CN115099096A - Vehicle collision waveform prediction method based on data driving, electronic equipment and readable storage medium - Google Patents

Vehicle collision waveform prediction method based on data driving, electronic equipment and readable storage medium Download PDF

Info

Publication number
CN115099096A
CN115099096A CN202210736800.4A CN202210736800A CN115099096A CN 115099096 A CN115099096 A CN 115099096A CN 202210736800 A CN202210736800 A CN 202210736800A CN 115099096 A CN115099096 A CN 115099096A
Authority
CN
China
Prior art keywords
collision
waveform
prediction
model
curve
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210736800.4A
Other languages
Chinese (zh)
Inventor
李锐阳
毛溶洁
崔泰松
张彬
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chongqing Changan Automobile Co Ltd
Original Assignee
Chongqing Changan Automobile Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Chongqing Changan Automobile Co Ltd filed Critical Chongqing Changan Automobile Co Ltd
Priority to CN202210736800.4A priority Critical patent/CN115099096A/en
Publication of CN115099096A publication Critical patent/CN115099096A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/23—Design optimisation, verification or simulation using finite element methods [FEM] or finite difference methods [FDM]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G06Q10/047—Optimisation of routes or paths, e.g. travelling salesman problem
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Human Resources & Organizations (AREA)
  • Economics (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Strategic Management (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Artificial Intelligence (AREA)
  • Development Economics (AREA)
  • Geometry (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Game Theory and Decision Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a vehicle collision waveform prediction method based on data driving, an electronic device and a storage medium, wherein the method comprises the following steps: s1 builds a collision waveform database: establishing a collision waveform database by adopting a finite element method according to the collision scene parameters and the boundary thereof; s2 obtains a feature value of the waveform: extracting characteristic values of the waveform of each simulation sample in the collision waveform database, wherein the characteristic values comprise a real acceleration curve in the X, Y, Z direction and a real rotary displacement curve around the X, Y, Z axis; s3 constructs a machine learning model: a prediction model 1 for predicting the characteristic value of the acceleration curve and a prediction model 2 for predicting the characteristic value of the rotary displacement curve; s4 waveform prediction: and (3) respectively inputting the specific parameters of the scene into the prediction models 1 and 2 to respectively obtain the characteristic value of an X, Y, Z direction acceleration curve and the characteristic value of a X, Y, Z axis rotation displacement curve, and further obtain a prediction acceleration curve and a prediction rotation displacement curve. According to the method, the collision waveforms of the vehicles in different scenes can be quickly obtained by using fewer sample sets, the prediction precision is ensured while the cost is reduced, the method can be used as the input of the passenger damage prediction, and finally the path planning and the triggering of the pre-collision system in the automatic driving danger scene are realized.

Description

Data-drive-based vehicle collision waveform prediction method, electronic device and readable storage medium
Technical Field
The invention relates to the field of vehicle collision, in particular to a vehicle collision waveform prediction method based on data driving.
Background
Automatic driving path planning, especially path planning in an emergency situation, needs to consider the damage condition of passengers if collision occurs in a real-time situation, in other words, passenger damage needs to be predicted, meanwhile, triggering of a pre-collision system is based on collision risk and personnel damage risk, and when the collision risk and the personnel damage risk are high, the pre-collision system with the highest level is triggered, so that safety of the passengers is protected in advance. At present, passenger damage prediction algorithms are mainly divided into two types, firstly, machine learning algorithms such as regression, support vector machine and random forest are adopted based on real traffic accident data, collision scene parameters (relative speed, overlapping rate, collision angle and collision target substance quantity) and characteristic parameters (sex, age, size, riding posture and the like) of passengers are used as input, and personnel damage levels are used as output; secondly, the waveform of the vehicle after collision (the time history curve of the vehicle deceleration in the collision process) is predicted as input, the dynamic response of the passenger in the whole collision process is predicted by adopting a deep learning algorithm, the damage grade of the passenger is obtained from the dynamic response curve, and the method has higher prediction precision and strong model generalization capability compared with the former method and avoids the problem of sample imbalance.
How to obtain the collision waveform of the vehicle from the scene parameters is a difficult problem that must be solved, and the traditional method is to use a Lumped Parameter Model (LPM) to make the stiffness of the vehicle body equivalent to a few or more simple spring-damping models, so as to predict the waveform. However, in different scenes, due to different force transmission directions, the vehicle body stiffness is also different, the same LPM model cannot be adopted, and meanwhile, the precision of the LPM model is low, so that the LPM model is only suitable for auxiliary development of a single standard working condition.
In summary, a method is needed to realize a prediction method from collision scene parameters (several scalars) to collision waveforms (vectors), obtain high-precision collision waveforms as input of personnel injury prediction, further evaluate injury risks of passengers under an automatic driving planned path, and simultaneously, serve as a trigger condition of a pre-collision system.
Disclosure of Invention
In view of the above technical problems, an object of the present invention is to provide a data-driven vehicle collision waveform prediction method, a computer device and a readable storage medium, so as to quickly obtain collision waveforms of vehicles in different scenes with fewer sample sets, reduce cost and ensure prediction accuracy.
The invention is realized by the following technical scheme:
the invention provides a vehicle collision waveform prediction method based on data driving, which mainly comprises the following steps:
s1 constructs a collision waveform database:
and according to the collision scene parameters and the boundary thereof, building a complete vehicle finite element model, and building a collision waveform database in which the scene parameters and the real collision waveforms are in one-to-one correspondence.
Further, the step of establishing the database comprises the steps of establishing a high-precision finished automobile collision finite element model, analyzing a traffic accident database, determining the boundary of the scene parameters, randomly sampling according to the scene parameters and the boundary thereof, generating a simulation matrix table, updating the finite element model according to the matrix table, carrying out simulation calculation, and obtaining the real collision waveform of the automobile. The real collision waveform refers to a real acceleration curve in the X, Y, Z direction under a vehicle coordinate system and a real rotary displacement curve rotating around the X, Y, Z axis. And when all the simulations in the matrix table are completed, a collision waveform database with scene parameters corresponding to real collision waveforms one to one is formed.
The finite element model comprises a high-precision model of a vehicle and a target object, wherein the vehicle needs to be subjected to waveform prediction, and the target object is a target object colliding with the vehicle.
The scene parameters include collision speed, collision angle, overlap rate and object type. The collision angle refers to the difference value of the course angle of the self-vehicle and the target vehicle at the time of collision 0; the collision speed refers to the speed difference between the self vehicle and the target vehicle; the overlapping rate refers to the percentage of the overlapping length of the self vehicle and the target vehicle in the collision 0 moment in the width of the self vehicle; the type of object refers to the classification of the object vehicle.
In the step, through deep analysis of traffic accident data, scene parameters are discretized according to the principle that the damage of passengers is obviously influenced.
The random sampling refers to random sampling in the same space formed by all scene parameters, and a matrix table with the sample size of N is generated.
S2 obtains a feature value of the waveform: respectively extracting characteristic values of the real acceleration curve and the real rotary displacement curve of each sample to obtain a characteristic value F1 of the real acceleration curve true And the characteristic value F2 of the real rotary displacement curve true 。
Further, the step S2 of obtaining the eigenvalue of the waveform includes:
