The machine learning (ML) for classification has been used widely in the engineering design for mapping from response to design variables, such as the feasible domain recognition and hidden pattern discovering. To train an ML model sufficiently requires big data to fill each corner of the design space. However, it is expensive to generate a big dataset for ML training by the frequently used numerical simulations, especially for complex systems. After fitted by a relatively small dataset, the surrogate model (SM), as the substitute of expensive simulations, can generate a big dataset efficiently. This big dataset may benefit the ML classification accuracy. However, the noise and uncertainty introduced by the SM approximation may degrade data quality and then the ML accuracy. Their complex effect has not been explored. In this study, a preliminary discussion is made to explore if the SM can provide more information for classification-based engineering design. A classification-based data mining method is introduced based on the classification and regression tree (CART). This method could recognize the design subspace efficiently and generate explicit decision-making rules for the design. Using the CART as the basic algorithm, an energy-absorbing structure under impact load is designed. The structure is parameterized and design datasets with different sizes are synthesized by the design of experiment (DoE). Their energy-absorbing responses are evaluated by the finite element (FE) method to form the FE datasets. One another fixed-size dataset is also generated as the FE validation dataset. Each FE dataset is applied to train a Gaussian process regression (GPR) model as the surrogate model. For each GPR model (corresponding to each FE dataset), several datasets with different sizes are generated by DoE with responses predicted by this GPR models to form SM datasets. Each FE dataset and corresponding SM datasets are adopted to train the CART with the classification accuracy validated by the fixed-size FE validation dataset. The comparison is made between the FE dataset and corresponding SM datasets for their CART accuracy. The results showed that the CART-based design method could help to recognize preferable design domains efficiently and generated clear and brief decision-making rules for generating designs with anticipated performance. After training by a small dataset, the surrogate model can provide more information for this CART-based design method by generating a big dataset efficiently. Especially for a pretty small FE dataset, the CART accuracy is increased significantly (by 33.7% at maximum), which is a great improvement for applying the data mining method to the engineering design, especially for expensive complex systems. The side-effect of noise and uncertainty induced by the surrogate modeling can be eliminated by a big enough and low-cost SM dataset.

 


Generating Pseudo-Data to Enhance the Performance of Classification-Based Engineering Design: A Preliminary Investigation

Division / Sponsors: 
Aerospace, Bioengineering, Applied Mechanics, Design Engineering Division, Advanced Energy Systems, Manufacturing Engineering Division, Noise Control and Acoustics, Nondestructive Evaluation, Diagnosis, and Prognosis (NDPD), Nuclear Engineering Division, Process Industries Division, Fluids Engineering, Heat Transfer, Materials, Microelectromechanical Systems, Safety Engineering and Risk Analysis, Computers & Information in Engineering


Paper Proceedings: https://asmedigitalcollection.asme.org/IMECE/IMECE2020/volume/84478

IMECE2020 - Virtual Conference

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.