Virtual product development especially in car development requires the evaluation of multiple sensors signals in the simulations as one of the tasks; the sensor data is also needed for comparison with the real product. Comparing many virtual sensors manually from many simulations turns out to be a time consuming and challenging task. We propose a methodology and workflow setting that address this challenge, allowing a similarity comparison of hundreds of sensors from hundreds of simulations detecting similar events (clusters) or very different behavior as outliers. The approach uses a method of dimensionality reduction combined with different type of clustering methods including hierarchical clustering. The dimensionality reduction reduces the virtual sensor data information such that a visual comparison of thousand sensor signals can easily be performed in 3D, the hierarchical clustering on the other hand allows a localized comparison of sensor signals. The approach is demonstrated using binout Ls-Dyna data from a frontal crash example with many model variants containing many sensor data per simulation as well as for head impact computation.
To cope up with the ever growing amount of simulation runs being performed, tools and techniques are needed to store the huge amount of simulation data and to make use of data being stored. While current Simulation Data Management systems allows managing and accessing datasets and would facilitate putting this into action for analysis, the demand on bandwidth and storage increases. Even with SPDMs, the users usually only had tools and time to make rather straight forward model to model comparisons, between current model versions and their immediate predecessors. To take analysis capabilities and model development a leap forward, it is necessary to also make use of whole model development branches to learn from the gathered simulation information. With the availability of such tools, the value of past simulation data increases.
The car of the future thinks ahead - and, based on the analysis and processing of data, can do far more than humans alone behind the wheel. Radar, lidar sensors as well as cameras, already collect and evaluate large amounts of information in real time and detect hazards such as black ice, objects such as stationary cars or the ends of traffic jams. If desired, the automobile of tomorrow will even drive autonomously. But it will be some time before all offline vehicles have disappeared from the roads and most of all cars are navigating through traffic completely autonomously. A study by the Prognos research institute for the ADAC [1] shows that autonomous driving is not expected to become established until 2040. By the time the first driverless cars are on the roads in Germany, both the volume and variety of data will have exploded. According to estimates by the international market research and consulting firm IDC [2], the volume of data worldwide will grow to as much as 143 Zbytes by 2024. And an end to the flood of data is not in view. In its latest study, IDC forecasts a globally generated data volume of around 284 Zbytes by 2027.
Machine learning (ML) approaches for geometric part recognition have been evaluated with 3D automotive data in [1], where only one vehicle was used (Toyota Yaris with around 200 parts) and the exact match was tested, which means that the model was able to identify only the particular part shown regardless of the other classes (one-to-one match).