(1) x, Y, Z direction real acceleration curve characteristic value extraction:
because the acceleration curve is small in number of peaks/troughs, any waveform can be obtained by superposing a plurality of sine waves according to the Fourier theory, the sine waves are unique, and the Fourier equivalent method is adopted for better fitting the trend of the acceleration curve. In the traditional method, the waveform is equivalent to square wave/trapezoidal wave/half sine wave, the wave/trapezoidal wave can lose all the wave crests/wave troughs and other key information of the original waveform, only one wave crest of the half sine wave can not reflect the variation trend of the waveform in the time course, and the Fourier equivalent method can keep the main characteristics of the original waveform.
In particular, the superposition approximation of the n-order sine wave is adopted to replace the original waveform,
Figure BDA0003716201410000021
Figure BDA0003716201410000022
Figure BDA0003716201410000031
wherein:
a (t) is the approximate acceleration curve, a (t) is the true acceleration curve in the direction X, Y, Z; k is the total number of discrete points comprised by A (t); a is i Is the eigenvalue to be solved; t is t k Is the effective duration of the crash (the time at which the acceleration curve returns to zero after time 0); t is t j Is the time corresponding to the j point of A (t); omega is the natural angular frequency of the own vehicle; m is j Is the slope of the line formed by the j-1 th and j-th points in A (t).
By adopting the method, the characteristic values of the real acceleration curve in the X, Y, Z direction can be respectively solved and are respectively marked as a x1 ,a x2 ,...,a xn 、a y1 ,a y2 ,...,a yn And a z1 ,a z2 ,...,a zn For the convenience of description, we construct these eigenvalues into a vector [ a ] x1 ,a x2 ,...,a xn ,a y1 ,a y2 ,...,a yn ,a z1 ,a z2 ,...,a zn ]Is marked as F1 true 。
(2) Characteristic value extraction of a rotational displacement curve around an axis X, Y, Z:
the disclination curve increases either monotonically during a crash or presents a single peak and many peaks in the absence of similar acceleration waveforms, because the effective crash time is typically very short (about 150ms), which is the time that a vehicle generally moves in a single direction about an axis. Therefore, the original curve can be well characterized by directly taking points on the original curve for the rotational displacement curve. By true rotation about the X-axisTaking a displacement curve as an example, taking n points to represent an original curve, and respectively taking the original curve t as a first point and a last point 0 And t k The rotation displacement value corresponding to the moment, the corresponding points on the original curve with equal time intervals are taken from the rest n-2 points, and the characteristic value is r x1 ,r x2 ,...,r xn Similarly, the characteristic values r in Y and Z directions can be obtained respectively y1 ,r y2 ,...,r yn And r z1 ,r z2 ,...,r zn Vector r composed of these eigenvalues x1 ,r x2 ,...,r xn ,r y1 ,r y2 ,...,r yn ,r z1 ,r z2 ,...,r zn ]Is marked as F2 true 。
S3, constructing a machine learning model:
the machine learning model includes a prediction model 1 and a prediction model 2. The inputs (features) of the training samples of prediction model 1 and prediction model 2 are both scene parameters, and the outputs (labels) are respectively F1 true And F2 true (ii) a Model 1 and model 2 are trained until the prediction accuracy requirement is met. Model 1 characteristic value F1 for predicting X, Y, Z directional acceleration pre . Model 2 used to predict eigenvalues of a rotational displacement curve around X, Y, Z F2 pre 。
In order to ensure that the model can better understand data, the characteristic values are subjected to data preprocessing before the characteristics are input into the model for training.
Further, the 2 prediction models all adopt multilayer perceptrons (MLPs), the input activation function of the hidden layer adopts a Relu function, and the structure (the number of layers and the number of neurons in each layer) of the MLPs belongs to a hyper-parameter and can be optimized.
In order to solve the overfitting problem caused by a small sample, a drop method (drop) is adopted to set a part of elements in the output of a hidden layer to be 0 according to the probability of p, and meanwhile, the variance of the output of the hidden layer after the drop action is ensured to be unchanged, wherein
Figure BDA0003716201410000041
x is a one-dimensional vector of input samples, x is the ith value of x, p is the drop probability, 0< p < 1. Is a hyper-parameter, and optimizes the p-value according to whether the model over-fits the validation set.
Back propagation employs a small batch gradient descent method (SGD) to update the weights W 1 ,W 2 ,W 3 And b 1 ,b 2 ,b 3 。
The database was randomly divided into 2 parts, one part was used for training and the rest was used for validation. After parameter adjustment, if the verification precision does not meet the requirement, returning to the step S1, randomly increasing the sample points again to increase the sample amount N, and retraining until the precision meets the requirement.
S4 waveform prediction:
after the model precision meets the requirement, in practical application, the scene parameters are respectively input into the model 1 and the model 2 to obtain a characteristic value F1 of the prediction characteristic of the X, Y, Z direction acceleration curve pre And predicted eigenvalues of the rotational displacement curve around X, Y, Z F2 pre And from F1 pre And F2 pre And respectively reconstructing a predicted acceleration curve in the X, Y, Z direction and a predicted rotary displacement curve around X, Y, Z, namely predicting the injury of the person.
Specifically, the characteristic value F1 is obtained from the model 1 pre ，F1 pre The first n eigenvalues are eigenvalues of the acceleration in the X direction according to:
Figure BDA0003716201410000042
obtaining an acceleration waveform, further, doubly integrating the acceleration waveform:
d translation (t)＝∫∫a(t)dt
