Data-driven deviation generation for non-ideal surfaces of Skin Model Shapes

The digitalization techniques enable quality control as well as performance simulation of mechanical products in the Digital Twin era. In the product design process, realistic models regarding surface geometrical deviations are essential for further functional analysis. In the context of ISO standards on geometrical product specifications and verification (GPS), the Skin Model Shapes (SMSs) concept stemmed from the Skin Models paradigm is put forward to represent deviations of mechanical parts in accordance with the nature of geometric deviations on the surfaces. Contributions have been made to generate SMSs by combining both systematic deviations and random deviations on the nominal model. However, the non-ideal surface generation process is limited by the reuse of geometric deviation's knowledge. Existing geometric deviations that are predicted or observed on the parts cannot be applied for other SMSs generation without new parameter settings using different methods. In this paper, a framework for data-driven geometric deviation generation is proposed for non-ideal surface modeling in tolerancing. A database that contains a variety of deviations generated by different methods is constructed. Deviations that are classified as systematic and random types are represented and stored as samples in the dataset. Deviation pattern identification is addressed by transfer learning using AlexNet and it is used for different types of non-ideal surface generation. Different approaches regarding both prediction and observation stage within SMS schema are included for constructing the deviation samples. The mapping process for both systematic and random deviation samples are considered. Therefore, there is no need to re-design deviations with parameters for the surfaces. A case study is used to illustrate the non-ideal surface generation process by the propose method.