Detecting anomalous motions in ultraprecision shell-polishing process combining unsupervised spectral-band identification and Explainable-AI

This study presents a multi-sensor fusion approach to monitor the fine-abrasive polishing process for finishing small, diamond-coated spherical Si shells. These shells serve as fuel capsules for the Inertial Confinement Nuclear Fusion (ICF) experiments at Lawrence Livermore National Laboratory (LLNL). The success of the nuclear fusion reaction to an ignition stage critically depends on ensuring the ultrafine surface finish of the shell surfaces, mostly devoid of microscale pits. However, collisions among shells during polishing may damage their surface quality. Contemporary sensor-based and imaging monitoring methods have severe shortcomings in tracking the interactions of highly occluded and dynamically complex polishing processes. Towards addressing this challenge, we equipped the polishing setup with a vibration sensor and a camera to synchronously monitor the shell interactions to 0.1 s precision. We employed a bayesian optimization approach, EGO-MDA, to identify frequency bands that differentiate between the energy levels when the shells are together and far apart. Our findings reveal that an two frequency bands of vibration signals in 1–2.5 KHz range suffices to achieve this classification, with an accuracy of 80%. We adapted the Local Interpretation Model-agnostic Explanations (LIME) to identify the important frequency signature from the vibration signals that can identify the onset of a prominent collision event. Our findings indicate that vibration signals exhibit a 25% higher energy over a 0.1 s interval in the 1.3 KHz and 2.25 KHz bands at the time of the onset of a prominent collision event. These results establish the feasibility of employing vibration sensors to monitor severe interaction events.