and obtaining the displacement curve predicted in the X direction.
Similarly, an Y, Z directional displacement curve can be obtained.
Here, the acceleration curve is converted into the displacement curve so as to be unified with the rotational displacement curve into the displacement curve, and is preferably used as an input of the human injury prediction.
Obtaining a characteristic value F2 according to the model 2 pre The first n eigenvalues constituteOrdinate of the curve of the rotational displacement about the X-axis.
Similarly, a Y, Z-directional rotational displacement curve can be obtained.
The injury of the person can be predicted by obtaining a displacement curve in the X, Y, Z direction and a rotation displacement curve around X, Y, Z.
By this, the prediction of the collision waveform is ended.
Another aspect of the present invention provides a computer apparatus including a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the data-drive-based vehicle collision waveform prediction method described above when executing the computer program.
Still another aspect of the present invention is directed to a computer-readable storage medium storing a computer-readable storage medium that executes the data-drive-based vehicle collision waveform prediction method described above.
The invention has the following beneficial effects:
according to the method, the characteristic values of the waveform can be predicted from scene parameters through a small sample quantity and a multilayer perceptron model, and the waveform is reconstructed from the characteristic values of the waveform, so that a good effect is achieved. The invention avoids the problems that the value of each moment of the waveform is directly predicted by the scene parameters, the directly predicted value is too many, the sample size demand is huge, and no good algorithm is realized.
According to the method, the characteristics of the waveform are manually extracted, only n characteristic values are required to be predicted in a single direction, generally n is 2-5, a good effect can be obtained, and the final waveform prediction is realized. The waveform prediction result obtained by the method can be used as the input of subsequent passenger damage prediction, and finally provides input for path planning and pre-collision system triggering under the dangerous state scene of the automatic driving working condition, and the method is low in cost, high in precision and high in engineering feasibility.
Drawings
FIG. 1 is a flow chart of collision waveform prediction;
FIG. 2 is a flow chart of waveform characteristic value acquisition;
FIG. 3 is a schematic view of an impact angle;
FIG. 4 is a schematic view of an overlap ratio;
FIG. 5 calculation of m j Schematic diagram of
FIG. 6 is a schematic diagram of a multi-level perceptron;
FIG. 7 is a schematic diagram illustrating an effect of extracting acceleration characteristic values in the X direction;
fig. 8 is a schematic diagram illustrating the effect of extracting characteristic values of a disclination curve around X.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the following detailed description of the embodiments of the present invention is provided with reference to the accompanying drawings and examples. Other advantages and effects of the present invention will be readily apparent to those skilled in the art from the disclosure herein. The invention is capable of other and different embodiments and of being practiced or of being carried out in various ways, and its several details are capable of modification in various respects, all without departing from the spirit and scope of the present invention. It should be understood that the preferred embodiments are illustrative of the invention only and are not limiting upon the scope of the invention.
It should be noted that the drawings provided in the following embodiments are only for illustrating the basic idea of the present invention, and the components related to the present invention are only shown in the drawings rather than drawn according to the number, shape and size of the components in actual implementation, and the type, quantity and proportion of the components in actual implementation may be changed freely, and the layout of the components may be more complicated.
In one embodiment of the present invention, the main steps of collision waveform prediction are given, referring to fig. 1, a collision waveform prediction flowchart, and the prediction process is mainly completed by 4 steps, which respectively include: s1 constructs a collision waveform database: establishing a collision waveform database by adopting a finite element method according to the collision scene parameters and the boundary thereof; s2 obtains a feature value of the waveform: extracting characteristic values of the waveform of each simulation sample in the collision waveform database, wherein the characteristic values comprise a real acceleration curve in the X, Y, Z direction and a real rotary displacement curve around the X, Y, Z axis; s3 constructs a machine learning model: a prediction model 1 for predicting the characteristic value of the acceleration curve and a prediction model 2 for predicting the characteristic value of the rotary displacement curve; s4 waveform prediction: and respectively inputting the specific parameters of the scene into a prediction model 1 and a prediction model 2 to respectively obtain the characteristic value of an acceleration curve in the X, Y, Z direction and the characteristic value of a rotary displacement curve around the X, Y, Z axis, and further obtain a predicted acceleration curve in the X, Y, Z direction and a predicted rotary displacement curve around the X, Y, Z axis.
The following is a detailed description of the above steps:
and S1, constructing collision waveform database samples.
A finite element method is adopted to establish a simulation data set mainly according to collision scene parameters and boundaries thereof.
