This content is not included in your SAE MOBILUS subscription, or you are not logged in.

Feature-Based Response Classification in Nonlinear Structural Design Simulations

Journal Article
ISSN: 2380-2162, e-ISSN: 2380-2170
Published July 24, 2018 by SAE International in United States
Feature-Based Response Classification in Nonlinear Structural Design Simulations
Citation: Andersson, N., Rinaldo, R., and Abrahamsson, T., "Feature-Based Response Classification in Nonlinear Structural Design Simulations," SAE Int. J. Veh. Dyn., Stab., and NVH 2(3):185-202, 2018,
Language: English


An applied system design analysis approach for automated processing and classification of simulated structural responses is presented. Deterministic and nonlinear dynamics are studied under ideal loading and low noise conditions to determine fundamental system properties, how they vary and possibly interact. Using powerful computer resources, large amounts of simulated raw data can be produced in a short period of time. Efficient tools for data processing and interpretation are then needed, but existing ones often require much manual preparation and direct human judgement. Thus, there is a need to develop techniques that help to treat more virtual prototype variants and efficiently extract useful information from them. For this, time signals are evaluated by methods commonly used within structural dynamics and statistical learning. A multi-level multi-frequency stimulus function is constructed and simulated response signals are combined into frequency domain functions. These are associated with qualitative system features, such as being periodic or aperiodic, linear or nonlinear and further into subcategories of nonlinear systems, such as fundamental, sub or super harmonic and even or odd order types. Appropriate classes are then determined from selected feature metrics and rules-of-thumb criteria. To automate the classification of large data sets, a support vector machine is trained on categorised responses to determine whether a single feature, or combinations of features, applies or not. The trained classifier can then efficiently process new sets of data and pick out cases that are associated with possible vibrational problems, which subsequently can be further analysed and understood. This article describes elements of the analysis, discuss the effectiveness of evaluated feature metrics, reports practical considerations and results from two separate training study examples.