Available online at www.sciencedirect.

com

Advances in Engineering Software 39 (2008) 459465 www.elsevier.com/locate/advengsoft

Vehicle crash accident reconstruction based on the analysis 3D deformation of the auto-body
Xiao-yun Zhang *, Xian-long Jin, Wen-guo Qi, Yi-zhi Guo
Department of Mechanical Engineering, Shanghai Jiao Tong University, Room 909, No. 800 Dongchuan Road, Shanghai 200240, PR China Received 17 July 2006; received in revised form 17 April 2007; accepted 3 May 2007 Available online 19 June 2007

Abstract The objective of vehicle crash accident reconstruction is to investigate the pre-impact velocity. Elasticplastic deformation of the vehicle and the collision objects are the important information produced during vehicle crash accidents, and the information can be fully utilized based on the nite element method (FEM), which has been widely used as simulation tools for crashworthiness analyses and structural optimization design. However, the FEM is not becoming popular in accident reconstruction because it needs lots of crash simulation cycles and the FE models are getting bigger, which increases the simulation time and cost. The use of neural networks as global approximation tool in accident reconstruction is here investigated. Neural networks are used to map the relation between the initial crash parameter and deformation, which can reduce the simulation cycles apparently. The inputs and outputs of the articial neural networks (ANN) for the training process are obtained by explicit nite element analyses performed by LS-DYNA. The procedure is applied to a typical trac accident as a validation. The deformation of the key points on the frontal longitudinal beam and the mudguard could be measured according to the simulation results. These results could be used to train the neural networks adapted back-propagation learning rule. The pre-impact velocity could be got by the trained neural networks, which can provide a scientic foundation for accident judgments and can be used for vehicle accidents without tire marks. 2007 Elsevier Ltd. All rights reserved.
Keywords: Crash; Simulation; Finite element method; Neural networks; Accident reconstruction

1. Introduction Vehicle accidents bring dramatic tragedy to people and have become serious social problems that threaten people and their property with an accelerating rate, particularly in developing countries such as China. Vehicle collision analyses and accident reconstructions have become increasingly popular in recent years due to the seemingly expanding rate of civil litigation. The most concern of vehicle crash accident reconstruction is to retrieve the pre-impact velocity. The whole process of a vehicle accident can be divided into three phases: (1) post impact the usual starting point of the reconstruction developed depending on the informa*

tion obtained from the accident scene. (2) Impact determination of the vehicles impact speeds, impact directions and impact location, and so on. (3) Pre-impact determination of speeds and trajectories (usually the unknowns). The main guideline of accident analysis is nding out the vehicle motions in collision phase through vehicle rest positions in post-collision phase, and then the vehicle motion in pre-collision phase, based on the collection, recording, investigation and analysis of the accident scene. There are a variety of ways to evaluate a vehicles speed. The evaluation of speed will usually be conducted with one or more of the following methods: 1. 2. 3. 4. Momentum/energy analysis [1,2]. Damage/energy analysis [3,4]. Centrifugal force analysis. Launch, fall or vault analysis [5,6].

Corresponding author. E-mail address: general_zhang@sjtu.edu.cn (X.-y. Zhang).

0965-9978/$ - see front matter 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.advengsoft.2007.05.002

460

X.-y. Zhang et al. / Advances in Engineering Software 39 (2008) 459465

5. Geometry and timing analysis [7]. 6. Event data recorders [8]. The momentum/energy analysis method is limited in the accidents with xed objects or between vehicles with large weight dierences. On the other hand, the method need clear tire marks or the rest position of the vehicles in the accidents, but the tire marks form ABS are much fainter or non-existent now. The damage/energy analysis method reconstruct the accidents based on the 2D deformation of the auto-body. The other methods can only be applied in the extra trac accidents. Due to the variety, complexity and instantaneity of the trac accidents, the precision of the quantitative analyses of the accidents manually is fairly low. For trac accident, the objective is the pre-impact velocity and trajectory. Elasticplastic deformation of the vehicle and other collision objects are the important information in the vehicle crash accidents, however, the computer simulation models now available for the reconstruction of the vehicle accidents seldom consider the deformation. With the development of simulation technology, the deformation can be fully utilized based on the nite element method, which plays an important role for the reconstruction of vehicle crash accidents. The explicit nite element algorithm takes not only the strain rate of materials in high speed but also the elastic and plastic characteristics into consideration by using a very small integration time step. Thus, researches on reconstruction of vehicle accidents based on the nite element method can get a relatively high precision of calculation. In addition, for an eective presentation of the results, 3D animation can be created using the post-processing software based on the nite element method, and direct observations of the deformation behavior of key energy absorbing parts can be taken in dierent types of accidents as well, such as mudguard and the front longitudinal beam. This cannot be nished well by other accident reconstruction programs. At present, the nite element method has been widely used for vehicle crashworthiness simulation. Unfortunately, the FEM never got a good application for accidents reconstruction because it needs lots of crash simulation cycles that will cost a great deal of time [911]. Therefore, this paper deals with the development of crash accident reconstruction technique by using neural network and the nite element method. In the present method, the non-linear explicit nite element code LSDYNA is adopted to simulate the crash accidents. On the basis of the numerical results of the crash accidents, a three-layer feedforward neural network is applied to generate an approximated function of the initial crash parameter and the deformation index E. The approximated function of E is then maximized under velocity and angle constraints by genetic algorithm. The procedure is applied to a typical trac accident: the deformation of the key points on the frontal longitudinal beam and the mudguard could be measured according to the simulation results. These results could be used to train

