, which has the problem of sample imbalance, making it difficult to train the machine learning model; the accident data only shows the injury level of the occupant after the collision and cannot provide specific injury values (such as head injury criterion HIC). The second method is to use simulation waveform as input and get occupant injury response curve as output through numerical simulation. This method adds the embedded representation of the input scalar value (such as driver and passenger information, restraint system state information) to the embedded representation of the waveform vector as the input of the deep learning model (such as LSTM, TCN) for pre-training to realize the prediction of occupant injury. This method has good prediction accuracy, but it does not incorporate prior knowledge and needs a large number of training samples to achieve end-to-end prediction. SUMMARY

[0004] In view of the above technical problems, the present application aims to provide a passenger injury prediction method and device and vehicle, based on data driving, using collision waveform, and adopting deep learning method to realize prediction of passenger injury, reduce sample data quantity, reduce calculation complexity, improve prediction accuracy, and provide accurate passenger injury level before collision, so as to provide input for path planning and pre-crash system triggering in automatic driving working condition dangerous scene, and provide accurate information for post-crash rescue.

[0005] The technical scheme of the present application is as follows:

[0006] In the first aspect, the present application provides a passenger injury prediction method, which comprises:

[0007] Obtaining a collision waveform: the collision waveform is a predicted collision waveform obtained before collision or a real collision waveform collected in real time after vehicle collision. That is, if the passenger injury prediction is used before collision, the collision waveform refers to a predicted collision waveform obtained by a waveform prediction module, and if it is used after collision, it refers to a real collision waveform actually collected by EDR / ADR.

[0008] Human injury prediction: according to the collision waveform, waveform feature values are extracted in advance; then the collision waveform, waveform feature values, driver and passenger information and restraint system information are input into a human injury prediction model to obtain a human injury kinematics curve Y' and a real injury curve Y through inverse mapping. pre

[0009] Personnel injury level determination: according to the real injury curve Y pre , the injury values of different parts of the human body are calculated, and the injury level of the driver and passenger is determined according to the injury values.

[0010] Specifically, when predicting human injury, waveform feature values are calculated according to the obtained collision waveform. That is, waveform feature values that have a significant impact on human injury are extracted in advance from the obtained collision waveform. The waveform feature values serve as prior knowledge and as one of the inputs of the passenger injury prediction algorithm.

[0011] According to one embodiment of the present application, the waveform feature values can include: vehicle collision speed v0, speed change amount Δv, rebound time t rebound , collision energy E, maximum X, Y, Z direction acceleration a x_max , a y_max , a z_max , average X, Y, Z direction acceleration a x_mean , a y_mean , a z_mean , maximum X, Y, Z axis rotation displacement d x_max , d y_max , d​z_max , the sliding window average acceleration SM W , the occupant load index OLC, the vehicle speed index VPI and the acceleration strengthening index ASI. In a specific implementation, if the acceleration curves in the Y and Z directions and the rotational displacement curves around the Y and Z axes cannot be obtained, the characteristic values can not include a y_max , a z_max , a y_mean , a z_mean d y_max , d z_max one or more of them.

[0012] According to an embodiment of the present application, the human body injury prediction model is obtained by constructing a human body injury database, using a human body injury prediction algorithm for calculation, and finally obtaining an occupant injury curve Y'.

[0013] According to an embodiment of the present application, the human body injury database is obtained by the following method:

[0014] First, a high-precision finite element model of the driver and passenger or a multi-rigid-body constraint system model is established.

[0015] Then, random sampling is performed in a space composed of a collision waveform, driver and passenger information, and constraint system state information to generate a simulation matrix table.

[0016] Then, the multi-rigid-body constraint system model is updated according to the simulation matrix table, and simulation calculation is completed, i.e., the occupant injury curves of different parts of the driver and passenger are obtained.

[0017] Finally, when all simulations in the matrix table are completed, the human body injury database is obtained.

[0018] The occupant injury curves of each part of the human body refer to the head acceleration curve, the neck axial force curve, the neck bending moment curve, the neck injury index curve, the chest compression curve, the chest acceleration curve, the femur axial force curve, the tibia axial force curve, the tibia shear force curve, and the bending moment curves of the tibia around the x and y axes.

[0019] According to an embodiment of the present application, the information, i.e., data, input into the human body injury prediction model is preprocessed, and the method is as follows:

[0020] For the numerical type data in the input features, the value range of all samples is mapped to the interval [0, 1]; for the classification data in the input features, the one-hot encoding method is used for representation.

[0021] For the output label, i.e. the occupant injury curve, the injury curve (e.g. head acceleration curve) data of all samples are mapped to [0, A] and rounded, specifically, all sample injury curves are concatenated by row into a 2-dimensional matrix Y, the first dimension represents the number of samples, and the second dimension represents the number of time series steps. Y is converted as follows to obtain Y'.

[0022]

[0023] int represents rounding,

[0024] max(Y) and min(Y) represent the maximum and minimum values of Y, respectively, which are constants after the sample set is determined.

[0025] The value of A is determined according to the predicted injury type, indicating that the possible value distribution of the injury prediction value at each time is in [0, A], and if the predicted injury value distribution interval is relatively small, the value of A is set to be larger, so as to better classify the injury level.

[0026] According to an embodiment of the present application, the personnel injury prediction algorithm is built by using a deep learning model based on an encoder-decoder architecture. Traditional deep learning models for processing time series, such as TCN based on convolutional neural network, recurrent neural network (RNN), Seq2Seq, etc., mostly use an end-to-end method, without fusing prior knowledge, and are not good at processing both sequential input and heterogeneous scalar data input. The present application simultaneously extracts feature values in advance (i.e. calculates waveform feature values according to the collision waveform as prior knowledge) and automatically extracts features by machine (i.e. the process of the following encoder processing vector is the process of automatically extracting features by machine), and effectively fuses scalar values into collision waveform time series data through transposed convolution, which accelerates the learning process of the model, improves the robustness of the model, and significantly improves the accuracy of the model.

[0027] Specifically, the encoder includes a scalar processing module and a vector processing module.

[0028] The vector processing module of the encoder maps the collision waveform to the vector data hidden layer H v_RNN . Specifically, the 1D convolutional layer extracts waveform features through kernel function weight update, and then inputs the features to the max pooling layer to reduce data redundancy and reduce the sensitivity of the convolutional layer to position. Then, the RNN module is inputted to map the sequence data to the hidden layer H v_RNN , and the H v_RNNThe output of the last layer of the RNN module is referred to as the output of the RNN module. The RNN module can be composed of one or more RNN layers, which can be original RNN layers or gate-based GRU or LSTM. The RNN module further extracts waveform key features (here, the key features are automatically extracted by the algorithm, and are also important to the human body injury), while compressing the historical information of the entire time sequence, representing the cumulative effect of the collision process on the occupant injury.