In this step, first, the scene parameters that affect the collision waveform and the boundaries of these parameters are determined. The determination of the scene parameters can be determined by analyzing the traffic accident database, and a specific embodiment is to adopt the traffic accident database, take the scene parameters as independent variables (input characteristics), take the passenger injury results as response variables (labels), train a random forest model, and determine the selected scene parameters according to the importance of the input characteristics. One specific example is that collision speed, overlap ratio, collision angle and object type are used as parameters of the collision scenario.
The determination of the boundary of the scene parameter, discretizing the scene parameter on the basis of the principle that the damage to the passengers is significant, for example, the collision speed can be discretized into discrete points with the interval of 2km/h, and the collision angle can be discretized into discrete points with the interval of 5 degrees, so that the cost of finite element calculation in S1 can be reduced.
One specific example is:
collision speed: the interval range is [10km/h, 120km/h ], and the sampling interval is 2 km/h.
Collision angle: the frontal impact angle range [0 °, 30 ° ], [330 °,0 ° ] (the head direction is 0 °, clockwise positive), the lateral impact is (30 °, 150 °), (210 °, 330 °), the tail impact is [150 °, 180 ° ], [180 °, 210 ° ], the sampling interval is 5 °, see fig. 3.
Front face overlap ratio: left and right sides 25%, left and right sides 50%, 100%, and the side overlap ratio is defined as the extension of the boundary line of the target vehicle width as the middle part if covering the B-pillar boundary line of the own vehicle, as the front part if not covering the B-pillar boundary line and being biased toward the vehicle head, or as the rear part, and as the definition of the rear overlap ratio and the front overlap ratio, as shown in fig. 4.
Type of the target: classified as cars, SUVs, minitrucks, lorries, buses, columns.
Here, the random sampling refers to random sampling in the same space formed by all scene parameters, and a matrix table with the sample size N is generated.
In the finite element method, a high-precision finite element model of the self-vehicle and the target object is established, each sample parameter of the matrix table is updated to the finite element model, simulation calculation is carried out, and a real acceleration curve in the X, Y, Z direction of the lower end of the left B column of the vehicle and a real rotation displacement curve of the vehicle body (taking the front point and the rear point of the vehicle body which is not deformed) around the X, Y, Z axis are extracted. And after the simulation and the extraction of all samples in the matrix table are completed, storing the results to form a collision waveform database.
The S2, obtaining a feature value of the waveform, see fig. 2:
the method mainly comprises two parts, wherein one part is used for extracting characteristic values of a real acceleration curve in the X, Y, Z direction, and the other part is used for extracting characteristic values of a real rotary displacement curve around the X, Y, Z axis.
X, Y, Z extraction of characteristic value of real acceleration curve:
taking the X direction as an example, obtaining a real acceleration curve A (t) from a finite element calculation result, and calculating adjacent points (t) j-1 ,A t＝j-1 ) And (t) j ,A t＝j ) Slope m of j ，
Figure BDA0003716201410000071
Calculating the effective duration of the collision t k ，
t k ＝t A(t)＝0 ，
I.e. the time at which the acceleration curve returns to zero after time 0, see fig. 5.
From m j And t k Calculating the ith characteristic value a i ，
Figure BDA0003716201410000072
Where k is the total number of discrete points contained in the original wave form curve, the natural angular frequency is calculated
Figure BDA0003716201410000073
Finally, the acceleration curve is reconstructed, and a i And ω into:
Figure BDA0003716201410000074
obtaining an approximate collision waveform a (t), and performing double integration on the a (t) to obtain an approximate displacement curve in the X direction:
d translation (t)＝∫∫a(t)dt。
y, Z the extraction of direction feature values is similar to the X direction.
X, Y, Z extraction of characteristic value of true rotary displacement curve:
taking the X direction as an example, d (t) is the X direction true rotational displacement curve, and its characteristic value is directly taken as the point on the original curve. Taking n points to represent original curves, wherein the n points are respectively represented as r 1 、r 2 、…、r n First point r 1 And a last point r n Respectively taking original curves t 0 And t k Ordinate, r, corresponding to time 2 、r 3 、…r n-1 Taking the corresponding points r on the original curve at equal time intervals 1 、r 2 、…、r n The substitution function f () of these n points can reconstruct the rotation displacement curve, and f usually takes a unitary linear model.
Y, Z the extraction of direction feature values is similar to the X direction.
And S3, constructing a machine learning model.
The machine learning model is composed of 2 models which are respectively marked as model 1 and model 2, and model 1 completes X, Y, Z direction acceleration curve characteristic value F1 pre Model 2 completes the characteristic value F2 of the rotary displacement curve around X, Y, Z pre And (4) predicting. The inputs (features) of the training samples of prediction model 1 and prediction model 2 are both scene parameters, and the outputs (labels) are respectively F1 true And F2 true 。
Before inputting the features into the model training, the feature values are subjected to data preprocessing. As a specific example, the collision angle is converted into vector coordinates on a unit circle in a rectangular coordinate system, for example, coordinates (0, 1), (0.87, 0.50), (1, 0) correspond to 0 °, 30 °, and 90 °, respectively, and the collision angle feature becomes 2 features; the overlapping rate and the type of the target object are represented by single-hot coding, and 16 and 6 characteristics are obtained respectively; the collision velocity is normalized and mapped to the interval of [0, 1], 1 feature, so far, a total of 25 feature configurations are input.
