Scribd
Upload
95 views

Data Analysis For Parallel Car-Crash Simulation Results

This document discusses using data mining techniques to analyze the results of parallel car crash simulations. The simulations can produce varying results due to uncertainties from parallel computing. Clustering algorithms are used to group nodes with similar behavior and identify the origins of inconsistencies. Analysis of a benchmark model found scatter was caused by buckling of the motor carrier part during impact. Modifications to this part reduced scatter in subsequent simulations. Data mining is shown to help understand sources of uncertainty in parallel crash simulations.

Uploaded by

Goutham Mareeswaran B
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
95 views

Data Analysis For Parallel Car-Crash Simulation Results

This document discusses using data mining techniques to analyze the results of parallel car crash simulations. The simulations can produce varying results due to uncertainties from parallel computing. Clustering algorithms are used to group nodes with similar behavior and identify the origins of inconsistencies. Analysis of a benchmark model found scatter was caused by buckling of the motor carrier part during impact. Modifications to this part reduced scatter in subsequent simulations. Data mining is shown to help understand sources of uncertainty in parallel crash simulations.

Uploaded by

Goutham Mareeswaran B
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 22

Data analysis for parallel car-crash simulation results and model optimization

Submitted by - GOUTHAM MAREESWARAN B 1541210010 M TECH CIM 12-14

ABSTRACT
The paper discusses automotive crash simulation in a stochastic context, whereby the uncertainties in numerical simulation results generated by parallel computing. Since crash is a non-repeatable phenomenon, qualification for crashworthiness based on a single test is not meaningful, and should be replaced by stochastic simulation. But the stochastic simulations may generate different results on parallel machines, if the same application is executed more than once. For a benchmark car model, differences between the position of a node in two simulation runs of PAMCRASH or LS-DYNA of up to 10 cm were observed, just as a result of round-off differences in the case of parallel computing. In this paper, some data mining algorithms are described to measure the scatter of parallel simulation results of car-crash and then provide hints to overcome this scatter to get more stable car model.

The field of data mining and knowledge discovery is emerging as a new, fundamental research area with important applications to science, engineering, medicine, business, and education. Data mining attempts to formulate, analyze and implement basic induction processes that facilitate the extraction of meaningful information and knowledge from unstructured data. Simulation is now accepted as a third mode of science, supplementing theory and experiment. Today, not only do experiments produce huge data sets, but so do simulations. Data mining, and more generally data intensive computing, is proving to be a critical link between theory, simulation, and experiment.

NEED OF SIMULATION IN CAR MANUFCTURING


Nowadays the car manufacturing industry relies heavily on simulation results. By simulation the number of real prototypes is reduced, the insight into the features of the actual design is increased and the turn-around time between model changes is much shorter than in the case of real tests. Numerical crash simulation is the most computer-time consuming simulation task in car design.

Uncertainty of crash simulation results

Fig. 1 shows the maximal and average differences between several simulation runs of PAM-CRASH for a car model consisting of about 60.000 shell elements on a distributed memory parallel machine. In this paper, some practical data mining methods are introduced to measure the amount of the scatter of some parallel crash simulation results, the intention of this work is to find the origin of this scatter then provide hint to overcome this scatter.

Fig 2:Fresults for 15 LS-DYNA runs of NEON model.

Fig. 2 shows a typical F max result for 15 parallel simulation runs of LS-DYNA for the test case of NEON model at time state of 81 ms. The color indicates the scatter of the results at each point. Red areas are those with the largest values.1 The large scatter of the wheel in Fig. 2 is not important for design engineers. More important is the scatter of parts of the fire wall close to the driverss feet shown in Fig. 3. In order to improve the model, it is important to distinguish between local effects and impacts of scatter sources at other places, like the part of the motor carrier also shown in Fig. 3. In order to distinguish source and impact of indeterministic behavior, clustering analysis are introduced into the stability analysis of car-crash simulations.

Clustering algorithms for simulation data


Clustering in data mining is discovery process that group the data into classes or clusters such that objects within a cluster have high similarity in comparison to one another, but are very dissimilar to the objects in the other clusters. In other words, the data points in one cluster are more similar to one another; the data points in separate clusters are less similar to one another. In this paper, we try to describe how to find the origin of the in deterministic behaviour of parallel crash simulations using clustering analysis. One challenge for clustering process of crash simulation data is the problem size. A typical crash simulation code consists of 500,000 nodes and for a dependence analysis up to 150 time steps may be used of objects to be compared therefore contain 70 million objects.

Fig 3 :Scatter of simulation results on the motor carrier and in the interior.

