NEURAL NETWORK-BASED DESIGN OF AUTOMOBILES CONSIDERING SENSORY INFORMATION

点击翻译按钮获取中文摘要

In this research, we develop a new method for multi-objective satisfactory design of artifacts such as automobiles. In the method, we first specify a certain quantitative level of sensory information such as subjective evaluation for handling and stability of vehicle, and then determine sets of design variables which satisfy the level of design requirement. Another key issue here is how to derive a quantitative relationship between sensory information and physical response.In the first year, we performed experiments on subjective evaluation for handling and stability of vehicles in MAZDA's driving course using about 30 kinds of commercial cars. We invented a neural network-based method to derive the quantitative relationship and satisfactory design solutions using the relationship. The method is successfully applied to the experimental data. In addition, we developed a software system to automatically perform a series of data processes which runs in the Windows environment. Using the system, the processing time can be dramatically shortened. In the last year, we also implemented a recurrent type neural network to enhance learning capability. The system is applied some other types of problems such as prediction of corrosion life time, and promising results are obtained.