The 2 prediction models all adopt a multilayer perceptron (MLP), which is shown in fig. 6 and is formed as follows:
the prediction method is characterized by comprising 1 input layer (X), 2 hidden layers (H1, H2) and 1 output layer (O), wherein X is composed of g (dimension of input features) neurons, H1 is composed of u neurons, H2 is composed of v neurons, and O is composed of m (number of feature values to be predicted) neurons.
The forward propagation calculation process is as follows:
H 1 ＝σ(W 1 *X+b 1 )
H′ 1 ＝drop(H 1 )
H 2 ＝σ(W 2 *H′ 1 +b 2 )
H′ 2 ＝drop(H 2 )
O＝W 3 *H′ 2 +b 3
wherein X belongs to g multiplied by 1, O belongs to m multiplied by 1, W 1 ∈u×n,W 2 ∈v×u,W 3 ∈m×v,b 1 ∈n×1,b 2 ∈v×1,b 3 E m x 1, σ denotes the activation function, σ (x)Max (x,0), i.e., Relu function.
Adopting a drop method (drop) to set a part of elements in hidden layer output to be 0 according to the probability of p, and simultaneously ensuring the drop p The variance of the output of the post-action hidden layer is unchanged, wherein
Figure BDA0003716201410000081
Where x is a one-dimensional vector of input samples, x i Is the ith value of x, p is the discard probability, 0<p<1. Is a hyper-parameter, and optimizes the p-value according to whether the model over-fits the validation set.
Back propagation employs a small batch gradient descent method (SGD) to update the weights W 1 ,W 2 ,W 3 And b 1 ,b 2 ,b 3 The loss function adopts a root mean square error method.
The database was divided into 2 sections, 80% used for training and 20% for validation. After parameter adjustment, if the precision (R) is verified 2 And-adjusted) does not reach the requirement (more than or equal to 85%), returning to the step S1, randomly increasing the sample points again to increase the sample volume N, and retraining until the precision meets the requirement.
The R is 2 -adjusted is calculated as follows:
Figure BDA0003716201410000091
Figure BDA0003716201410000092
n is the number of samples, p is the number of input features,
Figure BDA0003716201410000093
is the average value of the values of y,
Figure BDA0003716201410000094
to predict value, y i Is the ith sample label value.
A specific oneExamples are (model 1), g 25, p 0.05, u 128, v 64, m 15, N500, R 2 -adjusted＝87.5％。
The S4, predicting a waveform.
After the model precision meets the requirements, the specific parameters of the scene can be respectively input to the prediction model 1 and the prediction model 2 in real time for actual prediction.
As a specific example, if n is 5, then prediction model 1 and prediction model 2 each obtain F1 pre And F2 pre 。
F1 pre The first 5 eigenvalues are eigenvalues of the acceleration in the X direction, according to:
Figure BDA0003716201410000095
obtaining an acceleration curve predicted in the X direction, and further integrating the acceleration curve predicted in the X direction:
d translation (t)＝∫∫a(t)dt
the displacement curve predicted in the X direction is obtained as shown in fig. 7.
Similarly, an Y, Z direction predicted displacement curve can be obtained.
According to F2 pre The first 5 eigenvalues constitute the ordinate of the rotational displacement curve around the X-axis. A specific example is a prediction of 150ms (t) k 150ms), the abscissa is 0ms, 37.5ms, 75ms, 112.5ms, 150ms, respectively, and the corresponding ordinate is r, r 1 、r 2 、r 3 、r 4 、r 5 The horizontal and vertical coordinates are in one-to-one correspondence to form 5 coordinate pairs (0, r) 1 )，(37.5，r 2 )，(75，r 3 )，(112.5，r 4 )，(150，r 5 ) The connecting lines of these 5 coordinate pairs constitute the X-axis predicted rotational displacement curve, as shown in fig. 8.
Similarly, an Y, Z axis predicted rotational displacement curve can be obtained.
After the above processing, the X, Y, Z predicted displacement curve and the X, Y, Z predicted rotational displacement curve are obtained and can be used as the input of the subsequent personnel injury prediction algorithm.
In another embodiment of the present invention, a computer apparatus is provided, which includes a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the data-driven vehicle collision waveform prediction method as described in the above embodiment when executing the computer program.
In still another embodiment of the present invention, a computer-readable storage medium is provided, which stores a computer program for executing the data-drive-based vehicle collision waveform prediction method described in the above embodiment.
The waveform prediction result obtained by the method has wide application. Specific examples are:
1. in the whole vehicle development, the change of the corresponding collision waveform when the vehicle parameters are changed can be quickly obtained through the method, so that the whole vehicle development can be quickly iterated.
2. The waveform is used as the input of the human body injury prediction model, the injury risk of the human body can be obtained before the collision, and when the injury risk is high, a pre-collision safety system is triggered, for example, the angle of the zero-gravity seat backrest is adjusted, and the injury risk of the lumbar of the human body is reduced.
3. Under the automatic driving danger state scene, the human body injury risks of different planned paths can be obtained, and the path with the minimum injury risk is selected.
It should be understood that the present application may use various forms of the flows shown above, with steps reordered, added, or deleted. For example, the steps described in the present application may be executed in parallel, may be executed sequentially, or may be executed in different orders, and are not limited herein as long as the desired results of the technical solutions disclosed in the present application can be achieved.
The above-described embodiments should not be construed as limiting the scope of the present application. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may be made in accordance with design requirements and other factors. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims (16)