[0029] The scalar processing module of the encoder is a multi-layer perception 1 (MLP-1), which is calculated in parallel with the vector processing module. The scalar data is mapped to the scalar data hidden layer by the multi-layer perception (MLP-1), denoted as H s_MLP . The scalar data refers to the waveform feature value, the information of the driver and passenger in the vehicle, and the information of the restraint system.

[0030] The decoder includes a vector / scalar fusion module and a multi-layer perception 2 (MLP-2).

[0031] The vector / scalar fusion module of the decoder first increases the dimension of the vector data hidden layer H v_RNN and the scalar data hidden layer H s_MLP by one, respectively, to represent the vector / scalar dimension. At this point, it is assumed that the dimensions of H v_RNN and H S_MLP are changed to (batch_size, num_step, 1, num_channels), batch_size represents the batch size, num_step represents the number of time steps, 1 represents the vector / scalar dimension, and num_channels represents the high-dimensional representation of the data, i.e., the number of channels. H v_RNN and H S_MLP are spliced according to the vector / scalar dimension to obtain a matrix H VS_stack . Then, H VS_stack is upsampled by 2d transpose convolution, and the convolution kernel size of the transpose convolution is set to (w, 2), and the output is denoted as H transpose . The last dimension of H transpose is changed from 2 to 3 (i.e., the last dimension is composed of three vectors), where the first vector represents the vector feature, the second vector represents the feature after the fusion of the vector and the scalar, and the third vector represents the feature of the scalar. In order to accelerate the fusion of the vector and the scalar data and avoid overfitting, the third vector is cropped to only keep the first two vectors, and w is a hyperparameter that can be optimized. Then, in order to input the data to the MLP-2, the last three dimensions of H transpose are flattened, denoted as H flatten .

[0032] The multi-layer perception 2 (MLP-2) of the decoder takes H flattenMapped to occupant lesions, the number of input layer neurons in MLP-2 is equal to H. flatten The second dimension value, the number of neurons in the output layer, can be defined according to specific needs. Generally, sampling one point per 1ms of occupant injury output is sufficient to calculate the injury value. Assuming the original collision waveform duration is 200ms, the number of neurons in the output layer can be set to 200. The output of these 200 neurons constitutes Y'.

[0033] According to one embodiment of the present invention, the human injury prediction model further includes initialization and training, including:

[0034] The weights of each layer of the model are initialized using random sampling. Specifically, the weights of the MLP and RNN layers are randomly sampled from a Gaussian normal distribution with a mean of 0 and a variance of 1.0; the kernel weights of the convolution and transposed convolution are initialized using random sampling from the interval [0,1]. All bias terms are initialized to 0.

[0035] The forward propagation of the model involves inputting training samples into the model. Specifically, the preprocessed collision waveform is input into the scalar processing module, and the preprocessed scalar data is input into the vector processing module for forward propagation.

[0036] Backpropagation: The weights in the model are updated using mini-batch gradient descent (SGD).

[0037] The loss function is the cross-entropy loss function. In this application, the response curve to be predicted is mapped to the interval [0, A], and a total of A classes are classified for each time step of the output. That is, the output at each time step belongs to an integer value in the interval [0, A]. In this way, a regression problem is transformed into a classification problem, which can significantly improve the prediction accuracy.

[0038] During training, the above personnel injury database was divided into two parts: 80% for training and 20% for validation. After parameter tuning, if the correlation between the predicted loss curve and the actual loss curve does not meet the requirements, sample points are randomly added again to increase the sample size and retraining is performed until the accuracy meets the requirements.

[0039] In real-time prediction, after obtaining the occupant's injury curve Yˊ through the human injury prediction model, this application needs to perform an inverse mapping according to the following formula to obtain the actual injury curve Y. pre .

[0040]

[0041] max(Y) and min(Y) represent the maximum and minimum values ​​of Y in the training set, respectively, and are constants once the sample set is determined.

[0042] Finally, based on the actual damage curve Y preThe injury values of each part can be calculated, and the injury level of the member can be determined according to the injury values.

[0043] In a second aspect, the application further provides a data-driven occupant injury prediction device, comprising:

[0044] A collision waveform acquisition module is configured to obtain a predicted collision waveform before a collision or a real collision waveform collected in real time after a vehicle collision;

[0045] An occupant injury prediction module is configured to input the waveform characteristic values that have a significant influence on occupant injury, which are calculated in advance according to the collision waveform, the occupant information in the vehicle and the restraint system information into an occupant injury prediction model, to obtain an occupant injury kinematic curve Y' and perform inverse mapping to obtain a real injury curve Y pre ;

[0046] An injury level determination module is configured to calculate the injury values of different parts of the human body according to the real injury curve Y pre , and determine the injury level of the occupant according to the injury values.

[0047] The device realizes the data-driven occupant injury prediction method of the first aspect of the application through the above modules.

[0048] In a third aspect, the application further provides a vehicle configured with the data-driven occupant injury prediction device of the second aspect.

[0049] The above method of the application can be used not only before a collision but also after an accident. Before a collision, the predicted collision waveform is obtained through the waveform prediction module, and the predicted collision waveform, the waveform characteristic values calculated according to the predicted collision waveform, the occupant information in the vehicle and the restraint system information are used as the input of the occupant injury prediction. After an accident, the real collision waveform is read from the ACU / EDR, and the real collision waveform, the waveform characteristic values calculated according to the real collision waveform, the occupant information in the vehicle and the restraint system information are used as the input of the occupant injury prediction. The occupant injury prediction module calculates the occupant injury kinematic curve according to the input, further calculates the injury level of different parts of the human body, and finally sends the injury information of the occupant to the cloud server through a wireless network, and further notifies the relevant parties such as maintenance, insurance and rescue.

[0050] The technical effects of the application are as follows:

[0051] The present application adopts the method of pre-extracting waveform features and machine automatic extraction features, adopts pre-extraction of waveform feature values (scalar features) significantly related to human body injury as prior knowledge, and takes the typical features (scalar features) of the driver and passengers, the typical features (scalar features) of the restraint system and the original waveform of the collision as inputs, that is, by extracting feature values from the original waveform, the feature values are taken as the input of the neural network, which significantly speeds up the convergence of the algorithm, and higher precision can be obtained with the same number of samples, and high-precision prediction of the occupant injury can be realized with less sample amount, and the prediction accuracy of the model is significantly improved.

