To detect abnormal patterns and predict remaining time to failure of semiconductor processing equipment ahead of time to minimize downtime losses and excessive maintenance costs
We formulated the problem as semi-supervised novelty detection to overcome lack of failure samples in production. On top of novelty detection results, we developed a method to estimate time-to-failures (TTF) of semiconductor processing equipment. To adapt to production environment change, a continual learning scheme was developed as well, and is now ready to apply.
Improved Time-to-Failure prediction with 90% + accuracy and less than 1% false alarm rate