1. A data-driven vehicle crash waveform prediction method, the method comprising:
s1 constructs a collision waveform database:
according to collision scene parameters and boundaries thereof, a complete vehicle finite element model is built, a collision waveform database with scene parameters and real collision waveforms in one-to-one correspondence is built, and the real collision waveforms refer to a real acceleration curve in the X, Y, Z direction and a real rotary displacement curve rotating around the X, Y, Z axis, which are obtained through finite element calculation under a complete vehicle coordinate system;
s2 obtains a feature value of the waveform:
respectively extracting characteristic values of the real acceleration curve and the real rotary displacement curve to obtain a characteristic value F1 of the real acceleration curve true And the characteristic value F2 of the real rotary displacement curve true ，F1 true And F2 true Is a one-dimensional vector;
s3, constructing a machine learning model:
the machine learning model comprises a prediction model 1 and a prediction model 2, wherein the input (characteristics) of training samples of the prediction model 1 and the prediction model 2 are scene parameters, and the output (labels) are respectively F1 true And F2 true (ii) a Training the model 1 and the model 2 until the prediction precision requirement is met;
s4 waveform prediction:
inputting scene parameters into a prediction model 1 and a prediction model 2 in real time to respectively obtain predicted characteristic values F1 pre And predicted eigenvalues F2 pre From F1 pre Reconstructing a predicted acceleration curve in the direction of X, Y, Z to obtain a predicted displacement curve; from F2 pre Reconstructing a predicted rotational displacement curve around an axis X, Y, Z; the predicted displacement curve and the predicted rotational displacement curve constitute a predicted collision waveform.
2. The data-drive-based vehicle collision waveform prediction method according to claim 1, wherein the S1 includes: establishing a complete vehicle collision finite element model, analyzing a traffic accident database, determining the boundary of scene parameters, randomly sampling according to the scene parameters and the boundary thereof, generating a simulation matrix table, updating the finite element model according to the matrix table, carrying out simulation calculation to obtain collision waveforms of the vehicle, and establishing a collision waveform database in which the scene parameters and the real collision waveforms are in one-to-one correspondence.
3. The data-drive-based vehicle collision waveform prediction method according to claim 2, characterized in that the finite element model includes a high-precision model of a host vehicle and a target object, the host vehicle being a vehicle for which waveform prediction is required, the target object being a target object with which the host vehicle collides.
4. The data-driven-based vehicle collision waveform prediction method according to claim 2, characterized in that the determination of the scene parameters is performed by using a traffic accident database, the scene parameters as independent variables, the occupant injury results as response variables, and the selected scene parameters are determined according to the importance of the input features through a training model.
5. The data-drive-based vehicle collision waveform prediction method according to claim 4, characterized in that the scene parameters include collision speed, collision angle, overlap ratio, and target type; the overlapping rate refers to the overlapping length of the self vehicle and the target vehicle at the time of collision 0, and accounts for the width of the self vehicle.
6. The data-drive-based vehicle collision waveform prediction method according to claim 2, characterized in that the random sampling refers to randomly sampling in the same space composed of all scene parameters, and generating a matrix table with a sample size of N.
7. The data-drive-based vehicle collision waveform prediction method according to any one of claims 1-6, characterized in that in S2, the characteristic value extraction of the true acceleration curve in the X, Y, Z direction employs a Fourier equivalent method.
8. The data-drive-based vehicle crash waveform prediction method of claim 7, characterized in that a real acceleration curve in X, Y, Z direction is approximately characterized by using a Fourier equivalent method, specifically: n-order Fourier equivalence is adopted in each direction, the characteristic values of the real acceleration curve in the X, Y, Z directions are respectively solved, and then n characteristic values exist in each direction, the total number of the characteristic values is 3n, and the characteristic values are marked as F1 true . The eigenvalues are calculated as follows:
Figure FDA0003716201400000021
Figure FDA0003716201400000022
wherein:
a (t) is the true acceleration curve in the direction of X, Y, Z; k is the total number of discrete points comprised by A (t); a is i Is the eigenvalue to be solved; t is t k The effective duration of the collision is the time corresponding to the acceleration curve returning to zero after the time 0; t is t j Is the time corresponding to the j point of A (t); ω is the natural angular frequency of the own vehicle; m is j Is the slope of the line formed by the j-1 th point and the j-th point in A (t).
9. The data-drive-based vehicle collision waveform prediction method according to any one of claims 1-6, characterized in that in S2, the characteristic values of the rotation displacement curve around X, Y, Z axis are extracted by directly taking points on the true rotation displacement curve, and n characteristic values in each direction, 3n characteristic values in total, are recorded as F2 true 。
10. The data-drive-based vehicle collision waveform prediction method according to any one of claims 1-6, characterized in that in S3, the prediction model employs a multilayer perceptron (MLP), and the activation function of the input of the hidden layer employs a Relu function.
11. The data-drive-based vehicle collision waveform prediction method according to claim 10, wherein in the step S3, in constructing the prediction model, a drop method (drop) is used to set a part of elements in the hidden layer output to 0 with the probability of p, and the variance of the output of the hidden layer after the drop action is ensured to be constant, wherein the drop method is used to set the part of elements in the hidden layer output to 0 with the probability of p, and the variance of the output of the hidden layer after the drop action is ensured to be constant
Figure FDA0003716201400000023
x is a one-dimensional vector of input samples, x i Is the ith value of x, p is the discard probability, 0<p<1 is a hyper-parameter that optimizes the p-value based on whether the model over-fits the validation set.
12. The data-drive-based vehicle collision waveform prediction method of claim 10, wherein the S4 updates the weight W using a small batch Stochastic Gradient Descent (SGD) method for back propagation in constructing the prediction model 1 ,W 2 ,W 3 And b 1 ,b 2 ,b 3 。
13. The data-drive-based vehicle collision waveform prediction method according to any one of claims 1 to 6, characterized in that the S4 specifically includes:
respectively inputting scene parameters into a prediction model 1 and a prediction model 2; obtaining a predicted characteristic value F1 according to the prediction model 1 pre Taking the X direction as an example, the predicted acceleration curve a (t) of the X direction:
Figure FDA0003716201400000031
obtaining an acceleration waveform, and further integrating the acceleration waveform:
d translation (t)＝∫∫a(t)dt
and obtaining a predicted displacement curve in the X direction, and obtaining predicted displacement curves in the Y direction and the Z direction in the same way.
Obtaining characteristic value F2 according to prediction model 2 pre Further, a rotational displacement curve around X, Y, Z is obtained.
14. The data-drive-based vehicle collision waveform prediction method according to any one of claim 13, wherein the value of n is selected according to the accuracy requirement, and the larger n is, the higher the accuracy is, the more samples are needed, and the cost is, usually, 2-5.
15. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the data-driven vehicle collision waveform prediction method of any one of claims 1 to 14 when executing the computer program.
16. A computer-readable storage medium storing a computer-readable program for executing the data-drive-based vehicle collision waveform prediction method according to any one of claims 1 to 14.
CN202210736800.4A 2022-06-27 2022-06-27 Vehicle collision waveform prediction method based on data driving, electronic equipment and readable storage medium Pending CN115099096A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210736800.4A CN115099096A (en) 2022-06-27 2022-06-27 Vehicle collision waveform prediction method based on data driving, electronic equipment and readable storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210736800.4A CN115099096A (en) 2022-06-27 2022-06-27 Vehicle collision waveform prediction method based on data driving, electronic equipment and readable storage medium