[0052] The personnel injury prediction algorithm of the present application is based on an encoder-decoder architecture, the encoder can effectively fuse scalar features into vector features, and the decoder maps the input hidden state to the injury of the human body. On the one hand, the personnel injury prediction algorithm adopted by the present application maps all scalars to the same space through a single-layer MLP, and obtains the contribution degree of different scalars through the weights of the hidden layer, which has better interpretability. On the other hand, the personnel injury prediction algorithm extracts features from the vector through a single-layer CNN, preserves the time relationship through a single-layer RNN, extracts features from the scalar through a single-layer MLP, and finally maps the features to the vector features through transposed convolution, the fusion features of the vector and the scalar, which has low computational complexity and preserves the spatial and temporal position relationship, ensures the accuracy of the algorithm under the condition of low algorithm complexity.

[0053] By adopting the technical scheme of the present application, after obtaining the occupant injury level before the collision, input can be provided for path planning in the automatic driving working condition and triggering of the pre-crash system in the dangerous scene, and accurate information can be provided for post-crash rescue. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 is a flowchart of the occupant injury prediction method of the present application;

[0055] Figure 2 is an implementation flowchart of the occupant injury prediction model;

[0056] Figure 3 is a collision waveform schematic diagram;

[0057] Figure 4 is a collision angle schematic diagram;

[0058] Figure 5 is a rebound moment definition schematic diagram;

[0059] Figure 6 is an OLC definition schematic diagram;

[0060] Figure 7 is a framework diagram of the personnel prediction algorithm;

[0061] Figure 8 is a schematic diagram of MLP;

[0062] Figure 9 is a schematic diagram of an occupant injury prediction device.

[0063] Figure 10 is a schematic diagram of vehicle hardware and software involved in configuring the occupant injury prediction device on a vehicle. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of the present application clearer, the embodiments will be described in detail below with reference to the drawings.

[0065] Reference is made to Figure 1 As an embodiment of the present application, the embodiment takes occupant injury prediction before vehicle collision as an example to illustrate the flow of prediction, including the following steps:

[0066] First, obtain the collision waveform: the collision waveform is obtained by prediction before the collision, or is obtained in real time after the vehicle collision. If the occupant injury prediction is used before the collision, the collision waveform refers to the predicted collision waveform obtained by the waveform prediction module, and if it is used after the collision, it refers to the real collision waveform actually collected by the EDR / ADR. In the prior art, there are many technologies disclosed for predicting the waveform before the collision, for example, the patent document CN107169235A equates the waveform to a second-order double-step waveform, and predicts the key parameters of the double-step waveform through a linear model. The waveform predicted by this method is a square wave, and the information of the original waveform is greatly lost, and it is limited to standard working conditions. The present applicant has also proposed a method for predicting a collision waveform (patent document No. CN115099096A), that is, the scene parameters (target object type, target object mass, target object speed, collision angle, and relative speed of two vehicles) obtained by the ADAS system can be input to the collision waveform prediction module to obtain the predicted collision waveform.

[0067] Then, human body injury prediction is performed: the waveform feature values are extracted in advance according to the collision waveform obtained in step 1, and are input into the human body injury prediction model together with the collision waveform obtained in step 1, the driver and passenger information detected by the in-vehicle occupant detection sensor, and the restraint system information from the CAN bus to obtain the human body injury kinematic curve Yˊ, and inverse mapping is performed to obtain the real injury curve Y pre；

[0068] Finally, the personnel injury level is determined: according to the real injury curve Y pre , the injury values of different parts of the human body are calculated, and the injury level of the driver and passenger is determined according to the injury values, which provides input for the pre-collision system triggering and automatic driving path planning.

[0069] For the present application, to realize the above occupant injury prediction method, the key is to realize the construction of a human body injury prediction model. In the following embodiment, the model construction process is described in detail.

[0070] Referring to Figure 2 The construction process of the human body injury prediction model mainly includes:

[0071] S1, constructing a collision waveform and calculating a waveform characteristic value;

[0072] S2, constructing an occupant injury database;

[0073] S3, data preprocessing;

[0074] S4, building a human body injury prediction algorithm;

[0075] S5, model initialization and training;

[0076] S6, real-time prediction.

[0077] The following will be described in detail:

[0078] S1, constructing a collision waveform and calculating a waveform characteristic value:

[0079] In the present application, the collision waveform database can be obtained by three ways: 1, obtained by whole vehicle finite element numerical simulation, 2, collected from the collision waveform recorded by the acceleration sensor of the accident vehicle, 3, constructed by the waveform equivalent formula.

[0080] 1, obtained by whole vehicle finite element numerical simulation, the process is as follows:

[0081] First, a high-precision self-vehicle finite element model and a finite element model of the collision target object are established.

[0082] Secondly, according to the target object type, target object mass, target object speed and the collision angle between the target object and the self-vehicle when the collision occurs, a simulation matrix table is generated by random sampling, which can make the target object impact the self-vehicle from different directions at different speeds. The collision speed boundary can be obtained according to the real accident statistical analysis, and the recommended value interval is [0, 120km / h].

[0083] Finally, the finite element model is updated according to the matrix table, and numerical calculation is carried out to obtain the collision waveform of the vehicle. The collision waveform refers to the acceleration curves of X, Y and Z directions in the whole vehicle coordinate system and the rotation displacement curves around X, Y and Z axes. When all the simulations in the matrix table are completed, different collision waveform databases can be obtained.

[0084] 2, collected from the collision waveform recorded by the acceleration sensor of the accident vehicle, the process is as follows:

[0085] The vehicle collision waveform collected by the acceleration sensor of the ego vehicle in a large number of accidents. The collection method can be through accident investigation and reconstruction by accident reconstruction software (such as PC-Crash), so as to obtain the collision waveform; or through cloud collection, that is, when the user's vehicle is involved in an accident, the collision waveform saved by the event data recorder (EDR) controller is automatically collected through the wireless network. After collecting a sufficient number of accidents (including different types of collision targets, different collision speeds and different collision angles, the specific number is determined according to the training effect of the algorithm), a database of different collision waveforms can be obtained.

[0086] 3. Construction by waveform equivalent formula

[0087] The equivalent formula refers to a simplified formula for approximating the collision waveform of the vehicle. The approximate formula can be a second-order waveform, a half-sine waveform, a Fourier waveform or a trapezoidal waveform. A specific example can adopt a Fourier waveform, which represents the real waveform by linear superposition of n sinusoidal waves, and the formula is as follows:

[0088]

[0089] a1 is the amplitude of the i-th sinusoidal waveform, ω is the natural frequency, t k is the effective collision time of the waveform. More specifically, n = 2, a1, a2, ω are sampled in the same space, and the value boundaries of a1, a2, ω are determined empirically, a1∈[0,0.5], a2∈[-0.1,0.5], ω∈[0.06,0.02]. After each sampling, a(t) is obtained according to

[0090] The final waveform is the first period greater than 0 of a(t).