the neural network by back-propagation learning rule. The pre-impact velocity could be achieved by the trained neural network, which can provide a scientic foundation for accident judgments. 2. Nonlinear nite element method and neural networks 2.1. The nonlinear nite element method in crash simulation For a point in a body, the time-dependent deformation can be found by seeking a solution to the momentum equation: rij;j qfi qi x 1

where rij is the Cauchy stress, q is the current density, f is the body force density, i is acceleration. x Eq. (1) satisfy the traction boundary conditions rij ni ti t 2

on boundary ob1, the displacement boundary conditions xi X a ; t Di t on boundary ob2, the contact discontinuity r r ni 0 ij ij along an interior boundary ob3 when x x . i i The energy equation _ _ _ E Vsij eij p qV 5 4 3

is integrated in time and is used for equation of state evaluations and a global energy balance. In Eq. (5), V is the relative volume, and sij and p represent the deviatoric stresses and pressure, sij rij p qdij 1 p rij dij q 3 6 7

respectively, q is the bulk viscosity, dij is the Kronecker _ delta, if i = j, dij = 1, otherwise dij = 0, and eij is the strain rate tensor [12,13]. It is known that, Z Z qi rij;j qf dxi dv x rij nj ti dxi ds v ob1 Z r r nj dxi ds 0 8 ij ij
ob3

where dxi satises all boundary conditions on ob2, and the integrations are over the current geometry. By using the divergence theory it can lead to a statement of the principle of virtual work, i.e.: Z Z Z dp qi dxi dv rij dxi;j dv qfi dxi dv x v v v Z ti dxi ds 0 9
ob1

X.-y. Zhang et al. / Advances in Engineering Software 39 (2008) 459465

461

The structures and components can be modeled by many elements. The geometry of the element is represented by Lagrangian description that is in terms of shape function. After assembly, the elemental mass is lumped at the nodes. Whereas, the nite element models for crash simulation contain very large number of elements and time step of explicit central dierence method is small, the computing time is considerable long [14,15]. 2.2. Back-propagation neural networks Neural networks as global approximation tool has been widely used because of their ability to process and map external data and information based on past experience. In a neural network the transmission and the processing of the input data are assigned to a net of simple computing units called neurons. Each neuron returns an output signal when the weighted sum of the inputs exceeds an activation value. The output value is computed by dening a transfer function, also called activation function. In this research a feedforward multilayer neural network is used. The network consists of a rst neuron layer, one or more hidden layers and an output layer as sketched in Fig. 1. The rst layer, called input layer, receives the external inputs and transfers signals to the hidden layers. The signals are modied, weighted and redistributed throughout the hidden layers till the output layer. The output signal of the ith neuron of the kth layer can be computed as outk;i fk gk;i 10