Publications (1)

Publication Number Publication Date
CN115099096A true CN115099096A (en) 2022-09-23

Family

ID=83295480

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210736800.4A Pending CN115099096A (en) 2022-06-27 2022-06-27 Vehicle collision waveform prediction method based on data driving, electronic equipment and readable storage medium

Country Status (1)

Country Link
CN (1) CN115099096A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP4397547A1 (en) * 2023-01-03 2024-07-10 Chongqing Changan Automobile Co., Ltd. Vehicle collision risk prediction device and method

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP4397547A1 (en) * 2023-01-03 2024-07-10 Chongqing Changan Automobile Co., Ltd. Vehicle collision risk prediction device and method

Similar Documents

Publication Publication Date Title
Zhang et al. Near-optimal rapid MPC using neural networks: A primal-dual policy learning framework
US11295517B2 (en) Generating realistic point clouds
Zhang et al. Vehicle crash accident reconstruction based on the analysis 3D deformation of the auto-body
EP4397547A1 (en) Vehicle collision risk prediction device and method
EP3747706B1 (en) Vehicle total safety performance application and method utilizing crash configuration information
Gu et al. Reliability-based robust assessment for multiobjective optimization design of improving occupant restraint system performance
CN103345554B (en) A kind of Fuzzy Reliability Evaluation method of vehicle occupant restraint system
WO2021071772A1 (en) Multi-agent simulations
WO2016106060A1 (en) Analyzing a collision with a vehicle having unknown damage
CN111259737B (en) Method and device for predicting failure of steering wheel of vehicle, electronic equipment and storage medium
CN115099096A (en) Vehicle collision waveform prediction method based on data driving, electronic equipment and readable storage medium
CN114418203B (en) Passenger damage prediction method and device based on collision key characteristic data
Horii Advancement of vehicle occupant restraint system design by integration of artificial intelligence technologies
CN110766213A (en) Automobile passenger collision damage prediction method and computer equipment
CN114418200B (en) Model training method and device for model prediction of vehicle type specific passenger injury
Müller et al. Real-time crash severity estimation with machine learning and 2d mass-spring-damper model
Müller et al. A statistical learning approach for estimating the reliability of crash severity predictions
CN114386691A (en) Occupant damage prediction method and device based on stress posture prediction
Seyedi et al. A comprehensive assessment of bus rollover crashes: Integration of multibody dynamic and finite element simulation methods
Zeng et al. Machine learning enabled fusion of CAE data and test data for vehicle crashworthiness performance evaluation by analysis
CN116484627A (en) Method for generating occupant injury prediction model, and occupant injury prediction method and device
WO2018167453A1 (en) Method and system of predicting the degree of bodily harm in an accident
CN115771506A (en) Method and device for determining vehicle driving strategy based on passenger risk cognition
Li et al. Investigation of the Prediction Model and Assessment Parameters of Head Injury of Children Occupants Based on Machine Learning
Jeon et al. A deterministic feedback model for safe driving based on nonlinear principal analysis scheme

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination