What is the purpose of the model?
This example is a rail constrained at the lower end and impacted by a heavy plate on the top. This model is derived from a test which simplifies a crashbox in a car. A crashbox is desired to have a stable folding behaviour in order to ensure passenger safety.
In total 176 parameters were varied in a Design of Experiments (DOE) study with 1000 runs (for academic purposes). Most parameters originate from the weld spots keeping both sheets together. Additionally geometric tolerances, impact parameters, as well as hardening of the material were varied. The variation followed a normal distribution since it was assumed that prototype manufacturing aims to satisfy the specified values. Almost all models had non-matching meshes, which is an additional difficulty level for machine learning algorithms.
Why use Machine Learning for Post-Processing?
Nowadays a models performance is measured through certain values tracked by sensors. In crash it is essentially the passengers safety, thus the acceleration of head, forces on body, etc. To make sure the passenger is safe robustness studies are performed to check the influence of manufacturing and model uncertainty. These studies usually involve 20-200 simulation runs, of which some usually don't match the performance. Analysing manually why certain deformation patterns occur is practically impossible given entire car crashes.
What is the Machine Learning approach?
In crash we currently evaluate models from single sensors and neglect the rest of information in the system. Its similar to analyzing an image from a few, hand-selected pixels. Machine learning and Artificial Intellgence give us the capabilities to use all pixels of an image and harvest it's information in a more complete approach.
How to compare 1000 simulation runs?
By analyzing the similarity of simulation result fields, we are able to determine not only the common behaviour of the simulation model, but also detect outliers. To illustrate our findings, we provide a variety of comprehensible and interactive plots, such as the similarity plot below, where every marker represents a simulation result. The closer markers are, the more similar their results have been. Discover the plot yourself and hover over the markers to see an image of the simulation.
This crashbox has the issue, that folding is not initiated nicely at the upper section of impact, but also at the lower end, which may cause further issues in a car crash.
How to investigate cause and effect?
Finding causes for certain types of deformation or failing targets can be a tedious task when done manually. Machine Learning (ML) can detect such dependencies in vast amounts of data. We therefore use our own Rule Mining Engine, which is designed specifically for simulation data, where one has only few samples in combination with a lot of variables. The originally complex and high-dimensional dependencies are simplified by the engine, so that an engineer can comprehend the recommendations. The outputs are understandable rules, as shown in the App below. Click on the rules to visualize them.
The rules below recommend the engineer to control the angle of impact to deviate from 0 to ensure stable folding initiation at the upper end. This can be enforced by inclining the upper angle of the plate. Note that the rules where found despite having only ~20 simulations with the undesired behaviour.
What to do against buckling_bottom?14:01
You have 2 reliable options:
- 0.333 ≤ impact_angle < 7.241 (100%)
- -5.983 ≤ impact_angle < -0.400 (100%)
Do you have questions regarding Artificial Intelligence?
We are happy to help you out: