Sequential Detection Techniques for Machine Degradation Monitoring

Research Update for 2021-22 15th July 2022 Our work addresses inference problems in a networked factory environment where sensors monitor rotating machines’ health parameters and communicate their measurements to a controller over a wireless network prone to congestion and packet drops. The controller then makes decisions on the health of the machines based on the data it receives. Early detection of trends to failure in machines helps prevent cascading failures in assembly lines. We want to design procedures to detect incipient machine failures in this setting with minimum delay and perform well in a networked decision-making system. We model the rotating machine’s transition to a faulty state as a change point in a stochastic system and design quickest change-point detection procedures. The focus of our work has been on developing parametric quickest change-point detection algorithms and analyzing their performance. We study the mechanical vibrations from rolling element bearings since faults in bearings are the most common causes of failures in rotating machines. We consider parametric models that describe the vibration signals from bearings, motivated by real-world bearing vibration signal datasets. The parameters of the model change at a random time when the bearing transits to a faulty state from a normal state. We develop sequential parametric change detection tests where the pre-change parameters are known and the post-change parameters are unknown. Our generalized likelihood ratio (GLR) type test is designed to be analytically tractable, implementable with low complexity, and extendable to a multiple-stream change point detection scenario. We study the performance of our detection methods by analysis and a simulation study. We also compare the performance of these tests against existing methods in the literature and show betterment in performance. A summary of the research contributions during the period of the fellowship is listed below. • We improve the performance of our defect detection algorithm using the second order spectral properties of the bearing vibration signal • We model the second order spectral properties of the vibration signal before and after the onset of a defect using parametric distributions • We setup this as a composite quickest change point detection algorithm with unknown post-change parameters and use the earlier developed theory of a sub-optimal GLR-CUSUM based detector to design algorithms.

• We perform a simulation study of the performance of our test. We generate synthetic data that models bearing vibration signals and analyze the performance of our scheme on these signals. We also demonstrate the efficiency of our algorithm on real-world bearing vibration signals from datasets and compare performance against popular algorithms in the literature.