Refine
Document Type
- Conference Paper (5)
- Article (2)
- Preprint (1)
Language
- English (8)
Institute
Review
- peer-review (6)
- nein (1)
Version
- published (2)
In road traffic, critical situations pass by as quickly as they appear. Within the blink of an eye, one has to come to a decision, which can make the difference between a low severity, high severity or fatal crash. Because time is important, a machine learning driven Crash Severity Predictor (CSP) is presented which provides the estimated crash severity distribution of an imminent crash in less than 0.2ms. This is 63⋅ 103 times faster compared to predicting the same distribution through computationally expensive numerical simulations. With the proposed method, even very complex crash data, like the results of Finite Element Method (FEM) simulations, can be made available ahead of a collision. Knowledge, which can be used to prepare occupants and vehicle to an imminent crash, activate and adjust safety measures like airbags or belt tensioners before of a collision or let self-driving vehicles go for the maneuver with the lowest crash severity. Using a real-world crash test it is shown that significant safety potential is left unused if instead of the CSP-proposed driving maneuver, no or the wrong actions are taken.
The development of finite element vehicle models for crash simulations is a highly complex task. The main aim of these models is to simulate a variety of crash scenarios and assess all the safety systems for their respective performances. These vehicle models possess a substantial amount of data pertaining to the vehicle's geometry, structure, materials, etc., and are used to estimate a large set of system and component level characteristics using crash simulations. It is understood that even the most well-developed simulation models are prone to deviations in estimation when compared to real-world physical test results. This is generally due to our inability to model the chaos and uncertainties introduced in the real world. Such unavoidable deviations render the use of virtual simulations ineffective for the calibration process of the algorithms that activate the restraint systems in the event of a crash (crash-detection algorithm). In the scope of this research, authors hypothesize the possibility of accounting for such variations introduced in the real world by creating a feedback loop between real-world crash tests and crash simulations. To accomplish this, a Reinforcement Learning (RL) compatible virtual surrogate model is used, which is adapted from crash simulation models. Hence, a conceptual methodology is illustrated in this paper for developing an RL-compatible model that can be trained using the results of crash simulations and crash tests. As the calibration of the crash-detection algorithm is fundamentally dependent upon the crash pulses, the scope of the expected output is limited to advancing the ability to estimate crash pulses. Furthermore, the real-time implementation of the methodology is illustrated using an actual vehicle model.