gk;i bk;i

N k1 X l1

xl outk1;i k;i

13

where Nk1 is the number of neurons in the previous layer and therefore it is equal to the number of links of the ki neuron. xl , with l = 1, . . ., Nk1 and bk,i are, respectively k;i the link weights and the bias value of the considered neuron. The learning process of the neural network involves a nite number of examples, the training set, characterized by inputoutput pairs known a priori. The link weights and the bias value of the neural network are changed so to minimize the dierences between the known and computed output. The neural network uses the back-propagation learning rule where the RMS between the known and computed outputs is minimized. The training process is performed using the LevenbergMarquardt algorithm. The algorithm was originally designed for least-squares minimization where the objective function is dened as a sum of square. As thus the objective function results always are non-negative. A batch mode is used to perform the training process. For each output of the neural network a single average square error is dened as the sum of the RMS errors attained by each case of the training set. Indeed, considering a training set of N inputoutput pairs, the error to be minimized is dened as v u N u1 X 2 RMS t 14 tn on N n1 where tn is the known output and on the corresponding value computed by the network. The use of batch training allows the learning process to be faster and reduces the probabilities to nd a local minimum. In order to improve the speed and the eciency of the learning phase, the input and output data are scaled. Herein, the input data are linearly scaled between 0 and 1 using the equation: x X X min X max X min 15

where fk denotes the transfer function of the considered layer. A sigmoid function is used for the neurons of the hidden layer, the function can be written as fk gk;i 1 1 egk;i 11

and the neurons of the output layer use a linear transfer function as fk gk;i gk;i 12

where gk,i is the activation signal of the ki neuron. The activation value is computed as
Hidden layer Output layer

Input layer
1

X1

X 1M

where x is the scaled value of the design variable vector X, where Xmax and Xmin are the maximum and minimum values of the design domain, respectively. The input variables of the neural network are chosen to describe the velocity and angle of the considered structure while the output values are mainly related to the deformations. 3. Accident reconstruction using FEM and neural networks

1 Xn

M Xn

Fig. 1. The architecture of neural network.

According to characteristics of crash accidents, the indices to evaluate the accident were given as x1, x2, . . ., xn, which reected the deformation of the key points in the main energy-absorbing parts during the crash.

462

X.-y. Zhang et al. / Advances in Engineering Software 39 (2008) 459465

The key points were chosen at the round holes and the bolt holes, which were easy to locate. The deformation of the center of the round hole can be measured using the deformations of the 3 points on the circle of the hole and the deformation of the bolt hole can be measured using the information of three coordinates of the nodes in that place. Using the measurement values of these indices from the survey of the accident scene: x1e, x2e, . . ., xne and the calculation values of these indices from computer simulations: x1c, x2c, . . ., xnc, the index E is dened to reect the diculty level of accident reconstruction E 1
n X 2 xie xic =xie i1

ing the trained neural network predictor, as shown in Fig. 2. 4. Application to a typical trac accident 4.1. Trac accident scene The procedure described above has been applied to an accident happened to one car in Hai Nan province, China in December 2001 as shown in Fig. 3. Three-coordinate-measuring instrument is used to measure the deformation of the car body. First the body ordinate system is set at the back of the car, where there is no deformation, and then 11 key points are selected at the front of the car where major deformation occurred: 7 bolts hold on the frontal longitudinal beam, 4 round holes on the mudguard. Because these two main energy-absorbing parts are hardly aected by the rough surface of the wall, they should be more reliable than the front wall. Fig. 4 is the photo of measuring the deformation of the key points. Fig. 5 is the geometry of the frontal longitudinal beam and the mudguard. 4.2. Finite element modeling The vehicle model is rst disassembled and grouped into several main groups: the frame, front inner, cabin, doors, etc. The three dimensional geometric data of each component are obtained by using UG and Catia software. Next, these les are imported into Hypermesh software for meshing and model assembly. Finally, the model is translated into LS-DYNA input les. It is known that the total structure is made up of more than 130 parts, which are connected by joints (including spot welding, arc welding, bolt connecting, rivet connecting and adhesives gluing). Many types of elements are used to describe the geometric shapes, connecting properties etc., including shell element, beam element, rod element, hexahedron element, spring element, rigid elements and specic welding elements which can fail when some specic conditions are met. A 3D dummy model used to simulate the dynamic behavior of the occupant in the full frontal crash is created. All parts of the model are connected using spherical and revolute joints. The numerical model is also validated by means of experimental results according to CMVDR294 (China Motor Vehicle Design Rule). Fig. 6 shows the comparison between the experimental and the numerical accelerationtime curve [16,17]. In this simulation, the deformation of the engine is not considered, but is treated as a rigid part. The doors are connected to the body using hinges and door locks. According to the photo of the trac accident scene, the nite element model of the wall is also included in the whole model. Fig. 7 shows the nite element model of the commercial vehicle and the wall. Fig. 8 shows the nite element model of the mudguard and the front longitudinal beam.

