Time Series Data-Driven Online Prognosis of Wind Turbine Faults in Presence of SCADA Data Loss

Abstract

While recent years have witnessed the high potential of various machine learning based schemes in wind turbine fault prognosis (WTFP), none of them effectively tackle the challenging yet practical issue of severe operational data loss during WTFP. To address this issue, this paper develops an intelligent time series (TS) data analytics approach for missing data-tolerant WTFP. Instead of struggling to impute missing data, a powerful TS subsequence nearest neighbor (SNN) profiling technique is introduced to directly characterize TS data with missing values and transform them into discerning SNN profiles. Taking such profiles as inputs, the long short-term memory based recurrent network is further leveraged to perform in-depth faulty feature learning from temporal evolution trends. This eventually results in a discriminative multi-fault classification model, which can implement reliable online WTFP in severe missing data contexts. Experimental test results with field data collected from an actual wind turbine demonstrate the efficacy of the proposed approach.