[0091] If the maximum value of all constructed waveforms cannot cover the maximum value of the real collision waveform (the maximum value of the real waveform can be obtained by finite element simulation according to the collision speed boundary), a1 and a2 should be increased.

[0092] The obtained collision waveform is shown in Figure 3 .

[0093] In this embodiment, the ego vehicle refers to a research vehicle

[0094] In this embodiment, the target object types include passenger cars, trucks, buses and columnar objects.

[0095] In this embodiment, the collision speed refers to the speed of the target object when colliding with the ego vehicle. ​

[0096] In the embodiment, the collision angle refers to the direction of the velocity of the target object when the target object collides with the vehicle, and is also the direction of the force on the vehicle, as shown in the following figure. Figure 4

[0097] Further, after the collision waveform is obtained, the waveform characteristic values are calculated.

[0098] For the occupant injury prediction algorithm, whether used before or after the collision, the input includes four parts: the collision waveform, the waveform characteristic values, the information of the driver and passengers in the vehicle, and the information of the restraint system. If the occupant injury prediction algorithm is used before the collision, the collision waveform here is the predicted collision waveform, and the waveform characteristic values are calculated according to the predicted collision waveform. If used after the collision, the collision waveform here is the real collision waveform, and the waveform characteristic values are calculated according to the real collision waveform.

[0099] The waveform characteristic values can be: the vehicle collision speed v0, the speed change Δv, the rebound time t rebound , the collision energy E, the maximum X, Y, Z direction acceleration a x_max , a y_max , a z_max , the average X, Y, Z direction acceleration a x_mean , a y_mean , a z_mean , the maximum X, Y, Z axis rotation displacement d x_max , d y_max , d z_max , the sliding window average acceleration SM W , the occupant load index OL, the vehicle speed index VPI and the acceleration strengthening index ASI. In specific implementation, if the acceleration curves in Y, Z directions and the rotation displacement curves around Y, Z axes cannot be obtained, one or more of a y_max , a z_max , a y_mean , a z_mean d y_max , d z_max may not be included in the waveform characteristic values.

[0100] The above waveform characteristic values can be obtained by using the following calculation method:

[0101] The collision speed v0 of the vehicle refers to the speed v0 of the vehicle at the collision time, and the acceleration curve can be integrated to obtain the speed curve v(t), v0=v(t) t＝0 .

[0102] The speed change Δv refers to the speed difference Δv=v0-v T , v T =t) at the beginning and end of the collision of the vehicle​t＝T , v T is the velocity at the end of the collision, T is the duration of the collision.

[0103] the rebound time t rebound is the time when the velocity of the vehicle reverses. As shown in Figure 4 , if the velocity of the vehicle does not reverse (velocity does not return to 0) after the collision, the variance of the velocity of the vehicle is calculated for 10 consecutive sampling points (1 ms sampling for a point of the velocity curve), and a velocity variance curve is obtained. If the variance values of 10 consecutive points (the time interval corresponding to the 10 points is defined as [t, t+10) ms) on the velocity variance curve are all less than a threshold θ, the rebound time t rebound = t+5 ms, and θ can be 0.005.

[0104] the collision energy

[0105] the maximum value of the X-direction acceleration a x_max = max(abs(a x (t)), and similarly, the maximum values of the Y and Z-direction accelerations can be solved.

[0106] the average value of the X-direction acceleration a and similarly, the average values of the Y and Z-direction accelerations can be solved.

[0107] the maximum value of the rotation displacement around the X-axis d x_max = max(abs(d x (t)), d x (t) is the rotation displacement around the X-axis, and similarly, the maximum values of the rotation displacements around the Y and Z axes can be solved.

[0108] the sliding window average acceleration W represents the window width of the sliding window, and is generally 3 ms.

[0109] the occupant load indicator OL C, the OL C indicator is proposed by Dr. Kübler, as shown in Figure 6 , in the initial stage of the collision, the restraint system has no restraining effect on the human body, and the occupant is in a free state (free flight stage), and the occupant maintains the motion in this free state to t1 time, and the motion distance is s. Thereafter, the restraint system begins to act, and the human body moves at a constant deceleration until it contacts the remaining chest space d-s, and the corresponding time is t2, and the value of the constant deceleration is the OL C value. Generally, d takes the X-direction distance of the chest of the driver and passenger from the steering wheel (dashboard), and s takes 60 mm. The OL C is obtained by iterative calculation according to the following formula.

[0110]

[0111]

[0112] v0-OLC·(t2-t1)=v(t2)

[0113] where v0 is the initial velocity of the collision and v(t) is the velocity curve of the collision.

[0114] The VPI is proposed by the International Organization for Standardization (ISO) for evaluating the degree of influence of the collision acceleration on the occupant injury, and the VPI is obtained by iterative calculation according to the following formula.

[0115]

[0116]

[0117]

[0118] where M is the mass of the occupant, y(t) is the displacement of the occupant, x(t) is the displacement of the vehicle, P(t) represents the force between the restraint system and the occupant, s is the slack of the restraint system, and k is the stiffness of the restraint system. Generally, k is 2500 N / m and s is 30 mm.

[0119] The acceleration reinforcement index ASI is calculated according to the following formula:

[0120]

[0121] SM 50ms,x ,SM 50ms,y ,SM 50ms,z respectively represent the 50 ms sliding window average acceleration values of the X, Y and Z directions of the ego vehicle.

[0122] S2 constructs a human body injury database:

[0123] First, a sample of n occupant injury matrix table is generated by random sampling in the same space composed of the collision waveform, the driver and passenger information, and the restraint system state information. The collision waveform refers to the collision waveform obtained by S1. The different driver and passenger information refers to the information of the gender, age, size, and pre-crash posture of the driver and passenger, and the pre-crash posture refers to the roll angle, the forward angle and the distance from the chest of the driver and passenger to the center of the steering wheel (for the driver), to the instrument panel (for the front passenger) or to the front seat (for the rear passenger) at the time of the collision. The restraint system state information refers to the information of the restraint system affecting the safety of the driver and passenger, such as whether the driver and passenger are wearing seat belts and whether the airbag is triggered.

[0124] Then, a high-precision restraint system model can be built, which can be a finite element model or a multi-rigid-body model. The model mainly includes components that can be contacted by the human body during the collision process, such as airbags, steering columns, seat belts, steering wheels, seats, instrument panels, door trims, roofs, carpets, and a body-in-white.

