Abstract:
Diagnosing characteristic industrial equipment characteristic behavior non-invasively
and in situ is an emerging field of study. An algorithm was developed to acoustically
monitor mechanical systems with minimal data labels. The methodology was evaluated
using a semiconductor device manufacturing process, consisting of a Selective Compliance
Assembly Robot Arm (SCARA) system, via an embedded microphone array. Combined unsupervised
and supervised data analysis techniques to identify critical processes for eventual
life-cycle tracking, was demonstrated. A spectrogram-based convolutional neural network
performed primary robotic motion segmentation with an average accuracy of 85% using
ground-truth validation data. Subsequent unsupervised analysis using similarity metrics
as well as k-means clustering on engineered features had mixed success in distinguishing
secondary robotic actuations. A semi-supervised technique was viable to differentiate
characteristics in robotic motions with limited available labeled data. Data visualizations
demonstrated potential limitations in engineered feature separability as well as probable
error sources. Further refinement is required for better segmentation accuracy as well as
identifying features that represent secondary characteristics in manufacturing systems.