Clustering algorithms for simulation data Hierarchical clustering


We are facing sea amounts parallel execution result data for actual car simulation model, direct clustering methods would take large CPU time. Fortunately, the geometric position of the nodes provides some structure, which can be exploited. The dependent relationship of the scatter of parallel execution results will be first related to the neighbour nodes, such that clustering process can be carried out using hierarchical methods. A hierarchical clustering method works by grouping data objects into a tree of clusters. Cluster(N) is executed with Nlist representing a list of all nodes of the car model. As a result all nodes are assigned to clusters and the whole model are grouped into some clusters. As well-known, clusters is not a deterministic process. One node might be related to two others, which are not related with each other. Therefore this node might be assigned to any of the two reference clusters. Inorder to reduce these effects, the clustering process is performed twice.

Therefore this node might be assigned to any of the two reference clusters. In order to reduce these effects, the clustering process is performed twice.
Perform clustering Cluster(N) for this reference time step. Find the center node for each cluster. Perform the second Cluster(N) to check for all related nodes to these center nodes. Sort center nodes by the size of clusters to get a new node sequence.

Clustering algorithm may be very in deterministic and might result in too many clusters. Therefore the clustering process is started again.

Clustering algorithms for simulation data Partitioning clustering for all time steps
The resulting set of clusters from the reference time step is ordered by size and all clusters smaller than Scluster are eliminated. By adjusting the size of threshold Csim and Scluster, a moderate number of clusters can be obtained for the whole car model, for example, the number of clusters Nclusters is dozens.

The algorithm used for crash simulation analysis is performed in the following steps:
(1) Select a reference time step, a minimal cluster size and a threshold. (2) Perform a preclustering for this time step. (3) For each time step. Perform parallel partitioning clustering based on the new node sequences created in Section 3.1. (4) Output.

Uncertainty analysis and optimization


Data mining is the process of knowledge discovery, both statistics and clustering are applied in sea amount data as mathematical tools to discover natural rules. The purpose of this paper is describe the indeterministic behavior of parallel crash simulations and search out the origin of this behavior. One cluster dominates the fire wall in the area of the driverss legs for those nodes with the largest scatter(Fig. 4 (right)).The evolution of clusters as time step was shown in Fig. 5. Fig. 6 contains a picture of the motor carrier at 35 ms, part of which is covered by the green cluster also. This cluster starts when the shock absorber pipe hits the rest of the motor carrier near 28 ms. At this point in time the shock absorber pipe puts pressure on the motor carrier parallel to the direction of this part. This causes a buckling effect.

Two simulation results for the motor carrier are shown in Fig. 7. The form of the motor carrier at the end of the simulation is quite different. The sim cluster function indicates, that this causes the substantial scatter at the fire wall. By analyzing the geometric structure of the part of motor carrier, where the shock absorber pipe puts pressure to at the critical time, we know the indeterministic behavior of parallel crash simulations was from contacting and buckling in some critical cases, see Fig. 8.

Fig. 4. Simcluster as color for CAMAS46 testcase of whole car and the interior at 80 ms.

Fig. 5. Development of the clusters with time (70, 60, 50, 40 ms).

Fig. 6. Simcluster for the motor carrier at time 35 ms (left) and at 28 ms (right) for its inner part.

Our work aims at stable car model. As a result of this analysis, some attempt was done to modify the parts in the areas of the origin of the instability. Here small change was done on the element of the part of motor carrier by changing the thickness, density or node position respectively. As a consequence the scatter of the result due to parallel computing on the fire wall was reduced substantially, which was shown in Fig. 9.

Fig. 7. View on the motor carrier at 80 ms for two extremely different simulation runs.

Fig. 8. Contact and buckling in critical cases.

Fig. 9. Scatter of several variants of the original model showing a substantial improved behavior.

CONCLUSIONS AND PERSPECTIVES


The paper has shown that data mining algorithms can be useful in describing the in deterministic behaviour of parallel crash simulations and identifying the origin of the scatter of simulation results. This indeterminacy was either due to the parallel computer architectures or buckling and certain contact in some critical cases. For example, the in deterministic behaviour in the part of drivers leg area are related to the part of motor carrier shown in Figs. 46. As a result of the analysis, hints are provided so that engineer can try to modify the motor carrier in the areas of the origin of the instability of wire wall. As a consequence the scatter of the result due to parallel computing on the fire wall was reduced substantially. This application is only a first step using data mining technology in the context of industrial crash simulation. Car manufacturing companies store the complete outputs of their crash simulation runs. This provides an excellent basis for data mining applications, like design of experiments, optimization, parameter fitting for coarser models used in concept studies.

You might also like