[0125] Finally, the driver and passenger information and the restraint system state information parameters in the occupant injury matrix table are updated to the finite element model, the collision waveform is taken as the boundary condition of the restraint system model, and numerical simulation calculation is performed. After the calculation is completed, the human body injury curve is extracted from the calculation result to form a human body injury database with the input (features) of the collision waveform, the driver and passenger information, and the restraint system state information, and the output (labels) of the injury curves of each part of the human body. The injury curves of each part of the human body refer to the head acceleration curve, the neck axial force curve, the neck bending moment curve, the neck injury index curve, the chest compression curve, the chest acceleration curve, the femur axial force curve, the tibia axial force curve, the tibia shear force curve, and the tibia bending moment curve around the x-axis and the y-axis.

[0126] S3 data preprocessing:

[0127] For the collision waveform in the input features, the value range is mapped to the interval [0, 1]. Specifically, the collision waveform data of all samples is spliced into a 2-dimensional matrix X by row, with the first dimension representing the number of samples and the second dimension representing the number of time series steps. X is converted as follows to obtain X′:

[0128]

[0129] max(X) and min(X) represent the maximum and minimum values of X, respectively, which are constants after the sample set is determined.

[0130] For the classification data in the input features, a one-hot encoding method is used to represent a vector with a length of the total number of categories, with the category corresponding component set to 1 and all other components set to 0. A specific example uses [1, 0] and [0, 1] to represent the two categories of airbag point explosion and non-point explosion, respectively.

[0131] For the numerical data in the input features, the same preprocessing method as the collision waveform is used.

[0132] For the output label, i.e., the injury curve of the occupant, the injury curve (such as the head acceleration curve) data of all samples is spliced into a 2-dimensional matrix Y by row, with the first dimension representing the number of samples and the second dimension representing the number of time series steps. Y is converted as follows to map the data to the interval [0, A] to obtain Y′.

[0133]

[0134] int represents integer division.

[0135] max(Y) and min(Y) represent the maximum and minimum values ​​of Y, respectively, and are constants once the sample set is determined.

[0136] The value of A is determined based on the predicted injury type, representing the possible distribution of the predicted injury value at each time step within the range [0, A]. If the predicted injury value distribution range is relatively small, the value of A is set larger to better classify the injury level. Based on experience, A can be set to 500 for head acceleration curves; 200 for neck injury index curves; and 100 for chest compression curves.

[0137] S4 is used to build a human injury prediction algorithm:

[0138] Personnel injury prediction employs a deep learning model based on an encoder-decoder architecture, such as... Figure 7 As shown. Traditional deep learning models for processing time series data, such as TCN, RNN, and Seq2Seq based on convolutional neural networks, mostly adopt an end-to-end approach, without incorporating prior knowledge, and are not adept at handling inputs of existing sequences and heterogeneous scalar data. This invention employs both pre-extracted feature values ​​and automatic machine feature extraction, and effectively integrates scalar values ​​into the collision waveform time series data through transposed convolution, accelerating the model's learning process, improving model robustness, and significantly enhancing model accuracy.

[0139] The encoder consists of a scalar processing module and a vector processing module.

[0140] The encoder's vector processing module maps the original collision waveform (the predicted collision waveform before a vehicle collision, and the real collision waveform acquired in real time after a vehicle collision) to a vector latent state.

[0141] Specifically, for the acceleration sequence X in the X direction v_0 The input is fed into a 1D convolutional layer, where waveform features are extracted and updated using kernel weights. This input is then fed into a max-pooling layer to reduce data redundancy and decrease the convolutional layer's sensitivity to position. The output of the pooling layer is X. v CNN After X v CNN The input is fed into the RNN module, which maps the sequence to the vector data hidden layer H. v RNNThe RNN module can consist of one or more RNN layers, which can be a primitive RNN layer, or a gated GRU or LSTM-based layer. The RNN module further extracts key waveform features and compresses historical information from the entire time series to characterize the cumulative damage effect on the occupants caused by the collision. The H... v_RNN H refers to the output of the last layer of the RNN module. v_RNN The dimensions are denoted as (batch_size, num_step, num_channels), where batch_size is the batch size, num_step is the number of time steps, and num_channels is the number of high-dimensional representations of the data, i.e., the number of channels.

[0142] The encoder's scalar processing module is a multilayer perceptron (denoted as MLP-1). A schematic diagram of the multilayer perceptron is shown below. Figure 8 As shown. The waveform characteristic values, occupant information, and restraint system information obtained previously are used as scalar data, scalar data X. s_0 Data is mapped to a scalar data hidden layer using MLP-1, denoted as H. s_MLP The number of neurons in the input layer of an MLP-1 is equal to the number of scalars, and the number of neurons in the last layer of an MLP-1 is equal to H. v_RNN The time step (num_step), assuming H s_MLP The dimension is (batch_size, num_step), in order to match the vector data hidden layer H v_RNN , H S_MLP Add one dimension and copy num_channels times to increase the dimension to a 3D matrix, H S_MLP The dimensions become (batch_size, num_step, num_channels).

[0143] Furthermore, the decoder consists of a vector / scalar fusion module and a multilayer perceptron (denoted as MLP-2).

[0144] The decoder's vector / scalar fusion module first hides the vector data in the H layer. v_RNN and scalar data hidden layer H s_MLP Add one dimension to each, representing the vector / scalar, and thus H v_RNN and H S_MLP The dimensions are all changed to (batch_size, num_step, 1, num_channels), and H is... v_RNN and H S_MLP Matrix H is obtained by concatenating along this dimension. VS_stack H VS_stackThe dimensions are (batch_size, num_step, 2, num_channels). To facilitate transpose convolution calculation, H is swapped. VS_stack The num_step dimension and num_channels dimension, H VS_stack The dimensions become (batch_size, num_channels, num_step, 2). Then, H is processed by a 2D transpose convolution. VS_stack Upsampling and transposing the convolution kernel size to (w, 2) yields an output denoted as H. transpose w is a hyperparameter that can be optimized, H transpose The last dimension changes from 2 to 3 (i.e., the last dimension consists of 3 vectors). The first vector represents the vector features, the second vector represents the features resulting from the fusion of vector and scalar data, and the third vector represents the scalar features. To accelerate the fusion of vector and scalar data and avoid overfitting, the third vector is pruned, retaining only the first two vectors. Then, to input the data into MLP-2, H... transpose The last three dimensions are flattened, denoted as H. flatten .

[0145] The input of the MLP-2 decoder is H flatten The number of neurons in the input layer of MLP-2 is equal to H. flatten The second dimension value, the number of neurons in the output layer, can be defined according to specific needs. Generally, sampling one point per 1ms of occupant injury output is sufficient to calculate the injury value. Assuming the original collision waveform duration is 200ms, the number of neurons in the output layer can be set to 200. The output of these 200 neurons constitutes Y'.

