A novel prognostics solution for accurate identification of degradation patterns in turbo machines with variable observation window

Abstract

The degradation of a system is a time bound phenomenon, which leads to the deterioration of turbomachinery, in terms of performance and reliability. If undetected and not acted upon in time, this could also lead to sudden system failure, resulting in unplanned unit downtime and maintenance. Unplanned downtime of a turbomachine leads to severe production loss for the end customer and consequent economic damages. Early detection of a degradation pattern would provide the customer with the opportunity to timely carry out corrective actions, preventing an unscheduled down time. The paper evaluates degradation identification methodology currently known from literature and finds them not accurate enough for general purpose application required by the solution. The paper discusses a novel methodology which can accurately detect degradation patterns of timeseries data. Critical features of this methodology are novel time-based correlation enabled regression model with variable observation window, autonomous training, and automatic adjusting capability to incorporate operating behavior change or physical system replacement. This leads to high accuracy, high generalization, and domain agnostic application capability. Moreover, particular focus is given to achieving high probability of detection and a low probability of false alarm. The paper demonstrates the performance achieved by the methodology when applied to the field of prognostics and diagnostics of IoT connected turbomachines through 50+ real application cases.