Prolego: Time-Series Analysis for Predicting Failures in Complex Systems

Failures in large, complex systems can be difficult to diagnose and expensive for both the system maintainers and users. We present techniques for predicting failures when there is sparse ground truth in sensor logs and usage modes of the system evolve rapidly. We demonstrate our methods using real-world logs from three different systems. Our method achieves over 80% prediction accuracy for various amounts of lead time.