[0146] The number of intermediate layers in a multilayer perceptron is a hyperparameter that can be optimized. After the hidden layers in an MLP, a dropout layer is connected. The dropout probability p-value of the dropout layer is a hyperparameter that can be optimized. The activation function type can be ReLU, Sigmoid, etc., which are hyperparameters that can be optimized.

[0147] S5 Model Initialization and Training:

[0148] The weights of each layer in the model are initialized using random sampling. Specifically, the weights of the MLP and RNN layers are randomly sampled from a Gaussian normal distribution with a mean of 0 and a variance of 1; the kernel weights of the convolution and transposed convolution are initialized using random sampling from the interval [0,1]. All bias terms are initialized to 0.

[0149] The forward propagation of the model involves inputting training samples into the model. Specifically, the preprocessed collision waveform is input into the scalar processing module, and the preprocessed scalar data is input into the vector processing module for forward propagation.

[0150] Backpropagation uses mini-batch gradient descent (SGD) to update the weights in the model.

[0151] The loss function is the cross-entropy loss function. This invention maps the response curve to be predicted to the interval [0, A], and performs a total classification of A classes for each output time step. That is, the output at each time step belongs to an integer value in the interval [0, A]. In this way, a regression problem is transformed into a classification problem, which can significantly improve the prediction accuracy.

[0152] The personnel injury database is divided into two parts: 80% for training and 20% for validation. After parameter tuning, if the correlation between the predicted loss curve and the actual loss curve does not meet the requirements, the process returns to step S2, randomly adding sample points to increase the sample size and retraining until the accuracy meets the requirements.

[0153] S6 Real-Time Prediction:

[0154] After obtaining the collision waveform, the waveform is preprocessed according to S3 and input into the prediction model pre-trained by S5 to obtain the occupant injury curve Yˊ. Then, the true injury curve Y is obtained by inverse mapping according to the following formula. pre .

[0155]

[0156] max(Y) and min(Y) represent the maximum and minimum values ​​of Y in the training set, respectively, and are constants once the sample set is determined.

[0157] The following examples illustrate how to determine the degree of occupant injury:

[0158] First, to obtain the actual injury curve Y of the human body pre Then, the damage value for each part of the human body can be calculated, as follows:

[0159] 1. Head Injury Index (HIC). This index represents the severity of skull fractures by measuring the acceleration at the occupant's head center of gravity, measured in meters (m). 2.5 / s 4 The terminal can acquire time-series data of occupant head center of mass acceleration from collision simulation data.

[0160]

[0161] Among them, a g (t) is the acceleration of the occupant's head center of mass at collision time t, expressed in units of gravitational acceleration g; t0 and t TThese are the collision start time and collision end time, respectively; collision time t1 and collision time t2 are the start time and end time used to calculate HIC, respectively. The difference between t1 and t2 is less than a preset threshold, for example, the preset threshold can be 15ms, that is, t2-t1 is less than 15ms.

[0162] 2. Neck Injury Index Nij. Nij is calculated from the combination of axial force and bending moment at the occiput condyle of the occiput neck during the collision. The terminal can obtain the time series data of axial force and bending moment at the occiput condyle of the occiput neck from the collision simulation data, as well as the load condition type of the neck at each collision moment. The terminal calculates the maximum value of Nij based on the axial force and bending moment at the occiput condyle of the occiput neck at each collision moment according to the following formula.

[0163]

[0164] Among them, F z and M y F represents the axial force and bending moment at the occipital condyle of the occupant's neck at the time t of the collision. int and M int These are the intercepts of axial force and bending moment intersecting the axis, respectively. Specifically, the reference values ​​differ for positive (tension) and negative (compression) axial force, and for negative (tension) and positive (bending) bending moment. For details, please refer to General Motors' Injury Assessment Reference Values ​​(IARVs), as shown in the table below:

[0165] Human dimensions Fz+(N) Fz-(N) My+(N.m) My-(N.m) 3-year-old child 2330 2130 67 29.3 6-year-old child 3080 2820 96 42 10-year-old child 3710 3390 125 54.8 5 th female 4260 3900 153 66.9 50 th male 6780 6200 305 133 95 th male 8180 7480 405 177

[0166] 3. Chest injury index C disp C disp This represents the maximum compression of the occupant's chest during the collision. The displacement-time series curve of chest feature points relative to the back can be obtained from the collision results.

[0167] 4. Maximum axial force of the femur (F) z-femur F z-femur This represents the maximum axial force borne by the occupant's femur during the collision. The maximum axial force can be obtained by acquiring the time-series data of the axial force borne by the occupant's femur from the collision simulation data and calculating its maximum value.

[0168] 5. Tibial Index (TI): TI is calculated from the combination of axial force and bending moment at the occupant's tibia during a collision. It can be obtained from the collision simulation data, specifically the time-series data of axial force and bending moment at the occupant's tibia (including bending moments in both the x and y axes). Then, the TI-time series curve is calculated using the following formula, and its maximum value is the TI value.

[0169]

[0170] Among them, F z M is the axial force at the tibia of the occupant at the moment of impact t. x and M y M represents the bending moment along the x-axis and y-axis at the occupant's tibia at the moment of impact t, respectively. c and F c This is the default value; please refer to FMVSS208, M. c =225 N·m, F c =35.9kN.

[0171] The above five items are injury indicators for various parts of the body adopted in practice. It should be noted that a single injury indicator can also be other indicators. For example, the head injury can be assessed based on the head rotation speed index Bric, the abdominal injury can be assessed based on the maximum abdominal compression, the upper limb injury can be assessed based on the maximum axial force and maximum bending moment of the upper limb (forearm or upper arm), and the foot injury can be assessed based on the maximum bending moment or maximum acceleration of the ankle joint, etc.

[0172] Furthermore, after obtaining the damage values ​​for each body part, the damage level for each body part is calculated based on the damage risk curve in FMVSS208. The specific calculation method is as follows:

[0173] After obtaining the injury values ​​for each site, these values ​​are converted into occupant AIS (Abbreviated Injury Scale) injury index information. AIS uses a scoring system of 1-6 points for each injury based on the anatomical extent of damage to tissues and organs, with AIS 3+ indicating severe injury.

[0174]

[0175]

[0176] For example, after obtaining the HIC value of the occupant's head injury index, the occupant's head injury index is converted into the probability corresponding to different injury levels according to the following formula. When a certain loss level first exceeds the threshold (such as 25%), that level of injury has occurred.