16

Obviously, E is closer to 1, and the crash reconstruction by computer simulation is closer to the real situation. So the equation to calculate the index E is set as the object function, the initial velocity and angle of the car are set as the variables, which would be optimized. That is to choose a right combination of the initial velocity and angle of the car to make the index E the maximum. How to maximize the index E is a process of repetitious computer simulations. The combination of computer simulation and neural network modeling is an enabling technology for crash reconstruction in shorter time. The nite element analyses are used to generate the examples for the training and test sets of the neural network and are performed using the commercial code LS-DYNA. The trained neural network can map the relation between the variable, that is, the initial velocity and angle of the car, and the objective function, that is, the index E. The settlement of the training patterns is performed by the technique of design-of-experiment in order to obtain a homogeneous allocation inside the normalized domain. The maximum of the index E is achieved using a genetic algorithm utiliz-

Fig. 2. The owchart of accident reconstruction.

X.-y. Zhang et al. / Advances in Engineering Software 39 (2008) 459465

463

Fig. 3. Photo of accident scene.

Fig. 4. Measuring the deformation of the key points.

Fig. 6. Acceleration curve of sub oor by FEM simulation and experimental result.

Fig. 5. Key points of the mudguard and the front longitudinal beam.

4.3. Accident reconstruction The measurements of the deformation of the longitudinal beam and mudguard are listed in Table 1. The velocity and direction of the commercial vehicle are taken as design variables, and the upper and lower bounds of the design variables are restricted as
Fig. 7. Finite element model of the commercial vehicle and the wall.

38 km=h 6 V 6 78 km=h 5 6 h 6 25

464

X.-y. Zhang et al. / Advances in Engineering Software 39 (2008) 459465

Fig. 8. Finite element model of the mudguard and the front longitudinal beam.

Table 1 Deformation of the key points (mm) No. 1 2 3 4 5 6 7 8 9 10 11 Coordinate before accident 116, 405.09, 88 132, 406.52, 73 124.54, 413.64, 245 85, 442.34, 307.97 263.36, 351.1, 106.5 606, 497.86, 174 630, 497.31, 200 681, 575, 128.8 600, 555, 138.45 462.54, 569.65, 138.44 230, 366, 0.012 Coordinate after accident 129.33, 404.85, 101.41 118.27, 412.76, 84.14 135.50, 411.12, 257.64 95.01, 438.28, 320.26 229.08, 346.89, 74.05 402.23, 499.92, 213.74 425.13, 512.77, 241.63 466.95, 594.62, 172.40 386.44, 571.39, 180.96 252.81, 579.84, 157.91 213.53, 383.75, 5.39 Deformation 18.91 18.75 16.92 16.36 47.39 207.62 209.63 219.32 218.37 210.88 24.81 Fig. 9. The training curve of neural network.

Table 3 Comparisons between E obtained by FE simulation and by the neural network Test points 1 2 3 4 Velocity (km/h) 46 58 40 62 Direction () 11 13 21 19 E (numerical results) 0.53 0.84 0.61 0.70 E (neural network results) 0.51 0.83 0.60 0.68

As shown in Table 2, six design levels of each design variable have the same intervals, respectively. The training set consists of a total number of 36 points in the design domain, and the test set consists of 4 points randomly chosen inside the design domain. A total number of 40 LS-DYNA runs are then performed. The average CPU time required for a single nite element analysis is about 8 h using 4 CPUs of SGI onyx3800 shared-memory parallel computer. The data of deformation of the 11 key points and the index E are calculated for each simulation as the output value of the neural network described above. The training parameter mainly includes the learning rate and momentum constant. Their values are 0.25 and 0.50, respectively. The neural network converges to an acceptable RMS error value of 1E03 after about 5000 cycles. It is shown in Fig. 9, where e is the error value and n is the data of the calculation cycles.
Table 2 Design variables and design levels in the design domain Design variables Velocity V (km/h) Direction h () Design levels 38 5 46 9 54 13 62 17 70 21 78 25

Fig. 10. Simulation result of 16 and 51 km/h condition.

