Abstract

Abstract

Precise knowledge of the load history of safety-relevant structures is a central aspect within the fatigue strength design of modern vehicles. Since the experimental measurement of load variables is complex and therefore associated with high costs, vehicles require estimation of these variables in order to design even more customer-orientedly in the future and thus consistently pursue sustainable lightweight construction. Hence the data measured by sensors in today's standard production vehicles is based on vehicle bus system signals which can be permanently retrieved. Due to the increasing availability of large quantities of recorded vehicle data, machine learning methods are moving into the focus of application. In this work, the implementation of Recurrent Neural Networks for the estimation of loadtime curves is investigated. In order to close existing gaps in the state of the art, successful concepts of machine learning for sequential data, such as speech processing, are to be transferred to this application case. Long Short-Term Memory cells [1] play a central role for this type of problem. In addition to the adaptation of the network architecture, the integration of engineering knowledge is pursued within the method development process in order to increase the quality of the model. Relevant input variables are specifically selected by feature engineering and new meaningful variables are generated by filtering. Statistical analysis is used to investigate the correlation of these input signals with the estimated quantities. The development of a robust load estimation takes place in the course of model development on the basis of the torque of the left-hand rear drive shaft. Results reveal that the Recurrent Neural Networks approach is justified in estimating the highly nonlinear load curve of a complexly loaded part ­ as a component of the dynamic system ­ by means of available sensor signals [2]. Subsequently, the model is validated using recorded measurement data for different chassis settings of the same vehicle. Finally, the transferability of the designed network configuration to other chassis components of the same vehicle is investigated and evaluated.

Abstract

In recent years, Topology Optimization (TO) has been increasingly gaining attention with the development of new constructing techniques. In the optimization process of a density-based TO method, the sensitivities of the design variables are often computed using numerical methods like Finite Element Method (FEM). Such operation is performed repetitively for tens or hundreds of iteration steps, therefore generating huge computational cost for large scale design scenarios. This paper proposes to accelerate TO by replacing the full-scale sensitivity analysis of FEM with a reduced-scale case, and adopting a deep neural network to map the reduced-scale sensitivity field back to fine scale. Three neural network models are trained and tested using training data generated by structures of three difference scales. The results show that the proposed network successfully reduced the time cost to a large extent, while preserving the topology of the optimized design.

Abstract