top of page


Mr. John Smith

Job title



By big data and technology developed in computer science, AI methodology is used to many branches and proved its efficiency, which can forecast precisely. Lately this methodology is tried in vehicle crash. But without considering the characteristic of crash data, forecasting with the common like LSTM was not enough because it is needed, the cooperation between vehicle crash engineer who understands well the crash data characteristic and data science programmer, AI professional. The purpose of this research is the development of AI model forecasting vehicle crash date. The manufacturer can save test vehicles numbers and fee by more precise forecast with crash engineer’s level up to AI scientist. This research is proceeded by Hyundai-Kia Namyang R&D and Postec data science LAB for 2 years. First year, 598 all front kinds of Hybrid III dummy crash test data were used and Second year 1067 56kph full frontal crash data and 30 kinds of static features to improve chest deflection prediction. First hypothesis was dummy is one closed system there would be some interrelation among external forces. Actually vehicle crash engineers analyze with the experience of interrelation between dummy’s each link part - head/neck/chest/pelvis/femur measured data. At the first year research, only front seat dummies’ data of 18 injuries, 2 belt load and 2 body pulse channels are trained as a simple feature selection. Postec advised AI model. Simply speaking, AI deep neural network is: 1) probability prediction pulling out patterns after training data 2) How to analyze pattern is, arbitrary matching a value on the mid of expectation by net propagation, comparing the value with real value of training set and moving its position by learning rate α to minimize the differential of value difference function. (Each step differential is stored on the cache. Then weight and bias are updated during back propagation) 3) From repeating as layer, a network like creatures’ recognizing is made and this is deep learning more complicated than machine learning. Representative common neural networks of time series data are CNN and RNN. CNN is made from mimicking cat’s eye cell structure. Crash test data is sequential but it is long to use RNN so more efficiency is needed. The Wavenet (2016, voice changer) and Bytenet (2017, translator) of Google Deepmind are good samples because RNN’s slow calculation is compensated with mixing CNN. From Bytenet, crash prediction was good but MSE was over 1 sigma. So as a nice inspiration Crash TSRnet has been created by deleting the Bytenet’s auto regressive inserting in decoder and using direct inserting from encoder. Then MSE marvelously became under 0.5 and maximum was 1.3 all through the various test modes, equally. This was proved as no overfitting from the research of NHTSA official test data including side tests, later. But to reduce MSE as much as possible training was accomplished in one test mode each and in case of wrong data using showed worse and the chest deflection prediction was not so good without inclination because it would be sensitive to local features. At the second year we collected 30 kinds of static features by engineering judgement from relevance, countable and possibility of collecting. The model was changed to separate sequence processing of static features, “Gated TSRnet” and MSE became better but not enough in worst case. Test data which has all static features were only in 129 tests among 1,067, one case only was in 49 tests. To break through this, static data were classified, grouped by ranking method and inserted by heuristic method but the chest prediction was not improved. Next, Crash TSRnet is quite good on the most of prediction, the new model “Residual Crash TSRnet” is created to lessen the frequency of static data by using tanh function. MSE was quite low even worst case (2mm). This would be useful to predict test result by only belt load and body pulse without testing, selecting next test restraint system and restoring missing channel. The limitation is current realistic data are not CAE but crash data because CAE is not accurate enough. So expansion to all kinds of field crashes will not be easy. More research could solve this as crash deepmind, something like “safe box”, control unit.

Mr. Unchin Park, Senior Engineer, Hyundai (Kia) Namyang R&D

Time Series Regression Network for Crash Test Data and its deep research for chest deflection advanced forecasting

FWC2023-CYB-002 • Integrated safety, connected & automated driving


Sign up or login to the ICC to download this item and access the entire FISITA library.

Upgrade your ICC subscription to access all Library items.

Congratulations! Your ICC subscription gives you complete access to the FISITA Library.


Retrieving info...

Available for purchase on the FISITA Store


bottom of page