From Table 3, it can be found that the neural network can map the relation between the initial crash parameter and the deformation index E in accurate manner. The maximum of the index E is achieved using a genetic algorithm utilizing the trained neural network. The velocity and angle corresponding to the maximum of the index E are 51 km/h and 16, respectively. The simulation result is shown in Fig. 10. It can be said that the simulation with the combination of the velocity and angle is quite coincident with the real accident. 4.4. Result verication Because of the clear tire marks on this accident scene (as shown in Fig. 3a), we can get the model of these marks by making use of photogrammetry technique by solving the collinear equations to obtain the object points set (see Fig. 11). So the angle between the direction of vehicle motion and the plane of wall can be measured from this

X.-y. Zhang et al. / Advances in Engineering Software 39 (2008) 459465

465

accident. The simulation with the pre-impact velocity and angle achieved by the trained neural network are coincident with the real accident, which can provide a scientic foundation for accident judgments. Acknowledgements The authors gratefully acknowledge the support from the National Natural Science Foundation of China (No. 60174023). The authors appreciate the assistance and suggestions of many individuals associated with the Shanghai Supercomputer Center and with SJTU High Performance Computing Center. References
[1] Stephan H, Moser A. The collision and trajectory models of PCCRASH. SAE 960886; 1996. [2] Zhang XY, Jin XL, Qi WG, et al. Analysis and reconstruction of the typical trac accident based on the tire marks. J Basic Sci Eng 2006;14(3):41826. [3] Mchenry BG, Mchenry RR. SMAC97 renement of the collision algorithm. SAE 970947; 1997. [4] Day TD, Hargens RL. An overview of the EDSMAC4 collision simulation model. SAE 1999-01-0102; 1999. [5] Leden L, Wikstrom P. Safety and accessibility eects of code modications and trac calming of an arterial road. Accident Anal Prevent 2006;38(3):45561. [6] Edosomwan JA, Khalil TM. Accident prevention in slips and falls. Prof Saf 1981;26(6):305. [7] Serrano A, Conde C. Computer vision application: real time smart trac light. Lecture Notes Comput Sci 2005;3643:52530. [8] Webb W. Black boxes capture car-crash data, controversy. EDN 2005;50(10):338. [9] Yamazaki K, Han J. Maximization of the crushing energy absorption of cylindrical shells. Adv Eng Software 2000;31(6):42532. [10] Allen RY, Terry DD. The DYMESH method for three-dimensional multi-vehicle collision simulation. SAE Trans 1999;108:36582. [11] Lozana JA, Vera C, Felez J. A computational dynamical model for trac accident reconstruction. Int J Vehicle Design 1998;19(2): 21327. [12] Livemore Software Technology Corporation, Compiled by John O. Hallquist. LS-DYNA theoretical manual; 1998. [13] Livemore Software Technology Corporation. LS-DYNA keyword users manual (version 970 beta); 2002. [14] Zhong ZH. Finite element procedures for contact-impact problems. Oxford University Press; 1993. [15] Zhong ZH, Nilsson L. Automatic contact searching algorithm for dynamic nite element analysis. Comput Struct 1994;50:18797. [16] Zhang XY, Jin XL, Qi WG, et al. Crashworthiness simulation and structure optimization design of auto-body based on nite element method. J Systems Eng Electron 2004;15(4):1126. [17] Zhang XY, Jin XL, Shen J. Virtual reconstruction of two types of trac accident by the tire marks. Lecture Notes Comput Sci 2006;4282:112835.

Fig. 11. 3D reconstruction of accident scene.

model, and its result is 16.89. Then this model would be imported into the PC-CRASH, and the velocity before the vehicle impacted the wall could be computed, the data is 49.3 km/h. PC-Crash contains several dierent calculation models, including an impulse-momentum crash model, a stiness based impact model, a kinetics model for realistic trajectory simulations, and a simple kinematics model for time-distance studies. For maximum versatility, PC-Crash simulation results can be viewed and outputted in scale plan and elevation views, 3D perspective view, and in numerous diagrams and tables. The simulation result by the software of PC-Crash is very close to the result obtained by the FE method. But the software does not utilize the information of vehicle deformation and it relies on the tire marks thoroughly. It is well known that nowadays many vehicles are installed with ABS so that the tire marks can hardly be got after braking. Furthermore, the weather condition or people can destroy tire marks easily. FE method resolves the velocity and angle according to vehicle deformations, and does not need any exterior conditions such as the tire mark. So, more reliable results could be got with the nite element method. 5. Conclusions This paper proposes a new technique to reconstruct crash accident using the combination of the nite element method and neural networks, which can handle the kind of accident without tire marks. The elasticplastic deformation and strain rate take into account adequately in this study using the FE simulation, so the results are more accurate. Using the neural network, the relation between the initial crash parameter and the key points deformation is mapped correctly, so it reduces the cycles of FE simulation. As verication, the procedure is applied to a typical trac