Enhancement of virtual data quality using pre-trained Bayesian transfer learning under inaccurate and insufficient measurement data

The accurate application of designed dimensions in manufacturing is crucial but complicated by certain limitations in the manufacturing process. Therefore, manufacturing and assembly conditions should be considered in product designs. However, predicting and managing the deformation quality of industrial products, composed of multiple components, is often challenging owing to part deformation during the manufacturing process and structural deformation at the assembly stage. Owing to the advancements in artificial intelligence, highly precise predictive models can be created if sufficient data are available. However, the acquisition of extensive labeled data required for model training can be expensive and time-consuming in the industrial sector. Traditional approaches such as computer-aided engineering, which are commonly used to produce virtual data, have limitations owing to uncertainties in the experimental data, making consistency correction both time-consuming and expensive. This study proposes a method for enhancing the quality of virtual data created in contexts with insufficient and inaccurate experimental data. Engineering properties were extracted from experimental data and integrated with transfer learning using a pre-trained Bayesian neural network. We applied the proposed method to a micro-LED case to predict assembly and manufacturing variations during the design phase under inaccurate and insufficient data. Consequently, the accuracy was significantly improved over the predictive model that predicted the amount of assembly deformation using only experimental data, resulting in a predictive model with an average R-squared prediction accuracy of 98.9% for the deformation resulting from the assembly.