Filling Missing Surface Roughness Data for Grinding Process Using Physics-Guided Neural Network

Abstract

Missing data is often unavoidable in real-world datasets which may lead to poor data-driven analysis and unreliable prediction. Filling the missing data presents a nontrivial challenge in existing techniques. Current naïve methods for filling the missing data often lead to biased approximations or make false assumptions about the data and correlations of the data. This paper presents a physics-guided neural network (PGNN) method that integrates the deep learning method with the guidance of a physics-based model to effectively fill the missing data that caused by limitations in information collection. The PGNN method first determines the coefficients of a physics-based model of a certain process by using machine learning (ML)-based estimations instead of experiment-based calibration. The outputs of this estimated physics-based model will serve as additional inputs for the next-level deep learning model. A numerical case study is performed to demonstrate the effectiveness of the proposed method by comparing it with other standard data-driven methods. The simulation results show that the PGNN method can more accurately fill the missing information and capture the uncertainty in the environment.