[5K2-OS-38a] OS-38

12:15 PM - 12:30 PM JST(3:15 AM - 3:30 AM UTC)

[5K2-OS-38a-02] Root Cause Estimation of Faults based on High-Dimensional Inspection Data in Optical System Assembly Processes

〇Heya Ouyang¹, Tatsuya Kubota¹, Mineyuki Nishino¹, Chiaki Koike¹, Yosuke Otsubo¹, Naoya Otani¹, Masashi Sugiyama² (1. NIKON CORPORATION, 2. RIKEN Center for Advanced Intelligence Project)

Keywords:

Assembly Processes,Root Cause Analysis,High-Dimensional Data,Approximate Bayesian Computation,Density Ratio Estimation

The assembly of optical products demands high precision, and improving productivity requires the rapid identification of fault causes. However, the root causes of faults detected during inspection often lie in preceding processes where causal data is unavailable, leading to a reliance on human expertise. While we proposed a cause estimation method using simulation-based Approximate Bayesian Computation, it suffers from instability when applied to the high-dimensional, small-sample data characteristic of these processes. This study extends this method to achieve high-precision estimation with improved stability and efficiency. The enhancements include generating a prior distribution based on domain knowledge, refining the process of density ratio estimation, and performing dimensionality reduction using Singular Value Decomposition of the sensitivity matrix. We validated the proposed method using both artificial data simulating the assembly process and real-world data, demonstrating that it identifies potential fault causes consistent with expert knowledge.