Statistical methods or Experts Know-How
Plenty of methods to find anomalies in your data
- Signal Band and Statistics
- Differences to Reference or ML model
- Bivariate Relations
- Outlier Detection (IQR, grubbs, gesd, ...)
- Self-Organizing Maps
Deep Analysis
Detect Initial Penetrations or missing Contact Definitions over time
- detect missing contact defintions
- find bad contact parametrization
- initial penetrations might be known, but do they matter? Combine with other results like damage or strain
Mesh independent evaluations
- combine results and data from arbitrary entities (nodes, elements, parts, spotwelds, curves, ...)
Assessment of overall respones functions
- Intrusion
Time variant and load dependent assessment
Detect relevant Events at relevant Locations. Combine possible sensitivities with arbitrary results, e.g.
- Combine Events or model inaccuracies with load results -> Detect problems where they matter
- High quality meshes are crucial for valid results. Distorted elements at highly loaded areas may indicate relevant sensitivities
- Check if damage near to a torsional loaded spotweld is too high
- ...
Challenge
Finite Element Analysis is mass business at automotive OEMs with huge turnovers due to automated execution with help of Simulation Data Management Systems. High quality processes for the entire simulation chain are mandatory for
- „first time right“ with first hardware build phase,
- enabling less hardware prototypes.
Product development for passive safety is strongly based on simulation with 3 basic parts
- Finite Element Solver: The quality of the based algorithms is assured by use of proven software from commercial vendors and proper release management
- Material Properties are evaluated at an offline process from experts working in strict standards.
- Meshing and Model Assembly has often to be done on a manual step after each simulation run. In the assembly steps many developers, teams or even companies (OEM and subcontractors) are involved. This step is time critical and need high level experts which is often not available at the necessary scale.
Solution
The Andata Tool Expecator is our generic Tool for anomaly detection. The basic idea is that there is no single method covering all types of anomalies for the plausibilization of simulation results. Expectator uses easy to define, pragmatic and adaptive Expectations which can be seen as a kind of criterium the data must fullfill. The sum of single Expectations evaluated and presented in systematic and automatic manner leads to a powerful tool to support the operating engineer in his decisions. Definition of Expectation can vary from very easy to highly abstract/complex/nonlinear.
As a result of a Expectator check, the user gets a automatic generated report with structured and well-prepared data on the actual solution/loadcase and presents only relevant data. The report can be
- html for static use
- Integrated into Animator4 in Form of a Session file for interactive analyzes
- fully customized
The use of Expectator as a automatic quality check as post processing step enables more time for analyzing the simulation results instead of checking the data for plausibility. It speeds up the quality management in the simulation chain and enables generic, automated plausibilization of simulation results in SDM.
Expectator can also be interactively used with our Animator4 Matlab Live Coupling or as an embedded application which can be used as an automatic post processing step on a cluster machine.
References
- A. Kuhn: CAE & AI: An Overview and Field Report on Various Applications, NAFEMS Artificial Intelligence und Machine Learning in der CAE-basierten Simulation, 23. - 24. Oktober 2023, München
- T. Hinterdorfer, P. Eder, A. Kuhn, H.U. Mader: Systematic and automatic anomaly detection in crash simulation results