[0177]

[0178]

[0179]

[0180]

[0181]

[0182]

[0183] In a specific example, assuming HIC = 1000, substituting HIC into the above formula, we can obtain the risk values ​​for head injuries of grades 1, 2, 3, 4, 5, and 6 as 99.15%, 89.19%, 52.75%, 16.80%, 2.37%, and 0.11%, respectively. Therefore, the head injury was grade 3.

[0184] For injury indicators and probability curves of injury levels for other parts of the body, please refer to the US Road Traffic Safety Code FMVSS208.

[0185] The present invention also relates to an apparatus, such as Figure 9 As shown, the device includes:

[0186] The collision waveform acquisition module is used to obtain the collision waveform by prediction before the collision, or to collect the real collision waveform in real time after the vehicle collision.

[0187] The human injury prediction module is used to calculate waveform feature values ​​that have a significant impact on human injury based on the collision waveform, and input these values, along with the collision waveform, vehicle occupant information, and restraint system information, into the human injury prediction model to obtain the human injury kinematic curve Y', and then perform inverse mapping to obtain the actual injury curve Y'. pre .

[0188] The damage level determination module is used to determine the damage level based on the actual damage curve Y. pre It calculates the injury values ​​of different parts of the human body and determines the injury level of drivers and passengers based on the injury values.

[0189] The system implements the occupant injury prediction method disclosed in the previous embodiments through the above modules.

[0190] like Figure 10 As shown, the device is configured on a car and involves the vehicle's computer, data storage device, ADAS environmental perception sensor, occupant detection sensor, vehicle kinematic parameter sensor, and may also involve wireless network and cloud server.

[0191] Before a collision, ADAS environmental perception sensors and vehicle kinematic parameter sensors input scene parameters (target type, target mass, target speed, collision angle, and relative speed of the vehicle) into the waveform feature value prediction model to obtain feature values. The computer calculates the predicted collision waveform (e.g., patent document CN115099096A). The predicted collision waveform, the waveform feature values ​​calculated based on the predicted collision waveform, the information of the occupants and the restraint system are used as inputs for occupant injury prediction to obtain the human injury kinematic curve. Furthermore, the computer calculates the injury level of different parts of the human body, providing input for the pre-collision system triggering and autonomous driving path planning. The data storage module stores the constants and prediction models required during the calculation process.

[0192] If this method is to be used after an accident, the actual collision waveform is read from the ACU / EDR. The actual collision waveform, the waveform feature values ​​calculated based on the actual collision waveform, the information of the occupants and the restraint system are used as inputs for occupant injury prediction. The computer calculates the kinematic curve of human injury and further calculates the injury level of different parts of the human body. Finally, the injury information of the occupants is sent to the cloud server via wireless network, thereby notifying relevant parties such as maintenance, insurance and rescue.

Claims

1. A method for predicting occupant injury, characterized in that, The method includes: Acquiring collision waveforms: The collision waveforms are either obtained by prediction before the collision or are real collision waveforms collected in real time after the vehicle collision. Human Injury Prediction: Based on the collision waveform, waveform feature values ​​are extracted in advance. These waveform feature values ​​refer to waveform characteristics that significantly affect human injury when calculating the collision waveform, and include at least the vehicle collision speed v0, the speed change Δv, and the rebound time t. rebound Collision energy E, maximum acceleration a in the X, Y, and Z directions x_max a y_max a z_max The average acceleration a in the X, Y, and Z directions x_mean a y_mean a z_mean The maximum displacement d around the X, Y, and Z axes x_max d y_max d z_max Sliding window average acceleration SM W One or more of the following: Occupant Load Index (OLC), Vehicle Speed ​​Index (VPI), and Acceleration Intensity Index (ASI); The rebound time t rebound This refers to the moment when the vehicle's speed reverses. If the vehicle's speed direction does not reverse after the collision, the variance of the vehicle's speed over N consecutive sampling points is calculated to obtain a speed variance curve. If the variance values ​​at N consecutive points on the speed variance curve are all less than a threshold θ, then the rebound time t is considered a true time. rebound = t+5ms, θ is 0.005; the time interval corresponding to these N points is defined as [t, t+10)ms, where N is a positive integer; The acceleration enhancement index (ASI) is calculated using the following formula: SM 50ms,x SM 50ms,y SM 50ms,z These represent the average acceleration values ​​over 50ms in the X, Y, and Z directions of the vehicle during window sliding. The collision waveform, waveform feature values, occupant information, and restraint system information are input into the human injury prediction model to obtain the human injury kinematic curve Yˊ, and then inverse mapping is performed to obtain the actual injury curve Y. pre ; The personnel injury prediction model is constructed using a deep learning model based on an encoder-decoder architecture; The encoder includes a scalar processing module and a vector processing module; The vector processing module maps the collision waveform to the vector data hidden layer H through a 1D convolutional layer and an RNN module. v_RNN ; The scalar processing module is a multilayer perceptron-1 (MLP-1), which performs computations in parallel with the vector processing module. Scalar data is mapped to a scalar data hidden layer, denoted as H, through the MLP-1. s_MLP The scalar data refers to the waveform characteristic values, information on the occupants inside the vehicle, and information on the restraint system. The decoder includes a vector / scalar fusion module and a multilayer perceptron 2 (MLP-2); the vector / scalar fusion module first hides the vector data in the hidden layer H. v_RNN and scalar data hidden layer H s_MLP Add a dimension to each, representing the vector / scalar dimension, and change H. v_RNN and H S_MLP The matrix H is obtained by concatenating the vector / scalar dimensions. VS_stack Then, H is subjected to transposed convolution. VS_stack Upsampling, the resulting output is denoted as H. transpose , will H transpose The rear dimension flattening, denoted as H flatten ; Personnel injury level determination: based on the actual injury curve Y pre It calculates the injury values ​​of different parts of the human body and determines the injury level of drivers and passengers based on the injury values.

2. The occupant injury prediction method according to claim 1, characterized in that, Among the waveform feature values: The maximum acceleration in the X direction, a x_max =max(abs(a x (t)), where a x (t) represents the collision waveform of the vehicle in the X direction, where abs() indicates taking the absolute value and max() indicates taking the maximum value; the maximum acceleration values ​​in the Y and Z directions are calculated in the same way as in the X direction; the average acceleration value in the X direction... In the formula, a x (t) represents the collision waveform of the vehicle in the X direction, and T represents the moment when the collision ends; the average acceleration values ​​in the Y and Z directions are calculated in the same way as in the X direction. The maximum rotational displacement d around the X-axis x_max =max(abs(d x (t)), where d x (t) represents the rotational displacement about the X-axis, abs() represents taking the absolute value, and max() represents taking the maximum value; the maximum value of the rotational displacement in the Y and Z directions is calculated in the same way as in the X direction. The average acceleration of the sliding window In the formula a x (t) represents the collision waveform in the X direction of the vehicle, W represents the window width of the sliding window, and T represents the moment when the collision ends.

3. The data-driven occupant injury prediction method according to claim 1 or 2, characterized in that, The human injury prediction model is calculated using a human injury prediction algorithm based on a constructed human injury database, and finally obtains the occupant's injury curve Yˊ.

4. The occupant injury prediction method according to claim 3, characterized in that, The human injury database is constructed in the following manner: Establish a high-precision finite element model or multi-rigid-body constraint system model of the driver and passengers; Random sampling is performed in the space composed of collision waveform, information of different drivers and passengers, and restraint system status information to generate an occupant injury matrix table; The multi-rigid-body constraint system model is updated according to the occupant injury matrix table. The collision waveform is used as the boundary condition of the constraint system model. Numerical simulation calculation is performed, and the human injury curve is extracted from the calculation results to obtain the occupant injury curves of different parts of the driver and passenger. After completing all simulations in the matrix table, the system outputs a human injury database with input features including collision waveforms, occupant information, and constraint system status information, and output labels including injury curves of various parts of the human body.

5. The occupant injury prediction method according to claim 4, characterized in that, The different occupant information refers to the occupant's gender, age, size, and pre-collision posture information; the pre-collision posture refers to the occupant's lateral tilt angle, forward tilt angle, and distance from the occupant's chest to the steering wheel center, dashboard, or front seat at the moment of collision; the restraint system status information refers to restraint system information affecting occupant safety, such as whether the occupant is wearing a seatbelt and whether the airbag has deployed.

6. The occupant injury prediction method according to claim 4, characterized in that, The high-precision finite element model or multi-rigid-body constraint system model of the driver and passengers includes components that the human body may come into contact with during a collision, such as airbags, steering column, seat belts, steering wheel, seats, dashboard, door trim, headliner, carpets, etc., as well as the body-in-white.

7. The occupant injury prediction method according to claim 4, characterized in that, The injury curves for various parts of the human body refer to the head acceleration curve, neck axial force curve, neck bending moment curve, neck injury index curve, chest compression curve, chest acceleration curve, femoral axial force curve, tibia axial force curve, tibia shear force curve, and tibia bending moment curves around the x-axis and y-axis.

8. The occupant injury prediction method according to claim 4, characterized in that, The data input to the human injury prediction model needs to be preprocessed. The method is as follows: for numerical data in the input features, the value range of all samples is mapped to the interval [0,1]. For categorical data in the input features, one-hot encoding is used to represent them. For the output label, i.e. the occupant injury curve, the injury curve data of all samples are mapped to [0, A] and rounded. The value of A is determined according to the predicted injury type, indicating that the possible values ​​of the injury prediction at each time step are distributed in [0, A].

9. The occupant injury prediction method according to claim 4, characterized in that, After obtaining the occupant's damage curve Yˊ, the true damage curve Y is obtained by inverse mapping using the following formula. pre max(Y) and min(Y) represent the maximum and minimum values ​​of Y in the training set, respectively, and are constants once the sample set is determined.

10. The occupant injury prediction method according to claim 4, characterized in that, It also includes initializing and training the human injury prediction model, including: The weights of each layer of the model are initialized by random sampling; Forward propagation of the model: The preprocessed collision waveform is input to the scalar processing module, and the preprocessed scalar data is input to the vector processing module for forward propagation; Backpropagation: The weights in the model are updated using mini-batch gradient descent (SGD). The loss function is the cross-entropy loss function.

11. An occupant injury prediction device, characterized in that, include The collision waveform acquisition module is used to obtain the collision waveform by prediction before the collision, or to collect the real collision waveform in real time after the vehicle collision. The human injury prediction module is used to calculate waveform feature values ​​that have a significant impact on human injury based on the collision waveform. These waveform feature values ​​include the vehicle collision speed v0, the speed change Δv, and the rebound time t. rebound Collision energy E, maximum acceleration a in the X, Y, and Z directions x_max a y_max a z_max The average acceleration a in the X, Y, and Z directions x_mean a y_mean a z_mean The maximum displacement d around the X, Y, and Z axes x_max d y_max d z_max Sliding window average acceleration SM W One or more of the following: Occupant Load Index (OLC), Vehicle Speed ​​Index (VPI), and Acceleration Intensity Index (ASI); the rebound time t rebound This refers to the moment when the vehicle's speed reverses. If the vehicle's speed direction does not reverse after the collision, the variance of the vehicle's speed over N consecutive sampling points is calculated to obtain a speed variance curve. If the variance values ​​at N consecutive points on the speed variance curve are all less than a threshold θ, then the rebound time t is considered a true time. rebound = t+5ms, θ is 0.005; the time interval corresponding to these N points is defined as [t, t+10)ms, where N is a positive integer; The acceleration enhancement index (ASI) is calculated using the following formula: SM 50ms,x SM 50ms,y SM 50ms,z These represent the average acceleration values ​​over 50ms in the X, Y, and Z directions of the vehicle during window sliding. The collision waveform, occupant information, and restraint system information are input into the human injury prediction model to obtain the human injury kinematic curve Yˊ, and then inversely mapped to obtain the actual injury curve Y. pre ; The personnel injury prediction model is constructed using a deep learning model based on an encoder-decoder architecture; The encoder includes a scalar processing module and a vector processing module; The vector processing module maps the collision waveform to the vector data hidden layer H through a 1D convolutional layer and an RNN module. v_RNN ; The scalar processing module is a multilayer perceptron-1 (MLP-1), which performs computations in parallel with the vector processing module. Scalar data is mapped to a scalar data hidden layer, denoted as H, through the MLP-1. s_MLP The scalar data refers to the waveform characteristic values, information on the occupants inside the vehicle, and information on the restraint system. The decoder includes a vector / scalar fusion module and a multilayer perceptron 2 (MLP-2); the vector / scalar fusion module first hides the vector data in the hidden layer H. v_RNN and scalar data hidden layer H s_MLP Add a dimension to each, representing the vector / scalar dimension, and change H. v_RNN and H S_MLP The matrix H is obtained by concatenating the vector / scalar dimensions. VS_stack Then, H is subjected to transposed convolution. VS_stack Upsampling, the resulting output is denoted as H. transpose , will H transpose The rear dimension flattening, denoted as H flatten ; The damage level determination module is used to determine the damage level based on the actual damage curve Y. pre Calculate the injury values ​​of different parts of the human body, and determine the injury level of the driver and passengers based on the injury values; The device implements the occupant injury prediction method according to any one of claims 1-10 through the above modules.

12. A vehicle, characterized in that, The vehicle is equipped with the occupant injury prediction device as described in claim 11.

Images