Data-Driven Fault Detection for Wafer Scanner Cable Slabs using Koopman Operators

The reliability of precision motion systems, such as semiconductor wafer scanners, is often influenced by nonlinear dynamics originating from components such as cable slabs. This paper introduces a data-driven framework for early fault diagnosis in these systems. Koopman operator theory is employed to derive a linear state-space model from experimental data, capturing the complex, hysteretic behavior of the cable slab. This model serves as a digital twin, and by comparing its predictions with real-time sensor measurements, operational anomalies can be detected. A systematic process for selecting observable functions yields a high-fidelity model with a tracking error of approximately ±1% across the operational range. When the proposed approach is tested against a state-of-the-art neural network model, it demonstrates a 75.4% reduction in reaction force prediction error. The framework successfully identifies an injected sensor noise fault (SNR of 20) in just 0.35 s using only force data, validating its potential to improve wafer scanner reliability.