Multi-fidelity data-driven optimization design framework for self-piercing riveting process parameters

Abstract

Due to its excellent performance, the self-piercing riveting (SPR) process has been widely used in auto body manufacturing in recent years. The need for efficient and reliable design of SPR process parameters is growing in industry. Traditional trial-and-error design methods rely on historical experience and can be limited by physical experimental materials and equipment. Data-driven methods rely on data accuracy and data volume, and data from a single source is often inadequate to meet design needs. To address these issues, a multi-fidelity data-driven optimization design framework is proposed in this paper. The most significant feature of this framework is the fusion of physical experiment data and simulation data to build surrogate models. In this framework, a modified optimal Latin hypercube sampling method and a multi-fidelity surrogate model based on transfer learning and neural networks are proposed for SPR process. This multi-fidelity surrogate model can use a very small amount of experiment data to modify the surrogate model built on simulation data based on transfer learning, so that the model predictions can more closely match the physical experiment results at a lower modeling cost. Benefit from this method, the proposed framework can balance the contradiction between design accuracy and development cost compared to a single-fidelity data-driven framework. The application cases show that the prediction errors of the multi-fidelity models are <0.1 mm for the key geometric parameters of the SPR joint. As verified by physical experiments, the rivets and dies selected by the framework are the optimal solutions within the optional range. To the authors' knowledge, this is the first time that a multi-fidelity modeling method has been introduced to the field of SPR process, to solve the process parameter optimization design problem. Further, the method is not limited to the use of SPR process but can be applied as a paradigm in other engineering optimization design problems, especially in joining process parameter design problems.